Nature Neuroscience December 2005

www.nature.com/natureneuroscience

© 2005 Nature Publishing Group http://www.nature.com/natureneuroscience

EDITORIAL OFFICE [email protected] 345 Park Avenue South, New York, NY 10010-1707 Tel: (212) 726 9319, Fax: (212) 696 0978 Editor: Sandra Aamodt Senior Editor: Kalyani Narasimhan Associate Editors: Cara Allen, I-han Chou, Annette Markus, Charvy Narain Copy Editor: Dorothy Moore, Lavanya Reddy Senior Production Editor: Nicole D. Fournier Production Editor: Jessica Iannuzzi Cover Design: Kimberly Caesar Editorial Assistant: Jessica Chen MANAGEMENT OFFICES NPG New York 345 Park Avenue South, New York, NY 10010-1707 Tel: (212) 726 9200, Fax: (212) 696 9006 Publisher: Beatrice Renault Executive Editor: Linda Miller Chief Technology Officer: Howard Ratner Global Head of Advertising: Fabien Savenay Head of Nature Research & Reviews Marketing: Sara Girard Assistant Marketing Manager: Amy Maurer Production Coordinator: Diane Temprano Associate Director, Web Publishing: Timo Hannay New Technology Director: Greg Suprock Head of Web Services: Anthony Barrera Web Production Editor: Edwina Hay NPG London The Macmillan Building, 4 Crinan Street, London N1 9XW Tel: 44 207 833 4000, Fax: 44 207 843 4996 Managing Director: Annette Thomas Publishing Director: Peter Collins Editor-in-Chief, Nature Publications: Philip Campbell Marketing Director: Della Sar NPG Tokyo Chiyoda Building 2-37 Ichigayatamachi, Shinjuku-ku, Tokyo 162-0843 Tel: 81 3 3267 8751, Fax: 81 3 3267 8746 Asia-Pacific Publisher: Antoine E. Bocquet Manager: Koichi Nakamura Senior Marketing Manager: Peter Yoshihara Asia-Pacific Sales Director: Kate Yoneyama Asia-Pacific Sales Manager: Rinoko Asami DISPLAY ADVERTISING

[email protected] (US/Canada) [email protected] (Europe) [email protected] (Japan) US Head of Display Advertising: Stephen Schwartz, Tel: (212) 726 9256, Fax: (212) 696 9481 Global Head of Display Advertising Sales: John Michael, Tel: 44 207 843 4960, Fax: 44 207 843 4996 Head of Display Advertising—Europe: Gerard Preston, Tel: 44 207 843 4960, Fax: 44 207 843 4996 Business Development Manager: Claire Hines, Tel: 44 207 843 4960, Fax: 44 207 843 4996 Asia-Pacific Sales Manager: Rinoko Asami, Tel: 81 3 3267 8751, Fax: 81 3 3267 8746 Western Region Sales Manager: George Lui, Tel: (415) 781 3804, Fax: (415) 781 3805 Sales Executives: New England: Sheila Reardon, Tel: (617) 399 4098, Fax: (617) 426 3717 New York, Mid-Atlantic, Southeast: Jim Breault, Tel: (212) 726 9334, Fax: (212) 696 9481 Midwest: Mike Rossi, Tel: (212) 726 9255, Fax: (212) 696 9481 Northwest: Mathieu DesJardins, Tel: (415) 781 6422, Fax: (415) 781 3805 Eastern England/Scotland/Italy/Spain/Israel: Matthew Wilkinson, Tel: 44 207 843 4960, Fax: 44 207 843 4749 South/West United Kingdom/Scandinavia/Holland: Marianne Boulakas, Tel: 44 207 843 4969, Fax: 44 207 843 4749 Northern Germany: Gerard Preston, Tel: 44 207 843 4960, Fax: 44 207 843 4749 Southern Germany/Austria/Switzerland/France/Belgium: Sabine Hugi-Fürst, Tel: 41 52761 3386, Fax: 41 52761 3419 NATUREJOBS

[email protected] (US/Canada) [email protected] (Europe) [email protected] (Japan) Publisher: Ben Crowe, Tel: (212) 726 9245, Fax: (212) 696 9482 US Sales Manager: Peter Bless, Tel: (212) 726 9248, Fax: (212) 696 9482 Japan Sales Manager: Rinoko Asami, Tel: 81 3 3267 8751, Fax: 81 3 3267 8746 naturejobs Sales Director: Nevin Bayoumi, Tel: 44 207 843 4961, Fax: 44 207 843 4996 SITE LICENSE BUSINESS UNIT Americas: Tel: (888) 331 6288 Asia/Pacific: Tel: 81 3 3267 8751 Australia/New Zealand: Tel: 61 3 9825 1160 India: Tel: 00 91 11 324 4186 ROW: Tel: 44 207 843 4609

[email protected] [email protected] [email protected] [email protected] [email protected]

CUSTOMER SERVICE www.nature.com/help Senior Global Customer Service Manager: Gerald Coppin For all print and online assistance, please visit www.nature.com/help Purchase subscriptions: Americas: Nature Neuroscience, Subscription Dept., 303 Park Avenue South #1280, New York, NY 10010-3601. Tel: (866) 363 7860, Fax: (212) 689 9108 Europe/ROW: Nature Neuroscience, Subscription Dept., Macmillan Magazines Ltd., Brunel Road, Houndmills, Basingstoke RG21 6XS, United Kingdom. Tel: 44 1256 329 242, Fax: 44 1256 812 358 Japan: Nature Neuroscience, Nature Japan K.K., MG Ichigaya Bldg. 5F, 19-1 Haraikatamachi, Shinjuku-ku, Tokyo 162-0841. Tel: 81 3 3267 8751, Fax: 81 3 3267 8746 India: Harpal Singh Gill, Macmillan Magazines Ltd, 5A/12 Ansari Road, Darya Ganj, New Delhi, 110 002 India. Tel: 00 91 11 324 4186, Tel/Fax: 00 91 11 327 2010 REPRINTS [email protected] Nature Neuroscience Reprint Department, Nature Publishing Group, 345 Park Avenue South, New York, NY 10010-1707, USA. For commercial reprint orders of 600 or more, please contact: UK Reprints Sales Executive: Christine Fothergill, Tel: 44 207 843 4967, Fax: 44 207 843 4749 US Reprints Sales Executive: Sharda Tulsie, Tel: (212) 726 9631, Fax: (212) 679 0843


VOLUME 8 NUMBER 12 DECEMBER 2005

E D I TO R I A L 1627 The practice of theoretical neuroscience

BOOK REVIEW To understand the brain, theoretical and experimental approaches must be integrated to make sense of the enormous amount of existing data, and to guide future experiments. In this issue, we present a special focus on computational and systems neuroscience. Along with commissioned perspectives, the focus contains primary research articles featuring the best work presented at Cosyne, a meeting that brings together a diverse group of theoretical and experimental neuroscientists. Image of glass brain courtesy of Bret Lobree. (pp 1643–1656 and 1667–1711)

1629 Sensation & Perception by Jeremy M Wolfe, Keith R Kluender, Dennis M Levi, Linda M Bartoshuk, Rachel S Herz, Roberta L Klatzky & Susan J Lederman Reviewed by David Burr

NEWS AND VIEWS 1631 Neural mechanisms of attention and control: losing our inhibitions? Sander Nieuwenhuis & Nick Yeung see also p 1784 1633 Flipping the switch from electrical to chemical communication Karl Kandler and Edda Thiels see also p 1720 1635 Tinker to Evers to Chance: semaphorin signaling takes teamwork Paul A Garrity see also p 1712 1636 Did you feel that? Charvy Narain see also p 1698 1637 Odor here, odor there: chemosensation and reproductive function Eric B Keverne see also p 1660 1639 Nematodes learn: now what? William G Quinn

P E R S P E C T I V E S : C O M P U TAT I O N A N D S Y S T E M S F O C U S 1643 A natural approach to studying vision Gidon Felsen & Yang Dan 1647 In praise of artifice Nicole C Rust & J Anthony Movshon

Gating of dendritic spike propagation (p 1667)

1651 Analyzing receptive fields, classification images and functional images: challenges with opportunities for synergy Jonathan D Victor

Nature Neuroscience (ISSN 1097-6256) is published monthly by Nature Publishing Group, a trading name of Nature America Inc. located at 345 Park Avenue South, New York, NY 10010-1707. Periodicals postage paid at New York, NY and additional mailing post offices. Editorial Office: 345 Park Avenue South, New York, NY 10010-1707. Tel: (212) 726 9319, Fax: (212) 696 0978. Annual subscription rates: USA/Canada: US$199 (personal), US$1,809 (institution). Canada add 7% GST #104911595RT001; Euro-zone: €271 (personal), €1,558 (institution); Rest of world (excluding China, Japan, Korea): £175 (personal), £1,005 (institution); Japan: Contact Nature Japan K.K., MG Ichigaya Building 5F, 19-1 Haraikatamachi, Shinjuku-ku, Tokyo 162-0841. Tel: 81 (03) 3267 8751, Fax: 81 (03) 3267 8746. POSTMASTER: Send address changes to Nature Neuroscience, Subscriptions Department, 303 Park Avenue South #1280, New York, NY 10010-3601. Authorization to photocopy material for internal or personal use, or internal or personal use of specific clients, is granted by Nature Publishing Group to libraries and others registered with the Copyright Clearance Center (CCC) Transactional Reporting Service, provided the relevant copyright fee is paid direct to CCC, 222 Rosewood Drive, Danvers, MA 01923, USA. Identification code for Nature Neuroscience: 1097-6256/04. Back issues: US$45, Canada add 7% for GST. CPC PUB AGREEMENT #40032744. Printed by Publishers Press, Inc., Lebanon Junction, KY, USA. Copyright © 2005 Nature Publishing Group. Printed in USA.

i



B R I E F C O M M U N I C AT I O N S 1657 Switch in a long-term depression mechanism in successful aging H-K Lee, S S Min, M Gallagher & A Kirkwood Spike timing in storage and recall (p 1677)

1660 Deficits in sexual and aggressive behaviors in Cnga2 mutant mice V S Mandiyan, J K Coats & N M Shah see also p 1637 1663 Dissociation between physical and mental number line bisection in right hemisphere brain damage F Doricchi, P Guariglia, M Gasparini & F Tomaiuolo

A R T I C L E S : C O M P U TAT I O N A N D S Y S T E M S F O C U S 1667 Conditional dendritic spike propagation following distal synaptic activation of hippocampal CA1 pyramidal neurons T Jarsky, A Roxin, W L Kath & N Spruston 1677 Matching storage and recall: hippocampal spike timing–dependent plasticity and phase response curves M Lengyel, J Kwag, O Paulsen & P Dayan 1684 Neural population coding of sound level adapts to stimulus statistics I Dean, N S Harper & D McAlpine 1690 Independence of luminance and contrast in natural images and in the early visual system V Mante, R A Frazor, V Bonin, W S Geisler & M Carandini 1698 Neuronal correlates of subjective sensory experience V de Lafuente & R Romo see also p 1636 Developmental loss of gap junctions requires NMDA receptors (pp 1633 and 1720)

1704 Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control N D Daw, Y Niv & P Dayan

ARTICLES 1712 A triggering role for the FERM domain containing GEF protein FARP2 in the signals for class 3 semaphorin-mediated axonal repulsion T Toyofuku, J Yoshida, T Sugimoto, H Zhang, A Kumanogoh, M Hori & H Kikutani see also p 1635 1720 NMDA receptors regulate developmental gap junction uncoupling via CREB signaling H Arumugam, X Liu, P J Colombo, R A Corrivieau & A B Belousov see also p 1633 1727 Regulation of neuronal morphology and function by the tumor suppressors Tsc1 and Tsc2 S F Tavazoie, V A Alvarez, D A Ridenour, D J Kwiatkowski & B L Sabatini

Main olfactory epithelium regulates sexual and aggressive behavior (pp 1637 and 1660)

NATURE NEUROSCIENCE

1735 The prolactin-releasing peptide antagonizes the opioid system through its receptor GPR10 P Laurent, J A J Becker, O Valverde, C Ledent, A Kerchove d’Exaerde, S N Schiffmann, R Maldonado, G Vassart & M Parmentier

iii



1742 K-ATP channels promote the differential degeneration of dopaminergic midbrain neurons B Liss, O Haeckel, J Wildmann, T Miki, S Seino & Jochen Roeper Tumor suppressor Tsc genes in neuronal growth and function (p 1727)

1752 BK channel β4 subunit reduces dentate gyrus excitability and protects against temporal lobe seizures R Brenner, Q H Chen, A Vilaythong, G M Toney, J L Noebels & R W Aldrich 1760 Synaptic background activity controls spike transfer from thalamus to cortex J Wolfart, D Debay, G Le Masson, A Destexhe & T Bal 1768 Prior experience of rotation is not required for recognizing objects seen from different angles G Wang, S Obama, W Yamashita, T Sugihara & K Tanaka 1776 Actual versus predicted memory formation: an fMRI study on successful learning and the judgment-of-learning Y-C Kao, E S Davis, J D E Gabrieli 1784 Cognitive control mechanisms resolve conflict through cortical amplification of task-relevant information T Egner & J Hirsh see also p 1631

1791 ERRATA AND CORRIGENDUM

N AT U R E N E U R O S C I E N C E C L A S S I F I E D Mechanisms of axon repulsion by class 3 semaphorins (pp 1635 and 1712)

See back pages

Prefrontal cortex amplifies task-relevant signals (pp 1631 and 1784)

NATURE NEUROSCIENCE

v


E D I TO R I A L

The practice of theoretical neuroscience

“I

n theory, there is no difference between theory and practice. But, in practice, there is,” wrote Nobel laureate chemist Manfred Eigen. In neuroscience, unfortunately, there remains a considerable difference between the two—particularly in the number of people who appreciate these different ways of doing research. Thus, we have devoted part of this issue to a special focus on research presented at the Computational and Systems Neuroscience (Cosyne) meeting earlier this year, in an effort to illustrate how theoretical and experimental approaches can work together to provide insight into brain function. Theory has developed a bit of a bad reputation among experimentalists. Many scientists are skeptical of claims based on simulated data, feeling that such efforts are too far removed from biology to be informative. Others question the utility of attempts to assign machine-like or—worse still—anthropomorphic operating principles to the brain. Many find the dense language of theoretical papers exhausting and are frustrated when straightforward biological principles are obfuscated by impenetrable math. Hard experimental evidence is the key to understanding the brain, such scientists say, so why indulge in these mental exercises? In reality, theory is an integral part of all good neuroscience papers—including experimental papers. Any good paper includes an intuitive framework for its results and why they came out the way they did. For example, a study identifying a new protein involved in long-term potentiation is nothing more than a disconnected data set without a mechanistic framework for how it interacts with other elements in the pathway and an intuition for the functional consequences of these interactions. ‘Theoretical’ papers simply formalize and explore these intuitions and mechanisms—sometimes leading to the conclusion that our initial, hand-waving explanations do not provide a good fit to the data. Good theories can synthesize large quantities of empirical data, distilling them to a few simple notions, and can establish quantitative relationships between individual observations. They can generate predictions that can serve to validate current and future experiments. Given the vast number of empirical studies being generated by the field and the sheer complexity of the brain, it is clear that theoretical approaches have great potential for making sense of the problem. What makes for a computational paper that is not only a good study but one that will have wide impact among experimental neuroscientists? Fundamentally, a good theory paper contains the same elements as any good paper in cellular, molecular, systems or cognitive neuroscience. The paper should have a thought-provoking new hypothesis that is of potential interest to a wide audience. The model should be rigorously tested. Is it robust to biological variability? Can the model be falsified, and does it survive that test? Results, such as network simulations, should be quantified and not just demonstrated qualitatively. As with any other neuroscience paper,

NATURE NEUROSCIENCE VOLUME 8 | NUMBER 12 | DECEMBER 2005

the hypothesis and assumptions should be reasonably constrained by available evidence. Theories that are motivated by biology are the ones that are most likely to be influential with biologists. Bold, abstract theories may turn out to be right in the end, but if there is no way to conceive of how the brain can implement them nor a way to test them experimentally in the near future, then the audience that may be influenced by the work diminishes. As with any paper, but particularly so for computation papers, the essentials must be presented in an intuitive way that can be grasped by scientists outside the field. This means keeping jargon to a minimum, and presenting arguments in sentences, not equations written out in words. Esoteric quantifiers such as ‘model-dependent statistic’ may be mathematically more elegant than ‘mean’ and ‘standard deviation’ (and sound more impressive), but a paper using the latter terms is far more likely to reach its audience. As programs in computational neuroscience and annual workshops such as the advanced computational neuroscience courses offered at Woods Hole and in the European Union flourish, an increasing number of theorists and biologists are becoming more facile with the language of the complementary approach and are coming to appreciate the value of integrating the two disciplines. However, both fields have a long way to go before it will be commonplace for them to proceed hand in hand. We feel that the papers presented in this special issue, which was put together by Associate Editor I-han Chou, exemplify the application of theory to empirical studies. In a departure from our usual focus format, which normally includes only commissioned reviews, the focus also features primary research papers highlighting the best work presented at the Cosyne meeting (http://www.cosyne.org). This meeting was held in March in Salt Lake City, Utah, and brought together a broad range of theorists and experimentalists interested in systems neuroscience. Reflecting the diversity of attendees at the meeting, the papers span a variety of topics and contain different degrees of theoretical formalization. Every research article in this special issue was subjected to our regular peer-review process. We applied our usual stringent editorial standards to each paper, and each one met the criteria for publication in a regular issue of Nature Neuroscience. To accompany these papers, we have also commissioned several perspectives on quantitative approaches to probing neural data. Gidon Felsen and Yang Dan discuss the merits of using natural scenes to expand our understanding of the visual system, whereas Nicole Rust and Tony Movshon counter with a piece extolling the approach of using synthetic stimuli. Jonathan Victor discusses data-analysis techniques applied to different experimental disciplines, and possible ways to translate them across fields. We hope that this focus will highlight the value of increased dialogue between theorists and experimentalists, and spur future integrative efforts.

1627


NEWS AND VIEWS

Neural mechanisms of attention and control: losing our inhibitions? Sander Nieuwenhuis & Nick Yeung How are we able to focus our attention on the task at hand while ignoring myriad distractions? An elegant neuroimaging study in this issue finds that, contrary to a widely held view, the prefrontal cortex implements attentional control by amplifying task-relevant information, rather than by inhibiting distracting stimuli. As you read this article, you focus your attention on the page in front of you and you try (we hope) to think about the meaning of these words. All around you are myriad distractions: the habitually loud conversation of your colleague across the hall, your email program indicating you have new mail and the beckoning sunshine outside your window. How are we able to focus our attention on the task at hand and ignore these distractions? What are the neural mechanisms involved? Egner and Hirsch1 present important new evidence relevant to these questions, using brain imaging methods to look directly at cortical representations of attended and ignored information, and at neural activity correlated with this attentional focusing. Specifically, their research addresses a fundamental question about how attentional focusing is implemented in the brain. Does attention operate by amplifying relevant information, by inhibiting distracting information or both? A widely held view is that a critical function in the brain is the suppression of cognitive processing that is irrelevant to the current task2. Another view proposes that regions within prefrontal cortex (PFC) may be specialized for supporting inhibition of irrelevant stimuli and inappropriate responses3. The concept of inhibition provides an attractively simple account of the complex and wide-ranging cognitive deficits observed following damage to the frontal lobes: failing to inhibit distracting information

Sander Nieuwenhuis is at the Department of Cognitive Psychology, Vrije Universiteit, Van der Boechorststraat 1, 1081 BT, Amsterdam, The Netherlands. Nick Yeung is at the Department of Psychology, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA. e-mail: [email protected] or [email protected].

Figure 1 Possible attentional functions of prefrontal cortex (PFC) in the task used by Egner and Hirsch1. On this example trial, participants are required to classify the face as that of an actor or a politician and to ignore the superimposed written name. The fusiform face area (FFA) recognizes Robert de Niro, resulting in a tendency to respond “actor”. However, the automatic processing of the name Mao Ze Dong by word-processing areas tends to activate the conflicting response “politician”, thereby causing interference in the choice process. The PFC could resolve this stimulus-induced conflict either by inhibiting taskirrelevant representations in word processing areas or by amplifying task-relevant representations in the FFA. The results of Egner and Hirsch provide strong support for amplification as the primary means by which PFC exerts attentional control. Also shown is the putative link between a conflict monitoring system in anterior cingulate cortex (ACC) and the attentional control system in PFC. This conflict-control loop is thought15 to mediate the conflict-adaptation effects exploited by Egner and Hirsch to vary participants’ levels of attentional focus.

or inappropriate responses would impair all but the simplest cognitive processes and would be expected to have particularly strong impact on higher-level cognitive processes of flexible planning, task switching, problem solving and decision making—the kinds of tasks classically affected by frontal damage4. According to this view, PFC-guided attention may operate primarily (or even exclusively) by suppressing the processing of irrelevant information.


However, the idea that inhibition is a principal control function of the PFC has not gone unchallenged. First, computational modeling studies show that PFC-guided amplification of task-relevant information may be sufficient to account for behavioral effects of frontal damage that, at a phenomenological level, seem to require inhibition5,6. Second, and more generally, it has been questioned whether inhibition, as used in the present

1631


NEWS AND VIEWS context, is a useful explanatory concept. To update Richard Gregory’s famous example7, a short-circuit in your desktop computer might conceivably cause random blotches of color to appear on your monitor, but you would not infer that the damaged region usually functions as a “color inhibition” system. Instead, it would be more accurate to conclude that the damaged region—the video card— previously performed detailed, sophisticated computations that resulted in the delivery of meaningful, rather than random, patterns of color to your screen. By analogy, although PFC damage is often followed by a reduced ability to inhibit unwanted thoughts and behaviors, it does not logically follow that one function of the PFC is ‘inhibition’; instead, it might be more appropriate to understand PFC function in terms of the complex computations that it usually performs to support effective, goaldirected thought and action. Egner and Hirsch’s new study complements these theoretical challenges by attempting to empirically distinguish between amplification and inhibition as the mechanism of attention. To address this issue, the authors used a modified version of the Stroop task. In the standard Stroop task, participants are required to name the ink color of a printed color word. Performance is typically worse if word and color are incongruent (for example, RED printed in green ink) than when the two attributes are congruent (RED printed in red ink). Critically, participants could, in principle, perform the Stroop task successfully by amplifying relevant information (ink color), inhibiting irrelevant information (word) or both. To determine which approach is actually implemented in the brain, Egner and Hirsch used a variant of the Stroop task in which subjects responded to superimposed faces and words (rather than to colors and words). They then used brain imaging to measure activity in the fusiform face area (FFA), an extrastriate visual area activated by face stimuli8, to provide an index of the degree to which face stimuli are processed when they are relevant or irrelevant to the task at hand. Participants in the study were presented with a series of stimuli, each composed of a face of a familiar actor or politician with the written name of another actor or politician superimposed (Fig. 1). For half of the experiment, participants were instructed to attend to the face and to ignore the written name; in the other half, they attended to the name and ignored the face. In both cases, their task was to indicate with a button press whether the relevant attribute (face or name) belonged to an actor or a politician. The face and name could be associated with the same

(congruent) response or with a different (incongruent) response. Thus, on congruent trials (for example, an actor’s face with another actor’s name), unintended processing of the irrelevant attribute would tend to activate the same response as that required for the relevant attribute. In contrast, on incongruent trials (an actor’s face with a politician’s name, or vice versa), unintended processing of the irrelevant attribute would tend to activate the incorrect response, inducing processing conflict as in typical Stroop designs (Fig. 1). Of particular interest to Egner and Hirsch were the well-documented ‘conflict adaptation’ effects that are observed in the Stroop task: levels of attention are generally increased on trials immediately following incongruent (conflicting) trials, as demonstrated by improved performance on such trials9,10. Thus, by classifying trials according to whether the previous trial was congruent or incongruent, Egner and Hirsch were able to distinguish between physically identical trials depending on whether they were associated with low or high attention. Using this new design, Egner and Hirsch were able to contrast the following critical predictions. If attention involves inhibition of task-irrelevant information, then one should observe decreased FFA activation under conditions of high attentional focus on the printed names (that is, when faces are irrelevant). Conversely, if attention involves amplification of task-relevant information, then one should observe increased FFA activation under conditions of high attentional control in the face-relevant task condition. The results were clear-cut: the neural response of the FFA to faces varied with the degree of attention paid to the faces, but only when faces served as target stimuli and not when they served as distracting stimuli. That is, when participants were responding to faces, heightened attention led to increased FFA activity. However, when participants were responding to words, heightened attention did not lead to a decrease in FFA activity. Thus, the results showed clear evidence for amplification of task-relevant information, but no evidence for inhibition of task-irrelevant information. Extending this analysis, Egner and Hirsch found evidence that the observed amplification effects in the FFA were uniquely mediated by PFC. The correlation between the activity patterns in the FFA and right lateral PFC (but not in other control areas) increased reliably when participants had to classify the faces under conditions of heightened attention (previous trial incongruent), as opposed to when faces were distractors. These context-dependent increments in functional connectivity between

the PFC and FFA are consistent with the proposed role for PFC in guiding attention in posterior brain areas11,12. Notably, the observed activation focus in PFC overlaps with regions previously implicated in inhibitory function3. Thus, taken together, the results present a strong challenge to the theory that PFC-guided attention operates primarily through the inhibition of irrelevant stimulus information. A key question for future research is whether corresponding principles of attentional amplification operate in the selection of task-relevant responses, or whether inhibitory processes are more critical in action selection than perceptual selection. Computational considerations suggest that inhibition may be more feasible in systems responsible for the control of movement than in systems responsible for processing perceptual stimuli. Given the massive range of distracting stimuli that could, in principle, appear in any particular situation, it may simply be impractical to predict and inhibit all possible perceptual distractions. In contrast, the range of available responses is typically far more tightly constrained, and thus inhibition of incorrect actions may be a computationally feasible strategy. Existing evidence suggests a role for inhibitory processes in response selection13, and future studies might profitably extend Egner and Hirsch’s clever methodology to investigate this hypothesis further. Specifically, such studies might address the question of whether focused attention reduces activation of the cortical representations of irrelevant responses. Looking beyond these questions of amplification versus inhibition in perceptual and response selection, the broader question remains of how PFC ‘knows’ or ‘decides’ which information is currently relevant. To return to Gregory’s instructive example7, we would learn little from a theory that merely states that the video card of a computer promotes correct displays or inhibit incorrect ones, with no explanation of the circuits and computations that support this function. Correspondingly, our theories of PFC functioning will remain incomplete to the extent that they focus solely on whether this region serves to amplify taskrelevant information or inhibit distractions. Instead, the ultimate goal of research in this area must be to understand the nature of the computations within PFC that might support these attentional functions14. Answering this deeper question will require the development of detailed computational theories that are capable of explaining how PFC comes to represent task-relevant information, and the development of correspondingly sophisticated neuroimaging methods capable of evaluating the predictions of these theories. 1. Egner, T. & Hirsch, J. Nat. Neurosci. 8, 1784–1790

1632

VOLUME 8 | NUMBER 12 | DECEMBER 2005 NATURE NEUROSCIENCE

NEWS AND VIEWS 45–77 (1992). 7. Gregory R.L. in Current Problems in Animal Behaviour (eds. Thorpe, W.H. & Zangwill, O.L.) 547–565 (Cambridge University Press, Cambridge, UK, 1961). 8. Kanwisher, N., McDermott, J. & Chun, M.M. J. Neurosci. 17, 4302–4311 (1997). 9. Gratton, G., Coles, M.G. & Donchin, E. J. Exp. Psychol. Gen. 121, 480–506 (1992). 10. Botvinick, M.M., Braver, T.S., Barch, D.M., Carter, C.S. & Cohen, J.D. Psychol. Rev. 108, 624–652 (2001). 11. Duncan, J. Nat. Rev. Neurosci. 2, 820–829

(2001). 12. Miller, E.K. & Cohen, J.D. Annu. Rev. Neurosci. 24, 167–202 (2001). 13. Burle, B., Vidal, F., Tandonnet, C. & Hasbroucq, T. Brain Cogn. 56, 153–164 (2004). 14. Rougier, N.P., Noelle, D.C., Braver, T.S., Cohen, J.D. & O’Reilly, R.C. Proc. Natl. Acad. Sci. USA 102, 7338–7343 (2005). 15. Kerns, J.G. et al. Science 303, 1023–1026 (2004).

Flipping the switch from electrical to chemical communication Karl Kandler & Edda Thiels Immature neurons in many brain regions are electrically coupled through gap junctions, which are lost as chemical synaptic transmission matures. This developmental uncoupling is now shown to require NMDA receptor activation. Early in brain development, neurons communicate with one another, even before synapses have formed. At this stage, electrical coupling through gap junctions is widespread and may contribute to neuronal and circuit maturation1. A transition from electrical to chemical synapses has been documented for multiple brain areas, but the cues and mechanisms that mediate the switch have remained elusive. In this issue, Arumugam and colleagues2 provide exciting new insight into the signaling pathways involved in this developmental regulation of gap-junction coupling in hypothalamic neurons. Gap junctions are specialized cell-cell contacts composed of membrane channels that join the cytosol of neighboring cells, allowing electrical current to flow between them3. Because they are permeable to small molecules (<1 kDa), such as the second messengers inositol 1,4,5-triphosphate (IP3) and cyclic AMP (cAMP), gap junctions provide a means to coordinate not only electrical activity but also metabolic and signaling processes in coupled cells4. Although neuronal gap junction coupling is not exclusive to the immature brain, in mammals it is most extensive and widespread before and during the time of synapse formation and occurs in many areas, including the retina, cortex, thalamus, brainstem and spi-

Karl Kandler and Edda Thiels are in the Department of Neurobiology and Center for the Neuronal Basis of Cognition, University of Pittsburgh School of Medicine, 3500 Terrace Street, Pittsburgh, Pennsylvania 15261, USA. email: [email protected] or [email protected]

NMDA receptor activation CaMKII/IV PKC CREB

NMDA receptor

Cx36

Ann Thomson


(2005). 2. Dagenbach, D. & Carr, T.H. (eds.) Inhibitory Processes in Attention, Memory, and Language. (Academic, San Diego, 1994). 3. Aron, A.R., Robbins, T.W. & Poldrack, R.A. Trends Cogn. Sci. 8, 170–177 (2004). 4. Stuss, D.T. & Benson, D.F. Psychol. Bull. 95, 3–28 (1984). 5. Kimberg, D. & Farah, M. J. Exp. Psychol. Gen. 122, 411–438 (1993). 6. Cohen, J.D. & Servan-Schreiber, D. Psychol. Rev. 99,

Glutamatergic synapse

Figure 1 NMDA receptor–mediated uncoupling of developing hypothalamic neurons. During synaptic circuit development in the medial hypothalamus and many other brain regions, the primary mode of neuronal communication switches from one based on gap junctions (left) to one based on chemical synapses (right). Arumugam and colleagues2 now show that downregulation of the neuronal gapjunction protein connexin36 (Cx36) and dye coupling (yellow) require activation of NMDA receptors, along with CaMKII/IV, PKC and CREB. In the intact brain, additional glutamatergic inputs are provided by other extrinsic sources.

nal cord3. Neurons joined by gap junctions form functional assemblies with coordinated patterns of spontaneous activity and changes in intracellular calcium levels5. These early gap junction–mediated activity patterns are thought to be important for neurogenesis, neuronal maturation and synaptogenesis. Arumugam and colleagues examined developmental changes in gap junction coupling in magnocellular neurons of the rat hypothalamus. They measured expression levels of the neuronal connexin protein Cx36 and looked for functional gap junctions by assessing the ability of coupled neurons to pass small dyes or neuronal tracers between them (dye coupling). By both measures, they


found that electrical coupling increases in vivo during the first two postnatal weeks and then decreases during the third and fourth weeks, a time of intense synapse formation (Fig. 1). Blocking NMDA receptors attenuates the loss of gap junction coupling in developing spinal motoneurons6, so the authors asked whether signaling through NMDA receptors might be important for the developmental decrease in coupling in the hypothalamus. They chronically blocked NMDA receptors in vivo by injecting newborn rats daily with the antagonist dizocilpine (MK-801), and they measured gap-junction coupling two and four weeks later. NMDA receptor blockade had no effect on the initial increase in cou-

1633


NEWS AND VIEWS pling during the first two postnatal weeks, but it significantly reduced subsequent uncoupling: in four-week-old animals treated with MK-801, the expression of Cx36 as well as the incidence of dye coupling were higher than in untreated controls. This effect seemed to be neuron specific because the expression of the protein Cx43, a connexin expressed by glia, was not altered by MK-801. These results extend the findings in spinal motoneurons6 and support the idea that NMDA receptor activation contributes to developmental uncoupling. However, systemic MK-801 treatment can disrupt other developmental processes—such as weight gain, motor development and synaptic refinement—that may indirectly affect uncoupling. To avoid these potential confounds, Arumugam and colleagues repeated the experiments in primary cultures of hypothalamic neurons, which showed a sequence of coupling and uncoupling similar to that observed in the intact brain. The NMDA receptor antagonist DL-2-amino-5-phosphonovalerate (AP5) did not affect the initial increase in coupling during the first 16 days in vitro, but like MK-801 treatment in vivo, AP5 nearly abolished uncoupling during the subsequent two weeks. For additional confirmation, the authors examined gap junction development in cultures from mice lacking the NMDA receptor subunit NR1, which is required for functional receptors. Electrical coupling failed to decrease after 16 days in vitro in these cultures. Having established a crucial role for NMDA receptors, Arumugam and colleagues asked which intracellular signaling pathways might link NMDA receptor activation to Cx36 downregulation and functional uncoupling. Using a battery of drugs, they identified a requirement for calcium/calmodulin–dependent protein kinases II and IV (CaMKII/IV) and protein kinase C (PKC), both known for contributing to activity-dependent neuronal plasticity. In contrast, cAMP-dependent protein kinase (PKA) and extracellular signal-regulated kinase/mitogenactivated protein kinase (ERK/MAPK), which are also involved in activity-dependent plasticity, were not required for uncoupling. Calcium influx through NMDA receptors seems to be sufficient to trigger CaMKII/IV- and PKC-mediated uncoupling, because it was not prevented by blocking ionotropic non-NMDA glutamate receptors or by blocking voltage-gated calcium channels. What lies downstream of CaMKII/IV and PKC to mediate uncoupling? Calcium-cAMP response element binding protein (CREB), a transcriptional activator, is important for NMDA receptor–dependent, long-term neu-

1634

ronal plasticity7–9 and was thus a likely suspect in uncoupling. Both members of the CaMK family and PKC can—either directly (CaMKs) or indirectly (PKC)—phosphorylate CREB at residue S133 and enhance its transcriptional capacity10. Using a three-pronged approach that included CREB overexpression, overexpression of a dominant-negative form of CREB and administration of CREB antisense oligodeoxynucleotides, the authors showed that more CREB accelerated uncoupling, whereas less or mutant CREB attenuated it. This inverse correlation between CREB protein levels and electrical coupling raises interesting questions about the relationship between PKC- or CaMK-dependent CREBmediated gene expression and gap junctions. For instance, do the changes in coupling following the various types of CREB manipulation that the authors observed result from alterations in Cx36 expression or function? The authors note that the promoter region of Cx36 includes a CREB binding motif. However, CREB binding motifs are found in hundreds of other genes, only a subset of which are actually known to be regulated by CREB; therefore, CREB binding motifs do not guarantee CREBregulated expression10. The present observation of a tight relationship between NMDA receptor activation and gap-junction uncoupling invites a more rigorous dissection of the molecular and cellular events involved. Another interesting issue that remains to be addressed concerns the dynamics of the reciprocal relationship between gap-junctions and chemical synapses. In identified motoneuron pairs of the snail Heliosoma, gap junction coupling increases when cholinergic synaptic transmission is blocked and decreases when chemical synaptic transmission is increased11. Is there a similar bidirectional interaction between NMDA receptors and gap junctions in hypothalamic or other mammalian neurons? For example, are they related in a self-perpetuating feedback mechanism, in which increased NMDA receptor activation is associated with decreased gap junction coupling, which in turn is associated with further increases in NMDA receptor activation? Careful examination of the time course of changes in NMDA receptor expression and activation and gap junction coupling, both across development and in response to experimental manipulations, should shed light on this pertinent issue and provide further information about underlying mechanisms. In their cultured hypothalamic neurons, Arumugam and colleagues also found that blocking sodium-dependent action potentials with tetrodotoxin (TTX) prevented the

developmental reduction in coupling, similar to the effect of NMDA receptor blockade. Although the specific mechanisms by which action potentials contribute to uncoupling still need to be investigated, one can think of at least three major possibilities. Action potentials may be required presynaptically for synchronized glutamate release and consequent NMDA receptor activation. Backpropagating action potentials that are present in many neurons12 might also be required for sufficient depolarization in postsynaptic cells to relieve the voltage-dependent Mg2+ block of NMDA receptors. Action potentials may be required to trigger a retrograde synaptic signal that causes the presynaptic neuron to uncouple from its neighbors13. In any event, the TTX results suggest the interesting possibility that the coordination and/or the propagation of action potentials in neuronal assemblies may contribute directly or indirectly to uncoupling. These new findings from Arumugam and colleagues constitute a big step forward in our understanding of the mechanisms that underlie early neuronal network formation. NMDA receptors are crucial for the refinement of synaptic connections throughout development. It now seems that NMDA receptors also flip the switch from electrical to chemical communication and thereby fundamentally change the mode of interneuronal communication. Many aspects concerning the nature of the link between NMDA receptor activation and gap junction loss await further investigation, and it will be interesting to know whether similar mechanisms contribute to circuit development in other brain regions. The exciting findings from Arumugam and colleagues provide a promising starting point for addressing these important issues. 1. Montoro, R.J. & Yuste, R. Brain Res. Brain Res. Rev. 47, 216–226 (2004). 2. Arumugam, H., Liu, X., Colombo, P.J., Corriveau, R.A. & Belousov, A.B. Nat. Neurosci. 8, 1720–1726 (2005). 3. Bennett, M.V. & Zukin, R.S. Neuron 41, 495–511 (2004). 4. Kandler, K. & Katz, L.C. J. Neurosci. 18, 1419–1427 (1998). 5. Yuste, R., Peinado, A. & Katz, L.C. Science 257, 665– 669 (1992). 6. Mentis, G.Z., Diaz, E., Moran, L.B. & Navarrete, R. J. Physiol. (Lond.) 544, 757–764 (2002). 7. Ghosh, A., Ginty, D.D., Bading, H. & Greenberg, M.E. J. Neurobiol. 25, 294–303 (1994). 8. Silva, A.J., Kogan, J.H., Frankland, P.W. & Kida, S. Annu. Rev. Neurosci. 21, 127–148 (1998). 9. Deisseroth, K., Bito, H. & Tsien, R.W. Neuron 16, 89– 101 (1996). 10. Lonze, B.E. & Ginty, D.D. Neuron 35, 605–623 (2002). 11. Szabo, T.M., Faber, D.S. & Zoran, M.J. J. Neurosci. 24, 112–120 (2004). 12. Waters, J., Schaefer, A. & Sakmann, B. Prog. Biophys. Mol. Biol. 87, 145–170 (2005). 13. Pastor, A.M., Mentis, G.Z., de la Cruz, R.R., Diaz, E. & Navarrete, R. J. Neurophysiol. 89, 793–805 (2003).


NEWS AND VIEWS

Tinker to Evers to Chance: semaphorin signaling takes teamwork © 2005 Nature Publishing Group http://www.nature.com/natureneuroscience

Paul A Garrity The Rho-family GTPases act downstream of axon guidance receptors, controlling proteins that remodel the cytoskeleton. New work now suggests that these GTPases also team up to regulate activation of the semaphorin-3A receptor itself. The mind-boggling complexity of the brain’s wiring pattern emerges one growth cone navigational decision at a time. The semaphorin family of signaling proteins exerts a critical influence over many of these decisions, often providing extracellular cues that repel growth-cone advance1. Semaphorin receptors have been identified, but it remains largely mysterious how semaphorins activate their receptors and trigger the growth-cone responses that wire the nervous system. Semaphorin-3A, a secreted member of the semaphorin family, acts through a receptor complex containing the single-pass transmembrane proteins neuropilin-1 (which binds semaphorin-3A) and plexin-A1 (which transduces signals). A new paper in this issue2 suggests that activation of the neuropilin-1/plexin-A1 receptor by semaphorin-3A involves a sequence of deft interactions worthy of the classic Chicago Cubs’ double-play combination. Growth-cone responses to semaphorin-3A critically depend on the activities of small GTPases, including Rho-family GTPases Rac and Rnd1. This may not seem surprising, as Rho-family GTPases figure prominently in growth-cone responses to a host of guidance cues3. However, although Rho-family GTPases are usually pictured downstream of guidance receptors controlling effectors that remodel the cytoskeleton, recent work suggests that this is not the whole story, at least for semaphorin-3A. The new study by Toyofuku et al.2 now provides additional insight into how small GTPases remodel the semaphorin-3A receptor itself. The participation of one Rho-family GTPase in the activation of a semaphorin receptor was clarified when Rnd1 was shown to bind and activate the semaphorin-4D receptor plexin-B1 (ref. 4). Rnd1 also binds plexin-A1 and is implicated in activating this receptor as well5. Rac’s function in semaphorin signaling remained much less clear. Rac is

Paul A. Garrity is in the Department of Biology, Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, Massachusetts 02139-4307, USA. e-mail: [email protected]

activated by semaphorin-3A treatment6 and is important for semaphorin-3A–mediated growth-cone repulsion7,8. Furthermore, a mutant form of plexin-A1 that signals in the absence of semaphorin-3A does not require Rac to trigger growth-cone collapse6, suggesting that Rac may be involved in activation of plexin-A1 in response to semaphorin-3A binding. However, the mechanisms by which semaphorin-3A causes Rac activation and how Rac activation contributes to plexin-A1 activation were unknown. Toyofuku et al.2 initially address the mechanism of Rac activation by semaphorin-3A by examining the role of a previously characterized activator of Rac, the RacGEF FARP2 (ref. 9). Small GTPases like Rac are active in their GTP-bound state, but not in their GDP-bound state, and their transitions between states are usually under tight control. Small GTPases are turned on by guanine exchange factors (GEFs), which catalyze the exchange of GDP for GTP, and are turned off by GTPase activating proteins (GAPs), which accelerate the GTPases’ intrinsic ability to hydrolyze GTP. Toyofuku et al. use HEK cells and neurons cultured from chick dorsal root ganglia to show that FARP2 is a critical element of the semaphorin-3A receptor complex. Semaphorin-3A binding to the receptor causes FARP2 to dissociate from plexin-A1, activating FARP2’s RacGEF activity and transiently elevating the level of GTP-bound Rac. Thus, FARP2 provides a physical link between semaphorin-3A binding and Rac activation. How does Rac activation lead to semaphorin-3A–mediated growth-cone repulsion? To address this issue, the authors drew on insights into Rnd1 regulation of plexin-B1. The intracellular domain of plexin-B1, like plexin-A1, shows homology to GAPs for Ras-family GTPases. Binding of Rnd1 to plexin-B1 activates plexin-B1’s RasGAP activity and triggers RRas inactivation, contributing to repulsion4. In the current work, Toyofuku et al.2 show that Rnd1 also activates plexin-A1 RasGAP activity and promotes R-Ras inactivation in response to semaphorin-3A treatment. This observation raised an interesting possibility: could growth-cone repulsion by


semaphorin-3A involve the regulation of Rnd1 binding by FARP2 and Rac? Some steps remain hazy, but all signs point to yes. Although the evidence is indirect, the activation of Rac seems to trigger Rnd1 binding to plexin-A1, leading to growth-cone repulsion. First, the authors show that semaphorin-3A stimulation induces Rnd1 binding to plexin-A1. Second, they show that the disruption of FARP2’s RacGEF activity inhibits both semaphorin-3A–stimulated Rnd1 binding to plexin-A1 as well as semaphorin-3A–mediated repulsion. Because FARP2 GEF activity seems to specifically activate Rac and not other Rho GTPases, these data suggest that Rac activation may trigger Rnd1 binding. However, data that clearly resolve the role of Rac in Rnd1 binding are not available, and determining how Rac activation could trigger Rnd1 binding to plexin-A1 remains a critical question. This work supplies a plausible multistep scenario for FARP2’s involvement in semaphorin-3A signaling (Fig. 1). In this model, semaphorin-3A binds to neuropilin-1, releasing FARP2 from plexin-A1 and activating FARP2’s RacGEF. FARP2 presumably activates Rac, which in turn promotes the binding of Rnd1 to plexin-A1. Rnd1 binding then activates plexin-A1’s RasGAP activity, leading to the downregulation of R-Ras activity. This downregulation of R-Ras activity then contributes to growth-cone repulsion. Precisely how inhibition of R-Ras triggers growth-cone repulsion is an important open question. Current models suggest that integrin signaling is involved. R-Ras activation can promote integrin function10, and therefore R-Ras inhibition could facilitate repulsion by decreasing integrin signaling within the growth cone. Interestingly, semaphorin-3A inhibits integrin signaling in endothelial cells11. Here, Toyofuku et al. provide two interesting nuggets on the potential role of integrin signaling downstream of semaphorin-3A in neurons. First, consistent with a decrease in integrin signaling promoting repulsion, they demonstrate that artificial activation of integrins with an antibody partially suppresses the repulsive effects of semaphorin-3A on axons. Second, they coimmunoprecipitate

1635

NEWS AND VIEWS Plexin-A1 Integrins (on)

Sema3A

Integrins (off) 1

Npn-1 GDP

GDP

FARP2

Rac RasGAP

GTP

R-Ras GTP

Rnd1

RacGEF

3

GTP

GTP

GTP GTP

GTP

7

GTP

GTP

5

4

GTP

6

GDP

2

Repulsion Figure 1 A model for plexin-A1 activation and growth-cone repulsion by semaphorin-3A. Binding of semaphorin-3A (Sema3A) to neuropilin-1 (Npn-1; step 1) triggers FARP2 dissociation from plexin-A1 and activates FARP2’s RacGEF (step 2), leading to activation of Rac (step 3). Activation of Rac stimulates binding of Rnd1 to plexin-A1 (step 4), which activates plexin-A1’s latent RasGAP (step 5). The plexin-A1 RasGAP downregulates R-Ras (step 6), leading to inhibition of integrin function (step 7) and growth-cone repulsion. Further details of the model are described in the text.

FARP2 with PIPKIγ661, a phosphatidylinositol kinase that has been implicated in promoting the assembly of integrin-containing focal complexes12,13. Semaphorin-3A stimulation reduces PIPKIγ661 kinase activity, providing a second possible way for semaphorin-3A to inhibit integrin function. However, the significance of PIPKIγ661 in growth-cone guidance remains uncertain, as PIPKIγ661 knockdown had no detectable effect on repulsion. Nonetheless, these results are especially intriguing because MICALs—a family of plexin-A1–binding proteins that, like FARP2, are involved in semaphorin-3A–mediated repulsion14—interact with CasL, a focal adhesion protein involved in integrin-mediated motility. The route from the semaphorin-3A receptor complex to integrins could turn out to be well traveled. The model for semaphorin-3A receptor activation that emerges from this work raises important questions to address in future studies. First, how does Rac activation contribute to the binding of Rnd1 to

plexin-A1? Rac could interact directly with plexin-A1 to induce a conformational change needed for Rnd1 binding or, alternatively, could act more indirectly. Second, how does Rnd1 binding to plexin-A1 stimulate plexin-A1’s RasGAP activity? Because the Rnd1 binding site is located between the two regions that comprise the plexin-A1 RasGAP domain, one obvious model is that Rnd1 binding causes a conformational change in plexin-A1 that promotes RasGAP activity. Future biochemical and structural studies will undoubtedly clarify these issues. In addition, what are the critical molecular links between the semaphorin-3A receptor and integrins? It will be important to determine how inhibition of R-Ras decreases integrin function, and which aspects of semaphorin-3A–mediated repulsion involve changes in integrin signaling. Finally, the path for semaphorin-3A receptor activation uncovered by Toyofuku et al. suggests several molecules, such as FARP2, that could serve as targets through which other

signaling pathways could modulate a growth cone’s response to semaphorin-3A. The semaphorin-3A guidance receptor is shaping up to be a sophisticated signaling machine. 1. Pasterkamp, R.J. & Kolodkin, A.L. Curr. Opin. Neurobiol. 13, 79–89 (2003). 2. Toyofuku, T. et al. Nat. Neurosci. 8, 1712–1719 (2005). 3. Luo, L. Nat. Rev. Neurosci. 1, 173–180 (2000). 4. Oinuma, I., Ishikawa, Y., Katoh, H. & Negishi, M. Science 305, 862–865 (2004). 5. Zanata, S.M., Hovatta, I., Rohm, B. & Püschel, A.W. J. Neurosci. 22, 471–477 (2002). 6. Turner, L.J., Nicholls, S. & Hall, A. J. Biol. Chem. 279, 33199–33205 (2004). 7. Jin, Z. & Strittmatter, S.M. J. Neurosci. 17, 6256– 6263 (1997). 8. Kuhn, T.B., Brown, M.D., Wilcox, C.L., Raper, J.A. & Bamburg, J.R. J. Neurosci. 19, 1965–1975 (1999). 9. Kubo, T. et al. J. Neurosci. 22, 8504–8513 (2002). 10. Zhang, Z., Vuori, K., Wang, H., Reed, J.C. & Ruoslahti, E. Cell 85, 61–69 (1996). 11. Serini, G. et al. Nature 424, 391–397 (2003). 12. Ling, K., Doughman, R.L., Firestone, A.J., Bunce, M.W. & Anderson, R.A. Nature 420, 89–93 (2002). 13. Di Paolo, G. et al. Nature 420, 85–89 (2002). 14. Terman, J.R., Mao, T., Pasterkamp, R.J., Yu, H.H. & Kolodkin, A.L. Cell 109, 887–900 (2002).

Did you feel that? How does subjective experience correlate with neural activity? Activity in early sensory cortex reflects accuracy in discriminating different stimuli, and is thought to be key in determining perceptual judgments. Now a study on page 1698 of this issue shows that in monkeys judging stimuli at the threshold of detection, activity in frontal rather than sensory cortex correlates with subjective experience. Lafuente and Romo recorded single-neuron responses from primary somatosensory and medial premotor cortex in monkeys trained to report the presence or absence of a mechanical vibration applied to their fingertips. The strength of the vibration varied, and it was sometimes just at the threshold of detection, so that the monkeys made mistakes. Activity in the somatosensory cortex accurately reflected stimulus strength, but it did not correlate with the monkeys’ perceptual reports. Medial premotor cortex activity was independent of stimulus strength, but covaried with the monkeys’ subjective judgments, even when they gave an incorrect response. Electrically stimulating the medial premotor cortex caused the monkeys to respond almost exactly as they did to real mechanical stimulation. These results suggest that somatosensory cortex represents the physical properties of a stimulus, whereas frontal neurons are more involved in subjective perceptual judgments. Charvy Narain

Photo credit: Index Stock Imagery


GTP

1636


NEWS AND VIEWS

Odor here, odor there: chemosensation and reproductive function Pheromones acting through the vomeronasal system influence sexual behavior and neuroendocrine function. Two new studies show that the main olfactory system also contributes to behavioral and possibly endocrine regulation. In science, views are continually shifting. Nowhere is this better illustrated than in the study of mammalian pheromones and their influence on behavior and endocrine systems. The earliest proposals suggested the main olfactory system was important for chemosensory activation of behavior by volatile releaser pheromones, whereas the vomeronasal accessory olfactory system conveyed nonvolatile (primer pheromone) information to the neuroendocrine hypothalamus, affecting hormonal secretions1. These distinct, dual roles were radically modified by the introduction of molecular genetics to the study of chemosensory function. A new generation of chemosensory biologists demonstrated by targeted mutagenesis of receptor transduction mechanisms2 or deletion of receptor genes3 that activation of the vomeronasal organ itself is important for sexual and aggressive behavior. The vomeronasal organ thus assumed a leading role in the pheromone saga, and the main olfactory system became its understudy on the behavioral stage. Two recent papers, one in this issue4 and another in Cell5, bring the main olfactory system back under the spotlight, not only by showing its significant role in sexual and aggressive behavior4,5, but also by revealing an anatomical projection from the main olfactory epithelium to luteinizing hormone–releasing hormone (LHRH) neurons in the neuroendocrine hypothalamus5, which regulate gonadal hormone release. And so swings the pendulum of pheromone function in behavioral neuroscience…or does it? Have these results radically altered our understanding, or is there a potential story that complements rather than refutes previous work? To determine the contribution of the main olfactory system to sexual and aggressive behaviors in male mice, Mandiyan et al.4 tested mice lacking a cyclic nucleotide–gated ion channel subunit (Cnga2–/Y) essential for odor signaling in the main olfactory epithelium

E.B. Keverne is in the Sub-department of Animal Behaviour, University of Cambridge, Cambridge CB3 8AA, UK. e-mail: [email protected]

Photoperiod

Stress

Metabolic condition

AVPV

Glutamate GABA

MOE

Glutamate GABA

VNO

MOE

MPOA

LHRH neurons

LHRH neurons

Arcuate nucleus ME

ME TIDA neurons

VNO

Prolactin LH surge

TIDA neurons

Pulsatile LH release

Debbie Maizels


Eric B Keverne

Figure 1 Sites of chemosensory modulation in the male and female rodent hypothalamus. The sexually dimorphic anteroventral periventricular nucleus (AVPV) is a pivotal relay for signals that regulate LHRH neurons. GABAergic and glutamatergic inputs from the AVPV, together with estradiol, regulate LHRH neurons in the medial preoptic area (MPOA). The AVPV is larger in females and contains many neurons coexpressing GABA and glutamate. Chemosensory input from the vomeronasal organ (VNO) as well as other sensory inputs regulate endocrine responses in the hypothalamus (center). Yoon et al.5 show that LHRH neurons receive polysynaptic input that originates in the main olfactory epithelium (MOE). LHRH neurons project to the median eminence (ME) blood vessels, and in the male, synchronized release of LHRH produces pulsatile luteinizing hormone (LH) release from the pituitary. In the female, tuberoinfundibular dopamine (TIDA) neurons receive a polysynaptic input originating from the vomeronasal organ. These neurons also project to the ME, where dopamine release inhibits prolactin, a prelude to inducing female cyclicity and ovulatory LH surge.

but not the vomeronasal organ. The authors found that these essentially anosmic mice do not investigate other males, females or their urine marks by sniffing. In 30-minute observation tests, they fail to respond to females with investigation, mounting or intromission. When exposed to intruder males, again they fail to investigate and, unlike normal males, do not initiate aggressive behavior. These results clearly illustrate the importance of the main olfactory system for identifying the presence of either a male or female, and for activating appropriate aggressive or sexual behaviors.


In a second paper, Yoon et al.5 studied the chemosensory influence on reproductive function primarily by anatomically tracing the afferent pathways to LHRH neurons in the hypothalamus. They used an elegant methodology to infect only LHRH-expressing neurons with a modified pseudo–rabies virus. Owing to transneuronal retrograde transport of the virus, they were able to identify a discrete population of olfactory receptor neurons in the main olfactory epithelium that had polysynaptic connectivity to LHRH neurons in male and female mice, but they did not find any synap-

1637


NEWS AND VIEWS tic connections from the vomeronasal system (Fig. 1). In the male hypothalamus, exposure to female urine increased levels of phosphorylated MAP kinase in a subset of LHRH neurons. This functional modulation was impaired by chemical or genetic lesions (Cnga2–/Y) of the main olfactory epithelium but was unaffected in mutant mice lacking vomeronasal signaling. Consistent with the results of Mandiyan et al.4, mating behavior and sniffing investigation were also reduced in mice with lesions of the main olfactory epithelium. Yoon et al.5 thus provide anatomical evidence for polysynaptic drive to hypothalamic LHRH neurons from the main, but not the vomeronasal accessory, olfactory system, and they provide functional evidence in the male that urineevoked modulation of LHRH neurons depends in part on an intact main olfactory epithelium. Together, both studies also demonstrate a severe impairment in male investigatory and sexual and aggressive behavior in the absence of main olfactory epithelium signaling. The possibility therefore exists that the subpopulation of LHRH neurons activated via the main olfactory epithelium may be important for male sexual behavior. However, neither group reported significant decreases in plasma testosterone, and Cnga2–/Y mice housed continuously with females over 10 days achieve a 20% pregnancy rate5. Although a poor performance compared with normal males (100%), this suggests anosmic males are not entirely reproductively incapacitated. In any event, would such a possibility actually run counter to current views of chemosensory effects on reproductive function and behavior? Chemosensory cues in female mouse urine can induce luteinizing hormone release in males, and this endocrine response is blocked by removal of the vomeronasal organ6. However, there is no strong experimental evidence that LHRH promotes male sexual behavior in mice. Indeed, such evidence comes largely from studies in hamsters, in which chemosensory modulation of behavior and LHRH neurons does seem to involve the vomeronasal system. LHRH injections in the hamster can compensate for impaired sexual behavior caused by vomeronasal lesions7, and female pheromones stimulate the release of LHRH, which leads to an increase in plasma testosterone levels8. Furthermore, electrical stimulation of the vomeronasal organ induces cFos activation in pathways leading to LHRH neurons, as well as in LHRH neurons themselves9. How LHRH projections might influence behavior is by no means clear, however, even in the hamster. The viral anatomical tracing data in the mouse were pooled from males and females, but no functional studies were done on females. Vomeronasal influences on neuro-

1638

endocrine function are known in the female mouse (Fig. 1), although no studies have suggested they are mediated by direct synaptic input onto LHRH neurons. On the contrary, evidence suggests that LHRH release is modulated by a wide range of intrahypothalamic neuropeptides, including opioids, NPY, galanin and CRF, as well as by neurosteroids acting particularly on GABAA receptors and by the classical neurotransmitters dopamine and noradrenaline10. The female vomeronasal pathway projects to tuberoinfundibular dopamine (TIDA) neurons, and dopamine release by these neurons acts to reduce prolactin levels, which induces female estrous and blocks pregnancy11,12. TIDA neurons also influence LHRH release by acting in the median eminence where neuronal-glial-endothelial interactions and nitric oxide seem to regulate LHRH secretion13. Actions such as these would not be revealed by retrograde transneuronal tracing from LHRH neurons. Confused? One clear point is the need to consider both sex and species differences in neuroendocrine responses and behavior to understand pheromone action in rodents. Whereas the female mouse is cyclic and only spends a few hours of her lifetime sexually receptive and fertile, the adult male is permanently both sexually active and fertile. For most of the female’s life, LHRH is inhibited, either because she is pregnant or lactating. Switching on reproduction requires an active process in the female. If male pheromones, and hence potential mates, are around, females need to opportunistically advance their puberty or return to postpartum or postlactational estrous, and these neuroendocrine changes in the female mouse involve the vomeronasal organ system. On the other hand, in males, switching off reproduction is an active process brought about by aggression and subordination. Males, therefore, need chemosensory cues to choose to attack rather than mate. In this context, Trpc2–/– mice with impaired vomeronasal organ signaling illustrate the importance of the vomeronasal organ, as Trpc2–/– males do not display aggressive behavior toward other male mice and attempt to mate with both males and females2. Now, the new results in the studies by Mandiyan et al.4 and Yoon et al.5 demonstrate the additional importance of the main olfactory system for the investigation by sniffing that precedes the activation of such male behavior. Different species have different reproductive strategies, which may also contribute to variations in how chemosensory input affects reproductive behavior and function. Unlike in laboratory mice, which breed all year, in seasonally breeding species like the hamster, shortening of the photoperiod in winter actively induces

testicular regression, decreases plasma testosterone and eliminates male sexual behavior. Males enter a period of hibernation until early spring when spontaneous gonadal recrudescence takes place. Recovery of sexual behavior is accelerated when males interact with estrous females, but is not accelerated by distant visual, auditory or volatile cues14. Hence, such seasonally breeding males have their sexual behavior actively facilitated by pheromones. The significant outcome of the present studies is the common finding that male mice with a dysfunctional main olfactory system fail to investigate odors or conspecifics, a behavioral prelude to sexual and aggressive behavior. Active investigation, as opposed to distant sniffing, is also necessary for activation of the vomeronasal system15, which leads to sex-specific behaviors and neuroendocrine responses. Both chemosensory systems are, therefore, important for neuroendocrinology and reproduction, but how they are used depends on the reproductive strategy, which can be very different between the sexes and across species. Although there is good evidence that female pheromones act via the vomeronasal organ in the sexually naive male hamster, sexual experience brings into play the main olfactory system, enabling it to take over the functional role of the vomeronasal organ in the context of sexual behavior. Could this functional role of the main olfactory system be hardwired in the mouse? Future studies will undoubtedly reveal more about how signals activating the main olfactory epithelium influence sexual behaviors and reproductive function. In particular, the identification of a subpopulation of olfactory receptor neurons that make contact with LHRH neurons will enable their characterization, determination of potential sex differences and perhaps discovery of their ligands. 1. Keverne, E.B. Trends Neurosci. 1, 32–34 (1978). 2. Leypold, B.G. et al. Proc. Natl. Acad. Sci. USA 99, 6376–6381 (2002). 3. Del Punta, K. et al. Nature 419, 70–74 (2002). 4. Mandiyan, V.S., Coats, J.K. & Shah, N.M. Nat. Neurosci. 8, 1660–1662 (2005). 5. Yoon, H., Enquist, L.W. & Dulac, C. Cell published online 2005 (doi:10.1016/j.cell.2005.08.039). 6. Coquelin, A., Clancy, A.N., Macrides, F., Noble, E.P. & Gorski, R.A. J. Neurosci. 4, 2230–2236 (1984). 7. Westberry, J.M. & Meredith, M. Chem. Senses 28, 191– 196 (2003). 8. Richardson, H.N. et al. Gen. Comp. Endocrinol. 138, 211–217 (2004). 9. Meredith, M. & Fewell, G. Brain Res. 922, 87–94 (2001). 10. Genazzani, A.R., Bernardi, F., Monteleone, P., Luisi, S. & Luisi, M. Ann. NY Acad. Sci. 900, 1–9 (2000). 11. Li, C.S., Kaba, H., Saito, H. & Seto, K. Neuroscience 36, 773–778 (1990). 12. Rosser, A.E., Remfry, C.J. & Keverne, E.B. J. Reprod. Fertil. 87, 553–559 (1989). 13. Beauvillain, J.C. & Prevot, V. J. Soc. Biol 198, 68–72 (2004). 14. Honrado, G.I. & Fleming, A.S. J. Biol. Rhythms 11, 103–112 (1996). 15. Luo, M., Fee, M.S. & Katz, L.C. Science 299, 1196– 1201 (2003).


NEWS AND VIEWS

Nematodes learn: now what? William G Quinn

The author is in the Department of Brain and Cognitive Sciences, 46-5009, Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, Massachusetts 02139, USA. e-mail: [email protected]

A

0!

0!

4RANSFER

.AIVEWORMSON PATHOGENIC 0AERUGINOSA HOURS

$B

4RANSFER

!4#

7ORMSON NON PATHOGENIC 3MARCESCENS HOURS

B

0! 4RANSFER %

.EWSAMPLEOF NAIVEWORMSON NON PATHOGENIC 0AERUGINOSA HOURS

!4#

7ORMSON PATHOGENIC 3MARCESCENS HOURS

Raised on OP50 Exposed to PA14 for 4 hours Starved for 4 hours

D

0ATHOGENIC0! 0ATHOGENIC 3MARCESCENS

0.4 0.2 0 –0.2 –0.4

*

Jessica Iannuzzi

C

4RANSFER !4#

!VOIDANCEOF0!

The nematode worm Caenorhabditis elegans has been an important model for four decades now, because of its rapid genetics, simple nervous system (302 neurons) and stereotyped development. What the worm seemed to lack was sophisticated behavior—careers have been built and Nobel prizes gained studying genetically induced variants in swimming, egg-laying and death. Now Zhang et al.1, building on earlier work2,3, persuasively report in Nature that nematodes can learn to associate different chemosensory stimuli with illness, and to avoid these stimuli in a choice test. The learning is of an unusual type, discovered 60 years ago in rats. In the mid 1950s, John Garcia and his colleagues discovered that rats made ill (nauseated) by radiation or toxins avoided foods that they had tasted some time earlier4. This induced behavioral change differed from classical conditioning in several ways. The effect was strikingly strong and long lasting, the pairing was confined to tastes and nausea and, particularly, it lacked the requirement for close temporal association of stimulus and reinforcement—the taste could occur some hours before the induced nausea, rather than the interval of a second or so that is typical of classical conditioning or electric shock–reinforced learning. Garcia’s work was controversial for years but is now established as authentic learning, with useful applications in ecology and patient care5. The effect has been demonstrated in mammals, squid and garden slugs. The worms of Zhang et al. also seem to learn in this way. Nematodes feed on bacteria. Some strains of bacteria are pathogenic; worms that feed on them will eventually die. Worms raised for 4 hours on pathogenic bacteria will later avoid them when given a choice. Associative learning by definition depends on temporal pairing between stimulus and reinforcement. In the case of taste (or odor) aversion learning, the requirement for pairing is loosened

Choice index


A study in Nature reports that nematodes can learn to associate different chemosensory stimuli with illness and to avoid these stimuli in a choice test. Elevated serotonin in a particular type of neuron was critical for this learning.

Figure 1 Discriminative learning in C. elegans. (a) Training and testing scheme, part 1. A sample of worms was raised on pathogenic P. aeruginosa (PA14) bacteria for 4 hours, then transferred for a further 4 hours to non-pathogenic bacteria of a different species (S. marcescens), which presumably have a different smell and taste. The worms were then transferred to a test plate, placed between samples of the two bacterial species (both pathogenic). After 1–2 hours, they were immobilized with sodium azide, and the worms on each bacterial sample were counted. (b) Training and testing, part 2. A new sample of worms were trained and tested as above, but with the pathogenicity switched to the other species (pathogenic and non-pathogenic strains of each species were, as far as possible, coisogenic). A relative shift in choice bias from S. marcescens in a to P. aeruginosa in b indicates discriminative (associative) learning to diffusible cues. (In the actual experiments, two more samples of worms were trained and tested exactly as described above, but with the order of exposure to species switched during training—S. marcescens first.) (c) Altered choice bias during testing. Animals exposed to pathogenic PA14 and harmless S. marcescens avoid PA14 more than animals exposed to harmless P. aeruginosa and pathogenic S. marcescens. (d) Data from c, replotted to show raw avoidance of one of the species (P. aeruginosa). Panels c and d reproduced from ref. 1 with permission from the authors. ***P < 0.001, *P < 0.05, n ≥ 4 assays. Error bars represent s.e.m.


1639

NEWS AND VIEWS

A

B


!$&

!$&

!$&

Figure 2 Pathogenic bacteria elevate ADF neuron serotonin levels. (a,b) Serotonin immunoreactivity in wild-type animals fed either Escherichia coli (a) or E. coli and pathogenic P. aeruginosa (b). Note the increase in immunoreactivity (marked by fluorescence) in b, particularly in chemosensory neurons ADF. Figure reproduced from ref. 1 with permission.

from seconds to hours, making pairing studies less conclusive. Given this limitation, Zhang et al. did the necessary controls for associative learning as well as possible, using a discriminative learning assay with built-in controls (Fig. 1a,b). They used pathogenic and nonpathogenic variants of two species of bacteria (Pseudomonas aeruginosa and Serratia marcescens). Four groups of worms were tested in four different experiments. Each experiment involved placing the worms for 4 hours on pathogenic bacteria of one species and then for 4 hours on non-pathogenic bacteria of the other species, concluding with a choice test between species. In different experiments, the pathogenicity was switched from one species to the other, and the order of species presentations was reversed. When choice percentages were averaged over all the experiments, there was a significant overall tendency of the worms to avoid the bacterial odor (or taste) experienced with pathogenicity (Fig 1c). Switching pathogenicity from one bacterial species to the other during training altered the worms’ later choice percentage between bacteria by about 40% during testing, averaged over all the experiments. Discriminative conditioning assays of exactly this type have been successfully used to condition bees, fruit flies, slugs and Aplysia vaccaria, among other species6. Now nematodes can be trained in this way. Zhang et al. then went on to use available mutants and transgenic worms to identify molecular components of the learning pathway and to start to define the neuronal circuit involved. Worms with a mutation3 in the gene for tryptophan hydroxylase (necessary to synthesize the neurotransmitter serotonin) failed to learn, although they showed normal chemosensory choices in non-learning tests. Indeed, raising nematodes on pathogenic bacteria

1640

dramatically increased their serotonin levels in several neurons, including in chemosensory neurons denoted ADF (Fig. 2). Expression of tryptophan hydroxylase cDNA in this neuron partially rescued the learning defect, whereas expressing the cDNA in another neuron, NSM, did not. This suggests (but does not prove) that elevated serotonin in the ADF cells, at least in part, signals the existence of the negative reinforcement—malaise from ingesting pathogenic bacteria. Nematodes have genes for several serotonin receptors, with mutants available for some of them. One such gene, mod-1, encodes a serotonin-gated chloride channel that is known to function in another response to food7. Mod-1 mutants, lacking the receptor, are deficient in the aversive component of the learning behavior. Mod-1 is expressed in five interneurons in the worm chemosensory circuit. ADF synapses directly onto two of them, AIZ and perhaps AIY, giving a suggestive minimal circuit for aversive learning of three cells—not too shabby. However, most of the mutant effects and the rescues were only partial, and some of the cellular localizations were also incomplete. Moreover, transmitters including serotonin can act at a distance (like hormones), as well as directly across synapses. Laser ablation studies may help refine the circuit further, provided the learning assay can be made strong enough to score small numbers of worms. C. elegans is an elegant model organism, but is it a model mostly of itself? Certainly its genes, molecules and metabolic processes are quite similar to ours—the pathway for programmed cell death, present in all animals, was worked out largely from genetic studies in nematodes. In this regard, the implication of serotonin as a reinforcement signal is consistent particularly with work in

Aplysia vaccaria8. What is surprising from the worm work is the finding that serotonin acts through an ion channel–linked receptor, rather than a G protein–coupled one, together with the absence of any evidence for one of the second-messenger systems implicated in learning in so many other species. Given that second-messenger systems are a handy way to produce moderately long-lasting changes in cells, it will be interesting to see from future studies how C. elegans encodes its memories. In particular, it will be revealing to see how long the worms can remember the chemosensory discrimination they have learned. Another open question also promises new mechanisms: how does C. elegans handle stimulus pairing? Associative learning requires that the stimuli—a bacterial taste or smell and nausea in this case—converge synergistically on some molecular entity to initiate a lasting change in the relevant neurons. In vertebrates, this convergence occurs in most cases onto the voltage-and-ligand-gated NMDA type of glutamate receptor. In A. vaccaria and in Drosophila melanogaster, signals from stimuli converge primarily onto the calcium-activated and G protein–activated (type I) adenylyl cyclase enzyme8,9. In nematodes, serotonin seems to be the signal for reinforcement, but it must somehow produce lasting changes in the avoidance circuit only for the paired chemosensory stimulus (for example, from P. aeruginosa) but not the unpaired control (for example, from S. marcescens). It is not clear yet how the mod-1 receptor could do this without hidden properties or new molecular partners. The study by Zhang et al. reflects a real advance in worm sophistication, as well as in human sophistication in understanding them. Given the small nervous system of the worm and the powerful genetic, molecular and anatomical tools available to study it, it represents a foot in a very big door. 1. Zhang, Y., Lu, H. & Bargmann, C.I. Nature 438, 179– 184 (2005). 2. Bernhard, N. & van der Kooy, D. Learn. Mem. 7, 199– 212 (2000). 3. Sze, J.Y., Victor, M., Loer, C., Shi, Y. & Ruvkun, G. Nature 403, 560–564 (2000). 4. Garcia, J., Kimeldorf, D.J. & Koelling, R.A. Science 122, 157–158 (1955). 5. Rozin, P. & Kalat, J.W. Psychol. Rev. 78, 459–486 (1971). 6. Carew, T.J. & Sahley, C.L. Annu. Rev. Neurosci. 9, 435–487 (1986). 7. Ranganathan, R., Cannon, S.C. & Horvitz, H.R. Nature 408, 470–475 (2000). 8. Kandel, E.R. Science 294, 1030–1038 (2001). 9. Livingstone, M.S., Sziber, P.P. & Quinn, W.G. Cell 37, 205–215 (1984).



C O M P U TAT I O N A N D S Y S T E M S

PERSPECTIVE

A natural approach to studying vision Gidon Felsen & Yang Dan An ultimate goal of systems neuroscience is to understand how sensory stimuli encountered in the natural environment are processed by neural circuits. Achieving this goal requires knowledge of both the characteristics of natural stimuli and the response properties of sensory neurons under natural stimulation. Most of our current notions of sensory processing have come from experiments using simple, parametric stimulus sets. However, a growing number of researchers have begun to question whether this approach alone is sufficient for understanding the real-life sensory tasks performed by the organism. Here, focusing on the early visual pathway, we argue that the use of natural stimuli is vital for advancing our understanding of sensory processing. The visual system is a remarkable device, able to process complex spatiotemporal patterns of light signals collected by the eyes and provide information about a rapidly changing world, which is essential for the survival of the organism. The efficiency of the visual circuit is not accidental; it has been shaped by the forces of evolution and refined during development in an experience-dependent manner. The function of the system is thus intimately related to the properties of the visual stimuli commonly found in the natural environment. Much of what we know about visual processing has been obtained from studies using ‘artificial’ stimuli, including parametric sets of simple stimuli (such as spots of light or sinusoidal gratings) and stimulus ensembles with simple statistics (for example, white noise). However, recent studies have begun to incorporate ‘natural’ stimuli, such as photographs or video recordings of the natural environment1. Here, before assessing how the use of natural stimuli can advance our understanding of visual processing, we will first examine the strengths and weaknesses of the commonly used artificial stimuli. Artificial stimuli Well-designed simple stimuli have been pivotal to elucidating the neural basis of sensory processing. In the visual system, simple stimuli such as spots or bars of light have revealed the receptive field (RF) structure of retinal2 and thalamic3 neurons and the orientation selectivity of cortical neurons4, which constitute the foundation of modern vision research. A major advantage of these stimuli is that

Gidon Felsen is at the Cold Spring Harbor Laboratory, 1 Bungtown Road, Cold Spring Harbor, New York 11724, USA. Yang Dan is in the Department of Molecular and Cell Biology and the Helen Wills Neuroscience Institute, University of California Berkeley, Berkeley, California 94720-3200, USA. e-mail: [email protected] Published online 23 November 2005; doi:10.1038/nn1608


they are readily parameterized and therefore ideal for determining the dependence of the neuronal response on a particular stimulus parameter (for example, orientation of the bar), often represented as a tuning curve. However, these simple stimuli have a significant drawback: in the natural environment, sensory inputs rarely consist of isolated simple patterns. Because neuronal processing is largely nonlinear, the responses to the natural stimuli are often poorly predicted by the sum of the responses to the individual simple components. Thus, studies using simple stimuli alone do not necessarily tell us how a neuron responds to the more complex stimuli encountered in the natural environment. Another approach is to use a large set of random stimuli, which allows exploration of a larger stimulus space and a more comprehensive characterization of the neuronal response properties. By identifying the set of stimuli that precede neuronal spiking (the ‘spike-triggered’ stimulus ensemble) and comparing it with the entire stimulus ensemble presented in the experiment, one can identify the features of the stimuli that influence the neuronal response. In particular, white noise, which contains equal power at all frequencies and no spatiotemporal correlation, has been used to measure the spatial and temporal RFs of early visual neurons5. Although this statistical property of white noise is highly desirable for systems identification in general6, in practice it is often ineffective for driving sensory neurons, particularly in later processing stages, making it difficult to estimate the RF parameters accurately. Sparse random stimuli, which consist of a rapidly presented, random sequence of preselected stimulus patterns, can be more effective than dense white-noise stimuli for measuring certain response properties7–10. However, the choice of stimulus set in each experiment requires prior assumption about what stimulus parameters are functionally relevant for the cell being studied, which may result in biases in the interpretation of the experimental results. Given the limitations of the simple and random stimuli, natural stimuli provide a valuable complement for the study of visual processing. Compared to simple, parametric stimulus sets, the use of natural stimulus ensembles requires fewer a priori assumptions about relevant stimulus parameters. Compared to white noise, natural stimuli are functionally more relevant and, in some cases, more effective for driving the neurons11,12. Furthermore, because many sensory neurons show adaptation and contextual modulation, the response properties measured with artificial stimuli may not generalize to natural stimuli; certain properties may be apparent only under natural stimulation13. Although analyzing responses to natural stimuli may not be as straightforward as for the artificial stimuli, new methods are being developed to overcome these complications11,13–17. Indeed, recent studies using natural stimuli have yielded new insights into the function of the visual system.

1643


PERSPECTIVE What natural stimuli have taught us Arguably, to establish the functional significance of a response property characterized with artificial stimuli, this property should be validated under natural stimulation. If a response property does not contribute to explaining the responses to natural stimuli, its functional relevance in visual processing must be limited—even if the property is prominent under artificial stimulation. Conversely, the extent to which a given model explains the responses to natural stimuli provides an essential measure of the completeness of the model. Such validation analyses have been performed at several stages of the early visual pathway. For example, in the lateral geniculate nucleus (LGN), neuronal RFs measured with white noise have been used to predict the responses to natural stimuli. The prediction based on the linear spatiotemporal RF alone captured the basic temporal features of the responses to natural stimuli18. In the primary visual cortex, neuronal RFs were originally characterized with spots or bars of light4. Simple cell RFs were found to contain oriented ON and OFF subregions; complex cell RFs were modeled as the combination of multiple simple cell RFs with the same orientation preference but different ON-OFF locations. Recently, several studies have examined the structure of cortical RFs by analyzing the neuronal responses to natural images. Using spiketriggered average for simple cells15 and spike-triggered covariance11 or other nonlinear methods14,17 for complex cells, these studies showed that the RFs of V1 neurons during natural stimulation are similar to those measured with the simple or white-noise stimuli. Together, these studies have largely validated the basic spatial structure of early visual RFs measured with artificial stimuli. In addition to verifying the RF models measured with artificial stimuli, natural stimuli have been used to reveal properties that were not predicted by the results of previous experiments. In a recent study13, the RFs of visual cortical neurons in awake behaving monkeys were estimated from the recorded responses to both natural stimuli and dynamic gratings, and the RF structures were found to be affected by both the spatial and temporal statistics of the stimuli. Under natural stimulation, the RF showed a significant inhibitory component that was not observed in the responses to the artificial stimuli; moreover, the RF estimated with natural stimuli predicted the responses to a separate set of natural stimuli more accurately than did the RF estimated with artificial stimuli. Thus, natural stimuli have helped us to reveal a functionally relevant response component that has evaded characterization with artificial stimuli. Another example of a response property uniquely revealed by natural images comes from a study on cortical feature sensitivity—the ability of each neuron to detect the presence of its preferred features in the visual inputs12. By analyzing the cortical responses to ensembles of random stimuli, natural images and synthetic stimuli with either natural power or natural phase spectra, this study showed that the sensitivity of complex cells is higher when the feature is present in natural images than when it is present in random stimuli. Notably, this enhanced feature sensitivity was due to the phase spectra of natural images rather than to their power spectra, and it was not predicted by the standard models of complex cells4,19,20, which were derived from studies using artificial stimuli. This finding suggests that complex cells are functionally more specialized than described by the standard models; they respond more vigorously to the oriented edges and contours commonly found in natural images than to the random stimuli that, according to the standard model, should have activated them to the same extent. Of course, once this response property is revealed, one can design parametric, artificial stimulus sets (for example, sums of sinusoids with various phase relationships21) to further characterize this property. However, without the use of natural images, it would have been difficult to discover such a property in the

1644

first place. In addition to this example in the visual system, response properties that are specifically tuned to the statistics of natural stimuli have also been demonstrated in the auditory system22,23. Finally, studies on the statistics of natural stimuli have addressed the functional significance of a range of neuronal response properties previously characterized with artificial stimuli. For example, natural visual stimuli contain extensive spatial24 and temporal25 correlations (for instance, neighboring pixels tend to have similar luminance values), which gives rise to a high degree of redundancy in the input signals. Quantitative analyses of these correlations show that the center-surround antagonism of the spatial RFs26,27 and the biphasic impulse response functions28 of retinal and LGN neurons are well suited for decorrelation of natural stimuli, which significantly reduces the coding redundancy29. In the primary visual cortex, reducing redundancy by maximizing the statistical independence of the responses of neighboring neurons (rather than through simple decorrelation) depends on a form of contrast gain control that resembles the well-known divisive normalization model20. This gain control can account for a variety of cortical response properties such as cross-orientation suppression and tuning for stimulus size30. Related to redundancy reduction, another well-known conjecture of visual processing is sparse coding31, in which each stimulus elicits robust responses in only a small number of neurons31,32. Notably, when response sparseness (or, similarly, statistical independence) of a population of model neurons was maximized under natural stimulation, the RFs of these model neurons resembled those of cortical simple cells—in terms of both spatiotemporal profile33–35 and chromatic properties36. By demonstrating the functional advantages of particular response properties, these theoretical studies have addressed the question of why the neurons respond the way they do. This line of research thus complements the experimental investigations that focus primarily on how the neurons respond. Challenges in the use of natural stimuli The results described above have demonstrated that a combination of theoretical and experimental studies using natural stimuli can lead to new insights into visual processing. However, there are considerable challenges in the use of natural stimuli, which makes the skepticism of some researchers understandable. To make natural stimuli a more useful tool for vision research, we believe it is necessary to address the following major issues. Most theoretical studies of natural stimulus statistics have thus far led to explanations of known response properties rather than to predictions of new properties. Predictions usually count more than explanations, an intuitive notion that has also been endorsed by philosophers of science37. Although explaining the existing observations may be safer and less likely to be challenged, ‘sticking one’s neck out’ with bold predictions is more likely to make a strong impact. From the experimentalists’ point of view, the most attractive predictions are those that not only provide a rigorous test of a particular computational theory, but also lead to potential findings that are surprising in their own right, with functional or mechanistic implications reaching beyond the original theory. A good example of such a prediction comes from a recent study, in which independent component analysis (a method for maximizing coding efficiency34) was applied to the outputs of model V1 complex cells. The result predicted that, to represent step edges in images, each V2 neuron pools the responses of V1 neurons that have similar preferred orientations and RF locations but different spatial frequencies38. In another study, a hierarchical probabilistic model was used to learn higher-order statistical regularities in natural images. This algorithm yielded representations of abstract image properties such as object location, scale and texture39,



PERSPECTIVE which may reflect the response properties of extrastriate visual cortical neurons. Such predictive modeling studies are highly valued by the experimental community. In experimental studies using natural stimuli, a major difficulty is the analysis of the stimulus-response relationship, owing to the complex statistical properties of natural stimuli. In general, estimation of RFs involves correlating the stimulus and the response, which is relatively straightforward if the stimulus is Gaussian white noise6. Unfortunately, natural stimuli are neither white nor Gaussian, which can cause inaccuracies in the RF estimation. Although some properties of natural stimuli have been well characterized (for example, the predominance of low-frequency signals resulting from spatial and temporal correlations), their higher-order regularities are not well understood. For instance, although it has long been recognized that the phase structure of natural images is far from random and is important perceptually40, only a few studies have addressed the phase structures of natural images41,42. A better understanding of these regularities in natural stimuli will greatly facilitate unbiased experimental characterization of neuronal response properties using natural stimuli. At a more basic level, even the concept of ‘natural stimuli’ has not been defined rigorously. For example, scenes from city streets and forest environments are likely to have different statistical properties, such as the distribution of oriented energy43. Each species may be specifically adapted to its own natural environment, and even the same visual circuit may function differently when stimulated by different types of natural scenes. Yet, in most studies, the stimuli are ‘natural’ only from a human perspective, and the selection of images in the stimulus ensemble is largely arbitrary. In an effort to overcome such a limitation, one study used video cameras attached to the heads of freely roaming cats to record visual stimuli natural to the cat44. The resulting stimuli indeed appear quite different from those selected from a standard Hollywood production, with a larger representation of the visual objects closest to the ground and temporal dynamics reflecting the head and body movements of the cat. Similar studies have been performed on several other species45–47. To make future experimental studies more fruitful, it would be useful to standardize natural stimuli and to categorize them into subclasses based on functionally relevant statistical properties. There are multiple criteria to consider, such as whether the stimuli are indigenous to the natural environment of the particular animal and whether sampling of the stimuli reflects the body, head and eye movements of the animal during natural behavioral tasks. Although there may not exist a simple and obvious set of classification criteria, the field is likely to benefit from some level of standardization. Conclusion In summary, we believe that physiological experiments using natural stimuli and guided by theories derived from natural stimulus statistics constitute a powerful approach to understanding the visual system. Although in this article we have focused on the early visual pathway, the response properties of higher cortical neurons (that is, those beyond V1) are likely to be even more closely associated with the characteristics of natural stimuli48. Predictions of these properties based on natural stimulus statistics combined with experiments using ecologically relevant stimuli may prove to be indispensable for cracking the neural code in higher visual areas. In addition to the response properties of single neurons, the measurement of ensemble neuronal responses to natural stimuli—using multielectrodes49 or optical methods—may also reveal new principles of visual processing. Thus, despite the challenges, natural stimuli are bound to be critical in advancing our understanding of visual processing.


ACKNOWLEDGMENTS We thank F. Han, J. Touryan, B. Willmore and W. Vinje for helpful comments. This work was supported by a grant from the National Eye Institute (R01 EY12561). COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/

1. Simoncelli, E.P. & Olshausen, B.A. Natural image statistics and neural representation. Annu. Rev. Neurosci. 24, 1193–1216 (2001). 2. Hartline, H.K. The receptive fields of optic nerve fibers. Am. J. Physiol. 130, 690–699 (1940). 3. Hubel, D.H. & Wiesel, T.N. Integrative action in the cat’s lateral geniculate body. J. Physiol. (Lond.) 155, 385–398 (1961). 4. Hubel, D.H. & Wiesel, T.N. Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. J. Physiol. (Lond.) 160, 106–154 (1962). 5. Reid, R.C., Victor, J.D. & Shapley, R.M. The use of m-sequences in the analysis of visual neurons: linear receptive field properties. Vis. Neurosci. 14, 1015–1027 (1997). 6. Marmarelis, P.Z. & Marmarelis, V.Z. Analysis of Physiological Systems: The WhiteNoise Approach (Plenum, New York, 1978). 7. Ringach, D.L., Hawken, M.J. & Shapley, R. Dynamics of orientation tuning in macaque primary visual cortex. Nature 387, 281–284 (1997). 8. Mazer, J.A., Vinje, W.E., McDermott, J., Schiller, P.H. & Gallant, J.L. Spatial frequency and orientation tuning dynamics in area V1. Proc. Natl. Acad. Sci. USA 99, 1645–1650 (2002). 9. Felsen, G. et al. Dynamic modification of cortical orientation tuning mediated by recurrent connections. Neuron 36, 945–954 (2002). 10. Bredfeldt, C.E. & Ringach, D.L. Dynamics of spatial frequency tuning in macaque V1. J. Neurosci. 22, 1976–1984 (2002). 11. Touryan, J., Felsen, G. & Dan, Y. Spatial structure of complex cell receptive fields measured with natural images. Neuron 45, 781–791 (2005). 12. Felsen, G., Touryan, J., Han, F. & Dan, Y. Cortical sensitivity to visual features in natural scenes. PLoS Biol. 3, 1819–1828 (2005). 13. David, S.V., Vinje, W.E. & Gallant, J.L. Natural stimulus statistics alter the receptive field structure of V1 neurons. J. Neurosci. 24, 6991–7006 (2004). 14. Ringach, D.L., Hawken, M.J. & Shapley, R. Receptive field structure of neurons in monkey primary visual cortex revealed by stimulation with natural image sequences. J. Vis. 2, 12–24 (2002). 15. Smyth, D., Willmore, B., Baker, G.E., Thompson, I.D. & Tolhurst, D.J. The receptivefield organization of simple cells in primary visual cortex of ferrets under natural scene stimulation. J. Neurosci. 23, 4746–4759 (2003). 16. Sharpee, T., Rust, N.C. & Bialek, W. Analyzing neural responses to natural signals: maximally informative dimensions. Neural Comput. 16, 223–250 (2004). 17. Prenger, R., Wu, M.C., David, S.V. & Gallant, J.L. Nonlinear V1 responses to natural scenes revealed by neural network analysis. Neural Netw. 17, 663–679 (2004). 18. Dan, Y., Atick, J.J. & Reid, R.C. Efficient coding of natural scenes in the lateral geniculate nucleus: experimental test of a computational theory. J. Neurosci. 16, 3351–3362 (1996). 19. Adelson, E.H. & Bergen, J.R. Spatiotemporal energy models for the perception of motion. J. Opt. Soc. Am. A 2, 284–299 (1985). 20. Heeger, D.J. Normalization of cell responses in cat striate cortex. Vis. Neurosci. 9, 181–197 (1992). 21. Mechler, F., Reich, D.S. & Victor, J.D. Detection and discrimination of relative spatial phase by V1 neurons. J. Neurosci. 22, 6129–6157 (2002). 22. Rieke, F., Bodnar, D.A. & Bialek, W. Naturalistic stimuli increase the rate and efficiency of information transmission by primary auditory afferents. Proc. Biol. Sci. 262, 259–265 (1995). 23. Woolley, S.M., Fremouw, T.E., Hsu, A. & Theunissen, F.E. Tuning for spectro-temporal modulations as a mechanism for auditory discrimination of natural sounds. Nat. Neurosci. 8, 1371–1379 (2005). 24. Attneave, F. Some informational aspects of visual perception. Psychol. Rev. 61, 183–193 (1954). 25. Dong, D.W. & Atick, J.J. Statistics of natural time varying images. Netw. Comput. Neural Syst. 6, 345–358 (1995). 26. Srinivasan, M.V., Laughlin, S.B. & Dubs, A. Predictive coding: a fresh view of inhibition in the retina. Proc. R. Soc. Lond. B 216, 427–459 (1982). 27. Atick, J.J. Could information theory provide an ecological theory of sensory processing? Netw. Comput. Neural Syst. 3, 213–251 (1992). 28. Dong, D.W. & Atick, J.J. Temporal decorrelation: a theory of lagged and nonlagged responses in the lateral geniculate nucleus. Netw. Comput. Neural Syst. 6, 159–178 (1995). 29. Barlow, H.B. Possible principles underlying the transformation of sensory messages. in Sensory Communication (ed. Rosenblith, W.A.) 217–234 (MIT Press, Cambridge, Massachusetts, USA, 1961). 30. Schwartz, O. & Simoncelli, E.P. Natural signal statistics and sensory gain control. Nat. Neurosci. 4, 819–825 (2001). 31. Field, D.J. What is the goal of sensory coding? Neural Comput. 6, 559–601 (1994). 32. Willmore, B. & Tolhurst, D.J. Characterizing the sparseness of neural codes. Network 12, 255–270 (2001).

1645


PERSPECTIVE 33. Olshausen, B.A. & Field, D.J. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature 381, 607–609 (1996). 34. Bell, A.J. & Sejnowski, T.J. The “independent components” of natural scenes are edge filters. Vision Res. 37, 3327–3338 (1997). 35. van Hateren, J.H. & Ruderman, D.L. Independent component analysis of natural image sequences yields spatio-temporal filters similar to simple cells in primary visual cortex. Proc. R. Soc. B Biol. Sci. 265, 2315–2320 (1998). 36. Caywood, M.S., Willmore, B. & Tolhurst, D.J. Independent components of color natural scenes resemble V1 neurons in their spatial and color tuning. J. Neurophysiol. 91, 2859–2873 (2004). 37. Lipton, P. Testing hypotheses: prediction and prejudice. Science 307, 219–221 (2005). 38. Hyvarinen, A., Gutmann, M. & Hoyer, P.O. Statistical model of natural stimuli predicts edge-like pooling of spatial frequency channels in V2. BMC Neurosci. 6, 12 (2005). 39. Karklin, Y. & Lewicki, M.S. Learning higher-order structures in natural images. Network 14, 483–499 (2003). 40. Oppenheim, A.V. & Lim, J.S. The importance of phase in signals. Proc. IEEE. Inst. Electr. Electron. Eng. 69, 529–541 (1981).

1646

41. Thomson, M.G. Beats, kurtosis and visual coding. Network 12, 271–287 (2001). 42. Wang, Z. & Simoncelli, E.P. Local phase coherence and the perception of blur. in Advances in Neural Information Processing Systems, Vol. 16 (eds. Thrun, S., Saul, L. & Scholkopf, B.) (MIT Press, Cambridge, Massachusetts, USA, 2003). 43. Coppola, D.M., Purves, H.R., McCoy, A.N. & Purves, D. The distribution of oriented contours in the real world. Proc. Natl. Acad. Sci. USA 95, 4002–4006 (1998). 44. Kayser, C., Einhauser, W. & Konig, P. Temporal correlations of orientations in natural scenes. Neurocomputing 52, 117–123 (2003). 45. Passaglia, C., Dodge, F., Herzog, E., Jackson, S. & Barlow, R. Deciphering a neural code for vision. Proc. Natl. Acad. Sci. USA 94, 12649–12654 (1997). 46. Lewen, G.D., Bialek, W. & de Ruyter van Steveninck, R.R. Neural coding of naturalistic motion stimuli. Network 12, 317–329 (2001). 47. Lei, Y. et al. Telemetric recordings of single neuron activity and visual scenes in monkeys walking in an open field. J. Neurosci. Methods 135, 35–41 (2004). 48. Tanaka, K. Inferotemporal cortex and object vision. Annu. Rev. Neurosci. 19, 109–139 (1996). 49. Fiser, J., Chiu, C. & Weliky, M. Small modulation of ongoing cortical dynamics by sensory input during natural vision. Nature 431, 573–578 (2004).




PERSPECTIVE

In praise of artifice Nicole C Rust & J Anthony Movshon The visual system evolved to process natural images, and the goal of visual neuroscience is to understand the computations it uses to do this. Indeed the goal of any theory of visual function is a model that will predict responses to any stimulus, including natural scenes. It has, however, recently become common to take this fundamental principle one step further: trying to use photographic or cinematographic representations of natural scenes (natural stimuli) as primary probes to explore visual computations. This approach is both challenging and controversial, and we argue that this use of natural images is so fraught with difficulty that it is not useful. Traditional methods for exploring visual computations that use artificial stimuli with carefully selected properties have been and continue to be the most effective tools for visual neuroscience. The proper use of natural stimuli is to test models based on responses to these synthetic stimuli, not to replace them. There are two important commonly held fallacies that drive the new fashion of using natural stimuli in visual neuroscience experiments. The first is that in 40 years of experimentation and modeling, we have failed to capture important aspects of the behavior of neurons in primary visual cortex. The second is that an important reason for this ‘failure’ is that we have been trapped in the world of ‘simplistic’ artificial stimuli, which lack the richness of natural stimuli and have therefore prevented us from uncovering crucial facts of cortical organization1. Standard models for cortical cells based on synthetic stimuli We know a considerable amount about the response properties of neurons in primary visual cortex (V1). These neurons show stimulus selectivity simultaneously for a diverse set of parameters, such as the location, size, form and color of an object. An important goal of visual neurophysiology is to produce models of these neurons that describe how this stimulus selectivity arises. Ultimately, one hopes to integrate all these models into a single theory that can predict neuronal and population responses to any arbitrary stimulus. In the classical tradition, a visual physiologist observes the response of a neuron to a particular stimulus or class of stimuli and considers whether the current generation of models captures that behavior or whether these models require elaboration. In other words, the physiologist treats the current model as a hypothesis and designs experiments to

Nicole C. Rust is at the Howard Hughes Medical Institute and the Center for Neural Science, and J. Anthony Movshon is at the Center for Neural Science, New York University, New York, New York 10003, USA. e-mail: [email protected] Published online 23 November 2005; doi:10.1038/nn1606


test it. Naturally, these models are based on measurements made with stimuli that are simple and easily parameterized, such as bars, points of light and sinusoidal gratings, and are then tested with stimuli of increasing complexity. The history of research on neurons in V1 illustrates the success of this approach. V1 neurons simultaneously represent information both about the spatial structure of a stimulus and where it is located in the visual field (for a review see ref. 2). This multidimensional representation is based on selectivity for such stimulus features as the position, orientation, size, binocular disparity and color of stationary stimuli, as well as the direction and speed of moving stimuli. The first widely accepted formal models of these cells (referred to here as the ‘old standard models’) captured these tuning properties by passing an image through one or more linear spatiotemporal filters (Fig. 1a). In these models, stimulus selectivity arises from the shape of these filters: stimuli that resemble the filters produce high firing rates whereas stimuli that differ produce negligible firing rates. To capture this behavior, model spatial filters based on Gabor or related wavelet functions are typically used3,4. Within V1, neurons that are sensitive to the position of a stimulus within their receptive field (‘simple cells’) are commonly modeled as a single linear filter whose output is halfrectified by the inevitable threshold nonlinearity shared by all spiking neurons5. The position insensitivity of ‘complex cells’ is commonly modeled with two phase-shifted filters whose outputs are squared and summed (the energy model6,7) before being passed through the nonlinear spiking threshold. The responses of both kinds of neurons are irregular, and this variability can be reasonably approximated by a Poisson spiking process. These old standard models predict not only the selectivity of V1 neurons’ responses to bars, edges and gratings, but they also provide a credible account of responses to a variety of more complicated targets, including checkerboards8, random dot textures and Glass patterns9, and photographs of natural scenes10. This is not to say, however, that the old standard models are completely satisfactory. Research has uncovered a number of interesting ways in which they fail, resulting in the continuing evolution of a more elaborate and comprehensive ‘new standard model’. There are five important elements of V1 receptive fields not captured by the old standard model. First, although the old model postulated one input filter for simple cells and two for complex cells, more sensitive spike-triggered analysis has shown that additional filters are often required to account fully for the dimensionality of the stimulus set to which these cells respond11,12. Furthermore, the simple linear or quadratic transformations of filter outputs in the old model may need to replaced with more general point nonlinearities. Second, there are three more or less distinct gain control mechanisms that change responses depending on the combination of stimuli being presented. One of these regulates luminance gain when the average illumination

1647

t

1648

Complex cell

y x

y

Vm

t

Σ()2

Vm

x y x

Luminance Luminance gain control control gain Dynamic Dynamic gain gain control control

Contrast adaptation

ips

Contrast Contrast gain control control gain

Contextual modulation and surround suppression

b

N of the receptive field changes13. A second regulates contrast gain when the local average stimulus contrast varies14–17. A third regulates the temporal dynamics of responses depending on the temporal character of concurrent stimulation18. These gain controls in combination capture such effects as the ability of an otherwise ineffective stimulus to reduce the response to an effective one14, as well as the saturation of signals and decreases in latency at high stimulus contrasts17. The three gain controls may not be completely independent of one another, and their basis in neural circuits is an active area of study. Third, it has long been known that V1 responses depend on the history of stimulation because of the phenomenon of contrast adaptation19. This important time-dependent nonlinearity seems to depend partly on a neuron’s history of activity20,21 and partly on changes in synaptic input22–24, and it has a major influence on cortical responses. Fourth, the old standard model of a V1 neuron’s classical receptive field fails to deal with the existence of ‘hypercomplex’ cells25 and, more generally, with the suppression of responses by stimuli presented in regions outside the classical receptive field26. The addition of an inhibitory surround signal, originating in part from feedback signals from other cortical areas, extends the models to include this behavior27–29. Fifth and finally, capturing the behavior of neurons on short time scales (<50 ms) requires that the old standard Poisson-spiking models be extended to include more realistic (e.g., Hodgkin-Huxley) spike generation30–32. All of these elements can be combined to create a ‘new standard model’, schematically illustrated in Figure 1b as a synthesis and elaboration of the old standard models of Figure 1a. All these extensions of the old standard model of V1 neurons were discovered and characterized using combinations of synthetic stimuli like bars and gratings; none of them was found using natural stimuli. Moreover, there is no case in which the response of V1 neurons to natural stimuli has been shown not to be captured by the new standard model. However, the components of this new standard model were for the most part discovered in relative isolation from one another, and an important challenge is to develop a set of measurement techniques to recover all the components of the new standard model for a single cell. Modern spike-triggered techniques are making a start on this problem (for example, see ref. 33), but they still fall short of achieving this goal. For the moment, then, the new standard model represents our accumulated understanding of the mechanisms in play in visual cortical processing, but it cannot be specified for individual cells. This creates a complex challenge for those who would use the new standard model to predict responses to complex stimuli, like natural images, that engage most or all of the mechanisms diagrammed in Figure 1b.

t

Simple cell

ips

a

ips

Figure 1 ‘Standard’ models of visual cortical cells, old and new. (a) The standard models of simple and complex cells in V1, circa 1985. The responses of simple cells are predicted by convolving the stimulus with a single linear spatiotemporal filter (represented by the cube; each face of the cube schematizes a section through the center of the filter), and passing the output through a spiking threshold nonlinearity obeying Poisson statistics that converts membrane voltage Vm into spikes at a rate ips (impulses per second). The responses of complex cells are predicted by summing the squared outputs of two linear spatiotemporal filters that have an approximate quadrature phase relationship in both space and time, and then passing the resulting signal through the same spiking nonlinearity. (b) The new standard model of visual cortical cells, circa 2005. The various deficiencies of the models in a have been corrected by allowing for multiple initial filters combined with an arbitrary nonlinearity N, and including several additional mechanisms: gain controls for luminance, contrast and temporal dynamics; contrast adaptation; and surround suppression and context effects. The output is passed through a spiking nonlinearity incorporating more realistic (e.g., Hodgkin-Huxley) dynamics. See text for details.

Vm

...


PERSPECTIVE

Uses and abuses of natural stimuli There are two fundamentally different approaches that use natural stimuli. The principle behind the first approach is that one can deduce the properties of brain mechanisms for visual coding by reverse engineering: start with a set of natural scenes, and then infer the properties of the visual mechanisms that would best process those scenes. This is an appealing idea and seems both simple and direct, but the difficulty is that it is not clear what it means to say that a set of mechanisms are ‘best’ for processing natural scenes, because it is not clear what is to be optimized. One approach34 is based on the hypothesis that the early visual system is designed to reduce redundancy in the neural code35 and seeks the best set of mechanisms that satisfy a ‘sparse coding’ constraint. Others36,37 take a different approach and optimize their mechanisms for independence in the sense given by independent components analysis. A third approach finds mechanisms that satisfy a different criterion for independence38. Each of these approaches is self-consistent and compelling in the terms the authors define, but the problem is that goals the brain satisfies in choosing neural codes are unknown. Because the specific coding models that emerge from these approaches depend in detail on the optimization chosen, the results can give only a qualitative impression of the true mechanisms. Another limitation of most of these efforts is that the computations usually generate mechanisms that only resemble the simple cells of the old standard model (Fig. 1a) and rarely incorporate more than one or two isolated features of the new standard model (Fig. 1b). The second common use of natural images is more ambitious: to probe the visual system directly with natural images, motivated by the notion that the synthetic stimuli used in classic physiology experiments may not be sufficiently rich to uncover the full range of neuronal behavior1. Implicit in this approach is the assumption that synthetic stimuli are in some way impoverished or ‘simplistic’ and therefore somehow miss important features of visual response. The main—and in our view, crippling—challenge is that the statistics of natural images are complex and poorly understood. Without understanding the constituents of natural images, it is imprudent to use them to develop a well-controlled hypothesis-driven experiment. Ironically, the only way to know what importance



PERSPECTIVE to attach to different aspects of natural scene statistics is to have a comprehensive formal model of the neuron under study—but this is usually presented as the outcome rather than the foundation for the analysis. The most popular way to try to solve this problem is to use ‘reverse correlation’ or spike-triggered approaches. In a traditional reverse correlation experiment, an experimenter constructs a stimulus-response model by presenting random stimuli (such as Gaussian white noise) and determining the characteristics of the subset of stimuli that elicit spikes (such as the spike-triggered average or spike-triggered covariance; for reviews, see refs. 33, 39). To yield an unbiased result, these techniques require that the stimuli are uncorrelated (‘white’), that their intensities are chosen from a Gaussian distribution and (in whatever descriptive space the experimenter chooses) that they span all dimensions of interest. A number of authors have applied similar approaches using natural images by modifying their analysis techniques to try to compensate for the strongly non-white and non-Gaussian nature of these stimuli40–47. The problem is that when one extracts the subset of stimuli that are correlated with spikes, the correlations in the stimulus set make it difficult to determine why the stimulus-response correlation is present. Is it because of a mechanistic relationship between the stimulus and the neural circuit under study, or because of the correlations between the stimulus and other members of the stimulus set that may themselves be effective? Removing the effect of these correlations is straightforward if the system under study is simple. But if it has unknown architecture, it is necessary to make assumptions about the form of the underlying neural computation to distinguish the components of the response attributable to the neuron from those attributable to the correlations among the stimuli. The importance of these potential difficulties is difficult to determine with certainty, and it is certainly worthwhile to explore the question empirically. One ambitious attempt to use natural stimuli to study mechanism in cortical cells used a spike-triggered analysis and attempted to moderate the effects of stimulus correlations41. The data were fit to a model similar in form to the old standard model of Figure 1a, but this succeeded in capturing only a small fraction of the variance in the responses of V1 neurons to natural stimulus sequences, a result that falls well short of complete success. So why do these analyses provide such disappointing results? Some have argued that this reflects deep limitations in our understanding of V1, and they suggest that we must refocus our efforts away from traditional, simple stimuli toward stimuli with more naturalistic characteristics1 to overcome these limitations. But there is a simpler explanation. The significant and widely misunderstood limitation of spike-triggered approaches—using either synthetic or natural stimuli—is that they are not model-free; these techniques fit a specific model to the data, and in most cases this model is no more than a variant of the ‘old standard model’ shown in Figure 1a. Specifically, the data are used to fit a model in which the stimulus is first passed through one or more linear filters, the outputs of those filters are combined via an instantaneous nonlinearity, and noise is introduced into the system via a Poisson process. Without the elements of the new standard model (Fig. 1b), it is hardly surprising that the model performs poorly when put to a quantitative test. There is little doubt that the additional mechanisms represented in Figure 1b have a major role in natural scene responses. Yet the models evaluated in natural scene experiments lack these features; some work on subcortical processing suggests ways in which they might be incorporated48. Another major limitation of the spike-triggered approaches is the absence of a realistic spike generator. These models are used to predict firing rate over the course of one or two frames (10–50 ms), yet within this time frame, deviations from Poisson spiking have a profound impact, especially at high firing rates. Techniques have been


proposed to incorporate realistic, non-Poisson spiking into spike-triggered characterizations30,31, but these methods have not yet been generalized to models like the one in Figure 1b. So the parsimonious interpretation of the ‘failure’ of the old standard model when faced with natural scenes is not very grand. Instead of reflecting some special feature of natural images that can reveal hitherto unsuspected neural machinery, it may be that the failure reflects only the limitations of the models used to evaluate the data, limitations that have been made very clear by numerous experiments using synthetic stimuli. In particular, until methods exist to fit the full model shown in Figure 1b, we cannot know whether the deficiencies in our ability to predict responses to natural stimuli from spike-triggered analyses—using either natural or synthetic stimulus sets—are due to known mechanisms or to novel ones. Proper use of the natural and the artificial Fitting and testing models that are general enough to predict the responses to arbitrary stimuli remains a central goal of visual neurophysiology. This process is useful because it provides a quantitative analysis of how close we are to reaching our goal of describing the behaviors of these neurons, and responses to natural scenes will always be the standard against which models are tested. But it seems to us that the limitations on using natural stimuli to build models rather than to test them are too important to ignore. Consider a scenario in which we are able to fit all known mechanisms with an integrated model but find that this model fails to accurately predict the response properties to natural scenes. What then? We would want to establish what machinery is missing from our description, and at this point, natural stimuli themselves are of no assistance. To determine mechanism, we must return to the classical approach of presenting artificial stimuli that are carefully designed as efficient and principled tests of specific hypotheses. In constructing these stimuli, we would be foolish to ignore the composition and structure of natural images, but only with synthetic stimuli could we carefully control and test model elements of computational importance. The proof of success will be found in predicting the responses to natural stimuli, but the predictions themselves will be made from artificial ingredients. These disadvantages of natural stimuli do not invalidate their use in exploratory experiments. Indeed, for neurons with complex properties whose circuitry is unknown (such as those in higher cortical areas), these methods may be the best or even the only way to begin (for example, see ref. 49). But there are two phases to discovery. After the system is explored using stimuli that are chosen mostly for their effectiveness, model-building begins and our tools become the classical ones of hypothesis and test. In the study of primary visual cortex, we are fully engaged in that second stage. In our view, the proper use for studies of natural stimuli is to provide the benchmark against which success or failure can be measured. But success will require the model-guided use of artificial stimuli to uncover neuronal mechanism. ACKNOWLEDGMENTS The authors thank E.P. Simoncelli for discussions. COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/ 1. Olshausen, B.A. & Field, D.J. How close are we to understanding V1? Neural Comput. 17, 1665–1699 (2005). 2. Lennie, P. & Movshon, J.A. Coding of color and form in the geniculostriate visual pathway. J. Opt. Soc. Am. A 22, 1–21 (2005).

1649


PERSPECTIVE 3. Jones, J.P. & Palmer, L.A. The two-dimensional spatial structure of simple receptive fields in cat striate cortex. J. Neurophysiol. 58, 1187–1211 (1987). 4. Ringach, D.L. Spatial structure and symmetry of simple-cell receptive fields in macaque primary visual cortex. J. Neurophysiol. 88, 455–463 (2002). 5. Movshon, J.A., Thompson, I.D. & Tolhurst, D.J. Spatial summation in the receptive fields of simple cells in the cat’s striate cortex. J. Physiol. (Lond.) 283, 53–77 (1978). 6. Adelson, E.H. & Bergen, J.R. Spatiotemporal energy models for the perception of motion. J. Opt. Soc. Am. A 2, 284–299 (1985). 7. Spitzer, H. & Hochstein, S. A complex-cell receptive-field model. J. Neurophysiol. 53, 1266–1286 (1985). 8. De Valois, K.K., De Valois, R.L. & Yund, E.W. Responses of striate cortex cells to grating and checkerboard patterns. J. Physiol. (Lond.) 291, 483–505 (1979). 9. Smith, M.A., Bair, W. & Movshon, J.A. Signals in macaque striate cortical neurons that support the perception of glass patterns. J. Neurosci. 22, 8334–8345 (2002). 10. Creutzfeldt, O.D. & Nothdurft, H.C. Representation of complex visual stimuli in the brain. Naturwissenschaften 65, 307–318 (1978). 11. Touryan, J., Lau, B. & Dan, Y. Isolation of relevant visual features from random stimuli for cortical complex cells. J. Neurosci. 22, 10811–10818 (2002). 12. Rust, N.C., Schwartz, O., Movshon, J.A. & Simoncelli, E.P. Spatiotemporal elements of macaque V1 receptive fields. Neuron 46, 945–956 (2005). 13. Shapley, R.M. & Enroth-Cugell, C. Visual adaptation and retinal gain controls. Prog. Retinal Res. 3, 263–346 (1984). 14. Bonds, A.B. Role of inhibition in the specification of orientation selectivity of cells in the cat striate cortex. Vis. Neurosci. 2, 41–55 (1989). 15. Geisler, W.S. & Albrecht, D.G. Cortical neurons: isolation of contrast gain control. Vision Res. 32, 1409–1410 (1992). 16. Heeger, D.J. Normalization of cell responses in cat striate cortex. Vis. Neurosci. 9, 181–197 (1992). 17. Carandini, M., Heeger, D.J. & Movshon, J.A. Linearity and normalization in simple cells of the macaque primary visual cortex. J. Neurosci. 17, 8621–8644 (1997). 18. Bair, W. & Movshon, J.A. Adaptive temporal integration of motion in direction-selective neurons in macaque visual cortex. J. Neurosci. 24, 7305–7323 (2004). 19. Maffei, L., Fiorentini, A. & Bisti, S. Neural correlate of perceptual adaptation to gratings. Science 182, 1036–1038 (1973). 20. Carandini, M., Movshon, J.A. & Ferster, D. Pattern adaptation and cross-orientation interactions in the primary visual cortex. Neuropharmacology 37, 501–511 (1998). 21. Sanchez-Vives, M.V., Nowak, L.G. & McCormick, D.A. Membrane mechanisms underlying contrast adaptation in cat area 17 in vivo. J. Neurosci. 20, 4267–4285 (2000). 22. Movshon, J.A. & Lennie, P. Pattern-selective adaptation in visual cortical neurones. Nature 278, 850–852 (1979). 23. Chance, F.S., Nelson, S.B. & Abbott, L.F. Synaptic depression and the temporal response characteristics of V1 cells. J. Neurosci. 18, 4785–4799 (1998). 24. Muller, J.R., Metha, A.B., Krauskopf, J. & Lennie, P. Rapid adaptation in visual cortex to the structure of images. Science 285, 1405–1408 (1999). 25. Hubel, D.H. & Wiesel, T.N. Receptive fields and functional architecture of monkey striate cortex. J. Physiol. (Lond.) 195, 215–243 (1968). 26. Blakemore, C. & Tobin, E.A. Lateral inhibition between orientation detectors in the cat’s visual cortex. Exp. Brain Res. 15, 439–440 (1972). 27. Angelucci, A., Levitt, J.B. & Lund, J.S. Anatomical origins of the classical receptive field and modulatory surround field of single neurons in macaque visual cortical area V1. Prog. Brain Res. 136, 373–388 (2002).

1650

28. Cavanaugh, J.R., Bair, W. & Movshon, J.A. Nature and interaction of signals from the receptive field center and surround in macaque V1 neurons. J. Neurophysiol. 88, 2530–2546 (2002). 29. Sceniak, M.P., Ringach, D.L., Hawken, M.J. & Shapley, R. Contrast’s effect on spatial summation by macaque V1 neurons. Nat. Neurosci. 2, 733–739 (1999). 30. Keat, J., Reinagel, P., Reid, R.C. & Meister, M. Predicting every spike: a model for the responses of visual neurons. Neuron 30, 803–817 (2001). 31. Paninski, L., Pillow, J.W. & Simoncelli, E.P. Maximum likelihood estimation of a stochastic integrate-and-fire neural encoding model. Neural Comput. 16, 2533–2561 (2004). 32. Aguera y Arcas, B., Fairhall, A.L. & Bialek, W. Computation in a single neuron: Hodgkin and Huxley revisited. Neural Comput. 15, 1715–1749 (2003). 33. Simoncelli, E.P., Pillow, J.W., Paninski, L. & Schwartz, O. Characterization of neural responses with stochastic stimuli. in The Cognitive Neurosciences (ed. Gazzaniga, M.) 327–338 (MIT Press, Cambridge, Massachusetts, 2004). 34. Olshausen, B.A. & Field, D.J. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature 381, 607–609 (1996). 35. Barlow, H.B. Possible principles underlying the transformations of sensory messages. in Sensory Communication (ed. Rosenblith, W.A.) 217–234 (MIT Press, Cambridge, Massachusetts, 1961). 36. Bell, A.J. & Sejnowski, T.J. The independent components of natural scenes are edge filters. Vision Res. 37, 3327–3338 (1997). 37. van Hateren, J.H. & van der Schaaf, A. Independent component filters of natural images compared with simple cells in primary visual cortex. Proc. Royal Soc. Lond. B Biol. Sci. 265, 359–366 (1998). 38. Schwartz, O. & Simoncelli, E.P. Natural signal statistics and sensory gain control. Nat. Neurosci. 4, 819–825 (2001). 39. Chichilnisky, E.J. A simple white noise analysis of neuronal light responses. Network 12, 199–213 (2001). 40. Theunissen, F.E. et al. Estimating spatio-temporal receptive fields of auditory and visual neurons from their responses to natural stimuli. Network 12, 289–316 (2001). 41. David, S.V., Vinje, W.E. & Gallant, J.L. Natural stimulus statistics alter the receptive field structure of V1 neurons. J. Neurosci. 24, 6991–7006 (2004). 42. Touryan, J., Felsen, G. & Dan, Y. Spatial structure of complex cell receptive fields measured with natural images. Neuron 45, 781–791 (2005). 43. Sharpee, T., Rust, N.C. & Bialek, W. Analyzing neural responses to natural signals: maximally informative dimensions. Neural Comput. 16, 223–250 (2004). 44. Smyth, D., Willmore, B., Baker, G.E., Thompson, I.D. & Tolhurst, D.J. The receptivefield organization of simple cells in primary visual cortex of ferrets under natural scene stimulation. J. Neurosci. 23, 4746–4759 (2003). 45. Ringach, D.L., Hawken, M.J. & Shapley, R. Receptive field structure of neurons in monkey primary visual cortex revealed by stimulation with natural image sequences. J. Vis. 2, 12–24 (2002). 46. Prenger, R., Wu, M.C., David, S.V. & Gallant, J.L. Nonlinear V1 responses to natural scenes revealed by neural network analysis. Neural Netw. 17, 663–679 (2004). 47. Felsen, G., Touryan, J., Han, F. & Dan, Y. Cortical sensitivity to visual features in natural scenes. PLoS Biol. 3, e342 (2005). 48. Mante, V., Frazor, R.A., Bonin, V., Geislr, W.S. & Carandini, M. Independence of luminance and contrast in natural scenes and in the early visual system. Nat. Neurosci. 8 1690-1697 (2005). 49. Gross, C.G., Rocha-Miranda, C.E. & Bender, D.B. Visual properties of neurons in inferotemporal cortex of the macaque. J. Neurophysiol. 35, 96–111 (1972).




PERSPECTIVE

Analyzing receptive fields, classification images and functional images: challenges with opportunities for synergy Jonathan D Victor In neurophysiology, psychophysics, optical imaging and functional imaging studies, the investigator seeks a relationship between a high-dimensional variable, such as an image, and a categorical variable, such as the presence or absence of a spike or a behavior. The usual analysis strategy is fundamentally identical across these contexts—it amounts to calculating the average value of the high-dimensional variable for each value of the categorical variable and comparing these results by subtraction. Though intuitive and straightforward, this procedure may be inaccurate or inefficient and may overlook important details. Sophisticated approaches have been developed within these several experimental contexts, but they are rarely applied beyond the context in which they were developed. Recognition of the relationships among these contexts has the potential to accelerate improvements in analytic methods and to increase the amount of information that can be gleaned from experiments. Many systems neuroscience experiments are based around a common basic design—identifying an association between a high-dimensional variable, such as a complex stimulus, and a variable that can be easily categorized, such as the presence or absence of neural spiking. For example, in receptive field analysis, the investigator presents stimuli drawn from a large set of images1–8 or sounds9,10 and records one of two neuronal responses—the presence or absence of a spike. Psychophysical ‘classification image’11,12 studies take a conceptually related approach. In this case, the response is a subject’s detection or lack of detection of a target embedded in experimentally controlled noise. The response being measured is different, but the goal of the analysis is similar—to determine which aspects of the stimuli lead to a particular neural or behavioral response (Fig. 1a). Functional imaging studies also share this basic design, but the high-dimensional variable is no longer under the experimenter’s control, so some aspects of the problem are reversed. For example, the investigator repeatedly presents stimuli from one of two categories and records many examples of images of neural activity elicited by the stimuli. In a simple optical imaging experiment13 in

Jonathan D. Victor is in the Department of Neurology & Neuroscience, Weill Medical College of Cornell University, 1300 York Avenue, New York, New York 10021, USA. e-mail: [email protected]. Published online 23 November 2005; doi:10.1038/nn1607


visual cortex, the two categories might consist of vertical and horizontal gratings; in a functional brain imaging experiment14, the two categories might consist of a behavioral task and a baseline state or two contrasting sets of sensory stimuli. In these experiments, the goal of the analysis is to determine which aspects of the brain image are associated with each category (Fig. 1b). These are only the simplest prototypes. The multivariate quantity may be spatiotemporal sequences and not just static spatial images. The categorical quantity may have more than two possible values (for example, multiple orientations presented in an imaging experiment in visual cortex or temporal sequences of spikes in receptive field mapping). However, the essence of the analytic challenge in all of these experimental approaches remains one of relating a highly multivariate quantity to a categorical quantity. By far the most common, and perhaps most intuitive, strategy is fundamentally identical across these different experiments—it amounts to calculating the average value of the high-dimensional variable for each value of the categorical variable and subtracting one average from the other. This is also at the heart of reverse correlation techniques, which are pervasive in receptive field mapping or differential imaging. However, intuitive methods may be inaccurate or inefficient and may overlook important details. In this article, I will consider this class of problems abstractly, highlighting certain similarities and differences between the problems faced by the different experimental approaches. My first goal is to illustrate the reasons why the intuitive analysis strategy may not be the best, and the conceptual challenges that must be faced. I will then describe (without technical detail) several approaches to these challenges that have been recently developed and applied within individual experimental approaches. I suggest that wider recognition of the common conceptual problem being solved in the different contexts, including both a focus on aspects that are specific to one approach as well as application of methods beyond their original context, will benefit both the development of analytic tools and the analysis of data. Why analysis is challenging To understand what analysis of these datasets might entail, consider a geometric view of a highly reduced experiment in which the goal is to categorize an image consisting of only two pixels (Fig. 2). In each panel, each image is represented by a point whose horizontal (x) and vertical (y) coordinates represent the image value of the two pixels. The color assigned to each point corresponds to the category associated with that image, as determined experimentally.

1651


PERSPECTIVE Figure 1 Two kinds of experiments in which a highly multivariate quantity (yellow box) is to be related to a categorical quantity (purple box). (a) The multivariate quantity is the stimulus, and the categorical quantity is the response (as in receptive field mapping and classification image studies). (b) The multivariate quantity is the response, and the categorical quantity is the stimulus (as in optical imaging and functional brain imaging). In each case, the investigator determines the categories in advance (items outlined in red versus blue), but the instances of the multivariate quantity associated with each category (items outlined in pink versus pale blue) are determined from the experiment.

a

Response

Or

b

For definiteness, I will discuss the problem Or in terms of receptive field analysis, but the ideas apply equally well to classification image analysis. Consider an idealized (and highly simplified) receptive field mapping experiment (Fig. 2a) in which the stimulus consists of uncorrelated Gaussian noise at two pixels, and the neuron responds to only one pixel. The value of the stimulus in the pixel that drives the neuron is represented by the horizontal coordinate in Figure 2a. Because the pixel values in the stimuli are assumed to be uncorrelated, the stimuli form a circularly symmetric cloud. The neuron is assumed to respond to only one pixel, so the probability that a point is assigned to the two response categories depends only on its horizontal position. The demarcation between ‘blue’ responses (left half) and ‘red’ responses (right half) is not sharp, to represent the presence of neural noise that combines additively with the stimulus. The red and blue circles represent the results of analyzing these data by determining the average stimulus that led to each response. This is the approach taken in the standard method of reverse correlation. The horizontal displacement of these circles properly reveal a dependence of the neural response on only one pixel value. Moreover, the best partitioning of the stimuli into subsets that are associated with the two responses (based on the experimental data) is given by the perpendicular bisector of the line between these averages. This line separates the points that are closest to the center (average) of the red cloud from those that are closest to the center of the blue cloud. Thus, averaging suffices to answer two basic questions: what is the average stimulus corresponding to each response (Qcenter), and what is the best way to determine which response will be elicited by a particular stimulus (Qrule). Averaging will answer these questions when the multivariate data are independent and identically distributed, the system is linear and noise is additive. As the next panels show, relaxation of these conditions can lead to very different results. In cases where there is correlation between the two pixel values (Fig. 2b), averaging fails to capture the full relationship between stimulus and response. Such correlations are typically present in natural images. Even though the neuron only responds to the pixel represented by the horizontal coordinate, the average stimulus corresponding to each response class (circles) is now displaced along the diagonal, as a consequence of the pixel-to-pixel correlation present in the stimulus. Thus, the perpendicular bisector between the centers of the clouds, which is oblique, no longer represents the best rule for determining which response will be elicited by a given stimulus. Rather, the best rule remains a vertical line (as in Fig. 2a). In short, correlations within the stimulus set induce bias when averaging or reverse correlation is applied. This bias can be corrected if the stimulus correlations are known, and have

1652

Stimulus

a sufficiently simple form (for example, if the correlation is Gaussian). However, for stimulus sets such as natural stimuli, these conditions do not apply6,7. When the correlation structure of the stimulus set is sufficiently complex, bias correction is problematic and the ‘average’ stimulus may not be typical of any stimuli (Fig. 2b, inset). Even if the stimuli are uncorrelated, averaging will still be inadequate if the neuron is nonlinear. The next examples show how nonlinearity (non-additive combination of stimulus components) leads to additional complexities (Figs. 2c–e). Consider a neuron whose response depends on only one pixel, but this dependence has a small quadratic contribution in addition to the linear response (Fig. 2c). Because of this nonlinearity, stimuli with a large negative value at pixel x also lead to a response in the red category. More importantly, the optimal partitioning of the stimuli into classes corresponding to the two responses (that is, the description of what the two responses ‘represent’) consists of two vertical lines, not one. Averaging gives no hint of the bipartite distribution of red responses but rather misleadingly summarizes the distribution of red responses by a single point—which might even lie within the blue distribution. The latter situation would arise if the nonlinearity dominates, as would be the case in a stereotypical neuron with a symmetrical on-off response. In another example of nonlinearity, the model neuron is an ‘energy’ unit15 (Fig. 2d). It produces one response if x2 + y2 (the energy) exceeds a criterion and the other response if it does not. The optimal partitioning of the stimuli into the two classes is a circle that lies on the energy threshold. Yet another example of nonlinearity is an idealized edge detector (Fig. 2e): this neuron produces one response if x and y have opposite sign and the other response if they do not—that is, its response is determined by an interaction of the two pixel values, the product xy. In this case, the optimal rule for partitioning stimuli is described by intersecting lines that run along the axes. For both types of neuron, however, the means of the stimulus subsets that correspond to the two response classes coincide (red and blue circles in Fig. 2d,e). This means that a correlation analysis will not detect any signal at all, even though there is simple (but nonlinear) relationship between stimuli and response. When there is a mixture of linear and nonlinear contributions (Fig. 2c), a correlation analysis properly indicates the neuron’s dependence on input stimuli, but the signal-to-noise (that is, the separation of the mean stimuli of each class) is less than that in Figure 2a because the average includes stimuli with large negative x values.


PERSPECTIVE


Y

the clouds, in a way similar to the analogous nonlinear relationships shown in Figure 2c–e. X The challenges of analyzing a real dataset are substantially greater than these simple examples would suggest for several reasons. First, in a real dataset, variability would be much higher than in the illustration in X Y Figure 2, so that the clouds would overlap to a much greater extent. Second, although we f g h i j considered separately the effects of deviations from Gaussian uncorrelated noise (Fig. 2b,g), local nonlinearities (Fig. 2c,h), and spatial interactions (Fig. 2d,e,i,j), these phenomena are typically all present to some degree. Third, Figure 2 A geometric view of associations of multivariate and categorical data. In each panel, each the dimensionality of the multivariate datainstance of the multivariate (here, bivariate) data is represented by a point whose coordinates X and Y set is typically large: 100 to 1,000 in recepare the values of its two components (in this case, pixel intensities). The color assigned to each point tive field mapping or classification image indicates which of the two categories is associated with it. (a–e) Experiments in which the multivariate experiments, and 105 or larger in optical or quantity is the stimulus, and the categorical quantity is measured. (f–j) Experiments in which the multivariate quantity is measured, and the categorical quantity is the stimulus. a and f indicate the functional imaging experiments. Figure 2 simplest situation (uncorrelated and Gaussian bivariate data, with category linearly determined by considers only two-dimensional datasets. As one of its components). Other panels introduce correlation structure into the bivariate data (b,g), a result, a typical dataset represents only a nonlinearities (c,d,h,i), or both (e,j). The large colored symbols in each panel indicate the centroids of sparse sample of the multivariate distributhe respective clouds, and are superimposed in panels d, e, i and j. tion. Thus, in contrast to these examples where the dataset provides a good estimate For the converse problem, where the stimulus is the categorical vari- of the shape of the distribution of the multivariate quantity, good able and the multivariate quantity is measured (that is, in optical or estimates of the distribution of the multivariate quantity may not functional imaging), analogous situations can arise (Figs. 2f–j). In the be available. These various considerations have led to the development of sophissimple, ideal situation in which the stimulus activates only one pixel and there is uncorrelated, additive Gaussian measurement noise in both pix- ticated methods for analysis of such datasets. However, development els (Fig. 2f), the cloud of responses elicited by each stimulus is circularly of methods to deal with the effects of correlations have generally prosymmetric, and the horizontal displacement of these clouds represents ceeded along separate lines in receptive field mapping, image classifithe mean response. As described earlier (Fig. 2a), the mean response to cation, and imaging contexts. Attempts to analyze nonlinearities have each stimulus provides an unbiased estimate of the positions of these been developed primarily for the purpose of receptive field mapping, clouds, and the perpendicular bisector between these means is the opti- and not in the other contexts. mal partitioning of the responses into two categories. When a stimulus-driven response confined to one pixel adds to noise Two basic questions that is correlated across the two pixels (Fig. 2g)—for example, by vas- As highlighted in Figure 2, there are two basic questions that can be cular pulsations—the cloud of points that represents the pairs of pixel asked about the correspondence between a multivariate quantity and a values corresponding to each stimulus becomes elongated and oblique categorical one: Qcenter—what is the most typical value of the multivaribecause of the background correlation. The mean image elicited by each ate quantity that corresponds to each value of the categorical variable stimulus category is no different than that of Figure 2f. However, the (that is, where is the center of each cloud)—and Qrule—what is the best optimal way to discriminate between these sets of images is no longer rule for distinguishing these clouds. I do not mean to imply that the a vertical line: it is an oblique line whose slope is determined by the answers to these questions are the endpoints of the analysis or suffice to degree of correlation of the noise background. There is a second, more draw scientific conclusions, merely that they are common conceptual subtle, effect of the fact that the set of images elicited by each stimulus places to begin. forms an elongated cloud. The oblique axis does not contribute to sepaFor receptive field analysis, it is natural to focus on Qrule. Even for a ration of the two clouds, but variability along it reduces the reliability neuron with very simple properties, the center of the cloud will depend of the estimates of the clouds’ centers. Thus, noise correlation has two on the choice of stimuli used in an experiment. But the rule, which can effects: the optimal rule for discriminating the two image classes does be thought of as a computation performed on the stimuli and an indicanot correspond to the perpendicular bisector of the line between their tion of what a neuronal response represents, can in principle represent means, and a more reliable estimate of the difference between the means a more universal characterization of the neuron. can be obtained by eliminating dimensions that contain large variance For imaging experiments, Qcenter and Qrule are interesting and quite and small signal. distinct, even if the stimulus-response relationship is linear (Fig. 2g). As is the case for receptive field mapping, a nonlinear relationship This is because the multivariate quantities (the pixel values) are typibetween the stimulus and the multivariate response makes straightfor- cally highly correlated, both by the underlying physiology and the ward averaging inadequate for capturing the relationship between signal physics of imaging. The center of each cloud indeed indicates the and the stimulus that drove it (Fig. 2h–j). There is evidence that brain average response to each stimulus. However, the answer to Qrule prostates are manifest not only by mean activity but also by changes in power vides additional information—how best to ‘read out’ a pattern of and correlation structure16–18, suggesting that such nonlinear relation- activity. The distinction between the two questions cannot be avoided, ships indeed exist. These nonlinear relationships affect the mean position because the experimenter cannot control the correlation structure of of each cloud of points and the lines and curves that optimally separate the multivariate data.

a

b

c

d


e

1653


PERSPECTIVE The answers to Qcenter and Qrule are usually displayed as maps, but it is important to emphasize that these maps represent very different things. For Qcenter, the map is simply an instance of the multivariate quantity, but the answer to Qrule is a rule, not an instance of the multivariate quantity. If the rule is linear, it can be rendered as a map in the following manner: a linear rule is characterized as an assignment of weights (‘sensitivities’) that multiply each pixel of the stimulus; the partitioning is based on a sum of these products. Thus, the map that portrays a linear rule is a map of sensitivities (that is, quantities whose units are the reciprocal of the units of the multivariate data). Nonlinear rules can also be displayed in a map-like fashion, but here, too, the map describes rules to be applied to stimuli, not stimuli themselves. In the following sections, we consider a variety of approaches to answer Qcenter and Qrule. Uncorrelated multivariate data Standard analysis (subtraction of mean responses for imaging13, crosscorrelation for receptive field2,3,19 or classification image11 analysis), address Qcenter . Qcenter and Qrule are equivalent only under very special circumstances (Fig. 2). When the linearity condition is relaxed but the multivariate quantity remains independent and identically distributed (Fig. 2c–e), Qrule can be determined by generalizations of the crosscorrelation approach. The desired characterization corresponds to the ‘kernels’ of Wiener-like2,3,19 procedures. For example, the basic computation in the recently developed spike-triggered covariance method4,5,8 is equivalent to the standard Lee-Schetzen cross-correlation estimate for the second-order kernel19,20, followed by diagonalization. In settings such as receptive field or classification image analysis in which the investigator has control over the multivariate quantity, the use of ‘designer’ stimulus sets (sinusoidal sums21 and m-sequences1) are particularly advantageous. Such approaches are often effective in characterizing nonlinear (as well as linear) response properties. This is because these finite stimulus sets are, in some sense, more nearly uncorrelated than a random sample drawn from a large uncorrelated ensemble. However, ‘designer’ methods cannot be applied to imaging data or to situations in which natural scenes6,7,22 or sounds10,23 are used for receptive field determination. These multivariate stimulus sets typically contain strong correlations that cannot be controlled or fully characterized. Here, the answer to Qrule provides information about the stimulusresponse relationship that is not available from Qcenter . Moreover, as will be sketched below, Qrule, along with the covariance structure of the stimulus, can provide a better estimate of Qcenter than averaging. Correlated multivariate data, linear relationship Even when correlations within the high-dimensional variable are present, linear regression identifies the linear function of a stimulus sequence that does the best job (in the mean-squared sense) of predicting the binary response (spike versus no spike; target seen versus target not seen). Thus, it is a natural approach to finding Qrule under the assumption that the stimulus-response relationship is linear and correlations are present within the high-dimensional variable. Also, it can be extended in a Wiener-like fashion to nonlinear relationships, as has been done in the context of receptive field analysis24. Of note, linear regression was used in the original description of the classification image method12. Other than a few exceptions25, this approach is not often taken because there are statistical difficulties that confound direct application of linear regression to the experimental contexts we are considering. However, as we next describe, there are recently developed techniques that can surmount these difficulties. In essence, finding Qrule via linear regression requires two steps (Supplementary Note online): (i) estimating the covariance matrix

1654

S of the multivariate set, and (ii) multiplication of the mean difference between the two multivariate clouds by S–1, the matrix inverse of S. The covariance matrix S is a symmetric array whose entries sjk are the correlations of the jth and kth pixels. For independent, identically distributed data, S is proportional to the identity matrix, and linear regression reduces to simple subtraction, as S–1 is also proportional to the identity. In imaging data, pixel values are coupled by motion, light scatter, blood flow and other physiologic factors; in receptive field analysis, pixel values are coupled by the statistics of natural scenes23,26. Thus, linear regression and the related approaches described below differ fundamentally from the subtraction method, as S is far from a multiple of the identity matrix. The main pitfall in linear regression is that of overfitting: namely, the estimated answer to Qrule may work well for the particular experimental sampling of the multivariate dataset but does not generalize to a larger sample. Linear regression provides an answer to Qrule that generalizes if the correlation structure of the multivariate data is well characterized by the experimental sample. This characterization requires enough samples to estimate its covariance matrix S, along with an assumption, typically that the distribution of the multivariate dataset is Gaussian, to determine higher-order correlations from second-order correlations. For optical or functional imaging data, there are typically many more pixels (105 to 106) than samples (∼104). For receptive field or classification image data, the undersampling problem is present but less severe (103 to 104 pixels; number of samples in the same range), but the correlation structure (especially of natural images) is likely to be very non-Gaussian. Because of undersampling and/or non-Gaussian characteristics of the multivariate data, direct application of standard linear regression is likely to produce results that are worse than simple subtraction. However, extensions of linear regression27–29 developed for imaging are applicable to the undersampled regime by recasting the problem in a form that does not require explicit inversion of S. These approaches focus on the eigenvectors of the covariance matrix S and related matrices (Supplementary Note). The eigenvectors of S (which are its ‘principal components’) are a small set of images from which all images can be reconstructed. Eigenvectors can be ranked in importance according to their eigenvalue. The larger the eigenvalue corresponding to an eigenvector, the greater the extent to which the stimulus set explores the corresponding image direction. On the basis of the eigenvalues, one can select a subset of eigenvectors within which a linear regression–like procedure can be carried out accurately. In the example of Figure 2g, such a procedure would correspond to restricting the estimation process to the direction that crosses the narrow axes of the ellipse and forgoing an attempt to estimate the position of the ellipse centers along their long axes, where variability is greater. As detailed in the Supplementary Note, there are several useful variations on this theme. The ‘truncated inverse’ method28 selects the eigenvectors of S whose eigenvalues are sufficiently large. More elaborate approaches select a subspace on the basis of not only on the overall covariance structure of S but also on the covariances within the subset of stimuli that lead to each categorical response. This includes the classic Fisher27 discriminant method, which restricts analysis to the one-dimensional projection that optimally discriminates between Gaussian fits to the two clouds. The ‘indicator function’ method30 projects into several dimensions (the ‘canonical variates’), chosen on the basis of a significance criterion. Canonical variates are also the basis of a method for characterizing spatiotemporal aspects of images acquired in fast fMRI studies31,32. A further variation is the ‘generalized indicator function method’29, which considers the eigenvectors of a linear mix of S and the within-group covariances and weights these eigenvectors in a graded fashion (Supplementary Note).



PERSPECTIVE These approaches all represent methods of dimensionality reduction—selection of subspaces that are likely to contain a large signalto-noise ratio (SNR). Within this general framework, independent component analysis33 can be viewed as a strategy for using higher moments of the images to identify mixtures of subspaces that are likely to contain signal34. There are other approaches to enhance SNR in functional imaging data that exploit specific features of such images (such as vascular artifacts35), but because these approaches are domain specific, we do not discuss them here. Dimensional reduction can also be viewed as a form of ‘regularization’36–38 to avoid the pitfalls of estimating the covariance structure from very limited data. The generalized indicator function method29 can be viewed as a regularized determination of canonical variates37,38. ‘Ridge regression’38 (Supplementary Note) is a regularization strategy (widely applied outside of neuroscience) that chooses a compromise between the estimated covariance matrix and the identity matrix. Variations of ridge regression that take into account smoothness constraints have recently been used to extract receptive field maps from natural stimuli, both in the visual22 and auditory10 domains, but apparently have not been applied in optical or functional imaging. Correlated multivariate data, nonlinear relationship In functional imaging, analytical methods (including the recent developments reviewed above) assume that the relationship between the imaged signal and neural activity is linear39, and the analytical focus has been on determining this relationship when the activity-dependent signal represents a very small fraction of the image. In receptive field analysis, it is generally considered that the neural response will be substantially greater than background when the stimulus is appropriate, but it is recognized that the stimulus-response relationship may not be linear6,40. This has driven the development of efficient methods for identifying nonlinear relationships that succeed even in the presence of strongly non-Gaussian multivariate data6,40. Moreover, the neural response, even if considered categorical, is generated in a manner that has stochastic and dynamic aspects. Understanding the implications of spike generation for the convergence and bias credentials of various estimation techniques40, and the interpretation of the resulting receptive fields41,42 is another focus of current work. Time for a convergence? One can readily identify several reasons for the separate development of analytical techniques in receptive field and classification image analysis on the one hand, and optical and functional imaging on the other, owing to various differences between these settings. However, I argue that the implications of these differences are less compelling than generally assumed and, consequently, that there are likely to be substantial opportunities for synergy. Most obviously (Fig. 1), the categorical and multivariate characteristics of stimulus and response are swapped. This has two implications. If the experimenter is willing to choose a ‘designer’ stimulus set, then there are opportunities for improved experimental design in receptive field or classification image analysis that are not available for optical or functional imaging. We do not focus on these here, however, and these are irrelevant to receptive field and classification image studies using natural stimuli. The other implication is that this distinction leads to a difference in how noise is considered. In imaging, no threshold is typically postulated; rather, neuronal and measurement variability smoothly combine with the ‘signal’. For receptive field and classification image analysis, it is generally assumed that after a computation is performed on the multivariate quantity, there is a threshold (such as a firing threshold or a decision threshold) that may be in part stochastic.


However, many treatments seek a rule for the firing rate (or decision variable) that minimizes the mean-squared prediction error, rather than an explicit maximum-likelihood solution for a model with a threshold. Because a mean-squared error criterion is essentially a maximum-likelihood criterion for an assumed Gaussian noise, it is tantamount to ignoring the statistical consequences of the threshold. Consequently, unless detailed dynamics of spike trains are of interest, the stimulus-response inversion does not prevent application of analysis methods for receptive field mapping to imaging, and viceversa. Explicit modeling of spike train dynamics may result in further improvement and insight40–42. But dynamics (that is, the effect of stimulus sequence and time course on response) are also present in an optical or functional experiment, suggesting that techniques to examine dynamics developed for receptive field analysis might usefully be applied to imaging. Another apparent difference between these settings is what is typically considered limiting. In classification image experiments, the number of trials that can be obtained may be limited to 103 to 104, as each trial requires an explicit behavioral response. In imaging, the main hurdle to analysis is usually considered to be an intrinsically low SNR: 1 in 103 to 1 in 104 for optical imaging, and 1 in 102 for fMRI. There may also be sources of variability that lead to highly structured artifacts such as pulsatile movements of the tissue. Because receptive field studies often have the implicit goal of prediction of responses to stimuli outside of the experimental sample, undersampling of the stimulus space is often considered to be the main problem, as (even when the SNR is high) a far greater sampling is required to construct a nonlinear model than to construct a linear model. However, SNR, number of trials and sampling of the stimulus space are always limiting. Analytical approaches that make better use of a given data set to identify smaller signals, provide greater spatial detail or shorten the experiment time to obtain results of a given quality are always useful, especially given the increasing availability of computational resources and the costs (not just direct economic) of obtaining neurophysiological data. In sum, analyzing imaging data in the presence of structured noise is closely analogous to identification of receptive fields from ‘natural’ stimulus sets6, which have strong but incompletely determined statistical structure. In optical and functional imaging, the relationship of the multivariate quantity to a behavioral index is typically assumed to be linear39, whereas a linear relationship is not always assumed in classification image analysis43,44 and in receptive field mapping2,3. But there is increasing recognition16–18 that stimulus-induced changes in the correlation structure of brain activity, and not just its mean level, are behaviorally and mechanistically relevant. Identifying such changes in imaging data is closely allied with identifying nonlinear aspects of a neural stimulusresponse relationship in the receptive field mapping context. Even for imaging studies that do not seek a nonlinear relationship between stimuli and the imaging data, inclusion of a nonlinearity in the model might nevertheless benefit the goal of signal extraction: that is, by identifying the rule that distinguishes responses to the members of the stimulus set. An analytic procedure that forces a nonlinear relationship to be modeled as a linear one necessarily causes some ‘signal’ (deviation of a systematic nonlinear response from a linear one) to appear as ‘noise’. Application of traditional nonlinear kernel methods is problematic, as the introduction of a large number of free parameters would likely defeat any benefit of capturing more signal. However, it would be very worthwhile to explore the use of recent receptive field mapping methods that seek simple nonlinear relationships in an efficient manner6,40. It should be emphasized that even if the relationship between the imaging signal and neural activity were strictly linear39, one would still expect this approach to be of value. Such linearity applies to the mean signal,

1655


PERSPECTIVE not its fluctuations, and in the noisy, multivariate regime, the latter may dominate the stimulus-response relationship. Moreover, mean neural activity may not be linearly related to the categorical variable—an issue that becomes relevant in experiments in which the categorical variable can take on more than two values, such as an orientation or spatial frequency experiment. Conversely, to maximize the ability to test mechanistic or functional receptive field models, it is necessary to identify not only the large, readily resolvable components, but small contributions as well, as illustrated by the recent analysis of complex cell receptive fields in terms of spiketriggered covariances5,8. Moreover, it is evident that further insight into neuronal properties can be gleaned from receptive field characterization with stimulus sets have complex statistics, including natural scenes6,7,23. For both of these reasons, analytical methods developed for imaging that improve signal-to-noise by using subspace selection38 may be useful. One might even envision that such methods could be further refined (for receptive field characterization) by guiding the estimation of covariances by the known statistical regularities of natural images26,45,46. Conclusion As detailed above, the problem of identifying the relationship between a highly multivariate quantity and a categorical quantity (such as a discrete stimulus, a behavioral response or the presence of an action potential) is deceptively simple. It has been approached by a variety of analytical techniques, most often motivated by particular features of one of these contexts. We claim that the distinctions between experimental domains are not as deep as generally assumed, and we speculate that opportunities for progress will result from applying these techniques (or the ideas behind them) beyond their original domain. The above considerations are only starting points, not an exhaustive list, and it is likely that the benefits will be relatively specific to particular experimental situations and goals. Whether cross-application of such methods and ideas will result in new qualitative insights, or merely incremental advances, is difficult to predict. Nevertheless, recognition of the close relationships between the mathematical challenges in these domains will enrich and accelerate the development of improved analytical techniques. Note: Supplementary information is available on the Nature Neuroscience website. ACKNOWLEDGMENTS The author thanks S. Klein, T. Yokoo, L. Paninksi and P. Buzás for helpful discussions.This work is supported in part by grants EY7977 and EY9314 to J.D.V. COMPETING INTERESTS STATEMENT The author declares that he has no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/ 1. Sutter, E. in Nonlinear Vision: Determination of Neural Receptive Fields, Function, and Networks (eds. Pinter, R. & Nabet, B.) 171–220 (CRC Press, Cleveland, 1992). 2. Chichilnisky, E.J. A simple white noise analysis of neuronal light responses. Network 12, 199–213 (2001). 3. Marmarelis, P.Z. & Naka, K. White-noise analysis of a neuron chain: an application of the Wiener theory. Science 175, 1276–1278 (1972). 4. Simoncelli, E., Paninski, L., Pillow, J. & Schwartz, O. in The Cognitive Neurosciences 3rd edn. (ed. Gazzaniga, M.) (MIT Press, Cambridge, Massachusetts, 2004). 5. Touryan, J., Lau, B. & Dan, Y. Isolation of relevant visual features from random stimuli for cortical complex cells. J. Neurosci. 22, 10811–10818 (2002). 6. Sharpee, T., Rust, N.C. & Bialek, W. Analyzing neural responses to natural signals: maximally informative dimensions. Neural Comput. 16, 223–250 (2004). 7. David, S.V., Vinje, W.E. & Gallant, J.L. Natural stimulus statistics alter the receptive field structure of V1 neurons. J. Neurosci. 24, 6991–7006 (2004). 8. Rust, N.C., Schwartz, O., Movshon, J.A. & Simoncelli, E.P. Spatiotemporal elements of macaque V1 receptive fields. Neuron 46, 945–956 (2005).

1656

9. Escabi, M.A. & Schreiner, C.E. Nonlinear spectrotemporal sound analysis by neurons in the auditory midbrain. J. Neurosci. 22, 4114–4131 (2002). 10. Machens, C.K., Wehr, M.S. & Zador, A.M. Linearity of cortical receptive fields measured with natural sounds. J. Neurosci. 24, 1089–1100 (2004). 11. Eckstein, M.P. & Ahumada, A.J., Jr. Classification images: a tool to analyze visual strategies. J. Vis. 2, 1x (2002). 12. Ahumada, A.J., Jr & Lovell, J. Stimulus features in signal detection. J. Acoust. Soc. Am. 49, 1751–1756 (1971). 13. Grinvald, A. Optical imaging of architecture and function in the living brain sheds new light on cortical mechanisms underlying visual perception. Brain Topogr. 5, 71–75 (1992). 14. Kwong, K.K. et al. Dynamic magnetic resonance imaging of human brain activity during primary sensory stimulation. Proc. Natl. Acad. Sci. USA 89, 5675–5679 (1992). 15. Ohzawa, I., DeAngelis, G.C. & Freeman, R.D. Encoding of binocular disparity by complex cells in the cat’s visual cortex. J. Neurophysiol. 77, 2879–2909 (1997). 16. Pesaran, B., Pezaris, J.S., Sahani, M., Mitra, P.P. & Andersen, R.A. Temporal structure in neuronal activity during working memory in macaque parietal cortex. Nat. Neurosci. 5, 805–811 (2002). 17. Kenet, T., Bibitchkov, D., Tsodyks, M., Grinvald, A. & Arieli, A. Spontaneously emerging cortical representations of visual attributes. Nature 425, 954–956 (2003). 18. Rodriguez, E. et al. Perception’s shadow: long-distance synchronization of human brain activity. Nature 397, 430–433 (1999). 19. Lee, Y. & Schetzen, M. Measurement of the kernels of a nonlinear system by cross-correlation. Int. J. Control 2, 237–254 (1965). 20. de Ruyter van Steveninck, R. & Bialek, W. Real-time performance of a movement sensitive neuron in the blowfly visual system: coding and information transfer in short spike sequences. Proc. R. Soc. Lond. B. 234, 379–414 (1988). 21. Victor, J.D., Shapley, R.M. & Knight, B.W. Nonlinear analysis of cat retinal ganglion cells in the frequency domain. Proc. Natl. Acad. Sci. USA 74, 3068–3072 (1977). 22. Smyth, D., Willmore, B., Baker, G.E., Thompson, I.D. & Tolhurst, D.J. The receptivefield organization of simple cells in primary visual cortex of ferrets under natural scene stimulation. J. Neurosci. 23, 4746–4759 (2003). 23. Theunissen, F.E. et al. Estimating spatio-temporal receptive fields of auditory and visual neurons from their responses to natural stimuli. Network 12, 289–316 (2001). 24. Korenberg, M.J., Bruder, S.B. & McIlroy, P.J. Exact orthogonal kernel estimation from finite data records: extending Wiener’s identification of nonlinear systems. Ann. Biomed. Eng. 16, 201–214 (1988). 25. Levi, D.M. & Klein, S.A. Classification images for detection and position discrimination in the fovea and parafovea. J. Vis. 2, 46–65 (2002). 26. Simoncelli, E.P. & Olshausen, B.A. Natural image statistics and neural representation. Annu. Rev. Neurosci. 24, 1193–1216 (2001). 27. Fisher, R.A. The use of multiple measurements in taxonomic problems. Ann. Eugen. 7, 179–188 (1936). 28. Gabbay, M., Brennan, C., Kaplan, E. & Sirovich, L. A principal components-based method for the detection of neuronal activity maps: application to optical imaging. Neuroimage 11, 313–325 (2000). 29. Yokoo, T., Knight, B.W. & Sirovich, L. An optimization approach to signal extraction from noisy multivariate data. Neuroimage 14, 1309–1326 (2001). 30. Everson, R., Knight, B.W. & Sirovich, L. Separating spatially distributed response to stimulation from background. I. Optical imaging. Biol. Cybern. 77, 407–417 (1997). 31. Friston, K.J., Frith, C.D., Frackowiak, R.S. & Turner, R. Characterizing dynamic brain responses with fMRI: a multivariate approach. Neuroimage 2, 166–172 (1995). 32. Worsley, K.J., Poline, J.B., Friston, K.J. & Evans, A.C. Characterizing the response of PET and fMRI data using multivariate linear models. Neuroimage 6, 305–319 (1997). 33. Bell, A.J. & Sejnowski, T.J. An information-maximization approach to blind separation and blind deconvolution. Neural Comput. 7, 1129–1159 (1995). 34. Thomas, C.G., Harshman, R.A. & Menon, R.S. Noise reduction in BOLD-based fMRI using component analysis. Neuroimage 17, 1521–1537 (2002). 35. Carmona, R.A., Hwang, W.L. & Frostig, R.D. Wavelet analysis for brain function imaging. IEEE Trans. Med. Imaging 14, 556–564 (1995). 36. Tikhonov, A.N. & Arsenin, V.Y. Solutions of Ill-Posed Problems (Wiley, New York, 1977). 37. Hastie, T., Buja, A. & Tibshirani, R. Penalized discriminant analysis. Ann. Stat. 23, 73–102 (1995). 38. Hastie, T., Tibshirani, R. & Friedman, J. The Elements of Statistical Learning: Data Mining, Inference, and Prediction (Springer-Verlag, New York, 2001). 39. Boynton, G.M., Engel, S.A., Glover, G.H. & Heeger, D.J. Linear systems analysis of functional magnetic resonance imaging in human V1. J. Neurosci. 16, 4207–4221 (1996). 40. Paninski, L. Convergence properties of three spike-triggered analysis techniques. Network 14, 437–464 (2003). 41. Aguera y Arcas, B. & Fairhall, A.L. What causes a neuron to spike? Neural Comput. 15, 1789–1807 (2003). 42. Pillow, J. & Simoncelli, E. Biases in white noise analysis due to non-Poisson spike generation. Neurocomputing 52–54, 109–115 (2003). 43. Neri, P. Estimation of nonlinear psychophysical kernels. J. Vis. 4, 82–91 (2004). 44. Neri, P. & Heeger, D.J. Spatiotemporal mechanisms for detecting and identifying image features in human vision. Nat. Neurosci. 5, 812–816 (2002). 45. Field, D.J. Relations between the statistics of natural images and the response properties of cortical cells. J. Opt. Soc. Am. A 4, 2379–2394 (1987). 46. Dong, D.W. & Atick, J.J. Statistics of natural time-varying images. Netw. Comput. Neural Syst. 6, 345–358 (1995).


400 Behavioral index

Y A

6,000 4,000 2,000 0 0

1 2 3 Blocks of 5 trials

Percentage FP slope

100

4

Y

A

20 1 Hz

100 Y

75

AU

15

*

10 5

AI

40

0 0

20 Time (min)

40

60

Young R = 0.30

I

R 2 = 0.20 P > 0.01

P < 0.02

20

Y U

Aged 2

30

10 0 –10 100

200 300 Behavioral index

125

400

100

PP-1 Hz

100 75 Y AU

50 +APV

40

200 300 Behavioral index 25 20

20 Time (min)

40

10 5

60

Young

Y U

I

Aged 2

2

R = 0.15 P > 0.2

20

*

AI 0

30

400

15

0

25 –20

e

200

125

50 –20

c

300

Percentage LTD

b

d Figure 1 Switch in an LTD mechanism with age. (a) Aged (A) rats are impaired relative to young (Y) rats on training trials (left, F1,53 ¼ 26.12, P o 0.0001) and on probe trials (right, F1,53 ¼ 36.68, P o 0.0001). Note there is no difference in initial performance (left panel, first data point). Right panel shows behavioral index in probe trials. Dotted line indicates the cut-off at 240 for dividing aged rats into aged-unimpaired (AU) and aged-impaired (AI) groups. All behavioral scores derive from a proximity measure (distance from platform sampled 10 times per s). High values reflect less accurate search on training trials (search error) and probe trials (behavioral index) (see Supplementary Methods online). (b) Larger NMDAR-LTD in Y compared to AU and AI. Right, comparison of average LTD magnitude during the last 10 min. (c) NMDAR-LTD magnitude correlates with the behavioral index in young, but not aged, rats. (d) Non-NMDAR-LTD is larger is AU compared to Y and AI. Right, comparison of the non-NMDAR-LTD magnitude. (e) NonNMDAR-LTD magnitude correlates with the behavioral index in aged, but not young, animals. *P o 0.02, Fisher’s protected least squares difference (PLSD) post-hoc analysis. Gray circles, Y; white squares, AU; black squares, AI. Error bars indicate s.e.m.

8,000

Percentage LTD

On average, cognitive abilities decline with age, yet a recognizable subpopulation of individuals maintains mental abilities. At the level of neural networks, functional imaging studies have revealed that unique patterns of brain activation distinguish high-performing older individuals from younger adults1. Although deficits in synaptic plasticity have been identified in aged animals with cognitive impairment2, few studies have examined the cellular basis for plasticity in aged animals with preserved cognitive abilities. Here we examined whether differences in the magnitude or the underlying mechanisms of wellcharacterized forms of synaptic plasticity were characteristic of high-functioning older rats. An assessment of spatial cognition that depends on hippocampal function reveals reliable individual differences in the cognitive status of healthy aged rodents3–5. A previous investigation using this study population reported a significant correlation between the activity of

Search error (cm)

Some individuals maintain high cognitive functioning at older ages. Here we show that mechanisms for long-term depression differ in aged rodents that maintain cognitive performance compared to young adults. Our results imply that cognitive abilities may be sustained in aged individuals by a switch in synaptic plasticity mechanisms.

a

Percentage FP slope

Hey-Kyoung Lee1–4, Sun Seek Min1,2,4, Michela Gallagher2 & Alfredo Kirkwood1

phospholipase C (PLC) in the hippocampus and cognitive outcome across the spectrum of performance among aged rats5. Because PLC has been implicated in long-term depression (LTD)6,7, we tested whether the magnitude of LTD is similarly related to cognitive performance among aged animals.

Percentage LTD

NMDA receptor–independent long-term depression correlates with successful aging in rats

Percentage LTD


B R I E F C O M M U N I C AT I O N S

R = 0.76 P < 0.0001

10 0 –10 100


400

100


400

1Department

of Neuroscience, The Mind/Brain Institute, and 2Department of Psychological and Brain Sciences, Johns Hopkins University, Baltimore, Maryland 21218, USA. 3Present address: Department of Biology, Neuroscience and Cognitive Science (NACS) Program, University of Maryland, College Park, Maryland 20742, USA. 4These authors contributed equally to this work. Correspondence should be addressed to M.G. ([email protected]) or A.K. ([email protected]). Received 29 August; accepted 21 September; published online 13 November 2005; doi:10.1038/nn1586

NATURE NEUROSCIENCE VOLUME 8

[

NUMBER 12

[

DECEMBER 2005

1657


1.25 1.00

1.0

0.0

0.75

g

100 75 AU AU + U7

+APV 25 –20

0

20 40 Time (min)

125

1 2 FV amp (mV)

PP-1 Hz

25 –20

AU AU + M

0

20 40 Time (min)

60

25

20

0 20 40 Time (min)

60

80

Aged

Young 2 R = 0.5 P < 0.03

*

10

Y

AU

AI

All together R 2 = 0.23 P < 0.05

R 2 = 0.38 P < 0.001

10 0

20 15

150 200 300 400 100 Behavioral index

NMDAR-LTD

25

Y

10

AI AU

5 Age

200 300 400 Behavioral index

non-NMDAR-LTD

20

AU

15 10

AI

Y

5 Age

In this study, we measured LTD in area CA1 of hippocampal slices prepared from behaviorally characterized young (6 months of age) and aged (24 months of age) outbred Long-Evans rats (Fig. 1). The aged rats included a substantial subset that performed within the range of young adult rats in a hippocampal-dependent assessment of spatial learning (Fig. 1a, right panel). These agedunimpaired (AU) rats were compared to both young adults (Y) and aged rats that had deficits in the behavioral assessment (AI: agedimpaired). First we measured LTD induced with a standard 1 Hz, 15 min protocol (a train of current pulses of 0.2 ms duration), which produces an NMDA receptor–dependent LTD (NMDAR-LTD, Supplementary Fig. 1 online)8. Young animals had, on average, larger NMDAR-LTD than the aged animals, but there was no difference between the AU and AI groups (Fig. 1b) (at 1 hr after onset of 1 Hz, mean NMDAR-LTD ± s.e.m., relative to baseline, was as follows: Y: 86 ± 1.5%, n ¼ 63 slices, 20 rats; AU: 94 ± 2.7%, n ¼ 16 slices, 4 rats; AI: 90 ± 1.5%, n ¼ 39 slices, 10 rats; ANOVA: F2,115 ¼ 5.497, P o 0.006). More importantly, we found that the average magnitude of NMDAR-LTD per animal correlated significantly with the behavioral index only in young, but not in aged, rats (Fig. 1c).

1658

Figure 2 Basal synaptic transmission and PLC dependence of LTD. (a) Similar paired-pulse facilitation (PPF) ratio across different groups. (b) No difference in AMPAR–mediated basal synaptic transmission. Field potential (FP) slopes were plotted against the presynaptic fiber volley (FV) amplitudes. (c) Pharmacologically isolated NMDAR–mediated synaptic transmission (by 10 mM NBQX and 0 mM MgCl2) was less in AI, but not different between Y and AU. FP amplitudes were plotted against the FV amplitudes. (d) Non-NMDAR-LTD in AU is blocked by a PLC inhibitor (U7: 10 mM U73122). (e) Non-NMDAR-LTD in AU is significantly reduced (P o 0.003) by bath application of an inhibitor cocktail (M: 10 mM MPEP, 10 mM prazosin and 5 mM atropin, antagonists of mGluR5, a1-adrenergic receptor and M1-mAchR, respectively). (f) Carbachol (CCh)-induced LTD is less in AI compared to AU and Y animals. Right, comparison of the average CCh-LTD magnitude. (g) CCh-LTD magnitude correlates with the behavioral index across ages. (h) Age-dependent switch in an LTD mechanism. NMDARLTD declines with age (left), whereas non-NMDAR-LTD increases in AU rats (right). *P o 0.001, Fisher’s PLSD post-hoc analysis. Gray circles, Y; white squares, AU; black squares, AI. Error bars indicate s.e.m.

30

Y AU AI

20

0

+APV

60

50

30

Y AU AI

1

50

75

0

2

75

100

0 –20

3

100

CCh

125

–10 100 200 300 Behavioral index

h

e

PP-1 Hz

50

4

0 0.0 0.25 0.50 FV amp (mV) Percentage FP slope

125

40 Percentage LTD

100 1,000 10,000 ISI (ms)

Percentage LTD

Percentage FP slope Percentage FP slope

f

Percentage LTD


10

d

Y AU AI

0.5

5 NR-FP amp (mV)

1.50

c

1.5

Percentage LTD

PPF ratio

b

Y AU AI

1.75

FP slope (mV/ms)

a

Next we examined an NMDAR–independent form of LTD (non-NMDAR-LTD), which was induced by delivering paired pulses (50-ms interstimulus interval) repeated at 1 Hz for 15 min (PP-1Hz) in the presence of a blocker of NMDAR (100 mM DL-2-amino-5-phosphonopentanoic acid, DL-APV)9. The AU group showed a significantly larger non-NMDAR-LTD than either the young or the AI groups (Y: 90 ± 2.2%, n ¼ 33 slices, 11 rats; AU: 79 ± 2.2%, n ¼ 22 slices, 8 rats; AI: 87 ± 2.1%, n ¼ 20 slices, 8 rats; ANOVA: F2,72 ¼ 5.917, P o 0.005; Fig. 1d). Moreover, we found a significant correlation between the magnitude of non-NMDAR-LTD and the behavioral index in aged, but not in young, animals (Fig. 1e). Collectively, these results suggest that the behavioral index correlates with the magnitude of NMDAR-LTD in young animals and with that of non-NMDAR-LTD in aged animals. We then examined whether the differences in LTD mechanisms in young and aged rats could be due to changes in basal synaptic transmission. Paired-pulse facilitation ratios and input-output curves, which respectively reflect presynaptic function and AMPA receptor– mediated basal synaptic transmission, were similar between the three groups (Fig. 2a,b). The pharmacologically isolated NMDAR-mediated synaptic responses were slightly reduced in AI rats. However, there was no difference between young and AU animals (Fig. 2c), suggesting that alterations in NMDAR responses cannot account for the reduced NMDAR-LTD in AU animals. Non-NMDAR-LTD in AU was blocked by a specific inhibitor of PLC, U73122 (Fig. 2d), and was attenuated by combined antagonists for PLC-linked receptors (Fig. 2e). Application of single receptor antagonists had no effect (data not shown). Similar to a previous report6, NMDAR-LTD in young rats was also blocked by U73122 (Supplementary Fig. 1). The dependence of LTD on PLC activity prompted us to examine whether the magnitude of other forms of PLC-dependent LTD, such as carbachol (CCh)-induced LTD7,10, also changes with age. We found that the average magnitude of CCh-LTD was significantly less in AI rats than in young and AU rats (Y: 78 ± 1.5% at 70 min post-CCh, n ¼ 59 slices, 10 rats; AU: 80 ± 1.9%, n ¼ 33 slices, 8 rats; AI: 92 ± 2.8%, n ¼ 34 slices, 10 rats; ANOVA: F2,125 ¼ 14.559, P o 0.01; Fig. 2f). Moreover, the magnitude of CCh-LTD correlated with the behavioral index across ages (Fig. 2g). Our results show that behavioral performance correlates with NMDAR-LTD in young animals and with non-NMDAR-LTD in aged animals. NMDAR-LTD is reduced in aged rats, but unimpaired animals seem to show an increase in non-NMDAR-LTD (Fig. 1d,e). This suggests that high-functioning aged rats maintain the ability to generate LTD, but do so by different mechanisms than those used by young

VOLUME 8

[

NUMBER 12

[

DECEMBER 2005 NATURE NEUROSCIENCE


B R I E F C O M M U N I C AT I O N S adults (Fig. 2h). It remains to be determined whether the diminished NMDAR-LTD in aged animals may have resulted from a shift in the activity requirement for the recruitment of NMDARs or from a downregulation of the machinery that couples NMDARs to LTD. In any case, our data imply that the PLC signaling pathway is activated via NMDARs in young rats (Supplementary Fig. 1), but predominantly via other receptors in AU rats (Fig. 2d), suggesting that aged animals that fail to make this switch will show impaired performance. NMDAR-LTD has been reported previously to increase with age8, which is seemingly at odds with our data showing a decrease with age (Fig. 1b,c). The difference may be due to the strain of rats used (LongEvans rats versus Fischer 344 rats8). Nonetheless, the magnitude of NMDAR-LTD correlated favorably with the behavioral index only in young adults but not in aged animals. Indeed, an opposite trend among the aged rats indicated that the worst performers tended to have slightly larger NMDAR-LTD (Fig. 1c), consistent with an association between NMDAR-LTD and impairment in older animals. Notably, our findings show that age-related cognitive impairment may be rescued by switching from an NMDAR-dependent to a nonNMDAR-dependent LTD (Fig. 2h). We further speculate that this switch has a neuroprotective function. Many studies show that excessive NMDAR activation leads to excitotoxicity11,12, and this signaling is due to the association of a macromolecular signaling complex to the receptor13. A dissociation of the macromolecular signaling complex can effectively uncouple NMDAR activation and recruitment of downstream effectors14. Similar uncoupling in AU rats would be consistent with a reduced NMDAR-LTD without a change in the NMDAR responses. We propose that this type of uncoupling has consequences


[

NUMBER 12

[

DECEMBER 2005

for preventing cognitive decline with age. It will be of future interest to test this possibility and determine the factors that mediate an effective switch in LTD mechanisms that may promote successful aging. Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS This work was supported by a US National Institutes of Health grant (NIA PO1AG09973) to M.G. and A.K. COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/

1. 2. 3. 4.

Cabeza, R. Psychol. Aging 17, 85–100 (2002). Barnes, C.A. Phil. Trans. R. Soc. Lond. B 358, 765–772 (2003). Gallagher, M., Burwell, R. & Burchinal, M. Behav. Neurosci. 107, 618–626 (1993). Nicholson, D.A., Yoshida, R., Berry, R.W., Gallagher, M. & Geinisman, Y. J. Neurosci. 24, 7648–7653 (2004). 5. Nicolle, M.M., Colombo, P.J., Gallagher, M. & McKinney, M. J. Neurosci. 19, 9604– 9610 (1999). 6. Reyes-Harde, M. & Stanton, P.K. Neurosci. Lett. 252, 155–158 (1998). 7. Kirkwood, A., Rozas, C., Kirkwood, J., Perez, F. & Bear, M.F. J. Neurosci. 19, 1599– 1609 (1999). 8. Norris, C.M., Korol, D.L. & Foster, T.C. J. Neurosci. 16, 5382–5392 (1996). 9. Huber, K.M., Roder, J.C. & Bear, M.F. J. Neurophysiol. 86, 321–325 (2001). 10. Auerbach, J.M. & Segal, M. J. Physiol. (Lond.) 492, 479–493 (1996). 11. Arundine, M. & Tymianski, M. Cell. Mol. Life Sci. 61, 657–668 (2004). 12. Lynch, D.R. & Guttmann, R.P. J. Pharmacol. Exp. Ther. 300, 717–723 (2002). 13. Husi, H., Ward, M.A., Choudhary, J.S., Blackstock, W.P. & Grant, S.G. Nat. Neurosci. 3, 661–669 (2000). 14. Sattler, R. et al. Science 284, 1845–1848 (1999).

1659

Deficits in sexual and aggressive behaviors in Cnga2 mutant mice Vidya S Mandiyan, Jennifer K Coats & Nirao M Shah Odors detected by the vomeronasal organ or the main olfactory epithelium (MOE) trigger social behaviors in many animals. It is unknown whether MOE neurons detect cues that initiate mating or aggression. We demonstrate that mice lacking functional CNGA2 (cyclic nucleotide–gated channel a2), which is required for odor-evoked MOE signaling, fail to mate or fight, suggesting a broad and essential role for the MOE in regulating these behaviors. Activation of the MOE or the vomeronasal organ (VNO) by specific odors can trigger appropriate social behaviors in rodents and many other vertebrates. The MOE is thought to be involved in detecting cues that initiate nursing and suckling1,2. The VNO is required for aggression in mice. Targeted deletion of transient receptor potential cation channel c2 (Trpc2), a gene expressed in the VNO, attenuates the responsivity of VNO neurons to mouse urine and abolishes aggression3,4. The chemosensory regulation of mating may be more complex. Olfactory bulbectomy, which eliminates transmission of MOE and VNO signals, abolishes sexual behavior2. Trpc2/ males, however, mate normally with females3,4. Taken together with additional studies2,5, these experiments suggest that the MOE and the VNO may mediate mating in a redundant fashion. Alternatively, mating and aggression may be segregated such that the MOE regulates mating, whereas the VNO initiates fighting. To distinguish between these models, we examined mice deficient in odor-evoked activity in the MOE. Although there may be redundancy in the early steps of odorevoked signaling1,6, CNGA2 (also referred to as CNGa2 or OCNC1) is the only cyclic nucleotide–binding subunit of the CNG channel

expressed in most MOE neurons7. CNGA2 is essential for odor-evoked activity in the vast majority of MOE neurons8. Because CNGA2 is not expressed in vomeronasal neurons7 (Fig. 1a), mice lacking the Cnga2 gene permit analysis of the MOE’s contribution to mating independent of the VNO. We examined male mating in Cnga2–/Y mice (Cnga2 is X-linked) and their wild-type siblings. Male mating consists of several sequential subroutines, including chemoinvestigation, mounting and intromission3. We observed a striking deficit in each of these components of sexual behavior in Cnga2–/Y mice compared to wild-type males (Fig. 1b,c). The latency to first sniff the female was 64-fold longer in Cnga2/Y mice (mean ± s.e.m.: 419.4 ± 90.0 s, n ¼ 11) compared to wild-type male mice (6.5 ± 1.4 s, n ¼ 9, P ¼ 6.0 104). All wild-type males (9 of 9) mounted in these assays, whereas none of the mutants (0 of 11) were observed to mount. Such deficits in sexual behavior persisted in a mating assay lasting several days (Supplementary Note online). These data demonstrate a profound reduction in many components of sexual behavior in Cnga2 mutants. Male mice chemoinvestigate males as well as females9. We asked whether the sniffing deficit exhibited by Cnga2–/Y mice toward females would also extend to male intruders. Wild-type resident males typically sniff and attack intruder males9. Cnga2–/Y residents showed a substantial reduction in sniffing and aggression compared to the wild-type mice (Cnga2–/Y: 5 of 11 sniffed and 1 of 11 attacked; wild-type: 9 of 9 sniffed and 6 of 9 attacked; Fig. 2). The latency to first sniff the intruder was 32-fold longer for Cnga2/Y mice (mean ± s.e.m.: 364.0 ± 165.8 s, n ¼ 11) compared to wild-type male mice (11.4 ± 3.0 s, n ¼ 9, P ¼ 1.0 103). Together with the reduction in chemoinvestigation of females, these results suggest that Cnga2 mutant males show a deficit in sniffing conspecifics during mating and aggression. Unlike Trpc2/ males3,4, Cnga2–/Y mice did not mount intruder males. The mounting of male intruders by Trpc2/ residents has led to a model suggesting that in the absence of fighting, male mice revert to a ‘default’ mode of mating indiscriminately with all conspecifics4. Our data provide a dissociation

c

Mean duration per assay (s)

b

Mean frequency per assay

M

a

O R O T+ VN E R O T+ M R O T– E W R at T– er

Figure 1 Loss of sexual behavior in Cnga2–/Y mice. (a) Reverse transcriptase–polymerase chain * Wild type (n = 9) 40 200 * reaction (RT-PCR) experiments reveal expression Mutant (n = 11) 180 35 160 of olfactory marker protein (Omp) and Cnga2 in 30 # 140 the MOE but only of Omp in the VNO. Shown is a # 25 120 typical gel run from an RT-PCR (35 cycles). The 100 20 80 ++ 15 lowest band of the ladder is 200 bp with B100 60 ## 10 bp increments up to 600 bp. RT+, reaction with 40 Cnga2 5 20 reverse transcriptase; RT–, reaction without Omp 0 0 reverse transcriptase; water, reaction without Sniff Mount Intromission Sniff Mount Intromission cDNA. Resident males were exposed to a female for 30 min. (b,c) The frequency (b) and duration (c) of mating subroutines are significantly diminished in mutants compared to wild-type mice. Animal care and handling was done in accordance with Institutional Animal Care and Use Committee guidelines. *P, 2.0 104; #P, 4.6 105; ##P, 7.0 104; ++P, 2.3 103. Frequency represents the number of times a particular behavioral subroutine was observed in the assay. Error bars represent s.e.m. VN



Neuroscience Graduate Program, Department of Anatomy, Box 2722, University of California San Francisco, San Francisco, California 94143, USA. Correspondence should be addressed to N.M.S. ([email protected]). Received 21 July; accepted 3 October; published online 30 October 2005; doi:10.1038/nn1589

1660

VOLUME 8

[

NUMBER 12

[



14 12 10

Wild type (n = 9) Mutant (n = 11)

#

8 6

##

4 2 0

Sniff

Attack

b

50 45 40 35 30 25 20 15 10 5 0

Chase

a

* #

##

Sniff

Attack

Chase

Figure 2 Loss of aggressive behavior in Cnga2–/Y mice. Resident males were exposed to a wild-type intruder male for 15 min. (a,b) The frequency (a) and duration (b) of sniffs, attacks and chases directed toward the intruder were significantly reduced in mutants compared to wild-type mice. *P, 2.0 104; #P, 5.7 103; ##P, 2.3 103. Frequency represents the number of times a particular behavioral subroutine was observed in the assay. Error bars represent s.e.m.

between sexual behavior and the absence of aggression, suggesting that mating may involve a sensory contribution from the MOE. The deficits in Cnga2 mutants could result from a general avoidance of conspecifics. Alternatively, Cnga2–/Y mice may be unable to recognize conspecific odors that initiate social behaviors. Cnga2 mutants groomed males and females in a manner similar to that of the wild-type residents, suggesting that these mutants did not avoid conspecifics entirely (Supplementary Fig. 1 online). Moreover, we found that Cnga2–/Y mice fail to sniff male or female odors even when these were presented on a neutral substrate. We provided socially naı¨ve wildtype and mutant males with female and male urine simultaneously on separate cotton pads. Cnga2–/Y mice sniffed these odors significantly less than did the wild-type mice (Fig. 3a). As expected from previous studies10, wild-type males showed a preference for female urine. By comparison, Cnga2–/Y mice failed to sniff either pad preferentially. In addition to sniffing, wild-type and mutant males interacted with the pads by carrying them in their mouths, tearing them and pushing them around. Although Cnga2–/Y mice engaged in such non-sniff interactions with greater frequency than wild-type mice, the total duration of these interactions was similar between the two genotypes (Fig. 3b). Taken together, these results suggest that the mating and aggression deficits in Cnga2–/Y mice are unlikely to result solely from an avoidance of conspecifics or their odor cues. Our data suggest an essential role for the MOE in mating and aggression. One explanation for the dual requirement for the MOE and the VNO in aggression is that these epithelia function in a parallel fashion to regulate fighting. In another model, attractant volatiles from conspecifics detected by the MOE may provoke chemoinvestigation involving physical contact, thereby permitting the VNO to access odorant cues. Diminished sniffing in Cnga2 mutants may prevent the VNO from processing cues that initiate aggression. In support of such a sequential model of activation of first the MOE and then the VNO, conspecific volatiles activate synaptic target neurons of the MOE, whereas physical contact and chemoinvestigation seems essential for the activation of synaptic target neurons of the VNO11,12. Finally, the VNO may also detect volatile odors2,13. Thus, the reduced sniffing by Cnga2–/Y mice may prevent the VNO from gaining access to aggression-modulating volatile cues. It will be interesting to determine the mechanisms underlying the diminished sniffing in these mutants. In any event, an intact VNO is not required for mating3,4, suggesting that the MOE also processes cues that regulate sexual behavior. In future studies, it will be important to determine whether prior social experience bypasses the requirement for a functional MOE in


[


*

Mean duration per assay (s)

16

NUMBER 12

[

DECEMBER 2005

*

20 18

#

*

16 14

#

12 10 8 6

*

4 2 0 Sniff M

Sniff F

NSI-M

NSI-F

*

b

*

70 Mean duration per assay (s)



a

Wild type (n = 5) Mutant (n = 5)

60 50 40 30 20

**

10 0

Sniff M

Sniff F

NSI-M

NSI-F

Figure 3 Loss of preference for female urine odors in Cnga2–/Y mice. Resident mutant and wild-type males were exposed to two cotton pads wetted with male (M) or female (F) urine for 5 min. (a,b) Wild-type mice sniffed female urine more frequently (a) and for longer duration (b) than did the mutants. Cnga2–/Y mice displayed more non-sniff interactions (NSI) with urine-wetted pads compared to the wild-type mice, although the total duration of NSI was similar between the two genotypes. *P, 7.9 103; **P, 1.6 102; #P, 3.2 102. NSI-F and NSI-M, NSI with female urineand male urine-wetted swabs, respectively. Frequency represents the number of times a particular behavioral subroutine was observed in the assay. Error bars represent s.e.m.

chemoinvestigation, mating and aggression. Finally, some MOE neurons do not express CNGA2 and use a distinct odor-evoked signaling pathway2. Our data suggest that this subpopulation cannot initiate mating or fighting in the absence of MOE neurons expressing functional CNGA2. We cannot exclude the possibility that central deficits, including aberrant connectivity, produce the behavioral phenotypes we observe in Cnga2–/Y mice. CNGA2 is also expressed in several brain regions14. Nevertheless, Cnga2–/Y mice resemble wild-type mice in many behavioral tests, including grooming (Supplementary Fig. 1) and operant conditioning2. Finally, Cnga2–/Y mice have circulating levels of testosterone that are similar to those in wild-type mice (Supplementary Note), suggesting that the behavioral phenotypes are unlikely to arise from testosterone deficits in adults. The behavioral deficits observed in Cnga2–/Y mice resemble the phenotypes resulting from adult bulbectomy15. These surgically lesioned males fail to chemoinvestigate conspecifics and do not mate or fight. Our results are therefore consistent with a role for CNGA2-expressing MOE neurons in regulating chemoinvestigation, mating and aggression. Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS The authors thank D. Anderson, R. Axel, H. Baier, U. Heberlein, L. Jan, D. Julius and members of the Shah lab for comments on the manuscript. We thank J. Ngai for providing us with Cnga2 mutant females and J. Wong for administrative support. This work was supported by a grant from the National Institutes of Health (R01 NS049488), a Career Award in the

1661

B R I E F C O M M U N I C AT I O N S Biomedical Sciences from the Burroughs Wellcome Fund, the Sandler Family Supporting Foundation and funds from Research Evaluation and Allocation Committee (REAC) at University of California, San Francisco (N.M.S.). N.M.S. is a McKnight Scholar and a Sloan Fellow.


COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/ 1. Belluscio, L., Gold, G.H., Nemes, A. & Axel, R. Neuron 20, 69–81 (1998). 2. Restrepo, D., Arellano, J., Oliva, A.M., Schaefer, M.L. & Lin, W. Horm. Behav. 46, 247–256 (2004).

1662

3. Leypold, B. et al. Proc. Natl. Acad. Sci. USA 99, 6376–6381 (2002). 4. Stowers, L., Holy, T.E., Meister, M., Dulac, C. & Koentges, G. Science 295, 1493–1500 (2002). 5. Del Punta, K. et al. Nature 419, 70–74 (2002). 6. Wong, S.T. et al. Neuron 27, 487–497 (2000). 7. Berghard, A., Buck, L.B. & Liman, E.R. Proc. Natl. Acad. Sci. USA 93, 2365–2369 (1996). 8. Brunet, L.J., Gold, G.H. & Ngai, J. Neuron 17, 681–693 (1996). 9. Miczek, K.A., Maxson, S.C., Fish, E.W. & Faccidomo, S. Behav. Brain Res. 125, 167–181 (2001). 10. Pankevich, D.E., Baum, M.J. & Cherry, J.A. J. Neurosci. 24, 9451–9457 (2004). 11. Lin, Y., Zhang, S.Z., Block, E. & Katz, L.C. Nature 434, 470–477 (2005). 12. Luo, M., Fee, M.S. & Katz, L.C. Science 299, 1196–1201 (2003). 13. Sam, M. et al. Nature 412, 142 (2001). 14. Kingston, P.A., Zufall, F. & Barnstable, C.J. Synapse 32, 1–12 (1999). 15. Devor, M. & Murphy, M.R. Behav. Biol. 9, 31–42 (1973).

VOLUME 8

[

NUMBER 12

[


a

It has been suggested that thinking about and comparing numeric quantities shares functional properties and brain circuits with spatial thinking1,2. In particular, semantic representations of the size and distance between numbers are organized along an internal horizontal ‘mental number line’, with small numbers located to the left of larger ones1. There are two lines of evidence for this. First, normal human subjects show spatial compatibility effects when responding to numbers: responses to small numbers are faster with the left hand and those

20 10 0 –10 –20

b Figure 1 Bisection of horizontal physical lines and mental number lines. (a,b) Bisection of horizontal physical lines (length: 20, 100 and 200 mm): group data (a) and individual data (b). (c,d) Bisection of the mental number line (interval size: 3, 5, 7 and 9): group data (c) and individual data (d). Positive values indicate rightward deviation and negative values leftward deviation from the objective midpoint (0 value on y-axis). ‘N+H–’ represents subjects with neglect and no hemianopia; N+H+, subjects with neglect and complete hemianopia; and N–H–, subjects without neglect or hemianopia. Confidence interval limits (P ¼ 0.05) are shown by gray-shaded boxes. Number line bisection was administered in two sessions separated by 1 week. No between-session difference was present (that is, we found good test-retest stability; group number interval length session ANOVA: session effect and interactions, F6,39 o 1, P ¼ 0.8, n.s.). Bisections of physical lines were performed during one session between the two number line sessions. Experimental line bisections (200 mm) were equivalent to bisections performed during general neuropsychological screening administered on admittance to the hospital unit (that is, we found good test-retest stability: group session ANOVA; session effect and interactions, F2,8 ¼ 1, P ¼ 0.4, n.s.). Number and line bisection test-retest stability was still present when only the two subgroups of subjects showing double dissociation (N+H–1 to N+H–3, and N+H+1 to N+H+3) were considered in the ANOVAs (all F values ¼ 1, 0.3 o P o 0.9, n.s.).

c

30

Deviation from midpoint (units)

To compare numeric quantities, humans make use of a ‘mental number line’ with smaller quantities located to the left of larger ones; it is unclear, however, whether orienting along the number line is like orienting along a physical line. We found that in brain-damaged subjects with defective leftward orienting, rightward deviation in the bisection of numeric and physical intervals is doubly dissociated. Deviation in numeric interval bisection was associated with prefrontal damage and spatial working memory deficit.

d

30

1

0.5

Group N+H–

0

N+H+ N–H–

–0.5 –1 3 Case N+H–1

20

N+H–2

2 Deviation from midpoint (units)

Fabrizio Doricchi1,2, Paola Guariglia1,2, Marina Gasparini3 & Francesco Tomaiuolo4

to large numbers faster with the right hand3. Second, when required to indicate which number is halfway between two other numbers, right brain–damaged subjects neglecting the left side of space disregard the left side of large number intervals, shifting the subjective midpoint rightward4 (for example, they assign the interval ‘1–9’ a subjective midpoint of 7 instead of 5). The same subjects were reported to bisect short number intervals leftward4 (assigning the interval ‘1–3’ a subjective midpoint of 1 instead of 2), mimicking the paradoxical contralesional deviation (that is, ‘crossover’) shown by some neglect subjects in the bisection of short physical lines (20–25 mm)5. In neglect subjects, the rightward deviation in the bisection of long physical lines is significantly worsened by concomitant primary visual field defects (that is, contralateral hemianopia) and leftward ‘crossover’ is prevalent when short lines cross a blind sector of the neglected space5. Here, we verified whether the same happens when the mental number

Deviation from midpoint (%)

Dissociation between physical and mental number line bisection in right hemisphere brain damage

Deviation from midpoint (%)



10

0

–10

N+H–3 N+H–4 N+H–5 N+H–6

1

N+H–7 N+H–8 N+H+1 N+H+2

0

N+H+3 N–H–1 N–H–2 N–H–3

–1

N–H–4

–20

N–H–5

–30

20

100

200

Line length (mm)

–2

3

5

7

9

Numerical interval length (units)

1Fondazione Santa Lucia IRCCS - LENA (Laboratoire Europeen Neuroscience de l’Action), Rome, Italy. 2Dipartimento di Psicologia, Universita ` degli Studi di Roma ‘La Sapienza’, Rome, Italy. 3Dipartimento di Scienze Neurologiche–Va Cattedra, Universita` degli Studi di Roma ‘La Sapienza’, Rome, Italy. 4Auxilium Vitae Volterra, Volterra, Pisa, Italy. Correspondence should be addressed to F.D. ([email protected]).

Received 2 August; accepted 14 September; published online 30 October 2005; doi:10.1038/nn1563


[

NUMBER 12

[

DECEMBER 2005

1663


B R I E F C O M M U N I C AT I O N S Figure 2 Anatomical correlates of line number bisection. (a) Anatomical references. The coronal plane (y ¼ 22) is oriented through the vertical ramus of the sylvian fissure separating the pars opercularis of the inferior frontal gyrus (including Ba44) from the pars triangularis (including Ba45) and the pars pre-triangularis (including Ba47). In b–e, the same coronal reference plane is indicated by the dashed red line. (b) Lesion overlap in neglect subjects with ipsilesional deviation in mental line number bisection: cases N+H–1, N+H–2 and N+H–3. In all cases, the damage involves the inferior frontal gyrus (pars triangularis and pre-triangularis) and/or the white matter below the middle and inferior frontal gyrus. In N+H–3, the damage also involves part of the mid-frontal gyrus (Ba46 and Ba9). (c) Lesion overlap in neglect subjects with hemianopia, marked ipsilesional deviation in physical line bisection and no deviation in number line bisection: cases N+H+1, N+H+2 and N+H+3. (d) Subtraction of lesions from b minus lesions from c. (e) Lesion overlap in neglect subjects with no hemianopia and no ipsilesional deviation in number line bisection: cases N+H–4 through N+H–8 (see mapping methods in ref. 11). All N–H– subjects suffered small lesions not involving prefrontal areas. Informed written consent was obtained from all subjects. This study was approved by the Institutional Review Board of the Fondazione Santa Lucia.

line is bisected. We compared numeric interval bisection and physical line bisection in three groups of consecutive right brain–damaged subjects: a group with chronic neglect and complete hemianopia (that is, no macular sparing; three cases, referred to as N+H+1 through N+H+3), a group with chronic neglect and no hemianopia (eight cases, N+H–1 through N+H–8) and a group without neglect or hemianopia (five cases, N–H–1 through N–H–5). Critically, this systematic comparison should first complement previous observations in which the relationship between bisection of mental number and physical lines was not investigated4 or reported in single cases6. The three groups of subjects did not differ in age (analysis of variance (ANOVA): F2,13 ¼ 1.4, P ¼ 0.2; mean age ¼ 62.2 years) or time elapsed from stroke onset (ANOVA: F2,13 ¼ 1.2, P ¼ 0.3; mean ¼ 70.5 d). On the letter cancellation task, neglect was more severe in N+H+ (38/104 cancelled items) than in N+H– subjects (62.5/104) and was more severe in N+H– than in N–H– subjects (103.2/104; ANOVA, F2,13 ¼ 8, P o 0.01; all comparisons of the means, P o 0.01). We used 20-mm, 100-mm and 200-mm physical lines (five trials per length) and four sizes of numeric intervals: 3 (for example, ‘4–6’), 5 (‘3–7’), 7 (‘2–8’) and 9 (‘1–9’). The two numbers defining each interval (units, teens and the numbers ranging from 21 to 29) were presented orally both in ascending (48 trials) and descending order (48 trials). The three groups differed in the bisection of physical lines (Fig. 1a,b) depending on line length (group length interaction ANOVA on arcsin transform of percentage deviation; F4,26 ¼ 10, P o 0.001). N+H+ subjects showed the strongest rightward deviation with 200-mm and 100-mm lines and the strongest leftward deviation (crossover) with 20-mm lines (all comparisons with the two other groups, P o 0.01). With 200-mm lines, N+H– subjects deviated rightward as compared with N–H– subjects (P ¼ 0.02), whereas no difference was present for 20-mm lines. By contrast, we did not find any significant rightward deviation in the bisection of longer intervals (of length 5, 7 or 9) or leftward deviation in the bisection of short intervals (of length 3) in the groups of neglect subjects (Fig. 1c; group interval size order interaction ANOVA, F6,39 o 1). However, inspection of individual data (Fig. 1d) showed that in three out of the eight N+H– subjects (cases 1, 2 and 3) there was clear rightward deviation with longer intervals (that is, falling outside the 95% confidence interval calculated over all subjects in the study). In one case (N+H–1), deviation was present for intervals of length 7 and 9; in the other two cases (N+H–2 and N+H–3), it was present for

1664

a

Y = 22

Z = 32

Z = –12 0%

100%

b

Z = –12

–2

12

22

32

c

d

e

intervals of length 5, 7 and 9. In all of these subjects, the longer the numeric interval, the stronger the rightward shift of the subjective midpoint. With physical lines, the performance of the three same subjects was within the confidence range (Fig. 1b). In addition, in all three N+H+ subjects, bisection of numeric intervals was within or below confidence ranges (Fig. 1d). In the same subjects, physical line bisection was below the inferior confidence limit for 20-mm lines (that is, leftward ‘crossover’) and above the superior confidence limit for 100-mm and 200-mm lines (rightward deviation, Fig. 1b). In both the whole sample of subjects and in neglect subjects, deviation in the bisection of short numeric intervals (of length 3) was unrelated to deviation with short physical lines (20 mm; Pearson’s r = 0.2 (not

VOLUME 8

[

NUMBER 12

[


B R I E F C O M M U N I C AT I O N S Table 1 Individual spatial (Corsi span) and verbal (Digit span) working memory scores


Case

Corsi span

Digit span

N+H–1 N+H–2

– 2/9 (0)

– 6/9 (4)

N+H–3 N+H–4

3/9 (0) 4/9 (2)

5/9 (3) 5/9 (3)

N+H–5 N+H–6

4/9 (2) 4/9 (2)

4/9 (2) 6/9 (4)

N+H–7 N+H–8

4/9 (2) 4/9 (2)

6/9 (4) 6/9 (4)

N+H+1 N+H+2

– 4/9 (2)

– 5/9 (3)

N+H+3 N–H–1

4/9 (2) 4/9 (2)

6/9 (4) 8/9 (4)

N–H–2 N–H–3

6/9 (4) 4/9 (2)

6/9 (4) 5/9 (4)

N–H–4

6/9 (4)

6/9 (4)

N–H–5

4/9 (2)

5/9 (3)

Scores outside of parentheses are raw scores. Scores within parentheses are normalized for age and educational level. Range of normalized scores is 1–4.

significant; n.s.)), and deviation with long intervals (of length 9) was unrelated to deviation with long physical lines (200 mm; r ¼ 0.1, n.s.). Severity of neglect in a multiple-item letter cancellation task showed no correlation with deviation in the bisection of any number interval (0.08 o r o 0.3 n.s.). This data implies that mechanisms allowing estimation of numeric interval length along the mental number line are not identical to those allowing length estimation along physical lines. Comparative judgments of numeric quantities activate the horizontal segment of the intraparietal sulcus bilaterally, the precentral gyrus in the left hemisphere, and, depending on task, prefrontal areas1,2,7. In contrast, clinical8 and imaging9 data demonstrate that physical line bisection depends on the striate, the extrastriate visual cortex and the inferior and superior parietal lobe, with marked lateralization in the right hemisphere. The dissociation between severity of neglect and deviation in numeric interval bisection also suggests that moving between proximal mental numeric quantities is not a selective function of areas most frequently damaged in neglect subjects, such as the inferior parietal lobe and the parietal-frontal connections in the underlying white matter10–12. We therefore mapped and compared lesions in neglect subjects with and without rightward deviation in the bisection of number intervals. We found that in all subjects showing deviation, the brain damage involved prefrontal areas (cortically and subcortically in N+H–1 and N+H–3; subcortically in N+H–2;


[

NUMBER 12

[

DECEMBER 2005

Fig. 2b,d). None of the neglect subjects without deviation suffered prefrontal damage (Fig. 2c,e). On the basis of this anatomical data, we normalized available spatial working memory scores (Corsi span, administered in the ipsilesional space; Table 1). N+H–2 and N+H–3 were the only subjects with null scores (no scores were available for cases N+H–1 and N+H+1), and they suffered no analogous deficit in the short-term retention of verbally presented digit series (Table 1). These results suggest that right brain damage to prefrontal spatial working memory structures (which may itself contribute to severity of visual neglect13) is the main source of pathological rightward bias in building up and keeping the mental number line active, as proposed for other types of mental representations14. Ipsilesional bias can be specifically due to disruption of neuronal populations that, in the human ventrolateral prefrontal cortex of each hemisphere (Ba 45/47), show lateral selectivity in the short-term retention of contralateral spatial positions15. To conclude, our findings confirm that the mental number line is spatially organized and that brain damage can disrupt its organization; however, they also demonstrate that thinking of a numeric interval as a horizontal line segment can be misleading, as visuospatial operations and brain networks required to navigate along number lines are appreciably different from those recruited by navigation along physical ones. ACKNOWLEDGMENTS F.D. wishes to thank D. Bueti and A. Doricchi for discussing the paper and H. Bowles for text revision. This study was supported by grants from the Fondazione Santa Lucia and from the Ministero della Universita` e della Ricerca Scientifica. COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/

1. Dehaene, S. et al. Curr. Opin. Neurobiol. 14, 218–224 (2004). 2. Walsh, V. Trends Cogn. Sci. 7, 483–488 (2003). 3. Dehaene, S., Bossini, S. & Giraux, P. J. Exp. Psychol. Gen. 122, 371–396 (1993). 4. Zorzi, M., Priftis, K. & Umilta`, C. Nature 417, 138–139 (2002). 5. Doricchi, F. et al. Brain 128, 1386–1406 (2005). 6. Rossetti, Y. et al. Psychol. Sci. 15, 426–430 (2004). 7. Dehaene, S. et al. Cogn. Neuropsychol. 20, 487–506 (2003). 8. Doricchi, F. & Angelelli, P. Neurology 52, 1845–1852 (1999). 9. Fink, G.R. et al. Neurology 54, 1324–1331 (2000). 10. Doricchi, F. & Tomaiuolo, F. Neuroreport 14, 2239–2243 (2003). 11. Mort, D.J. et al. Brain 126, 1986–1997 (2003). 12. Leibovitch, F.S. et al. Neurology 50, 901–908 (1998). 13. Malhotra, P. et al. Brain 128, 424–435 (2005). 14. Della Sala, S. et al. Neuropsychologia 42, 1358–1364 (2004). 15. Rizzuto, D.S. et al. Nat. Neurosci. 8, 415–417 (2005).

1665


CO M P U TATI O N AN D SYSTE M S

ARTICLES

Conditional dendritic spike propagation following distal synaptic activation of hippocampal CA1 pyramidal neurons Tim Jarsky1,4, Alex Roxin2–4, William L Kath1,2 & Nelson Spruston1 The perforant-path projection to the hippocampus forms synapses in the apical tuft of CA1 pyramidal neurons. We used computer modeling to examine the function of these distal synaptic inputs, which led to three predictions that we confirmed in experiments using rat hippocampal slices. First, activation of CA1 neurons by the perforant path is limited, a result of the long distance between these inputs and the soma. Second, activation of CA1 neurons by the perforant path depends on the generation of dendritic spikes. Third, the forward propagation of these spikes is unreliable, but can be facilitated by modest activation of Schaffer-collateral synapses in the upper apical dendrites. This ‘gating’ of dendritic spike propagation may be an important activation mode of CA1 pyramidal neurons, and its modulation by neurotransmitters or long-term, activity-dependent plasticity may be an important feature of dendritic integration during mnemonic processing in the hippocampus.

Excitatory synapses on distal dendrites are common in the nervous system. For example, cortical pyramidal neurons receive corticocortical inputs in layer 1, often hundreds of microns from the soma. Mitral cells of the olfactory bulb receive input from olfactory receptor neurons on a tuft of apical dendrites, similarly far from the soma. Purkinje cells of the cerebellum receive inputs from mossy cells, many of which terminate on distal dendrites. Synapses like these are intriguing, because their long distance from the spike initiation zone, thought to reside in the axon1, suggests that special mechanisms may be required for these synapses to trigger action potentials. One possible mechanism to preserve the efficacy of distal synapses is through the generation of local dendritic spikes. Indeed, all dendrites studied to date express numerous voltage-gated channels2, and substantial evidence now supports the notion that spikes can be initiated in dendrites3. Unlike axonal action potentials, however, dendritic spikes do not propagate reliably over long distances in dendrites. A remaining challenge, therefore, is to determine what types of synaptic inputs can trigger dendritic spikes and how the spikes propagate along dendrites of various morphologies expressing unique combinations of channels. For such questions, a computational approach can place the available data in an integrated, quantitative framework and provide testable predictions concerning the function of various types of synapses and dendrites. In the hippocampus, CA1 pyramidal neurons receive two distinct excitatory synaptic inputs4: the perforant path (or temporo-ammonic path) provides direct input from layer 3 of entorhinal cortex to the apical dendritic tuft, and the Schaffer collaterals provide input from CA3 pyramidal neurons to basal and apical dendrites in CA1. Although

all of the Schaffer-collateral synapses are closer to the soma than are the perforant-path synapses, some of the Schaffer-collateral synapses are nevertheless hundreds of microns from the soma. The perforant-path projection to CA1 terminates in excitatory, glutamatergic synapses5,6, but it is controversial whether these synapses can depolarize the somatic membrane to action-potential threshold7–9. Distal perforantpath and Schaffer-collateral synapses elicit dendritic spikes, both in vitro and in vivo10,11, but some of these dendritic spikes do not propagate reliably to the soma, raising questions about the function of dendritic spikes and the conditions necessary for them to reach the soma and trigger an action potential in the axon10–13. Here we explored these questions by examining the impact of distal perforant-path and Schaffer-collateral synaptic inputs to CA1 pyramidal neurons, using a combined computational and experimental approach. We found that strong perforant-path activation resulted in dendritic spikes that could fail to propagate to the soma. Modest activation of Schaffer-collateral synapses in the upper apical dendrites, however, facilitated the forward propagation of tuft dendritic spikes, thus allowing action-potential output in the axon following distal synaptic activation. RESULTS Modeling distal synaptic activation in excitable CA1 dendrites To assess the ability of perforant-path synapses to activate CA1 pyramidal neurons, we used compartmental models of two reconstructed neurons. Each neuron included four active conductances: a voltagegated Na+ conductance (GNa), a delayed rectifier K+ conductance (GKdr)

1Institute for Neuroscience, Department of Neurobiology and Physiology, Northwestern University, 2205 Tech Drive, Evanston, Illinois 60208, USA. 2Department of Engineering Science and Applied Mathematics, Northwestern University, 2145 Sheridan Road, Evanston, Illinois 60208, USA. 3Unite´ Mixte de Recherche 8119, Centre Nationale de la Recherche Scientifique, Neurophysics and Physiology, Universite´ Rene´ Descartes, 45 Rue des Saints Pe`res, 75270 Paris Cedex 06, France. 4These authors contributed equally to this work. Correspondence should be addressed to N.S. ([email protected]).

Received 31 August; accepted 20 October; published online 20 November 2005; doi:10.1038/nn1599


[

NUMBER 12

[

DECEMBER 2005

1667

ARTICLES Strong

Weak

0.3

0.3 Max. conductance (S cm–2)

Max. conductance (S cm–2)

GNa

0.2

GKdr GKAp GKAd

0.1

0

200 400 600 Distance from soma (µm)

0.2

0.1

800

0

200 400 600 Distance from soma (µm)

800

b PP PP

40 30 20 10 0 –10 –20 –30 –40 –50 –60 –70 mV

100 µm

40 mV


a

5 ms

and two A-type K+ conductances (GKA). Consistent with experimental reports, GNa and GKdr were modeled with moderate conductance along the somato-dendritic axis, but a higher GNa in the axon made this the preferential site of action-potential initiation14. GKA was modeled with the reported sixfold increase in conductance along the somato-dendritic axis and lower half-inactivation voltage in proximal (GKAp) versus distal (GKAd) dendrites15,16. We used two versions of the model: a weak dendritic excitability model with uniform GNa in the soma and dendrites, and a strong dendritic excitability model with a slight gradient of increasing GNa with distance from the soma (Fig. 1a). These two models are simple compared to the full repertoire of voltagegated channels known to be expressed in pyramidal cells2,17; nevertheless, these models reproduce two populations of CA1 neurons with distinct profiles of action-potential backpropagation14. We used these models to make experimentally testable predictions about the ability of perforant-path inputs to activate CA1 neurons. The two models exhibited markedly different behavior (Fig. 1). In the strong dendritic excitability model, simulation of a strong perforant-path stimulus (10% of available synapses in the distal apical tuft) triggered the initiation of a dendritic spike, which propagated to the soma and triggered an action potential (Fig. 1b, left and Supplementary Video 1 online). Similar behavior was observed in the model of a second reconstructed neuron (Supplementary Fig. 1 online and Supplementary Video 2 online). In the weak dendritic excitability model, the same strong perforant-path input triggered a dendritic spike, but the spike failed to propagate to the soma, thus resulting in a subthreshold somatic depolarization (Fig. 1b, right and Supplementary Video 3 online). With a high-frequency burst of synaptic inputs (5 at 100 Hz), dendritic spikes sometimes propagated to the soma, but successful propagation required activating a large number of synapses (for example, 30% success at 15% of perforant-path synapses; data not shown). These results suggest that the only way perforant-path inputs can trigger an action potential in the axon is if a dendritic spike is

1668

Figure 1 Strong and weak dendritic excitability models of CA1 pyramidal neurons respond differently to perforant-path activation. (a) Channel distributions as a function of distance from the soma in the two models. GNa, voltagegated Na+ conductance; GKdr, delayed rectifier K+ conductance; GKAp, proximal A-type K+ conductance; GKAd, distal A-type K+ conductance. (b) Perforant-path (PP) activation (10% of available synapses) triggered dendritic spikes in the apical tuft. These spikes propagate to the soma and trigger an action potential in neurons with strong (left), but not weak (right), dendritic excitability. Scales apply to both panels. Color maps are of maximal (peak) voltage in each compartment of the model. Traces are voltage versus time plots at the three dendritic locations indicated by the electrodes. Animations of these simulations in this cell and another reconstructed cell are provided in Supplementary Videos 1–3 and 6.

initiated in the tuft and propagates down the apical dendrite. In the weak excitability model, forward propagation was unreliable and did not ensure an action-potential output. This finding is consistent with our previous experimental observation that dendritic spikes can occur in the absence of somatic action potentials10,18.

Interaction between perforant-path and Schaffer-collateral inputs In the models, when dendritic spikes failed to propagate from the distal apical tuft to the soma, they usually failed in the upper apical dendrite (Fig. 1b, right; Supplementary Figs. 2 and 3 online; Supplementary Videos 3–8 online). We therefore investigated how perforant-path synaptic responses were influenced by the activation of synapses in the upper apical dendrites, just below the apical tuft. These synapses correspond to distal Schaffer-collateral inputs. When perforant-path and Schaffer-collateral inputs are activated together, action-potential firing is probabilistic, depending on the exact number and location of the randomly selected synapses activated in each region. Actionpotential probability was determined from more than 170,000 simulations and plotted as a function of the percentage of perforant-path and Schaffer-collateral synapses activated (Fig. 2). In the strong dendritic excitability model, two modes of activation were apparent. In the first mode of the strong excitability model (Fig. 2b, left, region 1), activation of the Schaffer collaterals (43%) was on its own sufficient to trigger action potentials on at least some trials. With little or no activation of perforant path, action-potential probability increased from 0 to 1 over a narrow range of Schaffercollateral activation (about 3–5%). As the percentage of perforant-path inputs was increased from 0% to 2%, the percentage of Schaffercollateral inputs necessary to trigger an action potential 50% of the time (green in Fig. 2b, left) decreased from about 4% to 3%. In this mode, the Schaffer-collateral input was stronger than the perforantpath input and dendritic spikes almost always began in the upper apical dendrites, below the tuft and near the Schaffer-collateral inputs (green in Fig. 2c, left, region 1). In the second mode of the strong excitability model (Fig. 2b, left, region 2), perforant path was the dominant input, and action potentials could be elicited on at least some trials with 42% of the synapses activated. With little or no activation of the Schaffer collaterals, actionpotential probability increased from 0 to 1 over a narrow range of

VOLUME 8

[

NUMBER 12

[


ARTICLES 8

7

7

1

1

0.9 0.8

Tuft spike probability

Percentage Schaffer collateral

Percentage Schaffer collateral

8

Average site of spike initiation


b

Action potential probability

potentials only occurred on some trials (about 10% at 60–70% perforant-path activa6 6 1 0.7 1.0 tion), and even greater activation of perfor5 5 0.6 PP 0.5 4 4 ant-path inputs decreased the probability of 0.4 3 3 2 action-potential firing (owing to the activa0.3 2 2 0.2 SC tion of K+ channels and the inactivation of 2 1 1 0.1 0.5 0 Na+ channels by large excitatory post synaptic 0 0 0 2 4 6 8 10 12 0 2 4 6 8 10 12 Percentage perforant path Percentage perforant path potentials (EPSPs)). Because such large acti1 8 vation of perforant path is required, interac8 c 0.9 7 7 tion of perforant-path and Schaffer-collateral 0 0.8 1 6 6 0.7 inputs is a more plausible explanation for how 5 5 0.6 1 perforant path influences action-potential fir0.5 4 4 0.4 3 3 2 ing in CA1 pyramidal neurons with weakly 0.3 2 2 0.2 excitable dendrites. 2 1 1 0.1 In the weak dendritic excitability model, 0 0 0 0 2 4 6 8 10 12 0 2 4 6 8 10 12 perforant-path and Schaffer-collateral inputs Percentage perforant path Percentage perforant path interacted in two important ways. First, low 1 1 d levels of perforant-path activation slightly 0.8 0.8 reduced the percentage of Schaffer-collateral 0.6 0.6 inputs necessary to trigger an action potential 0.4 0.4 (Fig. 2b, right, region 1). Activation of 0.2 0.2 between 0% and 6% of perforant-path synapses reduced the percentage of Schaffer0 0 0 2 4 6 8 10 12 0 2 4 6 8 10 12 collateral synapses necessary to trigger an Percentage perforant path Percentage perforant path action potential 50% of the time (green in Figure 2 Modes of activation of CA1 pyramidal neurons by perforant-path and Schaffer-collateral Fig. 2b, right) from about 6.5% to 5.2% activation. (a) Schematic of model cell with perforant-path (PP) and Schaffer-collateral (SC) activation (slope ¼ 0.22). In this mode, most of the sites indicated. Electrodes at indicated locations were used to determine whether spikes began in the spikes began in the apical dendrite (below the soma (value 0, blue), mid apical dendrite (value 0.5, green) or apical tuft (value 1.0, red). (b) Probability tuft) and propagated to the soma (green in of action-potential initiation (in the axon) in response to activation of different percentages of available Fig. 2c, right, region 1), but in some cases, the PP (apical tuft) and SC synapses (upper apical dendrites). (c) Locations of spike initiation as described in a. Intermediate colors are determined by the fractions of trials yielding initiation at one or another Schaffer-collateral and perforant-path inputs site. In both b and c, regions 1 and 2 correspond to modes of activation where SC and PP inputs summed in the soma to trigger an action dominate, respectively (see text for details). (d) Probability of a dendritic spike in the apical tuft potential first in the axon (blue in Fig. 2c, as a function of the percentage of PP synapses activated (no SC activation). Columns in b, c right, region 1). Within this activation mode, and d correspond to the strong and weak dendritic excitability models, respectively. Probabilities were there was little difference between perforantdetermined from 200 trials of randomly placed synapses for each %PP, %SC pair. In b and c, horizontal path inputs that triggered dendritic spikes and guide lines correspond to the minimal percentage of SC inputs necessary to trigger an action potential; those that did not. Even when perforant-path in b, c and d, vertical guide lines correspond to the percentage of PP inputs necessary to trigger a spike in the apical tuft on 50% of trials. activation triggered a dendritic spike (for example, 20% of the time at 5% activation: Fig. 2d, right), the spikes did not spread into perforant-path activation (about 2–6%). In this mode, action the upper apical dendrites, below the tuft. Instead, the main effect in potentials were driven by dendritic spikes originating in the apical this mode of activation was that perforant-path inputs reduced the tuft (red in Fig. 2c, left, region 2). With perforant-path stimulation amount of Schaffer-collateral input necessary to trigger a dendritic alone, the probability of dendritic spike initiation in the tuft increased spike in the apical dendrites below the tuft. By contrast, stronger activation of perforant path (above about 6%) from near 0 to 1 over the same range (about 2–6%, Fig. 2d, left). Over this range of perforant-path activation, modest activation of the activated dendritic spikes in the tuft more frequently. These spikes Schaffer collaterals (0–3%) reduced the percentage of perforant-path tended to be larger and thus spread throughout the apical tuft. In this inputs necessary to trigger a propagating dendritic spike, but the spikes mode (Fig. 2b, right, region 2), activation of between 6% and 10% of almost always began in the apical tuft (innervated by perforant path) perforant-path synapses reduced the percentage of Schaffer-collateral and propagated to the soma (Fig. 2c, left). Stronger activation of either synapses necessary to trigger an action potential 50% of the time from or both pathways led to reliable dendritic spike initiation, with the about 5.2% to 2% (slope ¼ 0.80). Furthermore, in this mode, action spikes tending to begin near the strongest synaptic input. In summary, potentials almost always began as dendritic spikes in the apical tuft (red in the strong excitability model, each input was able to trigger dendritic in Fig. 2c, right, region 2). This activation mode thus revealed an spikes on its own, and each input could cooperate with the other to especially interesting interaction between perforant path and Schaffer collaterals. Dendritic spikes were triggered by strong perforant-path trigger a propagating dendritic spike. The behavior of the weak dendritic excitability model was drama- input, but their propagation to the soma was gated by Schaffertically different (Fig. 2b, right). Importantly, a single coincident collateral inputs. In the absence of Schaffer-collateral inputs, the apical activation of perforant-path synapses was almost never able to produce tuft dendritic spikes failed to reach the soma and did not produce an an action potential on its own. Only when perforant-path activation action potential. However, even modest levels of Schaffer-collateral was increased to very high levels (450% of all synapses) did an action input were able to facilitate the forward propagation of these dendritic potential occur in the soma (data not shown). Even then, action spikes, such that they could trigger a full action potential in the axon

a


[

NUMBER 12

[

DECEMBER 2005

1669

ARTICLES Figure 3 Gating of perforant-path–evoked dendritic spikes by Schaffercollateral evoked EPSPs. (a) Color map of peak depolarization and voltage versus time plots at three dendritic locations for activation of 10% of PP synapses (one trial) in the weak dendritic excitability model. (b) Response of the same model to activation of 3% of SC synapses in the upper apical dendrites. (c) Response of the same model to coincident activation of 10% PP and 3% SC synapses. Animations of these simulations in this and the other reconstructed cell are provided in Supplementary Videos 3–8. Scales in a apply to b and c as well.

a


40 30 20 10 0 –10 –20 –30 –40 –50 –60 –70 mV

40 mV

PP

5 ms

b SC

c PP

SC

and soma (Fig. 3) This gating of perforant path–evoked dendritic spikes by Schaffer-collateral synaptic inputs was observed in both of the CA1 cell models (Supplementary Figs. 2 and 3; Supplementary Videos 3–8 online). One notable feature of the behavior of both the strong and the weak dendritic excitability models is that the number of perforant-path inputs necessary to trigger a dendritic spike in the tuft was actually lower than the number of Schaffer-collateral inputs necessary to trigger a dendritic spike in the upper apical dendrites. In the strong excitability model, activation of a minimum of B2% of perforant-path inputs resulted in dendritic spikes on at least some trials, compared to at least 3% for the Schaffer-collateral input (Fig. 2c,d, left). About 4% of each input was necessary to produce an action potential on half of the trials (Fig. 2b, left), but because there were fewer synapses in the perforantpath pool (Methods), fewer perforant-path synapses were needed to produce half activation. In the weak excitability model, 43% of perforant-path synapses produced tuft spikes (Fig. 2d, right), compared to 45% of Schaffer-collateral synapses required to produce

1670

upper apical dendritic spikes (Fig. 2b, right). The lower number of perforant-path synapses required to produce dendritic spikes, in both the strong and weak excitability models, is attributable to the high input impedance of the small-diameter tuft branches. A possible limitation of our models, which were originally designed to reproduce action potential backpropagation14, is that they used a minimal repertoire of voltage-gated conductances. Although these models produced dendritic spikes without changes from their original forms, they are clearly very simple models. Nevertheless, the models make some interesting predictions concerning the response to distal synaptic activation. To determine if these predictions were robust, we used a more complex model, which reproduces many aspects of CA1 pyramidal neuron excitability17. This model contains a variety of conductances missing from our models, including H current, Ca2+ currents and Ca2+-activated K+ currents. Neither the presence of these additional conductances nor the longer dendrites in this model altered the basic behavior we have described. Strong activation of distal synaptic inputs triggered dendritic spikes that failed to propagate to the soma unless synapses on the apical dendrites below the tuft were also activated (Supplementary Fig. 4 online). A potentially important aspect of the interaction between perforantpath and Schaffer-collateral synaptic inputs is their timing. We used our models to explore the timing relationship necessary to obtain gating of forward-propagating dendritic spikes. Strong perforant-path activation (8%—sufficient to induce dendritic spikes on most trials) was simulated in the weak dendritic excitability model, and both the strength and timing of the simulated distal Schaffer-collateral input were varied. Action-potential probability was greatest when the two inputs were coincident, and the efficiency of gating decreased as the delay between the two inputs increased (Fig. 4). The exact nature of the interaction between the Schaffer-collateral and perforant-path inputs depended on synaptic strength and timing in a complex way. For 3.5% Schaffer-collateral activation, timing could vary by a few milliseconds, but Schaffer-collateral input preceding perforant-path input was slightly more effective than the reverse. For 5% Schaffer-collateral activation, the interaction was stronger and the range of effective timing was broader. In contrast to the weaker Schaffer-collateral stimulus, however, the 5% Schaffer-collateral input was most effective at gating the dendritic spikes if it came slightly after the perforant-path input. Mechanism of dendritic spike gating We explored the mechanism by which the Schaffer-collateral input facilitates the forward propagation of distally evoked dendritic spikes, by examining the voltage in the dendritic tree for Schaffer-collateral stimuli just below and just above threshold for successful propagation (Fig. 5). Perforant-path inputs were simulated on each apical branch and set to conductances large enough to evoke dendritic spikes in the apical tuft. A weak Schaffer-collateral input (3 nS) was too small to facilitate propagation of the dendritic spike into the main apical dendrites (Fig. 5a–c, left). In this case, the voltage near the main bifurcation in the apical dendrites peaked about 4 ms after the onset of

VOLUME 8

[

NUMBER 12

[


ARTICLES


Action potential probability

1

0.8

0.6

0.4

5% SC 3.5% SC

0.2

1.5% SC

0 –15

–10 PP first

–5

0 ∆t (ms)

5

10 SC first

15

Figure 4 Dependence of PP-evoked dendritic spike propagation on the strength and timing of SC activation. Action-potential probability (soma or axon) is plotted as a function of the time difference between PP activation (8%) and SC activation (1.5, 3.5 or 5%). Probabilities were determined from 200 trials of randomly placed synapses for each value of %SC and latency (1 ms intervals).

the synaptic inputs (6.3 ms from the beginning of the simulation). A slightly larger Schaffer-collateral input (4 nS) was sufficient to successfully facilitate propagation through the main bifurcation and into the primary apical dendrite (Fig. 5a–c, right). In this case, the spike in the branch point occurred about 1 ms later than the peak of the failed spike in the previous case. We found that the spike did not propagate beyond the large branch point that separates the primary apical dendrite from the apical tuft (Fig. 5a, left; note sharp drop in maximum voltage around 400 mm in Fig. 5b, left; see also Figs. 1b and 3). The low safety factor for propagation of spikes through branch points is a known consequence of the decreased impedance at branch points19,20. In our models of CA1 neurons, reductions in spike amplitude were observed at multiple branch points, but failures of propagation were most evident at the large branch point that divided the main apical dendrite from the apical tuft. Backpropagating action potentials fail at the same location in the CA1 dendritic tree14. The mechanism by which Schaffer-collateral inputs rescued the propagation of dendritic spikes through this branch point can be inferred by examining the currents flowing in the major apical branch, just before the spike (4.3 ms after the onset of the synaptic stimulus). Compared to the smaller Schaffer-collateral input, which resulted in failure of the dendritic spike, the larger Schaffer-collateral synaptic input approximately tripled the Na current at this critical time and location (Fig. 5d). The extra Na current was the key to the successful forward propagation of the dendritic spike. A similar mechanism underlies enhancement of action-potential backpropagation by EPSPs in cortical pyramidal neurons21. Experimental tests of model predictions We performed experiments to test three important predictions of the model. First, we tested the prediction that, on their own, perforantpath inputs have a limited ability to trigger action potentials in CA1 neurons. Second, we tested the prediction that when perforant-path inputs do trigger action potentials, they do so by eliciting dendritic spikes. Third, we tested the prediction that the propagation of perforant-path–evoked dendritic spikes from the apical tuft to the soma is facilitated by modest activation of Schaffer-collateral synaptic inputs. To test whether perforant-path stimulation could trigger action potentials, we placed a large bipolar stimulating electrode in stratum lacunosum-moleculare of hippocampal slices. The perforant-path projection to CA1 is visible in this region under infrared, differential-


[

NUMBER 12

[

DECEMBER 2005

interference contrast optics. In response to stimulation of the perforant path in the presence of GABA-receptor antagonists (Methods), we recorded monosynaptic EPSPs using whole-cell patch-clamp recordings from CA1 somata. As the stimulus intensity was increased, EPSPs increased in amplitude, but were never large enough to evoke an action potential (Fig. 6; average maximum EPSP ¼ 6.7 ± 1.0 mV, n ¼ 11). These results are consistent with the model’s prediction that it is difficult to produce an action potential using a single stimulus of perforant path alone. In fact, it was even more difficult to get action potentials than predicted by the model, because single stimuli of the perforant path were never effective, even though some of the recorded neurons should have had relatively strong dendritic excitability14. To further explore the ability of perforant path to evoke action potentials, we stimulated perforant path using high-frequency bursts (100 Hz) of five or ten pulses. In most cells, it was possible to evoke action potentials using high-intensity stimulation with five pulses; moreover, in all cells, ten pulses were effective (Fig. 6). When spikes were driven by perforant-path synaptic stimulation in this way, actionpotential threshold was lower than the threshold for evoking action potentials with current injection at the soma (Fig. 7). Specifically, we used somatic current injections resembling excitatory post synaptic currents (EPSCs) in the same ten-pulse pattern used for synaptic stimulation. Synaptically evoked action potentials had significantly lower thresholds than those evoked by synaptic current injection (Fig. 7b). In many cases, even subthreshold responses to current injection were larger than the threshold for synaptically evoked action potentials (Fig. 7a). These observations are consistent with the notion that perforant-path–evoked action potentials are driven by spikes that begin in the dendrites22. In some cases, subthreshold synaptic responses exhibited ‘spikelets’ (Fig. 7c). We have shown previously that these small spikes are the somatically recorded counterparts of large dendritic spikes18. Spikelets were identified more frequently during local application of tetrodotoxin (TTX) near the soma (Fig. 7d; Methods), suggesting that spikelets are often masked by action-potential firing. In two cells, we obtained enough trials to compare the occurrence of action potentials (control) and spikelets (local TTX) at the same perforantpath stimulus intensity; in these cells, the ratio of trials exhibiting spikelets to those exhibiting spikes was close to unity (1.04 and 0.79, 10 trials each). These results indicate that spikelets reflect dendritic spikes that have spread effectively to the soma and would usually trigger an axonal action potential. To test whether the forward propagation of perforant-path–evoked dendritic spikes can be gated by Schaffer-collateral stimulation of the upper apical dendrites, we used a second stimulating electrode to activate Schaffer-collateral inputs (single pulse) at the end of a burst of perforant-path inputs—where the spikelets were usually observed (Fig. 8). Stimulation strength was set to produce single Schaffercollateral EPSPs of 3–5 mV in the soma and single perforant-path EPSPs of 2-4 mV, which summated in bursts to 15–20 mV (at or just below threshold for spikelets). Schaffer-collateral and perforant-path responses were also simulated using EPSC-like current injections. We applied TTX near the soma so as to block action-potential firing and reveal spikelets. At the stimulus intensities used in these experiments, spikelets were never observed in response to Schaffer-collateral stimulation and were only occasionally evoked by perforant-path stimulation alone. When the two inputs were coactivated, however, the frequency of spikelets increased dramatically (Fig. 8c). To test whether this enhancement of spikelets was caused by summation of the somatic EPSPs, as opposed to dendritic integration of the events, we substituted a somatic, EPSC-like current injection for the Schaffer-collateral

1671

ARTICLES Figure 5 Mechanisms of conditional dendritic spike propagation. (a,b) Maximum voltage in upper half of the apical dendritic tree. A powerful synapse (3 nS) was placed midway along each reconstructed section of the dendritic tuft (PP, 19 synapses, red dots). A single synapse was placed near the end of the primary apical dendrite (SC, red dot with black arrow). On the left, the SC input has the same conductance as the tuft synapses and the dendritic action potentials initiated by the PP input in the tuft fail to invade the apical dendrites below the major bifurcation. On the right, the SC input has a slightly larger conductance (4 nS) and spike propagation is successful. Parameters as given in Figure 1a, right. (b) Plots showing the voltage (and maximum voltage, dashed lines) at 6.3 ms (time just before the successful dendritic spike in the main apical dendrite in b (right); green and red arrows in c along the primary apical dendrite and the tuft branch containing the synapse marked by blue arrow. Black and blue arrows correspond to locations indicated in a. (c) Voltage versus time at the tuft (blue) and apical (black) synapses. Green and red arrows indicate the apical voltage traces at 6.3 ms. (d) Comparison between dendritic currents at the apical synapse at 6.3 ms (time of maximum depolarization at the apical synapse for the case of propagation failure and just before the spike in the upper apical dendrite when propagation is successful). Green and red bars correspond to the points indicated by arrows in c. Isyn, instantaneous synaptic current; Qsyn, total synaptic charge transferred up to that time; Na, sodium current; Kdr, delayed rectifier potassium current; KA, A-type potassium current.

5 0 –5 –10 –15 –20 –25 –30 –35 –40 –45 –50 –55 –60 –65 –70 mV

5 0 –5 –10 –15 –20 –25 –30 –35 –40 –45 –50 –55 –60 –65 –70 mV

100 µm

0

200

Distance (µm) 400

600

800

–40

0

–80

0

0

Voltage (mV)

–45 –60 –75

600

800

Tuft Apical

–30 –45 –60 –75

0.5

0.04

0.4

Charge (pC)

0.05

0.03 0.02 0.01

20

25

Isyn

5

0.3 0.2 0.1

0.3 0.2 0.1 0.0

Qsyn

10 15 Time (ms)

0.8

0.4

0.0

0.00

0

Na

0.6 0.4 0.2 0

0.20

increase the frequency of spikelets above that observed by perforant-path stimulation alone (Fig. 8c), suggesting that spikelet enhancement by coincident stimulation of perforant path and Schaffer collaterals indeed reflects gating of dendritic spikes beginning in the apical tuft and propagating toward the soma.

0.15 0.10 0.05 0.00 KA

DISCUSSION The long distance between the perforant-path input and the axon of CA1 neurons has raised questions concerning the function of these synapses. Early reports suggested that the perforant path was mainly

b

c Peak depolarization (mV)

5 mV 50 ms

25

Kdr

stimulus. The amplitude of the current injection was set to yield simulated EPSPs larger than the Schaffer-collateral–evoked EPSPs (Schaffercollateral EPSP ¼ 4.5 ± 0.3 mV, simulated EPSP ¼ 7.7 ± 0.8 mV, perforant-path EPSP ¼ 3.5 ± 0.6 mV, n ¼ 10). This protocol did not

a

20

15 10 5 0 0

Cells with stimulus evoked APs (%)

10 15 Time (ms)

Current (µA cm–2)

5

Current (mA cm–2)

0

Current (nA)

Distance (µm)

400

–15 Tuft Apical

–30

d

200

–40

–80

–15

V (6.3 ms) Vmax

Current (mA cm–2)

Voltage (mV)

c

40

V (6.3 ms) Vmax

40

Voltage (mV)

b

Voltage (mV)


a

n=5

100 n=5

75 50 25 0

n = 11

1 pulse

2 4 6 Stimulus intensity (µA)

5 pulse 10 pulse Stimulus number

Figure 6 Experimental perforant-path stimulus-response plot. (a) Representative traces of EPSPs evoked by perforant path stimulation across a range of intensities. Left: EPSPs in response to a single 0.2 ms stimulus. Middle: EPSPs in response to five 0.2 ms stimuli at 100 Hz. Right: EPSPs in response to ten 0.2 ms stimuli at 100 Hz. Action potentials are truncated. (b) Stimulus response plot of EPSP amplitude for a representative neuron (mean ± s.e.m.; closedsquares, single stimulus; open-squares, five stimuli at 100 Hz; triangles, 10 stimuli at 100 Hz). (c) Summary bar graph indicating the percent of neurons in which perforant-path stimulation induced an action potential recorded at the soma in response to high-intensity stimulation.

1672

VOLUME 8

[

NUMBER 12

[


ARTICLES Our findings indicate that perforant-path excitatory synaptic inputs can be effective in 50 ms two ways. They can reduce the number of Schaffer-collateral inputs required to trigger a –30 dendritic spike. In this case, the Schaffer** –35 collateral synapses are the dominant input and the perforant path can be considered –40 modulatory. Alternatively, perforant path is –45 the dominant input, but even a small Schaffer-collateral input can be crucial in facilitat–50 ing the propagation of dendritic spikes from –55 the apical tuft toward the soma and axon. This Synaptic Current latter scheme indicates that perforant path can injection Iinj serve as the major input to CA1 neurons. Syn In models with strongly excitable dendrites, the perforant-path input is actually more efficacious than distal Schaffer-collateral c d * * 10 ms inputs, owing to the generation of reliably 100 ms 10 ms propagating dendritic spikes in the small, 100 ms high-impedance branches of the apical tuft. * In models with weakly excitable dendrites, dendritic spikes are also readily triggered by Syn Syn perforant-path input, but their propagation toward the soma is unreliable. Activation of Figure 7 Dendritic spikes underlie perforant-path–evoked action potentials. (a) Threshold for action potentials evoked by perforant-path stimulation (Syn: 10 pulses at 100 Hz) occurs at more Schaffer-collateral synaptic inputs, however, hyperpolarized potentials than those evoked by somatic current injection (Iinj). Top and bottom left: promotes the forward propagation of perforaction potentials evoked by ten-pulse PP stimulation. Dashed line indicates action-potential threshold ant-path–evoked distal dendritic spikes. This 1 (identified as the voltage at which dV/dt first exceeds 5 mV ms ). Top right: depolarization induced by synaptic gating of dendritic spike propagation a current injection mimicking a PP-evoked EPSP burst, which failed to elicit an action potential. Dashed is consistent with previous reports that denline indicates peak depolarization. Bottom right: current injection–evoked action potential. Dashed dritic spike propagation is facilitated by the line indicates action-potential threshold. Insets contain magnified views of the action potentials. (b) Threshold of synaptically evoked action potentials is significantly hyperpolarized relative to currentdepolarization of the apical dendrites33 and injection evoked action potentials. **P o 0.05, two-tailed Student’s paired t-test, n ¼ 8. Data from resembles a similar effect predicted to occur individual cells (connected points) and mean ± s.e.m. are shown. (c) Subthreshold spikelet (asterisk, stochastically during the background synaptic magnified in inset) in response to minimal (sufficient to elicit a single action potential) high-frequency activation of CA1 dendrites34. PP stimulation (Syn: three 10-pulse, 100-Hz bursts at 3 Hz). (d) Local application of TTX to the soma In our models, the gate seems to exist near reveals spikelets (asterisks, second one magnified in inset) in response to high-frequency PP stimulation the border between perforant-path and Schaf(Syn: three 10-pulse, 100-Hz bursts at 3 Hz). fer-collateral synaptic inputs. It exists, in large measure, because of the major bifurcation of inhibitory5,7,23, a result of strong feedforward inhibition5,8,9,24–28. the apical dendrites at this location in many CA1 pyramidal neurons. There is considerable evidence, however, for an excitatory influence Both the availability of voltage-gated channels and the types of synaptic of the perforant path on CA1 cells5,24,27,29,30. Furthermore, place-field inputs will influence whether the gate is open or closed. When the gate firing occurs in CA1 cells deprived of Schaffer-collateral inputs by CA3 is closed by default—as a result of a low ratio of Na+ to K+ channels or dentate gyrus lesions31,32; this suggests that perforant path is capable (weak dendritic excitability14)—it can be opened by a modest Schafferof driving firing of CA1 neurons in vivo. Our findings suggest that the collateral synaptic input. Our experiments also suggest that bursts of excitatory influence of the perforant path depends on both dendritic perforant-path input could lead to successful forward propagation of spikes and an interaction with Schaffer-collateral inputs. dendritic spikes—an effect that can be reproduced in the models (data Previous reports on the interaction between perforant-path and not shown). Thus, perforant-path inputs may open the gate on their Schaffer-collateral inputs are seemingly contradictory. One report own, without the help of Schaffer-collateral inputs. To determine shows that bursts of perforant-path activation can increase the prob- whether this is likely to happen in vivo, it will be important for future ability of Schaffer-collateral–evoked action potentials in CA127, work to determine the firing patterns of layer 3 pyramidal neurons of whereas another study showed that perforant-path activation did not entorhinal cortex in awake, behaving animals. In our models, we also increase the response to Schaffer-collateral activation5. These two show that the gate can exist in an open state by default (strong dendritic findings can both be understood in the context of dendritic excitability. excitability14). Whether neurons operate in a weak or strong dendritic Perforant-path stimuli that are below threshold for dendritic spikes excitability mode in vivo is likely to be influenced by a number of may have little effect in the soma, because of the enormous attenuation factors, including ionic conditions33, neuromodulatory state and of perforant-path EPSPs along the dendrites. Furthermore, perforant- activity-dependent plasticity. Changes in the availability of A-type K+ path inputs may have a negligible effect when Schaffer-collateral channels, which are abundant in CA1 dendrites15, is one mechanism by inputs are strong enough to evoke spikes on their own. However, which transitions between the two default states (gated closed or open) perforant-path inputs that trigger dendritic spikes on their own or in could occur15,35,36. combination with milder Schaffer-collateral input can contribute to Feedforward inhibition is likely to further enrich synaptic integraaction-potential firing. tion in distal dendrites. In the perforant path, inhibition is likely to

b

20 mV


[

NUMBER 12

[

DECEMBER 2005

5 mV

10 mV

Threshold (mV)

10 mV

2 mV


a

1673

ARTICLES

a PP

SC TTX

1 nA

simSC

10 mV

b

250 ms

SC

simPP

PP

simPP

*

*

PP + SC

simPP

PP + simSC

simSC

c

Spikelet frequency (% trials)


simPP

40

** n = 10

30 20 10 0

PP

PP + simSC

PP + SC

temporally limit the excitatory influence of the perforant path; in contrast, in the Schaffer collaterals, inhibition may regulate the gating of distally evoked dendritic spikes. Our simulations indicate that appropriately timed Schaffer-collateral inhibition can prevent the spread of perforant-path–evoked dendritic spikes (that is, close the gate), even in strongly excitable dendrites. In weakly excitable dendrites, feedforward inhibition narrowed the time window for facilitation of dendritic spike propagation (that is, opening the gate) by the Schaffercollateral EPSP (see Supplementary Fig. 5 online). Modulation of inhibition is likely to be an important means of regulating synaptic integration via these mechanisms. When TTX was locally applied near the soma, the spikelets we observed—in response to perforant-path stimulation or combined perforant-path and Schaffer-collateral stimulation—resembled the ‘D-spikes’ or ‘fast prepotentials’ observed in early intracellular recordings from hippocampal pyramidal neurons23,37,38. That these events are somatic indications of larger dendritic spikes is consistent with the early reports and our previous somatic and dendritic recordings18. More than forty years ago, it was suggested that the propagation of dendritic spikes is required for the activation of CA1 neurons by distal synaptic inputs37. A few years later, researchers hypothesized that the propagation of dendritic spikes might have a ‘‘low safety factor,’’ requiring interactions at ‘‘points of confluence’’ in the dendrites to trigger a somatic action potential23. Still later, it was suggested that spikes originating in the apical tuft of neocortical pyramidal neurons could be modulated by synaptic inputs along the apical trunk39. Our findings validate these predictions by demonstrating such interactions in the distal dendrites of CA1 pyramidal neurons. Whether such interactions occur in other neurons remains to be determined. In layer 5 pyramidal neurons, distal input can increase the efficacy (gain) of a more proximal input40. This resembles the

1674

Figure 8 Coincident Schaffer-collateral stimulation increases spikelet frequency in response to perforant-path stimulation. (a) Experimental schematic (left): TTX was applied locally near the cell body throughout the experiment to block somatic action potentials. Stimulus protocols (right): perforant path (PP)—three 10-pulse, 100-Hz bursts at 3 Hz; Schaffer collateral (SC)—single stimuli timed to occur in coincidence with PP stimulation; simulated Schaffer collateral (simSC)—somatic current injection designed to elicit depolarizations with identical timing, similar kinetics, and equal or greater amplitude than the SC stimulation (difference of exponentials: trise ¼ 2.5 ms, tdecay ¼ 10 ms); simulated perforant-path (simPP)—somatic current injection with kinetics and amplitude similar to a single burst of PP stimulation, followed by a 2 nA, 5 ms current injection to test TTX efficacy. Action potentials were never evoked with this stimulus during TTX application, even though the stimulus was well above threshold under control conditions. (b) Responses to various combinations of the stimuli represented in a. PP + SC elicited spikelets (asterisks and inset) in a representative neuron. PP + simSC was unable to mimic the effect that SC stimulation had on PP spikelet production (inset). (c) Summary graph (mean ± s.e.m.) of spikelet frequency in response to PP, PP + simSC and PP + SC stimulation. **P o 0.0001 (repeated-measures analysis of variance followed by Dunn’s test with Bonferroni correction, P o 0.01).

modulatory effect we describe, but the mechanism may be different. In the neocortex, this effect occurs at least in part through the enhancement of action potential bursting, via an interaction between backpropagating action potentials and distal synaptic inputs40,41, which has not yet been described in CA1 pyramidal neurons. The situation where the distal input can serve as the dominant input and be modulated by a modest proximal input has not been studied in cortical pyramidal neurons. Thus, it will be important for future studies to determine whether this gating of distal dendritic spikes is a general feature of pyramidal neurons in various cortical regions. METHODS Computational modeling. All simulations were performed using the NEURON simulation environment42 using a 64-processor Beowulf cluster. Data presented in Figures 1 through 4 are from one reconstructed CA1 pyramidal cell. An additional reconstructed cell was used to repeat the numerical experiments, yielding qualitatively similar results. The model cells, along with all code for our simulations, are freely available on the web (http://www.northwestern. edu/dendrite). A third reconstructed cell, using a separate model of active dendritic channels was obtained from B. Mel (Univ. Southern California, http://www-lnc.usc.edu/CA1-pyramidal-cell-model/)17 and used to verify the results of the other cells. Passive and active properties of the models. Our models were prepared from CA1 neurons reconstructed after staining following intracellular biocytin filling in hippocampal slices from 55–57-d-old Wistar rats. The resulting compartmental models included passive membrane properties (Rm ¼ 40,000 Ocm2, Cm ¼ 0.75 mF cm2, Ri ¼ 200 Ocm) and three voltage-gated conductances: a Na+ conductance, a delayed rectifier K+ conductance and two variants of an A-type K+ conductance, implemented as described14,16. Additional details regarding the models used here are provided in the Results section and Figure 1. Synapse distribution. Synapses were distributed throughout the dendritic arbor of the reconstructed cells, with total of about 27,000 synapses in each of the first two reconstructed cells. The number and distribution of these synapses was modeled in accordance with experimental data43. In our study, only synapses in the distal apical dendrites were activated. Densities of excitatory synapses were higher in distal stratum radiatum (that is, distal Schaffer collaterals) than in stratum lacunosum-moleculare (that is, perforant path). The third cell was larger than the others (straight-line distance from soma to distal tip approximately 1,100 mm as opposed to 700 mm for the other two cells) and therefore included more synapses (44,391 total). As a result, lower percentages of synapses were needed to elicit dendritic spikes and somatic actions potentials in the third model.

VOLUME 8

[

NUMBER 12

[


ARTICLES


Synapse properties. AMPA receptor–mediated synaptic conductances were modeled as the difference of two exponentials (trise of 0.2 ms and tdecay of 2.0 ms) with a reversal potential of 0 mV. The conductance of a single AMPA synapse was set to 0.18 nS, a value chosen because it produced a somatic EPSP of 0.2 mV when activated at a synapse 50 mm from the soma44. Schaffer-collateral and perforant-path inputs. Schaffer-collateral input was simulated as the activation of a percentage of total synapses present in the distal apical dendrites, approximately 250–500 mm from the soma. On each simulation trial, a percentage of synapses were chosen randomly from the available pool. For the first two model cells, the total number of synapses (100% activation) available for distal Schaffer-collateral input were 3,838 and 4,760. The third cell was larger and had 10,110 synapses available for distal Schaffercollateral input. Perforant-path input was modeled as the activation of a percentage of available synapses in the apical tuft. For the first two cells, these synapses were situated approximately 500–750 mm from the soma whereas for the third cell, this distance was approximately 700–1,100 mm. As for the Schaffer-collateral input, perforant-path synapses were chosen randomly from the available pool on any individual trial. The numbers of synapses for the three cells were 2,511, 1,407 and 2,968. Hippocampal slices preparation. All animal procedures were approved by the Northwestern University Animal Care and Use Committee. Transverse hippocampal slices were obtained from 4–5-week-old male Wistar rats. Rats were anesthetized with halothane (Sigma-Aldrich), perfused transcardially with icecold artificial cerebrospinal fluid (aCSF) and decapitated. The brain was then removed from the skull and transverse hippocampal slices (300 mm thick) were prepared using a vibratome. The extracellular solution used in slice preparation and incubation (20–30 min at 35 1C) was either standard aCSF (125 mM NaCl, 25 mM dextrose, 25 mM NaHCO3, 2.5 mM KCl, 1.25 mM NaH2PO4, 2 mM CaCl2 and 1 mM MgCl2) or a sucrose-based solution (75 mM sucrose, 75 mM NaCl, 25 mM dextrose, 25 mM NaHCO3, 7 mM MgCl2, 2.5 mM KCl, 1.25 mM NaH2PO4 and 0.5 mM CaCl2). All solutions were bubbled with 95% O2 and 5% CO2 to maintain a pH of 7.4. Following incubation, but before recording, slices were maintained at room temperature (22 1C). Three of the experiments shown in Figure 8 were repeated in aCSF containing higher K+ (3 mM) and lower Ca2+ (1.3 mM with 0.7 mM Mg2+), which enhances dendritic excitability33. Despite this increase in dendritic excitability, the results shown in Figure 8 were not affected by ionic conditions. Patch-clamp recording. Individual slices were held in a small chamber perfused with aCSF at 1–3 mL min1 (37 1C) and visualized with an upright, fixed-stage microscope (Zeiss Axioscop 2 FS plus) using differential interference-contrast, infrared video microscopy. Whole-cell current-clamp recordings were made with a BVC-700 amplifier (Dagan Instruments) and patch electrodes with an open tip resistance of 3–4 MO. Series resistance (3–25 MO) and capacitance were compensated using the amplifier. The intracellular solution contained 115 mM potassium gluconate, 20 mM KCl, 10 mM sodium phosphocreatine, 10 mM HEPES, 2 mM Mg-ATP, 0.3 mM Na-GTP and 0.1% biocytin. Only cells with membrane potentials less than 55 mV at the onset of the whole-cell recording were used for experiments. In most experiments, cells were maintained at 67 mV throughout the experiment with DC current injection as needed. In three of the experiments shown in Figure 8, however, cells were held at the resting potential (that is, no holding current); no significant differences in the results were noted. Synaptic stimulation. Activation of the perforant path and Schaffer collaterals was obtained with two-contact cluster electrodes (CE2D55, FHC). The Schaffercollateral stimulating electrode was placed 100 mm to the side of the cell body of the recorded neuron and approximately halfway between the cell body layer and the perforant path (visually identified by the high density of axons). The second stimulating electrode was placed in the perforant path, 200 mm lateral to the soma of the recorded neuron. Synaptic stimulation trials were repeated at intervals of 5–15 s, to minimize effects of activity-dependent plasticity. In initial experiments, to assess the localization of synaptic stimulation, before transferring a slice to the recording chamber, we made a longitudinal cut along the border between stratum radiatum and the stratum lacunosum-moleculare. In a separate experiment, to rule out the involvement of disynaptic activation, we


[

NUMBER 12

[

DECEMBER 2005

made a cut from the CA3/CA1 border to the CA3/granule cell border, thereby removing CA3. Cuts did not influence the efficacy of synaptic stimulation and were therefore not used in subsequent experiments. Drugs. In all experiments, 2 mM SR95531 (Sigma-Aldrich) and 3 mM CGP52432 (Tocris Bioscience) were added to the aCSF to block GABAA and GABAB receptors, respectively. In some experiments, 10 mM TTX (SigmaAldrich) was dissolved in aCSF and applied locally using positive pressure (0.04–0.2 psi) applied to the back of a glass pipette (tip resistance of 3–4 MO) positioned 10–50 mm from the soma. Data acquisition and analysis. Data were transferred to a computer during experiments by an ITC-18 digital-analog converter (Instrutech). Igor Pro software (Wavemetrics) was used for acquisition and analysis. Electrophysiological records were filtered at 5 kHz and digitally sampled at 50–100 kHz. Statistical tests were performed using Excel software (Microsoft) or GB-STAT (Dynamic Microsystems). All results are reported as mean ± s.e.m., and significance was determined at the P o 0.05 level. Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS We thank members of the Spruston and Kath labs and B. Mel for discussions. This work was supported by the US National Institutes of Health (NS-35180 to N.S., NS-46064 to N.S. and W.L.K., and NS-045437 to T.J.) and by the National Science Foundation (DGE-9987577 to A.R.). COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/

1. Stuart, G., Spruston, N., Sakmann, B. & Haüsser, M. Action potential initiation and backpropagation in neurons of the mammalian CNS. Trends Neurosci. 20, 125–131 (1997). 2. Johnston, D., Magee, J.C., Colbert, C.M. & Christie, B.R. Active properties of neuronal dendrites. Annu. Rev. Neurosci. 19, 165–186 (1996). 3. Haüsser, M., Spruston, N. & Stuart, G.J. Diversity and dynamics of dendritic signaling. Science 290, 739–744 (2000). 4. Amaral, D.G. Emerging principles of intrinsic hippocampal organization. Curr. Opin. Neurobiol. 3, 225–229 (1993). 5. Colbert, C.M. & Levy, W.B. Electrophysiological and pharmacological characterization of perforant path synapses in CA1: mediation by glutamate receptors. J. Neurophysiol. 68, 1–8 (1992). 6. Desmond, N.L., Scott, C.A., Jane, J.A., Jr. & Levy, W.B. Ultrastructural identification of entorhinal cortical synapses in CA1 stratum lacunosum-moleculare of the rat. Hippocampus 4, 594–600 (1994). 7. Levy, W.B., Colbert, C.M. & Desmond, N.L. Another network model bites the dust: entorhinal inputs are no more than weakly excitatory in the hippocampal CA1 region. Hippocampus 5, 137–140 (1995). 8. Soltesz, I. Brief history of cortico-hippocampal time with a special reference to the direct entorhinal input to CA1. Hippocampus 5, 120–124 (1995). 9. Yeckel, M.F. & Berger, T.W. Monosynaptic excitation of hippocampal CA1 pyramidal cells by afferents from the entorhinal cortex. Hippocampus 5, 108–114 (1995). 10. Golding, N.L. & Spruston, N. Dendritic sodium spikes are variable triggers of axonal action potentials in hippocampal CA1 pyramidal neurons. Neuron 21, 1189–1200 (1998). 11. Kamondi, A., Acsady, L. & Buzsaki, G. Dendritic spikes are enhanced by cooperative network activity in the intact hippocampus. J. Neurosci. 18, 3919–3928 (1998). 12. Colbert, C.M. & Johnston, D. Axonal action-potential initiation and Na+ channel densities in the soma and axon initial segment of subicular pyramidal neurons. J. Neurosci. 16, 6676–6686 (1996). 13. Spruston, N., Schiller, Y., Stuart, G. & Sakmann, B. Activity-dependent action potential invasion and calcium influx into hippocampal CA1 dendrites. Science 268, 297–300 (1995). 14. Golding, N.L., Kath, W.L. & Spruston, N. Dichotomy of action-potential backpropagation in CA1 pyramidal neuron dendrites. J. Neurophysiol. 86, 2998–3010 (2001). 15. Hoffman, D.A., Magee, J.C., Colbert, C.M. & Johnston, D.K. K+ channel regulation of signal propagation in dendrites of hippocampal pyramidal neurons. Nature 387, 869–875 (1997). 16. Migliore, M., Hoffman, D.A., Magee, J.C. & Johnston, D. Role of an A-type K+ conductance in the back-propagation of action potentials in the dendrites of hippocampal pyramidal neurons. J. Comput. Neurosci. 7, 5–15 (1999). 17. Poirazi, P., Brannon, T. & Mel, B.W. Arithmetic of subthreshold synaptic summation in a model CA1 pyramidal cell. Neuron 37, 977–987 (2003).

1675


ARTICLES 18. Golding, N.L., Staff, N.P. & Spruston, N. Dendritic spikes as a mechanism for cooperative long-term potentiation. Nature 418, 326–331 (2002). 19. Goldstein, S.S. & Rall, W. Changes of action potential shape and velocity for changing core conductor geometry. Biophys. J. 14, 731–757 (1974). 20. Manor, Y., Koch, C. & Segev, I. Effect of geometrical irregularities on propagation delay in axonal trees. Biophys. J. 60, 1424–1437 (1991). 21. Stuart, G.J. & Hausser, M. Dendritic coincidence detection of EPSPs and action potentials. Nat. Neurosci. 4, 63–71 (2001). 22. Anderson, P., Storm, J. & Wheal, H.V. Thresholds of action potentials evoked by synapses on the dendrites of pyramidal cells in the rat hippocampus in vitro. J. Physiol. (Lond.) 383, 509–526 (1987). 23. Andersen, P. & Lomo, T. Mode of activation of hippocampal pyramidal cells by excitatory synapses on dendrites. Exp. Brain Res. 2, 247–260 (1966). 24. Buzsaki, G., Penttonen, M., Bragin, A., Nadasdy, Z. & Chrobak, J.J. Possible physiological role of the perforant path-CA1 projection. Hippocampus 5, 141–146 (1995). 25. Empson, R.M. & Heinemann, U. The perforant path projection to hippocampal area CA1 in the rat hippocampal-entorhinal cortex combined slice. J. Physiol. (Lond.) 484, 707–720 (1995). 26. Empson, R.M. & Heinemann, U. Perforant path connections to area CA1 are predominantly inhibitory in the rat hippocampal-entorhinal cortex combined slice preparation. Hippocampus 5, 104–107 (1995). 27. Remondes, M. & Schuman, E.M. Direct cortical input modulates plasticity and spiking in CA1 pyramidal neurons. Nature 416, 736–740 (2002). 28. Pare, D. & Llinas, R. Intracellular study of direct entorhinal inputs to field CA1 in the isolated guinea pig brain in vitro. Hippocampus 5, 115–119 (1995). 29. Doller, H.J. & Weight, F.F. Perforant pathway activation of hippocampal CA1 stratum pyramidale neurons: electrophysiological evidence for a direct pathway. Brain Res. 237, 1–13 (1982). 30. Yeckel, M.F. & Berger, T.W. Feedforward excitation of the hippocampus by afferents from the entorhinal cortex: redefinition of the role of the trisynaptic pathway. Proc. Natl. Acad. Sci. USA 87, 5832–5836 (1990). 31. Brun, V.H. et al. Place cells and place recognition maintained by direct entorhinalhippocampal circuitry. Science 296, 2243–2246 (2002).

1676

32. McNaughton, B.L., Barnes, C.A., Meltzer, J. & Sutherland, R.J. Hippocampal granule cells are necessary for normal spatial learning but not for spatially-selective pyramidal cell discharge. Exp. Brain Res. 76, 485–496 (1989). 33. Gasparini, S., Migliore, M. & Magee, J.C. On the initiation and propagation of dendritic spikes in CA1 pyramidal neurons. J. Neurosci. 24, 11046–11056 (2004). 34. Fellous, J.M., Rudolph, M., Destexhe, A. & Sejnowski, T.J. Synaptic background noise controls the input/output characteristics of single cells in an in vitro model of in vivo activity. Neuroscience 122, 811–829 (2003). 35. Frick, A., Magee, J. & Johnston, D. LTP is accompanied by an enhanced local excitability of pyramidal neuron dendrites. Nat. Neurosci. 7, 126–135 (2004). 36. Johnston, D., Hoffman, D.A., Colbert, C.M. & Magee, J.C. Regulation of back-propagating action potentials in hippocampal neurons. Curr. Opin. Neurobiol. 9, 288–292 (1999). 37. Spencer, W.A. & Kandel, E.R. Electrophysiology of hippocampal neurons. IV. Fast prepotentials. J. Neurophysiol. 24, 272–285 (1961). 38. Wong, R.K. & Stewart, M. Different firing patterns generated in dendrites and somata of CA1 pyramidal neurones in guinea-pig hippocampus. J. Physiol. (Lond.) 457, 675–687 (1992). 39. Cauller, L.J. & Connors, B.W. Functions of very distal dendrites: experimental and computational studies of layer I inputs to layer V pyramidal neurons in neocortex. in Single Neuron Computation (eds. McKenna, T., Davis, J. & Zornetzer, S.F.) 199–230 (Academic Press, San Diego, 1992). 40. Larkum, M.E., Senn, W. & Luscher, H.R. Top-down dendritic input increases the gain of layer 5 pyramidal neurons. Cereb. Cortex 14, 1059–1070 (2004). 41. Larkum, M.E., Zhu, J.J. & Sakmann, B. A new cellular mechanism for coupling inputs arriving at different cortical layers. Nature 398, 338–341 (1999). 42. Hines, M.L. & Carnevale, N.T. The NEURON simulation environment. Neural Comput. 9, 1179–1209 (1997). 43. Megias, M., Emri, Z., Freund, T.F. & Gulyas, A.I. Total number and distribution of inhibitory and excitatory synapses on hippocampal CA1 pyramidal cells. Neuroscience 102, 527–540 (2001). 44. Magee, J.C. & Cook, E.P. Somatic EPSP amplitude is independent of synapse location in hippocampal pyramidal neurons. Nat. Neurosci. 3, 895–903 (2000).

VOLUME 8

[

NUMBER 12

[


ARTICLES


CO M P U TATI O N AN D SYSTE M S

Matching storage and recall: hippocampal spike timing–dependent plasticity and phase response curves Ma´te´ Lengyel1, Jeehyun Kwag2, Ole Paulsen2 & Peter Dayan1 Hippocampal area CA3 is widely considered to function as an autoassociative memory. However, it is insufficiently understood how it does so. In particular, the extensive experimental evidence for the importance of carefully regulated spiking times poses the question as to how spike timing–based dynamics may support memory functions. Here, we develop a normative theory of autoassociative memory encompassing such network dynamics. Our theory specifies the way that the synaptic plasticity rule of a memory constrains the form of neuronal interactions that will retrieve memories optimally. If memories are stored by spike timing– dependent plasticity, neuronal interactions should be formalized in terms of a phase response curve, indicating the effect of presynaptic spikes on the timing of postsynaptic spikes. We show through simulation that such memories are competent analog autoassociators and demonstrate directly that the attributes of phase response curves of CA3 pyramidal cells recorded in vitro qualitatively conform with the theory.

The task of storing memories and recalling them from partial or noisy cues is fundamental for the brain and has been a particular focus for empirical1,2 and theoretical work3 on the hippocampus. This naturally raises the key question as to how the properties of single cells and the overall hippocampal network support its proposed function. The CA3 region of the hippocampus has the densest recurrent collateral system in the mammalian cortex4, which is consistent with the central role accorded to recurrent connections in standard models of autoassociative memories5. However, with relatively few exceptions (in the hippocampus and elsewhere6–13), such models typically use a highly simplified treatment of the resulting collective dynamics of their model neurons14,15. They thereby fail to capture a salient characteristic of the activity of hippocampal neurons during memory states: namely, the central role played by spike timing. Evidence for the importance of timing in the hippocampus is widespread. Behaviorally relevant neural oscillations at different frequencies pace the activity of all hippocampal cell types. In addition, the timing of individual action potentials of principal cells is tightly regulated16,17, and temporal sequences of the ensemble firing pattern consistently reappear during both awake behavior18,19 and sleep20–22. Furthermore, synaptic plasticity is also critically sensitive to the precise timing of pre- and postsynaptic spikes23,24. A key idea for understanding networks such as CA3 in which information may be coded by spike times is the phase response curve (PRC). PRCs offer a precise characterization of the effect a presynaptic spike has on the timing of the succeeding postsynaptic spike depending on its time of arrival25–27. PRCs and related phase reduction and

analysis methods have wide application in everything from the analysis of cardiac rhythms28 and patterns of oscillatory coordination for motor pattern generation29 to the relationship between nonlinear and subthreshold intrinsic mechanisms within cells and various forms of synchrony30,31. However, to our knowledge, hitherto they have not been used to characterize oscillatory autoassociative memories. Here, we present a theory treating autoassociative recall as optimal probabilistic inference32,33, inferring the recurrent dynamics within a memory that are normatively matched to the form of the synaptic plasticity rule used to store traces. In the case of CA3, this encompasses memories encoded in spike timings relative to underlying neural oscillations, and it thus involves inferring the optimal PRC of neurons from their spike timing–dependent synaptic plasticity (STDP) rule. We thus make specific predictions about the shape of the PRC of CA3 pyramidal neurons based on the STDP reported in cultured hippocampal neurons34. We show that the theoretically derived PRC provides a good qualitative match to those recorded in hippocampal CA3 pyramidal cells in vitro. RESULTS Autoassociative recall as probabilistic inference The fundamental requirement for an autoassociative memory is to recall a previously encoded memory trace when cued with a noisy or partial cue. As synaptic plasticity, the key mechanism for long term storage in the brain, loses information about the traces, recall poses a complex problem of probabilistic inference. The optimal solution to this problem, which amounts to a normative theory of recall, depends

1Gatsby Computational Neuroscience Unit, University College London, 17 Queen Square, London WC1N 3AR, UK. 2University Laboratory of Physiology, Oxford University, Parks Road, Oxford OX1 3PT, UK. Correspondence should be addressed to M.L. ([email protected]).

Received 19 April; accepted 12 September; published online 30 October 2005; doi:10.1038/nn1561


[

NUMBER 12

[

DECEMBER 2005

1677

ARTICLES

1

W1N X1

2

X2

3

X3

4

b

N

X4

XN

W13

c

W32

WN4

d

Coupling function Phase response (rad)


W12

2 1 Delay 0 Advance –1

–2 0 2 φpre– φpost (rad)

Spike timing–dependent plasticity Synaptic weight change

i=

1

0

–1

–50

0 50 tpre– t post (ms)

Phase response curves Phase response (rad)

a

0.2 0.1 0 –0.1

–2 0 2 Phase of EPSP (rad)

Figure 1 Normative theory of spike timing–based autoassociative memory. (a) Schematic diagram of a recurrent network of neurons. Neurons are numbered i ¼ 1 y N and are characterized by their respective activities, x1 y xN. Presynaptic neuron j is connected to postsynaptic neuron i through a recurrent synapse with efficacy (weight) wij. Although all-to-all connectivity (excluding autapses) was assumed for the theoretical derivations, here only a few synapses are shown for clarity. (b) Memories are stored by a spike timing-dependent plasticity (STDP) rule derived from experiments on cultured hippocampal neurons34. tpre and tpost represent times of pre- and postsynaptic firing. Gray lines are exponential fits24 to data from ref. 34. Black line is a continuous fit taken to be the synaptic learning rule (O) in equation (1). (c) Optimal coupling function (H) for retrieving memories stored by STDP (black line in b), as derived in equation (3). This shows how the firing phase of the postsynaptic neuron should change as a function of the phase difference between pre- and postsynaptic firing, if neurons were to interact continuously. fpre and fpost represent firing phases of pre- and postsynaptic cells relative to a local field potential oscillation. (d) Optimal phase response curves (PRCs) derived from the optimal coupling function (shown in c), showing how neurons should interact through spikes. Different curves correspond to linearly increasing synaptic weights (in increasing order: red, yellow, green, blue). ‘Zero’ phase is the phase of the postsynaptic spike.

on the statistical characteristics of the traces and cues and the nature of the lossiness of storage. More concretely, consider a network of N neurons (Fig. 1a), fully connected by N (N – 1) synaptic connections, and storing M memory traces. Memories are represented by distributed patterns of neural activity, with scalar variable xim characterizing the ith neuron in the mth memory trace. Here, we treat xim as neuron i’s spiking time relative to a reference point of an ongoing field potential oscillation, such as the peak of theta oscillation16,35. Storage amounts to changing the synaptic weights between neurons according to their activities in a memory trace using a synaptic plasticity rule: Dwij ¼ Oðxim ; xjm Þ

ð1Þ

This rule is local in that the change to the synaptic weight wij between presynaptic neuron j and postsynaptic neuron i depends only on the activities of these two neurons and not those of other neurons. Except for this constraint, we allow O to be an arbitrary function. We also make the simplifying assumption that synaptic plasticity is additive across the memories: X wij ¼ Oðxim ; xjm Þ ð2Þ m

Local and additive plasticity loses information in storing memories, because O transforms pre- and postsynaptic activity into a single scalar

1678

contribution. Indeed, equation (2) is noninvertible, with many different combinations of synaptic weight changes potentially leading to the same total synaptic weight. This lossiness implies that recall from a noisy or partial cue involves a process of combining probabilistic information from (i) the general statistical properties of the traces (that is, the prior distribution), (ii) the cue itself and (iii) the synaptic weights. Finding the statistically most likely memory trace involves complex, nonlocal operations. However, it can be well approximated (Supplementary Note online) by a form of neural dynamics among the recurrently coupled neurons in which there are explicit terms corresponding exactly to each of these three sources of information. Under this account, synaptic interactions between neurons of the network implement a ‘search’ for the activity pattern associated with the original memory trace that was most likely to have led to the cue. Over the course of search, each neuron gradually changes its spike timing such that the activities of it and its synaptic partners are increasingly likely to reflect a pattern consistent with the corresponding synaptic weights. As a result, the contribution to the dynamics of neuron i associated with the synaptic weight term involves a linear sum of influences from its presynaptic afferents j, with each element in the sum taking the form Hðxi ; xj Þ ¼ wij

@ Oðxi ; xj Þ @xi

ð3Þ

This has the obvious appealing characteristic that the strength of the interaction between two neurons is scaled by the synaptic weight. Less obvious is our key suggestion that the interaction should be determined by the derivative of the synaptic plasticity rule that was used to store memories in the network. We can derive an intuition about this rule by considering what happens if the synaptic weight is positive; that is, above an average value. This large synaptic weight arises from the weight changes associated with the memory traces. Therefore, the neuron should change its activity to increase the contribution that it and its synaptic partner would have made to the synaptic weight had their present activity indeed been associated with one of the memories that caused this excess synaptic weight. Here, we study the case of area CA3 in the hippocampus, in which the synaptic plasticity rule involves STDP. We show that the optimal dynamics for neurons representing memory traces in terms of the phase of firing relative to an underlying oscillation is determined by a particular shape of PRC that we experimentally validate. Spike timing-based memories The hippocampus, as well as other areas involved in memory processing, demonstrates prominent local field potential oscillations (LFPOs) under a variety of conditions, including both awake and sleep states36. In such cases, the phases of the spikes of a neuron relative to the LFPO have been shown to be carefully controlled19 and even to convey meaningful stimulus information, such as information about the position of the animal in its environment16. The discovery of STDP, for which the relative timing of pre- and postsynaptic firings determines the sign and extent of synaptic weight change, offered new insights into how the information represented by spike times might be stored37. However, except for some interesting suggestions about neuronal resonance12, it is less clear how one might correctly recall this information. Our theory allows a systematic treatment of this case, if we interpret neuronal activities as the phases of firing relative to an ongoing LFPO. This description is valid in the limit that neurons are driven to spike by a perithreshold oscillation approximately once in each cycle of the LFPO. First, we interpret storage (O in equation (1)) in terms of an STDP rule (Fig. 1b; recorded in cultured hippocampal neurons34), with

VOLUME 8

[

NUMBER 12

[


ARTICLES Figure 2 Quality of memory retrieval in an Storage Retrieval optimally constructed spike timing–based autoassociative memory model: numerical simulations. (a) Retrieval of a memory pattern 10 10 defined by firing phases. Left: noise-free form of 20 20 the cued pattern as originally stored. Each row shows the firing phase of a cell (color code) and 30 30 its notional spike time (white squares) depending 40 40 on when the phase of the underlying theta 50 50 oscillation (top trace) coincided with the firing 100 150 200 250 19,750 19,800 19,850 19,900 19,950 20,000 0 50 100 0 50 phase of the cell. Cells in all three panels were Time (ms) –2 0 2 ordered according to their firing phases in the Firing phase (rad) 0.4 0.4 noise-free memory trace. One complete cycle of 2 theta oscillation (125 ms) is shown. Center: first 0.3 0.3 two theta cycles during retrieval. Right: two theta 0.2 0 0.2 cycles from the end of the retrieval process. (b) Retrieval performance of the network. Three 0.1 0.1 –2 networks were compared: a ‘prior only’ network 0 0 (red), an ‘input only’ network (blue) and the 10 100 500 –2 0 2 –2 0 2 complete optimal network (yellow; see text for Stored firing phase (rad) Error (rad) Number of stored memories further details). Left: retrieved firing phases (y-axis) as a function of noise-free firing phases of the associated cued traces (x-axis). Points near the diagonal indicate good retrieval. Center: histogram of errors (circular difference of retrieved and stored firing phases) for the three networks. Right: root mean squared error over neurons for the ‘input only’ and the complete network as a function of the number of memories stored. The ‘prior only’ network is omitted from this plot for clarity, because its average error was close to the maximally possible p/2 value.

the relative phases translated directly into relative spiking times assuming a theta-frequency oscillation. This rule prescribes that a synapse is strengthened if the presynaptic neuron spikes before the postsynaptic neuron and is weakened if the order of spiking is reversed. The data around zero time difference between pre- and postsynaptic firing are variable, so we consider a smooth, differentiable fit (Fig. 1b, black line) that captures the salient characteristics of hippocampal STDP. Second, given this firing phase interpretation of storage, the optimal recall dynamics of equation (3) also acts on firing times, requiring the phases of the spikes of a postsynaptic cell to be advanced or delayed relative to the LFPO on the basis of the phases at which the neuron itself and its presynaptic partners spike. The STDP rule specifies weight changes based on the difference between pre- and postsynaptic spike times. Thus, in recall, the interaction between neurons should also be a function of this difference, sometimes called a phase coupling function38. Specifically, equation (3) tells us that the STDP rule shown in Figure 1b requires the phase coupling function shown in Figure 1c. However, since influences must be based on discrete spikes, a correction is necessary (Supplementary Note). Interactions are then described in terms of a PRC25 (Fig. 1d), which indicates the extent to which the timing of the next spike of a cell is advanced or delayed as a function of the timing relative to its most recent spike of a small perturbation caused, for instance, by an excitatory postsynaptic potential (EPSP). This relationship between the PRC of a neuron and the STDP rule employed by its synapses provides a way of testing our theory. Since our derivations embody assumptions and approximations, we tested the recall performance of the network by numerical simulations (Fig. 2). Memories were stored using an STDP rule derived from the hippocampus (Fig. 1b, black line), and retrieved by a network using the optimally matching coupling function (Fig. 1c). During the course of retrieval of a single memory trace (shown with one whole theta cycle in Fig. 2a, left), firing phases smoothly changed from an initially noisy value reflecting the input to the network (Fig. 2a, center, first theta cycle) to the final, converged form (Fig. 2a, right). The noise-free version of the cue (Fig. 2a, left) was recovered with high fidelity. There were no discernible changes in spike times in the last two cycles,


[

NUMBER 12

[

DECEMBER 2005

Average error (rad)

Frequency

b

Retrieved firing phase (rad)


Cell number

a

demonstrating that the network had reached an attractor state. Note that some of the noise was already removed by the second theta cycle. The network incorporated information from the three sources of evidence that pertain to recall: the prior distribution, the input and the synaptic weights (Fig. 2b). We used two purely feed-forward networks as benchmarks of the optimal recurrent network to dissect the contribution of interactions through the recurrent synapses. One network used only information in the prior distribution and thus always retrieved only 0, the mean of the prior distribution, whereas the other used only information in the input and thus simply transmitted its input to its output. Compared with these networks, individual values recalled by the full network were closer to their ideal values (Fig. 2b, left). The distributions of errors were symmetrical around zero (correct retrieval) for all three networks, indicating that all of them were unbiased (Fig. 2b, middle). However, the complete network was markedly superior over the other two in terms of variance of the retrieved firing phases around the correct values; that is, it made smaller errors on average (Fig. 2b, right). The comparison with using only the input is especially important, as persistent input is known to improve recall performance by itself 39. Of course, performance did ultimately deteriorate as an increasing number of patterns was stored, and information in synaptic weights became negligible relative to information in the noisy input (Fig. 2b, right). This graceful degradation arose from optimal integration of available information sources (synaptic weights and input, in this case). We also tested the robustness of our results in a number of adversarial settings (Supplementary Fig. 1 online), including storage noise and limited connectivity. Recall performance was proportional to the degree of connectivity and inversely proportional to storage noise (Supplementary Fig. 1). Performance was also proportional to the size of the network (Supplementary Fig. 1), as is common for such memories3. Although both the network sizes (50–250 neurons) and connectivity ratios (100–20%) used in our simulations were outside the range of realistic values for rat CA3 (300,000 neurons with 5% connectivity4), these results together imply that the main determinant of recall performance is the number of synapses per neuron, and thus we predict that a network with realistic anatomy

1679

ARTICLES

a

Figure 3 Experimental measurement of the PRC in CA3 hippocampal neurons. (a) Diagram of a CA3 hippocampal neuron with patch recording electrode at the soma and extracellular stimulation electrode among recurrent fibers in the stratum oriens. Sinusoidal inhibitory conductance mimicking hippocampal theta oscillation (5 Hz) was injected through the patch pipette using dynamic clamp. An EPSP was evoked using extracellular stimulation. (b) Average somatic EPSP recorded in response to extracellular stimulation without oscillation (n ¼ 5). (c) Sample of current-clamp recordings showing the phase response of a CA3 neuron (gray trace) to the stimulated EPSP (arrows; times of stimulation) during 5 Hz oscillation (black trace). (d) Plot of phase delay and advancement of the spike as a function of the phase of the EPSP. ‘Zero’ phase was defined as the average phase at which spikes occurred during 5 Hz oscillation without EPSP (vertical dotted lines in c). The PRC (open circles) was subject to Gaussian smoothing (gray line). Horizontal dotted lines show ± 2 s.d. of the average spike phase without EPSP. (e) Smoothed PRCs (gray lines) and raw data points (filled black circles) normalized for n ¼ 7 cells. Note that there are virtually no data points in the second quadrant.

b

1 mV

Recording electrode

20 ms

Advancement

200 ms

e 2

1 Delay 0 Advance –1

Phase response (normalized)

d

25 mV 1 nS

Delay

c

Phase response (rad)


Stimulation electrode

1.0

0.5

0.0

–0.5

–1.0 –3

1 2 –2 –1 0 Phase of EPSP (rad)

3

–3 –2 –1 0 1 2 Phase of EPSP (rad)

3

(12,000 recurrent synapses per CA3 pyramidal cell4, as opposed to the maximal 200 synapses used in our simulations) will be a highly competent spike timing–based autoassociator. Further, performance degraded only weakly even if the STDP was asymmetric: for instance, with the potentiation having larger maximal amplitude and tighter time frame than depression34 (Supplementary Fig. 1). Performance was more sensitive to a mismatch between storage rule and recall dynamics. Characteristics of the optimal phase response curve The theoretically optimal PRC (Fig. 1d) for autoassociative memory recall has five salient characteristics. First, excitatory currents can cause both delay (positive parts) and advancement (negative parts) of the next spike. Second, spike delay is predicted for EPSPs that follow postsynaptic spiking. Third, EPSPs immediately preceding postsynaptic spikes should have negligible effect on postsynaptic spikes. Fourth, EPSPs before this insensitive period or after the interval where delay is predicted should result in advancement. Fifth, based on equation (3) and shown as different colored lines in Figure 1d, the effect of presynaptic spiking on the phase response should scale with the synaptic weight between the two cells. The optimal scaling of the PRC is not exactly linear, but its zero crossings (relative spike times for which no phase shift is predicted) should be unaffected by changing the synaptic weight. Type II oscillators, such as the Hodgkin-Huxley model, show spiking behavior that broadly complies with these criteria38, thus suggesting that real neurons may implement similar PRCs. These features are preserved (Supplementary Fig. 2 online) for a range of STDP curves that satisfy a few qualitative properties: potentiation for pre- before postsynaptic firings, depotentiation for

1680

post- before presynaptic firings, pre- and postsynaptic spikes required to appear within a limited time window for both, and a transitionary regime between the potentiation and depotentiation at around zero time difference. The optimal PRC also seems to be insensitive to inputs arriving in the middle of the spiking cycle (shown as the two flat flanks of the PRC in Fig. 1d), unlike most biophysically plausible PRCs26. This insensitivity is predicted because the function used to fit experimental STDP curves converges to zero and is therefore already flat at 25 ms (Fig. 1b, black line), whereas for the conversion from spike time to phase, the length of the theta cycle was assumed to be 125 ms (corresponding to the widely reported 8 Hz theta frequency35). A shallower fall-off of the STDP curve (as shown by the original exponential fit of the data; Fig. 1b, gray line) or a higher theta frequency would diminish this region, leading to the fusion of the two intervals where advancement is predicted (Supplementary Fig. 2). Phase response curves of hippocampal CA3 pyramidal cells We used somatic whole-cell patch-clamp recordings from CA3 pyramidal neurons in acute hippocampal slices to measure the PRC for comparison with the theory. Theta oscillation was simulated by a somatic oscillatory inhibitory conductance, as is also observed in vivo35, and excitatory synaptic input was delivered by extracellular stimulation (Fig. 3a,b). We confirmed experimentally that excitatory input could both delay and advance spikes (Fig. 3c). As predicted by theory, delay was observed in the next cycle when EPSP followed immediately after a spike, and advancement was observed when EPSP occurred before the expected spike or well after the previous spike (Fig. 3d,e; n ¼ 7 cells). To confirm that both phase advancement and phase delay are due to the EPSP itself, and not to some other extracellular stimulation-evoked modulatory or network event, we repeated the experiment using dynamic clamp to simulate an excitatory input conductance (Fig. 4). Indeed, the same effects of phase delay and phase advancement were observed using artificial excitatory postsynaptic conductances (EPSGs, Fig. 4a,b). Moreover, we also confirmed that the effect on phase advancement and phase delay increased with synaptic conductance, with the zero crossings of the PRC remaining relatively unaffected (Fig. 4c), as predicted by theory (Fig. 1d). Similar results were obtained in seven other cells. Finally, in order to test the generalizability of our findings, we recorded PRCs at a higher but still within-theta band frequency (Supplementary Fig. 3 online) as well as in response to bursts of EPSGs (Supplementary Fig. 4 online) and found that PRCs were preserved under these conditions. In sum, individual

VOLUME 8

[

NUMBER 12

[


ARTICLES

a

b

Delay

Phase response (rad)

1.0

c

20 mV 1 ns

200 ms

0 –0.5 –1.0 –1.5 –3

–2 –1 0 1 2 Phase of EPSG (rad)

3

1.0 Phase response (rad)


Advancement

0.5

0.5 Delay

0 Advance

–0.5 –1.0 –1.5 –3

–2

–1

0

1

2

3

Phase of EPSG (rad)

Figure 4 Effect of EPSG amplitude on the PRC. (a) Sample of current-clamp recordings showing the response of a CA3 neuron (gray trace) to the EPSG during 5 Hz oscillation (black trace). The vertical dashed line represents the average phase at which spikes occurred during 5 Hz oscillation without EPSG. (b) Plot of phase delay and advancement of spike as a function of the phase of the EPSG (2 nS). The PRC (open circles) was subject to Gaussian smoothing (gray line). Horizontal dotted lines show ± 2 s.d. of the average spike phase without EPSG. (c) Smoothed PRCs produced by EPSGs of different amplitude in the same cell; 0.5 nS (red), 1 nS (yellow), 1.5 nS (green) and 2 nS (blue).

CA3 pyramidal neurons demonstrate intrinsic dynamics that support optimal retrieval of information by phase coding according to our theory. DISCUSSION We report a normative theory of statistically sound recall in analog associative memory networks. We have shown that the theory makes a direct link between the rule governing spike timing–dependent synaptic plasticity and the neurons’ PRCs, and we have qualitatively confirmed the precepts of the theory by recording conformant PRCs from hippocampal CA3 neurons. Our technique treats analog autoassociative memory from a probabilistic viewpoint32,33, deriving a general relationship between (i) the nature and representational substrate of the memory traces and the rules governing neural plasticity, and (ii) the dynamical behavior during recall that would approximately solve a formally presented task such as pattern completion or noise removal. Applied to the case of memory traces represented as phases, and stored by an STDP rule (derived from data from cultured hippocampal neurons34), the resulting dynamics specified a form of PRC. Not only were the general characteristics of this PRC consistent with those in the CA3 data (for instance, the existence of delays and advances), but also the more detailed predictions were matched, such as the scaling of the PRC with the input (in the dynamic clamp experiments) and even the form of the delays and advances relative to the standard firing phase. It is not at all trivial that the resulting PRC that was expected had a biophysically reasonable form, let alone that it matched actual PRCs in CA3. Indeed, in contrast to the type II–like PRCs we recorded here in hippocampal


[

NUMBER 12

[

DECEMBER 2005

CA3 pyramidal cells, classical integrate-and-fire dynamics produce only phase advancement in response to excitatory inputs26, and even neocortical pyramidal cells show phase response characteristics of Type I membranes and thus lack a delay component in their PRCs40. It is also not trivial that the network performed recall competently, as analog autoassociative memory is hard15. Our theoretical framework embodied a number of simplifying assumptions that allowed for an analytical derivation of the optimal recall dynamics but whose biological plausibility may seem to be unclear. We explicitly tested the incorporation of storage noise, limited connectivity and asymmetry in the STDP rule (Supplementary Fig. 1), showing that none of these had an importantly deleterious effect on performance. As one might expect from the framework, the most marked sensitivity is to mismatch between storage and recall (Supplementary Fig. 1). One more holistic assumption was that, in line with the traditional theory of autoassociative memories3,5, we treated memory encoding and retrieval separately, as distinct modes of operation. Specifically, neural activities during encoding were clamped to the memory patterns being stored so that the intrinsic dynamics of the network did not contribute to this process. Although, in its extreme form, this assumption is certainly unrealistic, there is suggestive data that changing levels of acetylcholine neuromodulation may result in the separation of memory encoding and retrieval in the hippocampus and related structures by selectively suppressing transmission and plasticity in afferent or internal synaptic pathways during these two operational modes41. Another assumption was to have addressed only the simplest form of oscillatory memory in which all neurons fired once per cycle. This was a marked abstraction of the hippocampus, whose pyramidal cells often fire bursts of action potentials in vivo16,18. The induction of synaptic plasticity is also most effective when bursts rather than single spikes are used in the stimulation protocol23, and spike timing–dependent plasticity has been shown to encompass multi-spike interactions42 and to be sensitive to the firing rate of pre- and postsynaptic cells43. Thus, an extension to a joint rate- and phase-based code for information is pressing44–46. We suggest that the choice of the number of spikes fired in a cycle (including no spikes) could convey orthogonal information, characterizing the certainty a neuron has about its phase or, indeed, its relevance for the given pattern (M.L. and P.D., unpublished data). Under this account, the consequences of firing potentially multiple spikes per theta cycle for memory encoding are straightforward: some memories will be stored and therefore retrieved with greater efficiency. Retrieval dynamics would also have to take into account the extra information conveyed by instantaneous firing rates. Preliminary experimental results (Supplementary Fig. 4) are compatible with the conclusion from the extension of our theory that interactions between bursting cells should be scaled versions of single spike-based PRCs. An intriguing suggestion evident in the single-case figures (Figs. 3c and 4a) is that, after a stimulation, not only is the phase of the very next spike altered but also the phases of a few successive spikes change27. Depending on the assumptions postsynaptic neuronal mechanisms might embody about a neuron’s presynaptic cousins, the theory can predict various forms for these multistep PRCs; it would therefore be interesting to characterize these more fully. Finally, oscillations in one structure are only a small part of the overall puzzle of memory. There is increasing evidence for the involvement of multiple structures that undergo oscillations of potentially different frequencies and intermittencies36 but are nevertheless tightly and jointly regulated10,47,48. Perhaps a first step will be to generalize and abstract away from single-neuron PRCs to a form of population PRC,

1681

ARTICLES indicating the overall effect of one oscillatory structure on another to which it has moderately dense connections and with which it is coordinated.


METHODS Simulations. Networks of N ¼ 200 neurons were simulated (50 randomly chosen cells are shown in Fig. 2a). Memories were stored by an additive learning rule (equations (1) and (2)) that was a circular fit to experimental STDP curves24, using a Gabor function: O(xi,xj) ¼ A exp (s cosDfij ), with Dfij ¼ (2p / Ty (xi – xj )) and Ty ¼ 125 ms. Best-fitting parameters determined by minimizing the squared error between –62.5 and 62.5 ms were A ¼ 0.03 and s ¼ 4 (Fig. 1b). The number of stored memories was M ¼ 10 (Fig. 2a,b, left and center) or was varied between M ¼ 10 and M ¼ 500 (Fig. 2b, right). Firing phases in memory patterns were drawn from a von Mises (circular Gaussian) distribution with mx ¼ 0 mean and kx ¼ 0.5 concentration (the prior distribution), resulting in a distribution of firing phases that matched those recorded in vivo for hippocampal pyramidal cell populations49,50. At retrieval, a randomly selected pattern from the list of stored patterns was used as the recall cue corrupted with unbiased additive circular Gaussian noise of k ¼ 10 concentration. Retrieval dynamics of the network was parameterized accordingly to optimally match the form and parameters of the prior and the noise distribution, as well as of the synaptic plasticity rule (Supplementary Note), and involved a phase coupling function H(xi,xj) ¼ 2pwij A / Ty exp (s cos Dfij ) (cos Dfij – s sin2 Dfij ) (shown in Fig. 1c). Differential equations were solved numerically by using an adaptable step-size method, and the states of neurons were recorded every 5 ms (simulated time; Fig. 2a) or at the end of the simulation after 20000 ms simulated time (Fig. 2b). For Figure 2b 10 networks with different lists of stored patterns were simulated, 10 retrieval attempts were made in each network. Data points show results pooled over networks and retrieval attempts. Experiments. Somatic whole-cell patch-clamp recordings were made from CA3 pyramidal cells in hippocampal slices prepared from postnatal day 13–19 Wistar rats and maintained at 29–31 1C (for details of extracellular and intracellular solutions, see Supplementary Methods online). Theta oscillation was simulated by 5 Hz or 8 Hz oscillatory inhibitory conductance of 1–2 nS peak amplitude using dynamic clamp. A positive tonic current was superimposed on the oscillatory input so that the membrane potential was depolarized just enough to evoke a single action potential near the peak of every cycle of the oscillation. EPSP was evoked via extracellular stimulation in the presence of 1 mM gabazine (SR95531). EPSG of peak amplitude 0.5–4.5 nS was injected using dynamic clamp at 20 different phases of the inhibitory oscillation, starting at ‘zero’ phase. Each PRC data point is an average of ten repetitions with the same stimulation phase. Data were acquired online and analyzed with custom-made procedures in Igor Pro. In normalizing (Fig. 3e and Supplementary Figs. 3 and 4), both smoothed PRC and raw data points were divided by the peak advancement value of the smoothed PRC for each cell. Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS We thank B. Gutkin, D. MacKay and E. Shea-Brown for valuable discussions and E.O. Mann and D. McLelland for their help with Igor procedures. This work was supported by the Gatsby Charitable Foundation (M.L., P.D.), the European Bayesian-Inspired Brain and Artefacts project (M.L., P.D.), the Biotechnology and Biological Sciences Research Council (J.K., O.P.), the Kwanjung Educational Foundation, Korea (J.K.) and the Oxford University Clarendon Fund (J.K.). COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/ 1. Squire, L.R. Memory and the hippocampus: a synthesis from findings with rats, monkeys, and humans. Psychol. Rev. 99, 195–231 (1992). 2. Cohen, N.J. & Eichenbaum, H. Memory, Amnesia, and the Hippocampal System (MIT Press, Cambridge, Massachusetts, 1993).

1682

3. Rolls, E.T. & Treves, A. Neural Networks and Brain Function (Oxford Univ. Press, Oxford, 1998). 4. Amaral, D.G., Ishizuka, N. & Claiborne, B. Neurons, numbers and the hippocampal network. Prog. Brain Res. 83, 1–11 (1990). 5. Hopfield, J.J. Neural networks and physical systems with emergent collective computational abilities. Proc. Natl. Acad. Sci. USA 79, 2554–2558 (1982). 6. Baird, B. Bifurcation and learning in network models of oscillating cortex. Physica D. 42, 365–384 (1990). 7. Wang, D.L., Buhmann, J. & von der Malsburg, C. Pattern segmentation in associative memory. Neural Comput. 2, 94–106 (1990). 8. Li, Z. Modeling the sensory computations of the olfactory bulb. in Models of Neural Networks Vol. 2 (eds. Domany, E., van Hemmen, J.L. & Schulten, K.) 221–251 (Springer-Verlag, New York, 1995). 9. Hendin, O., Horn, D. & Tsodyks, M.V. Associative memory and segmentation in an oscillatory neural model of the olfactory bulb. J. Comput. Neurosci. 5, 157–169 (1998). 10. Li, Z. & Hertz, J. Odour recognition and segmentation by a model olfactory bulb and cortex. Network 11, 83–102 (2000). 11. Hasselmo, M.E., Bodelon, C. & Wyble, B.P. A proposed function for hippocampal theta rhythm: separate phases of encoding and retrieval enhance reversal of prior learning. Neural Comput. 14, 793–817 (2002). 12. Scarpetta, S., Li, Z. & Hertz, J. Hebbian imprinting and retrieval in oscillatory neural networks. Neural Comput. 14, 2371–2396 (2002). 13. Jensen, O. & Lisman, J. Hippocampal sequence-encoding driven by a cortical multi-item working memory buffer. Trends Neurosci. 28, 67–72 (2005). 14. Hopfield, J.J. Neurons with graded response have collective computational properties like those of two-state neurons. Proc. Natl. Acad. Sci. USA 81, 3088–3092 (1984). 15. Treves, A. Graded-response neurons and information encodings in autoassociative memories. Phys. Rev. A. 42, 2418–2430 (1990). 16. O’Keefe, J. & Recce, M.L. Phase relationship between hippocampal place units and the EEG theta rhythm. Hippocampus 3, 317–330 (1993). 17. Poe, G.R., Nitz, D.A., McNaughton, B.L. & Barnes, C.A. Experience-dependent phasereversal of hippocampal neuron firing during REM sleep. Brain Res. 855, 176–180 (2000). 18. Skaggs, W.E., McNaughton, B.L., Wilson, M.A. & Barnes, C.A. Theta phase precession in hippocampal neuronal populations and the compression of temporal sequences. Hippocampus 6, 149–172 (1996). 19. Harris, K.D., Csicsva´ri, J., Hirase, H., Dragoi, G. & Buzsa´ki, G. Organization of cell assemblies in the hippocampus. Nature 424, 552–556 (2003). 20. Skaggs, W.E. & McNaughton, B.L. Replay of neuronal firing sequences in rat hippocampus during sleep following spatial experience. Science 271, 1870–1873 (1996). 21. Na´dasdy, Z., Hirase, H., Czurko´, A., Csicsva´ri, J. & Buzsa´ki, G. Replay and time compression of recurring spike sequences in the hippocampus. J. Neurosci. 19, 9497–9507 (1999). 22. Louie, K. & Wilson, M.A. Temporally structured replay of awake hippocampal ensemble activity during rapid eye movement sleep. Neuron 29, 145–156 (2001). 23. Paulsen, O. & Sejnowski, T.J. Natural patterns of activity and long-term synaptic plasticity. Curr. Opin. Neurobiol. 10, 172–179 (2000). 24. Bi, G. & Poo, M.M. Synaptic modification by correlated activity: Hebb’s postulate revisited. Annu. Rev. Neurosci. 24, 139–166 (2001). 25. Rinzel, J. & Ermentrout, B. Analysis of neural excitability and oscillations. in Methods in Neuronal Modeling 2nd edn. (eds. Koch, C. & Segev, I.) 251–291 (MIT Press, Cambridge, Massachusetts, 1998). 26. Brown, E., Moehlis, J. & Holmes, P. On the phase reduction and response dynamics of neural oscillator populations. Neural Comput. 16, 673–715 (2004). 27. Gutkin, B.S., Ermentrout, G.B. & Reyes, A. Phase response curves determine the responses of neurons to transient inputs. J. Neurophysiol. 94, 1623–1635 (2005). 28. Guevara, M.R., Glass, L. & Shrier, A. Phase locking, period-doubling bifurcations, and irregular dynamics in periodically stimulated cardiac cells. Science 214, 1350–1353 (1981). 29. Ermentrout, B. & Kopell, N. Learning of phase lags in coupled neural oscillators. Neural Comput. 6, 225–241 (1994). 30. Lampl, I. & Yarom, Y. Subthreshold oscillations of the membrane potential: a functional synchronizing and timing device. J. Neurophysiol. 70, 2181–2186 (1993). 31. Ermentrout, G.B. & Kopell, N. Fine structure of neural spiking and synchronization in the presence of conduction delays. Proc. Natl. Acad. Sci. USA 95, 1259–1264 (1998). 32. MacKay, D.J.C. Maximum entropy connections: neural networks. in Maximum Entropy and Bayesian Methods, Laramie, 1990 (eds. Grandy, Jr, W.T. & Schick, L.H.) 237–244 (Kluwer, Dordrecht, The Netherlands, 1991). 33. Sommer, F.T. & Dayan, P. Bayesian retrieval in associative memories with storage errors. IEEE Trans. Neural Netw. 9, 705–713 (1998). 34. Bi, G.Q. & Poo, M.M. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. J. Neurosci. 18, 10464–10472 (1998). 35. Buzsa´ki, G. Theta oscillations in the hippocampus. Neuron 33, 325–340 (2002). 36. Buzsa´ki, G. & Draguhn, A. Neuronal oscillations in cortical networks. Science 304, 1926–1929 (2004). 37. Abbott, L.F. & Nelson, S.B. Synaptic plasticity: taming the beast. Nat. Neurosci. 3, 1178–1183 (2000). 38. Ermentrout, B. Type I membranes, phase resetting curves, and synchrony. Neural Comput. 8, 979–1001 (1996). 39. Engel, A., Englisch, H. & Schu¨tte, A. Improved retrieval in neural networks with external fields. Europhys. Lett. 8, 393–399 (1989).

VOLUME 8

[

NUMBER 12

[



ARTICLES 40. Reyes, A.D. & Fetz, F.E. Effects of transient depolarizing potentials on the firing rate of cat neocortical neurons. J. Neurophysiol. 69, 1673–1683 (1993). 41. Hasselmo, M.E. Neuromodulation: acetylcholine and memory consolidation. Trends Cogn. Sci. 3, 351–359 (1999). 42. Froemke, R.C. & Dan, Y. Spike-timing-dependent synaptic modification induced by natural spike trains. Nature 416, 433–438 (2002). 43. Sjostrom, P.J., Turrigiano, G.G. & Nelson, S.B. Rate, timing, and cooperativity jointly determine cortical synaptic plasticity. Neuron 32, 1149–1164 (2001). 44. Lengyel, M., Szatma´ry, Z. & E´rdi, P. Dynamically detuned oscillations account for the coupled rate and temporal code of place cell firing. Hippocampus 13, 700–714 (2003). 45. Huxter, J., Burgess, N. & O’Keefe, J. Independent rate and temporal coding in hippocampal pyramidal cells. Nature 425, 828–832 (2003).


[

NUMBER 12

[

DECEMBER 2005

46. Huhn, Z., Orbań, G., E´rdi, P. & Lengyel, M. Theta oscillation-coupled dendritic spiking integrates inputs on a long time scale. Hippocampus 15, 950–962 (2005). 47. Siapas, A.G. & Wilson, M.A. Coordinated interactions between hippocampal ripples and cortical spindles during slow-wave sleep. Neuron 21, 1123–1128 (1998). 48. Sirota, A., Csicsva´ri, J., Buhl, D. & Buzsa´ki, G. Communication between neocortex and hippocampus during sleep in rodents. Proc. Natl. Acad. Sci. USA 100, 2065–2069 (2003). 49. Csicsva´ri, J., Hirase, H., Czurko´, A., Mamiya, A. & Buzsa´ki, G. Oscillatory coupling of hippocampal pyramidal cells and interneurons in the behaving rat. J. Neurosci. 19, 274–287 (1999). 50. Siapas, A.G., Lubenov, E.V. & Wilson, M.A. Prefrontal phase locking to hippocampal theta oscillations. Neuron 46, 141–151 (2005).

1683


ARTICLES

CO M P U TATI O N A N D SYST E M S

Neural population coding of sound level adapts to stimulus statistics Isabel Dean1, Nicol S Harper1,2 & David McAlpine1 Mammals can hear sounds extending over a vast range of sound levels with remarkable accuracy. How auditory neurons code sound level over such a range is unclear; firing rates of individual neurons increase with sound level over only a very limited portion of the full range of hearing. We show that neurons in the auditory midbrain of the guinea pig adjust their responses to the mean, variance and more complex statistics of sound level distributions. We demonstrate that these adjustments improve the accuracy of the neural population code close to the region of most commonly occurring sound levels. This extends the range of sound levels that can be accurately encoded, fine-tuning hearing to the local acoustic environment.

The auditory system is required to code sounds that vary enormously in level. Accurately assessing the overall level of a sound has important survival value, for example, in judging the distance of a sound source, the urgency of an alarm call or the fitness of a competitor. Furthermore, if sound level is represented accurately, then the levels within an ongoing sound can be better discriminated, aiding analysis and recognition of the sound. At the threshold of hearing (approximately 0 dB sound pressure level (SPL) in humans), the ear drum may move by only a fraction of the width of an atom; some natural environments are characterized by similarly low levels of sound, averaging only a few dB SPL1. In comparison, groups of vocal animals can generate sounds of more than 109-fold higher average intensity (90 dB SPL)1–3, whereas the upper limit of human hearing is met by sounds that are twelve orders of magnitude higher in intensity (120 dB SPL) than sounds that are just audible. Over this vast range of levels, the auditory system achieves remarkable accuracy in detecting changes in sound level: humans can hear changes in level of about 1 dB across most of the full range of hearing4–6. In order for sound level coding to be achieved using neural firing rates, firing rates must change with level across the full range of hearing. However, the majority of primary auditory nerve fibers have low thresholds to sound stimulation, with firing rates that saturate at low to middle sound levels, giving neural dynamic ranges (the range of levels over which firing rates change) of just 35 dB or so7,8. Even by including the smaller population of fibers with higher thresholds and nonsaturating responses, it does not seem that neural firing rates can account for the accuracy of coding over the full range of hearing9. As the limited neural dynamic range does not cover the range of levels over which hearing operates (the so-called ‘dynamic range problem’10) mechanisms must exist that extend the range of coding beyond that observed in auditory nerve firing rates. To this end, several adjuncts to a rate code for sound level have been considered9. However, the means by

which neural responses are normally assessed as a function of sound level bears little resemblance to the demands placed on auditory neurons under natural listening conditions. Traditionally, firing rate versus sound level functions are obtained using sounds separated by long silent intervals, randomized in presentation order, such that consecutive sounds can vary enormously in level11–14. Conversely, although natural sound levels can vary extensively over the long term, over short time periods, within a given environment, they more often fluctuate over a relatively limited range1. Neurons throughout the auditory system are subject to adaptive processes, causing a change in response over time during sustained input to the neuron. A possible function of these adaptive processes is to tailor the neural code to match the local sensory environment. We hypothesize that the neural code for sound level is flexible enough to take account of the time-varying distribution of natural sound levels, such that coding is context-dependent, allowing an efficient representation of auditory stimuli. We have examined the ability of auditory neurons in the inferior colliculus, the major midbrain nucleus in the ascending auditory pathway, to adjust their coding for sound level to take account of stimulus statistics. Our results show that the responses of inferior collicular neurons are rapidly adjusted according to the statistical distribution of sound levels presented. These adjustments alter the coding properties of the neural population such that coding accuracy is increased near the most commonly occurring sound levels. RESULTS We presented a diotic (identical in each ear) white noise to anesthetized guinea pigs for B7 min, during which time the sound level was set every 50 ms to a new value drawn randomly from a defined distribution (Fig. 1a,b). The full range of sound levels presented was 21–96 dB SPL. The distribution of sound levels consisted of one or more regions of probable levels, referred to as stimulus ‘high-probability regions’, from

1Department of Physiology and University College London Ear Institute and 2Centre for Mathematics and Physics in the Life Sciences and Experimental Biology, University College London, Gower Street, London, WC1E 6BT, UK. Correspondence should be addressed to I.D. ([email protected]).

Received 14 April; accepted 8 August; published online 6 November 2005; doi:10.1038/nn1541

1684

VOLUME 8

[

NUMBER 12

[


40

0.06 0.05 0.04 0.03 0.02 0.01

3,000 Time (ms)

e

Neuron 1

40 30 20 10 40 60 80 Sound level (dB SPL)

200

0

50

100 150 Time (ms)

f 120

Neuron 2

150 100 50 0 20

100

5,000

40 60 80 Sound level (dB SPL)

100

200

0 20 30 40 50 60 70 80 90 100 Sound level (dB SPL)

g

Neuron 3

Firing rate (spikes/s)

1,000


0

50

0 20

0.07

Probability

60




60

c

80

20

d

b

100

Pressure

a

Sound level (dB SPL)

ARTICLES

100 80 60 40 20 0 20


100

150

Neuron 4

100

50

0 20


100

Figure 1 Adjustments in responses of inferior collicular neurons to the mean sound level. (a) Level variations over 5 s of stimulus with high-probability region centered at 63 dB SPL. (b) Stimulus waveform (200 ms shown). (c) Level distribution for stimulus with high-probability region centered at 63 dB SPL. (d–g) Each panel shows rate-level functions of one neuron for four different sound level distributions, plus the baseline function (gray). Colored functions in all figures obtained with high-probability regions at 39 dB SPL (green), 51 dB SPL (blue), 63 dB SPL (red) and 75 dB SPL (cyan). Filled circles and thick lines on abscissa indicate midpoint and extent of the high-probability region of each stimulus.

which levels were selected with 0.8 overall probability (Fig. 1c). The remaining levels were selected with an overall probability of 0.2. Adjustments of neural responses to the mean sound level We first examined the effect of the mean sound level on neural ratelevel functions, using a sound-level distribution with a single highprobability region of width 12 dB (Fig. 1c). Neural responses were recorded to level distributions with four different high-probability regions, centered at 39, 51, 63 or 75 dB SPL, referred to hereafter as the 39-, 51-, 63- and 75-dB stimuli. For each level distribution, the mean spike count of the neuron to each sound level was calculated, and the resulting rate-level function was plotted. ‘Baseline’ rate-level functions (Fig. 1d–g, gray) show the mean spike count to 50-ms noise bursts, separated by 300-ms intervals, at sound levels randomly selected from a flat distribution over 21–96 dB SPL. Sound stimulation of this form is typically used to record rate-level functions of auditory neurons11–14. In contrast, when rate-level functions were recorded using the fluctuating noise stimuli, which give each sound level a local statistical context, three observations were commonly made (Fig. 1d–g). First, rate-level functions shifted along the abscissa when the position of the high-probability region was changed. This shift resulted in a change in the threshold sound level. The largest change in threshold from the baseline function was approximately 35 dB. Notably, when the high-probability region was positioned below a neuron’s baseline threshold, no neuron shifted its threshold to sound levels below its baseline threshold (see Fig. 1f). Second, the shifts in rate-level functions often resulted in neurons’ thresholds lying within the range of sound levels encompassed by the high-probability region of the stimulus. Third, increasing the mean sound level often reduced the maximum spike rate and the slope of the rate-level functions, particularly over higher sound levels. Adaptation of the population code to the mean sound level We first investigated whether the adjustments in rate-level functions to different sound-level distributions improve coding of those level


[

NUMBER 12

[

DECEMBER 2005

distributions by single neurons. We calculated the Fisher information, a measure of coding accuracy15, for rate-level functions of single neurons obtained for different sound-level distributions (Fig. 2a). Assuming an optimal decoder, higher Fisher information reflects higher coding accuracy; that is, a more accurate representation of the sound level and thus a higher capacity to discriminate nearby sound levels. An intuitive approximation of Fisher information is the square of the slope of the rate-level function divided by the spike count variance. Thus, Fisher information is high when the spike count variance is low and when the spike count changes steeply with changes in sound level. As the variance in the neural spike counts that we recorded from inferior collicular neurons tended to be low when the spike counts themselves were low, the maximum Fisher information was typically at sound levels just above the threshold of the rate-level function. In some cases, this placed the peak Fisher information close to the high-probability region of the relevant stimulus distribution (Fig. 2a), suggesting that the function of the adjustments in rate-level functions might be to improve the accuracy of coding over this region. However, the rate-level functions of different neurons were diverse, and the peak Fisher information for individual neurons often did not cover the entire high-probability region of the stimulus, nor was it always adjusted to be close to that region. To understand the implications of the adjustments in rate-level functions for sound-level coding, it is necessary to examine responses across populations of neurons. Evidence from many levels and modalities of the vertebrate nervous system suggests that information is represented by the activity of neuronal populations16. An estimate of the coding accuracy for the neural population was obtained by summing the Fisher information functions of the individual neurons. This analysis assumes that correlation in spiking noise between neurons is low. Although data concerning such correlations in the auditory system are scarce, correlation in spiking noise is known to be low between auditory nerve fibers17, and recent data suggest this to be the case also in the inferior colliculus18 (see also C.V. Seshagiri & B. Delgutte, Assoc. for Research in Otolaryngology Abstr. 685, 2003).

1685

ARTICLES

40

0.02 80 40 60 Sound level (dB SPL)

0 100

0.01 0 20

0.03

100

0.02 0.01 0 20

40

60

80

100

0.03 0.02 0.01 0 20


40

60

80

100


f Mean Fisher information (1/dB2)

e Mean Fisher information (1/dB2)


d Mean Fisher information (1/dB2)

Mean Fisher information (1/dB2)

c

0.02

0.01

0 20

40 60 80 100 Sound level (dB SPL)

g

0.03 0.02 0.01 0 20


100


100

h 0.04




0 20

0.02

0.03 0.02 0.01 0 20


100

0.04 0.03 0.02 0.01 0 20

We compared the population Fisher information for neurons adjusted to pairs of different sound-level distributions (Fig. 2b–g). Within each comparison, the Fisher information was obtained for exactly the same population of neurons. Across all these paired comparisons, adjustments of rate-level functions to a given distribution improved sound-level coding just above the midpoint of the highprobability region of that distribution. For example, comparing the Fisher information curve for the 39-dB stimulus to that for the 75-dB stimulus (Fig. 2b), the most commonly occurring sound levels in the 39-dB stimulus were coded best by responses adjusted to that stimulus. Conversely, for the 75-dB stimulus, the most commonly occurring levels were coded best by responses adjusted to the 75-dB stimulus. Finally, we calculated the population Fisher information for 31 neurons from which responses to all four sound level distributions were obtained (Fig. 2h). Adaptation to stimulus variance Natural environments may present the auditory system not only with changes in the mean sound level but also with changes in the extent of sound level fluctuations1. Adaptation of inferior collicular neurons to Figure 3 Neural adjustments to stimulus variance. (a) Level distribution for stimulus with wider (24 dB) high-probability region (compare to Fig. 1c). (b,c) Each panel shows rate-level functions of one neuron for two widths of the high-probability region of the stimulus, 12 dB (gray) and 24 dB (black). (d) Population Fisher information for stimuli centered at 63 dB SPL, with high-probability regions of width 12 dB and 24 dB (n ¼ 25).

1686

sound level variance of pure tones has been reported in the cat19. However, the effect of sound level variance on neural rate-level functions and on the accuracy of the code for sound level has not been examined. We therefore investigated the effect of extending the range of common sound levels in the distribution by comparing responses to sound-level distributions with high-probability regions 12 dB and 24 dB in width, centered at 63 dB SPL (Fig. 3a). The rate-level functions of some neurons varied with the width of the high-probability region (Fig. 3b). However, changes in rate-level functions with increasing width of the high-probability region were typically not marked (Fig. 3c). Despite this, the Fisher information curve for the population of neurons was slightly wider when neural responses had adjusted to the wider (24 dB) range of common sound levels (Fig. 3d). For both widths, the region of highest coding accuracy was positioned just above the midpoint of the high-probability region; the adjustments of neural responses to the wider high-probability region resulted in a marked increase in Fisher information at the edges of the region of high coding accuracy. The high-probability region was widened on both sides by 6 dB, and the resulting Fisher information curve widened by a similar magnitude. No change was observed in the Fisher information curve when the width of the highprobability region was increased for the stimulus centered at 51 dB SPL (data not shown), possibly because the wider range of (lower) sound levels in this stimulus extended the distribution significantly below the neurons’ baseline thresholds. Adaptation to stimulus bimodality Finally, we examined how inferior colliculus neurons encode stimuli with more complex sound level distributions. In ‘bimodal’ stimuli, sound levels were selected from a distribution with two highprobability regions, centered at 51 and 75 dB SPL or at 39 and 63 dB SPL (Fig. 4a).

a

b

0.04 0.03 0.02 0.01 0 20

c


d

Neuron 2

40 30 20 10 0 20

VOLUME 8

Neuron 1 Firing rate (spikes/s)

0.04


80

0.03

Figure 2 Adjustments in neural responses improve population coding accuracy near the mean sound level. (a) Rate-level functions of one neuron for stimuli with high-probability regions at 39 and 63 dB SPL, and corresponding Fisher information functions (dotted lines). (b–h) Population Fisher information. In this and all figures, the total population Fisher information has been divided by the number of neurons in the population to facilitate comparison between plots. (b) High-probability regions at 39 and 75 dB SPL (n ¼ 34 neurons); (c) 39 and 63 dB SPL (n ¼ 35); (d) 51 and 75 dB SPL (n ¼ 50); (e) 63 and 75 dB SPL (n ¼ 42); (f) 51 and 63 dB SPL (n ¼ 54); (g) 39 and 51 dB SPL (n ¼ 38); (h) all four sound level distributions (n ¼ 31).

Probability


0.06

Fisher information (1/dB2)

0.08

120


b 160


a


[

NUMBER 12

[

60 40 20 0 20


0.03 0.02 0.01 0 20



ARTICLES

0.02 0.01

c 100 Firing rate (spikes/s)

0.03

0 20

0.015 0.01 0.005 0 20



50 0 100 50 0 100 50 0 20

Individual rate-level functions did not show any obvious tendency to adjust to both high-probability regions in response to bimodal stimuli. However, when the population Fisher information for bimodal stimuli was compared with that for the corresponding unimodal stimuli, we found that the region of highest coding accuracy was adjusted so as to incorporate both high-probability regions. In the case of the 51- to 75-dB SPL stimulus, two regions of high accuracy were apparent in the population Fisher information, with a dip in accuracy between them; these regions of high accuracy corresponded to the louder ends of each of the two high-probability regions (Fig. 4b). This double-peaked Fisher information curve appeared to result, largely, from some neurons positioning the thresholds of their rate-level functions near one of the high-probability regions, and others positioning their thresholds near the other high-probability region, thereby dividing the neural population’s resources. These divisions were observed within animals (Fig. 4c). Rate-level functions of neurons with high baseline thresholds tended to shift to the louder high-probability region; ratelevel functions of neurons with low baseline thresholds tended to shift to the quieter high-probability region (data not shown). Most sound levels presented at high probability within a bimodal stimulus were coded less accurately than the same sound levels presented at high probability within a unimodal stimulus. However, most sound levels presented at high probability within a bimodal stimulus were coded more accurately than the same sound levels presented at low probability within a unimodal stimulus. These results indicate that the neural population did not simply adapt to the mean sound level (which lay between the two high-probability regions of the bimodal stimuli) but rather could apportion its coding resources so as to take account of a more complex stimulus distribution.

Neuron 1 Neuron 2 Neuron 1

Neuron 2

Neuron 1

Neuron 2


100

Figure 4 Neural adjustments to stimulus bimodality. (a) Level distribution for bimodal stimulus centered at 51 and 75 dB SPL. (b) Population Fisher information for bimodal stimulus shown in a (n ¼ 46). (c) Rate-level functions from two neurons, from one animal. Filled circles on functions indicate point of each neuron’s maximum Fisher information. Top and center: responses to unimodal stimuli; bottom: responses to bimodal stimulus shown in a.

clearly differed from their baseline function within the first 5 s of presenting the stimulus (Fig. 5a); for 20/50 neurons, the rate-level function from the first 5 s was already very similar to the average function from the final 50 s, such that the r.m.s. data showed no apparent decay. The rate-level function of only one neuron did not stabilize but was continually adjusted throughout the 7 min of stimulus presentation. Neural sensitivity to changes in sound level Our data demonstrate the extent to which a neuron’s stimulation history influences its response to the current stimulus. Thus, the neuron’s response will be a function of the current sound level and past sound levels. One of the simplest possible functions that might provide for the shifting rate-level functions is one in which neural responses are determined not by the absolute sound level, but by the difference between the current and immediately preceding sound level (the ‘step size’), regardless of the sound level distribution. For this reason, we examined whether firing rate versus step size functions were invariant with sound level distribution (Fig. 6a–c). We found that few neurons’ firing rates were dependent on step size in a manner independent of the distribution of sound levels. The rate versus step size functions of such neurons were broadly aligned across different sound level distributions, although never invariant (Fig. 6a). However, for most neurons, rate versus step size functions differed considerably in threshold and shape between sound level distributions (Fig. 6b,c), as did the rate-level functions. Thresholds of individual neurons to step size varied by up to 30 dB between different sound level distributions (Fig. 6b), and, for some distributions, firing rates were relatively insensitive to step size (Fig. 6c). Thus, a simple dependence of neural responses on level step size, such that responses were invariant with the sound level distribution, did not account for the neural adaptation we observed.


[

NUMBER 12

[

DECEMBER 2005

Number of neurons

R.m.s. difference (spikes/s)

Time course of neural adaptation The time course of neural coding adjustments was examined by plotting rate-level functions separately for consecutive 5-s segments of the 63-dB (12-dB width) stimulus. For most neurons, rate-level functions were b a c adjusted rapidly to a final, stable position, 25 250 First 5-s segment reproducible in response to subsequent sti20 200 Tau = 4.7 s 100 Baseline mulus segments. The time taken for the rate15 150 level function to stabilize was assessed by 10 100 Average from final 50 s calculating the root-mean-squared (r.m.s.) 50 5 50 Seventh 5-s segment difference between the rate-level function from each 5-s segment and the average rate0 0 0 0 10 20 30 40 80 0 100 200 300 50 60 70 level function from the final 50 s of the Decay time constant (s) Time (s) Sound level (dB SPL) stimulus (Fig. 5a,b). Single exponential decay functions were fitted to the r.m.s. data Figure 5 Time course of neural adaptation. (a) Rate-level functions of one neuron for high-probability region of stimulus only, from 5-s segments of stimulus, compared with average function from final 50 s for each neuron (Fig. 5b). of stimulus and baseline function. (b) Same neuron as a: root-mean-squared (r.m.s.) differences between The time taken for neural responses to rate-level function from each 5-s segment and average function from final 50 s of stimulus. Times on stabilize varied between neurons (Fig. 5c), abscissa: segment number multiplied by the 5-s segment duration, minus 2.5 s. Dashed line: with a median time constant of 3.2 s single exponential decay; t ¼ 4.7 s. (c) Histogram of t values for all neurons for which t o 40 s (n ¼ 50). Most neurons’ rate-level functions (n ¼ 46/51 neurons). Firing rate (spikes/s)


b Mean Fisher information (1/dB2)

Probability

a 0.04

1687

ARTICLES Two-tone suppression is dependent on nearby frequencies in the background noise suppres40 sing the response to the tone on the basilar 100 30 100 membrane. As the relative power of our wideband stimulus was constant across all fre20 50 50 quency components for all sound level 10 distributions, we can discount two-tone sup0 0 0 pression as a potential mechanism contribut–40 –20 0 20 40 –40 –20 0 20 40 –40 –20 0 20 40 Step size (dB) Step size (dB) Step size (dB) ing to the adjustments in coding that we describe. Our data have shown that, despite Figure 6 Responses of inferior collicular neurons to changes in sound level. (a–c) Each panel shows the diversity of adaptive mechanisms that are firing rates of a neuron during the current 50-ms epoch of the stimulus, as a function of the likely to shape inferior colliculus responses, difference in sound level between the current epoch and the previous 50-ms epoch, for different the concerted action of such processes seems sound-level distributions. to function so as to improve the accuracy of the neural code for sound level. Neural adaptation has been attributed a number of functions. For DISCUSSION Our data demonstrate that although auditory rate-level functions show example, there is evidence that adaptive processes in auditory cortex a restricted dynamic range, the range of sound levels over which the facilitate novelty detection by single neurons30. The adaptation that we dynamic range lies is mutable. Neural responses were rapidly adjusted in have described may be the auditory analogue of gain control processes a manner that tended to improve coding of the most probable sound in the visual system31–33. These processes have been suggested to levels by the neural population. There was a slight bias of coding improve coding by adjusting neural responses to the statistics of visual accuracy towards levels louder than those occurring most commonly. stimuli. An improvement in coding resulting from neural adaptation This may arise because the brain needs not only to encode the ongoing has recently been demonstrated for single neurons in the blowfly visual noise stimulus, or the ambient environment, but also to be prepared for system34,35. Our study is the first demonstration of neural adaptation, encoding any additional stimuli that may be encountered. Further, dependent on stimulus statistics, that improves the coding accuracy of although rate-level functions shifted in an approximately parallel a neural population; we have shown that the population code adapts manner over some mean sound levels, parallel shifts were not main- to the mean sound level, the variance and even bimodality. The capacity tained across all mean levels. This suggests that the neural code is not of the auditory brain to fine-tune to the local acoustic environment invariant with mean sound level, but rather that it retains some allows high accuracy in level discrimination to be maintained over a information about overall level, which is consistent with evolutionary wide range of sound levels, despite the limited dynamic range of demands for some representation of overall level. It is possible that mean individual neurons. sound level–independent representation occurs at stages in the auditory METHODS pathway higher than the inferior colliculus, such as auditory cortex. Physiological recordings. Extracellular recordings were made from the A number of mechanisms, possibly acting in combination, might colliculi of urethane-anesthetized guinea pigs, using standard underlie the adaptive effects we have described. It is unlikely that the right inferior techniques26 approved by the UK Home Office. Single neuron responses middle-ear muscle reflex is involved, as this reflex is only weakly active were recorded using glass-coated tungsten microelectrodes (Tucker Davis in guinea pigs20, is inactive in anesthetized animals21 and predomi- Technologies System III). White noise (o25 kHz) was presented diotically nantly affects transmission of very high intensity, low-frequency via sealed, calibrated earphones. Only integer dB values of sound level, between sounds22. A more likely contributing mechanism is the action of the 21–96 dB SPL, were used, giving 76 sound level values in total. For a given medial olivocochlear system, which feeds back directly to the receptor sound-level distribution, the level sequence and noise token were the same for hair cells of the inner ear, causing suppression; however, the time all neurons. For baseline rate-level functions, noise bursts were separated by constant of action of this efferent system is approximately 100 ms 300-ms intervals; levels were selected randomly from a flat distribution, and (ref. 23), precluding it from fully accounting for the adaptive effects each level was presented ten times. For all other stimuli, sound was presented reported here, which have a time constant of several seconds. A further in 5-s segments, separated by o0.4 s. All rate-level functions were plotted contributing mechanism is likely to be neural spike-frequency adapta- from the mean number of spikes occurring during the 50-ms epochs corresponding to a given sound level. We used a standardized neural latency of tion, or the decline in firing rate over time during a sustained stimulus. 8 ms, the minimum latency measured, so we do not assume the operation Such adaptation may arise through mechanisms intrinsic to the of a system that can account for differing neural latencies. The first and second neuron, through synaptic depression or through network interactions. 5-s segments of each stimulus were excluded from all analyses, except analyses Spike-frequency adaptation is a characteristic of primary auditory of the time course of changes in neural responses. For time course analyses, nerve fibers24,25 as well as of higher-level auditory neurons26. Thus, it we calculated the r.m.s. difference between the rate-level function from is possible that neural substrates for the changes in coding we observed each 5-s segment and the average function from the final 50 s of the originate at low levels of the auditory system. Previous studies have stimulus. This analysis was performed for the high-probability region of the shown that rate-level functions to pure tones shift to higher sound stimulus only, because other sound levels were presented too few times levels if the tones are presented in a constant level of background noise. within each 5-s segment. We then fitted the r.m.s. data for each neuron These effects are observed at all levels of the auditory system that have with a single exponential decay (Fig. 5b) and obtained a decay time constant. For 20/50 neurons, the r.m.s. data were better fitted by a flat line been examined12–14,27,28. This phenomenon may share some mechanthan by an exponential decay (F-test, P o 0.05 to accept exponential fit), isms with the shifts in rate-level functions that we observed, although indicating that the rate-level function had already, within the first 5 s, reached the effect of such changes in rate-level functions on the population the position occupied in the final 50 s of the stimulus; the time constant coding accuracy for sound level has not been measured. At subcortical of these neurons was designated 0 s. The time constant of one neuron was levels, noise-induced changes in rate-level functions to tones are 4600 s, indicating that the rate-level function had not stabilized within the thought to result largely from cochlear, or ‘two-tone’, suppression12,29. time of recording.


1688

b 150

Neuron 2

c 150

Neuron 3


Neuron 1



a

VOLUME 8

[

NUMBER 12

[



ARTICLES Measuring the Fisher information of the neural population. The population Fisher information function F(s) was used to measure the accuracy with which sound level s was encoded by the spike counts of the recorded neural population when s was presented in a given sound level distribution. In the limit of a large number of neurons, the reciprocal of the Fisher information equals the variance of the representation of s by the neural population. The square root of this variance is proportional to the just-noticeable difference, another common measure of coding accuracy. Correlation in spike count of neuron pairs due to shared stimulus history was small for all stimuli. The mean correlation coefficient averaged over sound levels, for all neuron pairs and all stimuli, was 0.056 (s.d. across pairs ¼ 0.13). Furthermore, there is some suggestion that correlation in intrinsic spiking noise between inferior collicular neurons is low18 (see also C.V. Seshagiri & B. Delgutte, Assoc. for Research in Otolaryngology Abstr. 685, 2003). Thus, it is reasonable to approximate the Fisher information by assuming that the neurons generate spikes independently, giving X fa ðsÞ FðsÞ ¼ a

where fa(s) is the Fisher information function of each neuron a, for the sound level distribution of interest. The Fisher information function of a neuron is calculated from the probability Pa[r | s] of neuron a giving r spikes when sound level s is presented. Hence, fa ðsÞ ¼

X r

Pa ½rjs

dln Pa ½rjs 2 ds

where the differential was performed by a five-point centered numerical algorithm. To obtain Pa[r | s] for the stimulus with the level distribution of interest, the number of spikes was counted during each 50-ms epoch, with a standardized 8-ms neural latency. F(s) was also calculated using each neuron’s measured latency, rather than the 8-ms latency; this had no qualitative effect on the results. For each sound level, s, a histogram was constructed of the number of epochs in the stimulus containing r spikes. This gave an R 76 matrix, where R is the maximum number of spikes in any epoch. For the high-probability region of the stimulus, only a random sample of the epochs was used so that the average number of epochs per dB was the same in the high-probability region as elsewhere. For the unimodal stimuli, the number of epochs used for each sound level averaged at 23.5. Owing to the sampling, each neuron’s Fisher information function is the median of 20 functions, where each function was obtained using a different sample. The matrix was convolved with a Gaussian kernel with 0.5 spikes standard deviation and 4 dB standard deviation, in order to remove spurious fluctuations. Finally, the matrix was converted into a probability matrix of spike count given sound level, Pa[r | s], by normalizing the sum of each sound level column to 1. In addition, F(s) was calculated using an approximation of the single neuron Fisher information fa(s) ¼ ya¢(s)2/sa(s)2, where ya¢(s) is the differential of the spline fit to the rate-level function, and sa(s) is the spline fit to the standard deviation as a function of sound level, for neuron a. The splines were fitted using the BARS algorithm36. This alternative method had no qualitative effect on the results. ACKNOWLEDGMENTS We thank P. Latham and J. Linden for discussions and C. Micheyl, T. Marquardt and J. Ashmore for critical reading of the manuscript. This work was supported by the Royal National Institute for Deaf People and the Medical Research Council (UK). COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/


[

NUMBER 12

[

DECEMBER 2005

1. Christopher Kirk, E. & Smith, D.W. Protection from acoustic trauma is not a primary function of the medial olivocochlear efferent system. J. Assoc. Res. Otolaryngol. 4, 445–465 (2003). 2. Aubin, T. & Jouventin, P. Cocktail-party effect in king penguin colonies. Proc. R. Soc. Lond. B 265, 1665–1673 (1998). 3. Simmons, J.A., Wever, E.G. & Pylka, J.M. Periodical cicada: sound production and hearing. Science 171, 212–213 (1971). 4. Houtsma, A.J., Durlach, N.I. & Braida, L.D. Intensity perception XI. Experimental results on the relation of intensity resolution to loudness matching. J. Acoust. Soc. Am. 68, 807–813 (1980). 5. Viemeister, N.F. Auditory intensity discrimination at high frequencies in the presence of noise. Science 221, 1206–1208 (1983). 6. Viemeister, N.F. Intensity discrimination of noise in the presence of band-reject noise. J. Acoust. Soc. Am. 56, 1594–1600 (1974). 7. Evans, E.F. & Palmer, A.R. Relationship between the dynamic range of cochlear nerve fibres and their spontaneous activity. Exp. Brain Res. 40, 115–118 (1980). 8. Sachs, M.B. & Abbas, P.J. Rate versus level functions for auditory-nerve fibers in cats: tone-burst stimuli. J. Acoust. Soc. Am. 56, 1835–1847 (1974). 9. Colburn, H.S., Carney, L.H. & Heinz, M.G. Quantifying the information in auditorynerve responses for level discrimination. J. Assoc. Res. Otolaryngol. 4, 294–311 (2003). 10. Viemeister, N.F. Intensity coding and the dynamic range problem. Hear. Res. 34, 267–274 (1988). 11. Winter, I.M. & Palmer, A.R. Intensity coding in low-frequency auditory-nerve fibers of the guinea pig. J. Acoust. Soc. Am. 90, 1958–1967 (1991). 12. Rees, A. & Palmer, A.R. Rate-intensity functions and their modification by broadband noise for neurons in the guinea pig inferior colliculus. J. Acoust. Soc. Am. 83, 1488–1498 (1988). 13. Gibson, D.J., Young, E.D. & Costalupes, J.A. Similarity of dynamic range adjustment in auditory nerve and cochlear nuclei. J. Neurophysiol. 53, 940–958 (1985). 14. May, B.J. & Sachs, M.B. Dynamic range of neural rate responses in the ventral cochlear nucleus of awake cats. J. Neurophysiol. 68, 1589–1602 (1992). 15. Dayan, P. & Abbott, L.F. Theoretical Neuroscience: Computational and Mathematical Modeling of Neural Systems (MIT Press, Cambridge, Massachusetts, 2001). 16. Pouget, A., Dayan, P. & Zemel, R.S. Inference and computation with population codes. Annu. Rev. Neurosci. 26, 381–410 (2003). 17. Johnson, D.H. & Kiang, N.Y. Analysis of discharges recorded simultaneously from pairs of auditory nerve fibers. Biophys. J. 16, 719–734 (1976). 18. Popelar, J., Nwabueze-Ogbo, F.C. & Syka, J. Changes in neuronal activity of the inferior colliculus in rat after temporal inactivation of the auditory cortex. Physiol. Res. 52, 615–628 (2003). 19. Kvale, M.N. & Schreiner, C.E. Short-term adaptation of auditory receptive fields to dynamic stimuli. J. Neurophysiol. 91, 604–612 (2004). 20. Avan, P., Loth, D., Menguy, C. & Teyssou, M. Hypothetical roles of middle ear muscles in the guinea-pig. Hear. Res. 59, 59–69 (1992). 21. Carmel, P.W. & Starr, A. Acoustic and nonacoustic factors modifying middle-ear muscle activity in waking cats. J. Neurophysiol. 26, 598–616 (1963). 22. Nuttall, A.L. Tympanic muscle effects on middle-ear transfer characteristic. J. Acoust. Soc. Am. 56, 1239–1247 (1974). 23. Boyev, K.P., Liberman, M.C. & Brown, M.C. Effects of anesthesia on efferent-mediated adaptation of the DPOAE. J. Assoc. Res. Otolaryngol. 3, 362–373 (2002). 24. Westerman, L.A. & Smith, R.L. Rapid and short-term adaptation in auditory nerve responses. Hear. Res. 15, 249–260 (1984). 25. Yates, G.K., Robertson, D. & Johnstone, B.M. Very rapid adaptation in the guinea pig auditory nerve. Hear. Res. 17, 1–12 (1985). 26. Ingham, N.J. & McAlpine, D. Spike-frequency adaptation in the inferior colliculus. J. Neurophysiol. 91, 632–645 (2004). 27. Smith, R.L. & Brachman, M.L. Response modulation of auditory-nerve fibers by AM stimuli: effects of average intensity. Hear. Res. 2, 123–133 (1980). 28. Phillips, D.P. & Hall, S.E. Spike-rate intensity functions of cat cortical neurons studied with combined tone-noise stimuli. J. Acoust. Soc. Am. 80, 177–187 (1986). 29. Costalupes, J.A., Young, E.D. & Gibson, D.J. Effects of continuous noise backgrounds on rate response of auditory nerve fibers in cat. J. Neurophysiol. 51, 1326–1344 (1984). 30. Ulanovsky, N., Las, L. & Nelken, I. Processing of low-probability sounds by cortical neurons. Nat. Neurosci. 6, 391–398 (2003). 31. Ohzawa, I., Sclar, G. & Freeman, R.D. Contrast gain control in the cat’s visual system. J. Neurophysiol. 54, 651–667 (1985). 32. Ohzawa, I., Sclar, G. & Freeman, R.D. Contrast gain control in the cat visual cortex. Nature 298, 266–268 (1982). 33. Shapley, R.M. & Enroth-Cugell, C. Visual adaptation and retinal gain controls. Prog. Retin. Res. 3, 263–346 (1984). 34. Brenner, N., Bialek, W. & de Ruyter van Steveninck, R.R. Adaptive rescaling maximizes information transmission. Neuron 26, 695–702 (2000). 35. Fairhall, A.L., Lewen, G.D., Bialek, W. & de Ruyter Van Steveninck, R.R. Efficiency and ambiguity in an adaptive neural code. Nature 412, 787–792 (2001). 36. DiMatteo, I., Genovese, C.R. & Kass, R.E. Bayesian curve-fitting with free-knot splines. Biometrika 88, 1055–1071 (2001).

1689


ARTICLES


Independence of luminance and contrast in natural scenes and in the early visual system Valerio Mante1, Robert A Frazor1,2, Vincent Bonin1, Wilson S Geisler2 & Matteo Carandini1 The early visual system is endowed with adaptive mechanisms that rapidly adjust gain and integration time based on the local luminance (mean intensity) and contrast (standard deviation of intensity relative to the mean). Here we show that these mechanisms are matched to the statistics of the environment. First, we measured the joint distribution of luminance and contrast in patches selected from natural images and found that luminance and contrast were statistically independent of each other. This independence did not hold for artificial images with matched spectral characteristics. Second, we characterized the effects of the adaptive mechanisms in lateral geniculate nucleus (LGN), the direct recipient of retinal outputs. We found that luminance gain control had the same effect at all contrasts and that contrast gain control had the same effect at all mean luminances. Thus, the adaptive mechanisms for luminance and contrast operate independently, reflecting the very independence encountered in natural images.

In the early visual system, two rapid adaptive mechanisms control the gain of neural responses1: luminance gain control and contrast gain control. Luminance gain control (also known as light adaptation) occurs largely in the retina. It adjusts sensitivity to match the locally prevalent luminance (light intensity). Rather than responding linearly with luminance, which can potentially vary over an extremely wide range, the retina effectively divides luminance by the local mean luminance1–3. Contrast gain control begins in the retina1,4–7 and is strengthened at subsequent stages of the visual system8,9. It effectively divides10 the responses by a measure that grows with the locally prevalent root-mean-square (r.m.s.) contrast, the standard deviation of the stimulus luminance divided by the mean luminance. This division provides a degree of contrast invariance: rather than depending linearly on contrast, responses are reduced in those locations of the visual field where contrast is high and increased where contrast is low. At least some components of luminance and contrast gain control should operate rapidly, because the eyes typically fixate a given location for only 200–300 ms, and eye movements will often bring the receptive fields of neurons in the early visual system over image patches that differ in luminance and contrast. What ranges of luminance and contrast are typically encountered in natural scenes? To what extent do luminance and contrast vary together? We addressed these questions by measuring luminance and contrast for image patches selected from calibrated natural images. We found that the typical ranges are substantial and that luminance and contrast are largely statistically independent. We then turned to the gain control mechanisms in the early visual system and found that their operation reflects the independence encountered in natural images. This finding indicates a close match between the statistical properties of natural scenes and the processing of luminance and contrast in visual systems11–15. Moreover,

it directly supports the hypothesis1,2,16,17 that contrast is a fundamental independent variable encoded by the early visual system. RESULTS Luminance and contrast in natural scenes We studied the changes in luminance and contrast encountered during a simulated saccadic inspection of a natural scene and found that these changes cover more than an order of magnitude (Fig. 1). We analyzed 300 calibrated natural images18 and simulated scan paths based on the measured statistics of saccades19. We considered image patches such as would be covered by a receptive field of a neuron in the early visual system (we repeated this analysis for different patch sizes). For each patch, we measured local luminance (the mean luminance in the patch) and local contrast (the r.m.s. contrast in the patch). The typical distance between fixations was sufficient to cause large changes in the local luminance and local contrast that fall within the receptive field of visual neurons (Fig. 1b). An analysis of the population of images confirmed that within an image, local luminance and local contrast typically varied by more than a factor of 10 (Fig. 2a–d, marginal distributions). We found local luminance and contrast to be nearly statistically independent. Indeed, the joint distributions of luminance and contrast were approximately separable (Fig. 2b). Independence was confirmed by examination of the conditional probabilities: in the ranges of luminance and contrast that were most likely, all contrasts were roughly equally probable given a luminance (Fig. 2c), and all luminances were roughly equally probable given a contrast (Fig. 2d). This substantial degree of independence was reflected in the correlation between luminance and contrast, which was a low –0.2. Independence held regardless of patch size: for patches ranging from 0.11 to 1.01 (full-width

1The

Smith-Kettlewell Eye Research Institute, San Francisco, California 94115, USA. 2Department of Psychology and Center for Perceptual Systems, University of Texas, Austin, Texas 78712, USA. Correspondence should be addressed to M.C. ([email protected]).

Received 28 July; accepted 6 September; published online 13 November 2005; doi:10.1038/nn1556

1690

VOLUME 8

[

NUMBER 12

[


ARTICLES

a

b 0.4

8

Contrast

1 7 6 5 4

0.2

7 2

3

3 1

regions of high luminance and low contrast (bottom right of Fig. 2b) were indeed due to the sky. To gain more insight into the observed independence of local luminance and contrast, we evaluated, separately, the effects of the phase and amplitude spectra of the natural images on the joint distribution of luminance and contrast. We assessed the role of phase by measuring local luminance and contrast for artificial-phase images (Fig. 2e–h). These images had the same amplitude spectrum (that is, the same autocorrelation function) as the original images, but they had a random phase spectrum20. They were, approximately, samples of Gaussian noise with an amplitude spectrum that fell as 1/f, a distribution that is typically assumed to be representative of natural images21. In these artificial-phase images, luminance and contrast were far from independent (Fig. 2f); the average within-image correlation was –0.77 ± 0.01. The opposite was true in artificial-amplitude images (Fig. 2i–l). These images preserved the phases of natural images, but had an amplitude spectrum that decreased with frequency as 1/f n (as fitted to the amplitude versus frequency curve of each image, n ¼ 1.25 ± 0.02). In these images, luminance and contrast were much more independent (Fig. 2j); the average within-image correlation was –0.05 ± 0.02. A look at the conditional probabilities for luminance and contrast confirmed their marked dependence in artificial-phase images (Fig. 2g,h) and a much higher degree of independence in artificial-amplitude images (Fig. 2k,l). Thus, statistical independence of luminance and contrast was not trivial (1/f noise does not have this property), but was dependent on the typical phase structure of natural images (see Discussion).

5

6

4 8

0.0 100

1,000

10,000

Luminance (cd m–2)

Figure 1 Luminance and contrast in a natural scene. (a) A sequence of fixations in a natural scene. The crosses indicate fixation locations and the circles represent the corresponding locations of an arbitrary receptive field (diameter: 11). (b) Enlargements of the image patches falling within the receptive field as a function of their r.m.s. contrast (ordinate) and average luminance (abscissa).

at half-height), the within-image correlations were always low, ranging from –0.22 (±0.02, s.e.m., n ¼ 300) to –0.21 (± 0.02). The mild negative correlation between luminance and contrast was due to the portions of the images taken up by the sky, where local luminance was high but local contrast was, on average, very low. We measured local luminance and local contrast for some of the obvious constituents of natural images: sky, foliage, ground and backlit foliage. From each image, we selected, by hand, rectangular regions containing these constituents, while excluding ambiguous regions. Approximate statistical independence was found for each of these constituents—the correlations were 0.0, 0.2, –0.1 and –0.2 for sky, foliage, ground and backlit foliage. This analysis of the constituents also showed that the


[

a

c

b

Contrast

Natural images

Artificial phase

d

Prob (luminance | contrast)

0.2 0.1 0.05 0.02 50

e

Prob (contrast | luminance)

2 1 0.5

500

50

5,000


f

500

5,000


g

50

500

5,000


h

5 Contrast

Figure 2 Statistics of local luminance and contrast in natural images. (a) A natural image (same as in Fig. 1a). (b) Joint distribution of luminance and contrast as sampled from all 300 images. These distributions represent the variation of luminance and contrast within a typical image: specifically, we first computed the overall average luminance and contrast across images, and then rescaled each image so that its average luminance and contrast would match the overall average. The contours delineate the regions containing 90% (red), 65% (blue) and 40% (green) of the observations. The curves on the sides of the joint distribution indicate the marginal distributions of luminance and contrast. (c) Conditional probability of observing a certain contrast given a specified luminance. This distribution is obtained by normalizing vertical slices of the joint distribution in b. (d) Conditional probability of observing a certain luminance given a specified contrast. This distribution is obtained by normalizing horizontal slices of the distribution in b. (e) An artificial-phase image. This image has the same amplitude spectrum as the image in a, but a random phase spectrum. (f–h). Joint distribution and conditional probability distributions for the 300 artificial-phase images. Format as in b–d. (i) An artificial-amplitude image. This image has the same phase spectrum as the image in a, but the amplitude at each frequency is given by the 1/f n spectrum that best fits the spectrum of the original image. (j–l). Joint distribution and conditional probability distributions for the 300 artificial-amplitude images. Format as in b–d.

Luminance gain control and contrast gain control Thanks to luminance and contrast gain control, sudden changes in luminance or contrast that occured between fixations had a reduced

2 1 0.5 0.2 0.2 0.5 1 2 Luminance

i

NUMBER 12

Artificial amplitude

5

j

0.2 0.5 1 2 Luminance

5

k

0.2 0.5 1 2 Luminance

5

0.2

5

l

10 5 Contrast


2

2 1 0.5 0.2 0.1

[

DECEMBER 2005

0.2

1 Luminance

5

0.2

1 Luminance

5

1 Luminance

1691

ARTICLES

b 100 spikes s–1

64

20 ms

32 0 6.5

7.0

d 100 spikes s–1

c

19.5

20.0 Time (s)

20.5

40 ms

40.0 40.5 Time (s)

impact on the responses of the early visual system (Fig. 3). We recorded the responses of neurons in the lateral geniculate nucleus (LGN), which receives the output of the retina and provides input to the visual cortex. The recordings were performed extracellularly in anesthetized, paralyzed cats. LGN responses were barely affected by sudden steps in luminance (Fig. 3a) and were weakly affected by changes in contrast (Fig. 3c). The measured responses were much smaller and occured faster than the high-luminance responses predicted by low-luminance measurements of the receptive field (Fig. 3b) or the high-contrast responses predicted by low-contrast measurements of the receptive field (Fig. 3d). These reductions in gain and the changes in dynamics occured well within a cycle of the drifting grating (80 ms in Fig. 3a, 128 ms in Fig. 3c), confirming that the gain control mechanisms operate very quickly, in less than 100 ms1,5,6,22–26. Do the mechanisms of gain control for luminance and contrast reflect the independence of luminance and contrast seen in natural images? To react appropriately to the changes in luminance and contrast, the corresponding gain control mechanisms should be functionally independent. In other words, within the range of luminances encountered during natural viewing, luminance gain control should have the same effects at all contrasts, and contrast gain control should have the same effects at all mean luminances. Instead, if the gain control mechanisms were appropriate for statistics other than those in the natural environment—for example, for those of 1/f noise—one would expect that contrast gain control would be biased by local luminance or that luminance gain control would be biased by local Figure 4 Characterizing LGN responses at various luminances and contrasts. (a–c) Responses of an LGN neuron (X-type, on-center) to temporal frequency sweeps at (a) low luminance and low contrast (L ¼ 6 cd m–2, C ¼ 10%, Michelson contrast), (b) low luminance and high contrast (L ¼ 6 cd m–2, C ¼ 100%) and (c) high luminance and high contrast (L ¼ 54 cd m–2, C ¼ 10%). Histograms (gray) were obtained by averaging over ten stimulus presentations. Red curves are descriptions of the responses by the descriptive model (Fig. 5a). Stimuli were sinusoidal gratings at optimal spatial frequency (icons). The temporal profile of the stimuli is shown under the responses; drift rate increased exponentially with time, from 0.5 Hz to 40 Hz in 5 s, and back (not shown). (d–f) Impulse responses used for the predictions in a–c. The impulse response is smaller and faster at the higher contrast (e) or luminance (f) than at low luminance and contrast (d, and dotted curves).

1692

contrast. In other words, one would expect the visual system to exploit the redundancy implicit in any lack of independence. Independence of gain control mechanisms To test for independence, we characterized the effects of luminance and contrast gain control in the LGN. We recorded responses to drifting gratings (Fig. 4) with mean luminance (6–56 cd m–2) and contrast (10–100% Michelson contrast; 0.07–0.7 r.m.s. contrast) covering a range extending over a factor of 10, similar to the excursion seen in patches of natural images (Fig. 2b). To fully quantify the effects of gain control on both the amplitude and the dynamics of the responses1, we measured responses to a range of frequencies by increasing the drift rate exponentially with time from 0.5 Hz to 40 Hz in 5 s and back to 0 in the subsequent 5 s (Fig. 4a–c, and Supplementary Fig. 1 online). The responses to these stimuli can be read as tuning functions for stimulus temporal frequency. As expected1, the preferred temporal

a

L = 6 cd m–2 C = 10%

d

b

L = 6 cd m–2 C = 100%

e

L = 54 cd m–2 C = 10%

c 0.5

2

10

40

Spikes per cd s2

3.5

100 Spikes s–1

3.0

Figure 3 Effect and time course of gain control mechanisms in LGN. (a) Response of an LGN neuron to a drifting grating of constant contrast (14%), whose luminance steps from 32 cd m–2 to 56 cd m–2 (left) and back to 32 cd m–2 (right). Spatial frequency and temporal frequency (12.5 Hz) are optimal for this neuron. The temporal profile of the stimulus is shown below the responses. Histograms (gray) were obtained by convolving the spike trains with a Gaussian window (s ¼ 5 ms), and averaging over three stimulus presentations. From the histograms, we computed the average response to a cycle of the stimulus before (dashed) and after (black) the step in luminance. The linear prediction (green) was obtained by scaling the response before the step (dashed) by the ratio of the two luminances. (b) Comparison of average responses to low luminance (dashed) and high luminance (black), and of the response expected in the absence of gain control (green). (c) Response of an LGN neuron to a drifting grating of constant luminance (32 cd m–2) whose contrast steps from 31% to 100% (left) and back to 31% (right). Spatial frequency and temporal frequency (7.8 Hz) are optimal for this neuron. Histograms (gray) are the average over five stimulus presentations. The linear prediction (green) was obtained by scaling the response before the step (dashed) by the ratio of the two contrasts. (d) Comparison of average responses to low contrast (dashed) and high contrast (black), and of the response expected in the absence of gain control (green).

f 50 ms

Frequency (Hz) Luminance (cd m–2)



a

64 32 0 0

1

2

3

4

5

Time (s)

VOLUME 8

[

NUMBER 12

[


Luminance Contrast

c

d

Separable model

*

*

+

V0

0

11

0

Noise Luminance Contrast

0 20 40 60 80 100 Explained variance (%)

Figure 5 The two models used to describe LGN responses, and a measure of their performance. (a) The descriptive model: stimulus luminance is integrated by the spatial receptive field, filtered by the impulse response (red), added to Gaussian noise and rectified. (b) Quality of predictions for the descriptive model, measured by the percentage of stimulus-driven response variance explained by the model. N ¼ 40. The median is 85% (arrow). (c) The separable model. The impulse response is the convolution of three filters: a fixed filter (pink), a luminance gain filter (blue) and a contrast gain filter (green). (d) Quality of predictions for the separable model. The median explained variance is 81% (arrow).

Differences between predicted and measured responses were due mostly to short transients occurring at the onset of the rising phase of a cycle. These transients corresponded to bursts of action potentials27, which the model was not designed to produce. When either mean luminance or contrast was changed, the gain control mechanisms had potent effects on the impulse response4,28 (Fig. 4d–f). One effect of increasing mean luminance or contrast was on the impulse response’s amplitude, which decreased markedly. For example, the peak-to-peak amplitude was reduced by 59% by increasing the contrast (Fig. 4e) and by 71% by increasing the luminance (Fig. 4f), relative to the case of low luminance and low contrast (Fig. 4d). This decrease countered the increase in signal strength, reducing the dependence of the responses on mean luminance and contrast and adjusting the cell’s dynamic range to the prevalent stimulus conditions. The other effect of increasing mean luminance or contrast was on the impulse response’s time course, which became more transient. For example, the duration of the impulse response was reduced from 94 ms at low luminance and contrast (Fig. 4d) to 55 ms at the higher contrast (Fig. 4e) and to 48 ms at the higher luminance (Fig. 4f). This reduction modified the temporal frequency tuning of the responses, increasing the preferred temporal frequency as mean luminance and contrast increased (Fig. 4a–c). Our method for measuring impulse responses yielded robust measures, which were not affected by slow contrast-adaptation mechanisms6,29–32. The responses in the first 5 s, when frequency ramped up from 0.5 Hz to 40 Hz, were essentially identical to those in the subsequent 5 s, when frequency ramped back down (Supplementary Fig. 1). Moreover, responses commonly remained constant over 20 s of stimulation with a drifting grating of constant frequency (Fig. 3c).

frequency was substantially higher at higher contrast (Fig. 4b) or at higher luminance (Fig. 4c) than at lower luminance and contrast (Fig. 4a). To summarize the responses in each stimulus condition (that is, each fixed mean luminance and contrast), we fitted them with a descriptive model. In this model (Fig. 5a), stimulus luminance was filtered by a linear receptive field, whose output was rectified to yield positive firing rates. The impulse response (the temporal profile of the receptive field) was estimated independently for each stimulus condition (Fig. 4d–f). The descriptive model captured response amplitude and a 100 043.3.6 Time domain phase over the entire range of tested temporal frequencies (Fig. 4a–c and Supplementary Fig. 2 online). For over half of the cells in our population (N ¼ 40), it accounted 30 for more than 85% of the stimulusdriven variance in the responses (Fig. 5b).


[

10

100 ms

3 6

13 29 Luminance (cd m–2)

0.5 40 Frequency (Hz)

64

SVD

c

NUMBER 12

Frequency domain

FT

d

FT–1

100 ms

[

DECEMBER 2005

Spikes per cd s2

Figure 6 Independence of the effects of luminance gain control and contrast gain control. (a) Impulse responses of the cell in Figure 4, measured by fitting the descriptive model for all combinations of mean luminance and contrast. (b) The amplitude of the transfer functions corresponding to those impulse responses, as a function of frequency. (c) Impulse responses predicted by the separable model (yellow), compared to those predicted by the descriptive model (red, replotted from a for comparison). The latter are barely visible in the superposition, indicating that the predicted impulse responses are extremely similar. Each impulse response (yellow) is the convolution of the fixed filter (pink) with the luminance gain filter in the appropriate column (blue) and a contrast gain filter in the appropriate row (green). (d) The amplitude transfer functions corresponding to those impulse responses, as a function of frequency. The arrows between the panels indicate the sequence of operations: Fourier transform (FT, a to b), singular value decomposition (SVD, b to d), and inverse Fourier transform (FT–1, d to c).

b

Spikes per cd s2

Contrast (%)


V0

Noise

15

Gain

+

b

Gain

Descriptive model

a

Number of cells

ARTICLES

0.5 40 Frequency (Hz)

1693

ARTICLES

a

b

Gain (%)

100

30

c

d Integration time (ms)


10

75 70 65 60 55 10

100 30 Luminance (%)

10

30 100 Contrast (%)

Figure 7 Summary of the effects of luminance and contrast, and predictions of the separable model. Data points indicate the overall gain (top) and integration time (bottom) for the data of Figure 6. Lines show the predictions of the separable model. (a) Overall gain as a function of mean luminance, for different contrasts (black to white: 10% to 100%). (b) Gain replotted as a function of contrast, for different mean luminances (black to white: 10% to 84%). (c) Overall integration time as a function of luminance, for different contrasts (as in a). (d) Integration time as a function of contrast, for different luminances (as in b).

Thus, slow contrast adaptation mechanisms operating over time courses of seconds had little (if any) role in these responses. To test whether luminance gain control and contrast gain control operate independently of each other, we asked whether these responses could be explained by a separable model (Fig. 5c,d). This separable model is a special case of the descriptive model described above (Fig. 5a). In the model, the impulse response is described by a fixed filter followed by two variable filters: one for luminance gain control, which depends only on mean luminance, and one for contrast gain control, which depends only on contrast. This model makes a strong prediction: the effects on the impulse response of changing luminance should be the same for all contrasts, and the effects of changing contrast should be the same for all luminances. We tested this prediction on measurements made at a variety of luminances and contrasts (Fig. 6). To estimate the filters in the separable model, we started from the impulse responses measured with the descriptive model (Fig. 6a), and we applied a series of simple mathematical operations (Fig 6b–d; see also Supplementary Methods online). The first operation was the Fourier transform (Fig. 6b), which expressed each impulse response as a transfer function in the frequency domain; the advantage of this representation is that the transfer function of a series of filters (the three filters in the separable model, Fig. 5c) is the product of the transfer functions of the filters. According to the separable model, the matrix of transfer functions (Fig. 6b) should be separable: each transfer function should be the product of three transfer functions—a fixed one, one that depends only on luminance and one that depends only on contrast. The best estimates for the three transfer functions were obtained through singular value decomposition (Fig. 6d). This operation yielded (i) a fixed transfer function (pink), (ii) transfer functions that depended only on luminance (blue) and (iii) transfer functions that depended only on contrast (green). Finally, we performed an inverse Fourier transform to convert the transfer functions back into impulse responses in the time domain. The predicted impulse responses were very similar

1694

to the measured ones (Fig. 6c). For each combination of luminance and contrast, the model predicted that the impulse response was the convolution of the fixed filter with the corresponding snapshots of the filters for luminance gain control and contrast gain control (that is, the filters in the appropriate column and row). As expected, these filters become larger and slower as luminance or contrast is decreased (Fig. 6c). The separable model provided excellent fits to the data. First, it predicted impulse responses that were barely distinguishable from those estimated by the descriptive model (Fig. 6c). This was the case not only for the on-center, X-type cell (Fig. 6) but also for off-center cells and Y-type cells (Supplementary Fig. 3 online). Second, it predicted the firing rate responses almost as well as did the descriptive model (compare Fig. 5b and 5d). The percentage of stimulus-driven variance explained by the two models was comparable, with a median across cells of 81% for the separable model versus 85% for the descriptive model. In fact, we chose the example cell (shown in Fig. 6) because the quality of the fits was the same as the median values, 81% and 85%. This performance was notable, given that the separable model has many fewer degrees of freedom than the descriptive model. To predict the responses, the descriptive model requires 25 filters (one impulse response for each combination of mean luminance and contrast), whereas the separable model requires only 10 filters (the fixed filter plus the snapshots of the variable filters for five luminances and four contrasts; see Methods). Another way to gauge the quality of the separable model was to consider what would happen if the luminance and contrast gain mechanisms were instead matched to the statistics of 1/f noise, for which there is an inverse relationship between luminance and contrast (Fig. 2f). If this were the case, the full matrix of responses should be explained by only one gain control mechanism, which could operate based on luminance alone or contrast alone. We tested this hypothesis by trying to predict the full set of responses with a one-dimensional subset of impulse responses. We used the impulse responses estimated at combinations of luminance and contrast lying close to a line with the slope observed in a 1/f world (Fig. 2f). For each luminance, the impulse response nearest to the line was used to predict the responses obtained at all contrasts. This method yielded poor fits, explaining only 35% (median) of the stimulus-driven variance of the responses, implying that our analysis was sensitive enough to reject plausible alternatives to the independence assumption. Finally, an intuitive way to summarize the effects of gain control— and to gauge the performance of the separable model—is to consider overall measures of gain and integration time (Fig. 7). As a measure of overall gain, we took the average of the amplitude of the transfer function between 0.5 Hz and 15 Hz (at higher frequencies gain is barely affected by changes in luminance and contrast). This overall gain was plotted as a function of luminance (Fig. 7a). The slope of the curves was close to –1 in logarithmic axes, indicating that overall gain was inversely proportional to luminance1. On the other hand, overall gain decreased more modestly with contrast (Fig. 7b, slope of curves is shallower than –1). Indeed, at low temporal frequencies, LGN responses were largely independent of mean luminance (Fig. 3a,b), whereas they did grow with contrast (Fig 3c,d). As an overall measure of integration time, we took the slope of the best-fitting line relating the phase of the transfer function to frequency, weighted by the amplitude at each frequency33. As expected, integration time decreased with luminance (Fig. 7c) and with contrast (Fig. 7d). All of these effects were very well captured by the separable model (Fig. 7), confirming that the effects of luminance and contrast gain control are independent of each other.

VOLUME 8

[

NUMBER 12

[


ARTICLES


DISCUSSION We have demonstrated that luminance and contrast are largely independent in natural images, and that the mechanisms of luminance gain control and contrast gain control in the early visual system reflect this independence. Luminance and contrast in natural scenes The independence of luminance and contrast in natural images is not easily explained. According to the familiar view of image formation, the luminance that reaches the eye is the product of the reflectance of surfaces and an independent illuminant that changes slowly across the scene1. We simulated this layout and modeled the reflectance of a scene as a sample of 1/f noise and the illuminant as an independent sample of low-pass filtered noise (for example, 1/f n noise, n 4 1). In the resulting images, local luminance and local contrast invariably showed a strong negative correlation. The correlation was weaker than for a uniform illuminant, but still substantially more negative than for natural images; thus, independence of the illuminance and reflectance functions cannot be the whole explanation. A factor contributing to the statistical independence is that regions of a scene that have high luminance are likely to have more directional illumination, and hence higher-contrast shadows and shading. This factor contributes a positive correlation between luminance and contrast that partially cancels the expected negative correlation. Another possible factor is that the distribution of reflectance in natural environments might be skewed towards the higher values, instead of being symmetric (as in the case of 1/f noise). The analysis of natural images reported here does not take into account the longer-term changes in local luminance owing to the daynight cycle or to switching of environments (for example, moving from an open field to under a forest canopy). It seems likely that these longer term changes are also uncorrelated with local contrast. Nonetheless, our results are of most relevance to the more rapid components of luminance and contrast gain control that compensate for the changes in receptive field stimulation resulting from movements of eye, head, and body. Under these circumstances, we have shown that the visual system will mostly encounter changes in luminance of about one order of magnitude (but occasionally larger—see Fig. 2b), substantially less than the many orders of magnitude over which the eye is able to operate3. Independence of gain control mechanisms To study the effect of luminance and contrast on responses of the early visual system, we turned to simple stimuli, the drifting gratings that are classically used to study vision. Drifting gratings of optimal spatial frequency present many advantages for our purposes. First, drifting gratings are defined by few parameters, which explicitly include luminance and contrast. Second, drifting gratings drive neurons in the early visual system very well, so that we could test low luminances and low contrasts that would otherwise elicit very few spikes. Third, drifting gratings elicit, in the receptive field, responses that oscillate at the frequency of drift, so that it is straightforward to predict the responses of the receptive field and see how these responses are affected by gain control. In principle, it would have been desirable to use natural stimuli, but describing LGN responses to such stimuli requires a rather involved model that goes well beyond the simple temporal aspects discussed here. The independence of gain-control mechanisms for luminance and contrast that we have found may not seem surprising at first, given the common assumptions that the output of luminance gain control (i) removes all effects of mean luminance and (ii) is


[

NUMBER 12

[

DECEMBER 2005

completely independent of contrast. However, these assumptions are not accurate. First, the output of the luminance gain control mechanism does depend on mean luminance: it emphasizes the low temporal frequencies at low mean luminance and the high temporal frequencies at high mean luminance (Fig. 4). To interpret these responses and estimate stimulus contrast, a subsequent contrast gain-control stage would have to know the mean luminance. Second, the output of mechanisms performing luminance gain control does, to some extent, depend on contrast34: for a high-contrast stimulus, luminance gain can vary over time, because mean luminance is computed locally and rapidly1. One limitation of our study is that we attributed all the effects of contrast gain control to changes in the impulse response of the neurons. In principle, contrast gain control might also affect the spatial receptive field35 or the resting potential6,32. In practice, however, a model in which contrast gain control leaves constant both the spatial receptive field and the resting potential provides an excellent fit to the responses of LGN neurons to stimuli of different contrasts, sizes and spatial frequencies10. In ganglion cells, moreover, resting potential seems to be affected only very slightly by changes in contrast26. Similarly, we have not quantified the effects of luminance gain control on the spatial receptive field and on the resting potential. The surround of retinal ganglion cells becomes relatively weaker at lower luminances36,37, but this weakening takes place only at the lowest levels of luminance and is barely noticeable when one compares responses within a limited range of luminances, as we do here. Mean luminance might affect the resting potential, but this effect is small and variable, at least as gauged from the resting firing rate36. Indeed, in our models we have left the resting potential free to vary with mean luminance, but the resulting estimates did not vary by much and did not depend on luminance in an orderly fashion. Conclusions Our results add to the growing body of evidence for a close match between the statistical properties of natural scenes and the processing of contrast in visual systems. For example, there appears to be a close correspondence between the range of local contrasts in natural images and the dynamic range of single neuron responses in the eye11,12 and in the LGN13. Similarly, the statistics of natural images, together with the observation that signal strength, compared to noise strength, is smaller at low contrasts, can predict how the shape of the impulse response changes across contrasts14. There is also computational evidence that, for natural images, rapid local contrast adaptation enhances faint contours12 and increases statistical independence (reduces the redundancy) in the responses of orientation-selective and spatial frequency–selective neural populations in visual cortex15. Ultimately, however, a naturalistic explanation of the computational advantages of contrast gain control will have to account for the large differences seen across species and across the different retinogeniculate streams such as the M and P pathways in primates38. In summary, we have shown that there is little dependence between the local luminance and the local contrast in natural images and that this independence is elegantly reflected in the neural responses of the LGN. Our results provide direct support for the hypothesis1,2,16,17 that contrast is a fundamental independent variable encoded by the early visual system. They strengthen and validate a large body of neurophysiological, psychophysical and theoretical research that has implicitly assumed that luminance and contrast gain control are functionally independent, and provide a new example of how measuring the statistics of natural environments can provide insight into sensory systems.

1695

ARTICLES METHODS


Measurement of local luminance and contrast. From a publicly available image bank18, we selected 300 images (12-bit, gray scale) that did not include animals or artificial objects. For each image, we considered a sequence of eye movements by sampling from eye movement distributions measured separately19 (sampling from a uniform distribution gave similar results), with the constraint that no two samples could be closer than half the diameter of the patch. The local luminance of a patch was defined as L¼

N X

wi Li

i¼1

where N is the total number of pixels in the patch, Li is the luminance of the ith pixel and wi is the weight from a windowing function. The local contrast of a patch was defined as vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u N uX ðLi LÞ2 C¼t wi L2 i¼1 The weighting function was a circularly symmetric raised cosine qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2p ai ¼ cos ðxi xc Þ2 + ðyi yc Þ2 + 1 d where d is the patch diameter, (xi, yi) is the location of the ith pixel in the patch and (xc, yc) is the location of the center of the patch. The weights were P normalized to sum to 1: wi ¼ ai = N i ¼ 1 ai . The results described here are for a patch diameter of d ¼ 64 pixels, which corresponds to approximately 11 of visual angle18; however, we obtained very similar results (that is, small correlations) for a wide range of patch sizes. To determine the variation of local luminance and contrast within the typical single image (or constituent sub-image), we first computed the overall average luminance and contrast across all images (sub-images) and then rescaled each image (sub-image) so that its own average luminance and contrast would match the overall average. Correlations of luminance and contrast were measured on the individual images and then averaged across images. Physiological recordings. Methods for recording from single neurons in anesthetized cats have been described elsewhere39. The Animal Care and Use Committee of the Smith-Kettlewell Eye Research Institute approved all procedures. Extracellular signals were recorded with quartz-coated platinum-tungsten microelectrodes (Thomas Recording). Firing rates were obtained by convolving spike trains with a Gaussian window (s.d.: 5 ms). Stimuli were drifting gratings presented monocularly on a CRT-screen (refresh rate: 125 Hz). Gratings had optimal spatial frequency and position. For neurons that were strongly suppressed by large gratings, stimulus size was set to the optimal value, as measured with gratings at 50% contrast and 32 cd m–2; for the other neurons, the stimuli covered the entire receptive field. Combinations of mean luminance (4–6 values between 6 and 56 cd m–2) and contrast (3–5 values between 10% and 100%) were presented in a randomized order (12–25 stimuli repeated 6–12 times). The appearance of a grating was preceded by 2–2.5 s of uniform screen at the mean luminance of the stimulus. We classified neurons as on-center (26 cells) or off-center (14 cells) by mapping the receptive field with rapid sequences of flashed gratings. We classified neurons as X-type (34 cells) or Y-type (6 cells) based on standard criteria40. Models. For each combination of mean luminance L and contrast C, we estimated the impulse response fL,C(t) of a neuron by fitting a descriptive model (Fig. 5a). The first stage of the model is the convolution between the linear receptive field h(x,y,t) and the stimulus s(x,y,t), vdrive ðtÞ ¼ ½h sðx0 ; y0 ; tÞ; where vdrive(t) is the stimulus-driven membrane potential relative to rest, and (x0, y0) are coordinates of the receptive field center. The stimulus s(x,y,t) is the luminance

1696

distribution obtained by subtracting the mean from the luminance shown on the screen. The receptive field h(x,y,t) has center-surround organization: hðx; y; tÞ ¼ Gc ðx; yÞ f ðt Þ Gs ðx; yÞ f ðt dÞ; where Gc and Gs are Gaussian spatial profiles for center and surround, d is the delay between center and surround, and f(t) is a difference-of-gammas temporal impulse response (see Supplementary Methods). The latter is identical for center and surround. The parameters of Gc and Gs, and the delay d were fixed for a given neuron. The second and third stages of the model add Gaussian noise n(t) with fixed variance s2 to the visual response vdrive(t), and rectify the result to yield a firing rate r ðt Þ ¼ bv0 ðLÞ + vdrive ðtÞ + nðtÞc; where b c indicates rectification and v0 is the difference between the spiking threshold and the resting potential. The variance of the Gaussian noise was fixed for a given cell, whereas the resting potential was allowed to vary with mean luminance L to account for changes in spontaneous firing rate seen at different mean luminances. The sequence of fits used to estimate model parameters is described in Supplementary Methods. In the separable model (Fig. 5c), the impulse response fL,C is the convolution of three filters: fL;C ðtÞ ½f0 fL fC ðtÞ; where f0 is fixed, fL depends only on mean luminance (it describes the effects of luminance gain control) and fC depends only on contrast (it describes the effects of contrast gain control). We estimate the three filters from the impulse responses obtained with the descriptive model. This procedure is described in Supplementary Methods. Quality of predictions. To quantify how well the model predictions rt capture the measured responses st (both consisting of t ¼ 1, y, M samples) we estimated the fraction of stimulus-driven variance in the responses accounted for by the model41,42: b¼ where

* s2s

¼

s2s s2e s2s s2Z

1X 2 s M t t

+

is the power in the response, * s2e

¼

1X ðrt st Þ2 M t

+

is the mean square distance between data and model and "* # + d 1X 2 1 X 2 st st s2Z ¼ d1 M t M t is an estimate of the variance in st that is due to noise. Angular brackets indicate the average over d presentations of the same stimulus. The mean of the responses st was removed before these computations. The quantity b is an intuitive measure of fit quality, similar but superior to the commonly used ‘percentage of the variance’. A perfect model—that is, a model that is a perfect description of a system—could never account for 100% of the variance, because some of that variance is due to noise. By estimating the fraction of the variance that is due to the noise, the quantity sZ2, and subtracting it from the denominator in the definition of b, one eliminates this contribution. Rather reasonably, then, a perfect model is then one that yields b ¼ 1, or 100%. Values b 4 1 indicate overfitting, and b ¼ 0 as usual indicates that a constant value would have been better than the proposed model. The assumptions behind this estimation are minimal41: the noise should have zero mean and a non-infinite variance, and should be independent between trials. The estimator sZ2 is unbiased, and holds even if the variance or other property of the noise depends on the signal strength (as is common with neural signals). Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS We thank D. Ringach for suggesting the singular value decomposition method and J. Victor for helpful comments. The study of natural images was performed in W.S.G.’s laboratory, supported by National Eye Institute grant R01EY11747. The study of physiological responses was performed in M.C.’s laboratory,

VOLUME 8

[

NUMBER 12

[


ARTICLES supported by the James S. McDonnell Foundation 21st Century Research Award in Bridging Brain, Mind and Behavior. COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests.


Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/

1. Shapley, R.M. & Enroth-Cugell, C. Visual adaptation and retinal gain controls. in Progress in Retinal Research Vol. 3 (eds. Osborne, N. & Chader, G.) 263–346 (Pergamon, London, 1984). 2. Troy, J.B. & Enroth-Cugell, C. X and Y ganglion cells inform the cat’s brain about contrast in the retinal image. Exp. Brain Res. 93, 383–390 (1993). 3. Rodieck, R.W. The First Steps in Seeing (Sinauer, Sunderland, Massachussets, 1998). 4. Shapley, R.M. & Victor, J.D. The effect of contrast on the transfer properties of cat retinal ganglion cells. J. Physiol. (Lond.) 285, 275–298 (1978). 5. Victor, J. The dynamics of the cat retinal X cell centre. J. Physiol. (Lond.) 386, 219–246 (1987). 6. Baccus, S.A. & Meister, M. Fast and slow contrast adaptation in retinal circuitry. Neuron 36, 909–919 (2002). 7. Demb, J.B. Multiple mechanisms for contrast adaptation in the retina. Neuron 36, 781–783 (2002). 8. Kaplan, E., Purpura, K. & Shapley, R. Contrast affects the transmission of visual information through the mammalian lateral geniculate nucleus. J. Physiol. (Lond.) 391, 267–288 (1987). 9. Sclar, G., Maunsell, J.H. & Lennie, P. Coding of image contrast in central visual pathways of the macaque monkey. Vision Res. 30, 1–10 (1990). 10. Bonin, V., Mante, V. & Carandini, M. The suppressive field of neurons in lateral geniculate. J. Neurosci. (in the press). 11. Laughlin, S. A simple coding procedure enhances a neuron’s information capacity. Z. Naturforsch. [C] 36, 910–912 (1981). 12. Ruderman, D.L. The statistics of natural images. Network: Comput. Neural Syst. 5, 517–548 (1994). 13. Tadmor, Y. & Tolhurst, D.J. Calculating the contrasts that retinal ganglion cells and LGN neurones encounter in natural scenes. Vision Res. 40, 3145–3157 (2000). 14. Van Hateren, J.H. Spatiotemporal contrast sensitivity of early vision. Vision Res. 33, 257–267 (1993). 15. Schwartz, O. & Simoncelli, E.P. Natural signal statistics and sensory gain control. Nat. Neurosci. 4, 819–825 (2001). 16. Shapley, R. The importance of contrast for the activity of single neurons, the VEP and perception. Vision Res. 26, 45–61 (1986). 17. Shapley, R.M. & Man-Kit Lam, D. (eds.). Contrast Sensitivity (Bradford Books, 1993). 18. van Hateren, J.H. & van der Schaaf, A. Independent component filters of natural images compared with simple cells in primary visual cortex. Proc. Biol. Sci. 265, 359–366 (1998). 19. Najemnik, J. & Geisler, W.S. Optimal eye movement strategies in visual search. Nature 434, 387–391 (2005).


[

NUMBER 12

[

DECEMBER 2005

20. Oppenheim, A.V. & Lim, J.S. The importance of phase in signals. Proc. IEEE. 69, 529–541 (1981). 21. Field, D.J. Relations between the statistics of natural images and the response properties of cortical cells. J. Opt. Soc. Am. A 4, 2379–2394 (1987). 22. Saito, H. & Fukada, Y. Gain control mechanisms in X- and Y-type retinal ganglion cells of the cat. Vision Res. 26, 391–408 (1986). 23. Lankheet, M.J., Van Wezel, R.J., Prickaerts, J.H. & van de Grind, W.A. The dynamics of light adaptation in cat horizontal cell responses. Vision Res. 33, 1153–1171 (1993). 24. Yeh, T., Lee, B.B. & Kremers, J. The time course of adaptation in macaque retinal ganglion cells. Vision Res. 36, 913–931 (1996). 25. Lee, B.B., Dacey, D.M., Smith, V.C. & Pokorny, J. Dynamics of sensitivity regulation in primate outer retina: the horizontal cell network. J. Vis. 3, 513–526 (2003). 26. Zaghloul, K.A., Boahen, K. & Demb, J.B. Contrast adaptation in subthreshold and spiking responses of mammalian Y-type retinal ganglion cells. J. Neurosci. 25, 860–868 (2005). 27. Sherman, S.M. Tonic and burst firing: dual modes of thalamocortical relay. Trends Neurosci. 24, 122–126 (2001). 28. Tranchina, D., Gordon, J. & Shapley, R.M. Retinal light adaptation - evidence for a feedback mechanism. Nature 310, 314–316 (1984). 29. Smirnakis, S.M., Berry, M.J., Warland, D.K., Bialek, W. & Meister, M. Adaptation of retinal processing to image contrast and spatial scale. Nature 386, 69–73 (1997). 30. Chander, D. & Chichilnisky, E.J. Adaptation to temporal contrast in primate and salamander retina. J. Neurosci. 21, 9904–9916 (2001). 31. Brown, S.P. & Masland, R.H. Spatial scale and cellular substrate of contrast adaptation by retinal ganglion cells. Nat. Neurosci. 4, 44–51 (2001). 32. Solomon, S.G., Peirce, J.W., Dhruv, N.T. & Lennie, P. Profound contrast adaptation early in the visual pathway. Neuron 42, 155–162 (2004). 33. Reid, R.C., Victor, J.D. & Shapley, R.M. Broadband temporal stimuli decrease the integration time of neuron in cat striate cortex. Vis. Neurosci. 9, 39–45 (1992). 34. Lankheet, M.J., van Wezel, R.J. & van de Grind, W.A. Light adaptation and frequency transfer properties of cat horizontal cells. Vision Res. 31, 1129–1142 (1991). 35. Nolt, M.J., Kumbhani, R.D. & Palmer, L.A. Contrast-dependent spatial summation in the lateral geniculate nucleus and retina of the cat. J. Neurophysiol. 92, 1708–1717 (2004). 36. Derrington, A.M. & Lennie, P. The influence of temporal frequency and adaptation level on receptive field organization of retinal ganglion cells in cat. J. Physiol. (Lond.) 333, 343–366 (1982). 37. Troy, J.B., Bohnsack, D.L. & Diller, L.C. Spatial properties of the cat X-cell receptive field as a function of mean light level. Vis. Neurosci. 16, 1089–1104 (1999). 38. Lennie, P. Parallel visual pathways: a review. Vision Res. 20, 561–594 (1980). 39. Freeman, T.C., Durand, S., Kiper, D.C. & Carandini, M. Suppression without inhibition in visual cortex. Neuron 35, 759–771 (2002). 40. Hochstein, S. & Shapley, R.M. Quantitative analysis of retinal ganglion cell classifications. J. Physiol. (Lond.) 262, 237–264 (1976). 41. Sahani, M. & Linden, J.F. in Advances in Neural Information Processing Systems (eds. Becker, S., Thrun, S. & Obermayer, K.) 125–132 (MIT Press, Cambridge, Massachusetts, 2003). 42. Machens, C.K., Wehr, M.S. & Zador, A.M. Linearity of cortical receptive fields measured with natural sounds. J. Neurosci. 24, 1089–1100 (2004).

1697


ARTICLES


Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control Nathaniel D Daw1, Yael Niv1,2 & Peter Dayan1 A broad range of neural and behavioral data suggests that the brain contains multiple systems for behavioral choice, including one associated with prefrontal cortex and another with dorsolateral striatum. However, such a surfeit of control raises an additional choice problem: how to arbitrate between the systems when they disagree. Here, we consider dual-action choice systems from a normative perspective, using the computational theory of reinforcement learning. We identify a key trade-off pitting computational simplicity against the flexible and statistically efficient use of experience. The trade-off is realized in a competition between the dorsolateral striatal and prefrontal systems. We suggest a Bayesian principle of arbitration between them according to uncertainty, so each controller is deployed when it should be most accurate. This provides a unifying account of a wealth of experimental evidence about the factors favoring dominance by either system.

Diverse neural systems, notably prefrontal cortex, the striatum and their dopaminergic afferents, are thought to contribute to the selection of actions. Their differential and integrative roles are under active examination, and an important hypothesis is that subparts of these regions subserve two largely distinct and parallel routes to action. Such a division is the neurobiological scaffolding for an equivalent hypothesis about dual controllers that is prominent in psychological accounts of a range of behavioral phenomena in economic, social and animal-conditioning contexts1–5. The conventional idea is that the dorsolateral striatum and its dopaminergic afferents support habitual or reflexive control6, whereas prefrontal cortex is associated with more reflective or cognitive action planning7. (Following this convention, we will refer to the cognitive circuit as ‘prefrontal’, although it likely involves a number of additional regions, potentially including more medial striatal territories8.) This suggested dissociation is consistent with a range of electrophysiological9–11, functional magnetic resonance imaging (fMRI)12,13 and lesion studies14–17. The last are based on a clever behavioral approach to differentiating dual control strategies: namely, conditioning studies in which the values of rewards are unexpectedly changed. Outcome re-valuation affects the two styles of control differently and allows investigation of the characteristics of each controller, its neural substrates and the circumstances under which it dominates. Despite the wealth of evidence, there are few answers to two key normative questions: why should the brain use multiple action controllers, and how should action choice be determined when they disagree? For a framework for answers, we turn to reinforcement learning18, the computational theory of learned optimal action control. In reinforcement learning, candidate actions are assessed through

predictions of their values, defined in terms of the amount of reward they are expected eventually to bring about. Such predictions pose statistical and computational challenges when reward is contingent on performing a sequence of actions, and thus early action choices incur only deferred rewards. Approximations are essential in the face of these challenges; there are two major classes of reinforcement learning, which make different approximations, and so are differentially accurate in particular circumstances. One class involves ‘modelfree’ approaches such as temporal-difference learning, which underpin existing popular accounts of the activity of dopamine neurons and their (notably dorsolateral) striatal projections19,20. The other class involves ‘model-based’ methods18, which we identify with the second, prefrontal cortex system. We propose that the difference in the accuracy profiles of different reinforcement learning methods both justifies the plurality of control and underpins arbitration. To make the best decisions, the brain should rely on a controller of each class in circumstances in which its predictions tend to be most accurate. Here we suggest how the brain might estimate this accuracy for the purpose of arbitration by tracking the relative uncertainty of the predictions made by each controller. We show that this accounts for a range of factors shown in behavioral studies to favor either controller. To isolate our hypothesis, we develop the bulk of our account assuming strict separation between the systems; other aspects of their integration, particularly through learning21, are certainly also important. We interpret the two controllers as representing opposite extremes in a trade-off between the statistically efficient use of experience and computational tractability. Temporal-difference learning18 is a modelfree reinforcement learning method, which offers a compelling account of the activity of dopamine neurons in classical and instrumental

1Gatsby Computational Neuroscience Unit, University College London, Alexandra House, 17 Queen Square, London WC1N 3AR, UK. 2Interdisciplinary Center for Neural Computation, Hebrew University. P.O. Box 1255, Jerusalem 91904, Israel. Correspondence should be addressed to N.D.D. ([email protected]).

Received 15 April; accepted 12 September; published online 6 November 2005; doi:10.1038/nn1560

1704

VOLUME 8

[

NUMBER 12

[


ARTICLES

a

b

Tree System

Cache system

S0

S0 Initial state Press lever

Initial state

Enter magazine

Press lever Q=1

Enter magazine Q=0


S1 Food delivered S1

S2 Food delivered Press lever

S2

Press lever Q=0

No reward

Enter magazine

R=0

S3 No reward

Food obtained

R=0

R=1

learning tasks19,20. The foundation of this method is what we refer to as ‘caching’: namely, the association of an action or situation with a scalar summary of its long-run future value. A hallmark of this is the ubiquitous transfer of the dopaminergic response from rewards to the stimuli that predict them20. Working with cached values is computationally simple but comes at the cost of inflexibility: the values are divorced from the outcomes themselves and so do not immediately change with the re-valuation of the outcome. This is also the defining behavioral characteristic of habitual control. By contrast, we suggest that the prefrontal circuit subserves a modelbased reinforcement learning method. This constructs predictions of long-run outcomes, not through cached storage, but rather on the fly, by chaining together short-term predictions about the immediate consequences of each action in a sequence. Because this involves exploring a branching set of possible future situations, such methods are also known as ‘tree search’. Search in deep trees can be expensive in terms of memory and time and can also be error-prone. However, that the predictions are constructed on the fly allows them to react more nimbly to changed circumstances, as when outcomes are re-valued. This, in turn, is the behavioral hallmark of cognitive (or ‘goal-directed’) control. Here we develop these ideas in a formal, computational model and present simulation results that demonstrate the model’s ability to capture a body of animal conditioning data concerning the trade-off between controllers. Our results suggest that principles of sound, approximate, statistical reasoning may explain why organisms use multiple decision-making strategies and also provide a solution to the problem of arbitrating between them. RESULTS Post-training reinforcer devaluation We begin by discussing key experimental results suggesting the circumstances under which each controller dominates. Behavioral psychologists have investigated this issue extensively by post-training reinforcer devaluation (see a recent review5 for references). In a typical experiment, hungry rats are trained to perform a sequence of actions, usually a lever press followed by entry to a food magazine, to obtain a reward such as a food pellet. We formally depict this task (Fig. 1a) as a


[

NUMBER 12

[

DECEMBER 2005

Enter magazine Q=1

Figure 1 Task representations used by tree-search and caching reinforcement learning methods in a discrete-choice, discrete-trial representation of a standard instrumental conditioning task. (a) Structure of the task as represented by a tree-search controller. S0–S3 are the four possible states within the task; R ¼ {1, 0} represents whether or not reward was attained. (b) A caching reinforcement learning controller represents only the scalar expected future value (‘Q’) for each action in each state, divorced from the actual sequence and identity of future consequences.

tree of possible situations (states) that the subject can face in the task, the transitions No reward between those states engendered by the possible actions and the reward that is available Q=0 given an appropriate sequence of actions. Acquiring this arboreal representation of the S3 task from experience, and using it to choose Food obtained appropriate actions, are exactly the goals of Q=1 the tree-search controller. In the next phase of the experiment, the value of the food pellets is reduced, for instance by prefeeding the animal with them or by pairing them with illness to induce aversion. Then, animals are tested to see if they will continue to perform the actions previously associated with the newly devalued outcome. The test is performed without delivering outcomes (formally, in extinction) to prevent new learning about the value of the outcome during these trials. Outcome devaluation exploits a key distinction between tree search (Fig. 1a) and caching (Fig. 1b). Only tree search enumerates the specific consequences expected for some course of action, such as the identity of the food reward expected. The cached value of an action is, by its nature, independent of any such specific outcome information. Thus, if an animal acts based on a cached value, it will continue to do so even after the outcome has been devalued. In psychology, such outcome-insensitive behavior is known as ‘habitual’5,22. If, however, a behavior is determined by tree search, its propensity should be sharply reduced following devaluation. Psychologists term such outcome-sensitive behavior ‘goal-directed’5,22, as it changes when ‘goals’ are re-valued. Behavioral experiments (summarized in Fig. 2) demonstrate that, under different circumstances, animals show both profiles of devaluation sensitivity. Moderately trained lever presses are indeed sensitive to outcome devaluation (Fig. 2a, left)—suggesting control by a treesearch system. After extensive training, though, lever pressing becomes insensitive to devaluation (Fig. 2a, middle)—suggesting a transition to caching control23. Lesions or depletions of dopaminergic input to dorsolateral areas of the striatum evidently block this transfer of control to a caching system14,24. Such animals display relatively normal learning of the task, but despite over-training, their lever pressing is persistently sensitive to devaluation. This is consistent with choice relying only on an intact tree-search controller. The transition to caching with over-training is also tempered by two factors—the complexity of action choice and the proximity of the action to reward. In more complex tasks, in which an animal may, for instance, execute either of two different actions to obtain two different rewards, extensively trained actions remain sensitive to outcome devaluation (Fig. 2a, right), indicating a dominance of tree-search control25,26. Finally, though the evidence is perhaps less persuasive, actions closer to the reward are more sensitive to devaluation than S2

1705

ARTICLES

a

b

Distal action (lever press or chain pull)

Proximal action (magazine entry) 10

10


Actions per min

Actions per min

Non-devalued

5

Devalued 5

0

0 Moderate training

Extensive training, one action, one outcome

Extensive training, two actions, two outcomes

Moderate training

actions further away. For instance, when animals must press a lever and then enter a food magazine to obtain reward, the action more proximal to reward—magazine entry—remains devaluation-sensitive even after extensive training17 (Fig. 2b, right). In the experiment depicted here, this effect was significant only when collapsed over multiple lesion groups (which did not differ significantly among themselves), and there are only few other published reports of over-trained magazine behavior. However, actions more proximal to reward are more readily sensitive to devaluation in ‘incentive learning’ studies27, and extensively trained magazine responses remain devaluation-sensitive in a Pavlovian task without lever pressing28. The counterpart to lesions affecting the caching system is that lesions to a wide network of structures—including prelimbic cortex (a subarea of rat prefrontal cortex)15–17, prefrontal-associated regions of dorsomedial striatum8, basolateral amygdala29, gustatory insular cortex30 and, in a monkey study, orbitofrontal cortex31—seem to interfere with tree-search control. That is, they eliminate devaluation sensitivity even for moderately trained behaviors.

Extensive training

Figure 2 Behavioral results from reward devaluation experiments in rats. Actions per minute in an extinction test after devaluation of the outcome (black) or without devaluation (white). (a) Actions distal from the outcome (lever pressing and chain pulling) after moderate or extensive training and with one or two actions and outcomes, adapted from ref. 26, experiment 2. (b) Magazine entries (more proximal to the outcome), adapted from ref. 17. Data and error bars reproduced here are for a control group; differences were significant when collapsed with two additional lesion groups. Error bars: s.e.m.

chosen simply by comparing their values. The collection of values is called a ‘state-action value function’ or, for simplicity, a value function. The two classes of reinforcement learning methods can produce different, and differentially accurate, estimates of the value function. As in other cases of evidence reconciliation in neuroscience, such as multisensory integration32, we suggest that arbitration between values is based on the uncertainty or expected inaccuracy of each. Uncertainty quantifies ignorance about the true values (for example, about the probabilities of different payoffs); it should be distinguished from risk (which generically arises when payoffs are stochastic, but their probabilities may be known). For simplicity, we assume that the estimated value of each action is taken to be that derived from the controller that is more certain about the value. (Though most reliable, this estimate is not necessarily the largest.) The probability of choosing an action for execution is then proportional to this value. In addition to controlling estimation, uncertainty about an action’s value might, in principle, influence choice directly, as by promoting exploration to seek undiscovered rewards. In general, both controllers are uncertain about the values because they begin ignorant and have only a limited amount of noisy experience. Even given infinite experience, uncertainty persists, due to the possibility that the task itself (and hence the long-term values) can

Theory sketch The lesion studies indicate that each controller can substitute for the other even under circumstances when it would not normally dominate. This suggests a theory combining separate and parallel reinforcement learners (the implementation is detailed in Methods and Supplementary Methods online). S0 As in previous applications of reinforceInitial state ment learning to neural and behavioral data20, Press Pull Enter we work with a stylized description of the lever chain magazine experimental tasks (Figs. 1a and 3). This allows us to expose a unifying, normative interpretation of the pattern of experimental results discussed above, rather than focusing on a quantitative fit to rather qualitative data. S2 S3 S1 Food A delivered Food B delivered No reward Here, the goal of optimal control is to choose actions that maximize the probability of the Press Pull Enter Press Pull Enter lever chain magazine lever chain magazine ultimate receipt of a valued outcome (although it would be straightforward to include additional factors into the optimization, such as risk-sensitivity in the case of stochastic rewards). Optimization can be S3 S4 S3 S5 accomplished by calculating or learning the No reward Food A obtained No reward Food B obtained value of taking each action at each state, defined in terms of the probability that reward will later be received when starting from that action in that state. Given such Figure 3 Stylized tree representation of an instrumental conditioning task with two actions (a lever press information, advantageous actions can be and a chain pull) for two rewards. Note the additional states compared with Figure 1a.

1706

VOLUME 8

[

NUMBER 12

[


ARTICLES

S2

S3

Figure 4 Tree estimation at two stages of learning by the tree-search system on the task of Figure 1a. States are as in that figure; arrows represent the state transitions expected following a lever press or a magazine entry, with the width of each proportional to the estimated probability of the transition (averaged over 250 runs; s.e.m. error bars negligible). Before learning, all transitions were equally likely (data not shown).

change unexpectedly. We quantify uncertainty using approximate Bayesian versions of each reinforcement learning algorithm33,34. The differing methods of value estimation of the two systems give rise to differing uncertainty profiles. A (prefrontal) tree-search system uses experience with the task to estimate the nature of the state transitions and rewards (essentially, reconstructing the ‘trees’ of Figs. 1a and 3). Long-term reward probabilities are estimated by iteratively searching through this tree; uncertainty about which tree describes the task makes the value estimates uncertain. Furthermore, such tree search is computationally demanding in realistic (wide or deep) trees. Thus, in practice, approximations must be introduced at each iteration, such as ‘pruning’ or exploring only a subset of paths. We model the resulting inaccuracy or ‘computational noise’ in derived value estimates as an additional source of uncertainty that accumulates with each search step. A (dorsolateral striatal) caching system such as temporal-difference learning18 estimates the long-term values directly from experience, without explicitly constructing a tree. This relies on a different approximation: ‘bootstrapping’, or using the value estimates cached for subsequently encountered states as stand-ins for the actual longterm values at predecessor states. Initial ignorance and continual change make these values potentially inaccurate, and thus cause uncertainty. By contrast, though, the cached values make calculation straightforward, so there is little computational ‘noise’ associated with its output. To summarize, both the value predictions and the estimated uncertainties will differ between the tree-search and caching systems. Our account of action choice is based on an analysis of these two sets of quantities. Figure 5 Simulation of the dual-controller reinforcement learning model in the task of Figure 1a. (a) Distal action (lever press); (b) Proximal action (magazine entry). The topmost graphs show uncertainties (posterior variances) in the value estimates for different actions according to the cache (blue line) and tree (gold line), as a function of the number of rewarded training trials. The middle graphs show the value estimates themselves (posterior means); diamonds indicate the value estimates that would result after reward devaluation at various stages of training. Beneath the graphs are bar plots comparing the probability of choosing actions before and after their consequences were devalued, normalized to the non-devalued level. Bar color denotes which system (cache: blue; tree: gold) controlled the action in a majority of the 250 runs. All data reported are means over 250 runs; error bars (s.e.m.) are negligible.


[

NUMBER 12

[

DECEMBER 2005

a

b

Distal action (lever press)

Proximal action (magazine entry) 0.04

0.04 Cache Tree

0.02

Uncertainty

S3

S2

0.02

0

0

1

1

0.5

0

Non-dev. Dev.

Expected value

S1

1

1

0.5

0.5

Non- Dev. dev.

Nondev.

0.5

0

50 75 100 25 Rewarded trials Response rate relative to non-dev.

S2

Uncertainty


S1

S0

Simulations We simulated the two-controller reinforcement learning model on the action choice tasks, beginning with the task with one lever press for one outcome (Fig. 1a). The quantitative results conformed with the qualitative expectations adduced above. The prefrontal tree system learned, over experience, the structure of action-induced state transitions in the task, assigning high probability to the actual transitions (Fig. 4; the system additionally tracked uncertainty about its estimates of the transition probabilities, which is not illustrated). We studied each system’s action values (here, posterior means), along with its uncertainty (posterior variances) about those values, as functions of the amount of training and the position of the action in the behavioral sequence relative to the reward (Fig. 5a,b). Each system’s prior ignorance gradually resolved with experience. In all simulations, model-based reinforcement learning was more confident early in training, even though both systems had matched initial uncertainty. This is because under prefrontal tree search, any particular morsel of experience immediately propagates to influence the estimates of action values at all states; the effect of bootstrapping in dorsolateral striatal temporal-difference learning is to delay such propagation, making the system less data-efficient. Because the systems incorporate the expectation that actions’ values may change, past observations gradually become less relevant to present value estimates. This effective time horizon on the data implies that the uncertainties asymptote at finite levels for both systems. For the same reason, the value predictions can asymptote well short of the true payoffs. This asymptotic uncertainty has a harsher effect on the datainefficient cache. Thus, for the action proximal to reward (the magazine response), enhanced data efficiency allowed the tree-search system to be more certain, even asymptotically (Fig. 5b). However, an extra iteration of tree search was required to evaluate the action more distal from reward (the lever press), incurring additional uncertainty asymptotically due to the assumption of computational noise outlined above. The effect of this, asymptotically (Fig. 5a), was to favor the cache system, which suffers no such computational noise because it recalls values rather than computing them. We saw different results for a version of the task with two actions for two outcomes (Fig. 6a,b). Here, the agent’s experience was spread between more states and actions. Given the expectation of task change and the resulting temporal horizon on past experience, fewer relevant

Expected value

S0

Press lever Enter magazine

S2

After 50 rewards:

Response rate relative to non-dev.

After 1 reward:

Dev.

50 75 100 25 Rewarded trials

1

1

0.5

0.5

Non- Dev. dev.

Nondev.

Dev.

1707

ARTICLES

a

b

Distal action (lever press)

0.04

0.04

Proximal action (magazine entry)

0.02

0

0

1

1

0.5 Non-dev. Dev.

0

25

50

75

Expected value

Expected value

0.02

Uncertainty

Tree

0.5

0

100

1

1

0.5

0.5

Non- Dev. dev.


Rewarded trials



Uncertainty

Cache

Non- Dev. dev.

50 75 100 25 Rewarded trials

1

1

0.5

0.5

Non- Dev. dev.

Non- Dev. dev.

data were available to constrain any particular action value. The effect was asymptotically to preserve the tree system’s uncertainty advantage from the early training, low-data situation, even for the distal lever press (Fig. 6a). Whenever the tree system dominated, the overall system’s action choices were sensitive to outcome devaluation, whereas when the caching system dominated, they were not (Figs. 5 and 6, bar plots). This is because the underlying value predictions were sensitive to devaluation only in the tree system. The simulations, then, reproduced and explained the pattern seen in the behavioral experiments: overtraining promoted devaluation insensitivity, unless opposed by the countervailing effects of proximity to reward or task complexity. The results support the underlying hypothesis that the brain appropriately deploys each controller under those circumstances in which it is expected to be most accurate. DISCUSSION Our account builds on and extends existing ideas in several key ways. In contrast to the somewhat descriptive animal learning theories that are its foundation4,5,22, we have adopted a normative view, unifying the body of results on controller competition by appealing to uncertainty. This stance also contrasts with accounts of human behavioral data1,3: notably, ideas in economics2 suggesting that irrational, impulsive or emotional limbic influences (in our terms, the caching system) interfere with a more rational prefrontal controller. Under our account, both controllers are pursuing identical rational ends; in appropriate circumstances, the caching controller can more effectively accomplish the same functions as the prefrontal controller. Among reinforcement learning theories, there are various precedents for the idea of combining several controllers, including multiple caching controllers35–37 and also (partly cerebellar) model-based and model-free controllers38. However, normative, competitive interaction has not hitherto been investigated. Most other reinforcement learning theories that contemplate model-based control either completely replace caching with search39,40, or envision a hybrid blending features of both41,42. Such theories founder on lesion results indicating

1708

Figure 6 Simulation of the dual-controller reinforcement learning model in the task of Figure 3, in which two different actions produced two different rewards. One of the rewards was devalued in probe trials. (a) Distal action (lever press). (b) Proximal action (magazine entry). The same conventions and results are shown as in Figure 5, except that data reported are means over 1,000 runs; error bars (s.e.m.) are again negligible.

a dissociation between the neural substrates for tree-like and cache-like choice8,14–17,24. Of course, normativity only extends so far for us. The true reason for multiple controllers in our theory is the computational intractability of the complete Bayesian solution (roughly speaking, the tree-search system unencumbered by computational incapacity) and the resulting need for approximations. The cache system is an extreme example of an approximation that embraces potential inaccuracy to gain computational simplicity. Neural substrates We built on the classic idea that habitual control is associated with dopamine and dorsolateral striatum, and more cognitive search with prefrontal cortex. Because behavioral and lesion studies suggest these controllers can operate independently, for the purpose of modeling we made the simplifying approximation that they are strictly separate. However, their neural substrates are clearly intertwined—prefrontal cortex is itself dopaminergically innervated, and cortex and striatum are famously interconnected in ‘loops’43, including one that joins prefrontal areas with dorsomedial subregions of striatum. Indeed, recent pharmacological and lesion results implicate those prefrontalassociated striatal areas in tree-search control8. Competition between model-based and model-free control might, therefore, best be viewed as between dorsomedial and dorsolateral corticostriatal loops, rather than between cortex and striatum per se, a view that extends previous ideas about multiple caching controllers coexisting in different loops35,36. Although dopamine is hypothesized to support learning in the caching system, the role of dopamine in the tree-search controller remains wholly unresolved. Computational considerations also suggest that the systems should interact. Owing to computational costs in tasks involving deep trees, it is commonplace in reinforcement learning to search partway along some paths, then use cached values to substitute for unexplored subtrees18. Uncertainties can be compared at each stage to decide whether to expand the tree or to fall back on the cache44, trading off the likely costs (for example, time or calories) of additional search against its expected benefits (more accurate valuations allowing better reward harvesting). The essentials of our account would be preserved in a model incorporating such partial evaluation, and the resulting improvement in the tree system’s valuations due to learning in the cache system echoes other suggestions that learning in the basal ganglia might train or inform cortex11,21. There is limited evidence about the substrate for the uncertaintybased arbitration that has been our key focus. First, along with direct, population-code representations of uncertainty45, cholinergic and noradrenergic neuromodulation have often been implicated46. Second, two candidates for arbitration are the infralimbic cortex (IL; part of the prefrontal cortex) and the anterior cingulate cortex (ACC). Lesions to the IL reinstate tree-search from previously caching control16,17; however, because this area is not classically part of the habitual system, it has been suggested that it might support controller competition17. The involvement of the ACC in the competition-related functions of monitoring and resolving response error and conflict has been suggested in experimental and theoretical studies47,48.

VOLUME 8

[

NUMBER 12

[


ARTICLES


Complementary evidence about dual control arises from spatial tasks in both humans and animals37,49. Navigational decisions can arise from a flexible ‘cognitive map’ that supports latent learning and is associated with the hippocampus; with practice, they become habitized and evidently under dorsal striatal control. Experimental considerations One route to test our framework is neuronal recordings. We expect activity in areas associated with each controller to reflect its decision preferences, even when (as a result of arbitration) the other is actually directing behavior. Behavior should thus be better correlated with activity in whichever system is producing it. By manipulating factors such as the amount of training or the proximity of response to reward, it should be possible to transfer control between the systems and thereby to switch the behavioral-neural correlations. Researchers have recently recorded from striatum and prefrontal cortex (interpreted as parts of the cache and tree systems, respectively) in monkeys over-trained on an associative learning task with reversals11. Various features of this task could promote the dominance of either system—extreme over-training and task simplicity favor cache control, but action-reward proximity and frequent reversals promote tree search. The neural recordings are also inconclusive. A direct interpretation supports striatal control: neurons there are more strongly selective for the animal’s choices, earlier in trials, and more robustly after reversals. However, an alternative interpretation instead supports prefrontal dominance, because change in the prefrontal representation correlates with behavioral re-learning following reversal. A devaluation challenge or recordings under different task circumstances (over-training levels, etc.) could help to distinguish these possibilities. Because, in these recordings, representational changes occur faster in striatum, the authors suggest11 that relearning the correct responses following reversal might be more rapid in the striatum, and that this knowledge subsequently transfers to cortex21. This contrasts with some habitization models in which learning progresses in the other order, though our theory makes no specific claim about the relative ‘learning rates’ of the two systems. In any case, subsequent analysis of error trials shows that the striatal firing reflects the animal’s actual (and not the correct) choices (A. Pasupathy & E.K. Miller, Comput. Syst. Neurosci. Abstr., p. 38, 2005). Finally, because the striatal region recorded (caudate) includes areas likely corresponding to dorsomedial striatum in rats, it may be that this area too is part of the tree system and not the cache8, in which case properly interpreting the results will require a more finely fractionated understanding of the neural organization of tree search. Our theory provides additional testable factors likely to influence the trade-off between systems. Computational pressures might be increased, and tree search discouraged, in tasks that pose more strenuous cognitive demands (for example, delayed match to sample; such a strategy has been used with humans2 but not in animal devaluation studies). Introducing unexpected changes in task contingencies should also favor the data-efficient tree system, because relevant data thereby become more scarce. Further, although task complexity favors goal-directed control, the details of task structure may have subtler effects. It has long been known in reinforcement learning that caching is relatively advantageous in tasks with a fan-out structure (in which a state might be followed randomly by any of several others); conversely, tasks with linear or fan-in structure (several states leading to one) should favor search. Finally, our theory is applicable to several other phenomena in animal behavior. Stimuli can signal reinforcement that is available irrespective of the animal’s actions, and these ‘Pavlovian’ associations can affect


[

NUMBER 12

[

DECEMBER 2005

behavior. Such stimulus-reward predictions might originate from both cache and tree systems, with rich interactions and consequences. In ‘conditioned reinforcement’, animals learn to work to receive a stimulus that had previously been paired with reinforcement. Such learning might occur in either of our reinforcement learning systems. However, it is also a plausible test case for their potential interaction through partial evaluation, as the tree system might explore the consequences of the (new) response but defer to the cache’s evaluation of the (familiar) subsequent stimulus. Animals can acquire a new conditioned response even for a stimulus whose associated reinforcer had been devalued50, suggesting at least the value of the stimulus was cached. The hypothesized involvement of both systems might be investigated with lesions disabling each. Our theory also casts the phenomenon of ‘incentive learning’27 in a new light. In this, for some actions to be sensitive to outcome devaluation, the animal must previously have experienced the reinforcer in the devalued state. The predominant account of incentive learning5 holds that such experience is necessary for the goal-directed system (our tree) to learn about the new value of the reinforcer. We suggest instead that experience with the outcome decreases the tree system’s uncertainties (by confirming existing knowledge about the outcome’s value). This tends to promote its dominance over the cache, explaining interactions between outcome exposure and other factors such as over-training and reward proximity27. Because outcome exposure allows the tree system to overcome caching control, our theory makes the strong prediction (contrary to the standard account) that the need for such experience should vanish in animals with lesions disabling the caching system. METHODS Background. For simplicity, we modeled conditioning tasks using absorbing Markov decision processes (MDPs)18 (Figs. 1a and 3)—ones in which experience is structured as a set of trials, with a set of terminal states at which an episode can end. We assumed that outcomes were delivered only (if at all) in terminal states and identified particular terminal states with particular outcomes (for instance, different foods). Key to our account are two complicating factors. First, the agent started without knowing the exact MDP, which, furthermore, could change over time. These were the major sources of uncertainty. Second, although MDPs traditionally treat rewards with static, scalar utilities, here devaluation treatments explicitly changed some outcomes’ utilities. For convenience, we assumed that rewards were binary (0 or 1) and used the probability that the reward was 1 in a particular terminal state as a surrogate for the associated reward’s utility. Choice in both cache and tree systems depended on scalar values— predictions of the future utility of executing a particular action at a particular state. If an outcome was devalued, both could learn by experiencing it that its corresponding state had lower utility. However, only the tree system used that information to guide subsequent action choice at distal states, as it derived action values by considering what future states would result. The cache system’s values were stored scalars and were thus insensitive even to known changes in outcome value, absent new experience of the action actually producing the outcome. Fully optimal choice in unknown MDPs is radically computationally intractable. Tree and cache reinforcement learning methods therefore each rely on approximations, and we tracked uncertainties about the values produced by such systems to determine for what circumstances each method is best suited. Formal model. An absorbing MDP comprises sets S of states and A of actions, a ‘transition function’ T(s, a, s¢) P(s(t + 1) ¼ s¢ | s(t) ¼ s, a(t) ¼ a) specifying the probability that state s¢ A S will follow state s A S given action a A A, and (in our version) a ‘reward function’ R(s) P(reward(t) ¼ 1 | s(t) ¼ s) specifying the probability that reward is received in terminal state s. Here, the state-action value function Q(s, a) is the expected probability that reward will ultimately be received, given that the agent takes action a in state s

1709

ARTICLES and chooses optimally thereafter. The formal definition is recursive: ( RðsÞ s is terminal ða ¼ Þ P Qðs; aÞ Tðs; a; s0 Þ max½Qðs0 ; a0 Þ otherwise


s0

a0

Standard reinforcement learning methods18 do not track uncertainty in their estimates of Q. We consider Bayesian variations33,34, which estimate not simply the expected value Q(s, a) but a posterior distribution Qs,a(q) P(Q(s, a) ¼ q | data) that measures, for any 0 r q r 1, how likely it is that the true optimal probability of future reward (compounded over different paths through the states) equals q, given the evidence, ‘data’, about transitions and outcomes so far observed. A Bayesian tree-search (‘value iteration’) system34 uses experience to estimate a posterior distribution over the MDP (functions T and R) and explores it to derive distributions over Q(s, a) (Supplementary Fig. 1 online). A Bayesian caching (‘Q-learning’) system33 instead stores a distribution over Q(s, a) for each action and state and updates it for consistency with the stored value distributions of subsequently encountered states (Supplementary Fig. 2 online). Full equations appear in Supplementary Methods. If, for a particular controller, state and action, the distribution Qs,a is sharply peaked at some q, then the controller is fairly certain of the value; if it is instead spread out over a range of possible q’s, then the controller cannot identify the value with certainty. We thus arbitrated between the controllers’ estimates on the basis of their variance (mean squared error, ‘uncertainty’): given districache butions Qtree D Qs;aE from the cache, we took the winning s;a from the tree and ifD the variance of Qtree value Q(s, a) to be the mean Qtree E s;a s;a was smaller than cache cache otherwise. (Softer integration the variance of Qs;a , and the mean Qs;a schemes, such as a certainty-weighted average, are also possible.) Given winning estimates Q(s, a) for each action available in the current state, we chose an action stochastically using softmax probabilities, P(a(t) ¼ a | s(t) ¼ s) p ebQ(s,a) where the parameter b controlled the tendency of the system to choose exclusively the action deemed best. Experimentally, the effect of devaluation can be assessed either within or between animals (by comparing to another action or group for which the outcome was not devalued). In our simulations, we compared the probabilities of choosing the same action a in the relevant state s, with or without devaluation (similar to the between-group approach); softmax action selection ensured that a reduction in Q(s, a) for an action will reduce the probability that the action is chosen. Note that posterior uncertainty quantifies ignorance about the true probability of reward, not inherent stochasticity in reward delivery. For instance, reward may follow from some state randomly with 50% probability—but if the controller can precisely identify that the true probability is 50% rather than some other number, the value is not uncertain. Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS We are grateful to B. Balleine, A. Courville, A. Dickinson, P. Holland, D. Joel, S. McClure and M. Sahani for discussions. The authors are supported by the Gatsby Foundation, the EU Bayesian Inspired Brain and Artefacts (BIBA) project (P.D., N.D.), a Royal Society USA Research Fellowship (N.D.) and a Dan David Fellowship (Y.N.). COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/ 1. Kahneman, D. & Frederick, S. Representativeness revisited: attribute substitution in intuitive judgment. in Heuristics and Biases: the Psychology of Intuitive Judgment (eds. T. Gilovich, D.G. & Kahneman, D.) 49–81 (Cambridge University Press, New York, 2002). 2. Loewenstein, G. & O’Donoghue, T. Animal spirits: affective and deliberative processes in economic behavior. Working Paper 04–14, Center for Analytic Economics, Cornell University (2004). 3. Lieberman, M.D. Reflective and reflexive judgment processes: a social cognitive neuroscience approach. in Social Judgments: Implicit and Explicit Processes (eds. Forgas, J., Williams, K. & von Hippel, W.) 44–67 (Cambridge University Press, New York, 2003).

1710

4. Killcross, S. & Blundell, P. Associative representations of emotionally significant outcomes. in Emotional Cognition: from Brain to Behaviour (eds. Moore, S. & Oaksford, M.) 35–73 (John Benjamins, Amsterdam, 2002). 5. Dickinson, A. & Balleine, B. The role of learning in motivation. in Stevens’ Handbook of Experimental Psychology Vol. 3: Learning, Motivation and Emotion 3rd edn. (ed. Gallistel, C.R.) 497–533 (Wiley, New York, 2002). 6. Packard, M.G. & Knowlton, B.J. Learning and memory functions of the basal ganglia. Annu. Rev. Neurosci. 25, 563–593 (2002). 7. Owen, A.M. Cognitive planning in humans: neuropsychological, neuroanatomical and neuropharmacological perspectives. Prog. Neurobiol. 53, 431–450 (1997). 8. Yin, H.H., Ostlund, S.B., Knowlton, B.J. & Balleine, B.W. The role of the dorsomedial striatum in instrumental conditioning. Eur. J. Neurosci. 22, 513–523 (2005). 9. Jog, M.S., Kubota, Y., Connolly, C.I., Hillegaart, V. & Graybiel, A.M. Building neural representations of habits. Science 286, 1745–1749 (1999). 10. Holland, P.C. & Gallagher, M. Amygdala-frontal interactions and reward expectancy. Curr. Opin. Neurobiol. 14, 148–155 (2004). 11. Pasupathy, A. & Miller, E.K. Different time courses of learning-related activity in the prefrontal cortex and striatum. Nature 433, 873–876 (2005). 12. McClure, S.M., Laibson, D.I., Loewenstein, G. & Cohen, J.D. Separate neural systems value immediate and delayed monetary rewards. Science 306, 503–507 (2004). 13. O’Doherty, J. et al. Dissociable roles of ventral and dorsal striatum in instrumental conditioning. Science 304, 452–454 (2004). 14. Yin, H.H., Knowlton, B.J. & Balleine, B.W. Lesions of dorsolateral striatum preserve outcome expectancy but disrupt habit formation in instrumental learning. Eur. J. Neurosci. 19, 181–189 (2004). 15. Balleine, B.W. & Dickinson, A. Goal-directed instrumental action: contingency and incentive learning and their cortical substrates. Neuropharmacology 37, 407–419 (1998). 16. Coutureau, E. & Killcross, S. Inactivation of the infralimbic prefrontal cortex reinstates goal-directed responding in overtrained rats. Behav. Brain Res. 146, 167–174 (2003). 17. Killcross, S. & Coutureau, E. Coordination of actions and habits in the medial prefrontal cortex of rats. Cereb. Cortex 13, 400–408 (2003). 18. Sutton, R.S. & Barto, A.G. Reinforcement Learning: an Introduction (MIT Press, Cambridge, Massachusetts, 1998). 19. Houk, J.C., Adams, J.L. & Barto, A.G. A model of how the basal ganglia generate and use neural signals that predict reinforcement. in Models of Information Processing in the Basal Ganglia (eds. Houk, J.C., Davis, J.L. & Beiser, D.G.) 249–270 (MIT Press, Cambridge, Massachusetts, 1995). 20. Schultz, W., Dayan, P. & Montague, P.R. A neural substrate of prediction and reward. Science 275, 1593–1599 (1997). 21. Houk, J.C. & Wise, S.P. Distributed modular architectures linking basal ganglia, cerebellum, and cerebral cortex: their role in planning and controlling action. Cereb. Cortex 5, 95–110 (1995). 22. Dickinson, A. Actions and habits—the development of behavioural autonomy. Phil. Trans. R. Soc. Lond. B 308, 67–78 (1985). 23. Adams, C.D. Variations in the sensitivity of instrumental responding to reinforcer devaluation. Q. J. Exp. Psychol. 34B, 77–98 (1982). 24. Faure, A., Haberland, U., Conde´, F. & Massioui, N.E. Lesion to the nigrostriatal dopamine system disrupts stimulus-response habit formation. J. Neurosci. 25, 2771–2780 (2005). 25. Colwill, R.M. & Rescorla, R.A. Instrumental responding remains sensitive to reinforcer devaluation after extensive training. J. Exp. Psychol. Anim. Behav. Process. 11, 520–536 (1985). 26. Holland, P.C. Relations between Pavlovian-instrumental transfer and reinforcer devaluation. J. Exp. Psychol. Anim. Behav. Process. 30, 104–117 (2004). 27. Balleine, B.W., Garner, C., Gonzalez, F. & Dickinson, A. Motivational control of heterogeneous instrumental chains. J. Exp. Psychol. Anim. Behav. Process. 21, 203–217 (1995). 28. Holland, P. Amount of training affects associatively-activated event representation. Neuropharmacology 37, 461–469 (1998). 29. Blundell, P., Hall, G. & Killcross, S. Preserved sensitivity to outcome value after lesions of the basolateral amygdala. J. Neurosci. 23, 7702–7709 (2003). 30. Balleine, B.W. & Dickinson, A. The effect of lesions of the insular cortex on instrumental conditioning: evidence for a role in incentive memory. J. Neurosci. 20, 8954–8964 (2000). 31. Izquierdo, A., Suda, R.K. & Murray, E.A. Bilateral orbital prefrontal cortex lesions in rhesus monkeys disrupt choices guided by both reward value and reward contingency. J. Neurosci. 24, 7540–7548 (2004). 32. Deneve, S. & Pouget, A. Bayesian multisensory integration and cross-modal spatial links. J. Physiol. (Paris) 98, 249–258 (2004). 33. Dearden, R., Friedman, N. & Russell, S.J. Bayesian Q-learning. in Proceedings of the 15th National Conference on Artificial Intelligence (AAAI) 761–768 (1998). 34. Mannor, S., Simester, D., Sun, P. & Tsitsiklis, J.N. Bias and variance in value function estimation. in Proceedings of the 21st International Conference on Machine Learning (ICML) 568–575 (2004). 35. Nakahara, H., Doya, K. & Hikosaka, O. Parallel cortico-basal ganglia mechanisms for acquisition and execution of visuomotor sequences - a computational approach. J. Cogn. Neurosci. 13, 626–647 (2001). 36. Tanaka, S.C. et al. Prediction of immediate and future rewards differentially recruits cortico-basal ganglia loops. Nat. Neurosci. 7, 887–893 (2004).

VOLUME 8

[

NUMBER 12

[



ARTICLES 37. Chavarriaga, R., Strosslin, T., Sheynikhovich, D. & Gerstner, W. A computational model of parallel navigation systems in rodents. Neuroinformatics 3, 223–242 (2005). 38. Doya, K. What are the computations in the cerebellum, the basal ganglia, and the cerebral cortex. Neural Netw. 12, 961–974 (1999). 39. Suri, R.E. Anticipatory responses of dopamine neurons and cortical neurons reproduced by internal model. Exp. Brain Res. 140, 234–240 (2001). 40. Smith, A.J., Becker, S. & Kapur, S. A computational model of the functional role of the ventral-striatal D2 receptor in the expression of previously acquired behaviors. Neural Comput. 17, 361–395 (2005). 41. Dayan, P. & Balleine, B.W. Reward, motivation and reinforcement learning. Neuron 36, 285–298 (2002). 42. Daw, N.D., Courville, A.C. & Touretzky, D.S. Timing and partial observability in the dopamine system. in Advances in Neural Information Processing Systems 15, 99–106 (MIT Press, Cambridge, Massachusetts, 2003). 43. Alexander, G.E., Delong, M.R. & Strick, P.L. Parallel organization of functionally segregated circuits linking basal ganglia and cortex. Annu. Rev. Neurosci. 9, 357–381 (1986).


[

NUMBER 12

[

DECEMBER 2005

44. Baum, E.B. & Smith, W.D. A Bayesian approach to relevance in game playing. Artificial Intelligence 97, 195–242 (1997). 45. Pouget, A., Dayan, P. & Zemel, R.S. Inference and computation with population codes. Annu. Rev. Neurosci. 26, 381–410 (2003). 46. Yu, A.J. & Dayan, P. Uncertainty, neuromodulation, and attention. Neuron 46, 681–692 (2005). 47. Holroyd, C.B. & Coles, M.G. The neural basis of human error processing: Reinforcement learning, dopamine, and the error-related negativity. Psychol. Rev. 109, 679–709 (2002). 48. Botvinick, M.M., Cohen, J.D. & Carter, C.S. Conflict monitoring and anterior cingulate cortex: an update. Trends Cogn. Sci. 8, 539–546 (2004). 49. Hartley, T. & Burgess, N. Complementary memory systems: competition, cooperation and compensation. Trends Neurosci. 28, 169–170 (2005). 50. Parkinson, J.A., Roberts, A.C., Everitt, B.J. & Di Ciano, P. Acquisition of instrumental conditioned reinforcement is resistant to the devaluation of the unconditioned stimulus. Q. J. Exp. Psychol. B 58, 19–30 (2005).

1711


ARTICLES


Neuronal correlates of subjective sensory experience Victor de Lafuente & Ranulfo Romo When a near-threshold stimulus is presented, a sensory percept may or may not be produced. The unpredictable outcome of such perceptual judgment is believed to be determined by the activity of neurons in early sensory cortex. We analyzed the responses of neurons in primary somatosensory cortex, recorded while monkeys judged the presence or absence of threshold stimuli. We found that these responses did not covary with the monkeys’ perceptual reports. In contrast, the activity of frontal lobe neurons did covary with trial-by-trial judgments. Further control and microstimulation experiments indicated that frontal lobe neurons are closely related to the monkeys’ subjective experiences during sensory detection.

A fundamental goal of neuroscience is to understand how sensory experiences arise from activity in the brain. The detection of sensory stimuli is among the simplest perceptual experiences and is a prerequisite for any further sensory processing. Studies on the neuronal correlates of sensory detection showed that, in the case of vibrotactile stimuli, the responses of neurons in primary somatosensory cortex (S1) account for the measured psychophysical accuracy1. However, imaging and physiological studies show that, in addition to sensory cortices, areas of the frontal lobe are also active during sensory detection and discrimination2–5. This evidence raises an important question: what are the specific functional roles of primary sensory cortices and association areas of the frontal lobe in perception? We addressed this question by recording from single neurons in S1 and medial premotor cortex (MPC; neurons from this frontal lobe area are involved in decision processes during somatosensory discrimination)3, while trained monkeys reported the presence or absence of a mechanical vibration of varying amplitude applied to the skin of one fingertip. Here we report that the activity of S1 neurons covaried with stimulus strength, but not with the animals’ perceptual reports. In contrast, the activity of MPC neurons did not covary with stimulus strength, but did covary with the animals’ perceptual reports. We wondered whether, in addition to these neuronal correlates associated with the animals’ perceptual reports, the animals could also perform the detection task if their MPC neurons were activated (artificially) with electrical microstimulation, instead of with the mechanical vibrations delivered to one fingertip. This would provide unequivocal proof that the activity of MPC neurons is directly involved with a specific cognitive function. Psychophysical performance with artificial stimuli was almost identical to that measured with the mechanical stimuli delivered to the fingertips. These results suggest that perceptual judgments arise in the activity of frontal lobe neurons but not in sensory cortices. RESULTS Two monkeys (Macaca mulatta) were trained to perform a detection task (Methods). In each trial, the animal had to report whether the tip

of a mechanical stimulator vibrated or not (Fig. 1a). Stimuli were sinusoidal, had a fixed frequency of 20 Hz and were delivered to the glabrous skin of one fingertip; crucially, they varied in amplitude across trials. Stimulus-present trials were interleaved with an equal number of stimulus-absent trials in which no mechanical vibrations were delivered (Fig. 1a). Depending on the monkeys’ responses, trials could be classified into four types: hits and misses in the stimulus-present condition, and correct rejections and false alarms in the stimulusabsent condition (Fig. 1b). Stimulus detection thresholds were calculated from the behavioral responses (Fig. 1c). S1 responses during vibrotactile detection First, we simultaneously characterized the activity of S1 neurons and the monkeys’ psychophysical performance by recording the extracellular spike potentials of single S1 units while the monkeys performed the vibrotactile detection task (Methods). Thus we obtained each monkey’s psychometric curve and the spike trains of an S1 neuron in the same trials (Figs. 1c and 2a). The firing rate of this neuron varied smoothly as a function of stimulus amplitude, and no clear modulations in its firing rate could be appreciated during the stimulus-absent trials. To test whether the responses of S1 neurons accounted for the monkeys’ psychophysical performance, we calculated neurometric detection curves and compared them with the psychometric curves (Fig. 2b–d). The proportion of ‘yes’ responses for neurometric curves was defined, for a given amplitude, as the proportion of trials in which the neuron’s firing rate reached or exceeded a criterion value6,7 (Methods). For each neuron, this criterion was chosen to maximize the number of correct responses (Fig. 2b). The shape of the mean neurometric curve resulting from the activity of the S1 neurons (n ¼ 59) showed close correspondence with the shape of the mean psychometric curve (Fig. 2c). Pairwise comparisons of detection thresholds, obtained from logistic fits to the simultaneously obtained neurometric and psychometric data, showed that the detection thresholds of individual S1 neurons were not significantly different from the animals’ psychophysical thresholds (Fig. 2d; Wilcoxon signed rank

Instituto de Fisiologıá Celular, Universidad Nacional Autońoma de Me´xico, 04510 Me´xico, D.F., Me´xico. Correspondence should be addressed to R.R. ([email protected]). Received 12 July; accepted 5 October; published online 6 November 2005; doi:10.1038/nn1587

1698

VOLUME 8

[

NUMBER 12

[


ARTICLES Figure 1 Detection task. (a) Trials began when the stimulator probe indented 1.0 the skin of one fingertip of the right, restrained hand (probe down, PD). The Behavioral PU MT monkey then placed its left, free hand on an immovable key (key down, KD). response PD KD 0.5 Yes No On half of the randomly selected trials, after a variable pre-stimulus period Stimulus (‘‘Prestim’’,1.5 s to 3.5 s), a vibratory stimulus (‘‘Stim’’, 20 Hz, 0.5 s) was Hit Miss present 0 Prestim Stimulus presented. Then, after a fixed delay period (‘‘Delay’’, 3 s), the stimulator Stim FA CR 0 10 20 30 40 absent Delay probe moved up (probe up, PU), indicating to the monkey that it could make Stimulus amplitude (µm) the response movement (MT) to one of the two buttons. The button pressed indicated whether or not the monkey felt the stimulus (henceforth referred to as ‘yes’ and ‘no’ responses, respectively). (b) Depending on whether the stimulus was present or absent and on the behavioral response, the trial outcome was classified as a hit, miss, false alarm or correct rejection. Trials were pseudorandomly chosen; 90 trials were stimulus-absent (amplitude ¼ 0), and 90 trials were stimulus-present with varying amplitudes (9 amplitudes with 10 repetitions each). (c) Classical psychometric detection curve obtained by plotting the proportion of yes responses as a function of stimulus amplitude.

b

c

test8, P ¼ 0.15) and that the two threshold measures highly covaried (Pearson’s correlation coefficient9, r ¼ 0.6, t-test: P o 0.01). S1 responses do not covary with perceptual reports We then studied whether the activity of S1 neurons covaried with the perceptual ‘yes’ or ‘no’ judgments that the monkeys made on a trial-bytrial basis. To test this, we compared the mean normalized activity during hit and miss trials for the near-threshold stimulus, as well as the corresponding activity during correct reject and false-alarm trials in the stimulus-absent condition (Methods). We found no significant differences in the activity of S1 neurons between hits and misses (Fig. 2e, upper left panel; t-test: P ¼ 0.47), nor between correct rejections and false alarms (Fig. 2e, upper right panel: t-test, P ¼ 0.59). This indicated that the activity of individual S1 neurons did not predict the monkeys’ behavior. To further quantify this, we calculated a choice probability index, which estimates the probability with which the behavioral outcome can be predicted from the neuronal responses3,4,10. The results indicated that there were no significant differences between hits and misses or between correct rejections and false-alarm trials (Fig. 2e,f). The low choice probability values are consistent with a detection model in which the activity of S1 serves as input to an additional

20 Hz

b

PU MT

Proportion of trials

a

34.6 24.7 17.6 12.6 9.0 6.4 4.6 3.3 2.3

Criterion 34 9 0 µm

0.2

0

Proportion of 'yes'

1.0 Psychometric Neurometric

0.5

n = 59 0

Near-threshold stimulus 20 Hz 8 4

PU MT 8

Hits Misses n = 59

6

Stimulus-absent

PU MT 6 4

Correct rejections False alarms n = 59

2

2 0

0

1.0

1.0

0.5

0.5

0

0 0

1 2 Time (s)

3

f

0

1 2 Time (s)


[

3

Number of neurons

d

Psychometric threshold (µm)

0

e

20 40 60 Spikes per s

c

0.5 s

processing stage(s) to determine whether a stimulus has occurred or not. According to this hypothesis, correlation between S1 activity and the final decision about the stimulus presence or absence is highly dependent on the amount of correlated noise among sensory neurons11. We found that the mean ± s.e.m. noise correlation coefficient across pairs of S1 neurons was 0.16 ± 0.02 (n ¼ 51; see Methods). This amount of correlated noise is similar to those reported in previous studies12–14, and is also consistent with the near-chance choice probability values reported here. These results further support a detection model in which, to judge the stimulus presence or absence, a central area(s) with internal fluctuations must track the activity in S1. MPC responses covary with perceptual reports To test whether the neuronal correlates of the perceptual decisions associated with detection might reside outside S1, we recorded the responses of neurons in the MPC (Fig. 3a), a frontal cortical area known to be involved in the evaluation of sensory information and in decision-making processes3. We found that, in contrast to the graded dependence on stimulus amplitude observed in S1, MPC neurons responded in an all-or-none manner that was only weakly modulated by the stimulus amplitude (Fig. 3b,c) but that closely correlated with yes and no behavioral responses (Fig. 3b). The mean normalized activity across the 50 MPC neurons was strong and sustained and, with near-threshold stimuli, it was clearly different for hit and miss trials (Fig. 3d, upper left panel, t-test: P o 0.001; and Fig. 3e).

0.1

0

0.0 µm

Choice probability Normalized activity (s.d.)


Proportion of 'yes'

a

NUMBER 12

10 20 30 40 Stimulus amplitude (µm)

20

r = 0.6

10 n = 59 0 0 10 20 Neurometric threshold (µm) Hit vs. miss CR vs. FA 20 10 0 0 0.5 1 Choice probability

[

DECEMBER 2005

Figure 2 Activity of S1 neurons during the detection task. (a) Raster plot of the activity of an S1 neuron during the detection task. Each dot marks the time of spike occurrence, and each row is a trial. Trials are arranged by stimulus amplitude, shown at right. Red markers at the end of the trial denote misses in stimulus-present trials and false alarms in stimulus-absent trials. Gray box marks the time of stimulus presentation. (b) Activity distributions of the 59 neurons recorded in S1, grouped by stimulus amplitude (see calibration bar). Black vertical line marks the median criterion value (22 spikes per s) used to produce the neurometric proportion of yes responses for each neuron. Gray box indicates inter-quartile range. (c) Mean psychometric and neurometric detection curves (590 trials for each stimulus amplitude data point; 5310 trials for zero-amplitude data point). (d) Comparison of psychometric and neurometric detection thresholds, obtained from logistic fits (data not shown) to the proportion of yes responses for the neuronal and behavioral data obtained simultaneously. Diagonal marks the identity line; correlation coefficient, r ¼ 0.6. (e) Comparison of normalized neuronal population activity (s.d., standard deviation) during hits and misses for near-threshold stimuli, and during correct rejections and false alarms in stimulus-absent trials. Lower panels show the choice probability index as a function of time. Dotted lines mark significance levels (Methods). (f) Distributions of indices across the population of S1 neurons, calculated between the activity of hit versus miss trials (mean ± s.e.m.: 0.54 ± 0.02) and correct rejection versus false alarm trials (mean ± s.e.m. ¼ 0.50 ± 0.02 s.e.). n, number of neurons; CR, correct rejection; FA, false alarm.

1699

ARTICLES

Hits Misses n = 50

8 6

e

Stimulus-absent

PU MT

PU MT 8 6

4

4

2

2

0

0

1.0

1.0

0.5

0.5

Correct rejections False alarms n = 50

f

0

1 2 Time (s)

3

0

1 2 Time (s)

3

10

0 0 0.5 1 Choice probability

1.0

0.5

0 0

0

Hit vs. miss FA vs. CR

* * **

All trials Trials before FA

0 10 20 Stimulus amplitude (µm)

Moreover, almost 70% of the false alarm trials were predicted from increases in neuronal activity in stimulus-absent trials (Fig. 3d, upper right panel, t-test: P o 0.001; and Fig. 3e). We also found that the MPC activity preceding stimulus onset was higher during hits than during misses (Fig. 3d, upper right panel). These early increases in activity predicted detection success significantly above chance (Fig. 3d). Although we do not know the role of this increased pre-stimulus activity, we speculate that it might be associated with trial history during a run. To investigate this conjecture, we analyzed the behavioral responses on trials preceding false-alarm responses. We found that the probability of a yes response was increased in trials preceding a false alarm, supporting the notion that monkeys were biased toward yes

Figure 4 Sensory versus motor activity. (a) Responses of the MPC neuron shown in Figure 3b, to 10 repetitions of the 9 mm stimulus during the detection condition. (b) Responses of the same neuron to the same stimulus, but in a control condition in which the correct response button (left button) was illuminated at the beginning of each trial. In stimulus-absent trials, the right button was illuminated, so in this case the monkeys also knew the correct response button in advance (responses not shown). (c) Mean responses to a near-threshold stimulus (9 mm) during the standard detection task (continuous line) and during the two control conditions (dashed line, near-threshold stimulus + light; dotted line, near-threshold stimulus + reversed light). Each line is the mean of 140 trials from 14 neurons studied in these conditions (panels a, b and e). Responses to the stimuli were not significantly different across conditions (Kruskal-Wallis8, P ¼ 0.11). (d) Probability of burst response as a function of the stimulus amplitude during the detection task (open circles), detection task + light (squares) and detection + reversed light (asterisks). Symbols and small vertical bars, mean ± s.e.m. (e) Responses of the same neuron shown in a when the same correct response button (right button) was illuminated at the beginning of each trial. In this case, the positions of the buttons were reversed compared to the condition in b. (f) Choice probability indices for the population of neurons (n ¼ 14) tested in the condition shown in a (continuous line); dotted line denotes choice probability indexes of the same neuronal population in the condition shown in b versus the condition shown in e. n ¼ number of neurons.

1700

responses. We speculate that, because yes responses to three subthreshold amplitudes (Fig. 3f) were rewarded, monkeys could have been encouraged to respond yes in the next trial, producing a falsealarm response. The results indicate that, for all MPC neurons studied, increased responses were associated with stimulus presence or with false alarms: that is, with yes responses. We did not find neurons that increased their activity during no responses. We do not know the reason for this, but we speculate that no is a default response that the stimulus presentation needs to override. The close association between neuronal responses and behavioral responses, and the weak relationship between activity and stimulus amplitude, supported the interpretation that MPC neurons do not code the physical attributes of stimuli, but rather represent perceptual judgments about their presence or absence (Fig. 4a). As the monkeys reported their decisions by a motor act, a key question needed to be answered: was the MPC activity truly related to stimulus perception, or was it simply reflecting the different motor actions associated with the two response buttons? To test this, we designed a control task in which the correct response button was illuminated at the beginning of every trial3,4. In this variant of the detection task, the monkeys simply had to wait until the end of the trial to push the illuminated button, without needing to attend to the presence or absence of the mechanical

a

d 1.0 Near-threshold stimulus

0.5

n = 14 0 0

b

e

Near-threshold stimulus + light

c

f

10 20 30 Stimulus amplitude (µm)

40

Near-threshold stimulus + reversed light

0.5 s 100 80 60 40 20 0

n = 14 –1

VOLUME 8

[

0 1 Time (s)

2

NUMBER 12

3

[

Choice probability

20 Hz

Proportion of 'yes'

Normalized activity (s.d.)

Near-threshold stimulus

Proportion of burst

0.5 s

d

Number of neurons

0 10 20 30 40 Stimulus amplitude (µm)

Right

MPC (n = 50) S1 (n = 59)

Spikes per s

70 60 50 40 30 20 10 0

Figure 3 Activity of MPC neurons during the detection task. (a) Raster plot of the activity of an MPC neuron during the detection task; same conventions as in Figure 2a. (b) Responses of the S1 neuron shown in Figure 2a and the MPC neuron shown in Figure 3a, at near-threshold stimulus (9 mm). (c) Mean firing rates of hit trials for S1 (n ¼ 59) and MPC (n ¼ 50) neurons. (d) Comparison of normalized activity (s.d., standard deviation) during hits and misses in near-threshold trials, and during correct rejections and false alarms in stimulus-absent trials. Lower panels show the average choice probability index as a function of time. Dotted lines mark significance levels (Methods). (e) Distributions of indices across the population of MPC neurons (n ¼ 50), calculated between the activity of hit and miss trials (hit vs. miss, mean ± s.e.m.: 0.85 ± 0.02) and between correct rejection and false alarm trials (CR versus FA, mean ± s.e.m.: 0.81 ± 0.02). (f) Detection curves resulting from all trials (overall performance, black circles; except for stimulus amplitude 34.6 mm) and from trials that preceded false-alarm responses (gray circles). Asterisks indicate significant differences in the probability of ‘yes’ responses (Fisher’s exact test, P o 0.01)32. n ¼ number of neurons; FA, false alarm; CR, correct reject.

Right Left

Misses Hits

S1 MPC

0.5 s

c

0.0 µm

Choice probability


34.6 24.7 17.6 12.6 9.0 6.4 4.6 3.3 2.3

9.0 µm

Left

b

Misses Hits

PU MT

Spikes per s

20 Hz

a

1.0 0.5 n = 14 0 –1

0 1 Time (s)

2

3


ARTICLES 1.0

b

0.5 Mechanical + electrical Mechanical

0

10

20

30

Electrical Mechanical

0.5 n = 14

10

n = 10

0


1.0

Proportion of 'yes'

Proportion of 'yes'

a

20

30

40 µm

0

40

0

Stimulus amplitude (µm)

5

10

15 µA

Stimulus amplitude

Figure 5 Detection curves during microstimulation experiments. (a) The mean detection curves for mechanical stimuli (black) and for mechanicalplus-electrical stimuli (gray). Mechanical and mechanical-plus-electrical trials were randomly interleaved (Methods). Ten experiments were conducted with this protocol, and each experiment consisted of 10 repetitions of each kind of stimulus; thus, each point is the mean of 100 stimulus-present trials and 900 stimulus-absent trials . (b) The mean detection curves for mechanical stimuli (black) and for electrical stimuli (gray). Mechanical and electrical stimulation trials were randomly interleaved (Methods). Fourteen experiments were conducted with this protocol, and each experiment consisted of 10 repetitions of each kind of stimulus; thus, each point is the mean of 140 stimulus-present trials and 1260 stimulus-absent trials. n ¼ number of runs; symbols and small bars, mean ± s.e.m.

vibration. Raster plots of the neural activity for an example neuron showed that the responses to the stimulus in this control condition (Fig. 4b and dashed line in Fig. 4c) were very similar to the responses in the standard detection task (Fig. 4a and continuous line in Fig. 4c). In this test condition, all-or-none activity was still observed in relation to the near-threshold stimulus, and the probability of activation depended on the stimulus amplitude as in the standard detection task (Fig. 4d). Given that in the control test, the monkeys did not have to choose a response button based on the vibratory stimulus, the results are consistent with the interpretation that the activity of these MPC neurons is related to the subjective perception of sensory stimuli, rather than to the selection of the motor plan. To further examine whether MPC activity was associated with the preparation of movements in different directions, we did a second control experiment in which the correct response button was illuminated at the beginning of every trial, as before. In this case, however, we switched response buttons so that the yes button was now illuminated during stimulus-absent trials and, conversely, the no button was illuminated during stimulus-present trials. The results showed that reversing the direction of the arm movements did not change the all-ornone character of the evoked MPC activity (Fig. 4e and dotted line in Fig. 4c). To test whether the direction of movement had an influence on the responses of MPC neurons, we calculated the choice probability index between the activities observed during the light (left movement; Fig. 4b) versus reversed light (right movement, Fig. 4e) conditions. The analysis shows that the choice probability values of MPC neurons were close to chance levels (dotted line in Fig. 4f), suggesting that these activities were not associated with the animals’ hand and arm movements. Had these neurons participated primarily in movement choice or movement generation, their firing rates should have been consistently higher for one movement but not for the other. The observation of all-or-none responses during these control tasks favors the hypothesis that this MPC population reflects the failure or success of the nearthreshold stimulus in triggering a sensory percept. Microstimulation of MPC triggers perceptual reports Given the close association between MPC activity and the behavioral reports of stimulus detection, we wondered whether artificial activation


[

NUMBER 12

[

DECEMBER 2005

of MPC neurons through electrical microstimulation15,16 would increase the monkeys’ probability of detecting the vibratory stimuli. To test this, we injected a weak electrical current through the recording electrode in randomly selected stimulus-present and stimulus-absent trials (Methods). The resulting detection curves, separated into mechanical-plus-electrical and mechanical-only curves (Fig. 5a), show that monkeys tended to answer yes more often on microstimulation trials than with mechanical stimuli only. The increased probability of yes responses observed during microstimulation trials agreed with the hypothesis that MPC activity is related to perceptual judgments. Microstimulation experiments in the dorsal premotor cortex (n ¼ 5) using the protocol described above did not produce significant effects on the behavioral performance (data not shown). To further test whether artificial activation of MPC neurons could mimic neuronal activity related to sensory percepts, we did an experiment in which the mechanical vibrations were substituted by electrical stimuli of varying current strengths (Methods). We plotted detection curves for purely electrical stimuli, together with the detection curves for the mechanical stimuli that were randomly interleaved (Fig. 5b). Although it is difficult to compare these two stimulus quantities, the results show that psychometric performance based on microstimulation of MPC resembled that based on vibrotactile stimuli delivered to the skin. The same microstimulation protocol was used in dorsal premotor cortex (n ¼ 6), but in this case monkeys always reacted as in the stimulus-absent trials (data not shown). DISCUSSION Our results suggest that sensory and frontal lobe neurons have significantly different roles during perceptual judgments. The activity of MPC—but not S1—neurons covaried with the reported sensory percepts during the vibrotactile detection task. Therefore, the functional role of S1 in this and other perceptual tasks may be mainly to generate a neural representation of the sensory stimulus, for further processing in areas central to S17,12,13,17,18. However, a previous study found that functional magnetic resonance imaging (fMRI) signals in primary visual cortex (V1) reflect the percepts of human subjects, rather than the encoded stimulus features19. This result suggests that, in V1, top-down signals (non-sensory inputs delivered to visual cortex via feedback projections) can be combined with bottom-up (sensory) information and contribute to sensory percepts19. Our S1 data did not show evidence for this type of neural interaction; rather, it indicated that S1 represents the physical properties of stimuli and contributes little to near-threshold percepts. The discrepancy could be due to fundamentally different organizations across sensory cortices or to differences between species. Another possibility to consider is that the modulation revealed through fMRI may have an effect that is invisible from the point of view of single neurons. This would happen if, for instance, such modulation acted only to synchronize the spikes of multiple target neurons20,21. On the other hand, frontal lobe neurons, which are involved in decision-making3,4,22,23, working memory3,4,17,24 and motor planning25,26, did seem to be fundamental to perceptual judgments during sensory detection. This is consistent with the idea that perceptual judgments result from the interaction between internal signals (working memory, expectation, attention) and sensory inputs3–5,17,19, because the MPC is ideally situated to integrate these different types of information27. For the same reason, it is possible that other circuits of the frontal3,4,18 and parietal lobes17 also contribute to perceptual judgments in a similar way. An important clue about this interpretation is that the electrical current injected into MPC led to behavioral responses that resembled those elicited by mechanical vibrations,

1701


ARTICLES suggesting that although the artificial stimulus did not originate in S1, it was still interpreted as sensory evidence. We do not know whether microstimulation in MPC evoked the same somatosensory sensation as that evoked by natural stimuli, but it produced the same behavioral reactions. Another possibility is that microstimulation of MPC does not produce any somatic sensation but, instead, activates a task rule such as ‘a stimulus is present’. In this manner, varying the microstimulation strength could vary the probability of engaging a population of MPC neurons associated with this rule and, therefore, produce a psychometric detection curve similar to that produced by varying the mechanical stimulus strength. The contribution of different cortical areas to perceptual processing has also been investigated using binocular rivalry and other protocols in which a fixed but ambiguous visual stimulus gives rise to multiple, alternating percepts; that is, the same sensory input is consistent with multiple perceptual interpretations28,29. These studies agree with the present data in that high-order cortices show much stronger correlations with behavioral (perceptual) reports than do primary sensory areas. METHODS Detection task. Stimuli were delivered to the skin of the distal segment of one digit of the restrained hand, via a computer-controlled stimulator (BME Systems; 2-mm round tip). Initial probe indentation was 500 mm. Vibrotactile stimuli consisted of trains of 20 Hz mechanical sinusoids with amplitudes of 2.3–34.6 mm (Fig. 1). These were interleaved with an equal number of trials where no mechanical vibrations were delivered to the skin (amplitude ¼ 0). Animals pressed one of two buttons to indicate stimulus-present (left button) or stimulus-absent (right button). They were rewarded with a drop of liquid for correct responses. Performance was quantified through psychometric techniques1. Animals were handled according to institutional standards of the US National Institutes of Health and the Society for Neuroscience. Recording sessions and sites. Neuronal recordings were obtained with an array of seven independent, movable microelectrodes3,4 (2–3 MO) inserted into S1 and MPC. Recordings in S1 were made in areas 3b and 1, contralateral to the stimulated hand and ipsilateral to the responding hand and arm (two monkeys). Initially, we recorded S1 neurons with cutaneous receptive fields with quickly adapting or slowly adapting properties, but we found that the neurons with slowly adapting properties showed weak modulation in their firing rate during the stimuli (data not shown). We therefore focused on the quickly adapting neurons. Recordings in MPC (pre-supplementary motor area)3,27 were made in both hemispheres. Electrodes were advanced into the MPC to find neurons that responded during the task. MPC neurons preferentially responded during the stimulus and delay periods of the task. Recording sites in S1 and MPC changed from session to session. The locations of the electrode penetrations in S1 and MPC were confirmed with standard histological techniques. Data analysis. We analyzed the responses of 59 S1 neurons (area 3b: n ¼ 28; area 1: n ¼ 31). All the S1 neurons had small cutaneous receptive fields located in the distal segment of one digit (distal segments of fingertips 2, 3 or 4) and had quickly adapting properties. Stimuli were delivered to the center of the neuron’s cutaneous receptive field while the monkeys executed the detection task (Fig. 2a). A total of 127 responsive neurons were recorded in the MPC of both hemispheres during the detection task. These neurons were sorted in two groups according to their response dynamics: one that had transient responses lasting up to 1 s after stimulus offset (n ¼ 40), and another that showed persistent activity starting during the stimulus onset and continuing throughout the full delay period until ‘‘probe up’’ (PU) triggered the hand and arm movement (n ¼ 87). For analysis, we used 50 of the 87 neurons that had sustained activity (because recordings were stable for these 50); during the detection task, we collected 10 repetitions per stimulus amplitude and 90 repetitions of the stimulus-absent trials (Fig. 3a). The neurons that had

1702

transient responses (n ¼ 40) also had bimodal activity and generally behaved similarly, albeit for a limited time (data not shown). To calculate response distributions and the neurometric detection curves of S1 neurons (Fig. 2b–d), on each trial we obtained the maximum firing rate in a 500-ms window that was displaced every 1 ms in the period between 1.5 s before and 3.5 s after stimulus onset (the same period was used for stimulusabsent trials). Neurometric curves were calculated as the proportion of trials in which the maximum firing rate reached or surpassed a criterion level. For each neuron, this criterion was chosen to maximize the number of hits and correct rejections (that is, correct trials). From logistic fits, we calculated psychometric and neurometric detection thresholds as the probability that the proportion of ‘yes’ responses would be 0.5. For 59 S1 neurons and 50 MPC neurons, we calculated the firing rate as a function of time, using a 200-ms window displaced every 50 ms (Figs. 2e and 3d). Normalized activity was calculated by subtracting the mean activity and dividing by the standard deviation of the activity from a 200-ms window of the pre-stimulus period. (The same results were obtained using the raw data.) Normalized activity shown in Figure 2e was based on 370 hits and 370 misses (left panels) and 620 false alarms and 620 correct rejections (right panels), collected during the study of the 59 S1 neurons. Normalized activity shown in Figure 3d was based on 312 hits and 312 misses (left panels) and 494 false alarms and 494 correct rejections (right panels), collected during the study of the 50 MPC neurons. We used trials with 12.6, 9.0 and 6.4 mm stimulus amplitude. Choice probability index was calculated using methods from signal detection theory4,6,17. This quantity measures the overlap between two response distributions: in this case, between hits and misses and between correct rejections and false-alarm trials. Dashed lines in Figures 2e and 3d indicate P ¼ 0.01 significance limits, bootstrap technique30. To determine the differences between hit and miss responses and between correct rejection and false alarm responses, we used the two-tailed t-test on the distributions of the number of spikes found in a 500 ms window during the stimulus period for S1 neurons, and during a 500 ms window starting 250 ms after stimulus onset for MPC neurons. To estimate the amount of correlated noise activity across S1 neurons, the Pearson’s correlation coefficient was calculated for each pair of simultaneously recorded S1 neurons (n ¼ 51)9. We first standardized the firing rates of the two neurons by subtracting the mean and dividing by the standard deviation across the ten stimulus repetitions. Trials were sorted by stimulus amplitude. For each pair of neurons, we obtained a correlation coefficient as function of stimulus amplitude. The mean correlation coefficient, across the 51 pairs of neurons and across stimulus amplitude classes, was 0.16 ± 02 (mean ± s.e.m.). We found no relation between correlation coefficient and stimulus amplitude or trial type (hit, miss, correct rejection or false alarm). Trials in the control light task proceeded exactly as described in Fig. 1a, except that at the probe down, the correct target button was illuminated (Fig. 4b,e). Vibrotactile stimuli were delivered while the light was kept on; then the probe was lifted off from the skin (PU) and the light was turned off. The monkey was rewarded for pressing the previously illuminated button. Hand and arm movements were identical to those in the somatosensory detection task but were cued by visual stimuli. Under this condition, the choice probability indices (Fig. 4c) and burst proportion (Fig. 4d) were calculated by comparing response distributions for left versus right button presses. To estimate the proportion of bursts as a function of stimulus amplitude (Fig. 4f), we used a Poisson spike analysis31 that determined whether or not a burst occurred on each trial. First, we counted the spikes across the whole trial and divided them by the trial duration to obtain the mean trial firing rate. Second, we counted the number of spikes in a 500-ms window, beginning 250 ms after stimulus onset. Finally, using the Poisson cumulative density function, we estimated the probability of obtaining a firing rate equal or larger than that observed in the 500-ms response window31, given the mean firing rate across the whole trial. We considered that a burst occurred in a given trial if this probability was less than 0.05. Microstimulation. A computer-controlled pulse generator (Coulbourn), in series with an optical stimulus isolation unit, produced biphasic current pulses with the cathodal phase leading. Each phase lasted 0.2 ms, with 0.05 ms

VOLUME 8

[

NUMBER 12

[



ARTICLES between phases. Microstimulation consisted of 5-mA biphasic current pulses, delivered at 200 Hz and superimposed over the period that corresponded to either the stimulus-present or stimulus-absent mechanical trials (Fig. 5a). By stimulating in both stimulus-present and stimulus-absent trials, we did not reinforce the monkeys to answer yes in microstimulation trials. In fact, because of the more frequent false-alarm responses (non-rewarded trials), overall performance during microstimulation trials was slightly, but significantly, lower than performance during normal detection trials. In a second microstimulation protocol, mechanical stimuli were randomly substituted by electrical stimuli of 200 Hz with varying amplitudes (1–12 mA) in one half of the stimulus-present trials (Fig. 5b). In this experiment, monkeys were rewarded for answering yes in microstimulation trials. We cannot discard the possibility that microstimulation generated a non-natural sensation that could be used by the subjects to indicate detection. We believe, however, that this is unlikely because from the very first microstimulation trials, monkeys generated yes responses even if they were not rewarded for this behavioral report. ACKNOWLEDGMENTS We thank A. Hernańdez, L. Lemus, Y. Vazquez and R. Luna for technical assistance and C. Brody, J. Gold, J. Maunsell, E. Salinas, E. Seidemann and M. Shadlen for comments. R.R. was supported by an International Scholars Award from the Howard Hughes Medical Institute and by grants from the Direccioń General del Personal Acade´mico de la Universidad Nacional Autońoma de Me´xico and the Consejo Nacional de Ciencia y Tecnologıá. COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/ 1. Mountcastle, V.B., Talbot, W.H., Sakata, H. & Hyvarinen, J. Cortical neuronal mechanisms in flutter-vibration studied in unanesthetized monkeys. Neuronal periodicity and frequency discrimination. J. Neurophysiol. 32, 452–484 (1969). 2. Shulman, G.L., Ollinger, J.M., Linenweber, M., Petersen, S.E. & Corbetta, M. Multiple neural correlates of detection in the human brain. Proc. Natl. Acad. Sci. USA 98, 313–318 (2001). 3. Hernańdez, A., Zainos, A. & Romo, R. Temporal evolution of a decision-making process in medial premotor cortex. Neuron 33, 959–972 (2002). 4. Romo, R., Hernańdez, A. & Zainos, A. Neuronal correlates of a perceptual decision in ventral premotor cortex. Neuron 41, 165–173 (2004). 5. Romo, R. & Salinas, E. Flutter discrimination: neural codes, perception, memory and decision making. Nat. Rev. Neurosci. 4, 203–218 (2003). 6. Green, D.M. & Swets, J.A. Signal Detection Theory and Psychophysics (John Wiley, New York, 1966). 7. Hernańdez, A., Zainos, A. & Romo, R. Neuronal correlates of sensory discrimination in the somatosensory cortex. Proc. Natl. Acad. Sci. USA 97, 6191–6196 (2000).


[

NUMBER 12

[

DECEMBER 2005

8. Siegel, S. & Castellan, N.J. Non-parametric Statistics for the Behavioral Sciences. (McGraw-Hill, New York, 1988). 9. Press, W.H., Teukolsky, S.A., Vetterling, W.T. & Flannery, B.P. Numerical Recipes in C: the Art of Scientific Computing 2nd edn. (Cambridge University Press, Cambridge, UK, 1992). 10. Britten, K.H., Newsome, W.T., Shadlen, M.N., Celebrini, S. & Movshon, J.A. A relationship between behavioral choice and the visual responses of neurons in macaque MT. Vis. Neurosci. 13, 87–100 (1996). 11. Zohary, E., Shadlen, M.N. & Newsome, W.T. Correlated neuronal discharge rate and its implications for psychophysical performance. Nature 370, 140–143 (1994). 12. Salinas, E., Hernańdez, A., Zainos, A. & Romo, R. Periodicity and firing rate as candidate neural codes for the frequency of vibrotactile stimuli. J. Neurosci. 20, 5503–5515 (2000). 13. Romo, R., Hernańdez, A., Zainos, A. & Salinas, E. Correlated neuronal discharges that increase coding efficiency during perceptual discrimination. Neuron 38, 649–657 (2003). 14. Bair, W., Zohary, E. & Newsome, W.T. Correlated firing in macaque in visual MT: time scales and relationship to behavior. J. Neurosci. 21, 1676–1697 (2001). 15. Romo, R., Hernańdez, A., Zainos, A. & Salinas, E. Somatosensory discrimination based on cortical microstimulation. Nature 392, 387–390 (1998). 16. Romo, R., Hernańdez, A., Zainos, A., Brody, C.D. & Lemus, L. Sensing without touching: psychophysical performance based on cortical microstimulation. Neuron 26, 273–278 (2000). 17. Romo, R., Hernańdez, A., Lemus, L. & Brody, C.D. Neuronal correlates of decisionmaking in secondary somatosensory cortex. Nat. Neurosci. 5, 1217–1225 (2002). 18. Romo, R., Brody, C.D., Hernańdez, A. & Lemus, L. Neuronal correlates of parametric working memory in the prefrontal cortex. Nature 399, 470–473 (1999). 19. Ress, D. & Heeger, D.J. Neuronal correlates of perception in early visual cortex. Nat. Neurosci. 6, 414–420 (2003). 20. Fries, P., Neuenschwander, S., Engel, A.K., Goebel, R. & Singer, W. Rapid feature selective neuronal synchronization through correlated latency shifting. Nat. Neurosci. 4, 194–200 (2001). 21. Fries, P., Reynolds, J.H., Rorie, A.E. & Desimone, R. Modulation of oscillatory neuronal synchronization by selective visual attention. Science 291, 1560–1563 (2001). 22. Kim, J.N. & Shadlen, M.N. Neural correlates of a decision in the dorsolateral prefrontal cortex of the macaque. Nat. Neurosci. 2, 176–185 (1999). 23. Schall, J.D. Neural basis of deciding, choosing and acting. Nat. Rev. Neurosci. 2, 33–42 (2001). 24. Miller, E.K. & Cohen, J.D. An integrative theory of prefrontal cortex function. Annu. Rev. Neurosci. 24, 167–202 (2001). 25. Tanji, J. Sequential organization of multiple movements: involvement of cortical motor areas. Annu. Rev. Neurosci. 24, 631–651 (2001). 26. Cisek, P. & Kalaska, J.F. Neural correlates of reaching decisions in dorsal premotor cortex: specification of multiple direction choices and final selection of action. Neuron 45, 801–814 (2005). 27. Rizzolatti, G. & Luppino, G. The cortical motor system. Neuron 27, 889–901 (2001). 28. Leopold, D.A. & Logothetis, N.K. Activity changes in early visual cortex reflect monkeys’ percepts during binocular rivalry. Nature 379, 549–553 (1996). 29. Leopold, D.A. & Logothetis, N.K. Multistable phenomena: changing views in perception. Trends Cogn. Sci. 3, 254–264 (1999). 30. Efron, B. & Tibshirani, R.J. An Introduction to the Bootstrap (Chapman and Hall, New York, 1993). 31. Hanes, D.P., Thompson, J.G. & Schall, J.D. Relationship of presaccadic activity in frontal eye field to saccade initiation in macaque: Poisson spike train analysis. Exp. Brain Res. 103, 85–96 (1995). 32. Collett, D. Modeling Binary Data (Chapman & Hall, London, 1991).

1703


ARTICLES

FARP2 triggers signals for Sema3A-mediated axonal repulsion Toshihiko Toyofuku1–3, Junko Yoshida2,3, Tamiko Sugimoto2, Hong Zhang1,2,4, Atsushi Kumanogoh2,3, Masatsugu Hori1 & Hitoshi Kikutani2,3 Sema3A, a prototypical semaphorin, acts as a chemorepellent or a chemoattractant for axons by activating a receptor complex comprising neuropilin-1 as the ligand-binding subunit and plexin-A1 as the signal-transducing subunit. How the signals downstream of plexin-A1 are triggered upon Sema3A stimulation, however, is unknown. Here we show that, in the presence of neuropilin-1, the FERM domain–containing guanine nucleotide exchange factor (GEF) FARP2 associates directly with plexin-A1. Sema3A binding to neuropilin-1 induces the dissociation of FARP2 from plexin-A1, resulting in activation of FARP2’s Rac GEF activity, Rnd1 recruitment to plexin-A1, and downregulation of R-Ras. Simultaneously, the FERM domain of FARP2 sequesters phosphatidylinositol phosphate kinase type I isoform PIPKIc661 from talin, thereby inhibiting its kinase activity. These activities are required for Sema3A-mediated repulsion of outgrowing axons and suppression of neuronal adhesion. We therefore conclude that FARP2 is a key molecule involved in the response of neuronal growth cones to class-3 semaphorins.

In the developing nervous system, the axons and dendrites of differentiating neurons are guided toward their appropriate target regions by attractive and repulsive extracellular cues1. One such guidance cue is provided by semaphorins, which act through the plexin family of receptors. Plexins are a family of transmembrane receptors characterized by the presence of an extracellular sema domain2 and an intracellular sex-plexin (SP) domain3. Type-A plexins act together with ligand-binding neuropilins as signal transducers for soluble class-3 semaphorins, whereas without neuropilins, some type-A plexins can serve as receptors and sometimes ligands for transmembrane semaphorins such as Sema1a and Sema6D4,5. In the leading edge of the growth cone, the balance between dynamic membrane protrusion, driven by actin polymerization, and membrane withdrawal, driven by myosin-dependent retraction of the cortical actin network, governs axonal guidance. The required mechanical force is generated at the point where the actin network is temporarily anchored to the extracellular matrix via an integrin-based linkage6. By acting as chemorepellents and chemoattractants, semaphorins regulate a balanced progression of the mechanical steps involved in growth cone dynamics. Rho GTPases, including Rho, Rac and Cdc42, have been strongly implicated in the regulation of cell motility and axonal guidance7. Although the cytoplasmic domain is highly conserved among members of the plexin family, biochemical evidence suggests that different plexin receptors use different cytoplasmic signaling mechanisms after semaphorin binding. For example, binding of Sema4D to plexin-B1 induces Rho GTP–mediated contraction of actin-myosin filaments via PDZRhoGEF and leukemia-associated RhoGEF (LARG)8–10. In contrast, in

dorsal root ganglia (DRG) neurons and spinal motor neurons, Sema3A-induced collapse requires Rac-GTP11–14. Furthermore, another small G protein, Rnd1, has been found to interact with the cytoplasmic region of plexins15,16. How all of these molecules are coordinated to regulate axonal guidance, however, is largely unknown. In this study, we show that FARP2, which associates with the plexinA1–neuropilin-1 complex, is a GEF protein responsible for Rac activation in DRG neurons (Fig. 1). Sema3A-induced dissociation of FARP2 from plexin-A1 and activation of its Rac-GEF activity triggers a series of biochemical events such as Rac activation, Rnd1 binding to plexin-A1 and downregulation of R-Ras. In addition, FARP2 suppresses focal adhesion by inhibiting an isoform of type-I phosphatidylinositol phosphate kinase, PIPKIg661. Thus, FARP2 is critical not only in triggering but also in coordinating downstream signals of plexin-A1, which are necessary for Sema3A-mediated axonal repulsion. RESULTS Binding of FARP2 to plexin-A1 The amino acid sequences of the membrane-proximal cytoplasmic regions of plexin-A1, -A2, -A3, -A4 and -B1 are similar to the sequence of the FERM domain binding site of the hyaluronan receptor CD44 (Fig. 1a). In particular, the triplet of basic amino acids (KRK or RRK) that is essential for the binding of CD44 to the FERM domain of ezrin17 is well conserved in each of these plexins. To search for molecules that bind to this region of plexin-A1, we coexpressed various FERM domain–containing proteins with plexin-A1 and neuropilin-1 in HEK293 cells. Of the proteins we tested, the FERM domain–containing

1Department of Cardiovascular Medicine, Osaka University Graduate School of Medicine, 2-2 Yamada-oka, Suita, Osaka 565-0871, Japan. 2Department of Molecular Immunology, Research Institute for Microbial Diseases, Osaka University, 3-1 Yamada-oka, Suita, Osaka 565-0871, Japan. 3CREST, Japan Science and Technology Corporation, 4-1-8 Kawaguchi, Saitama 332-6612, Japan. 4Present address: Institute of Cardiovascular Sciences, Health Science Center, Peking University, Beijing 100083, China. Correspondence should be addressed to T.T. ([email protected]) or H.K. ([email protected]).

Received 10 August; accepted 12 October; published online 13 November 2005; doi:10.1038/nn1596

1712

VOLUME 8

[

NUMBER 12

[


ARTICLES

b

c

TM

myc

Plexin-A1

670–

–683

1262– 1258– 1240–

–1275 –1271 –1253

Plexin-A1 (AAA)

1254– 1509– 1109– 1292–

–1267 –1522 –1122 –1305

FARP2

KRK

e

+

+

+

+

Plexin-A1–myc Plexin-A1 (AAA)–myc

+ –

– +

+ –

+ –

FARP2-V5 FARP2∆C-V5

+ –

+ –

– +

– –

FARP2∆N-V5

–

–

–

+

AAA FERM DH PH V5

FARP2∆C

100

Blot: Anti-myc

FARP2∆N

IP: Anti-myc Anti-V5

Sema3A (min)

0

10

30

60

0

10

30

60

Neuropilin-1

+

+

+

+

+

+

+

+

Plexin-A1-myc

+

+

+

+

+

+

+

+

FARP2-V5

–

–

–

–

+

+

+

+

*

60

*

40

*

*

20 0 0

Blot:

IP: Anti-V5 Anti-myc

Anti-myc

f

GST-CRIB Pull-down

80

Anti-V5 Anti-V5

IP: Anti-myc

Rac Rho CDC42

Total Rac

Sema3A

–

+

–

+

FARP2 siRNA

–

–

+

+

GST-CRIB Pull-down

GST-RBD Pull-down

Total Rac

Total R-Ras

5 10 15 Time (min)

20

g 100 [3H]GDP bound (%)


Plexin-A1 Plexin-A2 Plexin-A3 Plexin-A4 Plexin-B1 Plexin-C1 Plexin-D1

d Neuropilin-1

[3H]GDP bound (%)

a CD44

*

80 60 40 20

0 FARP2-V5

+

+

Plexin-A1–myc

–

+

Neuropilin-1

–

+

Figure 1 Plexin-A1 binds to FARP2 and regulates its Rac GEF activity. (a) An amino acid comparison of the cytoplasmic regions of plexin family members and the FERM domain interaction site of CD44 showing the conserved triplet of basic residues (red). (b) Structures of wild-type plexin-A1, plexin-A1(AAA), wildtype FARP2, FARP2DN and FARP2DC (see Methods). (c) Immunoprecipitation (IP) and immunoblot (blot) results using transfected HEK293 cells show that plexin-A1, but not plexin-A1(AAA), binds to FARP2, whereas FARP2 and FARP2DC, but not FARP2DN, bind to plexin-A1. (d) Time course of nucleotide exchange of [3H]GDP-loaded Rac, Rho and Cdc42 caused by FARP2. Values are expressed as percentage of the value at t ¼ 0. *P o 0.05 compared with value at t ¼ 0. (e) Time course of amounts of FARP2 bound to plexin-A1, Rac-GTP and R-Ras-GTP in response to Sema3A (50 ng ml–1), measured by immunoprecipitation and immunoblot (for plexin-A1), a GST-fused CRIB domain of PAK (GST-CRIB) pull-down assay (for Rac-GTP) and GST-fused c-Raf-1 RBD (GST-RBD) pull-down assay (for R-Ras-GTP). (f) Rac activity in DRG neurons treated with or without FARP2 siRNA in response to Sema3A (50 ng ml–1) analyzed by a GST-CRIB pull-down assay. Knockdown of FARP2 abolished the Rac activity of neurons in response to Sema3A. (g) The nucleotide exchange of [3H]GDP-loaded Rac caused by FARP2 in the absence (square) or presence (circle) of plexin-A1 and neuropilin-1. Values observed at 20 min are expressed as percentages of the values at t ¼ 0. *P o 0.05, compared with the value in the absence of plexin-A1 and neuropilin-1.

GEF protein FARP2 (FERM, RhoGEF and pleckstrin domain protein 2, also known as FIR18) coimmunoprecipitated with plexin-A1 (Fig. 1b,c). Other type-A plexins also interacted with FARP2 in the presence of neuropilin-1 (Supplementary Fig. 1 online), whereas plexinB1 did not (data not shown). However, the interaction of plexin-A1 with FARP2 was substantially reduced in the absence of neuropilin-1 (Fig. 2a). The presence of the mutant neuropilin-1 lacking the cytoplasmic region was also sufficient for the binding of plexin-A1 to FARP2 (Supplementary Fig. 1), indicating that neuropilin-1 does not directly bind to FARP2 but rather may induce conformational change of plexin-A1, thereby making the FARP2 binding site more accessible to FARP2. When alanines were substituted for the basic KRK residues in plexin-A1, the mutant protein did not interact with FARP2 (Fig. 1c and Supplementary Fig. 1). Moreover, FARP2 deletion mutants revealed that the N-terminal FERM domain of FARP2 was necessary and sufficient for plexin-A1 binding (Fig. 1c). Thus, the plexin-A1– FARP2 interaction is dependent on the KRK sequence of plexin-A1, the FERM domain of FARP2 and the presence of neuropilin-1. Sema3A stimulates RacGEF activity of FARP2 Sema3A-induced growth cone collapse is known to require the activation of Rac. We confirmed a previous report that FARP2 shows GEF activity specific for Rac (Fig. 1d)18. A pull-down assay using a fusion protein containing the Cdc42/Rac interactive binding (CRIB) domain of p21-activated kinase (PAK), which selectively binds to Rac-GTP, was


[

NUMBER 12

[

DECEMBER 2005

then used to examine RacGEF activity in cells transfected with constructs encoding plexin-A1, neuropilin-1 and/or FARP2 before and after Sema3A stimulation (Fig. 1e). Sema3A stimulation increased RacGTP in cells expressing plexin-A1, neuropilin-1 and FARP2, which reached a maximal level after 10–30 min and then returned to the basal level within 60 min. In addition, FARP2 started to dissociate from plexin-A1 10 min after Sema3A stimulation, and this dissociation seemed to be irreversible in the presence of Sema3A. RacGEF activity, however, was not detected in cells expressing plexin-A1 and neuropilin-1 but not FARP2, even after Sema3A stimulation. Sema3A also increased Rac-GTP transiently in DRG neurons (Supplementary Fig. 2). The expression of dominant-negative forms of FARP2 such as FARP2DC and FARP2(AE), or knockdown of FARP2 by small interference RNA (siRNA), abolished Sema3A-induced Rac activation in DRG neurons (Fig. 1f and Supplementary Fig. 2). These results indicate that FARP2 is a RacGEF that is responsible for Sema3Ainduced Rac activation. In addition, the GEF activity of FARP2 in the complex with plexin-A1 and neuropilin-1 was significantly lower than that of free FARP2, suggesting that the GEF activity of FARP2 is downregulated by its interaction with plexin-A1 (Fig. 1g). The SP domain of plexins shows sequence similarity with Ras family–specific GTPase activating proteins (GAP). Rnd1, a member of the Rho GTPase family, binds to this region of both type-A and type-B plexin molecules15. This binding stimulates GAP activity of plexin-B1 for R-Ras, a Ras family member, which is critical for

1713

ARTICLES

a

Sema3A Neuropilin-1 Plexin-A1–myc FARP2-V5 Rnd1-Xpress

– – + + +

+ – + + +

– + + + +

+ + + + +

b

Sema3A Neuropilin-1 Plexin-A1–myc FARP2(AE)-V5 Rnd1-Xpress

Blot:

– + + + +

IP: Anti-myc

Anti-V5

Anti-V5


Anti-Xpress

Anti-Xpress Anti-myc

Anti-myc IP: Anti-V5

IP: Anti-V5

Anti-V5

Anti-V5

Anti-Xpress

Anti-Xpress

Anti-myc IP: Anti-Xpress

Anti-myc IP: Anti-Xpress

Anti-V5

Anti-Xpress

GST-RBD Pull-down

GST-RBD Pull-down

Total R-Ras

Total R-Ras

Sema3A FARP2 siRNA FARP2(AE) adeno. FARP2 adeno.

– – – –

+ – – –

– + – –

+ + – –

– – + –

+ – + –

– – – +

+ – – +

GST-RBD Pull-down Total R-Ras

plexin-B1–mediated growth cone collapse19. We also confirmed that Rnd1 enhanced the GAP activity of the cytoplasmic region of plexin-A1 in vitro (Supplementary Fig. 3), as reported for plexin-B1 (ref. 20). We then examined whether or not FARP2 is involved in the regulation of Rnd1 and R-Ras (Fig. 2a,b). Without expression of neuropilin-1, plexin-A1 bound to Rnd1. In the presence of neuropilin-1, which formed a complex with plexin-A1, the interaction of Rnd1 with plexinA1 was totally dependent on Sema3A stimulation; Rnd1 binding to plexin-A1 was hardly detected without Sema3A but induced by Sema3A (Fig. 2a). In cells stimulated with Sema3A, plexin-A1 dissociation from FARP2 and association with Rnd1 was accompanied by a downregulation of R-Ras (Figs. 1e and 2a). In contrast, neither Rnd1 binding to plexin-A1 nor downregulation of R-Ras was detected in Sema3A-stimulated cells expressing a mutant FARP2 that lacked GEF activity (FARP2(AE)), although dissociation of FARP2(AE) from plexin-A1 was observed (Fig. 2b). The requirement of Rac-GTP for Rnd1-R-Ras signaling was also supported by the inhibitory effect of dominant-negative Rac (RacT17N) on the Rnd1 binding to plexin-A1 (Supplementary Fig. 4). Furthermore, in DRG neurons, Sema3A-induced downregulation of R-Ras-GTP was abolished by either siRNA-mediated knockdown of endogenous FARP2 or expression of FARP2(AE) (Fig. 2c). These results indicate that activation of GEF activity of FARP2 is essential for the subsequent association of Rnd1 with plexin-A1 and downregulation of R-Ras in Sema3Astimulated cells. FARP2 interacts with PIPKIc661 Is the FERM domain of FARP2 necessary only for plexin-A1 binding? Previous reports that detailed the interaction between the FERM domain of talin and PIPKIg661 in cell adhesion21,22 led us to speculate that the binding of FARP2 to other molecules that interact with the FERM domain, such as PIPKIg661, may regulate integrin function in the leading edges of extending axons. To test this idea, we examined the physical interaction between FARP2 and PIPKIg661 (Fig. 3a,b). In immunoprecipitation experiments, the association between FARP2 and PIPKIg661 was detected independent of the molecule that was initially

1714

Figure 2 RacGEF activity of FARP2 is required for the Rnd1mediated reduction of R-Ras-GTP in Sema3A–plexin-A1 signaling. (a) Immunoprecipitation and immunoblot using transfected HEK293 cells shows that Rnd1 binds to plexin-A1, whereas FARP2 dissociates from plexin-A1 in either the absence of neuropilin-1 or the presence of Sema3A (50 ng ml–1). GST-fused c-Raf-1 RBD pull-down assay shows that binding of Rnd1 to plexin-A1 decreases R-Ras-GTP level. (b) Immunoprecipitation and immunoblot using HEK293 cells shows that Rnd1 cannot bind to plexin-A1, even though FARP2(AE) is dissociated from plexin-A1 in the presence of Sema3A (50 ng ml–1). GST-fused c-Raf-1 RBD pull-down assay shows that FARP2(AE) does not affect R-Ras-GTP level. (c) GST-fused c-Raf-1 RBD pull-down assay shows that in chick embryonic DRG neurons treated with FARP2 siRNA or subjected to adenovirus-mediated overexpression of FARP2(AE), the level of endogenous R-Ras-GTP was not affected by Sema3A as it was in untreated DRG neurons.

Anti-V5

Anti-Xpress

c

Blot: Anti-myc

Anti-myc IP: Anti-myc

+ + + + +

precipitated. Deletion of the FERM domain of FARP2 or the 28 Cterminal amino acids of PIPKIg661, which contain the interaction site for the FERM domain of talin21,22, abolished the association between FARP2 and PIPKIg661. Moreover, the FERM domain alone was sufficient for this association. We then examined whether or not this interaction was regulated by the binding of neuropilin-1 to Sema3A (Fig. 3c). A 30-min stimulation with Sema3A enhanced the interaction between FARP2 and PIPKIg661. We next sought to confirm our observations in vivo by examining neurons of chicken DRG, where plexin-A1, FARP2 and PIPKIg661 are expressed endogenously (Fig. 3d). From these cells, we were able to coprecipitate FARP2 with both plexin-A1 and PIPKIg661. Furthermore, consistent with the in vitro experiment, the interaction between FARP2 and plexin-A1 was suppressed, whereas FARP2-PIPKIg661 binding was enhanced by Sema3A stimulation. Notably, Sema3A stimulation also induced the dissociation of PIPKIg661 from talin (Fig. 3d). These results indicate that the binding affinity of FARP2 for plexin-A1 and PIPKIg661 are reciprocally regulated by Sema3A. We also measured the phosphatidylinositol4,5-biphosphates (PtdIns(4,5)P2)–synthesizing (kinase) activity of PIPKIg661 in transfected cells by thin layer chromatography (Fig. 3e, left). PIPKIg661 kinase activity was downregulated by Sema3A only when PIPKIg661 was coexpressed with FARP2. A similar effect was observed in chicken DRG neurons (Fig. 3e, right). Thus, FARP2 may contribute to Sema3A-mediated inhibition of integrin function by suppressing PIPKIg661. It is noteworthy that the GEF activity of FARP2 was suppressed by its interaction with PIPKIg661 (Fig. 3f), suggesting that PIPKIg661 may be also involved in downregulation of the GEF activity of FARP2 dissociated from plexin-A1. The role of FARP2 in Sema3A-mediated axonal repulsion To test whether FARP2 has a biological role in the responsiveness of neuronal axons to Sema3A proteins, we performed a collagen gel coculture using chicken DRG and Sema3A-expressing cell aggregates (Fig. 4). Sema3A repelled nerve growth factor (NGF)-dependent axonal outgrowth from untreated DRG neurons, but not that from DRG neurons infected with adenovirus encoding short hairpin RNA (shRNA) specific to chicken neuropilin-1. Axons of DRG neurons infected with adenovirus encoding shRNA against FARP2 almost completely lost their responsiveness to Sema3A. Furthermore, adenovirus-mediated expression of FARP2DC, FARP2(AE) or plexin-A1(AAA) (see Methods), which inhibited Rac activation and R-Ras downregulation after Sema3A stimulation (Supplementary Fig. 2), abolished the responses of DRG axons to Sema3A. These results were confirmed by a growth cone collapse assay (Supplementary Fig. 5). Our findings thus illustrate a critical role for FARP2, particularly for its RacGEF activity, in Sema3Ainduced repulsion. Overexpression of PIPKIg661 also significantly

VOLUME 8

[

NUMBER 12

[


ARTICLES

a

b PIPKlγ 661 Flag Kinase 28–aa tail

FARP2-V5 FARP2∆C-V5 FARP2∆N-V5 PIPKlγ 661-Flag PIPKlγ∆C-Flag

PIPKlγ∆C

+ – – + –

+ – – – +

– + – + –

– – + + – Blot:

IP: Anti-V5

Anti-V5

IP: Anti-Flag

Anti-Flag

Sema3A Neuropilin-1 Plexin-A1–myc FARP2–V5 PIPKlγ 661-Flag

– + + + +

+ + + + +

Blot: Anti-myc

IP: Anti-myc

Anti-V5 Anti-Flag

Anti-Flag

I

Anti-myc IP: Anti-V5

Anti-V5 Anti-Flag

Anti-V5

Anti-myc

Sema3A

e –

+

Blot: Anti-plexin-A1

IP: AntiFARP2

Anti-FARP2

Neuropilin-1 Plexin-A1 FARP2 PIPKIγ 661

Control + + + + + + + – + – + +

Sema3A + + + + + + + – + – + +

Anti-Flag

Anti-PIPKIγ Anti-talin

PtdP2

Anti-plexin-A1 IP: AntiPIPKIγ

Anti-FARP2

f

100 80

Original

*

60 40 20

0 FARP2-V5 PIPKlγ 661-Flag

Anti-PIPKIγ Anti-talin

Anti-V5

[3H]GDP bound (%)

d

IP: Anti-Flag

C on t Se rol m a3 A


c

+ + – +

Figure 3 FARP2 binding to PIPKIg661 reciprocally downregulates their activities. (a) Structure of wildtype PIPKIg661 and domain-lacking mutants (see Methods). (b) Immunoprecipitation and immunoblot using transfected HEK293 cells shows that FARP2 and FARP2DC, but not FARP2DN, bind to PIPKIg661, whereas PIPKIg661, but not PIPKIgDC, binds to FARP2. (c) Immunoprecipitation and immunoblot using transfected HEK293 cells shows that FARP2 binds to PIPKIg661, whereas FARP2 is dissociated from plexin-A1 in the presence of Sema3A (50 ng ml–1). (d) Immunoprecipitation and immunoblot using DRG from chick embryo shows that binding of endogenous FARP2 to endogenous PIPKIg661 is increased by Sema3A (50 ng ml–1), whereas binding of FARP2 to plexin A-1 is decreased. Sema3A also decreases the binding of PIPKIg661 to talin. (e) Thin-layer chromatographyautoradiography of PtdIns(4,5)P2 generated by PIPKIg661 in transfected HEK293 cells (left) or DRG from chick embryo (right) in response to Sema3A (50 ng ml–1). PtdIns(4,5)P2 synthesis of PIPKIg661 is suppressed by Sema3A only in the presence of FARP2. (f) The nucleotide exchange of [3H]GDP-loaded Rac caused by FARP2 in the absence (square) or presence (triangle) of PIPKIg661. Values observed for 20 min are expressed as a percentage of the values observed at t ¼ 0. *P o 0.05, compared with the value in the absence of PIPKIg661.

suppressed Sema3A-mediated repulsion, (P o 0.05) suggesting that the FARP2-PIPKIg661 pathway is involved in this effect. It is also noteworthy that adenovirus-mediated expression of intact FARP2 had a weak but significant suppressive effect on Sema3A-mediated axonal repulsion (P o 0.05). This may be due to uncontrolled Rac activation by overexpression of FARP2 (Supplementary Fig. 2). In fact, constitutively active Rac did not promote Rnd1 binding to plexin-A1 and R-Ras GAP activation; rather, it increased lamellipodia and protrusion in transfected HEK293 cells and DRG neurons (Supplementary Fig. 6)23–25. Thus, FARP2-mediated regulation of the local concentration of Rac-GTP in the vicinity of plexin-A1 may be critical for Sema3A-mediated repulsion. FARP2 regulates integrin-mediated adhesion of DRG neurons Both R-Ras and PIPKIg661 are involved in integrin-mediated cell adhesion21,22,26,27. The addition of stimulatory b1-integrin antibodies significantly blocked Sema3A-mediated axonal repulsion (P o 0.05; Fig. 4). Indeed, Sema3A has been reported to inhibit endothelial cell adhesion28. We examined the effect of Sema3A-induced signals on adhesion and spreading of DRG neurons on laminin (Fig. 5). Neurons demonstrated enhanced cell adhesion on laminin in the presence of NGF. This is consistent with the reported effect of NGF on b1-integrin accumulation at the leading edge of growth cones29,30 and on R-Ras activation (Supplementary Fig. 7). Addition of Sema3A suppressed


[

NUMBER 12

[

DECEMBER 2005

adhesion of both NGF-stimulated and unstimulated neurons to laminin (Fig. 5). Treatment with siRNA against chicken FARP2 completely abrogated these suppressive effects of Sema3A for both NGF-stimulated and unstimulated neurons, indicating an essential role for FARP2 in Sema3A-mediated regulation of neural cell adhesion. Knockdown of PIPKIg661 substantially suppressed adhesion of both NGF-stimulated and unstimulated neurons in the absence of Sema3A. A similar effect was observed in neurons overexpressing FARP2 or FARP2DC, which may be due to sequestration of PIPKIg661 from talin by the FERM domain of these molecules. Notably, knockdown of PIPKIg661 or overexpression of FARP2, which did not inhibit R-Ras GAP activity induced by Sema3A (Supplementary Fig. 2), abolished the effect of Sema3A on cell adhesion of unstimulated neurons but not of NGF-stimulated neurons. This suggests that NGF enhances neural cell adhesion independently of the PIPKIg-mediated pathway, possibly by activating R-Ras. In contrast, overexpression of FARP2DC, which inhibited R-Ras GAP activity induced by Sema3A (Supplementary Fig. 2), eliminated the effect of Sema3A on both NGF-stimulated and unstimulated neurons. In this assay system, cell adhesion of DRG neurons correlated well with the amounts of active b1-integrin after treatment with various reagents (Supplementary Fig. 8). These findings illustrate that two signaling pathways downstream of FARP2 differentially act in Sema3A-mediated regulation of integrin function and adhesion of neural cells.

DISCUSSION There are two known FERM domain–containing GEF proteins, FARP1 and FARP2, which are structurally related proteins18,31. FARP1, also known as CDEP (chondrocyte-derived ezrin-like domain containing protein), was originally isolated from differentiated chondrocytes. It functions as a RhoGEF, but its biological and functional role is unknown31. FARP2, however, exerts a GEF activity for Rac but not for Rho or Cdc42, and the ectopic expression of FARP2 in rat embryonic cortical neurons results in significantly shortened neurites and excessive growth cones18. Taken together with the fact that many FERM domain– containing molecules interact with membrane-associated proteins32, this observation has raised the possibility that FARP2 may be involved in cytoskeletal reorganization by transmitting receptor signals for guidance molecules. Here we show that FARP2 is an immediate downstream signaling molecule of the Sema3A receptor complex. FARP2 associates with plexin-A1 in the presence of neuropilin-1. Sema3A binding to neuropilin-1 seems to induce conformational change of plexin-A1 necessary for releasing FARP2. This suggests that neuropilin-1 is required not only for ligand binding, but also for signaling, by modulating the interaction of FARP2 with plexin-A1. Our present study demonstrates that FARP2 is critical for in Sema3Amediated axonal repulsion through two independent downstream signaling pathways (Supplementary Fig. 9). Sema3A-mediated

1715

ARTICLES

Control shRNA

Neuropilin-1 shRNA

FARP2 shRNA

PIPKIγ 661 shRNA

FARP2 shRNA β1 inhibit. Ab

β1 stim. Ab

Control adeno.

FARP2 adeno.

FARP2∆C adeno.

FARP2(AE) adeno.

PIPKIγ 661 adeno.

Plexin-A1(AAA) adeno.

β1 inhibit. Ab

Sema3A Control Sema3A

b Axon outgrowth ratio (P/D)

1.2

*

*

*

*

*

*

*

Control Sema3A

1.0 0.8 0.6 0.4 0.2 Ab

b

t. bi hi in

im

.A β1 -

st β1 -

PI P ad KIγ en 66 Pl o. 1 ex in ad A1 en (A o. AA )

C ad ont en rol o. FA R P2 ad en o. FA ad RP2 en ∆ o. C FA R ad P2 en (A o. E)

0 C sh ont R ro N l A N eu r o sh p R ilin N A 1 FA R P2 sh R N A PI PK sh Iγ R 66 N A 1 FA β1 RP 2 in s hi h bi RN t. Ab A


Control

a

Figure 4 FARP2 is required for Sema3A-mediated repulsion of neurite outgrowth. Chick embryo DRG neurons infected with adenovirus containing DNA templates for the synthesis of shRNAs or adenovirus encoding cDNA constructs (adeno.) were co-cultured with cell aggregates in a collagen matrix. In some experiments, the collagen matrix and culture medium contained P4G11 b1-integrin stimulatory antibodies (b1 stim. Ab) or V2E9 b1-integrin inhibitory antibodies (b1 inhibit. Ab). Explants were fixed after 48 h and stained with anti-neurofilament antibodies. (a) Representative responses of DRG to control and Sema3A-expressing cells. In these panels the explants were positioned to the right of the cell aggregates. Scale bar, 100 mm. (b) Quantification of the repulsion experiments. The repulsive activity was measured by the proximal/distal (P/D) ratio of axon outgrowth (see Methods). Data shown as mean ± s.e.m. from 20–30 explants for each condition. *P o 0.05, compared with the value in the absence of Sema3A in each condition.

dissociation of FARP2 from plexin-A1 is followed by activation of Rac by GEF activity of released FARP2, binding of Rnd1 to plexin-A1 and downregulation of R-Ras by GAP activity of plexin-A1. Downregulation of R-Ras, which is dependent on Rnd1 binding to plexin-B1, is required for Sema4D-induced growth cone collapse, and a similar mechanism has been proposed for Sema3A-plexin-A1 signaling because the constitutively active R-Ras mutant inhibits Sema3Amediated growth cone collapse19. Mutant FARP2 that lacks GEF activity acts as a dominant-negative molecule and abrogates not only Rnd1 binding but also downregulation of R-Ras. Therefore, our results indicate that the GEF activity of FARP2 and activated Rac trigger the Rac-Rnd1-R-Ras pathway of Sema3A-plexin-A1 signaling. It is still unclear, however, whether Rac-GTP directly binds to plexin-A1 (refs. 14,15,33). At the moment, we do not know how Rac-GTP is involved in Rnd1 binding and R-Ras GAP activation of plexin-A1.

1716

Upon Sema3A stimulation, FARP2 also directly interacts with PIPKIg661 via the FERM domain, which results in sequestration of PIPKIg661 from talin and downregulation of its activity. PIPKIg661 interacts with the FERM domain of talin and increases local levels of PtdIns(4,5)P2, which contributes to focal adhesion assembly34. Since R-Ras is also implicated in integrin activation26,27, both downregulation of R-Ras and inhibition of PIPKIg661 seem to contribute to Sema3A-induced axonal repulsion and growth cone collapse of DRG neurons by suppressing adhesion at the leading edge of growth cones. It is noteworthy that the Rac-Rnd1-R-Ras pathway seems to be essential for repulsion of outgrowing axons from neurons stimulated by NGF, whereas inhibition and sequestration of PIPKIg661 is necessary for suppressing the adhesion of unstimulated neurons. This may reflect a differential contribution of R-Ras and PIPKIg661 to cell adhesion that depends on the activation state of the neurons; R-Ras may be involved

VOLUME 8

[

NUMBER 12

[


ARTICLES


b

Control siRNA FARP2 siRNA PIPKIγ 661 siRNA FARP2 adeno. FARP2∆C adeno.

60

120 100

No. of attached cells

No. of attached cells

a

40

20

Control siRNA FARP2 siRNA PIPKIγ 661 siRNA FARP2 adeno. FARP2∆C adeno.

80 60 40 20

0 Control

Sema3A

0

β1 stim. Ab Sema3A & β1 stim. Ab

Control

Sema3A

β1 stim. Ab Sema3A & β1 stim. Ab

Figure 5 FARP2 and PIPKIg661 regulate the inhibitory effect of Sema3A on integrin function. DRG neurons that were either treated with rhodamine-labeled siRNAs or coinfected with adenoviruses containing various constructs and adenovirus-containing LacZ were plated on laminin-coated plates (200 neurons per well) for 30 min in culture medium (a) without or (b) with NGF (50 ng ml–1). Culture medium also contained combinations of Sema3A (50 ng ml–1), P4G11 b1-integrin stimulatory antibodies (b1 stim. Ab). After washing with culture medium, the number of rhodamine-positive cells or LacZ-positive cells attached on each well were counted. Data shown as mean ± s.e.m.

in adhesion of cells stimulated by neurotrophic factors such as NGF, whereas PIPKIg661 regulates basal adhesion of unstimulated cells. Another well-characterized semaphorin receptor, plexin-B1, transmits at least two downstream signals through the Rnd1-R-Ras and the PDZ-RhoGEF/LARG pathways2,19. The former regulates cell adhesion and the latter actin cytoskeleton reorganization. Both plexin-A1 and plexin-B1 share the Rnd1-R-Ras pathway. Although Rac-GTP binding to plexin-B1 is well established2,35, it is unclear whether Rac-GTP is involved in activation of the Rnd1-R-Ras pathway or how Rac is activated upon ligand stimulation. In this context, it is noteworthy that the cytoplasmic domain of plexin-B1 contains the FERM domain binding sequence. We could not detect the binding of FARP2 to plexinB1, although we cannot exclude the possibility that FARP2 may interact with plexin-B1 in the presence of unknown coreceptors and may participate in plexin-B1 signals by activating Rac. Alternatively, there may be a different GEF molecule that is involved in plexin-B1 signaling. Unlike plexin-B1 signals, which can activate Rho through PDZRhoGEF/LARG bound to its C-terminal region36, it is unclear how plexin-A1 signals regulate the actin cytoskeleton during axonal repulsion. Although it is likely that FARP2-mediated suppression of focal adhesion and integrin function may indirectly lead to a decreased turnover of the actin cytoskeleton, it would be interesting to know if FARP2 is directly involved in regulation of cytoskeletal dynamics. Several modulators for cytoskeletal dynamics, such as p21-activated kinase, LIM kinase and coffilin, have been implicated in Sema3Amediated growth cone collapse2. Furthermore, collapsin response mediator protein (CRMP)-2, a member of a small family of brainspecific proteins37, and MICAL, a member of the flavoprotein monooxygenases38, have been suggested to link semaphorin receptor complexes and actin cytoskeleton in Sema3A-mediated axonal repulsion. Future studies will determine if any of these molecules are regulated by FARP2. METHODS Construction of cDNAs and transfectants. Mouse Sema3A, neuropilin-1 and plexin-A1 cDNAs were prepared as described previously39. Mouse FARP2, PIPKIg661, Rnd1 and derived mutant cDNAs were synthesized by PCR. The following mutant proteins were produced: plexin-A1(AAA), plexin-A1 containing amino-acid substitutions at the potential FERM domain binding site (K1265A/R1266A/K1267A); FARP2DC, FARP2 lacking the Dbl homology


[

NUMBER 12

[

DECEMBER 2005

(DH) and Pleckstrin homology (PH) domains (Gly2 to Ile363); FARP2DN, FARP2 lacking the FERM domain (Arg362 to Leu1065); FARP2(AE), FARP2 containing amino-acid substitutions in the putative GEF domain (Q686A/L694E); and PIPKIgDC, PIPKIg661 lacking the putative FERM domain binding site (Met1 to Ile633). Constructs were ligated into the following expression vectors: pcDNA3.1/myc-HisA, pcDNA3.1/V5-HisA, pcDNA4/His.Max (Invitrogen), pFlag-CMV-3 or pFlag-CMV-4 (Sigma). The resulting plasmids were transfected into HEK293 cells by Fugene-6mediated transfection (Roche Molecular Biochemicals). A DNA fragment encoding the Ras-binding domain (RBD) of c-Raf-1 was synthesized by PCR and the construct was ligated into pGET-3X. The resultant plasmid was transformed into Escherichia coli, and c-Raf-1 RBD fused to glutathione-Stransferase (GST) was purified from bacterial lysates on a GST column.

Adenovirus-based overexpression of protein. To analyze the gain-of function of target proteins, we used an adenovirus-mediated system. For construction of adenoviral vectors containing FARP2, FARP2DC, FARP2(AE) or PIPKIg661, constructs were ligated into pShuttle2 and the expression cassette was excised and ligated into BD Adeno-X viral DNA (BD Biosciences Clontech). After transfection of adenoviral vector into HEK293 cells, adenovirus-containing medium was collected and concentrated. The efficiency of adenovirus incorporation into DRG neurons was confirmed by fluorescence microscopy and immunoblotting (Supplementary Fig. 10). RNA interference–based knockdown of protein function. To analyze the lossof-function of target protein, we used two methods of RNA interference (RNAi). In some experiments, including biochemical and cell adhesion assays, neurons on the plate were transfected with 21-nucleotide-long, doublestranded siRNA specific to chicken neuropilin-1, FARP2 or PIPKIg661 (Supplementary Table 1) using OligofectAMINE reagent (Invitrogen). In chemorepulsive experiments, DRG neurons cultured in the collagen gel were infected with adenovirus that directs the synthesis of siRNA. We constructed adenoviral vectors containing DNA templates for the synthesis of short hairpin RNAs (shRNAs), in which the sense and antisense sequences targeting chicken neuropilin-1, FARP2 or PIPKIg661 were linked with a nine-nucleotide loop (Supplementary Table 2). To accomplish this, the oligonucleotides encoding the shRNA were synthesized, and paired oligonucleotides were annealed and ligated into the pRNAT-H1.1/shuttle vector (GenScript), and the expression cassette was excised and ligated into BD Adeno-X viral DNA (BD Biosciences Clontech). After transfection of the adenoviral vector into HEK293 cells, adenovirus-containing medium was collected and concentrated. The efficiency of transfection with siRNA and infection with adenovirus into DRG neurons were confirmed by immunoblotting (Supplementary Figs. 10 and 11). GEF activity assays. GEF activity assays were carried out by the filter-binding method as described previously40. To prepare [3H]GDP-loaded GST-Rac, we incubated GST-Rho and GST-Cdc42 solutions containing 10 mM HEPES/ NaOH (pH 7.5), 100 mM NaCl, 7.5 mM EDTA, 15 mM GDP, 5.5 mM [3H]GDP (10 mCi mmol–1, Amersham Biosciences) and 12.5 mM Rho family GTPase (Cytoskeleton) for 25 min at 23 1C. The [3H]GDP-loaded GTPases were stabilized by supplementing the solution with 20 mM MgCl2. Nucleotide exchange reactions were performed at 25 1C by diluting the [3H]GDP-loaded GTPases to 4 mM in 250 ml reaction mixtures. Reaction mixtures contained FARP2 immunoprecipitates, 10 mM HEPES/NaOH (pH 7.5), 15 mM MgCl2, 100 mM NaCl, 1 mM DTT, 50 mg ml–1 BAS and 100 mM GTP. We sampled 30 ml of each reaction mixture at 0, 5, 19, 15 and 20 min and then quenched it with 1 ml of ice-cold dilution buffer containing 20 mM Tris (pH 7.5), 100 mM NaCl2 and 20 mM MgCl2. The amount of [3H]GDP bound to the GTPases was measured by filtering 0.9 ml of the quenched samples over BA85 nitrocellulose

1717

ARTICLES


filters and placing it in scintillation fluid. The percentage of [3H]GDP that remained bound at each time point for the GEF catalyzed and uncatalyzed reactions was compared to the zero time point of the uncatalyzed reaction. Rac activation assays. Because only activated GTP-bound Rac, not GDPbound Rac, binds to the CRIB domain of PAK41, Rac1 activity in vivo was measured by pull-down assays using the GST-fused CRIB domain of PAK. Rac1 activation assay kits (Upstate Biotechnology) were used in this study. Total Rac1 protein quantities were also measured by immunoblotting to ensure consistent loading. R-Ras activation assays. R-Ras activity in vivo was measured by pull-down assays using GST-fused c-Raf-1 RBD, which selectively interacts with GTPbound R-Ras42. Total R-Ras protein quantities were also measured by immunoblotting to ensure consistent loading. Lipid kinase assays. Immunoprecipitates were incubated in kinase buffer (50 mM HEPES, 30 mM NaCl, 10 mM MgCl2, 40 mM phosphatidylinositol 4-phosphate (PtdIns(4)P), 133 mM phosphoserine (PS) at pH 7.5). The reaction was started by adding 50 mM [g-32P]ATP for 10 min at 25 1C. The reaction was stopped by adding 1 N HCl, followed by a chloroform/methanol extraction (1:1) to extract the lipids. Phosphorylated lipids were separated by thin-layer chromatography using a mixture of chloroform, methanol, water and ammonium hydroxide (60:47:11:1.6) as a solvent. DRG neuron culture. DRG were dissected from embryonic day 7 (E7) to E8 chick embryos, trypsinized and triturated. Dissociated neurons were then plated on slides coated with laminin (20 mg ml–1, Gibco) in F12/DMEM containing 2 mM glutamine and 50 ng ml–1 NGF, and then incubated at 38 1C with 5% CO2. For Rac activation, R-Ras activation or lipid kinase assays, neurons were transfected with siRNA by OligofectAMINE reagent (Invitrogen). Alternatively, neurons were infected with adenoviruses containing various constructs. After 48 h, neurons were collected and cell lysates were used for assays. For the cell-matrix adhesion assay, siRNAs were labeled with rhodamine using the Label IT siRNA Tracker Intracellular Localization Kit (Mirus). Alternatively, neurons were coinfected with adenoviruses containing various constructs and adenovirus containing LacZ (molar ratio 3:1). After 48 h, neurons were trypsinized and triturated. In experiments, dissociated neurons were plated on laminin-coated slides (200 neurons per well of six-well chamber slides) for 30 min in culture medium with NGF (50 ng ml–1) or without NGF. Culture medium also contained combinations of Sema3A (50 ng ml–1), P4G11 b1-integrin stimulatory antibodies and V2E9 b1-integrin inhibitory antibodies. After five washes with culture medium, the number of rhodamine-positive cells or LacZ-positive cells was counted for each well. In vitro co-culture assays. DRG from E7–E8 chick embryos were dissected into L15 medium containing 5% fetal calf serum (FCS). The DRG and Sema3Aexpressing HEK293 cell aggregates were embedded at a distance of 200–400 mm from each other in collagen gels with F12/DMEM (1:1), 2 mM glutamine and 50 ng ml–1 NGF and incubated at 38 1C with 5% CO2 for 2 d. The explants were fixed with 4% paraformaldehyde in PBS, followed by staining using antibodies against neurofilament. For adenovirus-mediated RNAi or overexpression experiments, purified adenovirus that directs synthesis of siRNA in vivo or encodes FARP2, FARP2DC, FARP2(AE) or PIPKIg661 was applied to the DRG for 6 h. Then, the DRG were placed into the collagen gels with cell aggregates. The repulsive activity was measured by the proximal/distal (P/D) ratio of axonal outgrowth, where P is the extent of axonal outgrowth on the side proximal to the cell aggregate and D is the extent of axonal outgrowth distal to the cell aggregate. Therefore, a P/D ratio of 1 indicates no repulsion43. Antibodies, immunoprecipitation and immunohistochemistry. The following antibodies were obtained commercially: mouse monoclonal antibodies to Flag (anti-Flag) and anti-Flag bound to horseradish peroxidase (anti-FlagHRP) (Sigma), anti-V5 and anti–V5-HRP (Invitrogen), anti-myc and antimyc-HRP (Invitrogen), anti-Xpress and anti-Xpress-HRP (Invitrogen), goat polyclonal anti-PIPKIg and rabbit polyclonal anti-R-Ras (Santa Cruz), and mouse monoclonal b1-integrin antibodies P4G11 and V2E9 (Chemicon). Rabbit polyclonal anti-plexin-A1 and anti-FARP2 were produced by Kitayama

1718

Laboratories. Immunoprecipitation and immunohistochemistry were performed according to standard protocols. Accession codes. BIND identifiers (http://bind.ca): 335806–335816. Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS We thank K. Kubota for her assistance. This study was supported by research grants from the Ministry of Education, Science and Culture, Japan (to T.T., A.K. and H.K). COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/

1. Tessier-Lavigne, M. & Goodman, C.S. The molecular biology of axon guidance. Science 274, 1123–1133 (1996). 2. Pasterkamp, R.J. & Kolodkin, A.L. Semaphorin junction: making tracks toward neural connectivity. Curr. Opin. Neurobiol. 13, 79–89 (2003). 3. Maestrini, E. et al. A family of transmembrane proteins with homology to the MET-hepatocyte growth factor receptor. Proc. Natl. Acad. Sci. USA 93, 674–678 (1996). 4. Godenschwege, T.A., Hu, H., Shan-Crofts, X., Goodman, C.S. & Murphey, R.K. Bidirectional signaling by Semaphorin 1a during central synapse formation in Drosophila. Nat. Neurosci. 5, 1294–1301 (2002). 5. Toyofuku, T. et al. Guidance of myocardial patterning in cardiac development by Sema6D reverse signalling. Nat. Cell Biol. 6, 1204–1211 (2004). 6. Suter, D.M. & Forscher, P. Substrate-cytoskeletal coupling as a mechanism for the regulation of growth cone motility and guidance. J. Neurobiol. 44, 97–113 (2000). 7. Luo, L. Rho GTPases in neuronal morphogenesis. Nat. Rev. Neurosci. 1, 173–180 (2000). 8. Aurandt, J., Vikis, H.G., Gutkind, J.S., Ahn, N. & Guan, K.L. The semaphorin receptor plexin-B1 signals through a direct interaction with the Rho-specific nucleotide exchange factor, LARG. Proc. Natl. Acad. Sci. USA 99, 12085–12090 (2002). 9. Swiercz, J.M., Kuner, R., Behrens, J. & Offermanns, S. Plexin-B1 directly interacts with PDZ-RhoGEF/LARG to regulate RhoA and growth cone morphology. Neuron 35, 51–63 (2002). 10. Perrot, V., Vazquez-Prado, J. & Gutkind, J.S. Plexin B regulates Rho through the guanine nucleotide exchange factors leukemia-associated Rho GEF (LARG) and PDZ-RhoGEF. J. Biol. Chem. 277, 43115–43120 (2002). 11. Jin, Z. & Strittmatter, S.M. Rac1 mediates collapsin-1-induced growth cone collapse. J. Neurosci. 17, 6256–6263 (1997). 12. Kuhn, T.B., Brown, M.D., Wilcox, C.L., Raper, J.A. & Bamburg, J.R. Myelin and collapsin-1 induce motor neuron growth cone collapse through different pathways: inhibition of collapse by opposing mutants of rac1. J. Neurosci. 19, 1965–1975 (1999). 13. Vastrik, I., Eickholt, B.J., Walsh, F.S., Ridley, A. & Doherty, P. Sema3A-induced growthcone collapse is mediated by Rac1 amino acids 17–32. Curr. Biol. 9, 991–998 (1999). 14. Turner, L.J., Nicholls, S. & Hall, A. The activity of the plexin-A1 receptor is regulated by Rac. J. Biol. Chem. 279, 33199–33205 (2004). 15. Zanata, S.M., Hovatta, I., Rohm, B. & Puschel, A.W. Antagonistic effects of Rnd1 and RhoD GTPases regulate receptor activity in Semaphorin 3A-induced cytoskeletal collapse. J. Neurosci. 22, 471–477 (2002). 16. Oinuma, I., Katoh, H., Harada, A. & Negishi, M. Direct interaction of Rnd1 with Plexin-B1 regulates PDZ-RhoGEF-mediated Rho activation by Plexin-B1 and induces cell contraction in COS-7 cells. J. Biol. Chem. 278, 25671–25677 (2003). 17. Legg, J.W. & Isacke, C.M. Identification and functional analysis of the ezrin-binding site in the hyaluronan receptor, CD44. Curr. Biol. 8, 705–708 (1998). 18. Kubo, T. et al. A novel FERM domain including guanine nucleotide exchange factor is involved in Rac signaling and regulates neurite remodeling. J. Neurosci. 22, 8504–8513 (2002). 19. Oinuma, I., Ishikawa, Y., Katoh, H. & Negishi, M. The Semaphorin 4D receptor PlexinB1 is a GTPase activating protein for R-Ras. Science 305, 862–865 (2004). 20. Oinuma, I., Katoh, H. & Negishi, M. Molecular dissection of the semaphorin 4D receptor plexin-B1-stimulated R-Ras GTPase-activating protein activity and neurite remodeling in hippocampal neurons. J. Neurosci. 24, 11473–11480 (2004). 21. Ling, K., Doughman, R.L., Firestone, A.J., Bunce, M.W. & Anderson, R.A. Type-I gamma phosphatidylinositol phosphate kinase targets and regulates focal adhesions. Nature 420, 89–93 (2002). 22. Di Paolo, G. et al. Recruitment and regulation of phosphatidylinositol phosphate kinase type 1 gamma by the FERM domain of talin. Nature 420, 85–89 (2002). 23. Kozma, R., Sarner, S., Ahmed, S. & Lim, L. Rho family GTPases and neuronal growth cone remodelling: relationship between increased complexity induced by Cdc42Hs, Rac1, and acetylcholine and collapse induced by RhoA and lysophosphatidic acid. Mol. Cell. Biol. 17, 1201–1211 (1997).

VOLUME 8

[

NUMBER 12

[



ARTICLES 24. Gallo, G. & Letourneau, P.C. Axon guidance: GTPases help axons reach their targets. Curr. Biol. 8, R80–R82 (1998). 25. Lin, M.Z. & Greenberg, M.E. Orchestral maneuvers in the axon: trio and the control of axon guidance. Cell 101, 239–242 (2000). 26. Zhang, Z., Vuori, K., Wang, H., Reed, J.C. & Ruoslahti, E. Integrin activation by R-ras. Cell 85, 61–69 (1996). 27. Keely, P.J., Rusyn, E.V., Cox, A.D. & Parise, L.V. R-Ras signals through specific integrin alpha cytoplasmic domains to promote migration and invasion of breast epithelial cells. J. Cell Biol. 145, 1077–1088 (1999). 28. Serini, G. et al. Class 3 semaphorins control vascular morphogenesis by inhibiting integrin function. Nature 424, 391–397 (2003). 29. Grabham, P.W. & Goldberg, D.J. Nerve growth factor stimulates the accumulation of beta1 integrin at the tips of filopodia in the growth cones of sympathetic neurons. J. Neurosci. 17, 5455–5465 (1997). 30. Grabham, P.W., Foley, M., Umeojiako, A. & Goldberg, D.J. Nerve growth factor stimulates coupling of beta1 integrin to distinct transport mechanisms in the filopodia of growth cones. J. Cell Sci. 113, 3003–3012 (2000). 31. Koyano, Y. et al. Molecular cloning and characterization of CDEP, a novel human protein containing the ezrin-like domain of the band 4.1 superfamily and the Dbl homology domain of Rho guanine nucleotide exchange factors. Biochem. Biophys. Res. Commun. 241, 369–375 (1997). 32. Mangeat, P., Roy, C. & Martin, M. ERM proteins in cell adhesion and membrane dynamics. Trends Cell Biol. 9, 187–192 (1999). 33. Rohm, B., Rahim, B., Kleiber, B., Hovatta, I. & Puschel, A.W. The semaphorin 3A receptor may directly regulate the activity of small GTPases. FEBS Lett. 486, 68–72 (2000).


[

NUMBER 12

[

DECEMBER 2005

34. Nayal, A., Webb, D.J. & Horwitz, A.F. Talin: an emerging focal point of adhesion dynamics. Curr. Opin. Cell Biol. 16, 94–98 (2004). 35. Vikis, H.G., Li, W. & Guan, K.L. The plexin-B1/Rac interaction inhibits PAK activation and enhances Sema4D ligand binding. Genes Dev. 16, 836–845 (2002). 36. Li, Z., Aizenman, C.D. & Cline, H.T. Regulation of rho GTPases by crosstalk and neuronal activity in vivo. Neuron 33, 741–750 (2002). 37. Goshima, Y., Nakamura, F., Strittmatter, P. & Strittmatter, S.M. Collapsin-induced growth cone collapse mediated by an intracellular protein related to UNC-33. Nature 376, 509–514 (1995). 38. Terman, J.R., Mao, T., Pasterkamp, R.J., Yu, H.H. & Kolodkin, A.L. MICALs, a family of conserved flavoprotein oxidoreductases, function in plexin-mediated axonal repulsion. Cell 109, 887–900 (2002). 39. Toyofuku, T. et al. Dual roles of Sema6D in cardiac morphogenesis through regionspecific association of its receptor, Plexin-A1, with off-track and vascular endothelial growth factor receptor type 2. Genes Dev. 18, 435–447 (2004). 40. Zheng, Y., Hart, M.J. & Cerione, R.A. Guanine nucleotide exchange catalyzed by dbl oncogene product. Methods Enzymol. 256, 77–84 (1995). 41. Sander, E.E. et al. Matrix-dependent Tiam1/Rac signaling in epithelial cells promotes either cell-cell adhesion or cell migration and is regulated by phosphatidylinositol 3-kinase. J. Cell Biol. 143, 1385–1398 (1998). 42. van Triest, M., de Rooij, J. & Bos, J.L. Measurement of GTP-bound Ras-like GTPases by activation-specific probes. Methods Enzymol. 333, 343–348 (2001). 43. Chen, H., He, Z., Bagri, A. & Tessier-Lavigne, M. Semaphorin-neuropilin interactions underlying sympathetic axon responses to class III semaphorins. Neuron 21, 1283–1290 (1998).

1719


ARTICLES

NMDA receptors regulate developmental gap junction uncoupling via CREB signaling Harsha Arumugam1, Xinhuai Liu1, Paul J Colombo2, Roderick A Corriveau3 & Andrei B Belousov1 Signaling through gap junctions (electrical synapses) is important in the development of the mammalian central nervous system. Abundant between neurons during postnatal development, gap junction coupling subsequently decreases and remains low in the adult, confined to specific subsets of neurons. Here we report that developmental uncoupling of gap junctions in the rat hypothalamus in vivo and in vitro is associated with a decrease in connexin 36 (Cx36) protein expression. Both developmental gap junction uncoupling and Cx36 downregulation are prevented by the blockade of NMDA glutamate receptors, action potentials and the calcium–cyclic AMP response element binding protein (CREB), and are accelerated by CREB overexpression. Developmental gap junction uncoupling and Cx36 downregulation are not affected by blockade of non-NMDA glutamate receptors, and do not occur in hypothalamic neurons from NMDA receptor subunit 1 (NMDAR1) knockout mice. These results demonstrate that NMDA receptor activity contributes to the developmental uncoupling of gap junctions via CREB-dependent downregulation of Cx36.

Gap junctions are widespread in the developing mammalian CNS and have been implicated in many developmental events, including neurogenesis1, neuronal differentiation2, cell death3, cell migration4, synaptogenesis and neural circuit formation5–11. The passage of Ca2+, inositol 1,4,5-trisphosphate (IP3), cyclic AMP (cAMP) and other signaling molecules through gap junctions may coordinate metabolic and transcriptional activities in developing neurons9,12–14. Gap junctions also contribute to a primitive form of synchronized spontaneous activity that is a hallmark of the developing brain15,16. This networkdriven activity often involves cooperation between gap junctions and chemical synapses such as cholinergic synapses in the retina9,17,18, cholinergic and GABA synapses in the spinal cord19 and GABA synapses in the hippocampus15,20. The incidence of gap junction coupling decreases during postnatal development. In the spinal cord6,21,22, hippocampus20, neocortex23,24 and striatum25 of the rat, and the visual cortex of the ferret26, uncoupling occurs during postnatal weeks 1 to 3 and thus overlaps with the major period of chemical synapse formation and increased synaptic activity26. Here we tested the hypothesis that the maturation of glutamatergic transmission is responsible for developmental gap junction uncoupling. The model system for the present study is the hypothalamus, which expresses gap junctions and is critical for homeostatic regulation and coordination of cardiovascular, nervous and endocrine functions27. We demonstrate that NMDA receptor–mediated glutamatergic transmission is required for developmental gap junction uncoupling via Ca2+-dependent signal transduction pathways and CREB-dependent regulation of Cx36 expression.

RESULTS Developmental gap junction uncoupling in vivo Developmental changes in neuronal gap junction coupling were determined in acute slices of the paraventricular nucleus (PVN) and the supraoptic nucleus (SON) of the rat hypothalamus. We used the coupling tracer neurobiotin, which passes through gap junctions, and the dextran Alexa Fluor 594, which is gap junction impermeable (Fig. 1a–c). In parallel experiments, we used western blots on dissected hypothalamus to study developmental changes in the expression of Cx36, a gap junction protein that is neuron-specific and essential for functional gap junctions in the hypothalamus28–30. The incidence of dye coupling—that is, the percentage of primary-labeled neurons coupled to one or more secondary-labeled neurons—and the expression of Cx36 both increased during the first two weeks of postnatal hypothalamic development (Supplementary Fig. 1 online). This change indicates an initial developmental increase in neuronal gap junction coupling between the time of birth and postnatal day 15 (P15). The incidence of dye coupling and the expression of Cx36 decreased between P15 and P30 (Fig. 1d–f and Table 1), indicating subsequent developmental gap junction uncoupling. Chronic administration in vivo of the NMDA receptor antagonist dizocilpine (MK-801) (0.5–1.5 mg kg1 subcutaneously every 24 h from P1 to P30) suppressed this developmental uncoupling and Cx36 downregulation (Fig. 1d–f and Table 1). The initial increases (from P1 to P15) in gap junction coupling and Cx36 expression were not affected by MK-801 (Fig. 1d–f; saline versus MK801 on P15). In contrast to Cx36, levels of Cx43 (a presumptive glial connexin28) increased in the hypothalamus between P15 and P30, and this increase was not affected by chronic MK-801 (Fig. 1g,h).

1Department of Cell and Molecular Biology and 2Department of Psychology, Tulane University, New Orleans, Louisiana 70118, USA. 3Department of Cell Biology and Anatomy, Louisiana State University Health Sciences Center, New Orleans, Louisiana 70112, USA. Correspondence should be addressed to A.B.B. ([email protected]).


1720

VOLUME 8

[

NUMBER 12

[


b

c

d

Incidence of dye coupling

a

20 µm

P15

g

P30

f

0

P = 0.024

Saline MK-801

P15

P15

P30

h

P = 0.033

n = 8 per data point

P30

[

treated with AP5 (100 mM) and NMDAR1 knockout non-treated (genotyped by polymerase chain reaction (PCR) and verified by Ca2+ imaging; Supplementary Fig. 2 online). Developmental gap junction uncoupling and Cx36 downregulation were prevented in both AP5treated and knockout cultures (Fig. 3 and Table 1).

Cx43 1.0

P = 0.001

0.8 0.6 0.4 0.2 0.0

P = 0.002

Saline n = 4 per MK-801 data point

P15

P30

Developmental gap junction uncoupling in vitro Primary neuronal cultures express gap junctions31 and provide a useful model for studying the mechanisms of gap junction regulation. Using neurobiotin and dextran Alexa Fluor 594 (Fig. 2a), we first tested whether the uncoupling of gap junctions also occurs in rat hypothalamic cultures and is prevented by chronic inactivation of NMDA glutamate receptors. In non-treated (control) cultures, the incidence of dye coupling decreased between day in vitro 16 (16DIV) and 32DIV (Fig. 2b and Table 1), indicating developmental uncoupling of gap junctions in the hypothalamus in vitro. Western blots demonstrated a parallel decrease in the expression of Cx36 (Fig. 2c,d). We detected no differences in the incidence of dye coupling or the expression of Cx36 on 16DIV between the control cultures and cultures chronically (starting on 4DIV) incubated in DL-2-amino-5-phosphonovalerate (AP5, 100 mM), a competitive NMDA glutamate receptor antagonist (Fig. 2b–d and Table 1). This suggests that, as in vivo, the initial increase in gap junction coupling that naturally occurs in the developing primary cultures (Supplementary Fig. 1) is not affected by NMDA receptor blockade. Further, the developmental uncoupling of gap junctions and Cx36 downregulation were prevented by the chronic AP5 treatment. As in the hypothalamus in vivo, an increase in the expression of the putative glial gap junction protein Cx43 was detected in neuronal cultures, and Cx43 levels were not affected by the NMDA receptor blockade (Fig. 2e,f). Taken together, the in vivo and in vitro data suggest that developmental gap junction uncoupling in hypothalamic neurons (from P15 to P30) is mediated by NMDA receptor– dependent downregulation of Cx36, but that an initial developmental increase in gap junction coupling (from P1 to P15) is NMDA receptor independent. In addition, we studied gap junctions in hypothalamic cultures prepared from NMDAR1 knockout mice and their wild-type littermates. Three culture groups were tested on 16DIV and 32DIV: wildtype non-treated (control), wild-type chronically (starting on 4DIV)


Figure 1 Chronic NMDA receptor blockade by MK-801 reduces developmental gap junction uncoupling and Cx36 downregulation in the hypothalamus in vivo. (a–c) Images of (a) neurobiotin (green) and (b) dextran Alexa 594 (red) staining in an SON slice (P30, MK-801-treated) are superimposed in c. Yellow indicates dye colocalization in the primary-labeled neuron. Arrows indicate secondary-labeled neurons. (d) Incidence of dye coupling (21–32 primary-labeled neurons per data point (also see Table 1)). Statistical significance was calculated using the Fisher’s exact probability test; data for SON and PVN are combined. (e–h) Expression of (e,f) Cx36 and (g,h) Cx43 in the hypothalamus. Here and for all other western blots, optical density signals are normalized relative to actin, and normalized values are compared (mean ± s.e.m.) to controls (set at 1.0): in this case, P30 saline-treated samples. Statistical analyses were performed using the paired Student’s t-test.

P30

Actin

P = 0.003

P15

P = 0.039

5

Cx43

1.5

0.5

10

Actin

Cx36

Saline MK-801

15

Cx36

2.0

1.0

20

Saline MK-801 Saline MK-801

Saline MK-801 Saline MK-801

Normalized optical density

e



ARTICLES

NUMBER 12

[

DECEMBER 2005

Mechanisms of developmental gap junction uncoupling To elucidate signal transduction mechanisms for gap junction uncoupling during neuronal development, we applied additional drugs in rat hypothalamic cultures from 16DIV to 32DIV, and used dye coupling (Fig. 4a) and Cx36 western blot (Fig. 4b) analyses. Gap-junction coupling and Cx36 expression remained high in 32DIV cultures treated with MK-801 (20 mM; NMDA receptor channel blocker), KN-62 (2.5 mM; inhibitor of Ca2+/calmodulin–dependent protein kinases II and IV (CaMKII/IV)), GF 109203X (500 nM; inhibitor of protein kinase C (PKC)), and a combination of AP5 (100 mM), atropine (50 mM; muscarinic acetylcholine receptor antagonist) and

Table 1 Dye coupling in the hypothalamus in vivo and in vitro

Total number of

Number of dye-coupled

primary-labeled neurons

primary-labeled neurons (%)

1

2

3

4+

Slice, ratii P15, saline P15, MK-801

21 25

4 (19.0), P ¼ 0.024 5 (20.0), P ¼ 0.013

2 5

2 –

– –

– –

P30, saline P30, MK-801

32 27

0 (0) 4 (14.8), P ¼ 0.039

– 3

– 1

– –

– –

16DIV, control 16DIV, AP5

21 18

6 (28.6), P ¼ 0.005 5 (27.8), P ¼ 0.009

4 3

1 2

– –

1 –


40 20

1 (2.5) 5 (25.0), P ¼ 0.013

1 3

– 1

– 1

– –


16 15

4 (25.0), P ¼ 0.035 4 (26.7), P ¼ 0.029

3 2

1 1

– –

– 1

16DIV, NMDAR1 knockout

21

5 (23.8), P ¼ 0.049

4

1

–

–


19 17

0 (0) 4 (23.5), P ¼ 0.040

– 2

– 1

– 1

– –

32DIV, NMDAR1 knockout

15

4 (26.7), P ¼ 0.029

3

1

–

–

Conditions

Coupling indexi

Culture, ratiii

Culture, mouseiv

iCoupling

index is the number of secondary-labeled neurons coupled to the primary-labeled neuron. ii–ivStatistical significance was calculated using the Fisher’s exact probability test relative to P30 salineii and 32DIV controlsiii,iv.

1721

20 µm

30 25 20 15

P = 0.013

10 5 0

P = 0.005 Control AP5

d

32DIV

AP5

Control

P = 0.006

1.5 1.0

P = 0.001 Control AP5

16DIV

32DIV

AP5


32DIV

Control

AP5

Cx36

Cx43

Actin

Actin

f

2.0

32DIV

16DIV

Control

AP5

Cx36

2.5

0.5

e


16DIV Control



16DIV

c

Cx43 P = 0.014

1.0 0.8

P = 0.012

0.6 0.4

Control AP5

16DIV


32DIV

mecamylamine (50 mM; nicotinic acetylcholine receptor antagonist) (AP5 + A + M in Fig. 4). The results provide further evidence that NMDA receptor function is required for gap junction uncoupling in developing neurons and suggest that CaMKII/IV and PKC act downstream of NMDA receptors in downregulating gap junctions. Moreover, we found that function of acetylcholine receptors was not required for maintaining high levels of gap junction coupling due to NMDA receptor blockade (Discussion). No difference in dye coupling and Cx36 expression were detected between the control cultures and cultures treated with 6-cyano-7nitroquinoxaline-2,3-dione (CNQX) (10 mM; non-NMDA receptor antagonist), nifedipine (20 mM; L-type voltage-gated Ca2+ channel blocker), H89 (1 mM; blocker of cAMP-dependent protein kinase, (PKA)), PD98059 (30 mM; blocker of extracellular signal-regulated kinase/mitogen-activated protein kinase (ERK/MAPK)), or a combination of AP5 (100 mM) and KCl (20 mM; KCl causes cell depolarization and Ca2+ influx through voltage-gated Ca2+ channels32). These results suggest that non-NMDA ionotropic glutamate receptors, L-type voltage-gated Ca2+ channels, PKA and ERK/MAPK are not required for developmental uncoupling of gap junctions. In addition, even though L-type voltage-gated Ca2+ channels did not seem to be essential for downregulation of gap junctions, depolarization by KCl and, presumably, excess entry of Ca2+ through voltage–gated Ca2+ channels was sufficient to compensate for the loss of NMDA receptor function by facilitating developmental uncoupling of gap junctions (AP5 + KCl; Discussion). We also tested the importance of action potentials for the regulation of gap junction uncoupling during development. Both developmental gap junction uncoupling and Cx36 downregulation were prevented in cultures treated with tetrodotoxin (TTX) (2 mM; voltage-gated sodium channel blocker; Fig. 4a,b), suggesting that developmental uncoupling of gap junctions is action potential dependent. Further, the incidence of dye coupling and expression of Cx36 were significantly lower in cultures treated with TTX from 4DIV to 16DIV than in the corresponding controls (Supplementary Fig. 3 online), suggesting that the initial developmental increase in gap junction coupling and Cx36 expression in culture is also action potential dependent.

1722

Figure 2 Chronic NMDA receptor blockade by AP5 prevents developmental gap junction uncoupling and Cx36 downregulation in hypothalamic neurons in vitro. (a) Neurobiotin (green), dextran Alexa 594 (red) and bright field (32DIV, AP5-treated culture). Yellow indicates neurobiotin and dextran Alexa 594 colocalization in the primary-labeled neuron. Arrow indicates a secondary-labeled neuron. (b) Incidence of dye coupling (18–40 primarylabeled neurons per data point; also see Table 1). Statistical significance was calculated using the Fisher’s test. (c–f) Expression of (c,d) Cx36 and (e,f) Cx43. Optical density signals are normalized relative to actin, and normalized values are compared (mean ± s.e.m.) to 32DIV controls (set at 1.0). Statistical analyses were performed using the paired Student’s t-test.

The role of CREB In the nervous system, CREB is regulated by activation of NMDA receptors and is a target for phosphorylation by PKC and CaMKII/IV (refs. 33–35). Because NMDA receptors, PKC and CaMKII/IV are all required for the developmental uncoupling of gap junctions, we set out to determine whether, in developing hypothalamic neurons, NMDA receptor function regulates CREB and whether CREB regulates gap junctions. In control cultures (32DIV), a 20 min application of NMDA (20 mM) increased the amount of phosphorylated CREB (Ser133) detected on western blots (Fig. 5a). This increase in phosphorylation was significantly reduced by the CaMKII/IV and PKC inhibitors KN-62 (5 mM) and GF 109203X (1 mM), respectively. Overexpression of CREB using a herpes simplex virus CREB vector (HSV-CREB36) accelerated developmental gap junction uncoupling and Cx36 downregulation in cultures between 16DIV and 19DIV (Fig. 5b,c). Moreover, overexpression of a dominant-negative mutant form of CREB (HSV-mCREB; Ser133Ala36; 24DIV–28DIV) and chronic administration of CREB antisense oligodeoxynucleotide (CREB ODN37; 16DIV–32DIV) both prevented developmental gap junction uncoupling and Cx36 downregulation (Fig. 5d–g). After infection, cultured neurons were maintained with CREB viral vectors for 4 d, a time window during which cell survival is not affected by viral vectors (Methods). The relatively short time of the treatment may be why the observed increase in the incidence of dye coupling in

a

b

30 25 20 P = 0.040 (AP5) P = 0.029 (Knockout)

15 10

Control AP5 Knockout

5 0

P = 0.035

16DIV

c

AP5


2.25 2.00 1.75

P = 0.002 (AP5) P = 0.008 (Knockout)

1.50

Control AP5 Knockout

1.25 1.00

32DIV

P = 0.005

16DIV

32DIV

32DIV

16DIV Control


b


a


ARTICLES

Knockout Control

AP5

Knockout Cx36

Actin

Figure 3 NMDAR1 knockout prevents developmental gap junction uncoupling and Cx36 downregulation. (a) Incidence of dye coupling (15–21 primary-labeled neurons per data point; also see Table 1). Statistical significance was calculated using the Fisher’s test. (b,c) Expression of Cx36. Optical density values were normalized to actin and compared (mean ± s.e.m.) to 32DIV controls (set at 1.0). Statistical analyses were done using the paired Student’s t-test.

VOLUME 8

[

NUMBER 12

[


P = 0.005


1.5 1.0 0.5

a 3


2

1

P = 0.007 P = 0.002 P = 0.001

NMDA + GF 109203X

0 NMDA

DISCUSSION Transient coupling of large groups of neurons by gap junctions is a general phenomenon of the developing mammalian nervous system. An early increase in gap junction coupling suggests the importance of electrical synapses for a number of developmental events1–11. Subsequent uncoupling of gap junctions marks the transition of the CNS from an immature structure with a primitive form of synchronized network-driven spontaneous activity (which includes cooperation

between immature chemical synapses and gap junctions) to a mature structure in which chemical synaptic transmission predominates. Part of this transition is the reduction of gap junctions from a common mode of communication between different types of immature neurons (for example, pyramidal excitatory neurons and GABAergic interneurons in the neocortex)38 to a sparingly used mechanism for electrical and metabolic coupling among highly specific populations of mature neurons (for example, subtypes of GABA-releasing interneurons in the

Control

HSV-mCREB–treated cultures did not reach statistical significance (Fig. 5e). However, the change in the expression of Cx36 in these cultures was statistically significant, as were the changes in dye coupling and Cx36 expression in HSV-CREB–treated (Fig. 5c) and CREB ODN– treated (Fig. 5g) cultures. The optical density for Cx36 expression (normalized to actin and compared to the corresponding control, control value ¼ 1) was 0.58 ± 0.09 in HSV-CREB–treated cultures (P ¼ 0.040), 1.77 ± 0.12 in HSV-mCREB–treated cultures (P ¼ 0.024), and 1.81 ± 0.12 in CREB ODN–treated cultures (P ¼ 0.019; mean ± s.e.m., n ¼ 3; statistical analyses by paired Student’s t-test). Coupling indices in viral vector and CREB ODN experiments ranged from 1 to 4, with most values being 1 or 2, as in previous experiments (Table 1). The results suggest that NMDA receptors downregulate gap junction coupling in developing hypothalamic neurons by causing an increase in intracellular Ca2+ that activates PKC and CaMKII/IV; this in turn leads to phosphorylation (and activation) of CREB, followed by CREBdependent changes in gene expression, and down regulation of Cx36.


on

K-

tro l 80 G KN 1 F 1 -6 AP 092 2 5 03 + X A + C M N NQ ife X di pi ne H -8 P9 9 AP 805 9 5 + KC I TT X

0.0

NMDA + KN-62

P = 0.033

P = 0.012

P = 0.014

P = 0.044

2.0

M

M

K-

tro l 80 1 G KN F 1 -6 AP 092 2 5 03 + X A + C M N NQ ife X di pi ne H -8 P9 9 AP 80 5 59 + KC I TT X

0


1/14

1/16

1/15

1/14

1/40

5

1/14

15

2.5

C

4/19, P = 0.033

20

10

Figure 4 Signal transduction pathways in developmental gap junction uncoupling. (a,b) Pharmacological manipulations were performed on 16DIV– 32DIV followed by assessment of (a) incidence of dye coupling with neurobiotin and dextran Alexa 594 and (b) the expression of Cx36 by western blot. In a, the number of dye-coupled, primarylabeled neurons of the total number of primarylabeled neurons and the statistical significance (relative to untreated 32DIV control; Fisher’s test) are shown. Coupling indices ranged from 1 to 4, with most values being 1 or 2. In b, optical density values were normalized to actin and compared (mean ± s.e.m.) to untreated 32DIV controls (set at 1.0). Statistical analyses were performed using the paired Student’s t-test.

b

4/17, P = 0.024

4/18, P = 0.029

4/14, P = 0.013

25

on


30

C

P-CREB Actin

b

d Control HSV-CREB 19DIV 16DIV–19DIV

f

Control HSV-mCREB 28DIV 24DIV–28DIV

Control CREB ODN 19DIV 6 h 19DIV CREB


[

NUMBER 12

[

DECEMBER 2005

Actin

e

g

25

25

20

20

15

15

10

3/17

15

10 1/31 P = 0.037

5 0

5

10 1/20

5

0

1/40

0 HSV-mCREB 24DIV–28DIV

HSV-CREB 16DIV–19DIV

4/20 P = 0.038

20

CREB ODN 16DIV–32DIV

6/25

Control 28DIV

25

Control 32DIV

c

Control 19DIV

Figure 5 Developmental gap junction uncoupling and Cx36 downregulation are mediated via CREB-dependent mechanisms. (a) Relative amounts of phosphorylated CREB in DIV32 control culture following a 20 min treatment with NMDA (20 mM), NMDA plus KN-62 (5 mM) and NMDA plus GF 109203X (1 mM). The intensity values were normalized to actin and compared (mean ± s.e.m.) to controls (set at 1.0). Statistical determinations are by a one-way ANOVA with post-hoc Tukey. (b,c) Treatment with HSVCREB (16DIV–19DIV) (b) increases the amount of CREB and (c) accelerates gap junction uncoupling and Cx36 downregulation. (d,e) Treatment with HSV-mCREB (DIV24–DIV28) (d) increases the expression of mutant CREB and (e) suppresses developmental decrease in gap junction coupling and Cx36 expression. (f) A 6 h treatment with CREB ODN (on DIV19) decreases CREB; (g) the chronic CREB ODN treatment (16DIV–32DIV) prevents developmental gap junction uncoupling and Cx36 downregulation. Primary antibodies specific for (a) phosphorylated CREB and (b,d,f) for all forms of CREB were used. In b,d and f, western blots are representative of three experiments. Incidence of dye coupling (upper panels) and Cx36 western blots (lower panels) are shown in c,e and g. The number of dye-coupled, primary-labeled neurons of the total number of primary-labeled neurons and statistical significance (Fisher’s test) are indicated.



a

5/24, P = 0.025

ARTICLES

Cx36 Actin

1723


ARTICLES cerebral cortex, hippocampus, thalamus, striatum and cerebellum)39,40. Therefore, understanding the mechanisms of developmental gap junction uncoupling is an important step in understanding the events that are responsible for redistribution of gap junction coupling from wide groups of neurons to specific neuronal populations. Here we report that developmental uncoupling of gap junctions and Cx36 downregulation in the hypothalamus are prevented by inactivation of NMDA receptors, PKC, CaMKII/IV and CREB. An increase in CREB activity accelerated gap junction uncoupling and Cx36 decrease. Thus, the data suggest that the uncoupling is mediated by NMDA receptor function through PKC-, CaMKII/IV- and CREB-dependent signaling mechanisms that downregulate Cx36. Phosphorylation of CREB by Ca2+-dependent protein kinases (on serine 133 and perhaps on other sites) causes the formation of a CREB dimer that binds to a Ca2+/cAMP response element (Ca/CRE) on the promoter of CREB target genes33–35. Our search in the National Center for Biotechnology Information (NCBI) GenBank database revealed the core Ca/CRE sequence (CGTCA) in the 5¢ regulatory region of the Cx36 mouse and rat genes, suggesting that Cx36 expression may be directly regulated by CREB. Blockade of voltage-gated Na+ channels with TTX also prevented developmental gap junction uncoupling and the downregulation of Cx36, suggesting that action potential–dependent synaptic glutamate release is involved in these regulatory mechanisms. Moreover, the early developmental increase in gap junction coupling was also prevented by TTX, but was not affected by NMDA receptor blockade. Therefore, NMDA receptor–independent, action potential–dependent mechanisms control the initial upregulation of gap junctions. Finally, PKA and ERK/MAPK, both of which can be regulated by calcium34,35,41, are not required for gap junction uncoupling during development. Blockers of L-type voltage-gated Ca2+ channels, one of the two major pathways for Ca2+ entry into neurons (NMDA receptors form the other)33,34, did not prevent developmental gap junction uncoupling in cultured hypothalamic neurons. However, the influx of Ca2+ through voltage-gated Ca2+ channels (induced by KCl depolarization) downregulated gap junction coupling in the absence of NMDA receptor function. This effect probably reflects compensation for the loss of NMDA receptor function by higher than normal levels of Ca2+ influx through voltage-gated channels, resulting in downregulation of Cx36. MK-801 is widely used to block NMDA receptor function in vivo because it passes freely through the blood-brain barrier. Our finding— that chronic MK-801 treatment in vivo prevents developmental gap junction uncoupling among rat hypothalamic neurons—agrees with results of a similar in vivo experiment demonstrating that transient (6–7 d) administration of MK-801 increases the incidence of gap junction coupling among developing rat spinal motor neurons22. However, MK-801 can cause side effects that limit the interpretation of such studies. These include motor disturbances, weight loss, psychotic symptoms and memory impairment42,43. Therefore, in addition to MK-801 in vivo, we performed cell culture studies using both pharmacological NMDA receptor blockade (AP5, MK-801) and genetic deletion of NMDAR1, so as to solidify the conclusion that NMDA receptor function is required for downregulation of gap junction coupling. We found that all four manipulations prevented gap junction uncoupling, effectively ruling out non-specific effects as an explanation for the results. The present results together with our previous study44 demonstrate that chronic inactivation of NMDA receptors prevents developmental gap junction uncoupling and induces cholinergic properties in developing hypothalamic neurons. However, increased cholinergic transmission is unlikely to be the factor that prevents developmental gap

1724

junction uncoupling and Cx36 downregulation during NMDA receptor blockade. First, in AP5-treated cultures where cholinergic activity is also suppressed by acetylcholine receptor antagonists, developmental gap junction uncoupling and Cx36 downregulation is prevented (AP5 + A + M in Fig. 4), just as it is in the presence of AP5 or MK-801 alone. Second, the downstream mechanisms that regulate gap junction uncoupling and the induction of cholinergic properties in non-cholinergic hypothalamic neurons seem to be distinct, although both are NMDA receptor dependent. For example, chronic inactivation of L-type voltage-gated Ca2+ channels and PKA both induce cholinergic properties in cultured hypothalamic neurons44 (A.B. Belousov et al., Soc. Neurosci. Abstr. 413.1, 2002), but neither prevent developmental gap junction uncoupling and Cx36 downregulation (Fig. 4). Furthermore, the cholinergic phenotype is not induced in hypothalamic neurons by chronic inactivation of voltage-gated Na+ channels44, whereas, as discussed above, developmental gap junction uncoupling requires action potentials. Our experiments in the developing hypothalamus in vivo and in vitro confirm that there is an increase in the expression of Cx43, a presumptive glial connexin, in the rat hypothalamus during development45. The developmental increase in Cx43 levels, however, was larger in vivo than in neuronal cultures. This presumably was because the antimitotic drug cytosine b-D-arabinofuranoside, which reduces glial proliferation, was constantly present in the cell-culture medium. The level of Cx43 was not affected by the NMDA receptor blockade either in vivo nor in vitro, suggesting that the expression of this connexin is not NMDA receptor dependent. In conclusion, we have characterized the cellular and molecular mechanisms underlying developmental gap junction uncoupling in the hypothalamus. This should allow us to manipulate, either pharmacologically or genetically, the signaling pathways that are involved in uncoupling to determine the role of gap junctions in the development of neuronal circuits. METHODS Animal care. The use of animal subjects in these experiments was approved by the Tulane University Animal Care and Use Committee. All experiments were conducted in accordance with guidelines issued by the US National Institutes of Health. MK-801 treatment and slice preparation. Sprague-Dawley rat male pups received daily subcutaneous injections of MK-801 (dissolved in sterile saline): 0.5 mg kg1 in 15 ml on P1–P5, 0.75 mg kg1 in 40 ml on P6–P10, 1 mg kg1 in 75 ml on P11–P22 and 1.5 mg kg1 in 100 ml on P23–P29. The control group received saline injections (volumes as above). Following treatment, the rats were anesthetized with halothane and killed; the brain was removed and 400-mm thick coronal hypothalamic slices were cut using a vibroslicer as described46. Slices were prepared (at 2–4 1C) and kept in artificial cerebrospinal fluid (ACSF) containing 124 mM NaCl, 3.0 mM KCl, 2.0 mM CaCl2, 2.0 mM MgCl2, 1.23 mM NaH2PO4, 26 mM NaHCO3 and 10 mM glucose (aerated, pH 7.4, 20–22 1C). The body weight and weight of the forebrain of P30 MK801-treated animals were, respectively, 59.1% and 82.0% of P30 controls (n ¼ 9). To determine whether the number of neurons in the hypothalamus is affected by the MK-801 treatment, hypothalamic slices (40-mm thick) were immunostained with mouse anti-neuron-specific nuclear protein (NeuN) monoclonal antibody (1:100; Chemicon International, cat.# MAB377) and Texas Red conjugated rabbit anti-mouse secondary antibody (1:100; Jackson ImmunoResearch Lab, cat.# 315–075–045). Using an Olympus fluorescent microscope and OpenLab 3.0 software (Improvision), we calculated the number of NeuN-labeled cells inside a rectangular field (325 245 mm) in the hypothalamus that included one of the nuclei studied in this research, the PVN. The number of NeuN-labeled cells in the field was not significantly different between P30 saline-treated (146.4 ± 7.4) and MK-801-treated (142.8 ± 8.7) rats (n ¼ 9).

VOLUME 8

[

NUMBER 12

[



ARTICLES Culture preparation. Cultures were prepared as described47 from embryonic day 18–19 (E18–19) medial hypothalamus. Pregnant Sprague-Dawley rats were anesthetized with halothane before embryos were removed. After disaggregation using papain, neurons were plated on glass coverslips and raised in glutamate- and glutamine-free minimal essential medium (Life Technologies) with supplements47 and cytosine b-D-arabinofuranoside (1 mM). The culture medium was changed twice a week. We determined cell survival using a toxicity assay (Live/Dead Kit, Molecular Probes) as described previously47; cell survival was not affected under the chronic treatments studied (only CREB viral vectors induced neurodegeneration in cultures after 7–10 d; therefore, cells were exposed to viral vectors for no more than 4 d). We performed pharmacological treatments using sister cultures. Dye coupling. The pipette solution contained 145 mM KMeSO4, 10 mM HEPES, 2 mM MgCl2, 0.1 mM CaCl2, 1.1 mM EGTA, 2 mM Na-ATP and 0.3 mM Na-GTP, 0.2% neurobiotin (Vector Laboratories, cat.# SP–1120; MW323; gap junction–permeable dye) and 0.4% dextran Alexa Fluor 594 (Invitrogen, cat.# D22913; MW10,000; gap junction–impermeable dye) (pH 7.2, 3–7 MO electrode resistance). In slices: On the day of preparation, slices were randomly numbered and the number and condition (control or MK-801 treatment) were documented in a database (Access) for later identification. Magnocellular PVN and SON neurons were visualized using an Olympus BX51 microscope and an infrared camera and identified, as described48, based on their location, relatively large size and electrophysiological properties (for example, delayed spiking while depolarizing from below the resting membrane potential). After we obtained whole-cell access (using Axoclamp-2B amplifier and pCLAMP9 software, Axon Instruments), the neuron (one per slice) was held at the resting membrane potential for 25–30 min and then (after removal of the recording pipette) incubated at 37 1C in a low-calcium (0.2 mM) ACSF for an additional 60 min to allow dye diffusion. The slice was then fixed in 4% paraformaldehyde in 0.01 M phosphate buffer overnight, washed and stored at –80 1C. Between two and three weeks later, 10–12 slices from several experiments and conditions were blindly processed and analyzed in a group. We washed the slices with phosphate buffer, treated them with 0.2% Triton X-100 to permeabilize the membranes and stained them with fluorescein avidin D (FITC, 1:200, Vector Laboratories). We visualized Alexa 594 autofluorescence and neurobiotin staining using, respectively, Texas Red and FITC filters and an Olympus fluorescent microscope. We analyzed the signals using a Qicam-fast camera and Qcapture software (QImaging). The results (number of secondarylabeled neurons per slice) and the slice number were added in a separate window in the database; the software automatically matched the results with the condition (control or MK-801-treated) and generated the summary table. In cultures: Coverslips with neurons were transferred to ACSF and visualized using Axoivert Zeiss microscope. Because in cell cultures, different cell types are morphologically indistinguishable and electrophysiological characterization of cell types is not elaborated, neurons were chosen randomly. Cells were labeled, stained and analyzed as described above (for slices). Western blots. Freshly dissected tissue or cultured cells were homogenized in a lysis buffer (0.5 M HEPES, 3 M NaCl, 1 M MgCl2, 0.5 M EDTA, 0.1 M DTT, 10% SDS, 10% deoxycholate and 0.3% Triton X-100). Total protein in the homogenate was determined using the Bio-Rad DC protein assay method. In each lane, 20 mg of protein was loaded, resolved by 10% SDS-PAGE and transferred to 0.2 mm polyvinylidene difluoride membrane. The membrane was blocked in tris-buffered saline (TBS) with 0.1% Tween 20 and 5% bovine serum albumin (BSA), and then probed with a primary antibody in TBS containing 0.1% Tween 20 and 1% BSA. The following primary antibodies were used: rabbit anti-Cx36 (1 mg ml1; Zymed, cat.# 51–6300); rabbit anti-Cx43 (2 mg ml1; Zymed, cat.# 71–0770); rabbit anti-phospho-CREB (Ser133), which recognizes only the phosphorylated (active) form of CREB (0.5 mg ml1; Upstate Group, cat.# 06–519); and rabbit anti-CREB (0.5 mg ml1; Upstate Group, cat.# 06–863), which is raised against amino acids 5–24 of human CREB and recognizes both wild-type CREB and mutant CREB (introduced by HSV-mCREB) (Fig. 5b,d). The primary antibody was visualized with goat alkaline phosphatase–conjugated anti-rabbit antibody (1:1000–1500, Vector, cat.# AP–1000) and signals were enhanced using an Immun–Star AP chemiluminescence substrate (Bio-Rad). Band optical density was determined


[

NUMBER 12

[

DECEMBER 2005

using a frame grabber and Scion imaging software (Scion). All optical density signals were normalized relative to actin, and experimental samples were compared to controls (set at 1.0). Actin levels per unit total protein did not vary significantly among samples used in this study. NMDAR1 knockout mice. NMDAR1 knockout animals and their wild-type littermates were generated by timed breeding of NMDAR1 heterozygous adults using an established NMDAR1 knockout line49. Plug day was designated E0.5. Because NMDAR1 knockout mice die 12–24 h after birth due to respiratory distress (they are born alive and appear pink and healthy for B10 h), it was necessary to prepare hypothalamic neuronal cultures from embryonic mice to study developmental changes in gap junctions. At E18.5, dams were anesthetized, embryos were removed and the hypothalamus was dissected and used for culture preparations. Embryos were genotyped blindly by PCR as described50. Drugs, reagents and viruses. All drugs were obtained from Sigma-RBI unless otherwise specified. CREB ODN was obtained from Oligo’s Etc. This 20-mer (5¢-TGG TCA TCT AGT CAC CGG TG-3¢) is reverse and complementary to base pairs 27–46 of rat CREB1 mRNA and is phosphorothioate modified (substitution of a sulfur for an oxygen) between each of the last three bases on both the 5¢ and 3¢ ends (this modification slows the degradation of ODN by exonucleases). CREB ODN decreases CREB by blocking CREB translation37. It was administered to the cultures daily (2 nmol per treatment). CREB viral vectors were supplied by R.L. Neve. Viruses were prepared as described36. The estimated titer of recombinant viral stocks was 1 108 infectious units ml1. Viruses were added to cultures once (0.25 ml stock per ml), HSV-CREB on 16DIV and HSV-mCREB on 24DIV, and the cultures were tested on, respectively, 19DIV and 28DIV. The mutant form of CREB (which contains a single point mutation Ala for Ser at residue 133) cannot be phosphorylated but still binds to the Ca/CRE, thus preventing the binding of active CREB molecules36. Statistical analysis. Data were analyzed using the two–tailed paired Student’s t-test, analysis of variance (ANOVA), Fisher’s exact probability test and InStat software. Data are reported as mean ± s.e.m. for the number of samples indicated. Each date point represents indicated day ± 1, except for the CREB experiments (Fig. 5) where all dates are as indicated. Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS We are grateful to J. Denisova, E. Leininger, I.R. Popescu and T. Stuart for technical contributions and to R.L. Neve (Harvard Medical School) for supplying CREB viral vectors. This research was supported by a Louisiana Board of Regents Research Competitiveness Subprogram award to R.A.C.; and by a US National Institutes of Health grant (RO1 DA015088-01A1), a National Science Foundation grant (IBN-0117603) and an American Heart Association grant (0350530N) to A.B.B. COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/

1. Becker, D.L. & Mobbs, P. Connexin alpha1 and cell proliferation in the developing chick retina. Exp. Neurol. 156, 326–332 (1999). 2. Bani-Yaghoub, M., Underhill, T.M. & Naus, C.C. Gap junction blockage interferes with neuronal and astroglial differentiation of mouse P19 embryonal carcinoma cells. Dev. Genet. 24, 69–81 (1999). 3. Lin, J.H. et al. Gap-junction-mediated propagation and amplification of cell injury. Nat. Neurosci. 1, 494–500 (1998). 4. Lo Turco, J.J. & Kriegstein, A.R. Clusters of coupled neuroblasts in embryonic neocortex. Science 252, 563–566 (1991). 5. Allen, F. & Warner, A. Gap junctional communication during neuromuscular junction formation. Neuron 6, 101–111 (1991). 6. Walton, K.D. & Navarrete, R. Postnatal changes in motoneurone electrotonic coupling studied in the in vitro rat lumbar spinal cord. J. Physiol. (Lond.) 433, 283–305 (1991). 7. Peinado, A., Yuste, R. & Katz, L.C. Gap junctional communication and the development of local circuits in neocortex. Cereb. Cortex 3, 488–498 (1993). 8. Personius, K., Chang, Q., Bittman, K., Panzer, J. & Balice-Gordon, R. Gap junctional communication among motor and other neurons shapes patterns of neural activity and synaptic connectivity during development. Cell. Commun. Adhes. 8, 329–333 (2001).

1725


ARTICLES 9. Roerig, B. & Feller, M.B. Neurotransmitters and gap junctions in developing neural circuits. Brain Res. Brain Res. Rev. 32, 86–114 (2000). 10. Bennett, M.V. & Zukin, R.S. Electrical coupling and neuronal synchronization in the mammalian brain. Neuron 41, 495–511 (2004). 11. Naus, C.C. & Bani-Yaghoub, M. Gap junctional communication in the developing central nervous system. Cell Biol. Int. 22, 751–763 (1998). 12. Kandler, K. & Katz, L.C. Neuronal coupling and uncoupling in the developing nervous system. Curr. Opin. Neurobiol. 5, 98–105 (1995). 13. Kandler, K. & Katz, L.C. Coordination of neuronal activity in developing visual cortex by gap junction-mediated biochemical communication. J. Neurosci. 18, 1419–1427 (1998). 14. Meyer, T. Cell signaling by second messenger waves. Cell 64, 675–678 (1991). 15. Ben-Ari, Y. Developing networks play a similar melody. Trends Neurosci. 24, 353–360 (2001). 16. Garaschuk, O., Linn, J., Eilers, J. & Konnerth, A. Large-scale oscillatory calcium waves in the immature cortex. Nat. Neurosci. 3, 452–459 (2000). 17. Feller, M.B., Wellis, D.P., Stellwagen, D., Werblin, F.S. & Shatz, C.J. Requirement for cholinergic synaptic transmission in the propagation of spontaneous retinal waves. Science 272, 1182–1187 (1996). 18. Wong, R.O. Retinal waves and visual system development. Annu. Rev. Neurosci. 22, 29–47 (1999). 19. Milner, L.D. & Landmesser, L.T. Cholinergic and GABAergic inputs drive patterned spontaneous motoneuron activity before target contact. J. Neurosci. 19, 3007–3022 (1999). 20. Strata, F. et al. A pacemaker current in dye-coupled hilar interneurons contributes to the generation of giant GABAergic potentials in developing hippocampus. J. Neurosci. 17, 1435–1446 (1997). 21. Chang, Q., Gonzalez, M., Pinter, M.J. & Balice-Gordon, R.J. Gap junctional coupling and patterns of connexin expression among neonatal rat lumbar spinal motor neurons. J. Neurosci. 19, 10813–10828 (1999). 22. Mentis, G.Z., Diaz, E., Moran, L.B. & Navarrete, R. Increased incidence of gap junctional coupling between spinal motoneurones following transient blockade of NMDA receptors in neonatal rats. J. Physiol. (Lond.) 544, 757–764 (2002). 23. Connors, B.W., Benardo, L.S. & Prince, D.A. Coupling between neurons of the developing rat neocortex. J. Neurosci. 3, 773–782 (1983). 24. Peinado, A., Yuste, R. & Katz, L.C. Extensive dye coupling between rat neocortical neurons during the period of circuit formation. Neuron 10, 103–114 (1993). 25. Venance, L., Glowinski, J. & Giaume, C. Electrical and chemical transmission between striatal GABAergic output neurones in rat brain slices. J. Physiol. (Lond.) 559, 215–230 (2004). 26. Kandler, K. & Katz, L.C. Relationship between dye coupling and spontaneous activity in developing ferret visual cortex. Dev. Neurosci. 20, 59–64 (1998). 27. Hatton, G.I. Synaptic modulation of neuronal coupling. Cell Biol. Int. 22, 765–780 (1998). 28. Rash, J.E. et al. Immunogold evidence that neuronal gap junctions in adult rat brain and spinal cord contain connexin-36 but not connexin-32 or connexin-43. Proc. Natl. Acad. Sci. USA 97, 7573–7578 (2000). 29. Belluardo, N. et al. Expression of connexin36 in the adult and developing rat brain. Brain Res. 865, 121–138 (2000). 30. Long, M.A., Jutras, M.J., Connors, B.W. & Burwell, R.D. Electrical synapses coordinate activity in the suprachiasmatic nucleus. Nat. Neurosci. 8, 61–66 (2005).

1726

31. Srinivas, M. et al. Functional properties of channels formed by the neuronal gap junction protein connexin36. J. Neurosci. 19, 9848–9855 (1999). 32. Bessho, Y., Nawa, H. & Nakanishi, S. Selective up-regulation of an NMDA receptor subunit mRNA in cultured cerebellar granule cells by K(+)-induced depolarization and NMDA treatment. Neuron 12, 87–95 (1994). 33. Hardingham, G.E., Fukunaga, Y. & Bading, H. Extrasynaptic NMDARs oppose synaptic NMDARs by triggering CREB shut-off and cell death pathways. Nat. Neurosci. 5, 405–414 (2002). 34. Greenberg, M.E. & Ziff, E.B. Signal transduction in the postsynaptic neuron: activitydependent regulation of gene expression. in Synapses (eds. Cowan, M.W., Sudhof, T.C. & Stevens, C.F.) 357–391 (Johns Hopkins Univ. Press, Baltimore, 2001). 35. Lonze, B.E. & Ginty, D.D. Function and regulation of CREB family transcription factors in the nervous system. Neuron 35, 605–623 (2002). 36. Carlezon, W.A., Jr. et al. Regulation of cocaine reward by CREB. Science 282, 2272– 2275 (1998). 37. Guzowski, J.F. & McGaugh, J.L. Antisense oligodeoxynucleotide-mediated disruption of hippocampal cAMP response element binding protein levels impairs consolidation of memory for water maze training. Proc. Natl. Acad. Sci. USA 94, 2693–2698 (1997). 38. Venance, L. et al. Connexin expression in electrically coupled postnatal rat brain neurons. Proc. Natl. Acad. Sci. USA 97, 10260–10265 (2000). 39. Connors, B.W. & Long, M.A. Electrical synapses in the mammalian brain. Annu. Rev. Neurosci. 27, 393–418 (2004). 40. Galarreta, M. & Hestrin, S. Electrical synapses between GABA-releasing interneurons. Nat. Rev. Neurosci. 2, 425–433 (2001). 41. Xia, Z. & Storm, D.R. Regulatory Properties of the Mammalian Adenylyl Cyclases Ch. 7, 105–134 (R.G. Landes Austin, Texas, 1996). 42. Kornhuber, J. & Weller, M. Psychotogenicity and N-methyl-D-aspartate receptor antagonism: implications for neuroprotective pharmacotherapy. Biol. Psychiatry 41, 135–144 (1997). 43. Facchinetti, F. et al. Structural, neurochemical and behavioural consequences of neonatal blockade of NMDA receptor through chronic treatment with CGP 39551 or MK-801. Brain Res. Dev. Brain Res. 74, 219–224 (1993). 44. Belousov, A.B., Hunt, N.D., Raju, R.P. & Denisova, J.V. Calcium-dependent regulation of cholinergic cell phenotype in the hypothalamus in vitro. J. Neurophysiol. 88, 1352–1362 (2002). 45. Aberg, N.D., Ronnback, L. & Eriksson, P.S. Connexin43 mRNA and protein expression during postnatal development of defined brain regions. Brain Res. Dev. Brain Res. 115, 97–101 (1999). 46. Belousov, A.B. & van den Pol, A.N. Local synaptic release of glutamate from neurons in the rat hypothalamic arcuate nucleus. J. Physiol. (Lond.) 499, 747–761 (1997). 47. Belousov, A.B., O’Hara, B.F. & Denisova, J.V. Acetylcholine becomes the major excitatory neurotransmitter in the hypothalamus in vitro in the absence of glutamate excitation. J. Neurosci. 21, 2015–2027 (2001). 48. Luther, J.A. & Tasker, J.G. Voltage-gated currents distinguish parvocellular from magnocellular neurones in the rat hypothalamic paraventricular nucleus. J. Physiol. (Lond.) 523, 193–209 (2000). 49. Li, Y., Erzurumlu, R.S., Chen, C., Jhaveri, S. & Tonegawa, S. Whisker-related neuronal patterns fail to develop in the trigeminal brainstem nuclei of NMDAR1 knockout mice. Cell 76, 427–437 (1994). 50. Sugiura, N., Patel, R.G. & Corriveau, R.A. N-methyl-D-aspartate receptors regulate a group of transiently expressed genes in the developing brain. J. Biol. Chem. 276, 14257–14263 (2001).

VOLUME 8

[

NUMBER 12

[



ARTICLES

Regulation of neuronal morphology and function by the tumor suppressors Tsc1 and Tsc2 Sohail F Tavazoie1,3,4, Veronica A Alvarez1,4, Dennis A Ridenour1, David J Kwiatkowski2 & Bernardo L Sabatini1 Mutations in the TSC1 or TSC2 tumor suppressor genes lead to tuberous sclerosis complex (TSC), a dominant hamartomatous disorder that often presents with mental retardation, epilepsy and autism. The etiology of these neurological symptoms is unclear and the function of the TSC pathway in neurons is unknown. We found that in post-mitotic, hippocampal pyramidal neurons of mice and rats, loss of Tsc1 or Tsc2 triggered enlargement of somas and dendritic spines and altered the properties of glutamatergic synapses. Furthermore, loss of a single copy of the Tsc1 gene was sufficient to perturb dendritic spine structure. Morphological changes required regulation of the actin-depolymerization factor cofilin at a conserved LIM-kinase phosphorylation site, the phosphorylation of which was increased by loss of Tsc2. Thus, the TSC pathway regulates growth and synapse function in neurons, and perturbations of neuronal structure and function are likely to contribute to the pathogenesis of the neurological symptoms of TSC.

TSC1 and TSC2 are tumor suppressor genes whose protein products, hamartin (TSC1) and tuberin (TSC2), negatively regulate cell growth in a variety of systems. In humans, heterozygous mutations in either TSC1 or TSC2 lead to TSC, an autosomal-dominant hamartomatous disorder characterized by benign tumors in multiple organs including the brain, kidneys, heart and eyes1. TSC also typically presents with a constellation of neurological deficits that include epilepsy, mental retardation and autism. Biochemical and genetic analyses in mammalian systems and Drosophila melanogaster have revealed that TSC1 and TSC2 participate in a conserved growth-regulating pathway involving the mammalian target of rapamycin (mTOR)2–5. In brief, the activation of growthpromoting receptor tyrosine kinases, such as the insulin receptor, stimulates phosphoinositide 3-kinase (PI3K) and the serine/threonine kinase Akt. In vitro, Akt phosphorylates TSC2 at conserved consensus phosphorylation sequences and downregulates its GTPase-activating protein (GAP) activity6–8. Reduced GAP activity allows the buildup of GTP-bound Rheb9,10 and upregulates mTOR, which, through multiple actions, enhances protein translation and cell growth11. Thus, given the loss of heterozygosity of TSC1 or TSC2 found in hamaratomas of TSC patients12, the growth of these benign tumors is thought to result from the increased mTOR activity and uncontrolled cell growth that accompanies interruption of the TSC pathway. The pathogenesis of the neurological symptoms of TSC is unclear, and loss of heterozygosity is not seen within the brains of TSC patients13. The function of TSC1 and TSC2 in mammalian neurons and the defects that arise from hemizygosity of TSC1 or TSC2 are unknown. Many parallels exist, however, between the TSC pathway and

those that link extracellular stimuli to synaptic refinement in neurons. For example, activation of the TrkB receptor tyrosine kinase by brainderived neurotrophic factor (BDNF) stimulates PI3K and Akt to promote dendritic growth14. BDNF also triggers long-term potentiation of synaptic strength in an mTOR-dependent manner15. Similarly, strong activation of metabotropic glutamate receptors depresses synaptic transmission through a PI3K- and mTOR-dependent pathway16,17. Here we examine the role of the TSC pathway in regulating the growth of post-mitotic, differentiated neurons. We show that the TSC pathway regulates soma size, the density and size of dendritic spines, and the properties of excitatory synapses in hippocampal pyramidal neurons. These morphological effects are independent of regulation of Tsc2 by Akt at conserved phosphorylation sites but do require regulation of cofilin at a conserved LIM-kinase (LIMK) phosphorylation site. Furthermore, the TSC pathway is sensitive to gene-dosage effects, such that loss of a single copy of Tsc1, as is present in all neurons of TSC patients, is sufficient to perturb dendritic spine structure. Our results indicate that the TSC pathway regulates neuronal structure and function and suggest that the neurological symptoms of TSC are, at least in part, due to cell-autonomous perturbations of synapse function. RESULTS To uncover defects in neuronal structure and function caused by perturbation of the TSC pathway, we used transgenic mice carrying a conditional Tsc1 allele (Tsc1C) in which exons 17 and 18 are flanked by loxP sequences18. Loss of Tsc1 protein in neurons following transfection with a plasmid encoding a Cre recombinase–nuclear localization sequence fusion protein (Cre) was confirmed in dissociated hippocampal

1Department of Neurobiology, Harvard Medical School, 220 Longwood Avenue, Boston, Massachusetts 02115, USA. 2Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, 221 Longwood Avenue, Boston, Massachusetts 02115, USA. 3Present address: Department of Medical Oncology/Hematology, Memorial Sloan-Kettering Cancer Center, 1275 York Avenue, New York, New York 10021, USA. 4These authors contributed equally to this work. Correspondence should be addressed to B.L.S. ([email protected]).

Received 28 June; accepted 15 September; published online 6 November 2005; doi:10.1038/nn1566


[

NUMBER 12

[

DECEMBER 2005

1727

ARTICLES cell-autonomous neuronal defects caused directly by loss of Tsc1, we generated geneti1.0 1.0 cally mosaic brain tissue in which a small number of neurons lacking Tsc1 were located 0.5 0.5 C/C in otherwise normal brain tissue. This was Tsc1 Tsc1C/C Tsc1C/C; Cre accomplished by sparse transfection of pyraTsc1C/C; Cre 0.0 0.0 midal neurons with Cre and green fluorescent 0 5 10 0 1 2 Tsc1 immunostaining (au) pS6 immunostaining (au) protein (GFP) in organotypic hippocampal slices prepared from Tsc1C/C mice. Tsc1C/C; GFP Tsc1C/C; GFP; Cre Tsc1C/C; GFP; Cre e f Tsc1C/C; GFP GFP-transfected pyramidal hippocampal neurons were identified by their characteristic morphology (Fig. 1e) and location in a celldense band. Somas of Cre-transfected Tsc1C/C pyramidal neurons were larger than those of 15 µM GFP-transfected neurons, reaching a roughly * twofold enlargement at 20 DPT (n ¼ 20–31, g * 600 P o 0.05) (Fig. 1e–g). Changes in soma size were prevented by cotransfection with a plas# 400 # mid encoding Tsc1 (n ¼ 8). To control for 200 nonspecific effects of Cre-mediated DNA 0 recombination, similar measurements were made in tissue prepared from B6;129-Gt(ROSA)26Sortm2Sho/J mice that carry a floxed Tsc1C/C RosaC/C transcriptional stop upstream of the EGFP coding sequence in the Rosa26 locus Figure 1 Cre expression in hippocampal neurons of Tsc1C/C mice leads to loss of Tsc1 protein, increased (RosaC/C). Cre expression in RosaC/C neurons phosphorylation of S6 and increased soma size. (a) Untransfected (top) and Cre-transfected (bottom) neurons in a dissociated hippocampal culture from Tsc1C/C mice with immunoreactivity to MAP2 and had no effect on soma size (n ¼ 7–8) (Fig. 1g), Cre (red) and to Tsc1 (green). Arrow highlights red nucleus indicative of expression of the Cre-NLS fusion confirming that the increased soma size in protein. (b) Cumulative distribution of Tsc1 immunostaining in Cre-transfected and untransfected Tsc1C/C neurons transfected with Cre was due neurons at 6 DPT. (c) Untransfected (top) and Cre-transfected (bottom) neurons with immunoreactivity to loss of Tsc1. to MAP2 and Cre (red) and to pS6 (green). Arrow highlights red nucleus indicative of expression of the Soma size was also measured in rat neurons Cre-NLS fusion protein. (d) Cumulative distribution of pS6 immunostaining in Cre-transfected and in which loss of Tsc2 was induced by RNA untransfected neurons at 6 DPT. (e) Images of GFP-transfected (left) and GFP/Cre cotransfected interference (RNAi; see Methods and Fig. 1g). (right) pyramidal neurons in organotypic hippocampal slice from Tsc1C/C mice. Scale bar, 25 mm. (f) Enlarged views of the somas from neurons shown in e. (g) Summary of effects of loss of Tsc1 or Tsc2 Transfection of dissociated rat hippocampal on cross-sectional area of somas of hippocampal pyramidal neurons. Expression of Tsc1 or human TSC2 pyramidal neurons with a dual-promoter suppressed the effects of Tsc1 loss at 20 DPT and Tsc2 knockdown at 10 DPT, respectively. Expression plasmid encoding cytomegalovirus (CMV) of Cre in neurons from control ROSA26 locus (RosaC/C) mice had no affect on soma size. *P o 0.05 promoter-driven GFP and a U6-driven # compared to control, and P o 0.05 for rescue experiments compared to perturbed phenotype short-hairpin RNA targeting Tsc2 (shTsc2) (Cre- or shTsc2-expressing cells, as appropriate). reduced Tsc2 and increased pS6 (Supplementary Fig. 1 online). In organotypic hippocamcultures prepared from mice homozygous for the conditional allele pal slice cultures, somas of pyramidal neurons transfected with shTsc2 (Tsc1C/C). Because transfection efficiency of neurons is low, we mon- were enlarged relative to control neurons at 10 DPT (n ¼ 7–13, P o 0.05) itored Tsc1 levels by fluorescence immunohistochemistry (fIHC) (Fig. (Fig. 1g). This effect was occluded by cotransfection with human TSC2, 1a,b). In neurons expressing Cre, cytoplasmic Tsc1 levels were sig- which contains nine base pair changes within the region targeted by nificantly reduced at 6 days post-transfection (6 DPT) compared to shTsc2 (n ¼ 9) (Fig. 1g). Thus, soma size of post-mitotic, hippocampal untransfected neighboring neurons (P o 0.05, n ¼ 62–67) (Fig. 1b). In pyramidal neurons is controlled by the TSC pathway. Somatic enlargeseveral cell types, downregulation of TSC1/TSC2 upregulates mTOR ment occurred with a shorter latency after shTsc2 transfection than activity and increases phosphorylation of the ribosomal protein S63,5,8. after Cre transfection of Tsc1C/C neurons, likely reflecting the rapid Immunostaining against phosphorylated S6 (pS6) revealed a peri- degradation of mRNA triggered by RNAi. nuclear cytoplasmic signal in neurons (Fig. 1c) that was abolished by application of rapamycin, a selective pharmacological inhibitor Tsc1/2 regulate density and size of dendritic spines of mTOR (data not shown). PS6 levels were substantially higher in In pyramidal neurons, the vast majority of excitatory synapses are Cre-transfected Tsc1C/C neurons than in neighboring control neurons made onto the heads of dendritic spines, and the morphology and (P o 0.05, n ¼ 68) (Fig. 1c,d). Thus, Cre transfection in post-mitotic, density of dendritic spines reflect the properties and number of differentiated Tsc1C/C neurons induces recombination of the Tsc1C synapses. At 20 DPT, dendritic spines of Tsc1C/C neurons expressing GFP alone displayed roughly spherical spine heads separated from the allele, loss of Tsc1 protein and upregulation of mTOR. dendrite by thin necks (Fig. 2a). In contrast, dendrites of Tsc1C/C neurons coexpressing Cre and GFP possessed elongated spines with Tsc1/2 regulate cell growth in differentiated neurons Widespread loss of Tsc1 in the mouse brain, even when limited greatly enlarged, bulbous heads. Quantification of these changes to astrocytes, leads to pronounced seizures18, which may trigger showed that Tsc1 loss increased spine length and head width changes in gene expression and synaptic transmission. To identify and decreased the density of dendritic spines (n ¼ 9–16 cells and

b

c

MAP2/Cre

Phospho-S6

d

Tsc1C/C

Tsc1

Tsc1C/C; Cre

MAP2/Cre

G F c2 shT P + sc hT 2 SC 2

FP C re

G

sh

Ts

C

C

G F re P 10 re d C re 20 + d Ts c1

Soma area (µm2)


Tsc1C/C; Cre

Tsc1C/C

a

1728

VOLUME 8

[

NUMBER 12

[


ARTICLES

b 1.0

a Apical

Basal

GFP GFP, Cre GFP, Cre, Tsc1

0.5

0.5

0.0

0.0

Tsc1C/C; GFP

0

c

2 Spine length (µm)

4

(µm–1)

Density

#

0

1 Spine head width (µm)

#

*

0.5

*

*

* #

*

*

*

*

2,441–4,344 spines, P o 0.05; Fig. 2b,c). Similar effects on spine size and morphology were seen at 10 DPT (n ¼ 7–8 cells and 2,577–3,338 spines, P o 0.05; Fig. 2), a time point at which loss of Tsc1 has no effect on soma size (Fig. 1g). Changes in spine morphology and density were prevented by expression of Tsc1 (n ¼ 8 cells and 2,960 spines; Fig. 2b,c) and no changes in spine density or morphology were seen with Cre expression in RosaC/C neurons (Supplementary Table 1 online), confirming that the morphological changes seen with Cre expression in Tsc1C/C neurons were due to loss of Tsc1 protein. The perturbations of spine morphology triggered by loss of Tsc1 were phenocopied by RNAi-mediated knock-down of Tsc2 in rat hippocampal neurons (Fig. 2c). Thus, at both 10 DPT (n ¼ 7–13 cells and 1,715–5,023 spines) and 20 DPT (5–13 cells and 993–4,262 spines), dendritic spines of shTsc2-expressing neurons were elongated

G FP Ts 10d c hT 2 1 SC 0d 2 10 d sh

sh

Tsc1C/C

Ts

c2

+

C

re

FP

C

G

C

FP G

20 re d 20 + d Ts c1

0.6

# G FP Ts 20d c2 20 d

#

*

0.8

10 re d 10 d

Head width (µm)

0

*

#

sh

Length (µm)

Tsc1C/C; Cre; GFP

*

*

1

2

*

1.0

0.0

with enlarged spine heads (P o 0.05). These effects were rescued by expression of hTSC2 (n ¼ 9 cells and 2,682 spines; Fig. 2c). Summaries of morphological parameters and numbers of cells and spines analyzed for each experimental condition are given in Supplementary Table 1. Functional defects associated with loss of Tsc1 Do the enlarged spine heads seen after loss of Tsc1 contain functionally perturbed synapses? To address this question, we obtained whole-cell voltage-clamp recordings from Tsc1C/C pyramidal neurons transfected with Cre and from untransfected neighboring neurons. Spontaneous miniature excitatory postsynaptic currents (mEPSCs) were recorded (Fig. 3a). A red fluorophore was included in the recording pipette solution in order to fill the neuron and confirm its identity. At 10 DPT, mEPSC amplitude was roughly 20% higher in Cre-expressing neurons

Figure 3 Loss of Tsc1 increases AMPAR-mediated stim synaptic currents. (a) Left, laser-scanning rec DIC image of the pyramidal cell layer of an organotypic slice culture from a Tsc1C/C mouse with superimposed green fluorescence from a GFP-expressing neuron. Note gold particle (arrow) pcl in the nucleus of the transfected neuron. Image of a Cre/GFP-transfected pyramidal neuron before Control (middle) and after (right) filling with a red * 4 fluorophore through the whole-cell recording electrode. Scale bars, 15 mm. (b) Representative traces of mEPSCs recorded from control and CreCre 2 transfected Tsc1C/C neurons. Scale bars: 20 pA, 20 pA 200 ms. (c) Cumulative distributions showing 20 ms increased mEPSC amplitudes in Cre-transfected 0 (black trace) Tsc1C/C neurons compared to control 2 1.0 neurons (gray trace) recorded at 10 DPT (left). 0.2 Frequency of mEPSCs in untransfected and CreControl transfected Tsc1C/C pyramidal neurons at 10 DPT A1 (right). (d) Wide-field image of a hippocampal A2 0.5 1 0.1 organotypic slice culture from a Tsc1C/C mouse Cre showing sparse GFP transfection of neurons in the 20 pA pyramidal cell layer (pcl) and a schematic of 0.0 0.0 0 PPF = A2/A1 50 ms relative positions of the recording electrode (rec) 20 40 and stimulating electrode (stim) within the slice. Amplitude (pA) Scale bar, 100 mm. (e) Evoked EPSCs recorded at –60mV (negative current) and +40mV (positive current) with arrowheads showing the times at which AMPAR (negative) and NMDAR (positive) mediated currents were measured (left). Mean AMPAR/NMDAR current ratios for control (gray bar) and Cre-transfected (black bar) Tsc1C/C neurons at 20 DPT (right). (f) Paired-pulse facilitation measured at a 50 ms interpulse interval in a Cre-transfected Tsc1C/C neuron at 20 DPT and summary data for control and Cretransfected neurons. *P o 0.05.

d

b

e

c

f

AMPA/NMDA ratio

a


[

NUMBER 12

[

DECEMBER 2005

re

C

tro l C re

C on

C

on

tro

l

PPF

mEPSC freq (Hz)


Figure 2 Loss of Tsc1 leads to increased size but decreased density of dendritic spines. (a) Images of apical (left) and basal (right) dendrites from Tsc1C/C pyramidal neurons in hippocampal organotypic slice culture. Neurons were transfected with GFP alone (top) or GFP and Cre (bottom). Scale bar, 10 mm. (b) Cumulative distribution of spine length and head width from Tsc1C/C neurons transfected with GFP alone; GFP and Cre; or GFP, Cre and Tsc1 at 20 DPT. (c) Summary of effects of Tsc1 or Tsc2 loss on dendritic spine density, length and head width. *P o 0.05 compared to control neurons. #P o 0.05 for rescue experiments compared to perturbed phenotype (Cre- or shTsc2-expressing cells as appropriate).

1.0

1729

ARTICLES

*

*

* 50 0

Soma area

Spine density

Spine length Spine head width

Figure 4 Neuronal morphology is sensitive to hemizygosity of Tsc1. (a) Dendrites of Cre/GFP (left) and GFP (right) transfected pyramidal neurons in hippocampal organotypic slice cultures prepared from Tsc1C/+ mice at 20 DPT. Scale bar, 5 mm. (b) Relative soma cross-sectional area, spine density, length and head width for Cre-transfected Tsc1C/+ and Tsc1C/C neurons compared to GFP-transfected neurons of each genotype. *P o 0.05.

compared to neighboring control neurons (n ¼ 10–12 cells and 660–781 events, P o 0.05; Fig. 3b,c), indicating an enhanced sensitivity to released neurotransmitter. No changes in mEPSC frequency (Fig. 3c) or in resting membrane resistance (Rin) and cell capacitance (Cm) were seen (controls: Rin ¼ 197 ± 43 MO, Cm ¼ 256 ± 61 pF; Cre-expressing neurons: Rin ¼ 200 ± 63 MO, Cm¼ 182 ± 34 pF). In contrast, at 20 DPT, cells lacking Tsc1 had greatly reduced Rin (127 ± 15 MO vs. 185 ± 20 MO in control neurons, P o 0.05) and increased Cm (234 ± 24 pF vs. 171 ± 17 pF in control neurons) (n ¼ 13 in each condition), making the comparison of mEPSCs between Tsc1-lacking neurons and control neurons difficult. At 20 DPT, EPSCs evoked by stimulation of Schaffer collaterals (Fig. 3d) were monitored in CA1 pyramidal neurons at a holding potential of –60 mV, at which AMPARs are activated, and at +40 mV, at which the block of NMDA-type glutamate receptors (NMDARs) by Mg2+ is relieved and the long-lived NMDAR current is revealed (Fig. 3e). The ratio of evoked AMPAR- to NMDAR-mediated currents normally increases during development as well as after induction of long-term potentiation. In Cre-transfected neurons, the AMPAR/ NMDAR current ratio was significantly increased relative to that in controls (3.7 ± 0.71 vs. 2.0 ± 0.35 in controls, n ¼ 13 in each condition, P o 0.05), indicating an aberrant relative enhancement of synaptic AMPARs. To examine possible changes in presynaptic function induced retrogradely by postsynaptic loss of Tsc1, we recorded responses to a pair of stimuli and measured the paired-pulse facilitation (PPF). PPF was similar in untransfected and Cre-transfected neurons (1.73 ± 0.22 and 1.75 ± 0.15, respectively, n ¼ 11–13; Fig. 3f), suggesting that release probability is not affected by postsynaptic loss of TSC1. Figure 5 mTOR-dependent regulation of neuronal morphology. (a,b) Lowpower images of the whole cell (left) and enlarged views of spiny dendrites (right) of (a) GFP- and (b) shTSC2-transfected pyramidal neurons in control media at 20 DPT (top) or for 14 d in control media followed by 6 d in 100 nM rapamycin (bottom). Scale bars, 30 mm (left) and 5 mm (right). (c) Summary of measured spine density, length, head size and soma area following application of rapamycin. Shaded areas indicate the mean ± 2 s.e.m. of each parameter for control (dark gray) and shTsc2 (light gray) neurons determined in Figures 1 and 2. *P o 0.05 compared to cells in control media.

1730

Regulation of neuronal morphology by mTOR Rapamycin is a potent and specific inhibitor of mTOR, and is predicted to act downstream of TSC1 and TSC2. To determine whether the effects of loss of Tsc1 or Tsc2 on neuronal morphology are mediated by increased activity of mTOR, we examined the ability of rapamycin to reverse defects in neuronal morphology (Fig. 5). Neurons in hippocampal organotypic slices were transfected and maintained in culture for 14 d, a time point at which the effects of Cre-mediated loss of Tsc1 and RNAi-mediated loss of Tsc2 were apparent. Rapamycin (100 nM) was then added to the culture media, and the cultures were maintained for six more days before morphological analysis (total of 20 DPT). Application of rapamycin to control GFP-expressing rat neurons had no effect on soma size but induced the growth of long, thin spines (n ¼ 6 cells and 1,702 spines; Fig. 5a,c). Surprisingly, the mean length of dendritic spines on control cells in the presence of rapamycin was similar to that of Tsc2 knockdown cells (Fig. 5a,c). Unlike the phenotype resulting from loss of Tsc2, however, the rapamycin-induced spine phenotype did not show enlarged spine heads. Application of rapamycin to shTsc2-expressing cells reversed the enlargement of the soma and spine heads but further increased spine length (n ¼ 8 cells and 1,163 spines, P o 0.05; Fig. 5b,c). In these cells, it was difficult to distinguish long filopodia and spines from nascent dendritic branches, artifactually lowering the measured spine density. Similar effects were seen in Cre-expressing Tsc1C/C cells, such that application of rapamycin

a

c

GFP

b

shTsc2

Spine length (µm) Spine density (µm–1)

*

shTsc2

* 0.0

* 2

*

1 0 0.8

*

0.7 0.6

400

*

200 0 ap

r P+

GF

VOLUME 8

[

NUMBER 12

[

GFP

0.5

Head width (µm)

* *

100

Soma area (µm2)

150

Control

Tsc1C/+; Cre Tsc1C/C; Cre

200

Rapamycin

*

b

Morphological effects of haploinsufficiency of Tsc1 Individuals afflicted with TSC carry heterozygous mutations in either the TSC1 or TSC2 genes and loss of heterozygosity, although present in hamartomas, is not seen within the brain. Therefore we tested whether loss of a single copy of Tsc1 is sufficient to trigger changes in neuronal morphology (Fig. 4). In tissue prepared from Tsc1C/+ mice, Creexpressing neurons had increased soma size, decreased spine density, increased spine length and increased spine head width relative to GFPexpressing neurons (n ¼ 11 cells and 2,996 spines, P o 0.05). The morphological changes were less pronounced after loss of a single copy than after loss of both copies of Tsc1 (Fig. 4b), suggesting that the TSC pathway is sensitive to gene dosage.

Control

Tsc1C/+; GFP

Rapamycin

Tsc1C/+; Cre

Percentage of control


a

/6d /6d 20 rap

20

+ c2

Ts

sh


ARTICLES

b

GFP Spine density (µm–1)

a

0.5

*

0.0

Head width (µm) Soma area (µm2)

A

* #

0.0 0.8

*

*

0.7

#

0.6

*

500

#

0 A

Ak Ak t hT C t A sh CA + SC Ts hT 2 AA c2 SC + hT 2 AA SC 2A

sh Ts c2

+

hT SC + A C

Ak t

*

0.5

hT SC

2A

2A

A

A C

*

1.0

*

Ak t


Spine length (µm)

*

shTsc2 #

Figure 6 Akt upregulation phenocopies Tsc1/Tsc2 loss but does not require phosphorylation of Tsc2 at conserved Akt phosphorylation sites. (a) Images of pyramidal neurons (bottom) in organotypic slice cultures of rat hippocampus and enlarged views of spiny dendrites (top). Neurons were transfected with an AktCA (left); AktCA and hTSC2AA (middle); or shTsc2 and hTSC2AA (right) and imaged at 10 DPT. Scale bars, 5 mm (top) and 30 mm (bottom). (b) Summary of effects on spine density, length and head width. Shaded areas represent mean ± 2 s.e.m. of each parameter for control (dark gray) and shTsc2 (light gray) neurons determined in Figures 1 and 2. *P o 0.05 compared to control neurons. #P o 0.05 for hTSC2AA rescue experiments compared to shTsc2 or AktCA expressing neurons, as appropriate.

prevented the increase in soma and spine head sizes but further increased spine length (Supplementary Table 1). Thus, rapamycinsensitive mTOR activity is epistatic to Tsc2 with respect to the regulation of soma and spine head sizes. Epistatic analysis of Tsc2 and Akt It has been proposed that Akt inhibits TSC2 by phosphorylation at two sites (S939 and T1462, numbered by the human sequence) that are conserved in D. melanogaster and mammals7. We investigated whether Akt regulates neuronal morphology and, if so, whether it occurs through phosphorylation of sites S939 and T1462 of Tsc2 (Fig. 6). Expression in rat hippocampal pyramidal neurons of a constitutively active Akt (AktCA), consisting of a myristoylated Akt lacking its pleckstrin homology domain (myrAktD4-129)19, phenocopied loss of Tsc1/Tsc2, resulting in large somas, long spines and increased spine head size (n ¼ 5 cells and 1,052 spines, P o 0.05; Fig. 6a,b). Three experiments were performed to determine if these effects are mediated by downregulation of Tsc2 by Akt through phosphorylation of S939 and T1462 (Fig. 6 and Supplementary Table 1). First, hTSC2AA was expressed alone and found to have no effect on spine or somatic morphology (n ¼ 5 cells and 1,014 spines). Second, expression of hTSC2AA did not occlude the changes triggered by overexpression of AktCA (n ¼ 5 cells and 1,149 spines). Third, hTSC2AA expression in Tsc2 knockdown cells rescued spine and soma morphology with an efficiency equal to that of wild-type hTSC2 (n ¼ 5 cells and 1,824 spines). Thus, phosphorylation of S939 and T1462 is not necessary for regulation of neuronal soma and spine size by Tsc2, and hTSC2AA is unable to act as a dominant negative with respect to the neuronal enlargement triggered by upregulation of Akt.


[

NUMBER 12

[

DECEMBER 2005

Morphological changes require signaling through cofilin The mechanism by which the TSC pathway regulates neuronal morphology is unknown. Cofilin, a protein that depolymerizes and severs actin filaments and is widely expressed in the mammalian brain, has recently been shown to regulate spine size20,21. Cofilin is negatively regulated by LIMK through phosphorylation at a conserved site (serine 3). We examined whether phosphorylation at this site is regulated by the TSC pathway using fIHC to monitor levels of Ser3-phosphorylated cofilin (p-cofilin) in cultures of dissociated hippocampal neurons (Fig. 7a,b). At 10 DPT, shTsc2-transfected cells had higher p-cofilin levels than control neurons (Fig. 7b; n ¼ 34–42 cells), whereas total cofilin levels were unchanged (n ¼ 26–32 cells). In addition, p-cofilin, but not total cofilin, was increased in Tsc2/ mouse embryonic fibroblasts (MEFs)22 compared to Tsc2+/+ MEFs (n ¼ 3–5, P o 0.05; Fig. 7c,d). Similar results were obtained with a second commercial p-cofilin-specific antibody (see Methods, data not shown). Furthermore, the intensity of the B19 kD band recognized by the p-cofilin antibody on western blots reflected levels of phosphorylated cofilin as its intensity was reduced in HEK293T cells transfected with the cofilin Ser3 phosphatase slingshot-1L23. No changes in intensity were noted following LIMK1 transfection, possibly reflecting additional regulation of the kinase activity independently of its expression level. To determine whether increased cofilin phosphorylation is necessary for the spine enlargement described above, we expressed wild-type cofilin or cofilin with the LIMK phosphorylation site mutated to alanine (cofilin S3A) in shTsc2-transfected neurons in rat organotypic slice cultures (Fig. 7e). At 10 DPT, neurons expressing cofilin and shTsc2 had slightly reduced spine length compared to those expressing shTsc2 alone (n ¼ 9 cells and 2,336 spines, P o 0.05) but similar soma size, spine head width and spine density. In contrast, expression of cofilin S3A in shTsc2-transfected neurons restored soma size, spine head width and spine density to control levels and decreased spine length to levels slightly below those of control neurons (n ¼ 6 cells and 3,200 spines, P o 0.05 compared to shTsc2 neurons; Fig. 7e). Cofilin S3A expression in control neurons had no effect on soma size, spine head width and spine density, but it did reduce spine length (n ¼ 6 cells and 4,243 spines, P o 0.05; Fig. 7f). The morphology of cofilin S3A/ shTsc2-expressing cells was identical to that of cofilin S3A/GFPexpressing neurons (Fig. 7g), indicating that cofilin lies downstream of Tsc2 and that phosphorylation at Ser3 is necessary for the morphological changes induced by loss of Tsc2. DISCUSSION We have shown that the TSC pathway regulates the morphology and function of post-mitotic, hippocampal pyramidal neurons. Loss of Tsc1/Tsc2 triggered multiple changes in neuronal morphology, including increased soma and dendritic spine size as well as decreased dendritic spine density. These effects required both rapamycin-sensitive mTOR activity and regulation of cofilin at a conserved LIMK phosphorylation site. Furthermore, phosphorylation of cofilin, an actindepolymerization factor not previously linked to the TSC pathway, was increased following loss of Tsc2. Synaptic strength and passive membrane properties were also altered in neurons lacking Tsc1, suggest that primary defects in neuronal function contribute to the neurological symptoms of TSC. Relationship of Tsc2 and Akt We found that expression of constitutively active Akt phenocopies the morphological defects seen with loss of Tsc1 or Tsc2. Increased Akt activity leads to enlarged neuronal size, in agreement with studies using overexpression of Akt or downregulation of PTEN, the lipid

1731

*

g

*

*

0.5

shTsc2 + cof

Control shTsc2

0.5 shTsc2 + cofS3A

0.0

K

H

M

–

Ts c2 + /+ Ts c2 – /

SS

LI

+

+

EK

EK

H

p-cof

Cof

α-tub

α-tub

*

0.5

*

f

GFP

GFP + cofS3A

200 100

0.7

*

0.6

500

*

3A

f co

fS

fS

+

G

FP

+

co

co +

c2 sh

c2

Ts

1 H

SS

K1 M LI

lin

Tu

bu

in

ilin

fil

of

co

C

p-

3A

0

*

0

sh

d

*

2 Head width (µm)

1 Cofilin (au) 1

0

2

H

c

1 p-cofilin (au)

*

1.0

0.0

0.0 0

0.5

0.0

1.0

1.0

shTsc2

*

Ts

b

GFP 1.0

*

*

Figure 7 Phosphorylation of cofilin is regulated by Tsc2 and is necessary for increased cell growth. (a) Cultured hippocampal neurons transfected with shTsc2 showing GFP fluorescence (green, left), immunostaining for MAP2 (blue, left) and immunostaining for phosphorylated cofilin (p-cofilin) (red, right). Arrows and asterisks highlight shTsc2-transfected neurons and nontransfected neighbors, respectively. Scale bar, 20 mm. (b) Cumulative distribution of p-cofilin (left) and total cofilin (right) immunostaining in shTsc2-transfected and GFP-transfected neurons at 10 DPT. (c) Left, western blots of p-cofilin (left) from Tsc2 +/+ MEFs, Tsc2/ MEFs, control HEK293T cells and HEK293T cells transfected with LIMK1 or slingshot-1L. Right, western blot of total cofilin in Tsc2 +/+ and Tsc2/ MEFs. (d) Summary of band intensities for p-cofilin, cofilin and tubulin in Tsc2/ MEFs relative to Tsc2 +/+ MEFs (left), and the intensities of p-cofilin bands in HEK cells transfected with LIMK1 or slingshot-1L relative to untransfected cells (right). *P o 0.05. (e) Dendrites of pyramidal neurons in organotypic hippocampal slices transfected with shTsc2 (top), shTsc2+cofilin (middle) or shTsc2+cofilinS3A (bottom). Scale bar, 5 mm. (f) Dendrites of pyramidal neurons transfected with GFP (top) or GFP + cofilin S3A (bottom) imaged at 10 DPT. (g) Summary of the effects of the manipulations shown in e and f on spine density, length, head width and soma area. Shaded areas represent the mean ± 2s.e.m. of each parameter for control (dark gray) and shTsc2 (light gray) neurons. *P o 0.05 compared to neurons transfected with GFP or shTsc2 alone, as appropriate.

Molecular mechanisms of morphological perturbations In addition to regulating mTOR activity, TSC1 and TSC2 are reported to participate in mTOR-independent signaling cascades that may regulate cell morphology. TSC1 interacts with and regulates the cytoskeleton via a Cterminal domain that binds ezrin-radixin-moesin (ERM) family proteins and an N-terminal domain that is capable of activating Rho GTPase in human endothelial cells30. TSC2 has been shown to activate Rho (in MDCK and ELT3 cells)30, and increased Rho activity decreases spine length in hippocampal pyramidal neurons31. Furthermore, Rho GDP/GTP exchange factors (Rho-GEFs) regulate dendritic spine morphology and density32,33. TSC2 also interacts with 14-3-3 proteins34,35, whose many cellular functions include trafficking of ion channels from the endoplasmic reticulum36 and regulation of the actin cytoskeleton37. Importantly, TSC1 and TSC2 proteins stabilize each other38, and thus reduced expression of either protein may induce a common set of morphological perturbations mediated through disruption of these mTOR-independent signaling cascades. Which of these pathways is responsible for the changes in neuronal morphology observed here? Application of rapamycin to neurons with reduced Tsc1 or Tsc2 levels restores neuronal soma and dendritic spine head sizes to control levels, consistent with mTOR being downstream of Tsc1/Tsc2. In contrast, in the presence of rapamycin, spine morphology is clearly different in control neurons than in neurons with reduced Tsc1/Tsc2, arguing against a strictly epistatic relationship. Thus, Tsc1/Tsc2 may regulate soma and spine size via a rapamycinsensitive mTOR-dependent pathway and spine length via a rapamycin-

1732

e shTsc2

p-cofilin

Spine density (µm–1)

MAP-2

Spine length (µm)

GFP

Soma area (µm2)

a

Ts c2 + /+ Ts c2 – /– H EK

phosphatase that antagonizes PI3K6,24–27. However, we found that the somatic and dendritic spine enlargement induced by increased Akt activity is independent of phosphorylation of TSC2 at S939 and T1462, the consensus sites conserved across D. melanogaster, rats, mice and humans. This is in contrast to findings in D. melanogaster in which expression of dTsc2S924A/T1518A suppresses the enlargement of ommatidia triggered by overexpression of Akt6. Furthermore, we found that expression of hTSC2AA does not perturb spine or soma size and that it reverses the effects of loss of endogenous Tsc2. Again, these results are in contrast with the marked reduction in ommatidia size following dTsc2S924A/T1518A expression6 but in agreement with recent findings that dTsc2S924A/T1518A is able to rescue the dTsc2-null phenotype in D. melanogaster development28. Lastly, our results are consistent with recent findings that indicate that Akt is regulated by mTOR29 and thus lies downstream of Tsc1/Tsc2. In summary, our data indicate that, although upregulation of Akt and downregulation of Tsc1/Tsc2 result in similar phenotypes, phosphorylation of Tsc2 at S939 and T1462 is not necessary for the normal control of neuronal size or for the neuronal enlargement induced by upregulation of Akt.

Intensity (% control)


ARTICLES

insensitive pathway. However, two caveats must be considered. First, long-term application of rapamycin may indirectly inhibit rapamycininsensitive mTOR activity by sequestering mTOR and preventing formation of rictor-mTOR complexes29,39. Second, rapamycin may have mTOR-independent effects through the sequestration of FK506binding proteins. The central role of actin dynamics in determining spine structure strongly suggests that perturbed regulation of the actin cytoskeleton underlies the morphological changes described above. Indeed, we find a link between the TSC pathway and regulation of the cytoskeleton through cofilin. Cofilin is an actin depolymerization factor whose activity is downregulated following phosphorylation by LIM kinase at Ser3 (S3)40. Haploinsufficiency of LIMK1, a downstream effector of the Rho family of GTPases, results in the cognitive disorder Williams syndrome41. Both cofilin and LIMK1, as well as upstream regulators of the pathway such as PAK and Rho, have been shown to regulate dendritic spine morphology21,42–45. We found that phosphorylation of cofilin at S3 is necessary for the morphological changes triggered by loss of Tsc2, as they are occluded by expression of cofilin with this site eliminated (S3A). Our results also demonstrate that the balance of active and inactive cofilin is regulated by the TSC pathway in both neurons and MEFs, such that loss of Tsc2 leads to increased

VOLUME 8

[

NUMBER 12

[


ARTICLES


S3-phosphorylated cofilin without changes in total cofilin. Possible mechanisms for this effect include mTOR-dependent translational control of a key regulatory protein as well as direct regulation of the enzymatic activity of PAK, LIMK1, or slingshot by mTOR or a downstream kinase such as Akt29. Alternatively, regulation of Rho through Tsc1-dependent protein-protein interactions30 might mediate downstream regulation of cofilin. Implication for tuberous sclerosis complex Individuals with TSC show perturbations of cortical architecture including tubers, which are disorganized regions of the brain with disturbed lamination containing increased numbers of astrocytes and sparse neurons1. The correlation between the number of cortical tubers and the severity of seizure symptoms has led to the idea that the neurological deficits in TSC could arise from disruptions of cortical architecture46. Here we show that loss of a single copy of Tsc1 results in defects in neuronal morphology, including increased soma size, decreased spine density and increased spine size. Therefore we propose that cell-autonomous neuronal defects due to haploinsufficiency of TSC1 or TSC2, in addition to the perturbations of brain architecture caused by cortical tubers, subependymal nodules and giant cell astrocytomas, contribute to the pathogenesis of the neurological symptoms of TSC. METHODS Animals. We used mice carrying a conditional Tsc1 allele18 (Tsc1C) consisting of loxP elements upstream of exon 17 and downstream of exon 18, as well as Sprague-Dawley rats (Charles River Laboratory). Tsc1C/+ mice were generated by crossing Tsc1C/C mice with C57Bl6 mice (Charles River Laboratories). RosaC/C mice (Jackson Laboratories) contain the EGFP coding sequence with an upstream loxP-flanked stop in the Gt(ROSA)26Sor locus. All animal procedures were conducted following Harvard Medical School guidelines. Tsc1C genotyping was performed using tail genomic DNA and primers F4536 (5¢-ACGAGGCCTCTTCTGCTACC-3¢) and R4830 (5¢-CAGCTCCGACCAT GAAGT-3¢), yielding 295-bp and 480-bp products from wild-type and conditional alleles, respectively. Plasmids. All enzymes were obtained from New England Biolabs. The following plasmids were gifts: pBS::b-actin-nls-cre (S. Dymecki, Harvard Medical School); hTSC2 and hTSC2AA (E. Henske, Fox Chase Cancer Center, Philadelphia, Pennsylvania); pBS/U6 (Y. Shi, Harvard Medical School); cofilin and cofilinS3A (A. Minden, Columbia University); Slingshot1L and LIM-kinase1 (K. Mizuno, Tohoku University, Japan). pEGFP-N1 (Invitrogen) was used as a GFP control. For construction of the dual promoter CMV-EGFP/U6-shRNA vector, pEGFP-N1 was digested with BamHI and BglII to remove the multiple cloning site (MCS) and designated as pEGFP-N1::DMCS. The U6 promoter and its downstream MCS from pBS/U6 were inserted into pEGFP-N1::DMCS at a filled-in AflII site downstream of the SV40 polyadenylation sequence in the forward (pGUF) or reverse (pGUR) orientation relative to EGFP. Four sets of shRNAs directed against Tsc2 were designed from 19–21 bp coding sequences conserved in mice and rats with roughly 50% guanosines and cytidines ratio. Oligonucleotides containing the target sequence, a restriction site, the reverse complement of the target sequence, five threonine residues for U6 polymerase termination and an EcoRI site were synthesized (Integrated DNA Technologies). The shTsc2 clone used in this study was constructed from the following oligonucleotides: GGTGAAGAGAGCCGTATCACAAAGCTTTGTGATACGGC TCTCTTCACCCTTTTTG and AATTCAAAAAGGGTGAAGAGAGCCGTATC ACAAAGCTTTGTGATACGGCTCTCTTCACC. For production of shTsc2, pGUF was digested with ApaI, klenowed, digested with EcoRI and treated with calf intestinal alkaline phosphatase. Oligonucleotides were annealed, phosphorylated with T4 polynucleotide kinase and ligated into the digested pGUF to yield shTsc2. Cultures and transfection. Dissociated hippocampal cultures were prepared from P3 rats and mice47 and plated at 8 104 cells/well on glial monolayers on


[

NUMBER 12

[

DECEMBER 2005

12-mm glass coverslips. Cultures were transfected with Lipofectamine 2000 (Invitrogen) after 3–5 days in vitro (d.i.v.) in neuronal NEU media lacking biotin. Organotypic slice cultures were prepared from P5–P7 mice and rats48 and biolistically transfected with a Helios Gene Gun (Biorad) after 2 d.i.v.. Bullets were prepared using 12.5 mg of 1.6 mm gold particles and 40–80 mg of plasmid DNA. Immunofluorescence. Dissociated hippocampal neurons were fixed in 3.7% paraformaldehyde/4% sucrose for 15 min at 20–24 1C, permeabilized with 0.1% Triton X-100/PBS (Sigma) for 10 min, blocked with 1% goat serum/PBS (Jackson ImmunoLabs) and incubated with the following primary antibodies: anti-phospho-S6 ribosomal protein (1:100, Ser235/236; Cell Signaling), antiMAP-2 (mouse, 1:500, Sigma), anti-MAP-2 (rabbit, 1:500, Chemicon), anti-Cre recombinase (mouse, 1:1,000, Chemicon), anti-Cre recombinase (rabbit, 1:1,000, Novagen), anti-p-cofilin1 (Ser3) (1:100, Santa Cruz) and anti-cofilin (1:200, Cell Signalling). Antibodies specific to tuberin and hamartin (1B2A8 and HF6, respectively) were kindly provided by V. Ramesh (Massachusetts General Hospital). The following secondary antibodies from Jackson Immunoresearch were used at a dilution of 1:500: Cy3-conjugated goat anti-rabbit, Cy3-conjugated goat anti-mouse, Cy5-conjugated goat anti-rabbit and Cy5conjugated goat anti-mouse. Secondary antibody fluorescence was measured in transfected and untransfected neurons using an LSM510 confocal (Zeiss). Western blotting. Lysates of Tsc2/ and Tsc2+/+ MEFs, and HEK293T cells (control or 24 h after transfection with LIMK1 or SSH1L), were separated by SDS-PAGE using 8–16% Tris-HCl gels (BioRad). Proteins were transferred onto a polyvinylidene difluoride membrane (PVDF, BioRad) overnight at 4 1C. Membranes were incubated in blocking solution (5% BSA, 0.1% Tween 20 in Tris-buffered saline) for 1 h at 20–24 1C, incubated overnight at 4 1C with primary antibodies (rabbit anti-p-cofilin S3 (1:200, Santa Cruz), rabbit anti-pcofilin S3 (1:1,000, Cell Signalling), rabbit anti-cofilin (1:1,000, Cell Signalling) or rabbit anti-a-tubulin (1 mg/ml, AbCAM), washed and incubated for 1 h at 20–24 1C with horseradish peroxidase (HRP)-conjugated goat anti-rabbit (1:20,000, BioRad). Membranes were washed and incubated for 5 min with chemiluminescent substrate (Pierce) before exposure to film. Band densitometry was performed using Quantity One 4.5.0 software (BioRad). Two-photon laser scanning microscopy and image analysis. Neurons were imaged with custom-built two-photon laser scanning microscopes49 with an excitation wavelength of 910 nm. Images of transfected pyramidal neurons were acquired at 0.8 zoom (image field, 300 270 mm), whereas spiny regions of basal and apical dendrites were imaged at 5 magnification (image field, 42 42 mm). Optical sections were taken at 1.0-mm spacing. Spine density, length and width, as well as soma size, were measured manually using custom software50 by observers who were blind to genotype. Spine lengths were measured from the junction with the dendritic shaft to the tip. To determine head width and primary dendrite thickness, the fluorescence was measured in a line across each structure and the width of the distribution where fluorescent intensity fell to 30% of maximum was calculated. Measurements performed on 100-nm diameter yellow-green fluorescent microspheres (FluorSpheres, Molecular Probes) indicated that the point-spread function placed a lower limit on measurable widths of 550 nm. The apparent width is the convolution of the true fluorescence distribution and point-spread function of the microscope and has a lower limit of B550 nm. Therefore, the summaries of spine head widths are plotted from this lower bound. Soma cross-sectional area was measured in the maximum intensity projection of a low-power image stack by counting the number of pixels within an outline drawn around the soma. Proximal dendrites of Tsc1C/C control and Cre-expressing neurons were of similar thickness (control, 1.26 ± 0.1 mm; Cre, 1.49 ± 0.1 mm; n ¼ 9–13 cells). Electrophysiology. Hippocampal slice cultures from Tsc1C/C mice were placed in a recording chamber perfused with artificial cerebrospinal fluid (ACSF) containing 127 mM NaCl, 25 mM NaHCO3, 1.25 mM Na2HPO4, 2.5 mM KCl, 2 mM CaCl2, 1 mM MgCl2, 25 mM glucose and saturated with 95% O2, 5% CO2 at 20–24 1C. Whole-cell voltage-clamp recordings were obtained using an Axopatch 200B amplifier (Axon Instruments) from Cre-transfected pyramidal neurons (green fluorescence and visible gold particle in soma) and

1733


ARTICLES untransfected neighbors at 10–12 DPT and 18–20 DPT. Transfection with GFP alone had no effect on membrane properties (data not shown). Borosilicate glass pipettes (3–5 MO tip resistance; Warner Instruments Inc.) were filled with 120 mM cesium methane-sulfonate, 10 mM HEPES, 10 mM EGTA, 4 mM MgCl2, 0.4 mM NaGTP, 4 mM MgATP, 10 mM phosphocreatine and 0.02 mM Alexa Fluor-594 (Molecular Probes) at a pH of 7.3 (290 mOsm). Bicuculline (20 mM, Tocris) was added to the bath to block GABAA receptors in all experiments and tetrodotoxin (1 mM, Sigma) was included to block sodium channels for mEPSC recordings. Series resistance (8–19 MO, o20 MO for inclusion in data set), input resistance and membrane capacitance were monitored online. mEPSC frequency and amplitude were analyzed in Igor Pro (Wavemetrics) using custom software. EPSCs were evoked at 0.125 Hz with a bipolar electrode placed in the stratum radiatum 250–350 mm from the soma. The AMPAR/NMDAR current ratio was calculated from the ratio of the EPSC peak amplitude at –60 mV to the current at +40 mV 100 ms after the peak. To measure PPF, paired stimuli at an interpulse interval of 50 ms were delivered and the ratio of peak amplitudes of the EPSCs was calculated. Statistics. Statistical significance was tested by analysis of variance (ANOVA) with, when appropriate, a Tukey-Kramer correction for multiple pairwise comparisons in Matlab (Mathworks) or Microsoft Excel. Distributions of mEPSC amplitudes and immunostaining intensities were compared using KolmogorovSmirnov tests. Band intensities of western blots were compared with paired ttests. Summary data is presented as mean ± standard error of the mean (s.e.m.). Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS We thank members of the Sabatini lab, D. Schmucker and D. Sabatini for comments on the manuscript; E. Hong, A. Carter and R. Witt for technical assistance and advice; and K. Mizuno, A. Minden, S. Dymecki, E. Henske, V. Ramesh, L. Cantley and Y. Shi for the gift of reagents. This work was supported by the US National Institutes of Health (5T32 NS07484) (to V.A.A.), a Burroughs Wellcome Fund Career Award, the Searle Scholars Fund, the Giovanni Armenise Foundation, the Smith Family Foundation and the US Department of Defense (TS030004). COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/ 1. Gomez, M.R., Sampson, J.R. & Whittemore, V.H. (eds.) Tuberous Sclerosis Complex (Oxford Univ. Press, New York, 1999). 2. Ito, N. & Rubin, G.M. Gigas, a Drosophila homolog of tuberous sclerosis gene product-2, regulates the cell cycle. Cell 96, 529–539 (1999). 3. Potter, C.J., Huang, H. & Xu, T. Drosophila Tsc1 functions with Tsc2 to antagonize insulin signaling in regulating cell growth, cell proliferation, and organ size. Cell 105, 357–368 (2001). 4. Tapon, N., Ito, N., Dickson, B.J., Treisman, J.E. & Hariharan, I.K. The Drosophila tuberous sclerosis complex gene homologs restrict cell growth and cell proliferation. Cell 105, 345–355 (2001). 5. Gao, X. & Pan, D. TSC1 and TSC2 tumor suppressors antagonize insulin signaling in cell growth. Genes Dev. 15, 1383–1392 (2001). 6. Potter, C.J., Pedraza, L.G. & Xu, T. Akt regulates growth by directly phosphorylating Tsc2. Nat. Cell Biol. 4, 658–665 (2002). 7. Manning, B.D., Tee, A.R., Logsdon, M.N., Blenis, J. & Cantley, L.C. Identification of the tuberous sclerosis complex-2 tumor suppressor gene product tuberin as a target of the phosphoinositide 3-kinase/akt pathway. Mol. Cell 10, 151–162 (2002). 8. Inoki, K., Li, Y., Zhu, T., Wu, J. & Guan, K.L. TSC2 is phosphorylated and inhibited by Akt and suppresses mTOR signalling. Nat. Cell Biol. 4, 648–657 (2002). 9. Stocker, H. et al. Rheb is an essential regulator of S6K in controlling cell growth in Drosophila. Nat. Cell Biol. 5, 559–565 (2003). 10. Garami, A. et al. Insulin activation of Rheb, a mediator of mTOR/S6K/4E-BP signaling, is inhibited by TSC1 and 2. Mol. Cell 11, 1457–1466 (2003). 11. Hay, N. & Sonenberg, N. Upstream and downstream of mTOR. Genes Dev. 18, 1926– 1945 (2004). 12. Sepp, T., Yates, J.R. & Green, A.J. Loss of heterozygosity in tuberous sclerosis hamartomas. J. Med. Genet. 33, 962–964 (1996). 13. Ramesh, V. Aspects of tuberous sclerosis complex (TSC) protein function in the brain. Biochem. Soc. Trans. 31, 579–583 (2003). 14. Dijkhuizen, P.A. & Ghosh, A. BDNF regulates primary dendrite formation in cortical neurons via the PI3-kinase and MAP kinase signaling pathways. J. Neurobiol. 62, 278–288 (2005).

1734

15. Tang, S.J. et al. A rapamycin-sensitive signaling pathway contributes to long-term synaptic plasticity in the hippocampus. Proc. Natl. Acad. Sci. USA 99, 467–472 (2002). 16. Hou, L. & Klann, E. Activation of the phosphoinositide 3-kinase-Akt-mammalian target of rapamycin signaling pathway is required for metabotropic glutamate receptordependent long-term depression. J. Neurosci. 24, 6352–6361 (2004). 17. Huber, K.M., Kayser, M.S. & Bear, M.F. Role for rapid dendritic protein synthesis in hippocampal mGluR-dependent long-term depression. Science 288, 1254–1257 (2000). 18. Uhlmann, E.J. et al. Astrocyte-specific TSC1 conditional knockout mice exhibit abnormal neuronal organization and seizures. Ann. Neurol. 52, 285–296 (2002). 19. Kohn, A.D. et al. Construction and characterization of a conditionally active version of the serine/threonine kinase Akt. J. Biol. Chem. 273, 11937–11943 (1998). 20. Meng, Y., Zhang, Y., Tregoubov, V., Falls, D.L. & Jia, Z. Regulation of spine morphology and synaptic function by LIMK and the actin cytoskeleton. Rev. Neurosci. 14, 233–240 (2003). 21. Zhou, Q., Homma, K.J. & Poo, M.M. Shrinkage of dendritic spines associated with longterm depression of hippocampal synapses. Neuron 44, 749–757 (2004). 22. Zhang, H. et al. Loss of Tsc1/Tsc2 activates mTOR and disrupts PI3K-Akt signaling through downregulation of PDGFR. J. Clin. Invest. 112, 1223–1233 (2003). 23. Mashimo, J., Maniwa, R., Sugino, H. & Nose, K. Decrease in the expression of a novel TGF beta1-inducible and ras-recision gene, TSC-36, in human cancer cells. Cancer Lett. 113, 213–219 (1997). 24. Tuttle, R.L. et al. Regulation of pancreatic beta-cell growth and survival by the serine/ threonine protein kinase Akt1/PKBalpha. Nat. Med. 7, 1133–1137 (2001). 25. Verdu, J., Buratovich, M.A., Wilder, E.L. & Birnbaum, M.J. Cell-autonomous regulation of cell and organ growth in Drosophila by Akt/PKB. Nat. Cell Biol. 1, 500– 506 (1999). 26. Kwon, C.H. et al. Pten regulates neuronal soma size: a mouse model of Lhermitte-Duclos disease. Nat. Genet. 29, 404–411 (2001). 27. Backman, S.A. et al. Deletion of Pten in mouse brain causes seizures, ataxia and defects in soma size resembling Lhermitte-Duclos disease. Nat. Genet. 29, 396–403 (2001). 28. Dong, J. & Pan, D. Tsc2 is not a critical target of Akt during normal Drosophila development. Genes Dev. 18, 2479–2484 (2004). 29. Sarbassov, D.D., Guertin, D.A., Ali, S.M. & Sabatini, D.M. Phosphorylation and regulation of Akt/PKB by the rictor-mTOR complex. Science 307, 1098–1101 (2005). 30. Lamb, R.F. et al. The TSC1 tumour suppressor hamartin regulates cell adhesion through ERM proteins and the GTPase Rho. Nat. Cell Biol. 2, 281–287 (2000). 31. Govek, E.E. et al. The X-linked mental retardation protein oligophrenin-1 is required for dendritic spine morphogenesis. Nat. Neurosci. 7, 364–372 (2004). 32. Penzes, P. et al. The neuronal Rho-GEF Kalirin-7 interacts with PDZ domain-containing proteins and regulates dendritic morphogenesis. Neuron 29, 229–242 (2001). 33. Penzes, P. et al. Rapid induction of dendritic spine morphogenesis by trans-synaptic ephrinB-EphB receptor activation of the Rho-GEF kalirin. Neuron 37, 263–274 (2003). 34. Li, Y., Inoki, K., Yeung, R. & Guan, K.L. Regulation of TSC2 by 14–3-3 binding. J. Biol. Chem. 277, 44593–44596 (2002). 35. Hengstschlager, M., Rosner, M., Fountoulakis, M. & Lubec, G. Tuberous sclerosis genes regulate cellular 14–3-3 protein levels. Biochem. Biophys. Res. Commun. 312, 676–683 (2003). 36. Nufer, O. & Hauri, H.P. ER export: call 14–3-3. Curr. Biol. 13, R391–R393 (2003). 37. Gohla, A. & Bokoch, G.M. 14–3-3 regulates actin dynamics by stabilizing phosphorylated cofilin. Curr. Biol. 12, 1704–1710 (2002). 38. Benvenuto, G. et al. The tuberous sclerosis-1 (TSC1) gene product hamartin suppresses cell growth and augments the expression of the TSC2 product tuberin by inhibiting its ubiquitination. Oncogene 19, 6306–6316 (2000). 39. Sarbassov, D.D. et al. Rictor, a novel binding partner of mTOR, defines a rapamycininsensitive and raptor-independent pathway that regulates the cytoskeleton. Curr. Biol. 14, 1296–1302 (2004). 40. Arber, S. et al. Regulation of actin dynamics through phosphorylation of cofilin by LIMkinase. Nature 393, 805–809 (1998). 41. Tassabehji, M. et al. LIM-kinase deleted in Williams syndrome. Nat. Genet. 13, 272– 273 (1996). 42. Hayashi, M.L. et al. Altered cortical synaptic morphology and impaired memory consolidation in forebrain-specific dominant-negative PAK transgenic mice. Neuron 42, 773–787 (2004). 43. Bryan, B. et al. GEFT, a Rho family guanine nucleotide exchange factor, regulates neurite outgrowth and dendritic spine formation. J. Biol. Chem. 279, 45824–45832 (2004). 44. Meng, Y. et al. Abnormal spine morphology and enhanced LTP in LIMK-1 knockout mice. Neuron 35, 121–133 (2002). 45. Chang, H. et al. Identification of a cDNA encoding a thiazide-sensitive sodium-chloride cotransporter from the human and its mRNA expression in various tissues. Biochem. Biophys. Res. Commun. 223, 324–328 (1996). 46. Shepherd, C.W., Houser, O.W. & Gomez, M.R. MR findings in tuberous sclerosis complex and correlation with seizure development and mental impairment. AJNR Am. J. Neuroradiol. 16, 149–155 (1995). 47. Boyer, C., Schikorski, T. & Stevens, C.F. Comparison of hippocampal dendritic spines in culture and in brain. J. Neurosci. 18, 5294–5300 (1998). 48. Stoppini, L., Buchs, P.A. & Muller, D.A. A simple method for organotypic cultures of nervous tissue. J. Neurosci. Methods 37, 173–182 (1991). 49. Carter, A.G. & Sabatini, B.L. State dependent calcium signaling in dendritic spines of striatal medium spiny neurons. Neuron (2004). 50. Trachtenberg, J.T. et al. Long-term in vivo imaging of experience-dependent synaptic plasticity in adult cortex. Nature 420, 788–794 (2002).

VOLUME 8

[

NUMBER 12

[



ARTICLES

The prolactin-releasing peptide antagonizes the opioid system through its receptor GPR10 Patrick Laurent1,4, Jerome A J Becker1,4, Olga Valverde2, Catherine Ledent1, Alban de Kerchove d’Exaerde3, Serge N Schiffmann3, Rafael Maldonado2, Gilbert Vassart1 & Marc Parmentier1 Prolactin-releasing peptide (PrRP) and its receptor G protein–coupled receptor 10 (GPR10) are expressed in brain areas involved in the processing of nociceptive signals. We investigated the role of this new neuropeptidergic system in GPR10-knockout mice. These mice had higher nociceptive thresholds and stronger stress-induced analgesia than wild-type mice, differences that were suppressed by naloxone treatment. In addition, potentiation of morphine-induced antinociception and reduction of morphine tolerance were observed in mutants. Intracerebroventricular administration of PrRP in wild-type mice promoted hyperalgesia and reversed morphine-induced antinociception. PrRP administration had no effect on GPR10-mutant mice, showing that its effects are mediated by GPR10. Anti-opioid effects of neuropeptide FF were found to require a functional PrRP-GPR10 system. Finally, GPR10 deficiency enhanced the acquisition of morphine-induced conditioned place preference and decreased the severity of naloxone-precipitated morphine withdrawal syndrome. Altogether, our data identify the PrRP-GPR10 system as a new and potent negative modulator of the opioid system.

Opiate drugs, the prototype of which is morphine, are used largely for the treatment of severe pain. However, the prolonged use of opiate drugs induces a behavioral adaptation that results in the development of tolerance and dependence1. Although these adaptive mechanisms have been known for decades, the underlying pathophysiological pathways have not been fully clarified. It has been proposed that the opioid receptor signaling pathway is adaptively regulated at the cellular level, although more recently a growing contribution of the plasticity of neuronal networks involving opioidergic neurons has been recognized. In support of this latter hypothesis, a number of neuropeptides, including cholecystokinin (CCK)2, neuropeptide FF (NPFF)3 and nociceptin (orphanin FQ)4, have been proposed as modulators of the opioid system. These various peptidergic pathways are collectively designated as an anti-opioid system. However, the action of individual peptides on the opioid pathway and on the processing of nociceptive signals is complex and some aspects remain controversial, particularly as the behavioral consequences can vary depending on the injection site of these peptides and the precise experimental setup5,6. Among these peptides, NPFF, which belongs to the RF-amide peptide family, was shown (among other actions) to regulate blood pressure7, prolactin release8 and nociceptive signal processing9. Prolactin-releasing peptide (PrRP) has recently been identified as an additional member of the mammalian RF-amide peptide family, following its isolation as the natural agonist of the previously orphan GPR10 (ref. 10) (also known as hGR3 in human or UHR-1 in rat). Intracerebroventricular (i.c.v.) administration of PrRP affects feeding

behavior11, blood pressure12 and neuroendocrine processes, such as corticotropin-releasing hormone (CRH)13 and oxytocin14 release. In addition, recent observations suggest that PrRP and GPR10 are involved in the processing of nociceptive information in vivo. Indeed, PrRP and GPR10 expression is restricted to a limited number of structures in the rat central nervous system known to be involved in the processing of nociceptive signals15,16. These include the hypothalamus, amygdala and brainstem areas. In the same line, the coexpression of PrRP and proenkephalin has been demonstrated in amygdala neurons17, and the activity of PrRP as a modulator of pain and allodynia has recently been described in rats18. In the present study, we generated Gpr10-deficient mice in order to determine the interaction of PrRP and its receptor GPR10 with the opioid system in vivo. Our observations identify the PrRP-GPR10 system as a new and potent negative modulator of the opioid system. RESULTS Knockout of GPR10 in mice In our targeting vector, part of the Gpr10 coding region was replaced by a tau-lacZ fusion gene (placed under control of the natural Gpr10 promoter) and selection cassettes (Fig. 1a–c). After the generation of the knockout mice, Gpr10 transcripts were amplified from brain by reverse transcription polymerase chain reaction (RT-PCR). As expected, no amplification occurred in knockout mice (Fig. 1d). Gpr10-knockout mice were fertile, transmitted the null allele with the expected mendelian frequency and did not show obvious morphological abnormalities.

1Institut de Recherche Interdisciplinaire en Biologie Humaine et Mole ćulaire (I.R.I.B.H.M.), Universite´ Libre de Bruxelles, Campus Erasme, Route de Lennik 808, Brussels,

Belgium. 2Laboratori de Neurofarmacologia, Facultat de Cie`nces de la Salut i de la Vida, Universitat Pompeu Fabra, Barcelona, Spain. 3Laboratory of Neurophysiology, Universite´ Libre de Bruxelles, Campus Erasme, Route de Lennik 808, Brussels, Belgium. 4These authors contributed equally to this work. Correspondence should be addressed to M.P. ([email protected]). Received 2 August; accepted 30 September; published online 20 November 2005; doi:10.1038/nn1585


[

NUMBER 12

[

DECEMBER 2005

1735

ARTICLES

S

H

X

B;H

GPR10

–– f r

Wild type H: 7.6 kb B: 11 kb

S

B Tau

LacZ

TK

Tau

LacZ

TK

H

B

Knockout

X

H

X

B;H

Neo

–– nr

11 kb

700 bp

6 kb

500 bp

d +/+

+/–

–/–

+/– –/– +/+

+/–

–/–

Hypothalamus pituitary adrenal axis PrRP has been found to control the secretion of adrenocorticotropic hormone (ACTH), through corticotropin releasing hormone (CRH), and a number of stressful conditions have been reported to activate PrRP-containing neurons in rats, indicating a likely role in the control of the stress axis13. We therefore assessed the glucocorticoid levels in plasma, both under basal conditions and following a hypoglycemic stress, in wild-type and knockout mice. The basal corticosterone levels measured at 8 a.m. and at 4 p.m. were slightly lower in knockout mice, and the difference became highly significant following insulin-induced hypoglycemia (Fig. 2a). We measured CRH transcripts hypothalamus using quantitative RT-PCR. We did not observe significant differences in basal conditions, but when the mice were subjected to insulin-induced hypoglycemia, CRH transcripts were less abundant in knockout mice than in wildtype mice (Fig. 2b). Behavioral characterization In a first set of experiments, we explored anxiety-related behaviors and locomotor activity for both genotypes. The elevated plus maze and the light and dark box protocols were used to evaluate the anxiety-related responses. In the elevated plus maze test, the time spent (wild type: 35.2 ± 2.1 s; knockout: 38.5 ± 2.1 s, n ¼ 17–20) and the number of entries (wild type: 17.8 ± 1.0; knockout: 14.7 ± 1.0; n ¼ 17–20) in the open arms were not significantly different between genotypes (P ¼ 0.28, P ¼ 0.051, respectively). Similarly, in the light and dark box test, no difference was observed for the latency of the first entry (wild type: 16.9 ± 1.9 s; knockout: 14 ± 1.9 s, n ¼ 17–18) or the total time spent in

Figure 2 Activity of the hypothalamus-pituitary-adrenal (HPA) axis. (a) Corticosterone levels were measured in wild-type (open bars) and knockout (black bars) animals in resting conditions at 8 a.m. and 4 p.m., and during insulin-induced hypoglycemia. n ¼ 5–13 depending on group. (b) Abundance of CRH-encoding transcripts, as determined by quantitative RT-PCR, in the hypothalamus of wild-type (open bar) and knockout (black bar) animals subjected to insulin-induced hypoglycemic stress. n ¼ 19 or 20 animals depending on group. Values represent the mean ± s.e.m. *P o 0.05; **P o 0.01. Statistical significance was determined using the Student’s t-test.

the lit compartment (wild type: 100.7 ± 5.2 s; knockout: 104.5 ± 3.9 s, n ¼ 17–18). Motor coordination was not affected in knockout mice as evidenced by the rotarod test (wild type: 294 ± 22 s; knockout: 281 ± 23 s, n ¼ 15). In actimetry boxes, the spontaneous locomotor activity (wild type: 457 ± 26 counts; knockout: 453 ± 32 counts, n ¼ 30) and activity of mice administered morphine (10 mg kg1) were similar for both genotypes (wild type: 402 ± 22 counts; knockout: 396 ± 39 counts, n ¼ 10) and were significantly greater than for mice administered saline (wild type: 221 ± 36 counts, P o 0.001; knockout: 229 ± 24 counts, P o 0.0001, n ¼ 10). In a second set of experiments, we measured the spontaneous thermal nociceptive threshold in wild-type and knockout mice using two tests reported to explore stimulus integration preferentially at spinal (tail-immersion) or supraspinal (hot-plate) levels. In the tailimmersion test, we observed similar nociceptive thresholds for both genotypes (Fig. 3a). In the hot-plate test (52 1C), we observed increased jump latency in knockout mice as compared to wild-type mice (Fig. 3b). Similar data were obtained when the hot-plate was set to 48 1C, 50 1C and 54 1C (data not shown), whereas no difference was seen at 56 1C. Corticosteroid supplementation did not reverse the modified nociceptive threshold of the knockout mice (Supplementary Fig. 1 online). Administration of the opioid receptor antagonist naloxone (1 mg kg1, intraperitoneally (i.p.)) fully reversed the relative analgesia observed in knockout mice in the hot-plate test, but did not affect the nociceptive threshold of wild-type animals. Notably, naloxone administration also decreased the nociceptive threshold of knockout mice in the tail-immersion test below that of treated or untreated controls (Fig. 3a). We then investigated the antinociceptive effects of morphine in the tail-immersion (52 1C) and hot-plate (56 1C) tests, conditions in which the basal nociceptive threshold is identical for both genotypes (Fig. 3c,d). As expected, morphine induced a dose-dependent antinociception in both genotypes, but this effect was significantly enhanced in knockout mice for the 10 mg kg1 (i.p.) dose. To complement these observations, we also investigated mechanical and

a

600

**

500 400 300

NS

*

200 100 0 8 a.m.

4 p.m.

Hypoglycemia

b

150

* 100

50

0

1736

pe oc ko ut

b

c

1 kb

ty

Pr

ild

H: 5.4 kb B: 6 kb

+/+

H

Neo

VOLUME 8

[

NUMBER 12

[

Kn

B

H

W


S

Figure 1 GPR10 gene targeting. (a) The structures of the wild-type Gpr10 locus (wild type), the targeting vector and the locus resulting from homologous recombination (knockout) are shown. The location of the restriction sites—BamHI (B), SacI (S), XhoI (X) and HindIII (H)—of the probe used in Southern blotting experiments (Pr), and of the forward (f), reverse (r) and Neo (n) primers used for the PCR genotyping, are indicated. The size of the bands expected after HindIII digestion and hybridization with the probe are 7.6 kb and 5.4 kb for the wild-type and knockout alleles, respectively. After BamHI digestion, the expected sizes are, respectively, 11.0 kb and 6.0 kb. (b) Southern blot illustrating the genotyping of wild-type (+/+), heterozygous (+/) or knockout (/) mice, following digestion of genomic DNA by BamHI. (c) Genotyping of mice by PCR using a three primers procedure. (d) Detection of GPR10 transcripts by RT-PCR in total RNA prepared from whole brain.

CRH ARNm (copy per 106β-actin copies)

B

Corticosterone (ng ml–1)

a


ARTICLES Behavioral effects of centrally administered peptides 6 3 Together, the data reported in the previous 200 sets of experiments suggest an interaction 100 4 2 between the opioid system and Gpr10. To 100 1 2 investigate this interaction further, we evalu0 0 0 0 ated the consequences of centrally adminis2 10 – – + + – – + + 0 0 2 10 Naloxone (1 mg kg–1) Naloxone (1 mg kg–1) Morphine (mg kg–1) Morphine (mg kg–1) tered PrRP (5 nmol, i.c.v.) on the nociceptive threshold recorded in the tail-immersion test e 800 f g h 4 NS 40 (50 1C, Fig. 4a). I.c.v. administration of PrRP 300 NS 3 NS promoted a hyperalgesia in wild-type mice 30 600 200 NS 2 and was able to reverse the analgesia induced 20 by morphine. Administration of PrRP to 100 1 400 10 knockout mice had no effects on nociceptive 0 0 0 thresholds, showing that the effects of the 1 2 0 2 10 0 Basal Stress Stress Basal Stress Stress +Nal +Nal Morphine (mg kg–1) Morphine (mg kg–1) peptide in vivo are exclusively mediated by the GPR10 receptor. Figure 3 Nociceptive thresholds in GPR10-knockout mice. (a–h) The nociceptive thresholds were We also investigated the potential interacmeasured in the tail-immersion test (a,c,g), the hot-plate test (jumping response: b,d,h), the tail-pressure tion of the PrRP-GPR10 system with pretest (e) and the writhing test (f) on wild-type (open bars) and knockout (black bars) animals. (a,b) Basal nociceptive threshold and effect of naloxone treatment. (c,d) Antinociceptive response to morphine viously characterized anti-opioid peptidergic (2 mg kg1 and 10 mg kg1 i.p.).(e) Nociceptive threshold in the tail-pressure test and response to systems, namely NPFF and nociceptin, using morphine analgesia (2 mg kg1 and 10 mg kg1 i.p.) (f) Number of writhes in the writhing test and the tail-immersion test as readout (Fig. 4b,c). response to morphine analgesia (1 mg kg1 and 2 mg kg1 i.p.). (g,h) Nociceptive threshold after a I.c.v. administration of nociceptin promoted 5-min forced swim stress (water temperature: 32 ± 1 1C, 5 min) and effect of naloxone treatment. hyperalgesia and reversed morphine-induced Values represent the mean ± s.e.m. n ¼ 7–15 animals depending on group. *P o 0.05; **P o 0.01; analgesia, without significant differences ***P o 0.001 (ANOVA followed by Dunnett t-test). (P ¼ 0.29, P ¼ 0.20, respectively) between wild-type and knockout mice (Fig. 4c). In our visceral pain paradigms, using the tail-pressure and writhing tests, conditions, NPFF did not modify the basal nociceptive threshold of respectively (Fig. 3e,f). In the tail-pressure test, we observed a relative wild-type mice, but antagonized the effects of morphine (Fig. 4b). analgesia for knockout mice as compared to control animals in basal Notably, in knockout animals, NPFF had an analgesic effect and was conditions (Fig. 3e). We observed the analgesic effect of morphine unable to reverse the effects of morphine. These observations suggest (2 mg kg1 and 10 mg kg1) for both genotypes, although the relative an interaction between the NPFF and PrRP systems, probably at the increments in latency were difficult to assess as the values got close neuronal network level. to the predefined cut-off of 800 s. In the writhing test, we obtained an analgesic effect of morphine, with no difference between the Opiate tolerance, dependence and rewarding properties Repeated opiate administration leads to the development of tolerance genotypes (Fig. 3f). Environmental stress is known to promote potent inhibition of of its analgesic effects. We evaluated morphine tolerance during a 5-d behavioral responses to nociceptive stimuli, and this stress-induced morphine exposure protocol in the tail-immersion test (52 1C, Fig. 5a). analgesia is mediated by the opioid system. As expected, a forced swim As expected, morphine (10 mg kg1, i.p.) induced a strong antinocitest19 (5 min at 32 1C) induced a marked analgesia in the hot-plate test ception on day 1; from day 2 to day 5, the antinociceptive response to (56 1C), which was reversed by naloxone in both genotypes (Fig. 3g). In morphine decreased slowly, down to basal level in both genotypes. The agreement with the response observed after morphine administration, kinetics of tolerance development were, however, delayed markedly in the stress-induced analgesia was stronger in knockout mice. No the knockout mice as compared to the wild-type mice. Chronic morphine exposure also produces a strong physical depenanalgesia was observed in the tail-immersion test, but the effect of dence syndrome as assessed by the characteristic set of behavioral naloxone on knockout mice was maintained (Fig. 3h).

a

b 200

4

c

d

8

Latency (s)

Latency (s)

b

NS

Latency (s)

Latency (s)

6

NS

4 2

0 PrRP i.c.v. Morphine i.p.

c

NS

10

8

8

NS

6 4

+ –

+ +

– +

– –

+ –

+ +

– +

0 NPFF i.c.v. Morphine i.p.

8 6 4 2

2

– –

10 Latency (s)

10

a

Latency (s)

Latency (s)

Latency (s)

Writhes (n)

Latency (s)


Latency (s)

300

– –

+ –

+ +

– +

– –

+ –

+ +

– +

0 Nociceptin i.c.v. Morphine i.p.

– –

+ –

+ +

– +

– –

+ –

+ +

– +

Figure 4 Anti-opiate properties of PrRP, NPFF and nociceptin. (a–c) Nociceptive thresholds were measured in the tail-immersion test. (a) PrRP (0 nmol or 5 nmol, i.c.v.) and morphine (0 mg kg1 or 10 mg kg1, i.p.), (b) NPFF (0 nmol or 5 nmol, i.c.v.) and morphine (0 mg kg1 or 10 mg kg1, i.p.) and (c) nociceptin (0 nmol or 25 nmol, i.c.v) and morphine (0 mg kg1 or 10 mg kg1, i.p.) were coadministered to wild-type (open bars) and knockout (black bars) mice. Results are expressed as mean ± s.e.m. n ¼ 7–14 animals depending on group. *P o 0.05; **P o 0.01; ***P o 0.001 (ANOVA followed by Dunnett t-test).


[

NUMBER 12

[

DECEMBER 2005

1737

ARTICLES Figure 5 Effects of the PrRP-GPR10 system on 5.5 morphine tolerance, dependence and rewarding. 5.0 250 (a) Tolerance was evaluated in the tail-immersion 150 4.5 test. Morphine (10 mg kg1, i.p, , J) or vehicle 150 4.0 (m, n) was administered twice daily (7 a.m. and 100 50 7 p.m.) for 5 d to wild-type (J, n) or knockout 3.5 50 (, m) mice. The antinociceptive effect of 3.0 –50 morphine was evaluated every morning. 2.5 0 0 2 6 –1 1 2 3 4 5 Saline Morphine (b) Dependence was evaluated in the withdrawal Days Morphine (mg kg–1, s.c.) test. Wild-type (open bars) and knockout (black bars) mice were treated for 5 d with morphine (from 20 mg kg1 to 100 mg kg1, i.p.) or with the saline vehicle as control, and withdrawal signs were evaluated after naloxone administration (1 mg kg1, s.c.). The global score was calculated for each mouse. (c) The morphine-induced rewarding response was evaluated in the place conditioning paradigm for wild-type (open bars) and knockout (black bars) mice. Morphine doses were 2 mg kg1 or 6 mg kg1, s.c. *P o 0.05; **P o 0.01; ***P o 0.001 when compared to the saline group of the same genotype. $P o 0.05; $$P o 0.01 for the comparison between genotypes (ANOVA followed by Dunnett t-test). Results are expressed as mean ± s.e.m. n ¼ 8–14 animals depending on group.

b

responses to naloxone-evoked withdrawal. This includes a number of somatic signs such as jumps, wet-dog shakes, paw tremors, sniffing, ptosis, mastication, piloerection and body tremor, and vegetative signs such as diarrhea. We observed evidence of physical dependence for both genotypes during abstinence. However, the incidence of signs of withdrawal was significantly lower in knockout mice than in wild-type mice (Supplementary Fig. 2 online). The global withdrawal score (Fig. 5b) illustrates the attenuation of the severity of naloxoneprecipitated morphine withdrawal syndrome in knockout mice. We investigated the rewarding properties of opiates in a conditioned place-preference paradigm. When morphine is administered to mice in this setting, the dose-response curve is reportedly bell shaped, with the maximal effect in mice at 6 mg kg1 (s.c.) of morphine20. In our experiments (Fig. 5c), the maximal effect of morphine was indeed obtained for the 6 mg kg1 dosage in wild-type animals. However, the knockout mice responded maximally to the lower dose of morphine (2 mg kg1, s.c.), whereas a higher dose did not induce any rewarding response. This observation suggests an increased sensitivity to the rewarding properties of morphine in the absence of GPR10. Acute or chronic opiate administration affects gene expression in various brain areas. We measured the PrRP transcript levels using quantitative RT-PCR in hypothalamus and brainstem following morphine treatment. In the tolerance test, we observed that 3 h after the last morphine injection, the PrRP transcript levels were upregulated in the hypothalamus (Supplementary Fig. 3 online). Evaluation of the opioid and anti-opioid systems GPR10 is expressed in enkephalin-containing neurons21, suggesting that it may be involved in the control of proenkephalin expression. To assess whether the relative analgesia observed in knockout mice might

Score

Global score

c

be related to increased proenkephalin expression, we investigated the distribution of proenkephalin transcripts in the brain of both genotypes by in situ hybridization. We measured similar levels of proenkephalin transcripts in various brain areas of wild-type and knockout animals, including several pain-associated areas such as the parabrachial nucleus, the dorsal raphe nucleus, the bed nucleus of stria terminalis and the central amygdala (Fig. 6a). We observed a modest upregulation in basal ganglia (caudate-putamen and nucleus accumbens), in which GPR10 is poorly expressed, thus suggesting an indirect consequence of Gpr10 inactivation. Using [3H]DAMGO as radioligand, we also characterized the m-binding sites in a saturation binding assay. Similar Kd (mean ± s.e.m.: 1.34 ± 0.25 nM versus 1.22 ± 0.36 nM) and Bmax (155 ± 32 fmol mg1 versus 163 ± 54 fmol mg1 protein) values were found on whole brain membranes from wild-type and knockout mice, respectively. We also tested the d-opioid receptor, using [3H]naltrindole as tracer in a binding assay on whole brain membranes, but we found no difference between wild-type and knockout mice in KD (0.39 ± 0.07 nM versus 0.44 ± 0.05 nM, respectively) or Bmax (93 ± 9 fmol mg1 versus 85 ± 12 fmol mg1 of protein). The functional response of the m and d receptors was evaluated on whole brain membranes and on brain slices in [35S]GTPgS binding assays. On whole brain membranes, the EC50 and Emax of the m receptor (378 ± 15 nM versus 374 ± 13 nM, and 130 ± 8% versus 132 ± 12% relative to basal values, for wild-type and knockout mice, respectively) and the EC50 and Emax of the d receptor (731 ± 45 nM versus 860 ± 43 nM; 121 ± 22% versus 128 ± 18% relative to basal values) were similar for both groups of mice. No differences in the binding distribution were detected on brain sections for the m receptor (Supplementary Table 1 online). We also evaluated pronociceptin and NPFF transcript levels by in situ hybridization, quantitative RT-PCR (Fig. 6b,c) or both. Both transcripts were found

Figure 6 Modulation of the opioid and anti-opioid 0.20 30 0.7 systems. (a) Proenkephalin transcript levels were 0.6 evaluated in various brain areas of wild-type (open 0.15 28 0.5 bars) and knockout (black bars) animals by in situ hybridization (n ¼ 5). (b) PrRP, NPFF, nociceptin 0.4 0.10 26 (Noc) and HPRT (as control) transcript levels 0.3 were evaluated in the brainstem of wild-type 0.2 (open bars) and knockout (black bars) mice by 24 0.05 quantitative RT-PCR. ‘‘Ct’’ values represent the 0.1 threshold cycle number at which fluorescence 22 0.0 0.00 increases above a fixed threshold value. n ¼ 8 wild-type and 14 knockout mice. (c) Nociceptin transcript levels were evaluated in two brain areas of wild-type (open bars) and knockout (black bars) mice by in situ hybridization (n ¼ 5 mice). Results are expressed as mean ± s.e.m. ***P o 0.001 (ANOVA followed by Dunnett t-test). D Raph, dorsal raphe; LPB, parabrachial nucleus; C Pu, caudate putamen; N Ac, nucleus accumbens; Olf T, olfactory tract; BNST, bed nucleus of stria terminalis; Amy, amygdala; LS, lateral septum; Rt, Reticular thalamus.

b

[

NUMBER 12

[

t R t

R

LS LS

Ct

VOLUME 8

Pr R Pr P R N P P N FF PF F N oc N o H c P H RT PR T

D R D aph R ap h LP B LP B C P C u Pu N A N c A O c If T O I BN f T S BN T ST Am y Am y

1738

c

Optical density

a

Optical density


Latency (s)

a


ARTICLES


at similar levels in several brain areas of wild-type and knockout mice. From these experiments, therefore, it seems that no gross alteration of the opioid and anti-opioid systems in pain-associated areas occur as a compensatory mechanism following Gpr10 knockout in mice. DISCUSSION Overall, our observations of knockout animals have confirmed that the PrRP-GPR10 system is involved in the control of neuroendocrine functions. Indeed, PrRP seems to regulate the level of CRH transcripts in hypothalamic structures, in agreement with the expression of the peptide precursor in catecholaminergic A1/A2 neurons. Together with the locus coeruleus, these neurons have been shown to control the release of CRH by the paraventricular nucleus (PVN) neurons13,22, which are known to express Gpr10. These results are also in line with the previous observation that i.c.v. administration of PrRP enhances neuronal activity in the PVN23, and increases ACTH and corticosterone levels in blood, through the release of CRH13. Moreover, we also demonstrated an important additional role of the PrRP-GPR10 system in the modulation of the various actions of opiates. Knockout mice showed a higher nociceptive threshold in some settings, increased analgesic and rewarding effects of morphine, increased stress-induced analgesia and reduced tolerance of morphine, as well as naloxone-precipitated withdrawal symptoms. Notably, naloxone reversed the analgesic effects of morphine, but also affected basal levels of the knockout animals in several tests, suggesting that Gpr10 disruption generates or unmasks a basal opioid tone that is not detectable in wild-type animals. In line with these observations, we also observed that i.c.v. administration of PrRP results in hyperalgesia and reverses the antinociceptive effects of morphine in wild-type mice. The absence of effects of the peptide in knockout mice demonstrates that GPR10 is the sole target of PrRP in vivo. Altogether, these observations establish the involvement of GPR10 in the modulation of pain-signal processing. The observed properties of the PrRP-GPR10 system are reminiscent of those of other opioid-modulating systems described previously. These other opioid-modulating systems (also referred to as anti-opioid systems) include the neuropeptides CCK, NPFF, nociceptin and dynorphin, and their respective receptors. Although the biological actions of these various peptides are not identical, they have all been reported to counteract some of the main behavioral effects of morphine in laboratory animals, including analgesia, tolerance, reward and dependence6. It is now well accepted that these neuropeptidergic systems have a major role in the adaptive mechanisms that characterize the chronic stimulation of the opiate system: namely, tolerance and dependence. Indeed, it is believed that these adaptive mechanisms involve not only regulations at the level of opioid receptors and their signaling cascades, but also long-term plasticity of neuronal networks24. As initially proposed25, opioid receptor stimulation would result in the concomitant activation of anti-opiate neuronal networks, counteracting the various actions of opiate drugs. The delayed but long-lasting stimulation of these anti-opiate systems would, following chronic opioid stimulation, decrease the drug’s effectiveness (tolerance) and contribute to adverse effects upon acute withdrawal (dependence). Despite their overlapping actions, each opioid-modulating peptide has its own set of biological activities, as an obvious consequence of differences in receptor distributions and regulation of peptide release. In relation to the opiate system, NPFF and nociceptin show essentially the same panel of biological activities in vivo. Supraspinal administration of these peptides results in hyperalgesia, decrease of morphineand stress-induced analgesia, reduction of rewarding properties (only demonstrated for nociceptin), enhancement of tolerance of morphine


[

NUMBER 12

[

DECEMBER 2005

and precipitation of morphine withdrawal (although a matter of conflict for nociceptin)5,6,26–28. Spinal effects include analgesia and potentiation of morphine-induced analgesia29. Inactivation of the nociceptin receptor ORL1 leads to a markedly attenuated tolerance to the analgesic effects of morphine but without modification of the basal nociceptive threshold30,31. These actions are markedly similar to those we observed for PrRP. A previous report18 described antinociceptive effects of PrRP following intrathecal administration. Despite the apparent contradiction with our findings, this observation also parallels the reported activities of NPFF and nociceptin. The activities of the two other anti-opioid peptides, dynorphin and CCK, are more dissimilar. Even though both peptides inhibit morphine-induced analgesia at the supraspinal level, they were reported to antagonize morphine action at the spinal level as well32,33, in contrast to nociceptin, NPFF and PrRP. Other differences include the decrease in tolerance of morphine caused by dynorphin and the increase of rewarding effects of morphine caused by CCK, which contrasts with the action of the other anti-opioid peptides (reviewed in ref. 6). Also, knockout of the CCK2 receptor gene led to spontaneous hyperalgesia and enhanced withdrawal signs34, which might be explained by an increase in central endogenous opioid peptide secretion with a paradoxical coupling of the m and d receptors. Given the similarity of the behavioral actions of PrRP, nociceptin and NPFF, we have investigated whether there might be interconnections between these systems. Inactivation of Gpr10 did not result in significant modification of NPFF and nociceptin transcripts in the central nervous system. Nociceptin was also shown to retain its effects (hyperalgesia and reversion of morphine analgesia) in knockout animals, demonstrating that GPR10 is not required for the central effects of this peptide. In contrast, NPFF administration in knockout mice led to paradoxical analgesic effects and did not reverse morphineinduced analgesia (an activity seen in wild-type mice). These results indicate that some of the central actions of NPFF require a functional PrRP system and that inactivation of this system unmasks analgesic properties of NPFF that probably correspond to the spinal effect described in the literature27. It seems, therefore, that the PrRPGPR10 system is located downstream of the NPFF-GPR10 system in a common central anti-opioid pathway. The absence of a receptor during development might cause adaptive changes in the central nervous system. The agreement between the data obtained with the knockout mice and those resulting from PrRP administration suggest, however, that the observed phenotype is not the consequence of such long-term adaptive mechanisms. Our result do not show significant modification in the expression and coupling of the of m- and d-opioid receptors and in the abundance of NPFF and pronociceptin transcripts. Despite the modest upregulation of proenkephalin transcripts in basal ganglia, no major modifications were observed in brain areas involved in pain management. However, the upregulation in nucleus accumbens might contribute to the modifications of reward- and withdrawal-related behaviors observed in the knockout mice. As this area does not constitute a major site of expression of GPR10, it is likely that the proenkephalin overexpression is indirect and reflects a functional modification of the local neuronal networks. Receptors involved in the action of opiate and anti-opiate peptides are essentially coupled to the same intracellular pathways. Indeed, opiate receptors, ORL1, NPFF2 and GPR10 receptors are all coupled to the Gi class of G proteins. Although the actual downstream signaling events have not been formally demonstrated in vivo for each of these receptors, Gi proteins are known to stimulate K+ channels, inhibit voltage-dependent Ca2+ channels and inhibit adenylyl cyclase, all of which result in an inhibition of signaling at the pre- or postsynaptic

1739


ARTICLES levels. The situation is somewhat more complex for CCK receptors, which have been shown to couple both to Gq and Gi family members. GPR10 has also been reported to couple to the Gq class of G proteins in some cell types35. The interference of anti-opioid systems with opiate peptide signaling therefore probably involves their action at different sites of the neuronal network, although interference between two receptors expressed in the same cell has been proposed as well6. Some of our observations might be explained by modulations of m-agonist release or by a modified sensitivity of the network to similar levels of agonists. However, only this latter hypothesis accounts for all of our observations. Indeed, decreased tolerance for morphine is only compatible with a deficit of anti-opioid tone at a post-receptor level. Similarly, the morphine-induced analgesia experiments are not compatible with a presynaptic hypothesis. Indeed, a potentially elevated endogenous opioid tone is expected to be negligible with regard to the i.p. administration of 10 mg kg–1 of morphine. We believe therefore that the effects of PrRP are essentially mediated through counterregulatory mechanisms acting post-synaptically in the opioid pathways. The loss of these counter-regulatory mechanisms in knockout mice would then reveal the weak opioid tone in basal situations and magnify the responses to endogenous or exogenous opioid agonists. PrRP is highly expressed in A1 and A2 noradrenergic neurons of the caudal medulla, which are activated by pain signals and opioid withdrawal36,37. These noradrenergic neurons are well known to be anatomically connected with other brain areas involved in nociceptive and emotional pathways, which might explain the large array of effects observed. It has recently been shown that the lesion of the ascending A1 and A2 projections reduced the withdrawal syndrome, demonstrating the role of these projections in this process38. The precise Gpr10 expression sites in which the observed anti-opioid effects of PrRP take place remain to be determined. However, potential sites include the amygdala, hypothalamus, brainstem and dorsal root ganglia, all regions of high Gpr10 expression and key players in the processing of pain signals. Altogether, the present study identifies the PrRP-GPR10 pathway as a new negative regulator of the opiate system and suggests that GPR10 might constitute a new pharmacological target for the clinical management of pain, opioid side-effects and addictive disorders.

diameter, 800 s cut-off, average of three responses, n ¼ 10–12). For the writhing test, mice (n ¼ 6–7) received an i.p. injection of 0.6% acetic acid, and the number of writhes was counted in 10-min periods starting 5 min and 15 min after the injection. For the forced swimming test, the hot-plate test was applied after a swimming session of 5 min (n ¼ 7–11). For assessment of morphine tolerance, mice (n ¼ 10–14) received 10 mg kg1 of morphine i.p. twice daily. The withdrawal syndrome was promoted by two daily injections of morphine (20–100 mg kg1 i.p. over 5 d) followed by a naloxone injection (1 mg kg1, s.c.), and the vegetative and somatic signs were scored over 30 min (n ¼ 8–14). The rewarding effects of morphine were evaluated by using the conditioned place preference paradigm (n ¼ 8–14)39,40. Additional details are available in the Supplementary Methods. PrRP (5 nmol), NPFF (5 nmol) or nociceptin (25 nmol) was administered i.c.v. into the left lateral ventricle in a volume of 4 ml using a modified Hamilton syringe. In situ hybridization. Coronal cryostat brain sections (15 mm) were hybridized with 3¢-end-labeled [a35S]dATP oligonucleotide probes for proenkephalin and pronociceptin as described in the Supplementary Methods. Quantification was performed for areas showing moderate to high expression levels. For each area, statistical analysis was performed for five mice of each genotype (3–9 slices per area and animal). Radioligand binding assays and GTPcS binding assay. Saturation binding assays were performed on membrane preparations, using [3H]DAMGO or [3H]naltrindole as tracers and 10 mM naloxone for determination of nonspecific binding. Nonlinear regression analysis was performed by using the PRISM software (Biosoft) and a single-site model. Agonist-stimulated [35S]GTPgS binding assays were performed on whole brain membrane preparations or on brain slices, using SNC80 or DAMGO as agonists. Details are provided in the Supplementary Methods. Corticosterone assay. Blood samples were collected on ice within 2 min of the removal of mice from their cage. Corticosterone was assayed by using an RIA kit (Amersham Pharmacia). The stress procedures are described in the Supplementary Methods. For the corticosterone supplementation, corticosterone (12.5 mg ml1) was added to drinking water for 5 d. Blood samples were collected on day 3; the hot-plate test was performed on the afternoon of day 5. Statistical analysis. The data were analyzed with the Instat software (GraphPad). For comparison with a theoretical value, one-sample t-test was used. For single comparisons, the two-tailed Student’s t-test was used. For multiple comparisons, ANOVA was used, followed by two-tailed Dunnett t-test. Note: Supplementary information is available on the Nature Neuroscience website.

METHODS Generation of GPR10-knockout mice. The targeting vector consisted in a 9-kb cassette containing a promoter-less tau-lacZ fusion gene, a PGK-Neo gene and a HSV-TK gene, flanked by 5 kb of 5¢ and 3.5 kb of 3¢ Gpr10 gene fragments. The cassette replaced the first 194 codons of the Gpr10 gene, encoding transmembrane segments 1–4. Homologous recombination was carried out in the R1 ES cell line, and the recombinant clone was aggregated with CD1 eight cell–stage embryos (Supplementary Methods). Heterozygous mutants were bred for seven generations on a CD1 background before generating the wild-type and knockout mice used in this study. Male mice (2–4 months of age) were used in all experiments. RT-PCR analysis. Total RNA was purified from brain regions using the RNeasy Mini Kit (Qiagen). According to the specific experiment, the animals were killed 1 h after the injection of 1.5 IU kg1 of human insulin (Novo Nordisc), 3–4 h after morphine or 1 h after naloxone injection. RNA samples were treated with RNase-free DNase I and reverse transcribed using the Superscript II kit (Invitrogen). The primers and methodological details are described in the Supplementary Methods. Behavioral studies. The elevated plus maze test, rotarod test, light and dark box test, hot-plate test and tail-immersion test were performed as described39,40. For the tail-pressure test, increasing local pressure was applied until a withdrawal response was elicited (Basile analgesia meter, 1 mm tip

1740

ACKNOWLEDGMENTS We thank C. Contet, L. Cuvelier and D. Houtteman for technical assistance and F. Simonin for helpful discussion. M.P. and S.N.S. were supported by the Fondation Me´dicale Reine Elisabeth and the Fonds de la Recherche Scientifique Me´dicale. M.P. was also supported by the Belgian ‘Interuniversity Attraction Poles’ program, initiated by the Belgian State, Prime Minister’s Office, Science Policy Programming and the LifeSciHealth (grant LSHB-CT-2003-503337) programs of the European Community. O.V. and R.M. were supported by the Spanish Ministry of Science and Technology (SAF 2001-0745 to O.V. and GEN2003-20651-C06-04 to R.M). P.L. was supported by fellowships from the Fonds pour la Recherche dans l’Industrie et l’Agriculture and Te´le´vie. C.L and A. de K. are Chercheurs Qualifie´s of the Fonds National de la Recherche Scientifique. Scientific responsibility is assumed by the authors. COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/

1. Inturrisi, C.E. Clinical pharmacology of opioids for pain. Clin. J. Pain 18, S3–13 (2002). 2. Noble, F. & Roques, B.P. CCK-B receptor: chemistry, molecular biology, biochemistry and pharmacology. Prog. Neurobiol. 58, 349–379 (1999). 3. Panula, P. et al. Neuropeptide FF and modulation of pain. Brain Res. 848, 191–196 (1999).

VOLUME 8

[

NUMBER 12

[



ARTICLES 4. Calo, G., Guerrini, R., Rizzi, A., Salvadori, S. & Regoli, D. Pharmacology of nociceptin and its receptor: a novel therapeutic target. Br. J. Pharmacol. 129, 1261–1283 (2000). 5. Meunier, J.C. Utilizing functional genomics to identify new pain treatments: the example of nociceptin. Am. J. Pharmacogenomics 3, 117–130 (2003). 6. Mollereau, C., Roumy, M. & Zajac, J.M. Opioid-modulating peptides: mechanisms of action. Curr. Top. Med. Chem. 5, 341–355 (2005). 7. Allard, M., Labrouche, S., Nosjean, A. & Laguzzi, R. Mechanisms underlying the cardiovascular responses to peripheral administration of NPFF in the rat. J. Pharmacol. Exp. Ther. 274, 577–583 (1995). 8. Sundblom, D.M., Heikman, P., Naukkarinen, H. & Fyhrquist, F. Blood concentrations of vasopressin, neuropeptide FF and prolactin are increased by high-dose right unilateral ECT. Peptides 20, 319–326 (1999). 9. Roumy, M. & Zajac, J.M. Neuropeptide FF, pain and analgesia. Eur. J. Pharmacol. 345, 1–11 (1998). 10. Hinuma, S. et al. A prolactin-releasing peptide in the brain. Nature 393, 272–276 (1998). 11. Lawrence, C.B., Celsi, F., Brennand, J. & Luckman, S.M. Alternative role for prolactinreleasing peptide in the regulation of food intake. Nat. Neurosci. 3, 645–646 (2000). 12. Samson, W.K., Resch, Z.T. & Murphy, T.C. A novel action of the newly described prolactin-releasing peptides: cardiovascular regulation. Brain Res. 858, 19–25 (2000). 13. Maruyama, M. et al. Prolactin-releasing peptide as a novel stress mediator in the central nervous system. Endocrinology 142, 2032–2038 (2001). 14. Maruyama, M. et al. Central administration of prolactin-releasing peptide stimulates oxytocin release in rats. Neurosci. Lett. 276, 193–196 (1999). 15. Roland, B.L. et al. Anatomical distribution of prolactin-releasing peptide and its receptor suggests additional functions in the central nervous system and periphery. Endocrinology 140, 5736–5745 (1999). 16. Passik, S.D. & Weinreb, H.J. Managing chronic nonmalignant pain: overcoming obstacles to the use of opioids. Adv. Ther. 17, 70–83 (2000). 17. Lin, S.H., Leslie, F.M. & Civelli, O. Neurochemical properties of the prolactin releasing peptide (PrRP) receptor expressing neurons: evidence for a role of PrRP as a regulator of stress and nociception. Brain Res. 952, 15–30 (2002). 18. Kalliomaki, M.L. et al. Prolactin-releasing peptide affects pain, allodynia and autonomic reflexes through medullary mechanisms. Neuropharmacology 46, 412–424 (2004). 19. Mogil, J.S., Sternberg, W.F., Balian, H., Liebeskind, J.C. & Sadowski, B. Opioid and nonopioid swim stress-induced analgesia: a parametric analysis in mice. Physiol. Behav. 59, 123–132 (1996). 20. Semenova, S., Kuzmin, A. & Zvartau, E. Strain differences in the analgesic and reinforcing action of morphine in mice. Pharmacol. Biochem. Behav. 50, 17–21 (1995). 21. Lin, S.H., Leslie, F.M. & Civelli, O. Neurochemical properties of the prolactin releasing peptide (PrRP) receptor expressing neurons: evidence for a role of PrRP as a regulator of stress and nociception. Brain Res. 952, 15–30 (2002). 22. Pacak, K. et al. In vivo hypothalamic release and synthesis of catecholamines in spontaneously hypertensive rats. Hypertension 22, 467–478 (1993).


[

NUMBER 12

[

DECEMBER 2005

23. Lawrence, C.B., Ellacott, K.L. & Luckman, S.M. PRL-releasing peptide reduces food intake and may mediate satiety signaling. Endocrinology 143, 360–367 (2002). 24. Ueda, H., Inoue, M. & Mizuno, K. New approaches to study the development of morphine tolerance and dependence. Life Sci. 74, 313–320 (2003). 25. Solomon, R.L. & Corbit, J.D. An opponent-process theory of motivation. I. Temporal dynamics of affect. Psychol. Rev. 81, 119–145 (1974). 26. Malin, D.H. et al. FMRF-NH2-like mammalian peptide precipitates opiate-withdrawal syndrome in the rat. Peptides 11, 277–280 (1990). 27. Panula, P. et al. Neuropeptide FF and modulation of pain. Brain Res. 848, 191–196 (1999). 28. Murphy, N.P., Lee, Y. & Maidment, N.T. Orphanin FQ/nociceptin blocks acquisition of morphine place preference. Brain Res. 832, 168–170 (1999). 29. King, M.A., Rossi, G.C., Chang, A.H., Williams, L. & Pasternak, G.W. Spinal analgesic activity of orphanin FQ/nociceptin and its fragments. Neurosci. Lett. 223, 113–116 (1997). 30. Ueda, H., Inoue, M., Takeshima, H. & Iwasawa, Y. Enhanced spinal nociceptin receptor expression develops morphine tolerance and dependence. J. Neurosci. 20, 7640–7647 (2000). 31. Ueda, H. et al. Partial loss of tolerance liability to morphine analgesia in mice lacking the nociceptin receptor gene. Neurosci. Lett. 237, 136–138 (1997). 32. Noble, F., Derrien, M. & Roques, B.P. Modulation of opioid antinociception by CCK at the supraspinal level: evidence of regulatory mechanisms between CCK and enkephalin systems in the control of pain. Br. J. Pharmacol. 109, 1064–1070 (1993). 33. Noble, F. & Roques, B.P. The role of CCK2 receptors in the homeostasis of the opioid system. Drugs Today (Barc.) 39, 897–908 (2003). 34. Pommier, B. et al. Deletion of CCK2 receptor in mice results in an upregulation of the endogenous opioid system. J. Neurosci. 22, 2005–2011 (2002). 35. Langmead, C.J. et al. Characterization of the binding of [(125)I]-human prolactin releasing peptide (PrRP) to GPR10, a novel G protein coupled receptor. Br. J. Pharmacol. 131, 683–688 (2000). 36. Laorden, M.L., Castells, M.T. & Milanes, M.V. Effects of morphine and morphine withdrawal on brainstem neurons innervating hypothalamic nuclei that control the pituitary-adrenocortical axis in rats. Br. J. Pharmacol. 136, 67–75 (2002). 37. Dayas, C.V., Buller, K.M., Crane, J.W., Xu, Y. & Day, T.A. Stressor categorization: acute physical and psychological stressors elicit distinctive recruitment patterns in the amygdala and in medullary noradrenergic cell groups. Eur. J. Neurosci. 14, 1143– 1152 (2001). 38. Delfs, J.M., Zhu, Y., Druhan, J.P. & Aston-Jones, G. Noradrenaline in the ventral forebrain is critical for opiate withdrawal-induced aversion. Nature 403, 430–434 (2000). 39. Martin, M., Ledent, C., Parmentier, M., Maldonado, R. & Valverde, O. Involvement of CB1 cannabinoid receptors in emotional behaviour. Psychopharmacology (Berl.) 159, 379–387 (2002). 40. Simonin, F., Valverde, O., Smadja, C., Slowe, S., Kitchen, I., Dierich, A., Le Meur, M., Roques, B.P., Maldonado, R. & Kieffer, B.L. Disruption of the kappa-opioid receptor gene in mice enhances sensitivity to chemical visceral pain, impairs pharmacological actions of the selective kappa-agonist U-50,488H and attenuates morphine withdrawal. EMBO J. 17, 886–897 (1998).

1741


ARTICLES

K-ATP channels promote the differential degeneration of dopaminergic midbrain neurons Birgit Liss1–3, Olga Haeckel1, Johannes Wildmann4, Takashi Miki5, Susumu Seino5 & Jochen Roeper2,6 The selective degeneration of dopaminergic (DA) midbrain neurons in the substantia nigra (SN) is a hallmark of Parkinson disease. DA neurons in the neighboring ventral tegmental area (VTA) are significantly less affected. The mechanisms for this differential vulnerability of DA neurons are unknown. We identified selective activation of ATP-sensitive potassium (K-ATP) channels as a potential mechanism. We show that in response to parkinsonism-inducing toxins, electrophysiological activity of SN DA neurons, but not VTA DA neurons, is lost owing to activation of K-ATP channels. This selective K-ATP channel activation is controlled by differences in mitochondrial uncoupling between SN and VTA DA neurons. Genetic inactivation of the K-ATP channel pore-forming subunit Kir6.2 resulted in a selective rescue of SN but not VTA DA neurons in two mechanistically distinct mouse models of dopaminergic degeneration, the neurotoxicological 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine (MPTP) model and the mutant weaver mouse. Thus, K-ATP channel activation has an unexpected role in promoting death of DA neurons in chronic disease.

Differential vulnerability of neuronal populations to cell death is found in most neurodegenerative diseases1. In Parkinson disease, this differential vulnerability is particularly notable within the clinically most relevant population of dopaminergic (DA) midbrain neurons. Here, highly vulnerable DA subpopulations within the substantia nigra (SN) and their axonal projections to dorsal parts of the striatum are almost completely lost, whereas neighboring DA subpopulations within the ventral tegmental area (VTA) and their projections to ventral striatum survive to a large extent2. For rare forms of familial Parkinson disease, several causal gene mutations have been identified3,4, but as their respective gene products are ubiquitously expressed, they alone do not explain the selective pattern of DA cell loss. For the common sporadic form of Parkinson disease, diverse candidate mechanisms have been discussed, including mitochondrial dysfunction, oxidative stress and proteasomal impairment4–6. However, as the stereotypical pattern of differential cell loss of DA neurons is reproduced in mechanistically distinct Parkinson disease animals models2,5,7,8, different toxic and molecular pathways might converge on a common but still unknown cellular mechanism for differential neuronal vulnerability3,9. One well-described cellular stressor in Parkinson disease and its animal models is mitochondrial dysfunction; SN neurons from Parkinson disease patients show reduced activity of complex I of the mitochondrial respiratory chain6. Accordingly, the neurotoxic complex I inhibitors rotenone and MPP+ (1-methyl-4-phenylpyridinium, the active metabolite of MPTP)5 induce differential degeneration of DA neurons and parkinsonism in animals9–11 as well as in humans7,12. The

relative importance of the various downstream consequences of reduced mitochondrial complex I activity for neurodegeneration is not fully understood; complex I inhibition might reduce mitochondrial ATP production and in turn could compromise proteasomal activity13. Equally important, complex I inhibition could increase production of reactive oxygen species (ROS), which in turn elevate oxidative stress and the load of misfolded proteins14. How these immediate biochemical consequences of complex I inhibition are translated into cellspecific pathophysiology and eventually result in differential cell death of DA neurons is unknown. Here, we investigated the role of a potential downstream target of complex I inhibition, the K-ATP channel. These channels, composed of discrete pore-forming (Kir6.1/Kir6.2) and regulatory subunits (SUR1/ SUR2), are called ‘metabolic sensors’, as their open probability depends on the metabolic state of a cell15–17. K-ATP channels in dopaminergic neurons have been shown to open in response to partial complex I inhibition as well as in response to ATP depletion and increased oxidative stress (ROS)18–20. Thus, cell type–specific K-ATP channel activation would provide a convergent downstream target that could integrate energy depletion and oxidative stress, offering a candidate mechanism for the differential vulnerability of DA neurons in Parkinson disease. RESULTS Identification of mesostriatal and mesolimbic DA neurons To investigate differential vulnerability in the mature dopaminergic midbrain with single-cell resolution, we combined in vitro

1Molecular Neurobiology, Department of Physiology, Marburg University Deutschhausstrasse 2, 35037 Marburg, Germany. 2MRC Anatomical Neuropharmacology Unit, Oxford University, Mansfield Road, Oxford OX1 3TH, UK. 3University Laboratory of Physiology, Oxford University, Parks Road, Oxford OX1 3PT, UK. 4Immunology, Department of Physiology, Marburg University, Deutschhausstrasse 2, 35037 Marburg, Germany. 5Cellular and Molecular Medicine, University Graduate School of Medicine, 7-5-1 Kusunoki-cho, 650-0017 Kobe, Japan. 6Neurophysiology, Department of Physiology, Marburg University, Deutschhausstrasse 2, 35037 Marburg, Germany. Correspondence should be addressed to B.L. ([email protected]).


1742

VOLUME 8

[

NUMBER 12

[


ARTICLES


a

Mesostriatal SN DA

Mesolimbic VTA DA

b

Figure 1 Properties of adult mesostriatal and mesolimbic DA neurons. (a) Identification of mesostriatal and mesolimbic DA neurons. For both images of SN neurons (left) and images of VTA neurons (right), upper left image shows injection sites (striatum for SN and nucleus accumbens for VTA, scale bars, 400 mm), upper right image shows midbrain localization of TH-positive (green) and retrobead-labeled (red) neurons (scale bars, 150 mm), and the lower row shows identification of retrobead-labeled SN und VTA neurons in living brain slices via infrared videomicroscopy and epifluorescence (scale bars, 5 mm). (b) Electrophysiological properties of retrogradely labeled mesostriatal and mesolimbic DA neurons. Upper row: spontaneous activities and responses to 100 pA current injection (scale bars: 20 mV, 400 ms; dashed line, 0 mV). Middle row: activation of Ih currents by hyperpolarizing voltage steps (10-mV increments) from 40 mV (scale bars: 200 pA, 400 ms). Bottom row: responses to 50 pA, 250 pA and 500 pA current ramps from SN and VTA DA neurons (scale bars: 20 mV, 400 ms). (c) Upper images: confocal analysis of traced (red), recorded and neurobiotin-filled (green), and TH-positive (blue) neurons. Lower images: morphology of traced, neurobiotin-filled and DAB-processed TH-positive neurons (scale bars, 10 mm).

limbic, 12.0 ± 1.2 mV, n ¼ 10; P o 0.01; Ih amplitudes at –120 mV: mesostriatal, 255 ± 46 pA, n ¼ 15; mesolimbic, 60 ± 15 pA, n ¼ 10; P o 0.01; maximum firing rate: mesostriatal, 8.3 ± 2.0 Hz, n ¼ 15; mesolimbic, 13.8 ± 3.0 Hz, n ¼ 10; P o 0.01). Functionally characterized, labeled neurons were filled with neurobiotin during whole-cell recordings (or by breaking into whole-cell mode after perforated-patch recordings) for subsequent triple labeling immunohistochemistry in order to confirm their dopaminergic identity (as measured by tyrosine hydroxylase (TH) expression) and morphology (Fig. 1c).

c

electrophysiology, gene expression profiling and immunohistochemistry of individual retrogradely labeled mesostriatal SN and mesolimbic VTA DA neurons in adult mice21. Figure 1a shows injection sites of retrograde tracer substances in dorsal striatum and nucleus accumbens; distribution of labeled mesostriatal and mesolimbic TH-positive cell bodies within SN and VTA (upper panels); and single, fluorescently labeled mesostriatal and mesolimbic neurons in living brain-slices from 3-month-old mice (lower panels). Adult fluorescence-labeled DA neurons showed typical electrophysiological properties (such as spontaneous pacemaker activity, broad action potentials and presence of hyperpolarization-activated current (Ih)–mediated sag components (Fig. 1b)) similar to those described for 14-d-old postnatal unlabeled DA neurons22. Physiological differences between DA midbrain neurons, such as differences in sag and Ih current amplitudes and maximal firing rates before onset of depolarization block, were associated with distinct DA projections to striatal or accumbal target areas (Fig. 1b; sag amplitudes at –100 mV: mesostriatal, 16.5 ± 0.9 mV, n ¼ 15; meso-


[

NUMBER 12

[

DECEMBER 2005

All DA midbrain neurons express functional K-ATP channels The presence of K-ATP channels in postnatal DA neurons is well described18,20,23,24, but potential differences between adult mesolimbic VTA and mesostriatal SN DA neurons have not yet been addressed. To probe for the presence of functional K-ATP channels, we dialyzed DA neurons in brain slices with ATP-free pipette solution. Our previous data suggested that Kir6.2 (KCNJ11) is the pore-forming subunit of K-ATP channels in all DA midbrain neurons18,25. Therefore, we also analyzed DA neurons of a K-ATP channel knockout (Kir6.2/) mouse in which the Kir6.2 gene was disrupted15,26. K-ATP currents were activated in both mesostriatal SN and, to a smaller degree, mesolimbic VTA DA neurons from adult wild-type (Kir6.2+/+) mice (Fig. 2a). The washout-induced current at –50 mV was 112.4 ± 16.3 pA for mesostriatal DA neurons (n ¼ 10) and 56.4 ± 16.9 pA for mesolimbic DA neurons (n ¼ 12). The activation of K-ATP channels in DA neurons was accompanied by hyperpolarization of the membrane potential below threshold (for mesostriatal DA neurons, –35.6 ± 1.8 mV for the control condition (n ¼ 10) and –51.9 ± 2.6 mV for the ATP washout condition (n ¼ 10); and for mesolimbic DA neurons, –35.5 ± 1.1 mV for the control condition (n ¼ 12) and –49.1 ± 5.2 mV (n ¼ 12) for the ATP washout condition). In contrast to results from wild-type mice, no K-ATP currents were activated in response to zero ATP dialysis in Kir6.2/ mice, nor did we observe any hyperpolarization in mesolimbic or mesostriatal DA neurons in Kir6.2/ mice. The mean current difference after washout at –50 mV was –8.0 ± 14.3 pA for mesostriatal DA neurons (n ¼ 6) and 7.2 ± 3.4 pA for mesolimbic DA neurons (n ¼ 6). Basic electrophysiological properties of single action potentials and spontaneous discharge rates, recorded in the perforatedpatch configuration at physiological temperatures and glucose concentrations, did not differ significantly in DA neurons from adult Kir6.2/ mice and wild-type mice (Fig. 2b). The action potential width for Kir6.2+/+ mice was 3.1 ± 0.34 ms (n ¼ 12) and for Kir6.2/ mice, 2.9 ± 0.24 ms (n ¼ 10). The spontaneous activity for Kir6.2+/+

1743

ARTICLES

Kir6.2–/–


Kir6.2–/–

Kir6.2–/–

Figure 2 Kir6.2 forms the K-ATP channel pore in all DA neurons. (a) Current amplitudes during ATP washout (at 50 mV) in mesostriatal and mesolimbic DA neurons from Kir6.2+/+ and Kir6.2/ mice (scale bars: 50 pA, 100 ms, dashed line, 0 pA). (b) Single action potentials and spontaneous discharges (insert) of Kir6.2+/+ and Kir6.2/ SN DA neurons recorded in the perforatedpatch configuration.

mice was 1.75 ± 0.20 Hz (n ¼ 36) and 2.03 ± 0.20 Hz for Kir6.2/ mice (n ¼ 15). These data are in accordance with our previous study on young postnatal mice and demonstrate that Kir6.2 is the essential poreforming subunit for functional somatodendritic K-ATP channels in mesostriatal as well as mesolimbic DA neurons from adult mice. Distinct molecular compositions of K-ATP channels might confer variable metabolic sensitivities to native channels16. Thus, we analyzed K-ATP channel subunit expression of individual SN and VTA DA neurons using RT-PCR. We have previously shown that in young mice (postnatal day 14 (P14)), K-ATP channels in DA midbrain neurons coexpressed Kir6.2 and SUR1 or SUR2B mRNA18. As previously described for rotenone18, DA neurons expressing SUR1 mRNA, but not SUR2B mRNA, from P14 mice underwent K-ATP channel activation after acute application of 1-methyl-4-phenylpyridinium (MPP+). Hyperpolarization to –53.8 ± 2.4 mV (n ¼ 7) was induced; six out of these seven neurons expressed SUR1. This confirmed SUR1-containing K-ATP channels as metabolic sensors in DA neurons. To study K-ATP channel subunit expression in adult mice, DA neurons (identified by fluorescence labeling) were harvested by means of photoablation and laser microdissection (Fig. 3). Notably, in microdissected DA neurons from adult mice, we found no evidence for differential expression of SUR subunits. Using qualitative RT-PCR, we detected SUR1 mRNA with similar frequencies in single SN and VTA DA neurons but detected SUR2B mRNA in o5% of individual DA neurons (Fig. 3a,b): 36% of SN neurons (9/25), and 41.6% of VTA neurons (n ¼ 10/24), were SUR1- and TH-positive. In contrast to the SUR isoforms, and in accordance with previous studies, mRNA for calbindin, a marker for more resistant DA neurons, was coexpressed to a higher degree in VTA DA neurons (62.5%, or 15/24) than in SN DA neurons (4%, or 1/25, Fig. 3b).

Figure 3 K-ATP subunit expression of individual laser-microdissected DA neurons. (a) Qualitative RT-PCR of individual laser-microdissected SN and VTA DA neurons. Top: detection of a retrobead-labeled mesolimbic VTA neuron (scale bar, 5 mm). Blue arrowhead is an adjustment symbol of laser software. Bottom: mRNA expression profiles of a single mesostriatal SN and a single mesolimbic VTA neuron. (b) Calbindin is more frequently detected in single VTA DA compared to SN DA neurons. (c) Top: photomicrograph of a section after laser microdissection of selective VTA and SN cell pools for quantitative PCR (scale bar, 100 mm). Bottom: qualitative PCR with onetenth aliquot of cDNA reaction from a VTA DA cell pool illustrates absence of glial (GFAP) or GABAergic (GAD65/GAD67) contamination (see Methods). Blue arrowhead as in a. (d) Quantitative real-time PCR. There are significantly higher mRNA expression (normalized to mean SN expression) of SUR1 in cell pools of SN DA neurons than in VTA DA neurons.

1744

Complex I inhibition selectively activates K-ATP channels To avoid perturbation of cellular metabolism, we studied sensitivities of K-ATP channel activation in DA neurons in response to the complex I inhibitors and neurotoxins rotenone and MPP+ in the perforated-patch configuration at physiological temperatures and glucose concentrations. In mesostriatal SN DA neurons from adult Kir6.2+/+ mice, application of MPP+ (10 mM for 5 min) induced a slowing followed by a complete cessation of spontaneous activity (n ¼ 14/15) and a tonic hyperpolarization (Fig. 4a,b; control: 1.7 ± 0.2 Hz, 44.7 ± 1.2 mV, n ¼ 15; post-MPP+: 0.1 ± 0.1 Hz, 61.3 ± 1.5 mV, n ¼ 15; P ¼ 1.6 108, P ¼ 5.3 104). By contrast, spontaneous activities and membrane potentials of mesolimbic DA neurons were not significantly affected by 10 mM MPP+ (control: 2.4 ± 0.4 Hz, 42.7 ± 3.5 mV, n ¼ 6; post-MPP+: 2.3 ± 1.1 Hz, –40.7 ± 4.9 mV, n ¼ 6). In Kir6.2/ mice,

a

c

VTA

SN

1,000 500 200

SN

AD 65 Bd /GA 28 D6 7 SU k R SU 1 R 2 G A/ FA B P

Kir6.2

Quantitative real-time PCR of homogeneous microdissected SN DA and VTA DA cell pools (ten neurons) for all four K-ATP channel subunits eliminated detection threshold problems and confirmed that Kir6.2 and SUR1, not Kir6.1 or SUR2, are the relevant K-ATP channel subunits in SN DA neurons as well as in VTA DA neurons from adult mice. (Detection frequencies were 100% for SUR1 and Kir6.2, o10% for SUR2, and 0% for Kir6.1.) Relative levels of SUR1 mRNA expression were about twofold higher in SN DA neurons than in VTA DA neurons, consistent with the twofold-larger K-ATP washout currents of mesostriatal SN DA neurons (Fig. 3d). (SUR1 expression was 100 ± 16.9 in SN (n ¼ 11) and 45.1 ± 5.5 in VTA (n ¼ 8; P ¼ 0.009). Kir6.2 expression was 100 ± 21.5 in SN (n ¼ 10) and 51.8 ± 12.4 in VTA (n ¼ 8; P ¼ 0.07). Mean SN expression was normalized to 100.) Thus, K-ATP channels have a uniform molecular makeup but distinct mRNA expression levels across DA midbrain populations. However, K-ATP subunit composition or expression levels alone do not determine the metabolic sensitivity of K-ATP channels in intact neurons, as K-ATP channel gating is controlled by the intracellular ATP/ADP ratio and is regulated by a number of complex signaling factors, such as reactive oxygen species, phosphorylation, phosphoinositides and long-chain acyl-CoA esters27–30.

C

+/+

G

Mesostriatal SN DA

TH

Kir6.2+/+

b

VTA

VTA

b

SUR1

d

CB

60 40 20 0 SN VTA

VOLUME 8

[

SN VTA

NUMBER 12

[

Relative mRNA expression levels

Kir6.2+/+

Mesolimbic VTA DA

TH G AD 6 C 5/G Bd 28 AD 67 SU k R SU 1 R 2A /B

Mesostriatal SN DA

Percentage of TH+ cells

a

150

SUR1

Kir6.2

100

50

***

0 SN VTA

SN VTA


ARTICLES

a

–20

Mesostriatal SN DA Kir6.2+/+

mV

–80

2 min MPP+

2 min MPP+

5 min MPP+

5 min MPP+ Washout

1

–60

4

2

–80 500

1,000

1,500

6

4

–40

4

2

–60 1,000 s

1,500

600

200

–20

500

–80

2 min MPP+

mV

5 min MPP+

2 0

2,000

600 s

200

1 6

2 min MPP+ 5 min MPP+

Washout

3

1

5 6

10 µM MPP+

f

0 200

600

1,000

6

6 100 nM rotenone

200

600

1,000

100 nM rotenone

4

2 0

5

2

2

4

3

10 µM MPP+

4

4

0

1,000

Kir6.2–/–

Washout

e

1,000

100 nM rotenone

Kir6.2–/–

–20

5

10 µM MPP+

0

2,000

100 nM rotenone

3

2

mV

Hz

mV

4

6 Hz

1 6

0

d

5 –40

0

c

3 10 µM MPP+

6

Hz


Washout

b

Figure 4 K-ATP channel activation in response to complex I inhibition only in SN DA neurons. (a,d) Spontaneous activities and responses to 10 mM MPP+ in DA neurons from Kir6.2+/+ and Kir6.2/ mice (current-clamp, perforatedpatch recordings in adult brain slices at 36 1C). (b,e) Spontaneous firing frequencies (Hz/10 s bins) and mean membrane potential (mV) (horizontal bars indicate MPP+ application). (c,f) Corresponding experiments with 100 nM rotenone (horizontal bars indicate rotenone application).

Mesolimbic VTA DA Kir6.2+/+

2 200

600 s

0

1,000

200

600 s

1,000

neither mesostriatal SN DA nor mesolimbic VTA DA neurons showed MPP+-induced silencing of spontaneous activity (Fig. 4d,e; mesostriatal: control, 2.0 ± 0.2 Hz, 44.3 ± 0.6 mV, n ¼ 6; post-MPP+, 1.8 ± 0.2 Hz, 43.9 ± 1.0 mV, n ¼ 6; mesolimbic: control, 2.6 ± 1.0 Hz, 42.5 ± 1.1 mV, n ¼ 6; post-MPP+, 2.7 ± 1.0 Hz, 41.7 ± 1.3 mV, n ¼ 6). We also observed selective silencing of mesostriatal but not mesolimbic Kir6.2+/+ DA neurons in response to 100 nM rotenone (Fig. 4c; mesostriatal: control, 1.8 ± 0.2 Hz, 44.1 ± 2.4 mV; postrotenone, 0.0 ± 0.1 Hz, 61.3 ± 1.5 mV, n ¼ 12; P ¼ 3.3 105, P ¼ 4.5 105; mesolimbic: control, 3.8 ± 1.1 Hz, 42.1 ± 3.4 mV; post-rotenone, 4.3 ± 1.1 Hz, –39.8 ± 5.1 mV, n ¼ 6). Again, activity of Kir6.2/ DA neurons was not significantly altered by rotenone (Fig. 4f; mesostriatal: control, 1.9 ± 0.3 Hz, 44.1 ± 1.5 mV; post-rotenone, 1.9 ± 0.4 Hz, 44.2 ± 1.7 mV, n ¼ 9; mesolimbic: control, 3.9 ± 1.0 Hz, –42.8 ± 2.1 mV; post-rotenone, 3.8 ± 1.2 Hz, 43.4 ± 2.3 mV, n ¼ 6). These data demonstrate that complex I inhibition induces selective hyperpolarization and electrical silencing of highly vulnerable mesostriatal SN DA neurons, but not VTA DA neurons, owing to activation of K-ATP channels made up of Kir6.2 and SUR1 subunits. As rotenone uptake (in contrast to MPP+ uptake) is passive, the differential responses of mesostriatal and mesolimbic DA neurons are independent of toxin loading effects. Mitochondrial uncoupling controls K-ATP channel activation As K-ATP channels in all DA neurons are composed of Kir6.2 and SUR1 subunits, other factors (including differences in ATP


[

NUMBER 12

[

DECEMBER 2005

consumption or ATP generation) must cause the observed differential channel activation in response to complex I inhibition. Extensive uncoupling of the mitochondrial respiratory chain by 2 mM carbonylcyanide-4-trifluoromethoxyphenylhydrazone (FCCP) activated K-ATP channels in both SN DA and VTA DA neurons (perforated-patch recording in 2 mM FCCP; SN: 0 ± 0 Hz, 75.8 ± 1.1 mV; VTA: 0 ± 0 Hz, 64.8 ± 9.0 mV, n ¼ 3; P ¼ 0.35). This indicates that K-ATP channel gating is primarily controlled by mitochondrial metabolism in all DA midbrain neurons. Consequently, quantitative differences in mitochondrial uncoupling might cause the differential sensitivity of K-ATP channels to complex I inhibition. Thus, we studied the effects of mild uncoupling of the mitochondrial respiratory chain induced by 50 nM FCCP. Mild neuronal uncoupling reduces the generation of reactive oxygen species (ROS) rather than impairing ATP production31,32. We found that 50 nM FCCP (20-min preincubation) alone did not activate K-ATP channels or alter firing rates of SN or VTA DA neurons (Fig. 5a; perforated-patch recording in 50 nM FCCP; SN: 1.9 ± 0.2 Hz, n ¼ 10; VTA: 2.4 ± 0.5 Hz, n ¼ 6). Notably, however, mild uncoupling inverted the response of K-ATP channels to complex I inhibition: in this case, VTA DA neurons, but not SN DA neurons, were hyperpolarized and functionally silenced due to K-ATP channel activation. In the presence of 50 nM FCCP, none of the SN DA neurons was significantly affected by 100 nM rotenone (Fig. 5a,b, left; perforatedpatch recording in 50 nM FCCP: 2.33 ± 0.29 Hz; FCCP + rotenone: 1.92 ± 0.36, n ¼ 6; P ¼ 0.40) or 10 mM MPP+ (data not shown). In contrast, the presence of 50 nM FCCP sensitized K-ATP channels of VTA DA neurons to complex I inhibition (Fig. 5a,b, right; 50 nM FCCP: 2.4 ± 0.55 Hz; FCCP + rotenone: 0 ± 0 Hz, n ¼ 6; P ¼ 0.0075). We found that 50 nM FCCP had no effects on membrane conductance of DA neurons from Kir6.2/ mice, in the absence or presence of rotenone (SN DA: control, 1.88 ± 0.36 nS; 50 nM FCCP, 1.95 ± 0.23 nS, n ¼ 6, P ¼ 0.88; VTA DA: control, 2.23 ± 0.27 nS; 50 nM FCCP, 2.07 ± 0.62 nS, n ¼ 3, P ¼ 0.83). To quantify the reciprocal effects of mild uncoupling on K-ATP channel activation in response to complex I inhibition in SN and VTA DA neurons, we generated dose-response curves for rotenone-induced K-ATP channel activation in the presence and absence of 50 nM FCCP. In response to a 5-min application of rotenone (concentrations from 1–1,000 nM), the subthreshold slope conductances (Gmin) of SN DA neurons increased fourfold to a maximum of about 7 nS (Fig. 5c, left; Gbase: 1.75 ± 0.1 nS; Gmax: 6.99 ± 0.1 nS; n ¼ 3–6), but increased only about 1.8-fold to a maximum of 3.3 nS in VTA DA neurons (Fig. 5c, right; Gbase: 1.78 ± 0.03 nS, Gmax: 3.28 ± 0.03 nS; n ¼ 3–6). Preincubation of SN DA neurons with 50 nM FCCP substantially reduced the sensitivity and extent of rotenone-induced K-ATP channel activation to a degree similar to that of VTA DA neurons under control conditions (Fig. 5c, left; control: rotenone concentration for halfmaximum response (EC50): 77.6 ± 3.1 nM, GK-ATP: 5.42 nS; 50 nM FCCP: EC50: 260 ± 17 nM, GK-ATP: 1.89 nS). In contrast, in VTA DA neurons, mild uncoupling substantially potentiated rotenone-induced K-ATP channel activation (Fig. 5c, right; control: EC50: 82.2 ± 4.3 nM, GK-ATP: 1.50 nS; 50 nM FCCP: EC50: 46.0 ± 1.8 nM, GK-ATP: 2.62 nS). In SN DA neurons, mild uncoupling not only prevented the loss of

1745

ARTICLES

a

Mesostriatal SN DA

Mesolimbic VTA DA

–20

50 nM FCCP

50 nM FCCP

–80

100 nM rotenone (2 min) + FCCP


Mesostriatal SN DA

b 6 4

Mesolimbic VTA DA 6

100 nM rotenone

100 nM rotenone

50 nM FCCP

4

mV

Hz

50 nM FCCP


2


0

Washout rotenone to 50 nM FCCP

Washout rotenone to 50 nM FCCP

2

200

600

1,000

0

1

c

8

3 Time (s) Control

5

1

4

3 Time (s)

600

1,000

Time (s)

5 In 100 nM rotenone

d

Control + 50 nM FCCP

4

0

200

8

+ 50 nM FCCP

gmin[nS]


Time (s)

In 100 nM rotenone + 50 nM FCCP

0 0.1

1

10 100 1,000 [Rotenone] (nM)

0.1

1

10 100 1,000 [Rotenone] (nM)

Figure 5 Mitochondrial uncoupling controls differential K-ATP channel activation. (a) Spontaneous activities and responses to 100 nM rotenone of SN DA and VTA DA neurons from Kir6.2+/+ mice in the presence of 50 nM of the mitochondrial uncoupler FCCP (current-clamp, perforated-patch recordings in adult brain slices at 36 1C). (b) Spontaneous firing frequencies (Hz/10 s bins) of SN and VTA DA neurons in 50 nM FCCP (horizontal bars indicate rotenone application). (c) Rotenone concentration-dependence of minimal slope conductances (Gmin) in SN and VTA DA neurons in the presence and absence of 50 nM FCCP. Points are mean ± s.e.m. of 3–6 experiments. Lines represent fits of the mean data by a Hill equation. (d) Electrophysiological properties of SN DA neurons after 5 min incubation with 100 nM rotenone in the absence (top) and presence (bottom) of 50 nM FCCP. Responses to depolarizing current ramps (50-pA steps; scale bars: 20 mV, 1 s; compare with Fig. 1b, lower panels).

spontaneous activity in the presence of complex I inhibition but also restored their excitability in response to current ramps (Fig. 5d; compare Fig. 1b, right panels). In accordance with a proposed reduction of ROS rather than reduced ATP levels as the main consequence of mild uncoupling31, preincubation of SN DA neurons with a superoxide dismutase mimetic (30 mM Mn(III)tetrakis(4-benzoic acid)porphrin chloride (MnTBAP)) also prevented rotenone-induced K-ATP channel activation and membrane hyperpolarization (100 nM rotenone + 30 mM MnTBAP: 33.68 ± 4.16 mV, 2.45 ± 0.42 nS, n ¼ 6). These data could indicate that VTA DA neurons, in contrast to SN DA neurons, already show a mild degree of constitutive uncoupling of the mitochondrial respiratory chain that reduces ROS and in turn prevents K-ATP channel activation20 in response to complex I inhibition. For further experimental support, using real-time quantitative RT-PCR, we quantified the mRNA expression of the two most relevant neuronal uncoupling proteins, UCP-2 and UCP-3 (ref. 31), both of which we found to be expressed in mouse brain. Consistently, we detected an about threefold-higher level of UCP-2 mRNA in laser-microdissected VTA DA neurons than in SN DA neurons. UCP-3 mRNA was not detected in DA neurons (UCP-2: SN, 100.0 ± 16.9, n ¼ 11; VTA, 316.4 ± 49.8, n ¼ 11; P ¼ 0.0014. Mean SN expression was normalized to 100). The EC50 values for K-ATP channel activation in response to chronic in vitro complex I inhibition (430 min preincubation) by MPP+ (EC50: 2.2 mM) and rotenone (EC50: 16.2 nM; ref 18) in SN DA neurons are within the range of the effective in vivo toxin concentrations in their respective rodent Parkinson disease models10,11. Thus, our data suggest the possibility that selective K-ATP channel activation in mesostriatal SN DA is also operative in vivo. Notably, in this context, mild mitochondrial uncoupling or increased expression of UCP-2 in vivo has been shown to be neuroprotective in animal models of Parkinson disease33,34. Selective SN DA rescue in MPTP-treated Kir6.2/ mice To investigate whether K-ATP channels in vivo are responsible for the preferential degeneration of mesostriatal SN DA neurons, we compared

1746

the extent and distribution of striatal DA fiber loss and dopaminergic cell death for Kir6.2+/+ and Kir6.2/ mice in two mechanistically independent mouse models of selective dopaminergic degeneration: a chronic low-dose MPTP model of Parkinson disease8 as well as the mutant weaver mouse35. Chronic low-dose MPTP administration replicated the differential loss of DA fibers and neurons typical of Parkinson disease (Fig. 6). We observed a reduction of about 70% in the density of TH-immunopositive fibers in the dorsal striatum of Kir6.2+/+ mice, whereas projections to the ventral striatum were not significantly reduced (Fig. 6a,b; optical densities: dorsal striatum: control, 33.4 ± 2.6, n ¼ 10; post-MPTP, 11.3 ± 1.3, n ¼ 6, P ¼ 4.2 106; ventral striatum: control, 17.4 ± 3.8, n ¼ 10; post-MPTP, 18.0 ± 3.4, n ¼ 6). In Kir6.2/ controls, TH fiber density in dorsal and ventral striatum was not significantly different from that in the wild-type condition. In contrast, after chronic MPTP, TH fiber density was selectively increased by about 100% in Kir6.2/ mice compared with wild-type mice, in the dorsal (P ¼ 0.004) but not in the ventral striatum (P ¼ 0.24; Fig. 6a,b; dorsal striatum: control, 33.3 ± 2.8, n ¼ 12; post-MPTP, 21.6 ± 1.9, n ¼ 6; P ¼ 0.005; ventral striatum: control, 21.2 ± 4.5, n ¼ 10; post-MPTP, 13.9 ± 2.2, n ¼ 6). The differential rescue of the mesostriatal DA system of Kir6.2/ mice was even more evident in the midbrain. Based on nonbiased stereology, we quantified the number of TH-positive neurons in SN and VTA in controls and after chronic MPTP across the entire caudo-rostral extent of the midbrain (Fig. 6c–e). In Kir6.2+/+ mice, MPTP-induced loss of TH-positive neurons within the SN was about 50% (control, 5,177 ± 212, n ¼ 5; post-MPTP, 2,740 ± 199, n ¼ 9; P ¼ 4.9 107). Cell loss was more extensive in caudal sections, consistent with the caudo-rostral pattern found in Parkinson disease2 (Fig. 6e, top). In the VTA, only about 25% of TH-positive cells were lost after MPTP treatment (control, 3,578 ± 175, n ¼ 5; post-MPTP, 2,602 ± 70, n ¼ 4; P ¼ 0.003). In Kir6.2/ mice, the number of TH-positive neurons in SN or VTA under control conditions was not significantly different from that in Kir6.2+/+ mice. Also, the moderate

VOLUME 8

[

NUMBER 12

[


ARTICLES

b

d

+/+

–/–

MPTP

4,000

*** 2,000 0

Control

Post-MPTP VTA

VTA

SN

HE + SN cells

+/+ –/–

c

8,000

0

TH + VTA cells

SN

–/–

30 20 10 0 +/+

e

+/+

–/–

Kir6.2+/+

Control

50

MPTP

0 –3.8 50 0 –50 –100

100

–3.8

–3.6

–3.4

–3.2

–3.0

–2.8

–3.6 –3.4 –3.2 –3.0 Caudorostral axis (bregma)

–2.8

Kir6.2–/–

+/+ –/– +/+ –/–

4,000 2,000

–/–

–/–

100

***

6,000 VTA

+/+

MPTP

40

4,000

SN

SN

+/+ –/–

12,000

+/+

VTA

–/–

Control

–/–

+/+

0

VS

VS –/–

10

6,000 TH+ SN cells

DS

DS

***

20

TH density, ventral striatum

+/+

30

Cell loss (%)

+/+

Control

40

Average SN (n)

VS

VS

VS

MPTP

Cell loss (%)

DS

DS


DS Control

Post-MPTP TH density, dorsal striatum

Control

Average SN (n)

a

0 +/+ –/– +/+ –/–

Control

50

MPTP

0 –3.8

–3.6

–3.4

–3.2

–3.0

–2.8

–3.8

–3.6 –3.4 –3.2 –3.0 Caudorostral axis (bregma)

–2.8

50 0 –50 –100

Figure 6 Selective rescue of SN DA neurons in Kir6.2/ in a MPTP model of Parkinson disease. (a) TH immunostaining of striatal sections. Preferential postMPTP loss of TH immunostaining in dorsal striatum (DS) but not ventral striatum (VS) of Kir6.2+/+ mice but not Kir6.2/ mice (scale bar, 1 mm). (b) Optical density quantification of TH immunostaining intensity of dorsal and ventral striatum of sham-injected Kir6.2+/+ and Kir6.2/ controls and post-MPTP Kir6.2+/+ and Kir6.2/ mice. (c) TH immunostaining of midbrain section from control and MPTP-treated Kir6.2+/+ and Kir6.2/ mice. Selective loss of THimmunostained cells in SN of Kir6.2+/+ mice (scale bar, 200 mm). (d) Nonbiased stereological analysis of TH-immunopositive and hematoxylin-eosin–positive cells in SN (top and center) and TH-immunopositive cells in VTA (bottom) from control and MPTP-treated Kir6.2+/+ and Kir6.2/ mice. Note complete rescue of DA SN cells in Kir6.2/ mice. (e) Upper panels: average number of unbiased sampled TH-positive SN neurons in 30 serial midbrain sections, covering the caudo-rostral axis (bregma –3.8 to –3.2) in control and MPTP-treated Kir6.2+/+ and Kir6.2/ mice. Lower panels: average loss (%) of TH-positive SN neurons along the caudo-rostral axis. Note absence of cell loss throughout the entire caudorostral extent of the SN in Kir6.2/ mice.

loss of TH-positive neurons in the VTA after chronic MPTP treatment (Kir6.2/ VTA: control, 3,438 ± 90, n ¼ 5; post-MPTP, 2,755 ± 89, n ¼ 4; P ¼ 0.001) was not significantly different between Kir6.2/ and wild-type mice. Notably, however, we found no evidence for cell loss of TH-positive SN neurons in Kir6.2/ after MPTP treatment (SN: control, 5,408 ± 172, n ¼ 6; post-MPTP, 5,179 ± 285, n ¼ 5; P ¼ 0.5). Compared with Kir6.2+/+ mice (P ¼ 1.1 107), we detected a complete and selective rescue of the highly vulnerable population of DA neurons in the SN across the entire caudo-rostral extent of the midbrain in Kir6.2/ mice (Fig. 6d,e, bottom). Stereological analysis of all SN pars compacta neurons in hematoxylin-eosin counterstained sections demonstrated genuine MPTP-induced neuronal death in wildtype mice and confirmed the complete rescue of SN neurons in the Kir6.2/ mice (Fig. 6d, middle panel; Kir6.2+/+ SN: control, 11,882 ± 222; post-MPTP, 8,061 ± 632, P ¼ 0.029; Kir6.2/ SN: control, 12,288 ± 231; post-MPTP, 12,619 ± 223; P ¼ 0.36; n ¼ 3 each). Striatal MPP+ levels 90 min after MPTP injection were not significantly different between wild-type and knockout mice (Kir6.2+/+: 16.0 ± 0.96 mM, n ¼ 10; Kir6.2/: 21.3 ± 2.4 mM, n ¼ 11; P ¼ 0.06).


[

NUMBER 12

[

DECEMBER 2005

We also measured striatal TH fiber density (Fig. 7b), and striatal dopamine and 3,4-dihydroxyphenylacetic acid (DOPAC) levels (Fig. 7c) before and 3.5 d after MPTP injections. These experiments excluded potentially confounding differences between Kir6.2/ and Kir6.2+/+ mice in MPTP metabolism, dopamine metabolism, or acute dopamine depletion after MPTP injection. In contrast, the mesostriatal system of Kir6.2/ mice seemed more vulnerable after only a single MPTP injection. TH fiber density was significantly reduced (P ¼ 0.004) in the dorsal striatum of Kir6.2/ mice (Fig. 7b; dorsal striatum control: Kir6.2+/+, 33.0 ± 3.0; Kir6.2/, 32.0 ± 2.8; dorsal striatum post-MPTP: Kir6.2+/+, 28.3 ± 3.4, Kir6.2/, 20.8 ± 1.2; ventral striatum control: Kir6.2+/+, 26.4 ± 4.0; Kir6.2/, 22.5 ± 4.0; ventral striatum post-MPTP: Kir6.2+/+, 27.0 ± 5.5; Kir6.2/, 27.6 ± 3.7; n ¼ 6 each). In addition, striatal levels of dopamine and dopamine/DOPAC ratios were slightly lower in Kir6.2/ mice than in controls after one dose of MPTP (Fig. 7c; dopamine controls (ng g–1): Kir6.2+/+, 6,869 ± 714; Kir6.2/, 6,292 ± 432, n ¼ 10/10; post-MPTP: Kir6.2+/+, 4,120 ± 565; Kir6.2/, 2,166 ± 217, n ¼ 10 (Kir6.2+/+) and n ¼ 6 (Kir6.2/); P ¼ 0.02; control dopamine/DOPAC ratio: Kir6.2+/+, 19.1 ± 0.7;

1747

ARTICLES 3.5 d after one dose of MPTP

b

Control

MPTP

DS

1.5

+/+

VS

TH density, VS

0

4,000

*

2,000 0

400

Control MPTP

300 200 100

25

***

20 10 0

+/+ –/–

+/+ –/–

Control

MPTP

+/+ –/–

+/+ –/–

30 20 10

Control

MPTP

*

15 10 5 0

+/+ –/– +/+ –/–

+/+ –/–

+/+ –/–

Kir6.2/, 17.3 ± 1.2; post-MPTP: Kir6.2+/+, 19.4 ± 1.7; Kir6.2/, 12.8 ± 0.9; P ¼ 0.01). In both Kir6.2+/+ and Kir6.2/ mice, DA depletion and TH fiber loss were reversible 1 week after injection (data not shown). Higher TH fiber loss and lower striatal dopamine level after one MPTP injection is in accordance with the well-described acute neuroprotective effect of K-ATP channel activation in response to metabolic stress28,36–39. Despite the loss of the acute neuroprotective effect in the Kir6.2/ mice, genetic inactivation of K-ATP channels leads to selective and complete rescue of vulnerable SN DA neurons in the MPTP mouse model of Parkinson’s disease. Selective SN DA rescue in Kir6.2/ weaver double mutants To investigate whether K-ATP channel activation is a general mechanism beyond the constraints of MPTP models, we analyzed an additional genetic model of selective dopaminergic degeneration, the weaver (wv/wv) mouse35. Owing to a point mutation in the Girk2 gene (Kir3.2, KCNJ6), homomeric Girk2 channels lose their potassium selectivity and G protein control and become constitutively active channels. Consequently, in homozygous weaver mice, dopaminergic midbrain neurons are affected by selective degeneration25,35. In addition, the cerebellum of homozygous weaver mice is markedly reduced

Control (wv+/+)

wv/wv

DS VS Kir6.2+/+

b

DS TH density, dorsal striatum

a

Figure 7 In vitro responses to MPP+ and short-term in vivo responses to MPTP in Kir6.2+/+ and Kir6.2/ mice. (a) MPP+ concentration–dependence of K-ATP channel conductance (GK-ATP) in DA SN neurons. Points are mean ± s.e.m. of four to eight experiments. Line represents a fit of the mean data by a Hill equation with a half-maximal effective concentration (EC50) of 2.2 mM and a Hill coefficient of 1. (b) Left: TH immunostaining of striatal sections from Kir6.2+/+ and Kir6.2/ controls and Kir6.2+/+ and Kir6.2/ mice 3.5 d after a single injection of MPTP (scale bar, 1 mm). Right: optical density quantification of TH immunostaining intensity of dorsal striatum and ventral striatum (compare with Fig. 4a). (c) Mean concentrations of dorsal striatal dopamine, DOPAC and DA/DOPAC ratios in Kir6.2+/+ and Kir6.2/ controls and 3.5 d after a single MPTP injection. There was significant DA depletion after a single MPTP injection in both Kir6.2+/+ and Kir6.2/ mice.

20

0

+/+ –/– +/+ –/–

MPTP

30

0

–/–

DA / DOPAC

6,000

[DOPAC] (ng/g fw)

c

8,000 Control MPTP

–/–

Control

40

0.5 0.1 10 1,000 [MPP+] (µM)

40

+/+

EC50 = 2.2 µM

[Dopamine] (ng/g fw)

Control

wv/wv

in size, owing to a loss of cerebellar granule cells40. We have previously shown that K-ATP channels containing SUR1 and Kir6.2 subunits are activated in SN DA midbrain neurons of homozygous weaver mice in response to active weaver mutant channels during the most active phase of postnatal DA degeneration25. Here, to study the effect of loss of K-ATP channels on dopaminergic degeneration in weaver mice in vivo, we crossed heterozygote weaver mice (wv/+) with Kir6.2/ mice to generate weaver Kir6.2+/+ (wv/wv Kir6.2+/+) and weaver Kir6.2/ mice (wv/wv Kir6.2/) and control mice of similar genetic background (wv+/+ Kir6.2+/+ and wv+/+ Kir6.2/). The cerebella of adult 3-month-old homozygous weaver mice were both significantly reduced in size, regardless of the presence (wv/wv Kir6.2+/+) or absence (wv/wv Kir6.2/) of functional K-ATP channels (data not shown), indicating that Kir6.2 K-ATP channels did not affect cell death of cerebellar granule cells. In contrast, lack of K-ATP channels in adult weaver mice (wv/wv Kir6.2/) led to partial but selective rescue of mesostriatal DA neurons. Comparison of striatal TH fiber density in wv/wv Kir6.2+/+ mice with that in wv+/+ Kir6.2+/+ mice confirmed the previously described pattern of differential vulnerability (dorsal striatum versus ventral striatum, Fig. 8a,b)35. This pattern was also similar to that of the chronic MPTP mouse model (compare Fig. 6a,b). Optical density of TH-positive fibers of homozygous double mutant mice (wv/wv Kir6.2/) was about 100% higher (P ¼ 0.005) than that of weaver mice with functional K-ATP channels (wv/wv Kir6.2+/+) in the dorsal striatum, but not in the ventral striatum, similar to our MPTP model mice (Fig. 8a,b; dorsal striatum: wv+/+

c

d

40

Control (wv+/+)

wv/wv

30

***

20

VTA

10 0

SN +/+ –/– +/+ –/–

SN

Kir6.2+/+

+/+

Control wv/wv 6,000

4,000

*** 2,000

+/+

Kir6.2

Kir6.2

+/+ –/– +/+ –/–

VS

Kir6.2 –/–

Kir6.2 –/–

40

Control

Control wv/wv 6,000

wv/wv

TH + VTA cells

TH density, ventral striatum


GK-ATP (nS)

2.5

TH density, DS

Chronic (>30 min) + in vitro MPP

TH + SN cells

a

30

Kir6.2 –/–

20

Kir6.2 –/–

10 0

+/+ –/–

+/+ –/–

4,000 2,000 0

VTA

+/+ –/– +/+ –/–

Figure 8 Selective rescue of SN DA neurons in Kir6.2/ weaver double mutant mice. (a) TH immunostaining of striatal sections. There is preferential loss of TH immunostaining in dorsal striatum (DS) but not ventral striatum (VS) of wv/wv Kir6.2+/+ mice but not wv/wv Kir6.2/ mice (scale bar, 1 mm). (b) Optical density quantification of TH immunostaining intensity of dorsal striatum and ventral striatum of control (wv+/+) and homozygous weaver (wv/wv) mice. (c) TH immunostaining of midbrain section from control (wv+/+) and homozygous weaver mice either in Kir6.2+/+ or Kir6.2/ mice. Note selective loss of THimmunostained cells in SN of wv/wv Kir6.2+/+ mice (scale bar, 200 mm). (d) Nonbiased stereological analysis of TH-immunopositive cells in SN and VTA from control and weaver Kir6.2+/+ and Kir6.2/ mice. Note partial rescue of DA SN cells in wv/wv Kir6.2/ mice.

1748

VOLUME 8

[

NUMBER 12

[



ARTICLES Kir6.2+/+, 38.0 ± 2.4; wv/wv Kir6.2+/+, 9.4 ± 0.6; wv+/+ Kir6.2/, 36.7 ± 2.5; wv/wv Kir6.2/, 18.2 ± 2.8; n ¼ 5 each; ventral striatum: wv+/+ Kir6.2+/+, 36.1 ± 1.9; wv/wv Kir6.2+/+, 37.3 ± 2.1; wv+/+ Kir6.2/, 33.6 ± 2.1; wv/wv Kir6.2/, 37.5 ± 2.2; n ¼ 5 each). Stereological analysis of TH-positive midbrain neurons and of respective hematoxylin-eosin–counterstained midbrain sections confirmed the differential loss of SN DA midbrain neurons and their selective, but only partial, rescue by genetic inactivation of K-ATP channels (Fig. 8c,d), in contrast to the MPTP model. The number of TH-positive neurons in SN or VTA from adult 3-month-old mice did not differ in wild-type control mice (wv+/+ Kir6.2+/+) and Kir6.2/ control mice (wv+/+ Kir6.2/). Loss of SN DA neurons in homozygous weaver mice was about 70% (wv+/+ Kir6.2+/+: 5,415 ± 131, n ¼ 3; wv/wv Kir6.2+/+: 1,687 ± 83, n ¼ 5; P ¼ 2.4 107). VTA DA neurons were significantly less affected by the weaver mutation, similar to the MPTP model; only B25% of VTA DA neurons were lost (wv+/+ Kir6.2+/+: 4,696 ± 152, n ¼ 3; wv/wv Kir6.2+/+: 3,472 ± 185, n ¼ 5; P ¼ 0.004, wv+/+ Kir6.2/: 4,157 ± 155, n ¼ 4; wv/wv Kir6.2/: 2,997 ± 185, n ¼ 4; P ¼ 0.003). Again, moderate cell death of VTA DA neurons in mutant weaver mice was not affected by loss of K-ATP channels: the number of surviving VTA DA neurons did not significantly differ between Kir6.2/ and Kir6.2+/+ weaver mice (P ¼ 0.12). In contrast, and in accordance with the MPTP model, the number of surviving SN DA neurons in double-mutant mice (wv+/+ Kir6.2/: 5,558 ± 203, n ¼ 5; wv/wv Kir6.2/: 2,126 ± 90, n ¼ 5; P ¼ 3.1 107) was significantly higher than in wv/wv Kir6.2+/+ mice (P ¼ 0.007), indicating that in the weaver mutant mice as well, genetic inactivation of the Kir6.2 gene resulted in a selective rescue of B25% of highly vulnerable SN DA neurons. Again, analysis of hematoxylineosin–counterstained sections confirmed genuine cell death and rescue (SN: wv+/+ Kir6.2+/+, 11,753 ± 82; wv+/+ Kir6.2/, 12,328 ± 295; wv/wv Kir6.2+/+, 5,387 ± 174; wv/wv Kir6.2/: 6,568 ± 66; n ¼ 3 each). DISCUSSION We demonstrate here that Kir6.2-containing K-ATP channels are causally linked to the differential degeneration of dopaminergic midbrain neurons in vivo in response to complex I inhibition as well as in response to mutant weaver Girk2 channels. In these two mechanistically unrelated, chronic mouse models of dopaminergic degeneration, the presence of functional K-ATP channels promotes cell death of SN DA but not VTA DA neurons. In both models, because the moderate loss of VTA DA neurons is independent of the presence of K-ATP channels, cell death of SN and VTA DA neurons is executed via divergent pathways. We have demonstrated in in vitro brain slices that selective K-ATP channels are selectively activated and electrical activity is lost in response to complex I inhibition in highly vulnerable SN DA neurons but that they are not activated in VTA DA neurons. Our mouse model data strongly suggest that distinct pathophysiological trigger factors (such as complex I inhibition in the MPTP model or constitutive activity of sodium-permeable mutant Girk2 channels in weaver mice25) might also converge in vivo on the selective activation of K-ATP channels in SN DA neurons, resulting in functional silencing. As about 70% of SN DA neurons are lost in adult weaver mice, it is likely that weaver SN DA neurons that we had previously studied in young (P14) mice25, might not have been long-term survivors, but rather neurons in the process of degeneration. Our in vivo animal data indicate that functional K-ATP channels are indeed necessary for cell death of SN DA neurons in the chronic MPTP model and promote neuronal death in the mutant weaver mouse. As we studied a general Kir6.2 knockout mouse, we cannot rule out a contribution from Kir6.2-expressing cell types other than SN DA neurons. However,


[

NUMBER 12

[

DECEMBER 2005

this would imply K-ATP channel activation in these other cells in the MPTP model as well as in the weaver mouse and selective interaction of these cells with SN DA neurons but not VTA DA neurons. K-ATP channels are selectively activated in response to complex I inhibition in SN DA neurons but not in VTA DA neurons; this is unexpected, as adult VTA DA neurons have functional K-ATP channels with the same molecular make-up. However, given the ubiquitous expression of K-ATP channels containing Kir6.2 and SUR1 subunits, as well as their distinct physiological functions in brain and other tissues15,16,41, cell-specific differences in K-ATP channel regulation are expected. Our results suggest that differential degrees of uncoupling of the mitochondrial respiratory chain are critical for the control of K-ATP channel activity in DA midbrain neurons. Although extensive uncoupling in vitro activates K-ATP channels in all DA neurons, we have found that mild uncoupling has opposite effects on K-ATP channel activity in response to complex I inhibition: it causes decreased K-ATP conductances in SN DA neurons but increased K-ATP channel conductances in VTA DA neurons. The latter express higher mRNA levels of the uncoupling protein UCP-2. In accordance, mild uncoupling has been shown to predominantly reduce generation of ROS31, and decreased ROS will reduce the open probability of K-ATP channels20. Accordingly, our data suggest that the differential activation of Kir6.2- and SUR1-containing K-ATP channels in response to complex I inhibition is controlled upstream by different degrees of mitochondrial uncoupling in SN DA and VTA DA neurons. Mild basal uncoupling in VTA DA neurons seems well suited to prevent K-ATP channel activation in response to complex I inhibition, as further uncoupling by FCCP in these neurons enhances K-ATP channel activation. These findings also suggest that the neuroprotective effects of increased mitochondrial uncoupling (owing to increased levels of UCP-2 or coenzyme Q) in rodent and primate models of Parkinson disease33,42 might act by reducing the open probability of K-ATP channels in vulnerable SN DA neurons. In essence, the convergence of mitochondrial complex I dysfunction and oxidative stress (both of which are present in Parkinson disease) on the activity of K-ATP channels provides a previously unknown candidate mechanism for differential vulnerability of DA neurons that would couple metabolic stress with electrophysiological failure and selective death of SN DA neurons. While we have identified a mitochondrial upstream mechanism for differential K-ATP channel control in response to complex I inhibition, our study did not address downstream events that could link K-ATP channel–mediated electrical changes with death of DA neurons. Future studies must show to what extent K-ATP channel activation does occur in vivo, and if resulting electrical changes (such as reduced activity or functional silencing) are sufficient to kill mature DA neurons in a cellautonomous fashion, as recently reported for maturing DA neurons in vitro43. Alternatively, functional silencing and lack of dopamine release could generate additional network effects, such as imbalances of synaptic inputs in the intact brain, that might eventually cause the death of DA neurons. However, dopamine production and release itself is not required for survival of DA neurons in vivo44. Given the well-documented function of Kir6.2-containing K-ATP channel activation in protecting complex brains against short-term stressors28,36–39, our findings define a previously unknown role for K-ATP channels in dopaminergic pathophysiology: the promotion of long-term neurodegenerative processes. Accordingly, factors that increase K-ATP channel open probability might be beneficial during short-term metabolic demands but could also promote chronic disease. Indeed, in type 2 diabetes, the insufficient closure of b cell K-ATP channels, which control insulin secretion, is a key event17,45. Conse-

1749


ARTICLES quently, K-ATP channel polymorphisms or a diabetic metabolic state might predispose individuals to Parkinson disease. In light of this, it is notable that early studies suggested a high prevalence of impaired glucose tolerance in Parkinson disease and exacerbation of this impairment by conventional pharmacotherapy for Parkinson disease (reviewed in ref. 46), although more recent evidence is missing. Thus, in the future, cell type–specific drugs might selectively enhance mitochondrial uncoupling or inhibit K-ATP channel activity for neuroprotection, but a clinical correction of a potential diabetic metabolic state in Parkinson disease patients might provide an already available neuroprotective strategy. METHODS Animals and MPTP treatment. Kir6.2/ and control (Kir6.2+/+) mice were generated and back-crossed into C57Bl/6 mice as described26. Homozygous weaver (wv/wv Kir6.2/) and wv+/+ Kir6.2/ control mice were obtained by crossing Kir6.2/ and heterozygous weaver mice (B6CBACa-Aw-J/A-wv, Jackson Laboratory). Genotyping for the weaver mutation and the Kir6.2/ construct was performed by PCR after DNA isolation from small ear punches (QiaAmp DNA Mini Kit, Qiagen). Unless otherwise stated, 3-month-old male mice were used. The chronic MPTP intoxication protocol similar to that described previously8: 20 mg kg–1 MPTP in 5 mg ml–1 saline was injected subcutaneously, and 250 mg kg–1 probenecid in DMSO was injected intraperitoneally every 3.5 d over a period of 5 weeks. Mice were killed 1 week after the final injection. For acute MPTP effects, a single dose was injected. Procedures were approved by the UK Home Office and the German Regierungspra¨sidium Giessen. MPTP handling and safety were accordance to published safety guidelines. Retrograde tracing and immunohistochemistry. Under stereotactic control (Cartesian Instruments), 100 nl retrobeads (Lumafluor) were injected in dorsal striatum (bregma: 0.98 mm, lateral: 1.9 mm, ventral: 2.2 mm) or nucleus accumbens (bregma: 1.54 mm, lateral: 1.0 mm, ventral: 4.0 mm) of mice under general ketamine-domitor anaesthesia. After 1 week recovery, animals were sacrificed and injection sites verified. Perfusion-fixation, immunocytochemical experiments and confocal analysis were as described previously22. TH-stained midbrain and control sections were rehydrated and hematoxylin-eosin counterstained (Hematoxiline-Eosine Quick Stain, Vector Labs), dehydrated and coverslipped for further stereological analysis. Optical density of striatal TH immunoreactivity was analyzed using digital imaging software (Neurolucida, MicroBrightField, Adobe Photoshop and IGOR Wavemetrics). Stereological analysis. Total numbers of TH-positive or hematoxylin-eosin– positive neurons in SN and VTA were determined using an unbiased optical fractionator method (Stereoinvestigator, MicroBrightField). Regions of interests (SN and VTA) were identified according to established anatomical landmarks (Paxinos mouse brain atlas). We analyzed 30 serial 30-mm THimmunostained sections (with or without hematoxylin-eosin counterstaining) from control and MPTP-treated mice as well as from double-mutant homozygous wv/wv Kir6.2/ and wv/wv Kir6.2+/+ mice and appropriate wv+/+ Kir6.2/ and wv+/+ Kir6.2+/+ controls, covering the entire caudo-rostral extent of the midbrain. Actual section thickness after mounting and shrinking was consistently about 11 mm. Estimated total number of TH-positive neurons (N) was calculated as follows: N¼

X

Q

t h asf ssf

where h is the height of optical dissector (5 mm), ssf is the section sampling fraction (1), asf is the area sampling fraction (0.69), t is the mean section P thickness, and Q – is the sum of counted neurons for all sections47. Sampling grid dimension were 120 120 5 mm (x, y and z-axes). Quality of estimated total numbers was assessed according to the Gundersen coefficient of error CE2 and was in all cases r0.05. High-performance liquid chromatography (HPLC). For quantification of MPP+ with HPLC with fluorescence detection, striatal punches (B5 mg) were

1750

sonicated (Bandelin Sonoplus HD 70, with a Microtip) for 1 s in 20 volumes of ice-cold 5% trichloroacetic acid and centrifuged. Supernatant aliquots of 10 ml were separated over a 5-mm Sulpelcosil LC-SCX column (25 0.46 cm) with mobile phase of 10% ACN, 90% 0.1 M acetic acid and 0.075 M triethylamine HCl, pH 2.3, and flow rate of 1 ml min–1. Eluate was monitored by a MerckHitachi fluorescence detector F-1000 at 295 nm excitation and 375 nm emission. Peaks were quantified by peak height evaluation and by area integration with Millennium evaluation software (Millipore). For quantification of DA and DOPAC by HPLC with electrochemical detection, brain tissues were homogenized in 40 volumes of the mobile phase, and detection was essentially as described previously48. Electrophysiological recordings from adult mouse brain slices. Coronal midbrain slices (250 mm) of adult mice were sectioned after intracardial perfusion with ice-cold sucrose–artificial cerebrospinal fluid (ACSF) (in mM: 50 sucrose, 75 NaCl, 25 NaHCO3, 2.5 KCl, 1.25 NaH2PO4, 0.1 CaCl2, 6 MgCl2, and 2.5 glucose, oxygenated with 95% O2/5% CO2). After 90 min of recovery, slices were transferred to a recording chamber and perfused continuously at 2–4 ml min–1 with oxygenated ACSF (2.5 mM glucose, 22.5 mM sucrose) at 36 1C. Fast excitatory and inhibitory synaptic transmission was inhibited by 20 mM CNQX (6-cyano-7-nitroquinoxaline-2,3-dione) and 10 mM gabazine. Labeled SN and VTA DA neurons were visualized by infrared–differential interference contrast (IR-DIC) video microscopy and epifluorescence for detection of retrobeads21. Whole-cell and perforated-patch recordings, data acquisition and analysis were essentially as described22. Minimal subthreshold slope conductances (Gmin) of DA neurons were determined by fitting the linear slope conductance in the subthreshold range between –50 and –70 mV of membrane currents in response to 1-s voltage ramps from –30 to –120 mV after a 200-ms conditioning step to –30 mV. Concentration-conductance data for rotenone (or MPP+) were fitted according to the Hill relationship: G ¼ Gbase +ðGmax 2Gbase Þ/ð1+ðEC50 /½rotenonenÞ where n is the Hill coefficient and [rotenone] is the concentration of rotenone in nM. Laser microdissection. A photoablation and laser microdissection system with fluorescence option (PALM) was used. Serial coronal cryosections (8–12 mm) of mouse midbrains were cut using a Leica cryostat CM1850 and mounted on RNAse-free ultraviolet light–treated membrane slides (1 mm polyethylenenaphthalate membrane, PALM). Sections were fixed with ethanol (four washes with 75%, 95% 100%, 100%, respectively), optionally stained with cresyl violet in 100% ethanol, and dried. Fluorescence-labeled neurons were visualized (63 objective) and marked under fluorescence illumination (rhodamine filter, 555 nm excitation) and were cut and catapulted under bright-field DIC microscopy. Individual neurons or cell pools were catapulted directly into an adhesive cap (PALM). A mixture for cell lysis (0.5% Nonidet P40, Roche Diagnostics, and 20 U SUPERaseIn, Ambion) and cDNA synthesis was added directly into the lid, the tube was incubated upside-down for 90 s at 65 1C, and the tube was quickly cooled on ice before adding reverse transcriptase. RT-PCR. cDNA synthesis and qualitative and quantitative real-time single-cell PCR were essentially as described49. For multiplex PCR, HotstarTaq (Qiagen) was used, and for nested PCR, RedTaq (Sigma) was used. Primer sequences for SUR1, SUR2, calbindin D28k, TH, and GAD67 were as described25; primers for GAD65 were 166-CATACGCAGACAGCACGTTT (F1), 1070-AAAAGATTCC ATCGCCAGAG (R1), 606-GGGATGTCAACTACGCGTTT (F2), 994-CAC AAATACAGGGGCGATCT (R2). For quantitative real-time PCR, a 10% aliquot of cDNA from laser-microdissected cell pools was used for qualitative multiplex nested PCR to confirm marker expression (TH and CBd28k for DA neurons) and to exclude possible contamination by glial cells (as detected by glial fibrillary acidic protein (GFAP) signal) or GABAergic neurons (as detected by GAD65/GAD67 signals). All real-time Taqman assays were predesigned by Applied Biosystems and labeled with FAM as a reporter and a nonfluorescent quencher. Several assays were tested to choose those with the best performance and reproducibility, as indicated by low standard deviation for replicates with serial dilutions of control cDNA over four orders of magnitude for generation

VOLUME 8

[

NUMBER 12

[


ARTICLES of standard curves49 This strategy was crucial for real-time single-cell analysis. UCP-3 was routinely detected in mouse brain cDNA (standard curve: slope ¼ –3.06, R2 ¼ 0.992) but not in DA neurons.


Statistics. All data are given as mean ± s.e.m. To evaluate statistical significance, we used Student’s t-tests. Normal, parametric data were compared by a twotailed, unpaired t-test. A value of P o 0.05 was considered to be statistically significant and is indicated by *. Accession codes. GenBank: GAD65, NM_008078.1. Applied Biosystems identifying assay identifiers: SUR1/ABCC8, Mm00803450_m1; SUR2/ABCC9, Mm00441638_m1; Kir6.1/KCNJ8, Mm00434620_m1; Kir6.2/KCNJ11, Mm00440050_s1; UCP-2, Mm00495907_g1, UCP-3, Mm00494074_m1. ACKNOWLEDGMENTS We are grateful to F.M. Ashcroft and R. Veh for support. We thank the animal facility of Marburg University for animal care and J. Clark, D. Meyer, E. Naudascher and H. Neuhoff for technical support. This work was supported by the Parkinson’s Disease Society, UK, the Medical Research Council, Bundesministerium fuer Bildung und Forschung (BMBF-NGFNII), Gemeinnu¨tzige Hertie Foundation, Royal Society, Deutsche Forschungsgemeinschaft (J.W.), and fellowships from New College, Oxford and the Royal Society (B.L.) and Exeter College, Oxford (J.R.). AUTHOR CONTRIBUTIONS B.L. performed molecular biology and animal model work; O.H. performed stereological analysis; J.W. performed HPLC experiments; T.M & S.S. provided the Kir6.2/ mouse; J.R. performed MPTP, tracing and electrophysiology experiments; and B.L. and J.R. designed the study and cowrote the ms. COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/

1. Morrison, B.M., Hof, P.R. & Morrison, J.H. Determinants of neuronal vulnerability in neurodegenerative diseases. Ann. Neurol. 44, S32–S44 (1998). 2. Damier, P., Hirsch, E.C., Agid, Y. & Graybiel, A.M. The substantia nigra of the human brain. II. Patterns of loss of dopamine-containing neurons in Parkinson’s disease. Brain 122, 1437–1448 (1999). 3. Greenamyre, J.T. & Hastings, T.G. Biomedicine. Parkinson’s-divergent causes, convergent mechanisms. Science 304, 1120–1122 (2004). 4. Moore, D.J., West, A.B., Dawson, V.L. & Dawson, T.M. Molecular pathophysiology of Parkinson’s disease. Annu. Rev. Neurosci. 28, 57–87 (2005). 5. Dauer, W. & Przedborski, S. Parkinson’s disease: mechanisms and models. Neuron 39, 889–909 (2003). 6. Schapira, A.H. Causes of neuronal death in Parkinson’s disease. Adv. Neurol. 86, 155–162 (2001). 7. Moratalla, R. et al. Differential vulnerability of primate caudate-putamen and striosome-matrix dopamine systems to the neurotoxic effects of 1-methyl-4-phenyl1,2,3,6-tetrahydropyridine. Proc. Natl. Acad. Sci. USA 89, 3859–3863 (1992). 8. Petroske, E., Meredith, G.E., Callen, S., Totterdell, S. & Lau, Y.S. Mouse model of Parkinsonism: a comparison between subacute MPTP and chronic MPTP/probenecid treatment. Neuroscience 106, 589–601 (2001). 9. Greene, J.G., Dingledine, R. & Greenamyre, J.T. Gene expression profiling of rat midbrain dopamine neurons: implications for selective vulnerability in parkinsonism. Neurobiol. Dis. 18, 19–31 (2005). 10. Betarbet, R. et al. Chronic systemic pesticide exposure reproduces features of Parkinson’s disease. Nat. Neurosci. 3, 1301–1306 (2000). 11. Przedborski, S. & Vila, M. The 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine mouse model: a tool to explore the pathogenesis of Parkinson’s disease. Ann. NY Acad. Sci. 991, 189–198 (2003). 12. Langston, J.W. The etiology of Parkinson’s disease with emphasis on the MPTP story. Neurology 47, S153–S160 (1996). 13. Hoglinger, G.U. et al. Dysfunction of mitochondrial complex I and the proteasome: interactions between two biochemical deficits in a cellular model of Parkinson’s disease. J. Neurochem. 86, 1297–1307 (2003). 14. Testa, C.M., Sherer, T.B. & Greenamyre, J.T. Rotenone induces oxidative stress and dopaminergic neuron damage in organotypic substantia nigra cultures. Brain Res. Mol. Brain Res. 134, 109–118 (2005). 15. Seino, S. & Miki, T. Gene targeting approach to clarification of ion channel function: studies of Kir6.x null mice. J. Physiol. (Lond.) 554, 295–300 (2004).


[

NUMBER 12

[

DECEMBER 2005

16. Bryan, J., Vila-Carriles, W.H., Zhao, G., Babenko, A.P. & Aguilar-Bryan, L. Toward linking structure with function in ATP-sensitive K+ channels. Diabetes 53, S104–S112 (2004). 17. Ashcroft, F. & Rorsman, P. Type 2 diabetes mellitus: not quite exciting enough? Hum. Mol. Genet. 13, R21–R31 (2004). 18. Liss, B., Bruns, R. & Roeper, J. Alternative sulfonylurea receptor expression defines metabolic sensitivity of K-ATP channels in dopaminergic midbrain neurons. EMBO J. 18, 833–846 (1999). 19. Liu, Y. & Gutterman, D.D. Oxidative stress and potassium channel function. Clin. Exp. Pharmacol. Physiol. 29, 305–311 (2002). 20. Avshalumov, M.V., Chen, B.T., Koos, T., Tepper, J.M. & Rice, M.E. Endogenous hydrogen peroxide regulates the excitability of midbrain dopamine neurons via atp-sensitive potassium channels. J. Neurosci. 25, 4222–4231 (2005). 21. Liss, B. & Roeper, J. Correlating function and gene expression of individual basal ganglia neurons. Trends Neurosci. 27, 475–481 (2004). 22. Neuhoff, H., Neu, A., Liss, B. & Roeper, J.I. (h) channels contribute to the different functional properties of identified dopaminergic subpopulations in the midbrain. J. Neurosci. 22, 1290–1302 (2002). 23. Jiang, C., Sigworth, F.J. & Haddad, G.G. Oxygen deprivation activates an ATP-inhibitable K+ channel in substantia nigra neurons. J. Neurosci. 14, 5590–5602 (1994). 24. Mercuri, N.B. et al. Effects of anoxia on rat midbrain dopamine neurons. J. Neurophysiol. 71, 1165–1173 (1994). 25. Liss, B., Neu, A. & Roeper, J. The weaver mouse gain-of-function phenotype of dopaminergic midbrain neurons is determined by coactivation of wvGirk2 and K-ATP channels. J. Neurosci. 19, 8839–8848 (1999). 26. Miki, T. et al. Defective insulin secretion and enhanced insulin action in KATP channeldeficient mice. Proc. Natl. Acad. Sci. USA 95, 10402–10406 (1998). 27. Tarasov, A., Dusonchet, J. & Ashcroft, F. Metabolic regulation of the pancreatic beta-cell ATP-sensitive K+ channel: a pas de deux. Diabetes 53, S113–S122 (2004). 28. Lin, Y.F., Raab-Graham, K., Jan, Y.N. & Jan, L.Y. NO stimulation of ATP-sensitive potassium channels: involvement of Ras/mitogen-activated protein kinase pathway and contribution to neuroprotection. Proc. Natl. Acad. Sci. USA 101, 7799–7804 (2004). 29. Lin, Y.F., Jan, Y.N. & Jan, L.Y. Regulation of ATP-sensitive potassium channel function by protein kinase A-mediated phosphorylation in transfected HEK293 cells. EMBO J. 19, 942–955 (2000). 30. Baukrowitz, T. et al. PIP2 and PIP as determinants for ATP inhibition of KATP channels. Science 282, 1141–1144 (1998). 31. Krauss, S., Zhang, C.Y. & Lowell, B.B. The mitochondrial uncoupling-protein homologues. Nat. Rev. Mol. Cell Biol. 6, 248–261 (2005). 32. Brookes, P.S. Mitochondrial H(+) leak and ROS generation: an odd couple. Free Radic. Biol. Med. 38, 12–23 (2005). 33. Horvath, T.L. et al. Coenzyme Q induces nigral mitochondrial uncoupling and prevents dopamine cell loss in a primate model of Parkinson’s disease. Endocrinology 144, 2757–2760 (2003). 34. Andrews, Z.B. et al. Uncoupling protein-2 is critical for nigral dopamine cell survival in a mouse model of Parkinson’s disease. J. Neurosci. 25, 184–191 (2005). 35. Roffler-Tarlov, S. & Graybiel, A.M. Weaver mutation has differential effects on the dopamine-containing innervation of the limbic and nonlimbic striatum. Nature 307, 62–66 (1984). 36. Ballanyi, K. Protective role of neuronal KATP channels in brain hypoxia. J. Exp. Biol. 207, 3201–3212 (2004). 37. Yamada, K. et al. Protective role of ATP-sensitive potassium channels in hypoxiainduced generalized seizure. Science 292, 1543–1546 (2001). 38. Hernandez-Sanchez, C. et al. Mice transgenically overexpressing sulfonylurea receptor 1 in forebrain resist seizure induction and excitotoxic neuron death. Proc. Natl. Acad. Sci. USA 98, 3549–3554 (2001). 39. Zingman, L.V. et al. Kir6.2 is required for adaptation to stress. Proc. Natl. Acad. Sci. USA 99, 13278–13283 (2002). 40. Patil, N. et al. A potassium channel mutation in weaver mice implicates membrane excitability in granule cell differentiation. Nat. Genet. 11, 126–129 (1995). 41. Liss, B. & Roeper, J. Molecular physiology of neuronal K-ATP channels. Mol. Membr. Biol. 18, 117–127 (2001). 42. Conti, B. et al. Uncoupling protein 2 protects dopaminergic neurons from acute 1,2,3,6methyl-phenyl-tetrahydropyridine toxicity. J. Neurochem. 93, 493–501 (2005). 43. Salthun-Lassalle, B., Hirsch, E.C., Wolfart, J., Ruberg, M. & Michel, P.P. Rescue of mesencephalic dopaminergic neurons in culture by low-level stimulation of voltagegated sodium channels. J. Neurosci. 24, 5922–5930 (2004). 44. Zhou, Q.Y. & Palmiter, R.D. Dopamine-deficient mice are severely hypoactive, adipsic, and aphagic. Cell 83, 1197–1209 (1995). 45. O’Rahilly, S., Barroso, I. & Wareham, N.J. Genetic factors in type 2 diabetes: the end of the beginning? Science 307, 370–373 (2005). 46. Craft, S. & Watson, G.S. Insulin and neurodegenerative disease: shared and specific mechanisms. Lancet Neurol. 3, 169–178 (2004). 47. Keuker, J.I., Vollmann-Honsdorf, G.K. & Fuchs, E. How to use the optical fractionator: an example based on the estimation of neurons in the hippocampal CA1 and CA3 regions of tree shrews. Brain Res. Brain Res. Protoc. 7, 211–221 (2001). 48. Alburges, M.E., Narang, N. & Wamsley, J.K. Alterations in the dopaminergic receptor system after chronic administration of cocaine. Synapse 14, 314–323 (1993). 49. Liss, B. et al. Tuning pacemaker frequency of individual dopaminergic neurons by Kv4.3L and KChip3.1 transcription. EMBO J. 20, 5715–5724 (2001).

1751


ARTICLES

BK channel b4 subunit reduces dentate gyrus excitability and protects against temporal lobe seizures Robert Brenner1,4, Qing H Chen1,4, Alex Vilaythong2, Glenn M Toney1, Jeffrey L Noebels2 & Richard W Aldrich3 Synaptic inhibition within the hippocampus dentate gyrus serves a ‘low-pass filtering’ function that protects against hyperexcitability that leads to temporal lobe seizures. Here we demonstrate that calcium-activated potassium (BK) channel accessory b4 subunits serve as key regulators of intrinsic firing properties that contribute to the low-pass filtering function of dentate granule cells. Notably, a critical b4 subunit function is to preclude BK channels from contributing to membrane repolarization and thereby broaden action potentials. Longer-duration action potentials secondarily recruit SK channels, leading to greater spike frequency adaptation and reduced firing rates. In contrast, granule cells from b4 knockout mice show a gain-of-function for BK channels that sharpens action potentials and supports higher firing rates. Consistent with breakdown of the dentate filter, b4 knockouts show distinctive seizures emanating from the temporal cortex, demonstrating a unique nonsynaptic mechanism for gate control of hippocampal synchronization leading to temporal lobe epilepsy.

Dentate granule cells are the first checkpoint for cortical information entering the hippocampal formation. These neurons are under heavy inhibitory synaptic input from local interneurons, which maintain granule cell firing at low frequencies, even when granule cells are strongly driven by afferent cortical input. Defects in these inhibitory synaptic inputs are widely regarded as a cause of breakdown in the ‘gate control’ function of the dentate granule cells that leads to temporal lobe seizures of the hippocampus1–4. However, very few studies have addressed the intrinsic firing behavior of hippocampal granule cells as an alternative mechanism that might contribute to the protective filtering properties of these cells. In many central neurons, calcium-activated channels have an important role in controlling action potential firing properties and intrinsic firing behavior. Small conductance (SK-type) calcium-activated potassium channels include a family of three voltage-insensitive potassium channels (SK1–SK3) that are blocked by the bee venom apamin. Apamin is uniquely selective for these channels (half-maximal inhibitory concentration (IC50) < 5 nM) and has been invaluable in understanding the role of SK channels in neuronal firing properties5–7. SK channels demonstrate a cumulative activation during trains of action potentials that causes an increased interspike interval and can terminate action potential firing8–11. The mechanism is due to the medium afterhyperpolarization (mAHP) that delays the onset to subsequent action potential firing. Using gene knockout studies, the mAHP has been unambiguously ascribed to SK channels12. The specific SK channel blocker apamin selectively affects this component of the AHP6,13.

In contrast to SK channels, BK channels contribute minimally to setting firing rates14. This is primarily because BK channels require coincident calcium entry and membrane depolarization to open channels under physiological conditions15 and therefore are not likely to be active during the interspike interval. Rather, BK channels contribute to action potential repolarization and, in some neurons, to an early component of the fast afterhyperpolarization, fAHP14. Generally, the contribution of BK channels in neurons can be determined using the highly specific BK channel blocker iberiotoxin5. More recently, the organic blocker paxilline, with IC50 in the submicromolar range, has been used to block BK channels and has similar effects on action potentials as iberiotoxin14,16,17. In these studies, it seems that biophysical properties of BK channels, such as fast deactivation or inactivation, restrict the contribution of BK channels to setting the neuronal firing rate. For instance, cerebellar Purkinje cells have BK channels with fast deactivation gating after action potential repolarization, minimizing their role in setting the interspike interval18, whereas hippocampal CA1 pyramidal cells and lateral amygdalar neurons have inactivating-type BK channels that contribute to action potential repolarization in early but not later spikes in a train19,20. However, an additional subtype of BK channel, the type II BK channel, is very slowly gated, is resistant to iberiotoxin and does not have intrinsic inactivation. This channel type has been detected in posterior pituitary nerve terminals and in synaptic vesicle membrane preparations from brain21,22. Heterologous coexpression of the accessory b4 subunit with BK a subunits confers many properties similar to those of type II BK channels23–27, including resistance to iberiotoxin block23 and very

1Department of Physiology, University of Texas Health Science Center at San Antonio, 7703 Floyd Curl Drive, San Antonio, Texas 78229, USA. 2Department of Neurology, Baylor College of Medicine, One Baylor Plaza, Houston, Texas 77030, USA. 3Howard Hughes Medical Institute, Molecular and Cellular Physiology, Stanford University School of Medicine, Stanford, California 94305, USA. 4These authors contributed equally to this manuscript. Correspondence should be addressed to R.B. ([email protected]) or R.W.A. ([email protected]).

Received 8 August; accepted 23 September; published online 30 October 2005; doi:10.1038/nn1573

1752

VOLUME 8

[

NUMBER 12

[


ARTICLES β4 gene targeting SfuI

b

ApaI

BglII

BamHI

β4 locus

BglII

2.8 kb

KpnI KpnI

a

c F1 mice

WT 12* 43* 53

–/– +/+ +/–

KpnI

BglII

eGFP SV40 neo

Northern analysis –/– +/– +/+

–/– +/– +/+

4.8 kb

Thymidine kinase

4.8 kb

2.8 kb

SfuI

BglII

BglII

2.8 kb eGFP SV40 neo

β4 probe

BgIII digest

eGFP probe

4.8 kb

Figure 1 Generation of b4 gene-targeted mice. (a) Top: the b4 locus flanking exon 1. Middle: the gene-targeting vector. Bottom: gene-targeted locus. Filled dark gray rectangle in b4 locus represents region of DNA used as a probe for genomic Southern analysis in b. (b) Genomic Southern analysis of left-arm polymorphism in the BglII restriction fragment demonstrating knockout of the b4 locus in two embryonic stem (ES) cell clones (12 and 43, marked with an arrow). (c) Genomic Southern analysis of F2 mice showing homozygous knockout (/) of the b4 locus. (d) Northern analysis showing elimination of b4 mRNA (left) in the b4/ mice and eGFP expression (right) in gene-targeted b4/ and b4+/ mice. Gel slice at left is an ethidium-stained RNA standard containing 2.4-kb (above) and 1.4-kb (below) markers.

slow gating kinetics (activation taking tens of milliseconds)28 that are inconsistent with the regulation of action potential repolarization29. In order to understand the physiological role of the b4 subunit, we generated b4-null mice. We found that the b4 subunit underlies type II BK channel properties. In hippocampus dentate gyrus granule cells, the b4 subunit prevents BK channels from contributing to action potential repolarization. Notably, downregulation of BK channels reduces excitability by indirectly sustaining SK channel activation during the interspike interval and reducing action potential spiking rates. Thus, regulation of intrinsic firing properties by the b4 subunit contributes to the protective ‘filter’ function of dentate granule cells against excessive downstream hippocampal synchronization. In contrast, gene-targeted b4 null mice show a gain-of-function for BK channels, a reduced contribution of SK channels, increased action potential firing rates and temporal lobe seizures. RESULTS Knockout of b4 subunit leads to an epileptic phenotype We designed a targeting vector to replace the first exon of the b4 gene (Kcnmb4) with enhanced green fluorescent protein (eGFP; Fig. 1a).


[

This serves to delete the b4 exon and also serves as a marker for expression driven by the b4 gene promoter. Deletion of the exon eliminates approximately 50% of the protein-coding region, including the first transmembrane domain. Downstream of eGFP, the neomycin resistance gene was inserted to select for transfected embryonic stem cells and terminate transcription of the gene. Genomic Southern analysis identified two embryonic stem cell clones that were gene targeted (Fig. 1b), and one clone produced germline transmission of the knockout allele (Fig. 1c). Northern analysis of brain mRNA showed an elimination of the b4 mRNA in the targeted mice and expression of eGFP mRNA in the heterozygous and homozygous knockout mice (Fig. 1d). Deletion of the b4 subunit produced mice (referred to as b4/ mice) with no overt neurological disorder. To determine the effect of b4 knockout on brain activity, we performed chronic video electroencephalogram (EEG) monitoring of freely moving mice using a ten– cortical electrode array (Fig. 2). EEG records show interictal spike discharges appearing initially over the temporal cortex that spread into a fully generalized seizure episode in the neocortex (Fig. 2a). Throughout the electrographic seizure episodes, the animals do not demonstrate

a Frontal

Figure 2 Deletion of the b4 subunit causes nonconvulsive partial onset seizures. (a) EEG recording with cortical surface electrodes (L, left; R, right). Seizure begins with a few interictal discharges in occipital, temporal and parietal regions, and then spreads frontally to involve the whole brain, illustrating partial onset. (b) Depth electrode recordings in right (R) hippocampus (lowest trace) show that hippocampal discharges (*), along with spikes in overlying temporal cortex, occur in relative isolation from other cortical regions (upper traces). Fr ctx, frontal cortex; Par ctx, parietal cortex; Temp ctx, temporal cortex; Hip, hippocampus. (c–e) Anti-GFP staining from brain sections of gene-targeted b4 mice and wildtype control mice. (c) Control wild-type brain lacks staining for the eGFP reporter (coronal section). (d) In b/ brain, b4 expression is highest in the dentate gyrus of the hippocampus (coronal section). e shows boxed area in d at a higher magnification. Scale bars: 100 mm for c and d, 50 mm for e.

Occipital Temporal Parietal


β4 targeted locus

SalII

Exon1

β4 KO vector

d

ES clones

b

L

L

R

Fr ctx R

L L Par ctx R

R L

L

R

Temp ctx L

L R

Hip

R

*

200 mV 1s

c

NUMBER 12

WT

[

DECEMBER 2005

d

β4–/–

e

β4–/–

1753

ARTICLES

Iberiotoxin

+40 Washout

Paxilline

0 4 pA –/–

β4 WT

Dwell time (ms)

40 30

0.4

20

0.2 0

d

10 0

20 40 Voltage (mV)

60

0

* 0

20 40 Voltage (mV)

* 60

any tonic or clonic behaviors, but remain either motionless or engage in repetitive grooming behaviors despite the paroxysmal synchronous discharge (Supplementary Video 1 online). Depth electrode recordings (Fig. 2b) show early isolated epileptiform discharges in the hippocampus with little abnormal activity in the cortex, consistent with a probable hippocampal origin. The electrographic uncoupling of abnormal network synchronization with convulsive behavior can occur during the preconvulsive stages in human partial seizures of temporal lobe origin30 and has not been previously observed in genetic mouse models of epilepsy, suggesting that the selective loss of the BK channel b4 subunit spares other extrahippocampal inhibitory mechanisms that prevent the recruitment of convulsive motor activity. Temporal lobe seizures are thought to originate frequently in the hippocampus or adjacent entorhinal cortex. Histochemical staining of the eGFP b4 reporter gene showed the highest expression in the dentate gyrus of the hippocampus (Fig. 2d,e; dentate granule cells shown in Supplementary Fig. 1 online) but also showed weaker expression in most other regions of the limbic system, including the CA3 region of the hippocampus, entorhinal cortex, subiculum (Supplementary Fig. 1) and the amygdala (data not shown). However, expression was not seen in some nonneuronal tissue, such as vascular smooth muscle (Supplementary Fig. 1). This is consistent with the expression pattern

Figure 4 Firing properties of dentate gyrus granule cells evoked by current injections. (a) Action potentials evoked by current clamp in b4/ cells (red) and wild-type cells (black). (b) Number of action potentials evoked at different current injections. (c) Representative tenth action potential waveform during a 300-pA current injection. The wild-type action potentials were broader and had a slower decay of the afterhyperpolarization (dotted line). (d) Action potential width measured at one-half height. (e) Amplitude of the fast afterhyperpolarization measured from the prespike voltage to peak after hyperpolarization. (f) Time constant of deactivation of the afterhyperpolarization. (g) Interspike intervals preceding the 15th action potential during a 300-pA current injection. b4/ cells (red) were compared with wild-type cells (black), either excluding wild-type cells with failed action potentials (‘No failures’) or including those cells (‘With AP failures’). *, P o 0.05; **, P o 0.005; ***, P o 0.0005. Error bars in all panels represent s.e.m.

1754

a

determined using in situ hybridization26. Type II BK channels have been well characterized in the posterior pituitary terminals31,32; we found very strong expression in this tissue as well as in the anterior pituitary (Supplementary Fig. 1). It is notable that b4 expression was highest in the dentate gyrus, because the dentate gyrus is generally regarded as a filter that prevents propagation of synchronous activity into the seizure-prone hippocampus and related entorhinal cortex33,34 and is consistent with the possibility that a defect in BK channel properties in dentate gyrus granule cells could result in the breakdown of the filtering capacity of this network. b4/ converts BK channels from type II to fast-gated type I To investigate whether b4/ alters BK channel properties in dentate gyrus granule cells, we obtained single-channel records from BK channels in these cells. Records showed large-conductance potassium channels that are voltage-dependent (Fig. 3a). In wild-type mice, these channels have long burst durations (slow gating; Fig. 3a,d) and are resistant to block from 100 nM iberiotoxin (Fig. 3c, N ¼ 4). The channels were, however, blocked by paxilline, a specific blocker of BK channels that does not discriminate between subtypes (Fig. 3c, right, N ¼ 4)17. These results indicate that granule cells from wild-type mice express iberiotoxin-resistant type II BK channels. In contrast, BK

β4–/–

b

WT

150 pA

250 pA

–/–

β4 WT

40 30 20

40 mV

**

*** ***

**

10 0

350 pA

**

50

100 150 200 250 300 350 400 Current injection (pA)

300 ms

c

d 10 mV 2 ms

β4–/– WT

e

2

VOLUME 8

f 10

* 1

5

0

[

NUMBER 12

0

[

*

g 6

4

2

0

100

*

80 60 40 20 0

*

With AP failures

–/–

β4 WT

Channel open probability

No failures

b

200 ms

Number of action potentials


4 pA 100 ms

Interspike interval (ms)

+20

Figure 3 Deletion of the b4 subunits converts slow-gated, type II BK channels to fast-gated type I BK channels. Records of single BK channels excised from b4/ and wild-type (WT) dentate gyrus granule cells. (a) Single BK channel openings at holding potential of 0 to +60 mV. (b) Channel open probability is not significantly different between wildtype and b4/ channels. (c) Example of single BK channel openings at +40 mV holding potential in 18 mM calcium before (top row) and after 100 nM iberiotoxin (second row). Channels from b4/ cells are sensitive to iberiotoxin block (N ¼ 4/4 channels), whereas wild-type channels are completely resistant to iberiotoxin (N ¼ 4/4 channels). Lower left trace shows that washout of iberiotoxin in the knockout restores BK channels. Lower right trace shows that iberiotoxin-resistant BK channels are sensitive to the BK channel blocker paxilline (5 mM). (d) Burst duration in wild-type cells (black) are longer, reflecting the slow gating kinetics of type II BK channels. Knockout of b4 (open bars) converts BK channels to short-burst durations. *, P o 0.05 (unpaired Student’s t-test). All single-channel records were with outside/out excised patches and 18 mM buffered calcium solution in the cytoplasmic side. Error bars in all panels represent s.e.m.

AHP deactivation tau (ms)

+60

WT

Control

AHP size (mV)

β4–/–

c

WT

Action potential width (ms)

β4–/–

a


ARTICLES

N

Resting membrane potential (mV)

Action potential amplitude (mV)

Membrane time constant (ms)

Input resistance (GO)

WT

19

87.6 ± 0.7

89.2 ± 1.9

19.4 ± 1.0

0.60 ± 0.1

b4/ WT, paxilline

10 9

86.8 ±1.4 86.9 ± 0.9

94.9 ± 1.9 90.9 ± 3.2

21.8 ± 1.5 17.0 ± 5.4

0.61 ± 0.1 0.58 ± 0.1

b4/, paxilline WT, apamin

12 7

83.5 ±1.4 88.3 ± 0.4

89.4 ± 1.8 96.5 ± 1.7

19.7 ± 6.4 17.4 ± 1.7

0.62 ± 0.1 0.62 ± 0.1

BK, apamin WT, BAPTA

7 7

87.9 ± 1.1 89.6 ± 0.7

96.3 ± 1.2 92.0 ± 3.4

18.0 ± 1.8 25.9 ± 2.3

0.58 ± 0.1 0.7 ± 0.1

b4/, BAPTA

6

92.5 ± 1.0

97.0 ± 3.5

32.6 ± 4.4

0.8 ± 0.1

Values are mean ± s.e.m. Measurements were made according to Methods. Unpaired Student’s t-test showed no significant difference between wild-type and b4/ cells or between wild-type and b4/ cells with different treatments.

channels from b4/ mice had shorter burst durations (Fig. 3a,d) and were blocked by 100 nM iberiotoxin (Fig. 3c, N ¼ 4). At these calcium concentrations (18 mM) we did not see a significant difference in channel open probability between knockout and wild-type cells (Fig. 3b). On average, we saw a higher density of BK channels per patch in b4/ patches than in wild-type patches. Wild-type mice had 1.8 ± 0.4 channels per patch, whereas b4/ mice had 3.3 ± 0.6 channels per patch (P ¼ 0.04, Student’s unpaired t-test). In summary, these results provide direct evidence that the b4 subunit forms the basis for gating kinetics and toxin sensitivity of type II BK channels.

a

b

β4–/–

WT

β4–/–, paxilline

[

β4–/–, paxilline WT, paxilline

40

β4–/– WT

**

*

*

30

NS NS

20

NS

10 0

100 pA

40 mV

200 pA

300 pA

400 pA

300 ms

1.5 1.0 0.5

*

d

NS

* 7.5 5.0 2.5

NUMBER 12

[

DECEMBER 2005

0

e

50 40 30 20 10 0

*

f Number of action potentials

2.0

NS

Interspike interval (ms)

c 2.5

0


WT, paxilline

200 pA


Figure 5 Effect of BK channel block on firing properties and action potential waveforms. (a) Examples of action potentials during 200-pA current injection without (top) and with (bottom) the BK channel blocker paxilline. (b) Number of action potentials evoked at different current injections. Paxilline reduces the firing frequency in b4/ cells but has no effect on firing frequency in wild-type cells. (c,d) Paxilline broadens action potentials (c) and slows the AHP repolarization (d) of b4/ cells but not wildtype cells. Measured at tenth action potential after 250-pA current injection. (e) Effect of intracellular 10 mM BAPTA on interspike interval measured at tenth action potential during a 3,000-pA current injection. (f) Effect of intracellular 10 mM BAPTA on action potential firing frequency. *, P o 0.05; **, P o 0.01. Error bars in all panels represent s.e.m.


interspike intervals (Fig. 4g and Supplementary Fig. 2). Even when compared with wild-type cells lacking action potential failure, b4/ cells had a significantly shorter interspike interval (Fig. 4g). Moreover, the b4/ cells continued to fire action potentials at much higher current injections with very little action potential failure (Fig. 4a, 350 pA; no failures in 10/10 cells). Thus, the shorter interspike intervals and the absence of action potential failures led to a substantial increase in neuronal firing rates in b4/ cells (Fig. 4b). These results indicate that granule neurons are limited in their firing frequency, whereas conversion of type II BK channels to type I–like, fast-gated BK channels in the b4/ neurons permit action potential firing at much higher rates. To further understand the mechanisms regulating firing frequency in b4/ cells, we analyzed changes in membrane properties and action potential waveforms. Resting membrane potentials and input resistance of granule cells were not affected by b4 knockout (Table 1). Representative action potentials for b4/ and wild-type granule cells are shown (Fig. 4c). The wild-type action potentials were broader and had a slower decay of the afterhyperpolarization. We analyzed the action potential waveforms at the tenth action potential during a 300-pA current injection, and we found that the wild-type cells had broader action potentials (Fig. 4d). The sharpening of the action potential in b4/ cells indicates that the b4 subunit inhibits BK channel activity during the repolarization phase of the action potential. This is likely to be a consequence of the very slow activation kinetics described for BK channels coexpressed with b4 subunits, similar to observations from

b4/ speeds AHP decay and increases spiking frequency We used current-clamp recordings in granule cells to investigate the consequence of b4/ on firing properties. Examples of b4/ and wild-type action potential traces are shown in Figure 4a. Current injections into wild-type granule cells elicited action potentials that seem to be limited in their firing frequency by two mechanisms. First, the interspike interval during the current injection increased over time and resulted in fewer action potentials (Fig. 4g and Supplementary Fig. 2). On average, the interspike interval approximately triples during a 900-ms-duration, 300-pA current injection. Second, higher current injection in some cells resulted in failure of action potential electrogenesis by the end of the current step (Fig. 4a, 350-pA trace; failure seen in 8/19 cells). In contrast, the b4/ mice showed very little spike frequency adaptation, which resulted in much shorter



Table 1 Summary of membrane properties of wild-type and b4/ dentate granule cells

β4–/– BAPTA WT BAPTA

β4–/– WT

* *

40

* 30 20 10 0

100 pA

200 pA

300 pA

400 pA

1755

β4–/–

a

β4–/–, apamin


40 WT, apamin

250 pA

40 mV 300 ms

c3

d

e

15

β4–/– WT

*

10

100 pA

b4 prevents BK channel repolarization of action potentials Two theories may explain the lower firing frequencies in wild-type mice and higher firing frequencies in the b4/ mice. One possibility is that the b4 subunits cause BK channel currents to have properties similar to SK-type channel currents. Such channels would have a sustained activation and would result in spike frequency adaptation in wildtype but not knockout mice. In this case, BK channels containing b4 subunits would contribute directly to the slow-deactivating, afterhyperpolarizing current. Another possibility is that the b4 subunit reduces BK channel activity during the spike, resulting in action potential broadening. This would result in greater calcium influx during the action potential and would indirectly produce greater activation of calcium-activated, SK-type potassium channels in wildtype but not in b4/ cells. We could distinguish between these possibilities by using paxilline to block BK channels in the b4/ and wild-type cells. If BK channels contribute directly to the slow decay of the AHP, then we would expect that paxilline would affect this component of the action potential and increase firing frequency in wild-type cells. Our results support the latter hypothesis (Fig. 5). Paxilline had little effect on wild-type cells, indicating that BK channels in wild-type cells do not control the firing rate directly (Fig. 5a,b). As expected from the firing rate, the slow decay of the AHP was unaffected by paxilline block (Fig. 5d). With the addition of paxilline, action potentials became broader in wild-type cells, but this was not statistically significant (Fig. 5c). In contrast, paxilline block markedly reduced the firing rate of b4/ cells (Fig. 5a,b). Consistent with the reduced firing rate, the AHP decay was slowed to wild-type levels (Fig. 5d). Furthermore, paxilline block significantly broadened action potentials to similar levels as in wild-type cells (Fig. 5c). These results indicate that type II BK channels contribute little to the action potential repolarization or

10

5

0

4

2

300 pA

40 Interspike interval (ms)


2

200 pA

*

f

8

6

*

20

* *

β4–/–, apamin WT, apamin

30

0

1 heterologous expression studies28. In some neurons, BK channels also contribute to an AHP after the spike29. We found that b4/ 0 decreased the AHP size (Fig. 4e) and had a marked effect on the kinetic decay of the AHP (Fig. 4f), allowing the AHP in the knockout to rise to firing threshold much more quickly. This shows that b4/ mice have a significantly smaller time constant to reach threshold for firing, and this property is likely to explain the higher firing frequency in b4/ cells compared with wild-type cells.

1756

b

WT

fAHP amplitude (mV)

Figure 6 SK channel block affects firing properties of wild-type but not b4/ cells. (a) Examples of action potentials during 250-pA current injection without (top) and with (bottom) the SK channel blocker apamin (100 nM). (b) Number of action potentials evoked after injection of different currents. Apamin increases the firing frequency in wild-type cells but does not have a significant effect on firing frequency in b4/ cells. (c) Apamin broadens action potentials in b4/ cells. (d) Apamin does not have a significant effect on fAHP amplitude. (e) Apamin speeds AHP repolarization in wild-type cells but has less of an effect in b4/ cells. (f) Apamin reduces the interspike interval in wild-type cells but not in b4/ cells. c–f were measured during a 300-pA current injection at the tenth action potential. *, P o 0.05 (wild-type cells versus b4/ cells). Error bars in all panels represent s.e.m.



ARTICLES

30

20

10

0

0

the firing rate, whereas BK channels lacking the b4 subunit sharpen action potentials and support abnormal high-frequency firing. b4 downregulation of BK channels recruits SK channels Because BK channels are negative feedback regulators of calcium influx in many neurons35, a potential mechanism by which downregulation of BK channels, either by the b4 subunit or paxilline, could reduce firing frequency is by increased calcium influx leading to recruitment of a calcium-activated afterhyperpolarizing current between action potential spikes. Thus, it is reasonable to assume that these effects are mediated through activation of calcium-activated, SK-type potassium channels that have an established role in regulating interspike intervals and spike frequency adaptation8–11. We addressed this hypothesis in two ways. First, we recorded action potentials with intracellular solutions containing the fast calcium chelator BAPTA (10 mM) to prevent the accumulation of calcium during action potential firing. BAPTA shortened the interspike interval (Fig. 5e) and converted the low firing frequency typical of wild-type mice to frequencies very similar to those in knockout mice (Fig. 5f). This indicates that the difference in firing frequency is largely accounted for by increased calcium in wild-type cells. Moreover, the fact that the firing frequency of b4/ cells and b4/ cells with BAPTA were very similar indicates that gain-of-function by BK channels in b4/ cells functionally eliminates all contribution of calcium influx to control of firing rates. Second, we used the specific SK channel blocker apamin (100 nM)6,13 to determine if increased SK channel activity contributes to lower firing rates in wild-type cells. Action potential firing rate of wildtype cells was markedly increased by SK channel blockade (Fig. 6a,b). We found that SK channel block converted the AHP decay rate (Fig. 6e) and the interspike interval in wild-type cells (Fig. 6f) to b4/ levels. We also observed a similar effect with the organic SK channel blocker UCL1684 (Supplementary Fig. 2). As predicted based on our hypothesis that gain of function of BK channels in b4/ cells reduces recruitment of SK channels between spikes, we found that apamin had no significant effect on firing rates, AHP tau or interspike interval in b4/ cells (Fig. 6b,e,f). In addition, application of apamin

VOLUME 8

[

NUMBER 12

[



ARTICLES to a b4/ cell had no effect on post-train AHP (Supplementary Fig. 2). Notably, apamin had a significant effect on action potential width in b4/ cells (P o 0.05; Fig. 6c) but no significant effect on fAHP amplitude, decay or interspike interval (Figs. 6d–f). In b4/ cells, the selective effect of apamin on the spike and not on interspike properties or on firing frequency (Fig. 6b) indicates that gain-of-function BK channels do not completely restrict access of calcium to SK channels during the spike. Rather, BK channels in b4/ cells prevent sustained calcium between pulses and thereby prevent sustained SK channel activation leading to spike frequency adaptation. In summary, these results indicate that a major effect of the b4 subunit is to indirectly recruit SK-type potassium channels between action potential spikes, which mediate spike frequency adaptation and reduced firing rates. DISCUSSION Here we have shown that the b4 subunit represses the contribution of the BK channel to action potential repolarization. The conversion from a slow-gated type II wild-type BK channel to a fast-gated, type I–like BK channel in b4/ cells has profound effects, reducing spike frequency adaptation and increasing the firing rate. The concept that fast-gated BK channels would lead to increased excitability may seem counterintuitive for the conventional role that potassium channels play in reducing membrane excitability. However, some fast-gated potassium channels serve a dual role, both repolarizing the action potential and, by quickly resetting the membrane potential, allowing a subsequent spike with a short interspike interval. An example is provided by Kv3.2 potassium channels. These channels are gated at relatively depolarized voltages, activating and deactivating quickly during the action potential and thereby allowing high tonic firing36. This is qualitatively similar to the fast-gated BK channels in b4/ cells. In fact, knocking out the gene encoding Kv3.2 channels in fast-spiking interneurons leads to reduced firing rates37. Supporting these conclusions, examples exist in which blockade of BK channel activity either does not increase excitability or protects against seizure activity. For instance, injection of a Kv channel blocker into ventricles of mice initiates epileptiform seizure activity, whereas injection of the BK channel blocker paxilline does not produce any convulsant effects38. It has been shown that BK channels such as the iberiotoxin-sensitive type I channels in cortical neurons are required for epileptic bursting behavior in cortical neurons, and blockade with iberiotoxin inhibits bursting behavior39. In contrast, block of the SKtype channels does not protect against bursting behavior39. More recently, a gain-of-function mutation of the a subunit of BK channels in humans was found to be linked to a syndrome of generalized epilepsy with paroxysmal movement disorders40. The mutation increases BK channel activation rates and channel open probability. These effects are consistent with our hypothesis that gain-of-function of BK channels are proepileptic. In the a subunit gain-of-function mutation, the broader, cortical expression pattern of the a subunit may account for the more generalized neurological phenotype. In the case of the b4 subunit, which is enriched in the hippocampus, gain of BK channel function is mostly restricted to temporal brain regions, suggesting that pharmacology directed against this subunit might be preferred for selective targeting of temporal lobe dysfunction. The mechanism by which wild-type BK b4 subunit channels recruit an afterhyperpolarizing current and reduce firing frequency requires further investigation. Our hypothesis is that the broader action potentials in the wild-type cells allow a greater calcium influx during each action potential. Consequently, calcium can accumulate to levels that sustain SK-type channels during high-frequency spiking, thereby prolonging the interspike interval and increasing spike frequency


[

NUMBER 12

[

DECEMBER 2005

adaptation. A larger AHP associated with spike broadening has been observed in CA1 hippocampal cells41 and in the early component of an action potential train in the lateral amygdala14. The dentate gyrus is regarded as an important low-pass filter that prevents high-frequency synchronized input from the neocortex into the hippocampus33,34,42. Low-frequency firing is not effective in transmission from dentate granule cells to their postsynaptic target cells in the CA3 pyramidal region, but higher-frequency firing effectively switches the granule cell input into a ‘detonator synapse’ that strongly synchronizes presynaptic activity with action potential firing in the postsynaptic target43. Moreover, firing frequency of the dentate gyrus also dictates whether postsynaptic CA3 cells are dominated by a disynaptic inhibitory circuit or by a monosynaptic excitatory circuit44. At frequencies above 20 Hz, the excitatory transmission becomes dominant44. During spatial stimuli, granule cells can fire at frequencies higher than 10 Hz45. Our data demonstrate that the loss of b4 subunits extends the dynamic range of dentate gyrus cells into a higherfrequency firing mode, thereby promoting excitatory synaptic transmission and abnormal synchronization leading to seizures. Our findings highlight an important role of intrinsic firing properties of dentate gyrus granule cells in epileptogenesis. However, inhibitory synapses are known to have a role in regulating excitability of these cells1–4, and reduced inhibition of dentate granule cells is observed in rat models of temporal lobe epilepsy46. Indeed, increased firing by dentate granule cells is not mutually exclusive with a role for local inhibitory interneurons in augmenting downstream effects of the b4/ defect in granule cells. For instance, mossy fibers also activate an inhibitory feedback loop via hilar interneurons. Prolonged depolarization of these cells by higher-frequency firing could lead to activity-dependent decreases in GABA release47 and disinhibition of granule cells, thereby further promoting abnormal activation of hippocampal circuits2. METHODS Gene-targeted mice. Genomic DNA encompassing the first exon of the gene encoding b4 was cloned from a 129svj l Fix genomic library (Stratagene). The left arm included a 1.7-kb fragment upstream of exon 1, including its first two amino acids. This was subcloned upstream of the Clontech eGFP-N1 BamHI/SspI fragment. This placed the upstream regulatory sequences and the first two amino acids of the b4 protein in translational frame with eGFP. The left arm–eGFP fusion was cloned into the ppNT NotI/SalI site (SalI site blunted). The right arm included a 7.2 kb BglII/KpnI fragment beginning approximately 120 nucleotides into the first intron. This was cloned into the BamHI/KpnI site of ppNT. The knockout vector was electroporated into 129svj embryonic stem (ES) cells, and recombinant ES clones were identified with positive and negative selection. Gene-targeted ES clones were confirmed by genomic Southern blot across regions encompassing the left arm (Fig. 1) and the right arm (data not shown). For brain slice electrophysiology, knockout mice were used that were congenic five generations with C57BL/6J mice (Jackson Laboratory). C57BL/6J mice were used as wild-type controls. EEG recordings were performed on b4/ mice on a mixed 129svj/C57BL/6 background. All animal protocols complied with US National Institutes of Health guidelines and were reviewed and approved by the University of Texas Health Science Center and Stanford University Animal Care and Use Committees. Immunohistology. Rabbit anti-GFP antibody (ab6556) was from Abcam. Frozen brain sections (40 mm) were stained according to previous methods48 using a 1:20,000 dilution of anti-GFP antibodies and 1:500 dilution of biotin, goat anti-rabbit secondary antibody. For paraffin sections (10 mm, Supplementary Fig. 1), we used 1:1,000 dilution of anti-GFP antibody. Video EEG monitoring. Adult mice were implanted with cortical surface electrodes under avertin anesthesia (1.25% tri-bromoethanol-amyl alcohol) by intraperitoneal injection (0.02 ml g1). Silver wire electrodes (0.005 inches in

1757

ARTICLES


diameter) soldered to a microminiature connector were implanted into the subdural space over the left and right cortical hemispheres. After several days of recovery, behavioral and EEG activity was recorded daily from freely moving mice for 7–10 d using a digital video electroencephalograph system (Stellate Systems). Electrophysiology. Brain slice (400 mm) protocols using mice (5–7 weeks) were as described previously49. Whole-cell recordings were conducted with internal solutions of 120 mM potassium gluconate, 20 mM KCl, 2 mM MgCl2, 10 mM HEPES, 2 mM ATP, 0.25 mM GTP and 0.1 mM EGTA (pH 7.4; free calcium was calculated as approximately 50 nM by MaxC software (Chris Patton, Stanford University)). Granule cells were identified using infrared microscopy and a charge-coupled device (CCD) camera. Cells were clamped in whole-cell mode using a HEKA EPC10 amplifier and Pulse software. Recordings were considered acceptable if cells had a resting membrane potential of 70 mV and input resistance of 350 MO or larger. Series resistance was compensated using the amplifier hardware. If capacitance or series resistance changed more than 20% during the recordings, the cell was excluded from further analysis. Data was collected with a hardware filter of 3 kHZ. Current-clamp records were made from a 80 mV holding current and incremental positive current injections for 900 ms durations. Current injections seemed to affect the difference in spiking frequency between b4/ and wild-type cells (see Results) but did not seem to affect the relative depolarization. For example, the mean interspike voltage during a 300-pA current injection (tenth action potential) was 17.1 ± 1.4 for wild-type dentate granule cells and was 15.8 ± 2.0 mV for b4/ dentate granule cells (P ¼ 0.6 in unpaired Student’s t-test). Spike width was measured at half the spike amplitude. The fAHP size was measured as the difference between the spike threshold and voltage minimum after the action potential peak. AHP decay was measured from the peak fAHP to the following 5–10 ms using an exponential function. Interspike interval was the time between action potentials peaks. Input resistance was measured from a hyperpolarizing current injection of 20 pA from a 80 mV holding current. Action potential height was measured during a 100-pA current injection while holding at 80 mV. For single-channel recording, the patch was excised from the whole-cell configuration to generate an outside/out patch. The pipette contained the internal solution described above and HEDTA (5 mM) with added calcium to result in 18.2 mM free calcium (determined by calibration with a calciumsensitive electrode). The patch was withdrawn from the slice to allow constant flow of 2 ml min1 of PSS containing normal solution, 100 nM iberiotoxin or 5 mM paxilline. To confirm the presence of channels after drug block, the patches were washed free of toxin until channel opening was recovered. For iberiotoxin-resistant (wild-type) channels, iberiotoxin perfusion was followed with paxilline to ensure that channels were indeed BK channels. Open probability and dwell times were measured using TAC and TACFIT software (Bruxton). Data were filtered at 3 kHz and were acquired continuously at 30-ms intervals for 5 s at each voltage. A 50% threshold was used to detect events, and each event was visually inspected before acceptance. Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS We thank S. Wiler for technical assistance and B.S. Rothberg for critical reading of the manuscript. This work was supported by US National Institutes of Health grants NS29709 and HD24064 to J.L.N., American Heart Association grant 02250724 to Q.H.C. and a University of Texas Health Science Center Executive Research Committee grant to R.B. R.W.A. is an investigator with the Howard Hughes Medical Institute. COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/ 1. Sloviter, R.S. et al. ‘‘Dormant basket cell’’ hypothesis revisited: relative vulnerabilities of dentate gyrus mossy cells and inhibitory interneurons after hippocampal status epilepticus in the rat. J. Comp. Neurol. 459, 44–76 (2003). 2. Ratzliff, A.H., Santhakumar, V., Howard, A. & Soltesz, I. Mossy cells in epilepsy: rigor mortis or vigor mortis? Trends Neurosci. 25, 140–144 (2002).

1758

3. Coulter, D.A. Mossy fiber zinc and temporal lobe epilepsy: pathological association with altered ‘‘epileptic’’ gamma-aminobutyric acid A receptors in dentate granule cells. Epilepsia 41 (Suppl.), S96–S99 (2000). 4. Buhl, E.H., Otis, T.S. & Mody, I. Zinc-induced collapse of augmented inhibition by GABA in a temporal lobe epilepsy model. Science 271, 369–373 (1996). 5. Garcia, M.L., Gao, Y., McManus, O.B. & Kaczorowski, G.J. Potassium channels: from scorpion venoms to high-resolution structure. Toxicon 39, 739–748 (2001). 6. Stocker, M. Ca(2+)-activated K(+) channels: molecular determinants and function of the SK family. Nat. Rev. Neurosci. 5, 758–770 (2004). 7. Stocker, M., Hirzel, K., D’Hoedt, D. & Pedarzani, P. Matching molecules to function: neuronal Ca2+-activated K+ channels and afterhyperpolarizations. Toxicon 43, 933–949 (2004). 8. Greffrath, W. et al. Contribution of Ca2+-activated K+ channels to hyperpolarizing afterpotentials and discharge pattern in rat supraoptic neurones. J. Neuroendocrinol. 16, 577–588 (2004). 9. Engel, J., Schultens, H.A. & Schild, D. Small conductance potassium channels cause an activity-dependent spike frequency adaptation and make the transfer function of neurons logarithmic. Biophys. J. 76, 1310–1319 (1999). 10. Teshima, K., Kim, S.H. & Allen, C.N. Characterization of an apamin-sensitive potassium current in suprachiasmatic nucleus neurons. Neuroscience 120, 65–73 (2003). 11. Yen, J.C., Chan, J.Y. & Chan, S.H. Involvement of apamin-sensitive SK channels in spike frequency adaptation of neurons in nucleus tractus solitarii of the rat. J. Biomed. Sci. 6, 418–424 (1999). 12. Bond, C.T. et al. Small conductance Ca2+-activated K+ channel knock-out mice reveal the identity of calcium-dependent afterhyperpolarization currents. J. Neurosci. 24, 5301–5306 (2004). 13. Blank, T., Nijholt, I., Kye, M.J. & Spiess, J. Small conductance Ca2+-activated K+ channels as targets of CNS drug development. Curr. Drug Targets CNS Neurol. Disord. 3, 161–167 (2004). 14. Faber, E.S. & Sah, P. Physiological role of calcium-activated potassium currents in the rat lateral amygdala. J. Neurosci. 22, 1618–1628 (2002). 15. Cui, J., Cox, D.H. & Aldrich, R.W. Intrinsic voltage dependence and Ca2+ regulation of mslo large conductance Ca-activated K+ channels. J. Gen. Physiol. 109, 647–673 (1997). 16. Sanchez, M. & McManus, O.B. Paxilline inhibition of the alpha-subunit of the highconductance calcium-activated potassium channel. Neuropharmacology 35, 963–968 (1996). 17. Hu, H. et al. Presynaptic Ca2+-activated K+ channels in glutamatergic hippocampal terminals and their role in spike repolarization and regulation of transmitter release. J. Neurosci. 21, 9585–9597 (2001). 18. Swensen, A.M. & Bean, B.P. Ionic mechanisms of burst firing in dissociated Purkinje neurons. J. Neurosci. 23, 9650–9663 (2003). 19. Shao, L.R., Halvorsrud, R., Borg-Graham, L. & Storm, J.F. The role of BK-type Ca2+dependent K+ channels in spike broadening during repetitive firing in rat hippocampal pyramidal cells. J. Physiol. (Lond.) 521, 135–146 (1999). 20. Faber, E.S. & Sah, P. Ca2+-activated K+ (BK) channel inactivation contributes to spike broadening during repetitive firing in the rat lateral amygdala. J. Physiol. (Lond.) 552, 483–497 (2003). 21. Reinhart, P.H., Chung, S. & Levitan, I.B. A family of calcium-dependent potassium channels from rat brain. Neuron 2, 1031–1041 (1989). 22. Wang, G. & Lemos, J.R. Tetrandrine blocks a slow, large-conductance, Ca(2+)-activated potassium channel besides inhibiting a non-inactivating Ca2+ current in isolated nerve terminals of the rat neurohypophysis. Pflugers Arch. 421, 558–565 (1992). 23. Meera, P., Wallner, M. & Toro, L. A neuronal beta subunit (KCNMB4) makes the large conductance, voltage- and Ca2+-activated K+ channel resistant to charybdotoxin and iberiotoxin. Proc. Natl. Acad. Sci. USA 97, 5562–5567 (2000). 24. Lippiat, J.D., Standen, N.B., Harrow, I.D., Phillips, S.C. & Davies, N.W. Properties of BK(Ca) channels formed by bicistronic expression of hSloalpha and beta1–4 subunits in HEK293 cells. J. Membr. Biol. 192, 141–148 (2003). 25. Behrens, R. et al. hKCNMB3 and hKCNMB4, cloning and characterization of two members of the large-conductance calcium-activated potassium channel beta subunit family. FEBS Lett. 474, 99–106 (2000). 26. Weiger, T.M. et al. A novel nervous system beta subunit that downregulates human large conductance calcium-dependent potassium channels. J. Neurosci. 20, 3563–3570 (2000). 27. Brenner, R., Jegla, T.J., Wickenden, A., Liu, Y. & Aldrich, R.W. Cloning and functional characterization of novel large conductance calcium-activated potassium channel beta subunits, hKCNMB3 and hKCNMB4. J. Biol. Chem. 275, 6453–6461 (2000). 28. Ha, T.S., Heo, M.S. & Park, C.S. Functional effects of auxiliary beta4-subunit on rat large-conductance Ca(2+)-activated K(+) channel. Biophys. J. 86, 2871–2882 (2004). 29. Sah, P. Ca(2+)-activated K+ currents in neurones: types, physiological roles and modulation. Trends Neurosci. 19, 150–154 (1996). 30. Williamson, P.D. & Engel, J. in Epilepsy: a Comprehensive Textbook (eds. Engel, J. & Pedley, T.A.) 557–566 (Lippincott-Raven, Philadelphia, 1997). 31. Tian, L., Knaus, H.G. & Shipston, M.J. Glucocorticoid regulation of calcium-activated potassium channels mediated by serine/threonine protein phosphatase. J. Biol. Chem. 273, 13531–13536 (1998). 32. Bielefeldt, K. & Jackson, M.B. A calcium-activated potassium channel causes frequency-dependent action-potential failures in a mammalian nerve terminal. J. Neurophysiol. 70, 284–298 (1993). 33. Nadler, J.V. The recurrent mossy fiber pathway of the epileptic brain. Neurochem. Res. 28, 1649–1658 (2003).

VOLUME 8

[

NUMBER 12

[



ARTICLES 34. Heinemann, U. et al. The dentate gyrus as a regulated gate for the propagation of epileptiform activity. Epilepsy Res. Suppl. 7, 273–280 (1992). 35. Sah, P. & Faber, E.S. Channels underlying neuronal calcium-activated potassium currents. Prog. Neurobiol. 66, 345–353 (2002). 36. Rudy, B. & McBain, C.J. Kv3 channels: voltage-gated K+ channels designed for highfrequency repetitive firing. Trends Neurosci. 24, 517–526 (2001). 37. Lau, D. et al. Impaired fast-spiking, suppressed cortical inhibition, and increased susceptibility to seizures in mice lacking Kv3.2 K+ channel proteins. J. Neurosci. 20, 9071–9085 (2000). 38. Juhng, K.N. et al. Induction of seizures by the potent K+ channel-blocking scorpion venom peptide toxins tityustoxin-K(alpha) and pandinustoxin-K(alpha). Epilepsy Res. 34, 177–186 (1999). 39. Jin, W., Sugaya, A., Tsuda, T., Ohguchi, H. & Sugaya, E. Relationship between large conductance calcium-activated potassium channel and bursting activity. Brain Res. 860, 21–28 (2000). 40. Du, W. et al. Calcium-sensitive potassium channelopathy in human epilepsy and paroxysmal movement disorder. Nat. Genet. 37, 733–738 (2005). 41. Kamal, A., Artola, A., Biessels, G.J., Gispen, W.H. & Ramakers, G.M. Increased spike broadening and slow afterhyperpolarization in CA1 pyramidal cells of streptozotocininduced diabetic rats. Neuroscience 118, 577–583 (2003).


[

NUMBER 12

[

DECEMBER 2005

42. Mody, I., Kohr, G., Otis, T.S. & Staley, K.J. The electrophysiology of dentate gyrus granule cells in whole-cell recordings. Epilepsy Res. Suppl. 7, 159–168 (1992). 43. Henze, D.A., Wittner, L. & Buzsaki, G. Single granule cells reliably discharge targets in the hippocampal CA3 network in vivo. Nat. Neurosci. 5, 790–795 (2002). 44. Mori, M., Abegg, M.H., Gahwiler, B.H. & Gerber, U. A frequency-dependent switch from inhibition to excitation in a hippocampal unitary circuit. Nature 431, 453–456 (2004). 45. Jung, M.W. & McNaughton, B.L. Spatial selectivity of unit activity in the hippocampal granular layer. Hippocampus 3, 165–182 (1993). 46. Kobayashi, M. & Buckmaster, P.S. Reduced inhibition of dentate granule cells in a model of temporal lobe epilepsy. J. Neurosci. 23, 2440–2452 (2003). 47. Mott, D.D., Xie, C.W., Wilson, W.A., Swartzwelder, H.S. & Lewis, D.V. GABAB autoreceptors mediate activity-dependent disinhibition and enhance signal transmission in the dentate gyrus. J. Neurophysiol. 69, 674–691 (1993). 48. Strassle, B.W., Menegola, M., Rhodes, K.J. & Trimmer, J.S. Light and electron microscopic analysis of KChIP and Kv4 localization in rat cerebellar granule cells. J. Comp. Neurol. 484, 144–155 (2005). 49. Moyer, J.R., Jr. & Brown, T.H. Methods for whole-cell recording from visually preselected neurons of perirhinal cortex in brain slices from young and aging rats. J. Neurosci. Methods 86, 35–54 (1998).

1759


ARTICLES

Synaptic background activity controls spike transfer from thalamus to cortex Jakob Wolfart1,3,4, Damien Debay1,4, Gwendal Le Masson2, Alain Destexhe1 & Thierry Bal1 Characterizing the responsiveness of thalamic neurons is crucial to understanding the flow of sensory information. Typically, thalamocortical neurons possess two distinct firing modes. At depolarized membrane potentials, thalamic cells fire single action potentials and faithfully relay synaptic inputs to the cortex. At hyperpolarized potentials, the activation of T-type calcium channels promotes burst firing, and the transfer is less accurate. Our results suggest that this duality no longer holds if synaptic background activity is taken into account. By injecting stochastic conductances into guinea-pig thalamocortical neurons in slices, we show that the transfer function of these neurons is strongly influenced by conductance noise. The combination of synaptic noise with intrinsic properties gives a global responsiveness that is more linear, mixing single-spike and burst responses at all membrane potentials. Because in thalamic neurons, background synaptic input originates mainly from cortex, these results support a determinant role of corticothalamic feedback during sensory information processing.

Thalamocortical neurons in the dorsolateral geniculate nucleus relay visual input from retinal ganglion cells to the cortex, from which they receive massive feedback1,2. The function and mechanisms of corticothalamic feedback are still a matter of discussion, but it is generally agreed that it has a strong influence on the transfer of sensory information by thalamocortical cells3–7. According to the classical view, thalamocortical cells function in two intrinsically generated firing modes. At depolarized membrane potentials, these neurons fire single action potentials, faithfully transmitting synaptic inputs. This relay or ‘single-spike’ mode is mainly found during the awake state6,8. At hyperpolarized potentials, activation of low-threshold calcium (T-type) channels triggers high-frequency bursts of action potentials9,10. The burst mode is mostly found during slow wave sleep and epileptic absence seizures, when thalamocortical cells participate in the synchronous bursting of the thalamic network, functionally uncoupling the cortex from visual input8,11. In vivo, neurons generally experience a noisy high-conductance state that is likely to interact with their built-in integrative properties12,13. Recent studies have shown that in vivo-like synaptic noise changes specific aspects of signal integration in cortical neurons14–18. Thalamocortical cells recorded in vivo are also in a high-conductance state, in particular during corticothalamic barrages12,19. A given thalamocortical cell receives between 4,000 and 8,000 synapses20,21, of which B30% have direct cortical origin1,2,21,22. In addition, they are innervated by intrathalamic inhibitory neurons (interneurons and reticular thalamic neurons), which also receive direct cortical inputs and account for B30% of synapses onto thalamocortical cells1,2,21–23. Thus, B60% of

synapses of thalamocortical neurons are directly or indirectly related to the activity of corticothalamic axons (in addition to that of other afferents), but it remains uncertain whether, overall, this feedback is excitatory or inhibitory for thalamocortical cells4. It has been proposed that corticothalamic feedback could switch thalamocortical neurons between burst and single-spike modes7,22,24,25, but it is also unclear whether corticothalamic input increases or decreases thalamocortical cell bursting7. We hypothesize that corticothalamic feedback exerts its function not (only) by exciting or inhibiting thalamocortical cells, but by using a separate ‘channel’ of modulatory information16: the variance of background synaptic input. To explore how synaptic noise affects thalamic neurons, we recorded from thalamocortical cells in brain slices using dynamic-clamp injection of stochastic background conductances. We found that background conductance noise significantly changed the ‘burstiness’ (percentage of burst responses per spikeevoking input) and the input-output transfer function of thalamic relay neurons. RESULTS Conductance noise and input simulation We recorded from thalamocortical neurons in dorsolateral geniculate nucleus (LGNd) slices of guinea pigs, using intracellular electrodes and the dynamic-clamp technique26. We analyzed 52 neurons from 36 animals. These neurons had a resting potential ± s.e.m. of –63 ± 0.4 mV and an input resistance of 61 ± 4 MOhm and showed rebound burst discharges accompanied by low-threshold calcium spikes (LTS) upon repolarization after hyperpolarization (Fig. 1a, inset in ‘Quiescent’ graph). These properties8,9,23, as well as morphological reconstructions

1Unite ´ de Neurosciences Integratives et Computationnelles, Centre National de la Recherche Scientifique, 91198 Gif-sur-Yvette, France. 2Institut National de la Sante´ et de la Recherche Me´dicale (INSERM) 358, Universite´ Victor Segalen Bordeaux 2, Bordeaux, France. 3Present address: Neurozentrum, Department of Neurosurgery, University Hospital Freiburg, Breisacher Strasse 64, 79106 Freiburg, Germany. 4These authors contributed equally to this work. Correspondence should be addressed to T.B. ([email protected]) or J.W. ([email protected]).

Received 14 July; accepted 30 September; published online 30 October 2005; doi:10.1038/nn1591

1760

VOLUME 8

[

NUMBER 12

[


ARTICLES

a

Quiescent

Noise

Static

Vm

10 mV –60 mV

ginput

20 nS 20 nS

g iNoise

0 nS

g eNoise

0 nS

c

0.0

***

Noise Static

0.2

0.8 Weak noise

–1

Spike probability

0.4

0.1

Noise

0.6

Quiescent

0.8

1.0 Strong noise

–1

Log slope (nS )

1.0

Slope (nS )

1.0

Static

b Spike probability


100 ms

0.6 0.4

0.08

0.04 0.02

0.2

***

0.06

1

Quiescent 0.0

3 s.d. (mV)

5

0.0 0

10

20

30

40

Input (nS)

0

10

20

30

40

50

Input (nS)

Figure 1 The influence of conductance (g) noise changes the transfer function of thalamocortical cells recorded in vitro. (a) Voltage during injection of input g alone (ginput, Quiescent) and with additional inhibitory plus excitatory background g strengths (giNoise, geNoise) that were either non-fluctuating (Static) or stochastically fluctuating (Noise). Combined noise g reduced the input resistance to B50% (insets in Quiescent, Static). (b) Probabilities of input g strengths to evoke Z1 spike, fitted to sigmoid functions. Noise, but not Static, induced a gain (slope) reduction of the response curve (inset). (c) Decreasing the variance of noise g values (Strong, Weak) increased the input-output slope. The response gain was correlated with the noise-induced voltage variance (s.d.). Error bars, s.e.m.

of biocytin-filled cells, unambiguously identify the neurons as thalamocortical relay neurons. The standard protocol used to assess the integration properties of thalamocortical cells consisted of a recording at resting potential, where excitatory signal input conductances were injected at 5 Hz without additional background conductances (Fig. 1a, Quiescent). Subsequently, we added excitatory and inhibitory background conductances without fluctuations (Fig. 1a, ‘Static’). The total background conductance was adjusted such that the cells’ input resistance was approximately 50% of its initial value (Fig. 1a, compare insets of Quiescent and Static traces), similar to the ‘shunting’ effect observed in thalamocortical cells in vivo during the activation of corticothalamic projections19. We then injected the same background conductances with stochastic fluctuations (Fig. 1a, ‘Noise’). To separate the effect of noise from simple depolarization or hyperpolarization, we compensated, when necessary, for the effect of background conductances by injecting a small DC current such that the mean potential was similar under these conditions (Fig. 1a). The variance of the membrane potential was low in the quiescent (0.71 ± 0.03 mV) and static conditions, and was increased with noise (3.65 ± 0.13 mV, n ¼ 24) to an amplitude consistent with voltage fluctuations in vivo19,27. Noise affects the gain of thalamocortical neurons We assessed whether inputs were transmitted and how this transfer was affected by synaptic background noise by evaluating the probability that a given input magnitude would evoke at least one action potential within a 20-ms delay17 for the quiescent, static and noise conditions (Fig. 1b). The slope (gain) of the input-output relation was determined by fitting a sigmoid function to spike probability values and extracting its slope at the 0.5 probability. Although a step-like transfer function characterized the quiescent and static conditions, under the influence of noise the response probability was linearized, adopting intermediate values between 0 and 1 over a


[

NUMBER 12

[

DECEMBER 2005

larger input range (Fig. 1b). The gain was not significantly different between the quiescent and static conductance injections, but was strongly reduced with noise (inset in Fig. 1b; quiescent: 0.413 ± 0.029 nS1, n ¼ 41; noise: 0.046 ± 0.002 nS1, n ¼ 27; static: 0.409 ± 0.029 nS1, n ¼ 24; quiescent versus noise: P o 0.001; quiescent versus static: P ¼ 0.72). We obtained similar results using a measure of the total spike output (see Methods). Thus, consistent with previous results from the cortex14,16,17, the variance of background conductance reduced the input-output gain of thalamocortical cells and increased the cells’ sensitivity to small inputs. Because the effect of corticothalamic feedback is expected to vary considerably with the state of the animal and the signals that are processed7,22,28, we explored different variances of the fluctuating conductances. In the ‘strong noise’ condition (our default for the results described above), the conductance variances for excitatory and inhibitory noise were 3 nS and 12 nS, respectively. A decrease in noise conductance variances (to 1 and 4 nS, respectively; ‘weak noise’ voltage variance: 2.61 ± 0.14 mV, n ¼ 5) increased the input-output gain (Fig. 1c, 0.070 ± 0.007 nS1, n ¼ 5; versus ‘strong noise’: 0.046 ± 0.002 nS1, n ¼ 27; P o 0.01). The gain in the various noise conditions correlated with the actual degree of voltage variance induced by the noise (Fig. 1c, inset, n ¼ 33, r ¼ 0.63, P o 0.01). These results show that background synaptic activity is able to modulate the response curve of thalamocortical neurons in a multiplicative manner. Gain depends on membrane potential and input frequency The responsiveness of thalamic neurons at depolarized membrane potentials is entirely different from that at hyperpolarized potentials8,9,29. At hyperpolarization, incoming excitatory postsynaptic potentials (EPSPs) can be amplified by T-type channels producing LTS burst discharges29. We found that, unexpectedly and unlike what is observed for cortical cells17, the gain of thalamocortical cells at hyperpolarized potentials was markedly lower than that at depolarized

1761

ARTICLES

a

b

Quiescent, 5 Hz

Hyperpolarized, quiescent

1.0 Dep

5 Hz

Hyp

Rest

1 Hz

1.0 Dep Rest Hyp –1

Slope (nS )

0.6 0.4 0.2

***

0.5

0.0

0.0

c

0

10

20

30 Input (nS)

40

50

60

Quiescent

1.0

1.0 0.8

Rest

20 Hz

Dep

Spike probability

Hyp Slope (nS–1)


Spike probability

0.8

**

0.6

*

0.4

1 Hz

5 Hz

0.6 0.4 0.2

0.2 0.0

0.8

20 Hz 0.0 1 5 10

20

1 5 10 20 Input frequency (Hz)

1 5 10

20

0

10

20

30

40

50

60

70

Input (nS)

Figure 2 Voltage and frequency dependence of the response gain of thalamocortical cells, without noise. (a) Response curves during quiescent mode, 5 Hz stimulation, at depolarized (Dep), resting (Rest) and hyperpolarized potentials (Hyp). The response gain (slope) was equal at Dep and Rest but reduced at Hyp (inset). (b) Different input frequencies (5 Hz, 1 Hz, 20 Hz) in the Hyp, quiescent condition. Response curves showed lower gain only in the 5 Hz condition. Scale bars in b: horizontal below traces: 5 Hz, 0.1 s; 1 Hz, 0.5 s; 20 Hz, 20 ms; between traces: –70 mV; vertical (upper, lower): 10 mV, 20 nS. (c) Summary of experiments as in a and b. Response gains were all-or-none with all input frequencies at Rest and Dep and with 1 Hz at Hyp, but they were reduced at Hyp with 5 Hz and 10 Hz inputs. Error bars, s.e.m. * P o 0.05, **P o 0.01, ***P o 0.001.

and resting potentials (Fig. 2a, hyperpolarized: 0.146 ± 0.028 nS1, n ¼ 12; resting: 0.413 ± 0.029 nS1, n ¼ 41; P o 0.001). What could cause the voltage dependence of gain? T-type channel activation at hyperpolarization is not expected to reduce the gain but rather to increase the all-or-none character of thalamocortical cell responses22. Yet, because de-inactivation time constants of T-type channels are in the range of hundreds of milliseconds9,30, varying degrees of T-type channel recruitment are to be expected at a 5-Hz input rate with randomized EPSP amplitudes. This could account for the gain reduction. Further analysis showed that much of the LTS variability at hyperpolarization could indeed be attributed to the recent LTS activation history (Supplementary Fig. 1 online). These results suggest that the voltage dependence of gain in the quiescent neuron is itself dependent on the input frequency. We tested four input frequencies (1, 5, 10 and 20 Hz) at all voltage conditions. At resting and depolarized potentials, there was no frequency dependence of gain (ANOVA, P ¼ 0.29). However, at hyperpolarization, a frequency dependence of gain was clearly visible (Fig. 2b,c, ANOVA, P o 0.001). This is in agreement with the idea that the voltage and frequency dependence of gain were due to the T-type channel gating behavior: the low gain at hyperpolarization reverted to an all-or-none gain when the input frequency was lowered to 1 Hz such that T-type channels could recover from inactivation between stimuli30 (Fig. 2b,c; 1 Hz: 0.468 ± 0.072 nS1, n ¼ 4; 5 Hz: 0.146 ± 0.028 nS1, n ¼ 12; 1 Hz versus 5 Hz: P ¼ 0.008). On the other hand, increasing the input frequency to 10 Hz and 20 Hz gradually increased the gain at hyperpolarization (Fig. 2b,c; 20 Hz: 0.270 ± 0.033 nS1, n ¼ 4; 20 Hz versus 5 Hz: P ¼ 0.039). This effect could be explained by a cumulative inactivation of the T-type channels at higher frequencies: the inability of T-type channels to follow high frequencies endows thalamocortical cells with low-pass filter properties when they are hyperpolarized29. Thus, our results suggest the following scenario: at

1762

hyperpolarization and input frequencies around 5 Hz, T-type channels favor low gain transfer functions, whereas at more depolarized potentials, thalamocortical cells are expected to show all-or-none gain regardless of the input frequency. Noise renders gain independent of voltage and frequency To test how conductance noise may interfere with the voltage and frequency dependence of gain, we performed the characterization described earlier (Fig. 2) in the presence of noise (Fig. 3). The injection of noise reduced the gain at all voltage conditions (Fig. 3a). Unlike in the quiescent condition, the gain was very similar at all membrane potentials, although still slightly reduced at hyperpolarization with a 5 Hz input (inset in Fig. 3a; see also inset in Fig. 3c: ANOVA voltage dependence: P ¼ 0.007; across all frequencies, P ¼ 0.076; with frequencies pooled, P ¼ 0.145). Thus, the voltage dependence of gain was strongly reduced with noise. We next asked if the frequency dependence of gain that was found in the quiescent conditions would endure in the noise condition. In accordance with the hypothesis that voltage and frequency dependence resulted from the same mechanism, noise nearly abolished the frequency dependence of gain. We compared three input frequencies at hyperpolarized potentials in the presence of noise (Fig. 3b; compare with Fig. 2b). The marked distinction in gain between 1 Hz and 5 Hz and between 5 Hz and 20 Hz (Fig. 2c) was much reduced with noise (Fig. 3c, Hyp). At all other frequencies and membrane potentials, the gain was equally low when noise was present (Fig. 3c, ANOVA across all conditions; frequency dependence, P ¼ 0.049; with potentials pooled, P ¼ 0.087). Further analysis of the underlying events showed that noise increased the variability of subthreshold responses to an overall high level, overwhelming the variability due to LTS, thereby reducing the response gain of thalamocortical cells to an overall low level that was equal across different membrane potentials and input frequencies

VOLUME 8

[

NUMBER 12

[


ARTICLES Noise, 5 Hz (

), quiescent (—)

b

1.0

5 Hz Dep.

Rest. Slope (nS–1)

Hyp.

0.4 0.2

*

0.1

0.0

0

10

20

30 Input (nS)

40

c

50

0.6 Slope (nS–1)

Noise 0.20 0.15

5 Hz

1.0 Quiescent

20 Hz

0.2 Noise

Rest

Hyp

60

0.4

0.0

Hyp

Rest

Dep

Dep 0.05 0.00

1 Hz

0.2 Dep Rest Hyp

0.6

0.0

Slope (nS–1)


Spike probability

0.8

0.10

Hyperpolarized, noise

0.8 Spike probability

a

0.6

10

20

1 5 10 20 Input frequency (Hz)

1 5

10

20

1 Hz

0.4

20 Hz

0.2 0.0

1 5

5 Hz

0

10

20

30 40 Input (nS)

50

60

70

Figure 3 Influence of noise on voltage and frequency dependence of gain. (a) Thick gray curves: response probabilities during noise mode, 5 Hz stimulation, at depolarized (Dep), resting (Rest) and hyperpolarized potentials (Hyp). Thin black lines: quiescent curves from Figure 2a (same cell). The response gain (slope) during noise was almost equally low at all potentials (inset). (b) Different input frequencies (5 Hz, 1 Hz, 20 Hz) in the Hyp, noise condition. Response curves showed almost equally low gain at all input frequencies. (c) Summary of experiments as in a and b. Response gains were low at all input frequencies and membrane potentials. Compare to Figure 2c (note: y-axis scaling different). The inset shows loss of voltage dependence of gain in the noise compared to the quiescent condition (at 5 Hz input). Scale bars in b: horizontal below traces: 5 Hz, 0.1 s; 1 Hz, 0.5 s; 20 Hz, 20 ms; between traces: –70 mV; vertical (upper, lower): 10 mV, 20 nS. Error bars, s.e.m.

(Supplementary Fig. 1). Thus, noise effectively masked the intrinsic, nonlinear response behavior of thalamocortical cells and equipped them with a robust, voltage-independent transfer function. Stimulation with physiologically realistic inputs Mean firing frequencies of retinal ganglion cells in vivo are in the range of 5–50 Hz and are gamma or Poisson distributed31. Even if the magnitude of a single retinogeniculate EPSP may vary little, the ‘effective’ retinogeniculate EPSPs depend, among other factors, on variable degrees of temporal summation, such that the effective input has a larger magnitude range32,33. In the experiments described so far, the variability of the effective input was achieved by randomizing the input conductances while fixing the input frequency, thereby separating magnitude from frequency and thus allowing better control and comparison with cortical neurons17. To check whether the spike probability is voltage and frequency dependent under more physiological input conditions, we compared the response properties of thalamocortical cells during stimulation with Poisson-distributed ‘retinal’ input using a mean frequency of 10 Hz. Even though the retinogeniculate input conductance was fixed, Poisson-rate stimulation led to varying effective input EPSP magnitudes as a result of summation (Fig. 4a). At the beginning of each experiment, the conductance magnitude was adjusted such that evoked subthreshold EPSPs at resting potential were in a physiological range (5–15 mV)32. Because the degree of input summation is dependent on the frequency, we used the inter-stimulus interval (ISI) immediately preceding the response to measure input strength (Fig. 4b). At resting potential, with subthreshold input magnitude, spikes were evoked only by summed inputs at smaller ISIs (Fig. 4a,b): ISIs in the range of 50–600 ms were related to spike probabilities in the range of 0 to 0.021 ± 0.016 (n ¼ 4), whereas ISIs shorter than 50 ms were


[

NUMBER 12

[

DECEMBER 2005

associated with a spike probability of 0.530 ± 0.123 (n ¼ 4). In contrast, at hyperpolarization, spike response could be evoked not only by input summation, but also by long ISIs (Fig. 4a,b; ANOVA, P o 0.001): ISIs in the range of 300 ms to 600 ms were related to spike probabilities in the range of 0.253 ± 0.099 to 0.847 ± 0.061 (n ¼ 4), even higher than those evoked by ISIs shorter than 50 ms (0.398 ± 0.089, n ¼ 4). Thus, consistent with our experiments using fixed input frequencies and randomized input magnitudes, spike probabilities induced by Poisson rate input had an all-or-none character at resting potential but adopted intermediate values, depending on input frequency, at hyperpolarized potentials. We repeated the Poisson rate experiment described above, in the presence of noise (Fig. 4c,d). With noise, input summation also increased the spike probability at resting and hyperpolarized potentials. However, unlike in the quiescent condition, larger ISIs did not lead to different spiking probabilities at the hyperpolarized potential as compared to the resting potential (ANOVA, P ¼ 0.43). Although no direct comparison between fixed rate and Poisson rate experiments (such as comparing the gain) is feasible, these results match: in both experimental conditions, voltage and frequency dependencies were abolished with synaptic background noise. Noise increases burst firing Action potential–burst firing in thalamocortical cells is classically considered to be the result of LTS activation after hyperpolarization34–36. However, in the noise condition, even at resting and depolarized potentials, high-frequency burst responses often occurred (Fig. 5a, left). Because noisy voltage fluctuations hamper LTS identification, we used a burst detection algorithm35 (see Methods) to assess burstiness. Indeed, not only the voltage but also the presence of noise influenced burstiness (ANOVA, P o 0.001). At resting and depolarized

1763

ARTICLES

a

Quiescent Rest

a

Hyp

Resting, noise Avg total spike output

3

*

2 Noise 1 Quiescent

Quiescent

b

0.6 80

0.4 0.2 Rest 500

400

300

200

100

0

ISIi – ISIi–1 (ms)

60

12 10

Dep

60 5 Hz 40

Noise 20

Quiescent Hyp

Rest

Dep

6

5

5

20 Hyp

0.8 0.6

6

5

2

Noise

5

16 25

Quiescent

11 10 4 0

0

5

20

5

20

4

6

5

20

Input frequency (Hz)

0.2 0.0 600

Dep

*

20

0.4

6

20

40

Rest

Hyp

5

Rest

80 Burst threshold (nS)

Noise

1.0

7

8

5 20 Input frequency (Hz)

***

d

*

** 9

60

7

40 27

Hyp

c

*** 7

***

20

Noise

Rest

60

0

40

0

500

400

300

200

100

0

ISIi – ISIi–1 (ms)

Figure 4 Physiologically realistic Poisson-distributed inputs. (a) Retinogeniculate input conductances (lower traces) were injected at a fixed magnitude with Poisson-distributed rate (mean frequency, 10 Hz). At resting potential (Rest) without synaptic noise, action potentials were only evoked by summation of inputs occurring at short interstimulus intervals (ISIs) whereas inputs with long ISIs did not trigger action potentials (left panel, asterisk). At hyperpolarized potential (Hyp), subthreshold inputs led to variable degrees of EPSP summation occasionally accompanied by LTS activation. Large ISIs led to the activation of LTS-driven bursts (right panel, arrow). (b) The probability of evoking at least one spike was plotted against the ISIs. Whereas at resting potential spiking probability was increased only with high-frequency inputs, at hyperpolarized potentials the spiking probability increased strongly with low frequency inputs. (c,d) During the injection of noise, the difference in frequency-dependent response behavior of thalamocortical cells was strongly reduced. Spiking probabilities were approximately equal at all input frequencies in the presence of noise. Scale bars in a,c, 100 ms, 10 mV; lower trace, 10 nS; between traces, –80 mV. Error bars, s.e.m.

potentials, there was an increased burstiness with noise in 24 of 27 cases (Fig. 5b, inset; resting quiescent, 5 Hz: 5.7 ± 1.9%, n ¼ 40; versus noise: 21.2 ± 3.4%, n ¼ 27; P o 0.001; depolarized quiescent, 5 Hz: 0%, n ¼ 7; versus noise: 27.6 ± 6.1%, n ¼ 7; P ¼ 0.004). As expected, burst firing was more pronounced at hyperpolarization (5 Hz quiescent: 44.5 ± 9.6%, n ¼ 12), and there was no clear change of burstiness with noise (Fig. 5c, left). Another approach to quantifying burstiness is to determine the ‘burst threshold’: that is, the input level at which bursts first occur. This analysis again showed that at resting and

1764

20 40 Input (nS)

Rest Burstiness (%)

Hyp

100

Burstiness (%)

Spike probability

b

Hyp

0.8

0.0 600

c

0

1.0

Spike probability


0

Figure 5 Noise increases occurrence of bursts at resting potential. (a) With noise, high-frequency bursts of action potentials occurred at resting potential (left). Plotting the average total number of spikes per burst response (Supplementary Methods online) against the input shows that noise linearized the staircase-like transfer function across the whole input range (right). (b) Percent bursts per spike-evoking stimulation with different input frequencies and membrane potentials. Noise increased burstiness at resting and depolarized potentials (for example, at 5 Hz, inset). This effect was smaller with higher input frequencies. Noise did not increase the burstiness at hyperpolarization. (c) The input value at which bursting occurred (burst threshold) was decreased with noise at rest and depolarization but not at hyperpolarization. Scale bars in a, 100 ms, 10 mV; lower trace, 10 nS; before trace, –80 mV; inset, 10 mV, 1 ms. Error bars, s.e.m. *P o 0.05, **P o 0.01, ***P o 0.001.

depolarized potential, bursting was more likely with noise because the bursting threshold was much lower than in the quiescent condition (Fig. 5c, resting quiescent, 5 Hz: 58.5 ± 3.8 nS, n ¼ 16; versus noise: 33.3 ± 3.4 nS, n ¼ 25; P o 0.001; depolarized quiescent, 5 Hz: no bursting; versus noise: 13.5 ± 8.0 nS, n ¼ 6). What could be the reason for the increased burstiness with noise? In principle, a decrease in the membrane time constant resulting from conductance increase could have a role; however, no increased burstiness was detected in the static condition (burstiness at resting quiescent, 5 Hz: 6.2 ± 3.3%; versus at static: 6.5 ± 3.3%; n ¼ 22; paired test: P ¼ 0.44). The mean membrane potentials in the noise and quiescent conditions were not significantly different (resting quiescent, 5 Hz: –65.3 ± 0.8 mV; versus noise: –65.7 ± 0.8 mV; n ¼ 24; paired test: P ¼ 0.80), arguing against a role of T-type channels. Indeed, the gating parameters of T-type channels suggest little involvement in these

VOLUME 8

[

NUMBER 12

[


ARTICLES

a

Quiescent, rest

Second spike

60

Number

First spike 40 20 0

b

10

20

30 Input (nS)

Noise, rest

40

First spike

60

50

60

Second spike

40 Overlap first/second

20 0 0

Overlap first/second (%)

c

40

20

30 Input (nS)

Hyp

40

50

60

Rest

Dep

Noise 30

*

Quiescent

11/12 10/10

*

0/5 6/7

4/7 4/6

5

20

6/6 5/5

***

20

16/40 25/27 2/9 6/8

10

0

d

10

5

20

Hyp

40

Rest

30

20


Overlap first/second (%)

Number

Third spike

Overlap first/third (%)


0

Overlap first/second

Figure 6 Noise mixes single-spike and burst responses of thalamocortical neurons. (a) Example of input strength distributions evoking single-spike (first spike) and burst responses (second spike, with interspike intervals r5 ms). Without noise, only very strong inputs could evoke spike doublets. (b) With noise, most input levels could evoke single-spike and burst responses. Different shades of grey mark areas of overlap between first and second spikes (light), second and third spikes (middle) and first and third spikes (dark). (c,d) Mixing (overlap) of single-spike and multiple-spike responses: (c) first and second spikes; (d) first and third spikes. Noise increased the mixing of single-spike and burst responses at most potentials and input frequencies (for example, at 5 Hz, inset). Error bars, s.e.m. *P o 0.05, ***P o 0.001.

*

Dep

5 Hz

30 20

Noise

10

Quiescent

0

Hyp

Rest

Dep

6/12 7/10

10 2/6 4/5

0/5 6/7 1/40 13/27

1/7 3/6

0/9 2/8

0

5

20


5

20

conditions30, although native T-type currents of thalamic neurons can be available at resting potential37. We compared spike-triggered averages (STAs) of single-spike and burst responses during noise at the resting potential and observed that there was a small but significant difference in the pre-response voltage (Supplementary Fig. 2 online). This suggests that even if, overall, noise has no effect on the membrane potential, short hyperpolarizations preceding inputs statistically recruit more T-type channels as compared to the situation in the quiescent state. In addition, the occasional occurrence of noise-induced depolarizations with retinal inputs clearly facilitated bursts, as seen from spike-triggered averages of excitatory and inhibitory conductances (Supplementary Fig. 2). These results show that noise increases the occurrence of burst responses at resting and depolarized potentials but not at hyperpolarized potentials. Noise mixes single-spike and burst responses If the number of spikes in the response grew proportionally with the strength of the input, the spike count could be used to reliably encode sensory information. Without synaptic background, such a reliable transfer function does not exist in the resting and depolarized states; the cell behaves as a high-pass filter, detecting only strong inputs with no


[

NUMBER 12

[

DECEMBER 2005

discrimination of strength past a threshold (see the step-like response in Fig. 5a, right). Noise made cells generate, on average, a number of spikes proportional to input strength (Fig. 5a, right), providing a more linear transfer function at all potentials. Thus, in the presence of synaptic background activity, probabilistic ‘mixing’ of single-spike and burst responses potentially provides better encoding capabilities. To quantify this mixing, we separated single-spike responses from responses with two- and three-spike bursts. We counted the different responses at the respective input levels and plotted their distribution (Fig. 6a,b). We calculated the relative overlap of the one-, two- and three-spike response curves by integration and compared the ‘percent mixing’ (overlap) for the different conditions (Fig. 6c,d). With noise, at resting potential, the overlap of one- and two-spike response curves was increased (Fig. 6a–c, 5 Hz quiescent: 4.3 ± 0.52%, n ¼ 16; versus noise: 14.6 ± 1.4%, n ¼ 25; P o 0.001). In the quiescent condition, only 16 of 40 cells showed two-spike bursts and one cell showed three-spike bursts with very strong inputs. In contrast, with noise, 25 of 27 cells showed two-spike bursts and 13 showed three-spike bursts; the overlap of the latter was zero in the quiescent condition but notable in the noise condition (Fig. 6d, 4.6 ± 1.4%). The same difference in the mixing of single-spike and burst responses was true for depolarized and hyperpolarized potentials, although it was less marked for the latter (Fig. 6c,d; inset in d). At 20-Hz input, the difference in percent mixing was generally reduced (Fig. 6c,d). Thus, although thalamocortical cells recorded in vitro are usually either in single-spike mode or in burst mode8, our data suggest that thalamocortical cells under the influence of synaptic background activity may show both burst firing and singlespike responses. DISCUSSION The main findings reported here are that (i) the duality of burst and single-spike modes in thalamic relay neurons is strongly affected by the presence of synaptic background activity and (ii) the input-output transfer function is determined by synaptic background activity combined with intrinsic properties. Previous work has shown that understanding the responsiveness of central neurons requires a detailed knowledge of their intrinsic properties, which are mediated by various calcium- and voltage-dependent conductances10. Our present results suggest that background activity alters this responsiveness fundamentally. We suggest that a complete characterization of the properties of central neurons requires the knowledge of intrinsic and synaptic background conductances, as well as the amount of conductance fluctuations (noise). The dichotomy of burst and single-spike firing modes is based on recordings performed in vitro or in vivo during states of deep anesthesia (with slow waves in the electroencephalogram (EEG))23. During EEGactivated states, there is a sustained synaptic activity in thalamocortical cells23,38, which accounts for about 50% of their input conductance19. We found here that if such states are recreated artificially using the

1765


ARTICLES dynamic clamp, the distinction of firing modes is dramatically reduced: single-spike and burst firing seem to be mixed. Classically, burst firing in thalamic neurons has been invariably associated with hyperpolarization, and observations of bursts in awake, behaving animals36,39 were difficult to reconcile with the fact that thalamocortical cells are depolarized in this state39,40. Our findings suggest that bursts should be observable if thalamocortical cells are depolarized, such as in awake animals, but only with sufficient levels of synaptic background activity. Another consequence of synaptic noise is that it causes qualitative and quantitative changes in the responsiveness of thalamic neurons to external inputs. The response curves of thalamic neurons are very different in these simulated activated states compared to in the quiescent states. There is a multiplicative scaling by noise, leading to gain reduction and increased responsiveness to small inputs, similar to model predictions14,27 and results from dynamic-clamp experiments in cortical neurons16,17. It may be explained by the fact that the probability for small-amplitude inputs to evoke a spike can only be enhanced by noise (floor effect), whereas for higher inputs, the probability can only be reduced by noise (ceiling effect)17. An important difference from the cortex is that in quiescent thalamocortical cells, the gain is highly dependent on membrane potential and input frequency; the low-threshold calcium current boosts the response specifically at hyperpolarized levels. With noise, this voltage-dependent response behavior is nearly eliminated, and the gain is similar at all potentials. Thus, the combination of intrinsic properties (T-type calcium current) and synaptic noise enables relay neurons to keep a uniform responsiveness for a large input range at different membrane potentials and, in this sense, to perform a particularly robust relay of information. During sleep, thalamocortical cell bursting is part of large-scale synchronized activity, and this ‘proper’ burst mode transmits statedependent information to the cortex, different from single spikes6,8,23,40. During activated states, however, single-spike and burst responses may relay the same visual information, only with different efficiencies41. The mixing of single-spike and burst responses to excitatory stimuli, as well as the more graded aspect of bursts, suggest that with background synaptic activity, there is indeed no clear distinction between single spikes and bursts. This may actually be advantageous for the relay of visual information: instead of a step-like input-output curve with limited coding capabilities for different input strengths, a linear response curve across the whole input range is generated by combining single and burst responses, suggesting that they convey the same type of information to the cortex during activated states. Thalamic neurons share with cortical neurons the probabilistic aspect of responses under synaptic noise. This probabilistic nature of the thalamic relays is indeed realistic15,42,43 if we consider that many thalamocortical cells converge on individual cortical cells33,44,45, enabling those ‘receiver’ cells to collect many thalamocortical inputs (each of them firing one or several spikes in response to their sensory input) and to extract the probability function for a given stimulus. Finally, these findings support the idea that corticothalamic synapses have a powerful role in controlling information transfer by the thalamus. It is known that corticothalamic feedback constitutes the primary source of synapses in the thalamus—one order of magnitude larger than synapses from peripheric axons1,2. Despite this anatomical fact, the feedforward view is still prominent: the thalamus is often considered as a ‘relay’ of information to cortex. By regulating the intensity of background activity, the cortex could exert a fast and efficient control of the thalamic relay—through instantaneous adjustment of gain and of bursting probability—which may be related to focused attention mechanisms (see also refs. 4, 28 and 46).

1766

METHODS We prepared LGNd slices (350 mm) from adult guinea-pigs as previously described11. We placed the slices at 34.5–35.5 1C in an interface style recording chamber (Fine Science Tools). The bathing medium contained (in mM): NaCl (126); KCl (2.5); MgSO4 (1.2); NaH2PO4 (1.25); CaCl2 (2); NaHCO3 (26); and dextrose (10) and was aerated with 95% O2 and 5% CO2 to a final pH of 7.4. We made intracellular electrodes on a Sutter Instruments P-87 micropipette puller from medium-walled glass (World Precision Instruments, 1BF100). The electrodes were filled with 1.2 M potassium acetate and beveled (Sutter ` . We recorded at 10 kHz Instruments BV-10M) to resistances of 90–100 MU with an Axoclamp-2B amplifier (Axon instruments) in discontinuous current clamp (switching rate 2.5–3.5 kHz), using dynamic clamp26. Dynamic clamp and modeling techniques. Synaptic inputs were generated using a real-time version (Y. Le Franc, B. Foutry, F. Nagy & G. Le Masson, Soc. Neurosci. Abstr. 927.18, 2001) of the NEURON simulation environment47 in combination with a programmable digital signal processor board M67 (Innovative Instruments), to achieve fast calculation and injection of dynamic-clamp currents (0.1 ms time step). We injected two types of simulated synaptic inputs, ‘signal’ and ‘background’, as conductances into thalamocortical cells using dynamic clamp26. The excitatory input from retinal ganglion cells (signal)22 was modeled from the activation and inactivation kinetics of AMPA receptors48 with randomized peak conductances (1–80 nS; step size: 2 nS). We fixed the stimulation frequency (5, 10 and 20 Hz) and randomized input amplitudes to analyze frequency-dependent effects separately. We also used Poisson-distributed inputs31 of constant amplitude to approximate physiological input. To simulate the background synaptic input (‘synaptic noise’), we used a fluctuating conductance model, mimicking the effect of thousands of stochastically glutamate- and GABA-releasing synapses27. These excitatory and inhibitory conductances were simulated as independent stochastic processes, assuming that the influence of the cortex on relay cells is a mixture of uncorrelated excitatory and inhibitory postsynaptic potentials (Supplementary Methods online). We validated the accuracy of the dynamic clamp conductance–based noise injection (i) by directly comparing the voltage fluctuations in real and model cells in response to injection of identical noise (Supplementary Fig. 3 online), and (ii) by comparing natural and artificial (dynamic-clamp generated) synaptic noise in the same cells49. Data analysis and statistics. A spike was considered as a response to the input when it occurred within 20 ms after the stimulus onset. We analyzed inputoutput relations using spike probabilities instead of frequencies (see also Supplementary Methods). Spike probability was calculated as the number of EPSPs per given input level that generated at least one spike. We also considered the ‘average total spike output’ by taking into account all spikes of burst responses (interspike intervals o5 ms). To measure statistical significance, we performed single- and multifactorial analyses of variance (ANOVA; JMP software, SAS Institute). These tests were of the repeated-measures type in cases where all cells were submitted to the complete set of conditions, and of the independent type in the remaining cases. In the remaining cases, we assumed independence of conditions. We carried out post-hoc comparisons with non-parametric Wilcoxon rank-sum tests, using the software ‘Instat’ (University of Reading). Comparisons were of the unpaired, two-sided type, unless stated otherwise. We determined the significance of correlation according to a table of Pearson’s r-values. Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS We thank M. Rudolph, G. Sadoc and L. Focsa for help with computation and Z. Piwkowska and D. Shulz for comments on the manuscript. This work was supported by the Centre National de la Recherche Scientifique, the Human Frontier Science Program, the European Commission (IST-2001-34712) and the Action Concerteé Initiative ‘Neurosciences integratives et computationnelles’. J.W. is the recipient of a European Union Marie Curie fellowship. COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests.

VOLUME 8

[

NUMBER 12

[


ARTICLES


Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/ 1. Erisir, A., Van Horn, S.C. & Sherman, S.M. Relative numbers of cortical and brainstem inputs to the lateral geniculate nucleus. Proc. Natl. Acad. Sci. USA 94, 1517–1520 (1997). 2. Van Horn, S.C., Erisir, A. & Sherman, S.M. Relative distribution of synapses in the Alaminae of the lateral geniculate nucleus of the cat. J. Comp. Neurol. 416, 509–520 (2000). 3. Sherman, S.M. & Koch, C. The control of retinogeniculate transmission in the mammalian lateral geniculate nucleus. Exp. Brain Res. 63, 1–20 (1986). 4. Koch, C. The action of the corticofugal pathway on sensory thalamic nuclei: a hypothesis. Neuroscience 23, 399–406 (1987). 5. Ahissar, E., Haidarliu, S. & Zacksenhouse, M. Decoding temporally encoded sensory input by cortical oscillations and thalamic phase comparators. Proc. Natl. Acad. Sci. USA 94, 11633–11638 (1997). 6. Sherman, S.M. Tonic and burst firing: dual modes of thalamocortical relay. Trends Neurosci. 24, 122–126 (2001). 7. Sillito, A.M. & Jones, H.E. Corticothalamic interactions in the transfer of visual information. Phil. Trans. R. Soc. Lond. B 357, 1739–1752 (2002). 8. McCormick, D.A. & Bal, T. Sleep and arousal: thalamocortical mechanisms. Annu. Rev. Neurosci. 20, 185–215 (1997). 9. Jahnsen, H. & Llinas, R. Electrophysiological properties of guinea-pig thalamic neurones: an in vitro study. J. Physiol. (Lond.) 349, 205–226 (1984). 10. Llinas, R.R. The intrinsic electrophysiological properties of mammalian neurons: insights into central nervous system function. Science 242, 1654–1664 (1988). 11. Le Masson, G., Renaud-Le Masson, S., Debay, D. & Bal, T. Feedback inhibition controls spike transfer in hybrid thalamic circuits. Nature 417, 854–858 (2002). 12. Steriade, M. Impact of network activities on neuronal properties in corticothalamic systems. J. Neurophysiol. 86, 1–39 (2001). 13. Destexhe, A., Rudolph, M. & Pare, D. The high-conductance state of neocortical neurons in vivo. Nat. Rev. Neurosci. 4, 739–751 (2003). 14. Ho, N. & Destexhe, A. Synaptic background activity enhances the responsiveness of neocortical pyramidal neurons. J. Neurophysiol. 84, 1488–1496 (2000). 15. Anderson, J.S., Lampl, I., Gillespie, D.C. & Ferster, D. The contribution of noise to contrast invariance of orientation tuning in cat visual cortex. Science 290, 1968–1972 (2000). 16. Chance, F.S., Abbott, L.F. & Reyes, A.D. Gain modulation from background synaptic input. Neuron 35, 773–782 (2002). 17. Shu, Y., Hasenstaub, A., Badoual, M., Bal, T. & McCormick, D.A. Barrages of synaptic activity control the gain and sensitivity of cortical neurons. J. Neurosci. 23, 10388– 10401 (2003). 18. Larkum, M.E., Senn, W. & Luscher, H.R. Top-down dendritic input increases the gain of layer 5 pyramidal neurons. Cereb. Cortex 14, 1059–1070 (2004). 19. Contreras, D., Timofeev, I. & Steriade, M. Mechanisms of long-lasting hyperpolarizations underlying slow sleep oscillations in cat corticothalamic networks. J. Physiol. 494, 251–264 (1996). 20. Wilson, J.R., Friedlander, M.J. & Sherman, S.M. Fine structural morphology of identified X- and Y-cells in the cat’s lateral geniculate nucleus. Proc. R. Soc. Lond. B 221, 411–436 (1984). 21. Liu, X.B., Honda, C.N. & Jones, E.G. Distribution of four types of synapse on physiologically identified relay neurons in the ventral posterior thalamic nucleus of the cat. J. Comp. Neurol. 352, 69–91 (1995). 22. Sherman, S.M. & Guillery, R.W. The role of the thalamus in the flow of information to the cortex. Phil. Trans. R. Soc. Lond. B 357, 1695–1708 (2002). 23. Steriade, M., Jones, E.G. & McCormick, D.A. Thalamus Vol. 1 (Elsevier, Amsterdam, 1997). 24. Destexhe, A., Neubig, M., Ulrich, D. & Huguenard, J. Dendritic low-threshold calcium currents in thalamic relay cells. J. Neurosci. 18, 3574–3588 (1998).


[

NUMBER 12

[

DECEMBER 2005

25. Destexhe, A. & Sejnowski, T.J. The initiation of bursts in thalamic neurons and the cortical control of thalamic sensitivity. Phil. Trans. R. Soc. Lond. B 357, 1649–1657 (2002). 26. Prinz, A.A., Abbott, L.F. & Marder, E. The dynamic clamp comes of age. Trends Neurosci. 27, 218–224 (2004). 27. Destexhe, A., Rudolph, M., Fellous, J.M. & Sejnowski, T.J. Fluctuating synaptic conductances recreate in vivo-like activity in neocortical neurons. Neuroscience 107, 13–24 (2001). 28. Sillito, A.M., Jones, H.E., Gerstein, G.L. & West, D.C. Feature-linked synchronization of thalamic relay cell firing induced by feedback from the visual cortex. Nature 369, 479–482 (1994). 29. McCormick, D.A. & Feeser, H.R. Functional implications of burst firing and single spike activity in lateral geniculate relay neurons. Neuroscience 39, 103–113 (1990). 30. Perez-Reyes, E. Molecular physiology of low-voltage-activated t-type calcium channels. Physiol. Rev. 83, 117–161 (2003). 31. Troy, J.B. & Robson, J.G. Steady discharges of X and Y retinal ganglion cells of cat under photopic illuminance. Vis. Neurosci. 9, 535–553 (1992). 32. Turner, J.P., Leresche, N., Guyon, A., Soltesz, I. & Crunelli, V. Sensory input and burst firing output of rat and cat thalamocortical cells: the role of NMDA and non-NMDA receptors. J. Physiol. (Lond.) 480, 281–295 (1994). 33. Usrey, W.M., Reppas, J.B. & Reid, R.C. Paired-spike interactions and synaptic efficacy of retinal inputs to the thalamus. Nature 395, 384–387 (1998). 34. Lu, S.M., Guido, W. & Sherman, S.M. Effects of membrane voltage on receptive field properties of lateral geniculate neurons in the cat: contributions of the low-threshold Ca2+ conductance. J. Neurophysiol. 68, 2185–2198 (1992). 35. Ramcharan, E.J., Gnadt, J.W. & Sherman, S.M. Burst and tonic firing in thalamic cells of unanesthetized, behaving monkeys. Vis. Neurosci. 17, 55–62 (2000). 36. Weyand, T.G., Boudreaux, M. & Guido, W. Burst and tonic response modes in thalamic neurons during sleep and wakefulness. J. Neurophysiol. 85, 1107–1118 (2001). 37. Leresche, N., Hering, J. & Lambert, R.C. Paradoxical potentiation of neuronal T-type Ca2+ current by ATP at resting membrane potential. J. Neurosci. 24, 5592–5602 (2004). 38. Hirsch, J.C., Fourment, A. & Marc, M.E. Sleep-related variations of membrane potential in the lateral geniculate body relay neurons of the cat. Brain Res. 259, 308–312 (1983). 39. Guido, W. & Weyand, T. Burst responses in thalamic relay cells of the awake behaving cat. J. Neurophysiol. 74, 1782–1786 (1995). 40. Steriade, M. To burst, or rather, not to burst. Nat. Neurosci. 4, 671 (2001). 41. Reinagel, P., Godwin, D., Sherman, S.M. & Koch, C. Encoding of visual information by LGN bursts. J. Neurophysiol. 81, 2558–2569 (1999). 42. Lewis, J.E. Sensory processing and the network mechanisms for reading neuronal population codes. J. Comp. Physiol. A 185, 373–378 (1999). 43. Barber, M.J., Clark, J.W. & Anderson, C.H. Neural representation of probabilistic information. Neural Comput. 15, 1843–1864 (2003). 44. Tanaka, K. Organization of geniculate inputs to visual cortical cells in the cat. Vision Res. 25, 357–364 (1985). 45. Alonso, J.M., Usrey, W.M. & Reid, R.C. Precisely correlated firing in cells of the lateral geniculate nucleus. Nature 383, 815–819 (1996). 46. Montero, V.M. Amblyopia decreases activation of the corticogeniculate pathway and visual thalamic reticularis in attentive rats: a ‘focal attention’ hypothesis. Neuroscience 91, 805–817 (1999). 47. Hines, M.L. & Carnevale, N.T. The NEURON simulation environment. Neural Comput. 9, 1179–1209 (1997). 48. Destexhe, A., Mainen, Z.F. & Sejnowski, T.J. Kinetic models of synaptic transmission. in Methods in Neuronal Modeling 2nd edn. (eds. Koch, C. & Segev, I.) Ch. 1, 1–26 (MIT Press, Cambridge, Massachusetts, 1998). 49. Rudolph, M., Piwkowska, Z., Badoual, M., Bal, T. & Destexhe, A. A method to estimate synaptic conductances from membrane potential fluctuations. J. Neurophysiol. 91, 2884–2896 (2004).

1767


ARTICLES

Prior experience of rotation is not required for recognizing objects seen from different angles Gang Wang1,2, Shinji Obama1, Wakayo Yamashita1, Tadashi Sugihara2 & Keiji Tanaka2 An object viewed from different angles can be recognized and distinguished from similar distractors after the viewer has had experience watching it rotate. It has been assumed that as an observer watches the rotation, separate representations of individual views become associated with one another. However, we show here that once monkeys learned to discriminate individual views of objects, they were able to recognize objects across rotations up to 60°, even though there had been no opportunity to learn the association between different views. Our results suggest that object recognition across small or medium changes in viewing angle depends on features common to similar views of objects.

An unfamiliar object often cannot be distinguished from similar or metrically varying distractors when the viewing angle changes1–6. Recognition across changes in viewing angle develops as the viewer sees the object while it rotates. Rotations provide two kinds of experience: exposure to different views of the object, and sequential pairing of these views with each other. The latter aspect has been central to discussions about the formation of the capability to recognize objects across changes in viewing angle. It has been assumed that different views of each object become associated with one another during object rotation, either through active learning or through passive experiencing of the successive appearance of nearby views7,8. This association, which forms during rotations, is thought to underlie object recognition ability across changes in viewing angle9–12. In contrast, the former aspect of observing an object in rotation— that is, the exposure to different views—has been relatively neglected. However, single-cell recordings in monkeys have shown that the neuronal representation of visual images changes as the monkey repeatedly experiences and discriminates between images13–18. Thus, the experience of different views alone, even if they do not appear in succession, may be important to the formation of perceptual tolerance to changes in viewing angle. We designed the present study to investigate the effects of experiencing different views without the experience of their successive appearance. A small proportion of cells in the anterior superior temporal sulcus and inferotemporal cortex of monkeys respond to all views of a particular person’s head or of a particular object, rather than to views of other people or objects19,20. Such perfect object selectivity across different views, a property of a small number of cells, might underlie the tolerance of perception to changes in viewing angle. In other studies, inferotemporal cells show moderately broad tunings for viewing angle of objects15,21–23. The response remains larger than half-maximum over viewing angles of 15–50° from the optimal. This partial object selectivity of a large number

of inferotemporal cells might also contribute to the tolerance of perception to changes in viewing angle. However, in these previous studies, the recordings were made after the monkeys had repeatedly experienced object rotations, and thus they do not address our present question: to what extent does object recognition across changes in viewing angle develop without the viewer’s experiencing rotation of an object? To address this issue, we conducted an experiment in which monkeys were exposed to different views of objects in an object recognition task. Monkeys, rather than humans, were used in the present study so that the behavioral results could later be compared with single-cell recording data in the same species. Similarity between objects and distractors affects the extent to which changes in viewing angle can be tolerated in the recognition of unfamiliar objects1–6,24–26; therefore, we systematically controlled the similarity between objects and distractors used in the experiments. We found that once monkeys learned to discriminate objects at each viewing angle, they were able to recognize objects across rotations up to 60°, even though there had been no opportunity to learn the association between different views. RESULTS For each experiment, we made four artificial objects by deforming a prototype in four different directions in a feature space, which was spanned by three combined parameters of the shape (Fig. 1a; also see Methods). Sets of objects were created by deforming individual prototypes to different extents (Fig. 1b.) Four different views of each object were created by rotating the object by 30° intervals around an axis perpendicular to the visual axis connecting the viewer’s eyes and the object. In each experiment, this set of 16 images was used to test object recognition across different viewing angles. Monkeys performed a discrimination task (‘Object task’) in which they had to detect a change in the identity of the object (Fig. 2). Two to five object images were presented sequentially in each trial: one to four

1Department

of Bioengineering, Faculty of Engineering, Kagoshima University, Kagoshima, Kagoshima 890-0065, Japan. 2Cognitive Brain Mapping Laboratory, RIKEN Brain Science Institute, Wako, Saitama 351-0198, Japan. Correspondence should be addressed to K.T. ([email protected]). Received 19 April; accepted 21 October; published online 20 November 2005; corrected online 8 August 2006; doi:10.1038/nn1600

1768


ARTICLES a

Object 4

Prototype

b

0

Version +8

o

30

o

o

60

o

90

Version 0

Prototype

Versio n

Object 3

Versio n

Object


1 2 3

Object 2 0

4

+4

Object 1 Figure 1 Formation of object sets and example sets of stimulus images. (a) The four objects in a set were positioned at the vertices of a tetrahedron in a three-dimensional feature space. Seven parameters of the object shape were combined into three, which spanned the feature space. Object sets with various distances from the center (the ‘Prototype’) were made. On Version 0 human subjects showed ~80% correct performance. (b) The Prototype (top), sets of dissimilar (left) and similar objects (right). A set comprised 16 images, four views for each of four objects. Views are aligned horizontally, and objects are aligned vertically.

views of a first object were followed by one view of a second object. The monkey held down a lever while different views of the first object were presented. When an image of the second object appeared, the monkey had to release the lever. This task examined the monkeys’ capability to recognize objects while tolerating changes in viewing angle, as they had to distinguish between image changes resulting from changes in viewing angle alone and those resulting from changes in both viewing angle and the identity of object. Tests without prior experience of individual images In the first series of experiments, new stimulus sets were introduced in the object task while the knowledge of the task itself was maintained in the monkeys by using a consistent familiar set (Fig. 3a). Four different object sets with different levels of similarity were introduced to

Stimulus

on off

Fixation point

on off

Lever

press release

Eye fixation

inside outside

First object (1–4 times) 0.5 s 0.5 s

Second object

0.5 s 0.5 s

each monkey, one at a time. When the objects within the set were very different from one another, the monkeys showed good discrimination: the proportion of hits was significantly larger than that of false alarms (‘Version +8’ and ‘Version +4’ in Fig. 3b). They responded to object changes, but not to pure view changes. When the objects in the set were more similar to one another, both monkeys failed to discriminate object changes from pure view changes: there were no significant differences between the proportions of hits and false alarms at 30°, 60° and 90° rotations (Versions +2 and 0 in last three rows of Fig. 3b). These similar objects were still within the discrimination capability of the monkeys as, in post hoc tests, the monkeys could well discriminate the object images at each viewing angle (Fig. 3c). These findings, which are consistent with previous studies using human and monkey subjects1–6,24,25, gave the baseline performance without the prior experience either of individual views or of their sequential appearance.

0.5 s 0.5 s

0.5 s

0–1.0 s 0–1.5 s

Object task

First object

Preparatory task

Second object

First object

Second object

Figure 2 The time sequence of events in the task. The monkey had to hold down a lever while one to four views of a first object were presented and then release the lever when a view of a second object appeared. In the object task, but not in the preparatory task, the viewing angle of object images changed from presentation to presentation.


Tests with prior experience of individual images In the second series of experiments, we examined the monkeys’ capability to recognize objects across rotations after the monkeys had repeatedly experienced individual images of a new stimulus set in a preparatory task (Fig. 4a). In the preparatory task, the monkeys learned to discriminate images at each viewing angle, but there was no opportunity for them to learn the association between different views of each object. The training period lasted for at least 4 weeks before a new set was introduced to the object task. When new sets composed of similar objects (Fig. 4b) were introduced into the object task after the monkeys had experienced individual images, one monkey successfully discriminated object changes from pure view changes at 30° and 60° rotations, within the first 50 trials (Fig. 4c,

1769

a

Figure 3 Performance of monkeys immediately after the introduction of new stimulus sets into the object task, without prior experience of individual images. (a) The time course of the experiments. Each new object set was introduced while the monkey’s knowledge of the task was maintained by using a familiar set. The number within the box indicates the version number of the object set used in the experiment. (b) The white and black bars represent the hit and false alarm rates, respectively, at the second stimulus presentation in each trial. The values at the right (0°, 30°, 60° and 90°) indicate how much the viewing angle changed between the first and second presentations. Four sets of data, vertically aligned in a column, were obtained in one experiment. The number at the bottom indicates the version number of the object set used in the experiment. Smaller version numbers mean that objects within the set were more similar to one another. The data were pooled from the first 50 trials of each condition. *P < 0.01; **P < 0.001; ***P < 0.0001 (χ2 test). Note that the performance at 0° rotation did not require the tolerance of object recognition to changes in viewing angle. (c) The performance of monkeys in discriminating the object images within each viewing angle, examined in the preparatory task after the main test with the object task was completed. The hit and false alarm rates at the second presentation are plotted by the solid and broken lines, respectively. The x-axis represents the day after the start.

a

Monkey 1

New sets Familiar set

b

+4

0 +8 Object task Time

Monkey 1 1.0

***

*** **

***

*

0 1.0

Performance

0.5 0 1.0

***

*** ***

0 1.0

***

60°

*

0 Monkey 2 1.0

***

90°

Set 7, 0

0 1.0

***

0.5

*

0 Version +8

+4

Set 10, +2

0.5

+2

0

+8

+4

+2

0

0 1

Similarity

c

Monkey 1 *** ***

0 1.0 Performance

Set 1 series

Set 11 series

–2 +4 +2 0 Version

–2

***

Similarity

3 5 Day

7

The monkeys’ performance thus showed a clear contrast between the first and second series of experiments, for stimulus sets composed of similar objects. They successfully discriminated object changes from pure view changes at 30° and 60° rotations, when the test was conducted after they had experienced individual views of the objects (Fig. 4c); however, without prior experience of individual views, their performance was at chance level (Fig. 3b: Versions +2 and 0). Nevertheless, because different

Monkey 2 *** ***

0°

***

***

*** 30°

***

**

60°

0.5 0 1.0

***

0.5 0 1.0

90°

0.5 0 Set 2, 0 Set 8, 0

1770

Set 7, +2

0.5

Object task

0 1.0

+4 +2 0

***

0.5

0.5

0

0 1.0

30°

*

1.0

Set 8 series

0° *

Performance

***

1.0

0.5

Set 4, 0

0.5

Preparatory task Object task

Monkey 2

0.5 Set 2 series

Monkey 2 *** *

1.0

0.5 ***

Time Monkey 1

+2

0.5

New sets Familiar set

b

c

Object task

left). The second monkey also showed clear discrimination at 30° (Fig. 4c, right). At 60°, this monkey showed a significant difference between hit and false alarms in one experiment with one object set (‘Set 1, 0’), but not in the other experiment conducted with another set (‘Set 11, 0’). Neither monkey showed discrimination at 90°. The discrimination at 90° developed within a week after the introduction of the new stimulus sets (Supplementary Fig. 1 online).

Performance


ARTICLES

Set 1, 0 Set 11, 0

Figure 4 Performance when new stimulus sets were introduced into the object task after each monkey had discriminated the images from one another at each viewing angle in the preparatory task. (a) The time course of the experiments on each monkey. The same procedure was repeated twice with different object sets. (b) The performance of monkeys with the object set series in the preparatory task. The hit and false alarm rates at the second presentation are represented by the solid and broken lines, respectively. The x-axis represents the version number of stimulus set. Smaller version numbers indicate that objects were more similar to one another in the set. The vertical broken lines indicate the versions used later in the object task. (c) Performance on the initial 50 trials of each condition in the object task (notations as in Fig. 3b). The similarity between objects within the sets used in these tests was comparable to that of the most similar sets used in the test without prior experience of individual images (Version 0; see also Supplementary Fig. 2).


ARTICLES a

Preparatory task

Object task

Monkey 1 Set 9, +2 Familiar set Object task

Set 12, 0

Monkey 2 Set 9, +2

Set 12, 0 Familiar set Object task

Set 9 series

c

Set 12 series

Set 9, +2 1.0

Set 12, 0 ***

0° **

0.5

1.0

0 1.0

0.5

***

0

+4 +2 0 –2 +4 +2 0 –2 Version Version Monkey 2

Performance

Performance

Monkey 1

*** 30°

0.5 0 1.0

**

60°

0.5

Performance

prior experience, but on different monkeys (cross-design). For each set, only one mon90° key was given prior experience discriminating 0.5 individual images at each viewing angle in the 0.5 0 preparatory task (Fig. 5a). The object sets used +4 +2 0 –2 +4 +2 0 –2 0 in this test were composed of similar objects Version Version Without With Without With (Fig. 5b). With prior experience of individual Monkey 2 Monkey 1 Monkey 1 Monkey 2 images, the monkeys discriminated object changes from pure view changes of up to 60° or object sets were used in the two series of experiments, a small possibil- 30° when new sets were introduced into the object task (columns marked ity remained that the differences were due to unexpected differences in ‘With’ in Fig. 5c). There was no sign of discrimination ability without prior experience (columns marked ‘Without’ in Fig. 5c). This contrast of stimulus objects rather than to differences in prior experience. To exclude this possibility, we conducted an additional series of tests, performance with the same object set confirmed that the differences in in which the same object sets were used for the tests with and without performance were due to prior experience of individual images. 1.0

0 1.0

1

Set 1, 0

Set 2, 0

Set 8, 0

Set 11, 0

0 1

Distance


Time

b

Figure 5 Performance of monkeys in cross-design tests. The notations are the same as in Figures 3 and 4. (a) The performance on Set 9, +2 was first tested on Monkey 1 with prior experience of individual images in the preparatory task, and then on Monkey 2 without prior experience. Conversely, the performance on Set 12, 0 was first tested on Monkey 2 with prior experience, and then on Monkey 1 without prior experience. (b) Performance of monkeys on the Set 9 series (left) and Set 12 series (right) in the preparatory task. Two sets of data (Set 9 series in Monkey 1 and Set 12 series in Monkey 2) were obtained before the object test was conducted with the stimulus set, whereas the other two sets (Set 9 series in Monkey 2 and Set 12 series in Monkey 1) were obtained after the object test. (c) Initial performance of monkeys in the object task.

0 1

Set 12, 0

Set 9, +2 0

Control analyses and experiments One may suppose that the discrimination of object changes from pure view changes was due to differences in similarity between pairs of views within objects and pairs across objects. This was not the case in the stimulus sets we used to examine the effects of prior experience of individual images (sets of Versions 0 and +2). We calculated the similarity between two images as the Euclidean distance between coefficients of their wavelet transformations. The position and orientation of one image relative to the other was changed to obtain the minimum distance. When the distances were compared between pairs within objects and pairs across objects with the same differences in viewing angle, the two distributions corresponded well with each other at viewing-angle differences of 30–90° (Fig. 6); further, there were no significant differences in any comparison in any stimulus set used to test the effects of prior experience of individual images (P > 0.05, t-test). The same results were obtained when the similarity between two images was calculated as the Euclidean distance between the luminosity values in individual pixels with adjustment of position and orientation. Therefore, the monkeys’ discrimination of object changes from pure view changes did not depend on the general similarity of images.

1

Set 13, 0

Set 14, 0

0 0° 10° 20° 30° 60° 90°

0° 10° 20° 30° 60° 90°

Difference in viewing angle


Figure 6 Similarity between images for within-object and acrossobject pairs. The euclidean distance between coefficients of wavelet transformations of the images was calculated, normalized by the maximum distance in each set and plotted against the difference in viewing angle between the images. Dotted lines represent pairs within objects and solid lines represent pairs across objects. Note that only views with 30°, 60° and 90° intervals were used in the present study. Error bars represent s.d.

1771

a

b

α β γ δ

11

12

13

14

21

22

23

24

31

32

33

34

41

42

43

44

Pseudo association

Preparatory task

β

New sets

γ

Familiar set

Object task time

δ

0°

α

30° 60°

c

Set 3 series

d

1.0

Monkey 1

1.0

90° Set 3, +1

0.5 0.5 0 0 1.0

1 Set 6 series

0.5

Performance

Figure 7 Control experiments with pseudo combinations of object views. (a) Different views of different objects were combined. The first digits indicate the object number and the suffixes indicate the view number. The pseudo combinations are indicated by lines and marked by the Greek letters. The monkeys were required to neglect changes within the pseudo combinations and respond to other changes. (b) The time course of the experiments on Monkey 1. The test of pseudo association started after the monkey experienced individual images in the preparatory task. The same procedure was repeated twice with different stimulus sets on Monkey 1, whereas only one experiment was conducted on Monkey 2. (c) The performance of monkeys with the object set series in the preparatory task. (d) Daily performance of monkeys in the pseudo association task. The hit and false alarm rates at the second presentation in each trial are represented by the solid and broken lines, respectively. The performance at 0°, 30°, 60° and 90° changes from the first to the second presentation is indicated by the gray, red, green and blue lines, respectively. The x-axis represents the day after the start.

Performance

3

2

4

5

7

6

1.0 Set 6, +4

0.5 0

0 1.0

1

2

Set 2 series

3

5

4

Monkey 2

Set 2, +2

1.0

Performance

It is also possible that the monkeys learned the association between pairs of views of each 0.5 0.5 object in the first 50 trials after the introduction of a new set. To examine this possibility, we 0 0 artificially grouped object images across objects +4 +2 0 –2 1 3 5 2 4 (‘pseudo combinations’, Fig. 7a); we then Version Day trained the monkeys to hold the lever down when the image changed to another member of the pseudo combination and to release it when the image changed to the first appearance of within-object pairs (Fig. 8b), indicating that the one belonging to other combinations. New stimulus sets were introduced monkeys discriminated object changes from pure view changes at the to this pseudo association task after the monkeys had experienced indi- first appearance of each pair. To infer what visual cues corresponded to the changes in performance vidual images in the preparatory task (Fig. 7b). The stimulus sets used in this test were composed of similar objects (Fig. 7c). If the monkeys’ between a 30° rotation and a 90° rotation or between a 60° rotation and successful performance in the object task had been achieved by a quick a 90° rotation, we analyzed feature differences between views of each learning of associations between views, the pseudo associations should also have been learned quickly. We found that the monkeys completely failed to learn this task—not only within the first day, but also over a Monkey 2 Monkey 1 week (Fig. 7d). a To examine whether the monkeys discriminated object changes from 1.0 pure view changes from the beginning of the test with the object task, we also conducted an experiment in which we measured each monkey’s performance at the first appearance of each pair of views. New sets composed 0.5 of similar objects were introduced to the object task after the experience of individual images in the preparatory task (Fig. 8a). The change in viewing angle between successively presented stimuli was limited to 30°, and the stimulus sequences were controlled so that new pairs always 0 +4 +2 0 –2 +4 +2 0 –2 appeared as the first and second stimuli in the sequence within a trial. In Set 13 series both monkeys, the proportion of correct releases at the first appearance Set 14 series of across-object pairs was significantly larger than that of false releases at b 1.0 Figure 8 Performance on the first appearance of each pair of object views. Notations are the same as in Figures 3–5. (a) The monkeys’ performance with the object set series in the preparatory task. (b) The hit and false alarm rates on the first appearance of each pair of images. We controlled the stimulus sequence so that new pairs always appeared as the first and second stimuli in the sequence of a trial.

1772

Performance


ARTICLES

***

30° **

0.5

0

Set 13, 0

Set 14, 0



ARTICLES object in the stimulus sets used to examine the effects of prior experience of individual images. Because we did not know which features the monkeys used to perform the task, we arbitrarily focused on the seven shape parameters that we had systematically changed to create four objects from the prototype (see Methods). We measured the value of each parameter in the projected images, normalized the values so that the largest and smallest values of each parameter became 1 and 0, respectively, and calculated absolute differences of normalized values of each projected parameter between views of each object. The differences were, on average, 0.18 for view pairs with 30° difference in viewing angle, 0.32 for 60° pairs and 0.49 for 90° pairs. The large value for 90° pairs means that prominent differences in some parameters between objects at one viewing angle tended to disappear or change directions with 90° rotations. Therefore, features that varied with the parameters were useless in pairing views of each object across 90° changes in viewing angle. There were no such drastic differences for view pairs with 30° or 60° separation in our stimulus sets. DISCUSSION Our data indicate that the monkeys discriminated object changes from pure view changes without any prior experience of the pairing between views of each object. Repeatedly experiencing individual images and discriminating among these images at each viewing angle was necessary and sufficient for the formation of this capability. However, this was true only for viewing-angle changes up to 60°. Object recognition across viewpoint changes >60° required specific learning of the pairing. This finding applies only to object recognition among similar or metrically varying objects. When distractors were very different from the target object, recognition across views did not require prior experience of the objects at all1,4,5,24,26. One concern is that the difference in performance between the tests with and without prior experience of individual images was due to unexpected differences in the object similarity between the object sets used in the two types of tests. This possibility was excluded for the following three reasons. First, the amount of deformation used in the monkey experiments had previously been normalized by using human psychophysics (see Methods). Second, we confirmed, in the monkeys, that the object sets used for the tests without prior experience were not more difficult to discriminate than those used for the test with prior experience. For the stimuli used in the test with prior experience, the threshold of discrimination had been determined in the monkeys in the preparatory task, and we used the sets just above the discrimination threshold (Fig. 4b). For the stimuli used in the test without prior experience, we ran the preparatory task with the set after the test with object task was completed, thus confirming that the set was within the monkeys’ capability of discrimination (Fig. 3c). Third, the cross-design series of tests showed a clear contrast of performance on the same object sets with and without prior experience (Fig. 5c). The formation of object recognition across viewpoints without any prior experience of pairing between views of each object was not specific to particular object shapes, because similar results were obtained with eight different sets of objects (Supplementary Fig. 2 online). In addition, by analyzing the images, we excluded the possibility that the monkeys depended on differences in general image similarity between pairs of views within objects and pairs across objects. In the tested range of viewing-angle differences (30–90°), the similarity between images was comparable between pairs within objects and pairs across objects for all the stimulus sets we used to examine the effects of prior experience of individual views (Fig. 6). Previous single-cell recording studies have shown that as a monkey learns to discriminate among stimulus images, the number of cells that are responsive to these images increases in the inferotemporal cortex15,16.


Responses of inferotemporal cells become more selective for stimulus shape parameters that are useful for discrimination than for irrelevant parameters17,18. In the present study, during the prior experience in the preparatory task, the monkeys learned to discriminate between images at each of several viewing angles, but they did not have any opportunity to learn the pairing between different views of each object. The monkeys’ ability to recognize objects across views, immediately after the introduction of the stimulus sets into the object task, suggests that the features for which inferotemporal cells became tuned during the image discrimination were common to nearby views. Because the emergent representations of nearby views of each object had common features, the monkeys recognized the objects even across views when they were introduced to the object task. Alternatively, feature-based attention might have developed during the image discrimination. The features to which attention was directed for the detection of object changes within each viewing angle were common to nearby views. In either case, the monkeys discriminated features that did not change with rotation from view-dependent features that changed significantly with rotation. Cells in the monkey inferotemporal cortex tend to be more sensitive to rotation-independent types of changes in object images than to changes that occur as a result of rotation27,28. Monkey inferotemporal cells tend to respond more similarly to different views of same objects than to different views of different objects (T.S., S. Edelman & K.T., Soc. Neurosci. Abstr. 26, 2000). This tendency exists not only for objects for which the monkey has learned the pairing between views, but also for objects for which the monkey has only experienced individual views separately. Features that are common to nearby views may be preferentially used by inferotemporal cells to represent object images, regardless of prior experience of the pairing between views of each object. As the difference in viewing angle increases between two views of an object, there tend to be fewer features in common across the two views. In the object sets of Versions 0 and +2, with which the monkeys showed different performances depending on prior experience of individual views, common features largely disappeared as the viewing-angle difference approached 90°. This fact is consistent with the monkeys’ successful recognition of objects at viewing angles of 60° but not 90°. However, depending on the object shapes and view-range selections, discriminative features may disappear or lose the correspondence in parametric rank order even with smaller changes in viewing angle. Therefore, the precise viewing-angle ranges over which object recognition was retained in the current study cannot be immediately generalized to other cases or other object sets. In conclusion, the present results suggest that the tolerance of object recognition to changes in viewing angle of up to medium magnitude (~60° for the objects used in the present study) has a different mechanism than the tolerance to larger changes in viewing angle. Regularities existing between nearby views of each object underlie object recognition for medium changes in viewing angle, whereas specific learning of pairing between views of each object is necessary for object recognition across larger changes in viewing angle. METHODS We used two male macaque monkeys (Macaca fuscata) weighing 9 kg and 10 kg, respectively. In a preparatory surgery, we implanted a titanium head holder in the skull with titanium screws. All procedures on the monkeys were performed in accordance with the guidelines of the Japan Neuroscience Society and were approved by the Animal Experiment Committee of Kagoshima University. Stimulus objects. Stimulus objects were created using three-dimensional (3D) graphic software (Shade 6, e-frontier). A prototype was deformed in four directions in a 3D feature space to make four daughter objects. We used human psychophysics to adjust the directions of deformations and the relative amount of

1773


ARTICLES deformation in the four directions so that the positions of the daughter objects formed a tetrahedron with sides of equal lengths in psychological space (Fig. 1a). We also used human psychophysics to calibrate the amount of deformation in different object set series. The presentation condition in the human experiments was adjusted so that the percentages of correct responses roughly corresponded to those of the monkeys in the preparatory task. We defined several related object sets: Version 0 was the set yielding ~80% correct performance. The amount of deformation in Version 0 was multiplied by 0.4, 0.7, 1.3, 1.6, 1.9, 2.2 and 3.4 to make objects for Versions –2, –1, 1, 2, 3, 4 and 8, respectively. Because seven shape parameters were combined into three to span the 3D space and because none of six sides of the tetrahedron formed by the four objects was parallel to the axes of the feature space, differences between daughter objects were distributed among the seven parameters. We generated 14 series of object sets from 14 prototypes, which were quite different from one another (Supplementary Fig. 3 online). Object sets made from different prototypes were used in different experiments to prevent the generalization of learning from one experiment to another. The sizes of object images were 6.5° on average. Procedure. We formed a 3D shape by combining several geometrical primitives, by deforming some of the primitives, or both. We then selected seven parameters of the shape (for example, length, diameter and curvature of each part, and angle between two parts). Some of these parameters were combined by making the sum or ratio of two measures a constant, so as to reduce the number of parameters to three. We then determined the range for each parameter that did not provide any abrupt changes in shape. The three parameters were normalized so that a given amount of numerical change in each of the three parameters produced roughly the same perceptual changes. Next, four positions were selected in the 3D feature space spanned by the three parameters so that they formed a tetrahedron. None of six sides of the tetrahedron was parallel to the axes of the feature space. The prototype was determined by the geometrical center of the four positions. Many different objects were defined by positions along the lines connecting the prototype and the original four object positions. We then conducted human psychophysics with the original set of four objects, using a delayed match-to-sample task. On each trial, we presented an alert signal (500 ms), which was followed by a sample at the center (200 ms), a mask (500 ms), a target and distractor side by side (500 ms) and another mask (500 ms). The target was a different view of the object presented as the sample. The distractor was a view of another object from the same set. The subject reported the side of the target (left or right) by pushing one of two buttons within 1,000 ms from the onset of the target and distractor. There was a total of 120 trials, 20 trials for each object combination. If the percentage of correct responses was >80% for all three tests using one object, then that object was replaced with a new object located at a position 20% closer to the prototype along the same deformation axis. If the percentage of correct responses was <80% for all three tests, a new object was introduced from a position 20% farther from the prototype along the same axis. Another 120 trials were run with the new set of objects. We repeated the replacement of objects and psychophysical measures until we obtained 75–85% correct responses for all the object pairs. If the repetition had not met this performance criterion, we would have returned to the earlier step in which we normalized the three parameters, but this actually did not occur. Seven subjects (six for some object series) participated in the psychophysics experiment. The mean positions of the objects that provided 75–85% performance were taken to define the set of objects for Version 0. Each of the subjects underwent another 120 trials with this set of Version 0 to confirm that the percentage of correct responses was distributed around 80%. The distance from the prototype to the objects for Version 0 was multiplied by 0.4, 0.7, 1.3, 1.6, 1.9, 2.2 and 3.4 to make objects for Versions –2, –1, 1, 2, 3, 4 and 8, respectively. Object task. The head of each monkey was fixed by an implanted head holder. To initiate a trial, the monkey pressed a lever, at which the fixation spot appeared at the center of the screen. After continuous eye fixation for 500 ms, the first stimulus appeared. Up to five stimuli were presented within a trial, each for 500 ms with 500 ms inter-stimulus intervals. When the object changed (from ‘First object’ to ‘Second object’: see Fig. 2), the monkey had to release the lever within 1 s . Correct responses were rewarded with a drop of water. The inter-trial interval was 1.5 s after correct responses and 2.5 s after error responses. The monkey had to maintain eye fixation with an accuracy of ±2.5° until the last stimulus appeared. The object changed with one-third possibility at each of the second,

1774

third and fourth presentations. Therefore, the proportions of trial types were as follows: AB, 1/3; AAB, 2/9; AAAB, 4/27; and AAAAB, 8/27 (where A represents a view of the first object and B a view of the second object). Preparatory task. In the preparatory task, we exposed monkeys to individual images without the opportunity to associate different views of each object. In this task, each monkey was also trained to discriminate between the images within each viewing angle. Here, an identical image was repeated one to four times before a corresponding view of a second object appeared. The task was to respond to the object change. Versions –2, –1, 0, 1, 2, 3 and 4 were used, one per day, in a pseudo-random order. Each of the versions was used for at least 3 d in total. This procedure was also used to determine the discrimination threshold for each monkey. The version with the smallest deformations from the prototype—in which the monkey showed 70–80% difference between the hit and false alarm rates—was selected for the later test with the object task. Monkeys performed the preparatory task, with objects of the selected version, for an additional week before they started on the test with the object task. The training period on a new object set series lasted for at least 4 weeks. In the preparatory task, the position of the stimulus was changed randomly up to ±0.7° between presentations, so that the monkeys could not use local changes as cues. Throughout the whole course of tests, each monkey continued to perform the object task with a consistent familiar set (‘Set 1, 0’ in Monkey 1 and ‘Set 4, 0’ in Monkey 2). Trials with stimuli in the fixed set and those with stimuli in a new set were mixed in equal proportions. Analyses of performance. We focused on the monkeys’ performance at the second stimulus presentation in each trial. We calculated the proportion of correct releases (hits: trials in which the monkey correctly responded to an object change) and compared it with the proportion of incorrect releases (false alarms: trials in which only the view of the object changed, but the monkey incorrectly released the lever). The hit and false alarm rates were calculated separately for trials in which the viewing angle changed by 0°, 30°, 60° and 90° from the first to the second presentation. The hit rate was calculated in the initial 50 trials of each condition. The false alarm rate was calculated in the same time window. It contained about 100 relevant trials in which the false alarm might occur. The time window was always confined to the first day. We focused the analysis of the monkeys’ performance on the second presentation, where we could uniquely define the change in the viewing angle from the previous to the current object images. At later presentations, before which multiple views of the first object had been presented, the monkeys could compare the current object image with any of the previous images. A proportion of hits significantly larger than that of false alarms would indicate that the monkey discriminated object changes from pure view changes of a single object. No difference between these proportions would indicate no discrimination. Note: Supplementary information is available on the Nature Neuroscience website. ACKNOWLEDGMENTS This research was partly supported by the collaborative research grant of RIKEN Brain Science Institute and the Grant-in-Aid for Scientific Research on Priority Areas (17022047) from the Ministry of Education, Culture, Sports, Science and Technology of Japan. COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/ 1. Rock, I. & DiVita, J. A case of viewer-centered perception. Cognit. Psychol. 19, 280– 293 (1987). 2. Bülthoff, H.H. & Edelman, S. Psychophysical support for a two-dimensional view interpolation theory of object recognition. Proc. Natl. Acad. Sci. USA 89, 60–64 (1992). 3. Edelman, S. & Bülthoff, H.H. Orientation dependence in the recognition of familiar and novel views of 3D objects. Vision Res. 32, 2385–2400 (1992). 4. Humphrey, G.K. & Khan, S.C. Recognizing novel views of three-dimensional objects. Can. J. Psychol. 46, 170–190 (1992). 5. Logothetis, N.K., Pauls, J., Bülthoff, H.H. & Poggio, T. View-dependent object recognition by monkeys. Curr. Biol. 4, 401–414 (1994). 6. Tarr, M.J. Rotating objects to recognize them: a case study on the role of viewpoint dependency in the recognition of three-dimensional objects. Psychon. Bull. Rev. 2, 55–82 (1995).



ARTICLES 7. Földiák, P. Learning invariance from transformation sequences. Neural Comput. 3, 194–200 (1991). 8. Stryker, M.P. Temporal associations. Nature 354, 108–109 (1991). 9. Perrett, D.I., Mistlin, A.J. & Chitty, A.J. Visual neurones responsive to faces. Trends Neurosci. 10, 358–364 (1987). 10. Wallis, G. & Bülthoff, H. Learning to recognize objects. Trends Cogn. Sci. 3, 22–31 (1999). 11. Riesenhuber, M. & Poggio, T. Models of object recognition. Nat. Neurosci. 3, 1199– 1204 (2000). 12. Palmeri, T.J. & Gauthier, I. Visual object understanding. Nat. Rev. Neurosci. 5, 291–303 (2004). 13. Li, L., Miller, E.K. & Desimone, R. The representation of stimulus familiarity in anterior inferior temporal cortex. J. Neurophysiol. 69, 1918–1929 (1993). 14. Sakai, K. & Miyashita, Y. Neuronal tuning to learned complex forms in vision. Neuroreport 5, 829–832 (1994). 15. Logothetis, N.K., Pauls, J. & Poggio, T. Shape representation in the inferior temporal cortex of monkeys. Curr. Biol. 5, 552–563 (1995). 16. Kobatake, E., Wang, G. & Tanaka, K. Effects of shape-discrimination training on the selectivity of inferotemporal cells in adult monkeys. J. Neurophysiol. 80, 324–330 (1998). 17. Sigala, N. & Logothetis, N.K. Visual categorization shapes feature selectivity in the primate temporal cortex. Nature 415, 318–320 (2002). 18. Baker, C.I., Behrmann, M. & Olson, C.R. Impact of learning on representation of parts


and wholes in monkey inferotemporal cortex. Nature 5, 1210–1216 (2002). 19. Perrett, D.I. et al. Neurones responsive to faces in the temporal cortex: studies of functional organization, sensitivity to identity and relation to perception. Hum. Neurobiol. 3, 197–208 (1984). 20. Booth, M.C. & Rolls, E.T. View-invariant representations of familiar objects by neurons in the inferior temporal visual cortex. Cereb. Cortex 8, 510–523 (1998). 21. Perrett, D.I. et al. Visual cells in the temporal cortex sensitive to face view and gaze direction. Proc. R. Soc. B Biol. Sci. 223, 293–317 (1985). 22. Perrett, D.I. et al. Viewer-centred and object-centred coding of heads in the macaque temporal cortex. Exp. Brain Res. 86, 159–173 (1991). 23. Logothetis, N.K. & Sheinberg, D.L. Visual object recognition. Annu. Rev. Neurosci. 19, 577–621 (1996). 24. Biederman, I. & Gerhardstein, P.D. Recognizing depth-rotated objects and conditions for three-dimensional viewpoint invariance. J. Exp. Psychol. 19, 1162–1182 (1993). 25. Biederman, I. & Bar, M. One-shot viewpoint invariance in matching novel objects. Vision Res. 39, 2885–2899 (1999). 26. Edelman, S. Class similarity and viewpoint invariance in the recognition of 3D objects. Biol. Cybern. 72, 207–220 (1995). 27. Vogels, R., Biederman, I., Bar, M. & Lorincz, A. Inferior temporal neurons show greater sensitivity to nonaccidental than to metric shape differences. J. Cogn. Neurosci. 13, 444–453 (2001). 28. Kayaert, G., Biederman, I. & Vogels, R. Shape tuning in macaque inferior temporal cortex. J. Neurosci. 23, 3016–3027 (2003).

1775


E R R AT U M

Erratum: Prior experience of rotation is not required for recognizing objects seen from different angles Gang Wang, Shinji Obama, Wakayo Yamashita, Tadashi Sugihara & Keiji Tanaka Nature Neuroscience 8, 1768–1775 (2005); published online 20 November 2005; corrected after print 11 August 2006 In the version of this article initially published online, there was an error in the page numbers of the web PDF. The error has been corrected in the PDF version of the article.

NATURE NEUROSCIENCE


ARTICLES

Neural correlates of actual and predicted memory formation Yun-Ching Kao1, Emily S Davis1 & John D E Gabrieli1,2 We aimed to discover the neural correlates of subjective judgments of learning—whereby participants judge whether current experiences will be subsequently remembered or forgotten—and to compare these correlates to the neural correlates of actual memory formation. During event-related functional magnetic resonance imaging, participants viewed 350 scenes and predicted whether they would remember each scene in a later recognition-memory test. Activations in the medial temporal lobe were associated with actual encoding success (greater activation for objectively remembered than forgotten scenes), but not with predicted encoding success (activations did not differ for scenes predicted to be remembered versus forgotten). Conversely, activations in the ventromedial prefrontal cortex were associated with predicted but not actual encoding success, and correlated with individual differences in the accuracy of judgments of learning. Activations in the lateral and dorsomedial prefrontal cortex were associated with both actual and predicted encoding success. These findings indicate specific dissociations and associations between the neural systems that mediate actual and predicted memory formation.

A critical aspect of learning is knowing how to learn. Knowing how to learn reflects an interaction between memory processes that encode experience into long-term memory and introspective (or metamemory) processes that evaluate whether information has been learned successfully. Such judgments of learning (JOLs) guide the allocation of cognitive and mnemonic resources so that information that has been sufficiently learned is no longer studied, whereas information that has not yet been successfully learned can be further encoded into long-term memory. Functional neuroimaging studies have delineated neural systems that seem to determine whether information is successfully or unsuccessfully learned during encoding1,2, but there is no evidence as yet about the neural systems that, during study, mediate judgments about whether information has or has not been learned. JOLs are known to influence the success of learning3–5. Studies of JOL ask people to judge, during encoding, whether particular pieces of information or stimuli are successfully or unsuccessfully encoded (that is, whether they are likely to be remembered or forgotten in a later test of retention). Superior JOLs are associated with superior learning: students with higher scholastic performance evaluate their learning and predict their test performance better than students with lower scholastic performance scores3,4. The effects of JOLs on successful learning are most potent in situations when people are given the most opportunity to select study strategies, such as self-paced learning, and least potent when there is minimal opportunity for this, as in experimenter-paced learning6. With training, people can improve their ability to accurately assess what they will remember or forget, especially under self-paced learning circumstances7.

The unknown neural circuitry that mediates JOLs may be the same as that known to mediate learning itself1, or there may be some distinction between memory systems that encode experience and the metamemory systems that evaluate the success of that encoding. Indeed, psychopharmacological and neuropsychological studies indicate that these memory and metamemory processes depend upon at least partially distinct neural circuitry8–10. Nitrous oxide and the benzodiazepine lorazepam both impair memory performance, but do not impair the ability to form accurate JOLs8,9. Participants given lorazepam perform poorly in recalling previously learned word pairs, but actually perform as well as those given a placebo in forming accurate JOLs about the likelihood of later recall9. Neuropsychological studies have also reported dissociations between JOLs and memory performance. Individuals with frontal lobe lesions are impaired at making JOLs, despite normal recognition-memory performance10,11. In one study, individuals with unspecified frontal lesions and those with a variety of posterior lesions memorized a 4 4 matrix of faces10. The participants predicted the number of faces they would later be able to match to the original matrix location. Whereas subjects with right frontal lesions showed better memory performance than did those with right posterior lesions, they were less accurate in making memory predictions. These results suggest that JOLs may be dissociable from actual encoding success for experimenter-paced learning. Furthermore, these results point to the potential importance of prefrontal cortex (PFC) in forming accurate JOLs. Currently, there is no evidence about the neural basis of JOLs in normal healthy adults. Although little is known about the neural systems supporting the prediction of encoding success, there is

1Department

of Psychology, Stanford University, 420 Jordan Hall, Stanford, California 94305, USA. 2Present address: Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, 77 Massachusetts Avenue, Building 46-4033, Cambridge, Massachusetts 02139, USA. Correspondence should be addressed to Y.-C.K. ([email protected]).

Received 13 September; accepted 12 October; published online 13 November 2005; doi:10.1038/nn1595

1776

VOLUME 8

[

NUMBER 12

[


RESULTS Task performance Participants made an average of 144 (s.d. ¼ 51) ‘‘will remember’’ predictions (R) and 104 (s.d. ¼ 43) ‘‘will forget’’ (F) predictions. There was no significant difference in either the propor-


[

Retrieval outcome

Retrieval outcome

Retrieval outcome

Retrieval outcome

mounting evidence about the neural systems that mediate actual encoding success. The ability to successfully encode experiences into long-term memory has been associated with the functioning of medial temporal lobe (MTL) and PFC. MTL structures are essential for the formation of new declarative memWill you remember this at test? ories12,13, and MTL activation is greater during 0 the encoding of scenes1 and words2 that will later be remembered than for those that will later be forgotten. Activations in PFC during 4 encoding also predict subsequent remember1,2,14 Will you remember this at test? ing . One interpretation is that PFC contributes to memory formation by supporting Time (s) 8 semantic elaborations and executive operations that monitor, regulate and facilitate 12 memory processes15. Thus, convergent behavioral evidence from patients with frontal lesions and imaging evidence suggest that Figure 1 Task design. Scenes and fixations were presented for 4 s. For each scene, participants made judgments of learning by predicting whether or not they would remember the scene in a later recognitionJOLs, which involve the self-monitoring of memory test. encoding processes, may also depend on PFC functioning. Although neuropsychological studies have implicated PFC as impor- tion (t15 ¼ 1.90, P 4 0.05) or latency (t15 ¼ 0.91, P 4 0.05) of R versus tant in JOL accuracy, it is unknown which specific subregions support F responses. In the post-scan recognition test, participants made JOLs. One clue comes from a neuropsychological study of a meta- accurate old or new judgments 71% of the time (s.d. ¼ 7%) and memory judgment made at retrieval, the feeling-of-knowing (FOK) had a mean d-prime of 0.83 (s.d. ¼ 0.33). (D-prime is a measure judgment16. In FOK studies, people are asked whether they feel they of recognition memory sensitivity, independent of decision criteria.) would recognize information that they have failed to recall. The Using participants’ responses on the post-scan memory test, we accuracy of the FOK can be gauged by then asking a person to actually sorted trials at study based on predicted encoding success and actual make the recognition memory. Individuals with the most inaccurate encoding success. Items remembered with low confidence were FOK judgments had in common damage to the ventromedial pre- excluded from all analyses to minimize the influence of guessing; frontal cortex (VMPFC)16. The finding that VMPFC is essential in FOK however, items forgotten with low confidence were not excluded metamemory judgment at retrieval raises the possibility that the same because such exclusion would have resulted in an insufficient number brain region participates in JOL metamemory judgments at encoding. of misses for data analysis. During encoding, scenes were given either R In the present study, we used event-related functional magnetic or F predictions and were either subsequently remembered with high resonance imaging (fMRI) to investigate the neural basis of JOLs. confidence (r) or subsequently forgotten (f). Thus, there were four Specifically, we examined whether subjective predictions of encoding possible trial outcomes (Fig. 2a,b): (i) scenes were given a ‘‘will success (JOLs) depend on the same or different neural circuits under- remember’’ prediction and were later remembered (Rr), (ii) scenes lying actual encoding success. Participants were scanned while predict- were given a ‘‘will remember’’ prediction but were later forgotten (Rf), ing whether or not scenes were successfully encoded (that is, whether (iii) scenes were given a ‘‘will forget’’ prediction but were later they would be later remembered or forgotten) (Fig. 1). Afterwards, remembered (Fr) and (iv) trials were given a ‘‘will forget’’ prediction outside the scanner, participants were given an old or new recognition- and were later forgotten (Ff). A 2 2 repeated-measures analysis of memory test. For analysis, items were sorted on the basis of whether variance (ANOVA) revealed no significant main effects for prediction they were given a ‘‘will remember’’ (R) or a ‘‘will forget’’ (F) JOL, and or retrieval outcome, indicating suitable and unbiased measurement whether they were later actually remembered (r) or forgotten (f) during per trial type. the recognition-memory test. This design allowed us to investigate the neural substrates Prediction Prediction Prediction Prediction Prediction a b c d e underlying predicted encoding success (JOLs) R F R F R F R F R F compared to actual encoding success (learning r Rr Fr r 96 35 r Rr Fr r Rr Fr r Rr Fr itself). We found specific brain regions that were related exclusively to actual memory f Rf Ff f 48 69 f Rf Ff f Rf Ff f Rf Ff formation, or exclusively to predicted memory JOL accuracy Actual Trial type Mean Predicted formation, or to both actual and predicted encoding number of encoding memory formation. success responses success Retrieval outcome


ARTICLES

Figure 2 Schematics of the four trial types and trial combinations for statistical analyses. (a) At study, there were two possible JOL predictions: ‘‘will remember’’ (R) and ‘‘will forget’’ (F). At the recognitionmemory test, scenes were either remembered (r) or forgotten (f). (b) Mean numbers of responses for the four trial types. (c) JOL accuracy was assessed by comparing correct JOL trials (gray) to incorrect JOL trials (white). (d) In the actual encoding success analysis, scenes that were later remembered (gray) were compared to scenes that were later forgotten (white). (e) In the predicted encoding success analysis, R predictions (gray) were compared to F predictions (white).

NUMBER 12

[

DECEMBER 2005

1777

ARTICLES Table 1 Brain regions associated with actual and predicted encoding success MNI coordinates Study task


Actual encoding success

Predicted encoding success

x

y

z

Peak Z-score

Cluster size

Region

BA

–22 38

–72 –60

45 –15

4.74 4.72

97 135

L R

Parietal Fusiform

45 –45

6 –66

22 –4

4.34 4.27

11 142

R L

Inferior frontal (lateral PFC) Fusiform

38 –34

–84 30

15 –22

4.07 3.98

78 12

R L

Middle temporal Inferior frontal

Peak Z-score 7 37 44/6 37 19 47

–41

6

26

3.87

5

L

Inferior frontal (lateral PFC)

15 34

–54 –48

15 49

3.47 3.35

5 6

R R

Posterior cingulate Parietal lobule

44/6

–22

12

56

4.67

153

L

Superior frontal

6

–41 22

–66 30

0 –15

4.18 4.07

64 58

L R

Middle occipital Middle frontal

37 11

–22 30

–72 –54

34 34

4.23 4.02

53 49

L R

Precuneus Precuneus

7 7

–11 –45

42 36

–22 19

4.45 4.08

26 17

L L

Inferior frontal (VMPFC) Inferior frontal (lateral PFC)

–15 49

0 –66

–22 –4

3.49 4.79

10 8

L R

Amygdala Inferior temporal

34 37

–41 52

24 –78

–19 8

3.86 3.46

7 5

L R

Inferior frontal Middle temporal

47 39

30 40

11/47 44/6

Only clusters of five or more voxels and a significance of P o 0.001 uncorrected are reported. BA, Brodmann’s area; PFC, prefrontal cortex; VMPFC, ventromedial prefrontal cortex; L, left; R, right.

We calculated JOL accuracy to examine whether participants could reliably assess their learning to predict future retrieval performance. Correct JOL trials consisted of trials in which JOL predictions matched retrieval outcomes (Fig. 2c). Participants made JOL predictions with above-chance accuracy (mean ± s.d. 64 ± 5%). To investigate whether participants could discriminate between well-learned and

poorly-learned items, the Goodman-Kruskal gamma was calculated for each participant. The gamma statistic indicates the strength of association between two ordinal variables and is the preferred measure in metamemory research17. The mean gamma score was 0.53 (s.d. ¼ 0.04) and was statistically greater than zero (t15 ¼ 12.64, P o 0.001), indicating that participants were able to discriminate between items

Table 2 Distinct and overlapping regions between actual and predicted encoding success MNI coordinates Study task

x

y

z

Actual 4 predicted

–26

–48

–26 30

–66 –42

–52 56 30 –11 0

Predicted 4 actual

Actual ¼ predicted

Peak Z-score

Cluster size

Region

BA

–11

4.62

35

L

Fusiform/parahippocampal

37

–11 –19

4.17 4.16

5 44

L R

Posterior cingulate Fusiform/parahippocampal

31 37

6

15

3.73

5

L

Precentral

44

–24 12

–4 –19

3.71 3.68

11 17

R R

Superior temporal Inferior frontal

21 47

42 54

–26 8

3.61 3.61

5 7

L L

Orbital frontal (VMPFC) Medial frontal (DMPFC)

11 10

4 –26

42 12

–4 64

3.40 3.32

7 5

R L

Anterior cingulate Middle frontal

32 6

–41 –22

6 –72

26 41

3.87 4.23

18 32

L L

Inferior frontal (lateral PFC) Superior parietal lobule

–45 26

–66 –72

4 34

4.27 3.49

44 28

L R

Inferior temporal gyrus Precuneus

37 19

–22

6

49

3.18

14

L

Cingulate gyrus

32

44/6 7

Only clusters of five or more voxels and a significance of P o 0.005 uncorrected are reported. BA, Brodmann’s area; VMPFC, ventromedial prefrontal cortex; DMPFC, dorsomedial prefrontal cortex; PFC, prefrontal cortex; L, left; R, right.

1778

VOLUME 8

[

NUMBER 12

[


1.5

Rr Rf Fr Ff

1 0.5 0 –0.5

–2

0

2 4 6 Time (s)

8

10

Signal change (%)

Right MTL

Actual > predicted

Signal change (%)

a

1.2 1 0.8 0.6 0.4 0.2 0

Rr Rf Fr Ff

Trial types

Rr Rf Fr Ff

1 0.5 0 –0.5

–2

0

2 4 6 Time (s)

8

10

Predicted > actual

1.2 1 0.8 0.6 0.4 0.2 0

Rr Rf Fr Ff

Trial types

Rr Rf Fr Ff

0.2 0.1 0 –0.1 –2

0

2

–0.2

4

6

8

10

Time (s)

Signal change (%)

Left VMPFC Signal change (%)

b

1.5

Signal change (%)

Signal change (%)

Right MTL

0.15 0.1 0.05 0 –0.05 –0.1

Rr Rf Fr Ff

Trial types

–0.3

0.8 0.6 0.4 0.2 0 –0.2 –2

0.2 –2

0

2

4

6

8

10

Rr Rf Fr Ff

–0.8 Time (s)

Signal change (%)

Left DMPFC Signal change (%)

Figure 3 Statistical activation maps and percent signal change. Activation maps are rendered onto the MNI normalized canonical single-subject brain. Line graphs represent the percent signal change in brain activation as a function of time, for each of the four trial types. Bar graphs represent the mean percent signal change from 4 s to 8 s after stimulus presentation, for each of the four trial types. Error bars indicate s.e.m. R, ‘‘will remember’’ predictions; F, ‘‘will forget’’ predictions; r, later remembered; f, later forgotten. (a) Regions of interest (ROIs) defined from actual 4 predicted encoding success contrast. In bilateral medial temporal lobe (MTL), only the main effect of actual encoding success was significant (solid lines above dotted lines). (b) ROIs defined from predicted 4 actual encoding success contrast. Ventromedial prefrontal cortex (VMPFC) and dorsomedial prefrontal cortex (DMPFC) showed a significant main effect for predicted encoding success (red lines above blue lines) but not actual encoding success. However, the main effect for actual encoding success showed a trend toward significance in DMPFC. (c) ROIs defined from regions in which predicted encoding success matched actual encoding success. Left lateral PFC showed significant main effects for both actual and predicted encoding success. Coordinates in Table 2.

Signal change (%)

Rr Rf Fr Ff

0.2 0 –0.2 –0.4 –0.6 Trial types

c Actual = predicted that would be later remembered and items that would be later forgotten with abovechance accuracy. Response times for correct and incorrect JOLs differed such that correct JOL predictions were made faster than incorrect JOL predictions (mean response times of 1,509 and 1,574 ms, respectively; t15 ¼ 2.73, P o 0.05). Pearson’s correlation revealed a significant correlation between participants’ JOL accuracy (gamma score) and their recognition-memory performance (d-prime) (r15 ¼ 0.52, P o 0.05). Neural correlates of actual encoding success To assess the neural correlates of actual encoding success, we identified regions in which activation was significantly greater during encoding for scenes subsequently remembered with high confidence than for scenes subsequently forgotten (Rr and Fr trials 4 Rf and Ff trials) (Fig. 2d). Activation was significantly greater for subsequently remembered scenes than for subsequently forgotten scenes in bilateral MTL regions including posterior parahippocampal and fusiform gyri, but not in the hippocampus proper. In addition, there was greater activation for remembered than for forgotten scenes in bilateral inferior frontal gyrus (lateral PFC) corresponding to Brodmann’s area 44/6 and in right posterior cingulate gyrus (Table 1). There were no significant activations for the opposite contrast (subsequently forgotten scenes compared to subsequently remembered scenes). Neural correlates of predicted encoding success We assessed predicted encoding success by comparing R predictions to F predictions across actual retrieval outcomes (Fig. 2e). This contrast yielded significant activations (P o 0.001) in left lateral PFC, left VMPFC corresponding to Brodmann’s area 11/47, left amygdala, right


[

NUMBER 12

[

DECEMBER 2005

Left lateral PFC Rr Rf Fr Ff

0

2

4 6 Time (s)

8

10

Signal change (%)


ARTICLES

Rr Rf Fr Ff

0.6 0.5 0.4 0.3 0.2 0.1 0 Trial types

middle temporal and right inferior temporal regions, bilateral precuneus and left middle occipital regions (Table 1). The opposite contrast, comparing F prediction trials to R predictions trials, did not reveal any significant brain activations. Comparing actual and predicted encoding success To identify brain regions more specifically related to either actual encoding success or predicted encoding success, we carried out paired t-tests using the two contrasts discussed above. Performing a paired t-test is analogous to performing a linear contrast comparing Fr trials (which are associated with actual encoding success but not predicted encoding success) to Rf trials (which are associated with predicted encoding success but not actual encoding success). The brain regions with greater activation for actual encoding success than predicted encoding success included bilateral MTL and left posterior cingulate (Table 2). The bilateral MTL regions included parahippocampal and fusiform gyri (Fig. 3a). The bilateral MTL regions were submitted to regions-of-interest (ROI) analyses. Peak percent signal change for all four trial types were entered into 2 2 repeated-measures ANOVA to assess the relationship between actual encoding success (retrieval outcome) and predicted encoding success (JOL predictions). The ANOVA revealed a significant main effect for actual encoding success in both the right and left MTL regions (right MTL, F1,15 ¼ 25.37, P o 0.001; left MTL,

1779

Individual differences in JOL accuracy In addition to identifying separable networks for actual and predicted encoding success, we were also interested in assessing which brain regions support the ability to make accurate predictions. The contrast between correct and incorrect predictions did not yield any significant clusters. Therefore, to investigate whether differences in brain activations might reflect individual differences in JOL accuracy, we submitted the four ROIs identified in the previous analyses to correlational analyses: MTL regions identified in the actual 4 predicted encoding success analysis, VMPFC and DMPFC regions identified in the predicted 4 actual encoding success analysis and the left lateral PFC region identified in the predicted ¼ actual encoding success analysis. Of the four ROIs, only VMPFC activation was significantly correlated with variation in JOL accuracy (Fig. 4). Specifically, gamma scores were positively correlated with VMPFC activation on trials given accurate predictions (Fig. 4a; Rr trials, r15 ¼ 0.58, P o 0.01; Ff trials, r15 ¼ 0.43, P o 0.05). However, gamma scores were not correlated with activation on trials given inaccurate predictions (Rf and Fr trials) (Fig. 4b). Thus, individuals who made more accurate JOL predictions showed greater VMPFC activations. Additional correlation analyses between brain

1780

Accurate predictions

Inaccurate predictions

JOL accuracy (gamma)

F1,15 ¼ 15.58, P o 0.001). Subsequently remembered trials showed greater activation than did subsequently forgotten trials (Fig. 3a). However, neither the main effect for predicted encoding success nor the interaction effect reached significance. This suggests that these MTL regions are sensitive to actual encoding success and not to predicted encoding success. Conversely, several brain regions were more activated for predicted than actual encoding success. These regions included prefrontal cortices, specifically left VMPFC, left dorsomedial prefrontal cortex (DMPFC) corresponding to Brodmann’s area 10, right inferior frontal gyrus, left precentral gyrus and right anterior cingulate gyrus (Table 2). In the ROI analysis for VMPFC and DMPFC (Fig. 3b), there was a significant main effect for prediction, such that R predictions resulted in more activation than did F predictions (VMPFC, F1,15 ¼ 28.54, P o 0.001; DMPFC, F1,15 ¼ 11.57, P o 0.001). The main effect of actual encoding success and the interaction effect did not reach significance, although the main effect of actual encoding success showed a trend toward significance in DMPFC (F1,15 ¼ 3.71, P ¼ 0.07). Significant effects in VMPFC reflected increases in activation relative to baseline, whereas effects in DMPFC reflected decreases in activation relative to baseline18. VMPFC seemed to have a role in predicted encoding success, but not in actual encoding success. In contrast, DMPFC seemed to respond to both predicted and actual encoding success. In the above analyses, no region showed greater activation for correct predictions than for incorrect predictions (Rr and Ff trials 4 Rf and Fr trials). We performed a voxelwise analysis to search for any such interactions, and neither this nor the reverse interaction yielded significant activations (even at a lowered threshold of P o 0.01). We used a masking procedure to identify brain regions associated with both actual and predicted encoding success. Both the actual and predicted encoding success contrasts showed significant activations in left lateral PFC, left superior parietal lobule, bilateral inferior temporal gyrus and bilateral precuneus (Table 2), suggesting that these regions may mediate both actual and predicted encoding success. In the ROI analysis of left lateral PFC (Fig. 3c), there were significant main effects for both actual and predicted encoding success (F1,15 ¼ 27.22, P o 0.001; and F1,15 ¼ 13.38, P o 0.001, respectively; the interaction was not significant, F1,15 ¼ 0.109, P ¼ 0.75). The pattern of activation from the left lateral PFC region was representative of the other significant clusters.

1.0

JOL accuracy (gamma)


ARTICLES

1.0

Rr

1.0

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

r = 0.58*

0.0 –0.4 –0.2 0.0

0.2

0.4

Ff

0.6

Rf

r = 0.43*

0.0 –0.4 –0.2 0.0 1.0

0.8

0.8

0.6

0.6

0.2

0.4

0.6

Fr

0.4

0.4 r = –0.12*

0.2

0.0 –0.4 –0.2 0.0 0.2 0.4 Signal change (%)

0.6

0.2

r = –0.21

0.0 –0.4 –0.2 0.0 0.2 0.4 Signal change (%)

0.6

Figure 4 Ventromedial prefrontal cortex (VMPFC) and individual differences in JOL accuracy. The scatter diagrams are plots of individual gamma scores of JOL accuracy as a function of percent signal change in VMPFC, for each of the four trial types. R, ‘‘will remember’’ predictions; F, ‘‘will forget’’ predictions; r, later remembered; f, later forgotten. *P o 0.05.

activations and recognition-memory performance, as indexed by d-prime, did not reach significance for any of the four ROIs. DISCUSSION The present study investigated the neural basis of judgments of learning (JOLs) by asking participants to study scenes and predict whether or not each scene would be later remembered. The goals of this study were to examine the brain regions supporting the predictions of encoding success (participants’ judgments that they had successfully learned a scene) and compare those regions to brain regions supporting actual encoding success (whether participants actually remembered that scene when tested). We identified four patterns of brain-behavior relationships in brain regions known to be important for memory, thought and introspection. First, actual encoding success engaged MTL regions (that is, activations were greater during the encoding of subsequently remembered scenes than that of subsequently forgotten scenes). MTL activations were unrelated to the predicted encoding success. This suggests that the MTL processes engaged during encoding reflect memory but not JOL processes. Second, predicted encoding success (that is, greater activation for the scenes that subjects predicted they would remember than for the scenes subjects predicted they would forget) engaged a network of PFC regions including medial regions such as VMPFC and DMPFC. This suggests that medial PFC processes engaged during encoding may support JOL processes. Third, JOL accuracy—the strength of the relation between predicted and actual encoding success—may involve processes mediated by VMPFC: participants who made more accurate JOL predictions showed greater VMPFC activations than participants who made less accurate predictions. Fourth, both actual and predicted encoding success engaged left lateral PFC. Thus, there were both dissociations and associations between the neural systems that seemed to be engaged by objective memory encoding and subjective appraisals of that encoding. MTL and actual encoding success MTL regions seemed to support actual encoding success, but not the subjective evaluation of encoding success. There is substantial evidence that the hippocampus and surrounding cortices are crucial for declarative memory formation12,13. Individuals such as H.M. who have

VOLUME 8

[

NUMBER 12

[



ARTICLES extensive MTL injury show a severe anterograde amnesia that includes an apparent inability to successfully encode new experiences13. In normal adults, neuroimaging studies have frequently shown MTL activation during episodic encoding tasks, with greater activation for items successfully than for those unsuccessfully encoded1,2. For complex scenes, MTL activations associated with successful encoding have sometimes included the hippocampus19, but have typically been most robust in the parahippocampal and fusiform gyri1,14. This may reflect parahippocampal specialization for visual-spatial memory20. Thus, the current finding—that activations in the parahippocampal and anterior fusiform cortices are associated with successful scene encoding—is consistent with prior imaging findings. Unexpected and new, however, is the discovery that MTL regions seemed not to be involved in predicting encoding success (that is, there was no difference in activation for items that had a ‘‘will be remembered’’ prediction compared to items that had a ‘‘will be forgotten’’ prediction). The current imaging study examined the neural correlates of subjective metamemory judgments during encoding, but a number of studies have found MTL activation associated with subjective judgments during retrieval. In one study, hippocampal activation was found for stimuli that were consciously remembered, relative to stimuli that were accurately judged as having been studied but seemed merely familiar21. In another study, hippocampal activation at retrieval occurred equally for items thought to have been studied whether they were actually studied or not, whereas parahippocampal activations did differentiate between whether items were actually studied or not22. The present findings, however, suggest that MTL regions, despite their essential involvement in actual encoding, may not be involved in the subjective evaluation of encoding success. MTL, and especially hippocampal, activations associated with the subjective evaluations of retrieval may reflect a different role for MTL structures at encoding versus retrieval. Alternatively, other regions associated with retrieval evaluation, such as prefrontal23,24 or parietal cortices25, may actually mediate subjective evaluation at retrieval. Medial PFC and predicted encoding success Several regions in PFC, such as VMPFC and DMPFC, were engaged during the prediction of encoding success. JOLs involve introspective processes about the mental states associated with successful or unsuccessful memory formation, and it has been proposed that medial PFC supports the ability to represent mental states of the self and others26. Accordingly, neuroimaging studies have found medial PFC activations when people process information about themselves relative to other people27,28. Thus, the medial PFC activations in the present study that were associated with predictions of encoding success may reflect processes involved in the introspective evaluation of internal mental states. In addition to differentiating between JOL predictions, activations in DMPFC showed a trend toward a significant main effect for actual encoding success. The ROI analyses revealed that DMPFC activations decreased compared to baseline, consistent with other studies that find deactivations in medial PFC18. Greater deactivations were associated with subsequently remembered items than with subsequently forgotten items. This is consistent with the interpretation that deactivations for items later remembered reflect the successful disengagement of ongoing baseline mental processes that are task irrelevant and the allocation of neurocognitive resources to the task at hand29. For predicted encoding success, items given ‘‘will forget’’ predictions showed greater deactivation than did items given ‘‘will remember’’ predictions, perhaps reflecting participants’ reallocation of resources in attempts to remember items they deemed harder to remember.


[

NUMBER 12

[

DECEMBER 2005

VMPFC and JOL accuracy Only VMPFC activation correlated with individual differences in JOL accuracy. Ever since the case of Phineas Gage, who sustained bilateral VMPFC damage30, the VMPFC has been thought to play an important role in judgment and decision-making by integrating somatic markers or emotions with goals and expectations31. VMPFC is also involved in determining whether information is contextually appropriate, the noetic feeling of ‘rightness’32. For judgments about memory, VMPFC may support accurate JOL predictions by forming an internal model of what constitutes successful learning and comparing information available from the current state of learning to this internal model33. Cognitive theories have proposed that accurate JOLs result from monitoring information that is predictive of later remembering, such as encoding strategy, list position, word-pair relatedness or length of study time34. However, people can also use information that is not predictive of future remembering when forming their JOLs. This can lead to a false sense of successful encoding and thereby lead to inaccurate predictions. For instance, speed of encoding is a salient cue that increases people’s confidence that an item will be later remembered (JOL magnitude), but the speed of encoding does not predict differences in future memory retrieval35,36. One possible role of the VMPFC is to flexibly evaluate reliable indicators of encoding success to form accurate JOLs. Future research is needed to determine whether individual differences in JOL accuracy reflect transient changes during task performance (such as effort, alertness or JOL strategy)37 or stable differences in metamemory abilities3,4. The VMPFC has also been shown to support other metamemory judgments, such as the FOK judgments made at retrieval. Among individuals with PFC damage, those with the most inaccurate FOKs had damage to VMPFC16. In an fMRI study with normal adults, greater FOK accuracy was associated with greater VMPFC activation38. The VMPFC is also implicated in predictions about outcomes other than memory performance. For instance, individuals with orbitofrontal lesions are unable to weigh advice and make predictions about economic outcome39. These convergent lesion and imaging findings support the interpretation that the quality of the internal monitoring of memory success or failure depends upon processes mediated by VMPFC. Lateral PFC, and actual and predicted encoding success Whereas MTL and VMPFC were differentially involved in either actual or predicted encoding success, lateral PFC activation occurred for both actual and predicted encoding success. In broad terms, this is consistent with evidence that lateral PFC contributes to successful memory encoding2,14, perhaps through semantic elaboration and organizational strategies that enhance successful encoding40. It might, however, have been expected that a brain region activated by both actual and predicted encoding success would also support the accuracy of JOL predictions. Our results identified several brain regions, such as lateral PFC, that tracked both actual and predicted encoding success; however, none of these regions correlated with JOL accuracy. The finding that lateral PFC tracked predicted encoding success is consistent with results from other studies comparing objective and subjective memory processes23,24,41,42. In studies of subjective judgments made during retrieval, lateral PFC activations varied with the strength of FOK judgments (analogous to the ‘‘will remember’’ versus ‘‘will forget’’ JOL judgments), but not with FOK accuracy across participants23,24. During FOKs, participants are bringing online the information associated with the target item43. Graded activations in lateral PFC may be associated with the amount of partial retrieval of the target item brought into working memory24. Lateral PFC may have a

1781


ARTICLES similar role in JOLs, such that the degree of activation reflects how powerfully the information to be learned is held in working memory and contributes to the subjective experience of how successfully an item is learned. Lateral PFC activations may also reflect an increased encoding effort associated with ‘‘will remember’’ predictions; this is consistent with evidence that the magnitude of lateral PFC activation during encoding varies with both encoding effort and subsequent retrieval outcome, whereas MTL activations vary only with subsequent retrieval outcome42. A fundamental issue in human memory research is the relationship between objective (veridical) and subjective (experiential) dimensions of memory. The relationship between objective and subjective memory processes can be studied during learning (memory encoding at study) or retrieval (recall or recognition at test). A number of functional neuroimaging studies have examined the relationship between objective and subjective memory processes at retrieval16,23,24,38,41. The present study reports initial evidence about the neural substrates of subjective memory during learning. Such subjective or metamemory processes during learning are of particular interest, because these processes can actually enhance the effectiveness of learning by guiding the allocation of resources at a time when information remains available for learning. Indeed, in the present study, greater accuracy in JOLs (including greater accuracy in predicting both future remembering and future forgetting) was correlated with greater accuracy in recognition memory and greater VMPFC activation. Thus, these processes, and the neural circuits that mediate them, constitute a critical component of the way in which knowing how to learn empowers learning itself. METHODS Participants. Six female and 14 male native English speakers between the ages of 19 and 25 (mean ± s.d.¼ 21 ± 2) participated in the study. Participants were right-handed, with normal or corrected vision, and were without any neurological or psychiatric conditions or structural brain abnormalities. We discarded data from four participants because they made excessive head movements during data acquisition. We obtained informed consent from all participants according to the requirements of the Panel on Human Participants in Medical Research at Stanford University. Stimulus materials. Materials consisted of 700 pictures of indoor and outdoor scenes. Pictures were randomly distributed into ten lists of 70 pictures. Five lists with a total of 350 scenes were presented during study, and the remaining five lists were presented as foils during the recognition-memory test. Lists were counterbalanced across participants such that all lists were presented as old and new items. Task procedure. In a rapid event-related design, we scanned participants while they studied 350 scenes randomly intermixed with 350 fixations (Fig. 1). Participants were explicitly instructed to memorize the scenes for an upcoming memory test. Scenes were presented for 3 s with an inter-stimulus interval of 1 s. For each scene, participants made JOL predictions about the likelihood of remembering the scene during the memory test. The instructions (‘‘Will you remember this at test?’’) appeared on the bottom of the screen, prompting participants to make ‘‘will remember’’ or ‘‘will forget’’ JOLs. Participants were instructed to press a button with their right index finger to indicate a ‘‘will remember’’ prediction and with their middle finger to indicate a ‘‘will forget’’ prediction. After scanning, participants were given a recognition test consisting of 350 old and 350 new pictures. For each trial, participants made two judgments: (i) whether the item was old or new and (ii) whether their judgment was made with high or low confidence. In this self-paced recognition test, each trial lasted a maximum of 8 s. Imaging procedure. Magnetic resonance imaging was performed using a 3-T GE Signa scanner. Before functional imaging, a spin-echo T1-weighted

1782

anatomical image was acquired (30 coronal slices; slice thickness ¼ 6 mm; TE ¼ 30 ms; TR ¼ 2,000 ms; field of view ¼ 24 24 cm2). A shim procedure was used to improve B0 magnetic field homogeneity. Functional images were then obtained in the same slice location as the anatomical images using a T2*sensitive two-dimensional gradient-echo spiral-in/out sequence. Scenes were presented over five scanning sessions lasting approximately 9 min each. Bite bars made out of dental compress were used to restrict head movement. Imaging analyses. Imaging data were preprocessed and analyzed using SPM99. We corrected for differences in the acquisition time of the functional images and then performed a motion correction using sinc interpolation. The T2* anatomical image was co-registered to the mean functional image that was created during motion correction. The anatomical image was then segmented into gray matter, white matter and cerebral spinal fluid. Anatomical and functional images were spatially normalized based on parameters determined by normalizing the segmented gray matter image to a gray matter template from the MNI series using a 12-parameter affine transformation. Finally, images were resampled into 3.75 6 3.75 mm voxels and spatially smoothed with an isotropic Gaussian kernel of 7 mm full-width at half maximum (FWHM). Statistical models were constructed for individual participants using a general linear model. Regressor functions were constructed for each of the four trial types (Rr, Rf, Fr and Ff). Trials were modeled as events assuming a canonical hemodynamic response function. Subject-specific effects were estimated using a fixed-effects model. Linear contrasts were computed to generate subject-specific SPM(t) contrasts representing statistical differences in brain activation between conditions. Contrasts constructed at the single participant level were then input into a second-level group analysis using a random-effects model. Group contrasts were constructed by using a one-sample t-test. All reported clusters survived a P-threshold of 0.001 (uncorrected for multiple comparisons) and consisted of five or more significant voxels. To identify voxels that differed between group-level contrasts for actual and predicted encoding success, we conducted one-tailed paired t-tests of the contrasts. Reported clusters for the paired t-test survived a P-threshold of 0.005 and consisted of five or more significant voxels. In addition, the contrast for actual encoding success was masked by the contrast for predicted encoding success to identify voxels that were significant in both contrasts. ROI analyses were employed to characterize the statistical effects of each of the four trial types. ROIs were functionally defined and included all significant voxels in the cluster. Data were extracted by selective averaging with respect to peristimulus time out to 10 s after stimulus onset. ROI data are expressed as percent signal change calculated by taking the average signal from 4 s to 8 s after stimulus onset. These data were then subjected to a repeated-measures ANOVA and Pearson’s r correlations. Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTS The authors thank S. Gabrieli, P. Mazaika, J. Cooper, A.R. Preston and P. SokolHessner for their assistance or comments. This research was sponsored by grants from the US National Institute of Mental Health to Y.-C.K. (MH073234) and J.D.E.G. (MH59940). COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/

1. Brewer, J.B., Zhao, Z., Desmond, J.E., Glover, G.H. & Gabrieli, J.D. Making memories: brain activity that predicts how well visual experience will be remembered. Science 281, 1185–1187 (1998). 2. Wagner, A.D. et al. Building memories: remembering and forgetting of verbal experiences as predicted by brain activity. Science 281, 1188–1191 (1998). 3. King, J.F., Zechmeister, E.B. & Shaughnessy, J.J. Judgments of knowing: the influence of retrieval practice. Am. J. Psychol. 93, 329–343 (1980). 4. Maki, R.H. & Berry, S.L. Metacomprehension of text material. J. Exp. Psychol. Learn. Mem. Cogn. 10, 663–679 (1984).

VOLUME 8

[

NUMBER 12

[



ARTICLES 5. Thiede, K.W., Anderson, M.C. & Therriault, D. Accuracy of metacognitive monitoring affects learning of texts. J. Educ. Psychol. 95, 66–73 (2003). 6. Mazzoni, G. & Cornoldi, C. Strategies in study time allocation: why is study time sometimes not effective? J. Exp. Psychol. Gen. 122, 47–60 (1993). 7. Dunlosky, J., Kubat-Silman, A.K. & Hertzog, C. Training monitoring skills improves older adults’ self-paced associative learning. Psychol. Aging 18, 340–345 (2003). 8. Dunlosky, J. et al. Inhalation of 30% nitrous oxide impairs people’s learning without impairing people’s judgments of what will be remembered. Exp. Clin. Psychopharmacol. 6, 77–86 (1998). 9. Izaute, M. & Bacon, E. Specific effects of an amnesic drug: effect of lorazepam on study time allocation and on judgment of learning. Neuropsychopharmacology 30, 196–204 (2005). 10. Vilkki, J., Surma-aho, O. & Servo, A. Inaccurate prediction of retrieval in a face matrix learning task after right frontal lobe lesions. Neuropsychology 13, 298–305 (1999). 11. Kennedy, M.R. & Yorkston, K.M. Accuracy of metamemory after traumatic brain injury: predictions during verbal learning. J. Speech Lang. Hear. Res. 43, 1072– 1086 (2000). 12. Scoville, W.B. & Milner, B. Loss of recent memory after bilateral hippocampal lesions. J. Neurol. Neurosurg. Psychiatry 20, 11–21 (1957). 13. Corkin, S. What’s new with the amnesic patient H.M.? Nat. Rev. Neurosci. 3, 153–160 (2002). 14. Kirchhoff, B.A., Wagner, A.D., Maril, A. & Stern, C.E. Prefrontal circuitry for episodic encoding and subsequent memory. J. Neurosci. 20, 6173–6180 (2000). 15. Fletcher, P.C. & Henson, R.N. Frontal lobes and human memory: insights from functional neuroimaging. Brain 124, 849–881 (2001). 16. Schnyer, D.M. et al. A role for right medial prefrontal cortex in accurate feeling-ofknowing judgments: evidence from patients with lesions to frontal cortex. Neuropsychologia 42, 957–966 (2004). 17. Nelson, T.O. A comparison of current measures of the accuracy of feeling-of-knowing predictions. Psychol. Bull. 95, 109–133 (1984). 18. Gusnard, D.A., Akbudak, E., Shulman, G.L. & Raichle, M.E. Medial prefrontal cortex and self-referential mental activity: relation to a default mode of brain function. Proc. Natl. Acad. Sci. USA 98, 4259–4264 (2001). 19. Stern, C.E. et al. The hippocampal formation participates in novel picture encoding: evidence from functional magnetic resonance imaging. Proc. Natl. Acad. Sci. USA 93, 8600–8665 (1996). 20. Epstein, R. & Kanwisher, N. A cortical representation of the local visual environment. Nature 392, 598–601 (1998). 21. Eldridge, L.L., Knowlton, B.J., Furmanski, C.S., Bookheimer, S.Y. & Engel, S.A. Remembering episodes: a selective role for the hippocampus during retrieval. Nat. Neurosci. 3, 1149–1152 (2000). 22. Cabeza, R., Rao, S.M., Wagner, A.D., Mayer, A.R. & Schacter, D.L. Can medial temporal lobe regions distinguish true from false? An event-related functional MRI study of veridical and illusory recognition memory. Proc. Natl. Acad. Sci. USA 98, 4805–4810 (2001). 23. Kikyo, H., Ohki, K. & Miyashita, Y. Neural correlates for feeling-of-knowing: an fMRI parametric analysis. Neuron 36, 177–186 (2002).


[

NUMBER 12

[

DECEMBER 2005

24. Maril, A., Simons, J.S., Mitchell, J.P., Schwartz, B.L. & Schacter, D.L. Feeling-ofknowing in episodic memory: an event-related fMRI study. Neuroimage 18, 827–836 (2003). 25. Wheeler, M.E. & Buckner, R.L. Functional dissociation among components of remembering: control, perceived oldness, and content. J. Neurosci. 23, 3869–3880 (2003). 26. Frith, C.D. & Frith, U. Interacting minds–a biological basis. Science 286, 1692–1695 (1999). 27. Kelley, W.M. et al. Finding the self? An event-related fMRI study. J. Cogn. Neurosci. 14, 785–794 (2002). 28. Schmitz, T.W., Kawahara-Baccus, T.N. & Johnson, S.C. Metacognitive evaluation, selfrelevance, and the right prefrontal cortex. Neuroimage 22, 941–947 (2004). 29. Daselaar, S.M., Prince, S.E. & Cabeza, R. When less means more: deactivations during encoding that predict subsequent memory. Neuroimage 23, 921–927 (2004). 30. Damasio, H., Grabowski, T., Frank, R., Galaburday, A.M. & Damasio, A.R. The return of Phineas Gage: clues about the brain from the skull of a famous patient. Science 264, 1102–1105 (1994). 31. Tranel, D. Emotion, decision making, and the ventromedial prefrontal cortex. in Principles of Frontal Lobe Function (eds. Stuss, D.T. & Knight, R.T.) Ch. 22, 338–353 (Oxford University Press, London, 2002). 32. Moscovitch, M. & Winocur, G. The frontal cortex and working with memory. in Principles of Frontal Lobe Function (eds. Stuss, D.T. & Knight, R.T.) Ch. 12, 188–209 (Oxford University Press, London, 2002). 33. Nelson, T.O. & Narens, L. Metamemory: a theoretical framework and new findings. Psychol. Learn. Motiv. 26, 125–141 (1990). 34. Koriat, A. Monitoring one’s own knowledge during study: a cue-utilization approach to judgments of learning. J. Exp. Psychol. Gen. 126, 349–370 (1997). 35. Hertzog, C., Dunlosky, J., Robinson, A.E. & Kidder, D.P. Encoding fluency is a cue used for judgments about learning. J. Exp. Psychol. Learn. Mem. Cogn. 29, 22–34 (2003). 36. Benjamin, A.S., Bjork, R.A. & Schwartz, B.L. The mismeasure of memory: when retrieval fluency is misleading as a metamnemonic index. J. Exp. Psychol. Gen. 127, 55–68 (1998). 37. Kelemen, W.L., Frost, P.J. & Weaver, C.A., III. Individual differences in metacognition: evidence against a general metacognitive ability. Mem. Cognit. 28, 92–107 (2000). 38. Schnyer, D.M., Nicholls, L. & Verfaellie, M. The role of VMPC in metamemorial judgments of content retrievability. J. Cogn. Neurosci. 17, 832–846 (2005). 39. Gomez-Beldarrain, M., Harries, C., Garcia-Monco, J.C., Ballus, E. & Grafman, J. Patients with right frontal lesions are unable to assess and use advice to make predictive judgments. J. Cogn. Neurosci. 16, 74–89 (2004). 40. Petrides, M. Specialized systems for the processing of mnemonic information within the primate frontal cortex. Phil. Trans. R. Soc. Lond. B 351, 1455–1461 (1996). 41. Chua, E.F., Rand-Giovannetti, E., Schacter, D.L., Albert, M.S. & Sperling, R.A. Dissociating confidence and accuracy: functional magnetic resonance imaging shows origins of the subjective memory experience. J. Cogn. Neurosci. 16, 1131–1142 (2004). 42. Reber, P.J. et al. Neural correlates of successful encoding identified using functional magnetic resonance imaging. J. Neurosci. 22, 9541–9548 (2002). 43. Schacter, D.L. & Worling, J.R. Attribute information and the feeling of knowing. Can. J. Psychol. 39, 467–475 (1985).

1783


ARTICLES

Cognitive control mechanisms resolve conflict through cortical amplification of task-relevant information Tobias Egner & Joy Hirsch A prominent model of how the brain regulates attention proposes that the anterior cingulate cortex monitors the occurrence of conflict between incompatible response tendencies and signals this information to a cognitive control system in dorsolateral prefrontal cortex. Cognitive control is thought to resolve conflict through the attentional biasing of perceptual processing, emphasizing task-relevant stimulus information. It is not known, however, whether conflict resolution is mediated by amplifying neural representations of task-relevant information, inhibiting representations of task-irrelevant information, or both. Here we manipulated trial-by-trial levels of conflict and control during a Stroop task using face stimuli, while recording hemodynamic responses from human visual cortex specialized for face processing. We show that, in response to high conflict, cognitive control mechanisms enhance performance by transiently amplifying cortical responses to task-relevant information rather than by inhibiting responses to task-irrelevant information. These results implicate attentional target-feature amplification as the primary mechanism for conflict resolution through cognitive control.

To meet changing environmental demands, humans make rapid, strategic adjustments to how they deploy their attentional resources1,2, such that when we encounter increasing task difficulty, we tend to refocus our attention on task-relevant aspects of our surroundings while ignoring less relevant ones. However, it is not fully understood how the human brain is able to swiftly adjust processing priorities in response to changing circumstances. In the laboratory, the strategic control of attention to optimize performance is captured by ‘conflict-adaptation’ effects in classic selective-attention tasks3,4. For example, in the Stroop task5,6, subjects are required to name the ink color of a printed color name (the ‘target’ dimension of the stimulus), while ignoring the word’s meaning (the ‘distracter’ dimension of the stimulus). When target and distracter dimensions are incongruent (for example, the word RED printed in green ink), they induce conflicting response tendencies, and reaction times are slowed in comparison to trials where target and distracter information is congruent (for example, the word GREEN printed in green ink). However, this deleterious effect of incongruent distracters on the processing of target information is reduced (and often abolished entirely) after incongruent trials, as compared to after congruent trials7,8. This suggests that high conflict in an incongruent trial leads to a transient upregulation of selective attention in anticipation of the next trial, resulting in improved conflict resolution: that is, conflict adaptation4,9. It has been proposed that such context-sensitive regulation of attentional resources is mediated by a specialized conflict-monitoring system that gauges co-activation in the processing pathways associated with incompatible responses9–12. When conflict is detected, the conflict

monitor triggers a ‘cognitive control’ system that is assumed to resolve the conflict through the attentional biasing of perceptual processes7,9,13. To illustrate, in the context of the Stroop task, an incongruent stimulus would elicit incompatible response tendencies, leading the conflict monitor to alert the cognitive control system to the need for conflict resolution. The cognitive control system would then deploy selective attention mechanisms to bias perceptual processing toward taskrelevant stimulus properties and away from task-irrelevant, distracting stimulus properties, by modulating activity in the visual pathways involved in extracting target and distracter features of the stimulus. To dissociate the neural correlates of the conflict-monitoring and cognitive control systems, human neuroimaging studies have exploited a particularly attractive feature of conflict adaptation: namely, the possibility of comparing identical incongruent trials on the basis of whether they are associated with low control (the incongruent trial follows a congruent one) or with high control (the incongruent trial follows an incongruent one)7,8,12,14. By using variants of conflict adaptation to separate conflict and control processes, a number of studies have identified neural correlates of conflict detection predominantly in medial prefrontal cortex, particularly in the anterior cingulate cortex (ACC)7,10–15. Also in the context of conflict-adaptation tasks, correlates of cognitive control processes have mostly been localized to sites in dorsolateral prefrontal cortex (DLPFC)7,8,13,14. However, it is not known how the putative loci of cognitive control in DLPFC rapidly bias perceptual processing in response to trial-by-trial fluctuation in conflict. Specifically, it remains an open question, whether in tasks such as the Stroop protocol, conflict is resolved through excitatory modulation (amplifying the processing of target information), inhibitory

Functional MRI Research Center, Columbia University, Neurological Institute, Box 108, 710 West 168th Street, New York, New York 10032, USA. Correspondence should be addressed to T.E. ([email protected]). Received 14 September; accepted 11 October; published online 6 November 2005; doi:10.1038/nn1594

1784

VOLUME 8

[

NUMBER 12

[


ARTICLES Actor

a

Political figure

b Congruent

Previous trial

Incongruent

congruentcongruent

congruentincongruent

Incongruent

Category incongruent


Category congruent

Current trial Congruent

incongruentcongruent

incongruentincongruent

c

3–5 s

1s

3–5 s

1s

Figure 1 Experimental protocol. Subjects discriminated between actors and political figures, based on either the face stimulus (face-target condition), or the written name (face-distracter condition). (a) Stimuli could be either category-congruent (for example, an actor’s face accompanied by an actor’s name), or category-incongruent (an actor’s face accompanied by a politician’s name or vice versa). (b) Trials were presented pseudo-randomly so as to produce an equal number of each possible trial sequence with respect to previous-trial and current-trial congruency and were analyzed by trial type (congruent-congruent, congruent-incongruent, incongruent-congruent and incongruent-incongruent). (c) Stimuli were presented for 1,000 ms with a jittered interstimulus interval (ISI) of 3,000–5,000 ms (mean ISI ¼ 4,000 ms). Shown is an example trial in which a congruent stimulus is followed by an incongruent stimulus, resulting in a low control–high conflict trial.

modulation (suppressing the processing of distracter information), or a combination of the two16. The current study aimed to complement our understanding of cognitive control by unraveling the nature of this perceptual biasing process. Target amplification and distracter inhibition have both been described as feasible neural mechanisms of selective attention17. For example, neural responses are enhanced for attended spatial locations18,19, stimulus features20–22 or objects23,24. However, previous studies investigating attentional selection have typically required subjects explicitly, via external cues, to attend to a particular aspect of their visual environment. In contrast, the current study investigates which selection mechanism underlies rapid, ‘online’ performance adjustments arising endogenously from the subject interacting with the ongoing task. This is because conflict adaptation emerges from the task context, as defined by the stimulus history, and in the absence of any explicit external cues or instructions to the subjects to shift their focus of attention or improve their performance. Therefore, conflict adaptation serves as a model of attention regulation in many real-life situations where, typically, we lack explicit external guidance as to where we should attend or what we should attend to, for optimal task performance. To probe the neural mechanism underlying conflict adaptation, we used functional magnetic resonance imaging (fMRI) while subjects performed a new variant of the Stroop task, involving face stimuli. The face stimuli were expected to elicit blood oxygen level–dependent (BOLD) signals in the fusiform face area (FFA), an extrastriate visual region responsible for face processing25 that is known to be susceptible to attentional and contextual top-down modulation26–29. By using face stimuli as either target or distracter stimulus features in a Stroop-like task, we obtained FFA responses that provided a window into the perceptual processing of target and distracter dimensions, under


[

NUMBER 12

[

DECEMBER 2005

varying levels of cognitive control during conflict adaptation. We found that behavioral conflict-adaptation effects were exclusively associated with the amplified processing of task-relevant stimulus properties and not with the suppressed processing of task-irrelevant stimulus features. This enhanced cortical representation of target information in visual cortex was accompanied by an increased functional interaction with cognitive control loci in DLPFC. RESULTS Subjects categorized stimuli consisting of familiar faces of actors and politicians, with either category-congruent or category-incongruent names of other actors and politicians written across them (Fig. 1a). Trials were analyzed on the basis of previous- and current-trial congruency (Fig. 1b): the previous-trial congruency determined the level of ‘control’ on the current trial, and the current-trial congruency determined the level of ‘conflict’ on the current trial. Thus, successive congruent-congruent trials reflected low conflict under low control; congruent-incongruent trials reflected high conflict under low control; incongruent-congruent trials reflected low conflict under high control; and incongruent-incongruent trials reflected high conflict under high control conditions. Within this factorial design, conflict adaptation was represented by the interaction of previous- and current-trial congruency, where the effect of current-trial conflict was greater following congruent trials (that is, congruent-incongruent 4 congruentcongruent) than following incongruent trials (that is, incongruentincongruent 4 incongruent-congruent)4. The critical comparison in this analysis lies in demonstrating reduced interference (conflict) from incongruent distracters under conditions of high control as compared to conditions of low control (incongruent-incongruent versus congruent-incongruent trials)7,12. Subjects discriminated actors from political figures (Fig. 1c) using a two-alternative forced-choice button press, in two experimental contexts. In one condition, they responded according to the identity of the face stimulus (‘face-target’ condition); in the other condition, they responded according to the written name while ignoring the face stimulus (‘face-distracter’ condition). This design allowed us to compare subjects’ responses to identical face stimuli under conditions of low and high cognitive control, depending on whether faces represented the target or the distracter dimension of the task. If attention regulation during the Stroop task was mediated by the amplification of target processing, we would expect to observe increased FFA activation in the high-control condition as compared to the low-control condition, in the face-target condition. Conversely, if performance adjustments depended on the suppression of distracter processing, we would expect FFA activation to be inhibited under the high-control condition as compared to the low-control condition, during the face-distracter condition. Finally, if excitatory and inhibitory mechanisms contributed to optimize performance, both of the above predictions would hold. Behavioral data: conflict adaptation Reaction times (RTs) for correct trials showed Stroop-like interference effects from incongruent distracters, both in the face-target condition (congruent mean ¼ 711 ms; incongruent mean ¼ 725 ms; t21 ¼ 4.0, P o 0.001) and in the face-distracter condition (congruent mean ¼ 862 ms; incongruent mean ¼ 903 ms; t21 ¼4.1, P o 0.001). Furthermore, the data in both tasks bore out classic conflict-adaptation effects (Table 1 and Fig. 2a,b). In the face-target task (Fig. 2a), RTs to incongruent stimuli were faster in the high-control condition than in the low-control condition (incongruent-incongruent o congruentincongruent; t21 ¼ 2.2, P o 0.05). This resulted in a previous-trial current-trial interaction (F1,21 ¼ 6.7, P o 0.02), as current-trial conflict

1785

under low control (congruent-incongruent 4 congruent-congruent; t21 ¼ 3.8, P o 0.002) was abolished under high control (incongruentincongruent E incongruent-congruent; t21 ¼ 0.1, P 4 0.5). Similarly, in the face-distracter task (Fig. 2b) responses to incongruent trials were faster in the high-control condition than in the low-control condition (incongruent-incongruent o congruent-incongruent; t21 ¼ 3.0, P o 0.009). This led to a reduction in current-trial conflict from the lowcontrol condition (congruent-incongruent 4 congruent-congruent, t21 ¼ 4.5, P o 0.001) to the high-control condition (F1,21 ¼ 5.5, P o 0.03), although conflict was not abolished entirely (incongruent-incongruent 4 incongruent-congruent; t21 ¼ 2.4, P o 0.03). These effects were not related to speed-accuracy trade-offs, as the accuracy data reflected the same pattern of results (Table 1). fMRI data: conflict adaptation in visual cortex To assess how conflict adaptation was achieved at the level of perceptual processing in visual cortex, we analyzed task-related BOLD responses in individually defined regions of interest (ROIs) in the FFA (Fig. 2c; Methods). When faces served as target stimuli, conflict adaptation was evident in the FFA (Fig. 2d). FFA responses to incongruent trials were enhanced in the high-control condition as compared to the low-control condition (incongruent-incongruent 4 congruent-incongruent; t21 ¼ 2.3, P o 0.04). This resulted in a previous-trial current-trial interaction (F1,21 ¼ 7.3, P o 0.02), as a current-trial conflict effect in the low-control condition (congruent-congruent 4 congruent-incongruent; t21 ¼ 2.8, P o 0.02) disappeared under high control (incongruent-incongruent E incongruent-congruent; t21 ¼ 1.5, P 4 0.14). Note that FFA activation to face target stimuli showed the inverse pattern of the behavioral data, in that low behavioral interference from the name distracters was associated with high FFA activation and high behavioral interference with low FFA activation. Notably, conflict adaptation was associated with the amplification of neural perceptual responses to relevant target stimulus features. When faces served as distracters, on the other hand, we found no effects of cognitive control on FFA responses (F1,21 ¼ 0.4, P 4 0.5; Fig. 2e), suggesting that the behavioral conflict-adaptation effect obtained in this task (Fig. 2b) was not mediated by selective inhibition of distracter processing. To illustrate, more directly, the effects of

Current trial

a

Congruent

940

Incongruent

760 740 720 700 680

Current trial

b Reaction time (ms)

Reaction time (ms)

780

920

d

Current trial Congruent

Reaction

Standard

Percentage

Standard

time (ms)

deviation

accuracy

deviation

Congruent-congruent

705

74

99.1

1.6

Congruent-incongruent Incongruent-congruent

732 717

81 76

95.5 97.5

4.4 3.3

Incongruent-incongruent

717

77

97.2

3.1

Face distracter condition

e

Congruent

Incongruent

Previous trial

deviation

116

97.6

4.6

119

94.2

4.5

Incongruent-congruent Incongruent-incongruent

864 891

122 100

97.5 95.2

3.4 4.7

cognitive control on target versus distracter processing, FFA responses to only incongruent stimuli were compared under conditions of low and high control (congruent-incongruent versus incongruent-incongruent), depending on whether they constituted target or distracter features. A task control interaction (F1,21 ¼ 4.9, P o 0.04; Fig. 3a) was characterized by an increase in target-related responses from low to high control (t21 ¼ 2.5, P o 0.03), with no effects of control on distracter-related responses (t21 ¼ 0.6, P 4 0.5). To corroborate that this cognitive control–related increase in activation during face-target processing was specific to the FFA and not a generic effect on high-level visual regions, we conducted a control analysis comparing responses to congruent-incongruent and incongruent-incongruent trials in the FFA to those in the parahippocampal place area (PPA), an extrastriate visual region selectively responsive to natural scenes30 (Methods). An expected main effect of cortical region in the processing of the face stimuli (FFA 4 PPA, F1,21 ¼ 17.1, P o 0.001) was accompanied by an interaction effect (F1,21 ¼ 4.3,

c

Incongruent

Figure 2 Conflict adaptation in behavioral and fMRI data. (a,b) Mean group reaction times (± s.e.m.) for current congruent and incongruent trials plotted as a function of previous-trial congruency (x-axis) for (a) the face-target condition and (b) the face-distracter condition. (c) Illustration of FFA activation on a rostral (top panel) and axial (bottom panel) brain slice, derived from a group analysis of the face-area localizer scan (MNI x ¼ 46, y ¼ –54, z ¼ –24; 121 voxels, 968 mm3; x ¼ –42, y ¼ –54, z ¼ –22; 85 voxels, 680 mm3) displayed at P o 0.05 (corrected). (d,e) Mean group activation values (betas ± s.e.m.) from subject-specific FFA ROIs for current congruent and incongruent trials plotted as a function of previous-trial congruency (x-axis) for (d) the face-target condition and (e) the face-distracter condition.

Congruent Incongruent


2.8 2.6 2.4 2.2

2.2

Standard

accuracy

859

Current trial

FFA activation

2.4

Percentage

deviation

915

3

2.6

Standard

Congruent-incongruent

860

Incongruent

2.8

Reaction time (ms) Congruent-congruent

Previous trial

3

1786

Face target condition

880

840


Congruent

Table 1 Descriptive statistics of behavioral data

900

Previous trial

FFA activation


ARTICLES

Congruent

Incongruent

Previous trial

VOLUME 8

[

NUMBER 12

[


ARTICLES

Face target Face distracter

b

0.5

0.5

0.3

0.3

0.1

Activation

FFA activation

a

–0.1 –0.3

0.1 –0.1 –0.3 –0.5

–0.5 –0.7


FFA PPA

–0.7 Low control

High control

Low control

High control

Figure 3 Task- and region-specificity of cognitive control effects. (a) FFA activation (normalized betas ± s.e.m.) for incongruent trials only are plotted as a function of control (low versus high) and task (face-target versus face-distracter). (b) Neural responses from FFA and PPA (normalized betas ± s.e.m.) are shown for incongruent trials under conditions of low versus high control during the face-target condition.

P o 0.05), as responses increased with control in the FFA (t21 ¼ 2.3, P o 0.04) but not in the PPA (t21 ¼ 0.5, P 4 0.5; Fig. 3b). We probed the specificity of cognitive control effects further by assessing whether activity in early visual cortex (V1 and V2) was affected by task and control variables. To this end, we analyzed data from voxels in early occipital regions (all falling within V1 or V2) that were most highly activated by all visual stimuli during the face area localizer task (Methods). These early visual regions displayed no task (F1,21 ¼ 0.1, P 4 0.7), control (F1,21 ¼ 0.1, P 4 0.7), or task control interaction effects (F1,21 ¼ 1.2, P 4 0.2). These results confirmed that the effect of cognitive control was specific to neuronal populations involved in the processing of task-relevant target stimulus features in the FFA and did not extend to other high level or to lower level visual processing. fMRI data: top-down conflict resolution Are the enhanced perceptual responses to face target stimuli under high control a result of top-down modulation from DLPFC? If this is the case, putative cognitive control regions should display increased connectivity with the FFA in the high-control condition as compared to the low-control condition; further, this effect should be limited to the facetarget condition. To test these predictions, we first identified neural substrates of cognitive control during the face-target task with an incongruent-incongruent 4 congruent-incongruent contrast in a whole-brain group analysis. This analysis yielded clusters of activation in the right DLPFC (Brodmann’s area 46), right middle temporal gyrus and left anterior insula (Fig. 4a). To assess the functional interaction, during conflict adaptation, between these cortical loci of cognitive control and the FFA, we then carried out a psychophysiologic interaction (PPI) analysis31. PPI represents a measure of context-dependent connectivity, explaining regionally specific responses in one brain area in terms of the interaction between input from another brain region and a cognitive or sensory process31,32. In the current study, PPI allowed us to assess if the FFA displayed context-sensitive increments in functional integration with these cognitive control ROIs when going from low- to high-control trials. We calculated the degree of functional interaction during high- versus low-control incongruent trials (incongruent-incongruent versus congruent-incongruent) for both the facetarget and the face-distracter task; we then subjected the connectivity data to a task control interaction analysis (Methods). This revealed a cluster of voxels in the DLPFC ROI (Fig. 4b) that showed task-specific (face target 4 face distracter) and control-specific (incongruentincongruent 4 congruent-incongruent) increments in functional integration with the FFA. Thus, functional coupling between the right DLPFC and the FFA increased under high control in the


[

NUMBER 12

[

DECEMBER 2005

face-target condition, but not in the face-distracter condition— precisely as would be predicted for a region implementing conflictsensitive top-down biasing of perceptual target-feature processing. Conflict adaptation versus priming effects It is controversial whether conflict-adaptation effects truly reflect conflict-driven cognitive control processes, or whether they can be accounted for by lower-level priming effects. Priming might arise from different proportions of stimulus-response repetitions (and alternations) between congruent-congruent and incongruent-incongruent trials on the one hand, and congruent-incongruent and incongruentcongruent trials on the other hand7,8,33–35. For instance, in the Eriksen flanker task36, behavioral conflict-adaptation effects can be mediated entirely by particularly fast responses on congruent-congruent and incongruent-incongruent trials where identical stimulus-response pairings are repeated; this suggests that the effect may be due to repetition priming rather than to adjustments in cognitive control33,35. For this reason, we controlled for repetition-priming confounds by not including any direct repetitions of identical stimuli7,8. However, it could still be argued that the conflict-adaptation effects we obtained may have occurred because on 50% of congruent-congruent and incongruentincongruent trials, both the target and distracter categories (that is, actor or politician) remained the same, possibly facilitating performance through some form of category priming. In contrast, on congruent-incongruent and incongruent-congruent trials, at least one category always changed. Thus, the behavioral conflict reduction (incongruent-incongruent o congruent-incongruent) and the accompanying neural target-feature amplification (incongruent-incongruent 4 congruent-incongruent) that we observed during face-target processing could, in theory, stem from incongruent-incongruent trials being subject to category-priming effects rather than top-down control influences. To test this alternative interpretation, we re-analyzed the data, splitting up incongruent-incongruent trials into those where target and distracter categories were repeated (repetition trials) and those where these categories alternated (alternation trials). If the previous results were driven by priming effects, repetition RTs should be faster than alternation RTs, and conflict adaptation should be observed exclusively for analyses including only repetition trials but not for analyses including only alternation trials33,35. Reaction times for repetition trials (mean ± s.d., 723 ± 88 ms) and alternation trials (713 ± 73 ms) did not differ (t21 ¼ 1.0, P 4 0.3). Note that, descriptively, responses for repetitions were actually slower than those for alternation trials. Accordingly, the conflict-adaptation

a

b

Figure 4 Regions associated with top-down control processes. (a) Brain areas implicated in cognitive control during conflict adaptation (incongruentincongruent 4 congruent-incongruent) were identified in middle and inferior frontal gyri of the right DLPFC (MNI x ¼ 40, y ¼ 38, z ¼ 20; 180 voxels, 1,440 mm3), right middle temporal gyrus (MNI x ¼ 48, y ¼ –54, z ¼ 0; 153 voxels, 1,224 mm3) and left anterior insula (MNI x ¼ –46, y ¼ –6, z ¼ 4; 216 voxels, 1,728 mm3), with a cluster threshold of P o 0.05 (corrected). (b) Within these cognitive control ROIs (red), right DLPFC exhibits voxels (blue) (MNI x ¼ 26, y ¼ 30, z ¼ 10; 27 voxels, 216 mm3) that show a task-specific and context-specific increase in functional integration with the FFA, at a voxelwise threshold of P o 0.05.

1787


ARTICLES interaction effect (previous-trial current-trial congruency) was evident to a similar degree when we included only repetition (F1,21 ¼ 4.2, P ¼ 0.053) or only alternation trials (F1,21 ¼ 6.2, P ¼ 0.021) in the analysis. Next, we conducted corresponding control analyses for fMRI data extracted from the FFA during face-target processing, modeling the data to include separate regressors for repetition and alternation trials. (Note that the inclusion of an additional regressor in the model inevitably alters parameter estimates for other regressors as well.) Again, FFA activation during repetition trials (mean beta ± s.d., 2.4 ± 1.6) and alternation trials (2.3 ± 1.4) did not differ (t21 ¼ 0.5, P 4 0.6). Conflict-adaptation effects were of similar magnitude when we included only repetition trials (F1,21 ¼ 3.6, P ¼ 0.07) or only alternation trials (F1,21 ¼ 3.1, P ¼ 0.09) in the analysis. These results show that incongruent-incongruent category-repetition and categoryalternation trials contributed similarly to the overall conflict-adaptation effect in both the behavioral and fMRI data, and that these data thus cannot be explained by category priming effects, but rather can likely be attributed to conflict-driven adjustments in cognitive control. Finally, the current results could also have been driven primarily by one of the two stimulus categories. For example, it could be argued that the actors in the stimulus set displayed a greater similarity with each other than the political figures, both in terms of their faces and their names. From the viewpoint of a priming account of conflict adaptation, higher within-category similarity would be expected to lead to stronger priming effects—that is, greater conflict adaptation. To test whether our results may have stemmed from effects within a particular subset of stimuli, we split up the data set according to whether a given trial presented the picture (or name) of an actor or a politician. If overall conflict adaptation was driven disproportionately by one stimulus category, a three-way interaction effect between stimulus category, previous-trial congruency and current-trial congruency would be observed. In the RT data, pictorial category did not interact with conflict adaptation (F1,21 ¼ 1.6, P 4 0.2) and neither did the name category (F1,21 ¼ 0.1, P 4 0.8). Next, we re-analyzed the fMRI ROI data, this time splitting up each trial-type regressor according to whether it contained the picture (or name) of an actor or a political figure. In accordance with the behavioral findings, there were no effects of stimulus category on conflict adaptation in terms of FFA activation: neither for the pictorial category (F1,21 ¼ 0.65, P 4 0.4), nor for the name category (F1,21 ¼ 0.1, P 4 0.7). These data demonstrate that the current results were not primarily driven by one particular stimulus category. DISCUSSION We obtained significant behavioral conflict and conflict-adaptation effects in a variant of the Stroop task that used face stimuli as either target or distracter stimulus features. In both versions of the task, responses were faster in incongruent trials that followed incongruent trials than in incongruent trials that followed congruent trials. This represented successful conflict resolution through cognitive control7,9,12. By simultaneously imaging BOLD responses in individuals’ FFA, we showed that the strength of neural face representation varied with conflict and control under conditions where faces served as target stimuli, but not when they served as distracter stimuli. Specifically, FFA activation to face-target stimuli was increased in response to incongruent trials following incongruent trials, compared to when incongruent trials followed congruent ones. Thus, face processing was amplified when cognitive control was high (and conflict was reduced), compared to when the identical incongruent stimuli were processed under conditions of low control (and high conflict). Further, this effect was exclusive to the face-target condition. In addition, we contrasted

1788

the effect of cognitive control during face-target processing in the FFA to that in two other visual areas: a similar category-specific high-level visual area, the PPA, and early visual cortex. We showed that this amplification of neural target-feature representation was region-specific and did not represent a generic upregulation of activity in striate or extra-striate visual regions under conditions of high cognitive control. The conflict-monitoring and cognitive control model would predict that this transient modulation of target-feature processing should be directly related to input from cognitive control loci in DLPFC7,9,13. We tested this hypothesis by measuring context-sensitive functional integration between the FFA and functionally defined cognitive control ROIs in right DLPFC, the left insula and right superior temporal cortex. In support of the notion that FFA modulation was mediated by a topdown biasing signal, we found that a subregion of the DLPFC ROI showed task- and control-dependent functional integration with the FFA. This subregion showed increased coupling with the FFA under conditions of high control, but only when the face stimuli served as targets for attentional selection. Therefore, along with previous studies documenting neural substrates of conflict monitoring in the ACC7,10–15 and cognitive control in DLPFC7,8,13,14, the current data complement an emergent neural model of cognitive control9 by supplying the mechanism through which conflict resolution is implemented at the target site of attentional modulation: namely, through target facilitation rather than distracter inhibition. The exact way in which this neural modulation of target-feature processing is achieved during conflict adaptation raises important questions for future research. One possible mechanism for target-feature enhancement is that during the face-target condition, attentional top-down signals may lead to enhanced pre-stimulus baseline neural activity in the FFA22,37, thus favoring this area in the competition for processing resources during subsequent stimulus processing1,17. Our results demonstrate that rapid, online performance adjustments in response to high conflict are mediated by amplified neural processing of task-relevant (target) stimulus features but not by inhibited processing of task-irrelevant (distracter) stimulus features. These data are in agreement with electrophysiological and neuroimaging studies that have reported enhancement of target processing with respect to cued attention shifts between different spatial locations18,19, stimulusfeatures20–22 and objects23,24. These previous investigations demonstrated neural target-feature enhancement in the context of explicitly (and exogenously) cued attention shifts; in contrast, the current study provides evidence suggesting that target-feature enhancement constitutes the main selection mechanism when attention regulation is driven endogenously so as to optimize performance—as is likely the case in many real-life situations. In keeping with computational models9,16, our findings strongly suggest that performance on classic selectiveattention tasks, such as the Stroop task, may be accounted for without invoking a mechanism that actively inhibits the perceptual processing of task-irrelevant stimulus features (see also ref. 38). Rather, our data are consistent with the proposal that target-feature amplification represents the primary top-down mechanism of selective attention. Both behavioral and neuroimaging data suggest that perceptual suppression of task-irrelevant (distracter) information may not be possible unless attentional resources are entirely bound up by the processing of taskrelevant (target) information under highly demanding conditions39,40. The above interpretation of the current data hinges critically on whether conflict adaptation is truly a reflection of conflict-driven adjustments in cognitive control or results from priming effects within particular stimulus sequences33,35. Our study did not contain any direct stimulus repetitions and was therefore not confounded by repetition priming effects7,8. Higher-level priming effects, however, could feasibly

VOLUME 8

[

NUMBER 12

[



ARTICLES arise from repetitions of target- and distracter-feature categories, which occurred on 50% of congruent-congruent and incongruentincongruent trials but never on congruent-incongruent and incongruent-congruent trials. In control analyses, we found no effects of category repetition on behavioral or fMRI data, and we can therefore reject this alternative interpretation of our data. Another potential confound in the conflict-adaptation effect is that, typically, on 50% of congruent-incongruent and incongruent-congruent trials, a response repetition is accompanied by a stimulus alternation35,41; these trials may be associated with slower responses than trials in which either the stimulus and the response both alternate or both stay the same41. This confound was also controlled for in the current study, as congruentcongruent and incongruent-incongruent trials contained no exact stimulus repetitions (that is, the actual stimulus always alternated), but entailed response alternations on 50% of the trials—exactly as was the case for congruent-incongruent and incongruent-congruent trials. Although we are confident in concluding that the current data set reflects a demonstration of conflict-driven adjustments in cognitive control, the careful assessment of other factors that may contribute to such sequential trial effects remains important for gaining a better understanding of cognitive control mechanisms. For instance, it is not clear why conflict adaptation on the flanker task may be mediated entirely by priming effects33,35, whereas in other conflict tasks— such as the Simon task42, the color-naming Stroop task7,8 or the current task— this does not appear to be the case. An additional question of interest is the respective contribution of (and potential interaction between) preparatory processes that arise from conflict and those that may arise from specific expectancies regarding the nature of an upcoming stimulus. Notably, the original report of the conflict-adaptation effect envisaged subjects’ expectancies, rather than conflict, as the driving force behind attentional adjustments4. Such expectancy effects, as well as their potential interaction with conflict-driven cognitive control processes, have been successfully modeled in the context of nonconflict, two-alternative, forced-choice tasks43,44, but they have yet to be explored within the context of the conflict-adaptation protocol. In conclusion, we have shown that conflict adaptation, reflected in improved selective-attention performance following high-conflict trials, is mediated by the amplified neural representation of taskrelevant stimulus features, but is not related to the perceptual inhibition of task-irrelevant features. We propose that attentional target-feature amplification is the neural mechanism by which cognitive control optimizes performance. METHODS Subjects. 22 (14 females) healthy volunteers (mean age ¼ 28.7 years, range ¼ 20–40 years) gave written informed consent in accordance with institutional guidelines to participate in this study. Experimental protocols. Stimuli were presented with Presentation software (Neurobehavioral Systems, http://nbs.neuro-bs.com) and displayed with a back-projection screen that was viewed by the subjects via a mirror attached to the head-coil. The FFA localizer task was adopted from a previous study in our laboratory29: subjects passively viewed photographic face and house stimuli in 12 alternating blocks of 15 s, separated by 10 s resting (fixation) periods. Within each block, 15 faces or houses were presented for 750 ms, with an interstimulus interval (ISI) of 250 ms. Each run of the main task consisted of 148 presentations of photographic stimuli depicting the face of either an actor (Robert DeNiro, Al Pacino or Jack Nicholson) or a political figure (Fidel Castro, Bill Clinton or Mao Zedong), all of whom were readily identified by the subjects before the experiment. Faces were presented with congruent or incongruent names (Fig. 1) written across them in red letters. No face stimulus was paired with its own name. Stimuli were presented for 1,000 ms, with a varying ISI of 3,000–5,000 ms (mean ISI ¼ 4,000 ms), in pseudo-random order (counter-


[

NUMBER 12

[

DECEMBER 2005

balanced for equal numbers of congruent-congruent, congruent-incongruent, incongruent-congruent and incongruent-incongruent stimulus sequences). Stimulus occurrences were counter-balanced across trial types and response buttons, and the stimulus sequence included neither immediate stimulus repetitions nor instances of ‘negative priming’ (where the distracter feature of one trial turns into the target feature of the next trial). Subjects were instructed to respond as fast as possible, while maintaining accuracy, by pushing response buttons corresponding to ‘actor’ (right index finger) or ‘politican’ (right middle finger). In the ‘face-target’ run, subjects responded according to the face dimension of the stimuli, and in the ‘face-distracter’ run, they responded according to the name dimension, with the order of runs counter-balanced across subjects. We analyzed the behavioral data in terms of reaction times (excluding error and post-error trials, and condition-specific outlier values of more than 2 standard deviations from the mean), and accuracy rates. Image acquisition. Images were recorded with a GE 1.5–T scanner. Functional images were acquired parallel to the anterior commissure–posterior commissure (AC-PC) line with a T2*-weighted EPI sequence of 24 contiguous axial slices (TR ¼ 2,000 ms; TE ¼ 40 ms; flip angle ¼ 601; FoV ¼ 190 190 mm, array size 64 64) of 4.5 mm thickness and 3 3 mm in-plane resolution. Structural images were acquired with a T1-weighted SPGR sequence (TR ¼ 19 ms; TE ¼ 5 ms; flip angle ¼ 201; FoV ¼ 220 220 mm), recording 124 slices at a slice thickness of 1.5 mm and in-plane resolution of 0.86 0.86 mm. Image analysis. All pre-processing and statistical analyses were done using SPM2 (http://www.fil.ion.ucl.ac.uk/spm/spm2.html). Functional data were corrected for differences in slice-timing, spatially realigned to the first volume of the first run and smoothed with a Gaussian kernel of 8 mm3 FWHM. For whole-brain analyses, the realigned images were spatially normalized to the MNI template brain (re-sampled voxel size: 2 mm3) before smoothing, whereas for FFA ROI definition and related analyses, the data remained in native space. The first five volumes of each run were discarded before we built and estimated the statistical model. A 128-s temporal high-pass filter was applied to the data and models, and temporal autocorrelation in the fMRI time series was estimated (and corrected for) using a first-order autoregressive function. For the FFA and PPA localizers, epochs of face and house stimuli were modeled with two box-car functions convolved with a canonical hemodynamic response function (HRF). Subject-specific FFA and PPA ROIs from nonnormalized data were defined by voxels within the fusiform/parahippocampal gyri that displayed face 4 house or house 4 face selectivity at a voxel-wise threshold of P o 0.0001 (uncorrected), with a minimum cluster threshold of five contiguous voxels (202.5 mm3). Marsbar software (http://marsbar. sourceforge.net/) was used to convert these clusters into ROIs and to extract ROI data for the subsequent analyses. As a comparison region in early visual cortex, data were also extracted from the voxel (40.5 mm3) displaying peak activition to both face and house stimuli in the FFA localizer task. For the main task, regressors of stimulus events (convolved with a canonical HRF) were created for congruent-congruent, congruent-incongruent, incongruent-congruent and incongruent-incongruent trial types, with error and post-error trials modeled separately. Beta values for each regressor were extracted from individual FFA ROIs (from non-normalized data) using Marsbar and analyzed in analyses of variance (ANOVA), followed by planned comparison t-tests. For the purpose of control analyses (Results), these data were modeled once more with the incongruent-incongruent trials regressor split up into category (actor or political figure)- repetition and category-alternation trials. Also, the data were modeled another two times with each trial-type regressor split up into two regressors, separately modeling trials that contained pictures (or names) of actors versus politicians. For the whole-brain search of topdown control regions, the original model was applied to normalized data across subjects in a random-effects analysis for an incongruent-incongruent 4 congruent-incongruent contrast, and results were thresholded at a cluster-level P o 0.05 (corrected). For the psychophysiologic interaction (PPI) analysis31, we extracted the deconvolved time-course of FFA activity in each subject (from normalized data), based on a sphere of radius 5 mm around the peak-activation voxel from the group FFA analysis (MNI x ¼ 46, y ¼ –54, z ¼ –24). We then calculated the product of this activation time-course and the vector of the psychological

1789


ARTICLES variable of interest (incongruent-incongruent 4 congruent-incongruent) to create the psychophysiological interaction term. New SPMs were computed for each subject, including, as regressors, the interaction term, the physiological variable (that is, the FFA activation time course) and the psychological variable. We then identified areas where activation was predicted by the psychophysiological interaction term, with the FFA activity and the psychological regressor treated as confound variables. These analyses were carried out separately for both the face-target and face-distracter tasks. Individual PPI SPMs were then entered into a random-effects group analysis contrasting connectivity patterns between face-target and face-distracter conditions with a paired-samples t-test, within the cognitive control ROIs and thresholded at P o 0.05 (uncorrected) with a cluster size of 45 voxels (40 mm3). ACKNOWLEDGMENTS We thank A. Etkin, C. Summerfield, E. Stern, J. Grinband and J. Mangels for comments. This work was funded in part by Johnson & Johnson (J.H). COMPETING INTERESTS STATEMENT The authors declare that they have no competing financial interests. Published online at http://www.nature.com/natureneuroscience/ Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/ 1. Desimone, R. & Duncan, J. Neural mechanisms of selective visual attention. Annu. Rev. Neurosci. 18, 193–222 (1995). 2. Miller, E.K. & Cohen, J.D. An integrative theory of prefrontal cortex function. Annu. Rev. Neurosci. 24, 167–202 (2001). 3. Logan, G.D. & Zbrodoff, N.J. When it helps to be misled: facilitative effects of increasing the frequency of conflicting stimuli in a Stroop-like task. Mem. Cognit. 7, 166–174 (1979). 4. Gratton, G., Coles, M.G. & Donchin, E. Optimizing the use of information: strategic control of activation of responses. J. Exp. Psychol. Gen. 121, 480–506 (1992). 5. Stroop, J.R. Studies of interference in serial verbal reactions. J. Exp. Psychol. 18, 643–662 (1935). 6. MacLeod, C.M. Half a century of research on the Stroop effect: an integrative review. Psychol. Bull. 109, 163–203 (1991). 7. Kerns, J.G. et al. Anterior cingulate conflict monitoring and adjustments in control. Science 303, 1023–1026 (2004). 8. Egner, T. & Hirsch, J. The neural correlates and functional integration of cognitive control in a Stroop task. Neuroimage 24, 539–547 (2005). 9. Botvinick, M.M., Braver, T.S., Barch, D.M., Carter, C.S. & Cohen, J.D. Conflict monitoring and cognitive control. Psychol. Rev. 108, 624–652 (2001). 10. Carter, C.S. et al. Anterior cingulate cortex, error detection, and the online monitoring of performance. Science 280, 747–749 (1998). 11. Carter, C.S. et al. Parsing executive processes: strategic vs. evaluative functions of the anterior cingulate cortex. Proc. Natl. Acad. Sci. USA 97, 1944–1948 (2000). 12. Botvinick, M., Nystrom, L.E., Fissell, K., Carter, C.S. & Cohen, J.D. Conflict monitoring versus selection-for-action in anterior cingulate cortex. Nature 402, 179–181 (1999). 13. MacDonald, A.W., III, Cohen, J.D., Stenger, V.A. & Carter, C.S. Dissociating the role of the dorsolateral prefrontal and anterior cingulate cortex in cognitive control. Science 288, 1835–1838 (2000). 14. Durston, S. et al. Parametric manipulation of conflict and response competition using rapid mixed-trial event-related fMRI. Neuroimage 20, 2135–2141 (2003). 15. Casey, B.J. et al. Dissociation of response conflict, attentional selection, and expectancy with functional magnetic resonance imaging. Proc. Natl. Acad. Sci. USA 97, 8728–8733 (2000). 16. Cohen, J.D., Dunbar, K. & McClelland, J.L. On the control of automatic processes: a parallel distributed processing account of the Stroop effect. Psychol. Rev. 97, 332–361 (1990).

1790

17. Kastner, S. & Ungerleider, L.G. Mechanisms of visual attention in the human cortex. Annu. Rev. Neurosci. 23, 315–341 (2000). 18. Heinze, H.J. et al. Combined spatial and temporal imaging of brain activity during visual selective attention in humans. Nature 372, 543–546 (1994). 19. Luck, S.J., Chelazzi, L., Hillyard, S.A. & Desimone, R. Neural mechanisms of spatial selective attention in areas V1, V2, and V4 of macaque visual cortex. J. Neurophysiol. 77, 24–42 (1997). 20. Corbetta, M., Miezin, F.M., Dobmeyer, S., Shulman, G.L. & Petersen, S.E. Selective and divided attention during visual discriminations of shape, color, and speed: functional anatomy by positron emission tomography. J. Neurosci. 11, 2383–2402 (1991). 21. Treue, S. & Maunsell, J.H. Attentional modulation of visual motion processing in cortical areas MT and MST. Nature 382, 539–541 (1996). 22. Chawla, D., Rees, G. & Friston, K.J. The physiological basis of attentional modulation in extrastriate visual areas. Nat. Neurosci. 2, 671–676 (1999). 23. O’Craven, K.M., Downing, P.E. & Kanwisher, N. fMRI evidence for objects as the units of attentional selection. Nature 401, 584–587 (1999). 24. Serences, J.T., Schwarzbach, J., Courtney, S.M., Golay, X. & Yantis, S. Control of objectbased attention in human cortex. Cereb. Cortex 14, 1346–1357 (2004). 25. Kanwisher, N., McDermott, J. & Chun, M.M. The fusiform face area: a module in human extrastriate cortex specialized for face perception. J. Neurosci. 17, 4302–4311 (1997). 26. Wylie, G.R., Javitt, D.C. & Foxe, J.J. Don’t think of a white bear: an fMRI investigation of the effects of sequential instructional sets on cortical activity in a task-switching paradigm. Hum. Brain Mapp. 21, 279–297 (2004). 27. O’Craven, K.M. & Kanwisher, N. Mental imagery of faces and places activates corresponding stimulus-specific brain regions. J. Cogn. Neurosci. 12, 1013–1023 (2000). 28. Cox, D., Meyers, E. & Sinha, P. Contextually evoked object-specific responses in human visual cortex. Science 304, 115–117 (2004). 29. Summerfield, C., Egner, T., Mangels, J. & Hirsch, J. Mistaking a house for a face: neural correlates of misperception in healthy humans. Cereb. Cortex, published online 13 July 2005 (doi:10.1093/cercor/bhi129). 30. Epstein, R. & Kanwisher, N. A cortical representation of the local visual environment. Nature 392, 598–601 (1998). 31. Friston, K.J. et al. Psychophysiological and modulatory interactions in neuroimaging. Neuroimage 6, 218–229 (1997). 32. Friston, K.J. Functional integration in the brain. in Human Brain Function 2nd edn. (eds. Frackowiak, R.S. et al.) 971–997 (Academic Press, San Diego, 2004). 33. Mayr, U., Awh, E. & Laurey, P. Conflict adaptation effects in the absence of executive control. Nat. Neurosci. 6, 450–452 (2003). 34. Botvinick, M.M., Cohen, J.D. & Carter, C.S. Conflict monitoring and anterior cingulate cortex: an update. Trends Cogn. Sci. 8, 539–546 (2004). 35. Nieuwenhuis, S. et al. Accounting for sequential effects in the flanker task: Conflict adaptation or associative priming? Mem. Cognit. (in the press). 36. Eriksen, B.A. & Eriksen, C.W. Effects of noise letters upon the identification of a target letter in a nonsearch task. Percept. Psychophys. 16, 143–149 (1974). 37. Kastner, S., Pinsk, M.A., De Weerd, P., Desimone, R. & Ungerleider, L.G. Increased activity in human visual cortex during directed attention in the absence of visual stimulation. Neuron 22, 751–761 (1999). 38. Egner, T. & Hirsch, J. Where memory meets attention: neural substrates of negative priming. J. Cogn. Neurosci. 17, 1774–1784 (2005). 39. Lavie, N. Perceptual load as a necessary condition for selective attention. J. Exp. Psychol. Hum. Percept. Perform. 21, 451–468 (1995). 40. Rees, G., Frith, C.D. & Lavie, N. Modulating irrelevant motion perception by varying attentional load in an unrelated task. Science 278, 1616–1619 (1997). 41. Hommel, B., Proctor, R.W. & Vu, K.P. A feature-integration account of sequential effects in the Simon task. Psychol. Res. 68, 1–17 (2004). 42. Sturmer, B., Leuthold, H., Soetens, E., Schroter, H. & Sommer, W. Control over locationbased response activation in the Simon task: behavioral and electrophysiological evidence. J. Exp. Psychol. Hum. Percept. Perform. 28, 1345–1363 (2002). 43. Cho, R.Y. et al. Mechanisms underlying dependencies of performance on stimulus history in a two-alternative forced-choice task. Cogn. Affect. Behav. Neurosci. 2, 283–299 (2002). 44. Jones, A.D., Cho, R.Y., Nystrom, L.E., Cohen, J.D. & Braver, T.S. A computational model of anterior cingulate function in speeded response tasks: effects of frequency, sequence, and conflict. Cogn. Affect. Behav. Neurosci. 2, 300–317 (2002).

VOLUME 8

[

NUMBER 12

[


E R R ATA

Erratum: Lbx1 and Tlx3 are opposing switches in determining GABAergic versus glutamatergic transmitter phenotypes Leping Cheng, Omar Abdel Samad, Yi Xu, Rumiko Mizuguchi, Ping Luo, Senji Shirasawa, Martyn Goulding & Qiufu Ma Nat. Neurosci. 8, 1510–1515 (2005) This article contained a misspelling. Lhx1/2 should have read Lhx1/5 throughout the text.

Erratum: Why pictures look right when viewed from the wrong place Dhanraj Vishwanath, Ahna R Girshick & Martin S Banks Nat. Neurosci. 8, 1401–1410 (2005) On page 1402, the first two sentences of the second full paragraph in the second column were omitted. The paragraph should have begun as follows: “An alternative explanation, the local-slant hypothesis, suggests that location of the CoP is not recovered. Instead, the observed invariance is due to an adjustment of the retinal-image shape based on measurements of the local slant of the picture surface at the point of interest. This hypothesis does not require estimates of the location of or distance to the CoP.”

Erratum: Top-down suppression deficit underlies working memory impairment in normal aging Adam Gazzaley, Jeffrey W Cooney, Jesse Rissman & Mark D’Esposito Nat. Neurosci. 8, 1298–1300 (2005) The article contained typographical errors in the labeling of the y axes in Figures 2b and 2c. The label for the y axis on Figure 2b should read β, and the label for the y axis on Figure 2c should read ∆β.

CO R R I G E N D UM

Corrigendum: Orexin, drugs and motivated behaviors Thomas E Scammell & Clifford B Saper Nat. Neurosci. 8, 1286–1288 (2005) This article contains an incorrect citation. The name of the last author in reference 2 (“A role for lateral hypothalamic orexin neurons in reward seeking” Nature 2005) should read Aston-Jones, G., rather than Jones, G.A. The authors regret the error.


1791

E R R ATA




CO R R I G E N D UM



1791

E R R ATA




CO R R I G E N D UM



1791

E R R ATA




CO R R I G E N D UM



1791

Nature Neuroscience December 2005

Recommend Documents