www.nature.com/natureneuroscience

EDITORIAL OFFICE neurosci@natureny.com75 Varick Street, Fl 9, New York, NY 10013-1917Tel: (212) 726 9319, Fax: (212) 696 0978Editor: Kalyani NarasimhanAssociate Editors: Hannah Bayer, Min Cho, Annette Markus, Charvy NarainAssistant Editor: Kathleen DaveCopy Editors: Anita Gould, David LechtenbergProduction Editors: Sabina Eberle, Jamel WootenSenior Illustrator: Katie VicariIllustrator: Kimberly CaesarCover Design: Erin DewaltEditorial Assistant: Natasha Klushina

MANAGEMENT OFFICESNPG New York75 Varick Street, Fl 9, New York, NY 10013-1917Tel: (212) 726 9200, Fax: (212) 696 9006Publisher: Stephanie DimentExecutive Editor: Linda MillerChief Technology Officer: Howard RatnerHead of Nature Research & Reviews Marketing: Sara GirardMarketing Manager: Amy MaurerProduction Coordinator: Diane TempranoHead of Web Services: Anthony BarreraWeb Production Manager: Susan Kline

NPG LondonThe Macmillan Building, 4 Crinan Street, London N1 9XWTel: 44 207 833 4000, Fax: 44 207 843 4996Managing Director: Steven InchcoombePublishing Director: Alison MitchellEditor-in-Chief, Nature Publications: Philip CampbellMarketing Director: Della SarDirector of Web Publishing: Timo Hannay

NPG Nature Asia-PacificChiyoda Building, 2-37 Ichigayatamachi, Shinjuku-ku, Tokyo 162-0843Tel: 81 3 3267 8751, Fax: 81 3 3267 8746Publishing Director — Asia-Pacific: David SwinbanksAssociate Director: Antoine E. BocquetManager: Koichi NakamuraOperations Director: Hiroshi MinemuraMarketing Manager: Masahiro YamashitaAsia-Pacific Sales Director: Kate YoneyamaAsia-Pacific Sales Manager: Ken Mikami

DISPLAY ADVERTISING display@natureny.com (US/Canada) display@nature.com (Europe) nature@natureasia.com (Asia)Global Head of Advertising and Sponsorship: Dean Sanderson, Tel: (212) 726 9350, Fax: (212) 696 9482Global Head of Display Advertising: Andrew Douglas, Tel: 44 207 843 4975, Fax: 44 207 843 4996Asia-Pacific Sales Director: Kate Yoneyama, Tel: 81 3 3267 8765, Fax: 81 3 3267 8746Display Account Managers:Global Account Development Manager: Graham Combe, Tel: 44 207 843 4914, Fax: 44 207 843 4749New England: Sheila Reardon, Tel: (617) 399 4098, Fax: (617) 426 3717New York/Mid-Atlantic/Southeast: Jim Breault, Tel: (212) 726 9334, Fax: (212) 696 9481Midwest: Mike Rossi, Tel: (212) 726 9255, Fax: (212) 696 9481West Coast South: George Lui, Tel: (415) 781 3804, Fax: (415) 781 3805West Coast North: Bruce Shaver, Tel: (415) 781 6422, Fax: (415) 781 3805Germany/Switzerland/Austria: Sabine Hugi-Fürst, Tel: 41 52761 3386, Fax: 41 52761 3419United Kingdom/Ireland: Jeremy Betts, Tel: 44 207 843 4968, Fax: 44 207 843 4749Scandinavia/Iceland/Spain/Portugal: Evelina Rubio-Hakansson, Tel: 44 207 843 4079, Fax: 44 207 843 4749France/Belgium/The Netherlands/Italy/Israel/Eastern Europe: Nicola Wright, Tel: 44 207 843 4959, Fax: 44 207 843 4749Asia-Pacific Sales Manager: Ken Mikami, Tel: 81 3 3267 8765, Fax: 81 3 3267 8746Greater China/Singapore: Gloria To, Tel: 852 2811 7191, Fax: 852 2811 0743

NATUREJOBS naturejobs@natureny.com (US/Canada) naturejobs@nature.com (Europe) nature@natureasia.com (Asia)US Sales Manager: Ken Finnegan, Tel: (212) 726 9248, Fax: (212) 696 9482European Sales Manager: Dan Churchward, Tel: 44 207 843 4966, Fax: 44 207 843 4596Asia-Pacific Sales Manager: Ayako Watanabe, Tel: 81 3 3267 8765, Fax: 81 3 3267 8746

SITE LICENSE BUSINESS UNITAmericas: Tel: (888) 331 6288 institutions@natureny.comAsia/Pacific: Tel: 81 3 3267 8751 institutions@natureasia.comAustralia/New Zealand: Tel: 61 3 9825 1160 nature@macmillan.com.auIndia: Tel: 91 124 2881054/55 npgindia@nature.comROW: Tel: 44 207 843 4759 institutions@nature.com

CUSTOMER SERVICE www.nature.com/helpSenior Global Customer Service Manager: Gerald CoppinFor all print and online assistance, please visit www.nature.com/helpPurchase subscriptions:Americas: Nature Neuroscience, Subscription Dept., 342 Broadway, PMB 301, New York, NY 10013-3910, USA. Tel: (866) 363 7860, Fax: (212) 334 0879Europe/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 358Asia-Pacific: Nature Neuroscience, NPG Nature Asia-Pacific, Chiyoda Building, 2-37 Ichigayatamachi, Shinjuku-ku, Tokyo 162-0843. Tel: 81 3 3267 8751, Fax: 81 3 3267 8746India: Nature Neuroscience, NPG India, 3A, 4th Floor, DLF Corporate Park, Gurgaon 122002, India. Tel: 91 124 2881054/55, Fax: 91 124 2881052

REPRINTS reprint@boston.nature.comNature Neuroscience, Reprint Department, Nature Publishing Group, 75 Varick Street, Fl 9, New York, NY 10013-1917, USA.For commercial reprint orders of 600 or more, please contact:UK Reprints: Tel: 44 1256 302 923, Fax: 44 1256 321 531US Reprints: Tel: (617) 494 4900, Fax: (617) 494 4960

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

i

volume 12 number 8 AuGuST 2009

Nature Neuroscience (ISSN 1097-6256) is published monthly by Nature Publishing Group, a trading name of Nature America Inc. located at 75 Varick Street, Fl 9, New York, NY 10013-1917. Periodicals postage paid at New York, NY and additional mailing post offices. Editorial Office: 75 Varick Street, Fl 9, New York, NY 10013-1917. Tel: (212) 726 9319, Fax: (212) 696 0978. Annual subscription rates: USA/Canada: US$225 (personal), US$3,060 (institution). Canada add 7% GST #104911595RT001; Euro-zone: €287 (personal), €2,430 (institution); Rest of world (excluding China, Japan, Korea): £185 (personal), £1,570 (institution); Japan: Contact NPG Nature Asia-Pacific, Chiyoda Building, 2-37 Ichigayatamachi, Shinjuku-ku, Tokyo 162-0843. Tel: 81 (03) 3267 8751, Fax: 81 (03) 3267 8746. POSTMASTER: Send address changes to Nature Neuroscience, Subscriptions Department, 342 Broadway, PMB 301, New York, NY 10013-3910. 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 © 2009 Nature Publishing Group. Printed in USA.

e d i To r i A l

955 Changes in house rules

Co r r e S P o n d e n C e

957 PDLIM5 is not a neuronal CaV2.2 adaptor protein

n e w S A n d v i e w S

959 Who let the spikes out?Chris G Dulla & John R Huguenard see also p 996

961 Practice makes perfect, even for breathingJack L Feldman, Kaiwen Kam & Wiktor A Janczewski see also p 1028

963 Should I stay or should I go: genetic bases for uncertainty-driven explorationJérôme Sallet & Matthew F S Rushworth see also p 1062

965 Inactivating the activated: identifying functions of specific neural networksRachel J Smith & Gary Aston-Jones see also p 1069

b r i e f Com m u n i C AT i o n S

967 Resolving single cone inputs to visual receptive fieldsL C Sincich, Y Zhang, P Tiruveedhula, J C Horton & A Roorda

970 Representation of internal models of action in the autistic brainC C Haswell, J Izawa, L R Dowell, S H Mostofsky & R Shadmehr

A r T i C l e S

973 The genesis of cerebellar interneurons and the prevention of neural DNA damage require XRCC1Y Lee, S Katyal, Y Li, S F El-Khamisy, H R Russell, K W Caldecott & P J McKinnon

981 A trophic role for Wnt-Ror kinase signaling during developmental pruning in Caenorhabditis elegansY Hayashi, T Hirotsu, R Iwata, E Kage-Nakadai, H Kunitomo, T Ishihara, Y Iino & T Kubo

988 A discrete alcohol pocket involved in GIRK channel activationP Aryal, H Dvir, S Choe & P A Slesinger

Shu and colleagues show that two sodium channel subtypes, a high-

threshold Nav1.2 and a low-threshold Nav1.6, are asymmetrically distributed in the axon initial segment (AIS). This

asymmetrical distribution explains many of the unique properties of the

AIS, including its generation of backpropagating action potentials.

Cover design by Jiafeng Zhao.(pp 959 and 996)

Action models in children with autism(p 970)

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

iii

volume 12 number 8 AuGuST 2009

nATure neuroSCienCe

996 Distinct contributions of Nav1.6 and Nav1.2 in action potential initiation and backpropagationW Hu, C Tian, T Li, M Yang, H Hou & Y Shu see also p 959

1003 Ca2+ and calmodulin initiate all forms of endocytosis during depolarization at a nerve terminalX-S Wu, B D McNeil, J Xu, J Fan, L Xue, E Melicoff, R Adachi, L Bai & L-G Wu

1011 SAP97 and CASK mediate sorting of NMDA receptors through a previously unknown secretory pathwayO Jeyifous, C L Waites, C G Specht, S Fujisawa, M Schubert, E I Lin, J Marshall, C Aoki, T de Silva, J M Montgomery, C C Garner & W N Green

1020 Balanced gene regulation by an embryonic brain ncRNA is critical for adult hippocampal GABA circuitryA M Bond, M J W VanGompel, E A Sametsky, M F Clark, J C Savage, J F Disterhoft & J D Kohtz

1028 Genetic identification of an embryonic parafacial oscillator coupling to the preBötzinger complexM Thoby-Brisson, M Karlén, N Wu, P Charnay, J Champagnat & G Fortin see also p 961

1036 Cocaine-evoked synaptic plasticity: persistence in the VTA triggers adaptations in the NAcM Mameli, B Halbout, C Creton, D Engblom, J R Parkitna, R Spanagel & C Lüscher

1042 Synaptic inhibition of Purkinje cells mediates consolidation of vestibulo-cerebellar motor learningP Wulff, M Schonewille, M Renzi, L Viltono, M Sassoè-Pognetto, A Badura, Z Gao, F E Hoebeek, S van Dorp, W Wisden, M Farrant & C I De Zeeuw

1050 Disparity- and velocity-based signals for three-dimensional motion perception in human MT+B Rokers, L K Cormack & A C Huk

1056 Sensory transformations and the use of multiple reference frames for reach planningL M M McGuire & P N Sabes

1062 Prefrontal and striatal dopaminergic genes predict individual differences in exploration and exploitationM J Frank, B B Doll, J Oas-Terpstra & F Moreno see also p 963

T e C H n i C A l r e P o r T

1069 Targeted disruption of cocaine-activated nucleus accumbens neurons prevents context-specific sensitizationE Koya, S A Golden, B K Harvey, D H Guez-Barber, A Berkow, D E Simmons, J M Bossert, S G Nair, J L Uejima, M T Marin, T B Mitchell, D Farquhar, S C Ghosh, B J Mattson & B T Hope see also p 965

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

See back pages.

Cerebellar interneuron differentiation relies on a DNA repair mechanism

(p 973)

Wnt signaling prevents neurite pruning(p 981)

Three-dimensional motion perception in the human brain

(p 1050)

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature neuroscience volume 12 | number 8 | AuGuST 2009 955

e d i to r i a l

more important to be clear about the relative contributions of each author. There is no prescribed format for these statements, and authors are free to tailor them to their particular needs. Authors can state that all authors contributed to all aspects of the work or that specific members contributed to the design, commented on the manuscript and so forth. We only require that all authors in the study be mentioned in these statements.

Related to this issue, we have also clarified our policies on authorship (http://www.nature.com/authors/editorial_policies/authorship.html). For collaborative studies, we now require that at least one member of each collaboration group, typically the senior author, take responsibility for their group’s contribution. Minimally, these responsibilities include ensuring that the original data is preserved and obtainable for re- analysis, ensuring that the data reported are representative of the original data and that image manipulations are in accordance with the journal’s guidelines (http://www.nature.com/authors/editorial_ policies/image.html). Senior authors are also responsible for ensuring the proper sharing of data, materials, reagents or algorithms presented in the paper. Corresponding authors, although implicitly responsible for the accuracy and integrity of the data, are solely responsible for communication with the journal and in managing communication with the co-authors. They are required to ensure that all the co-authors are aware of the content of the manuscript and the author list. They must inform co-authors of any issues that arise pre- or post- publication and are solely responsible for ensuring the accuracy of the proofs, including ensuring that the names of all of the co-authors are spelled accurately and that their affiliations are correctly listed. We now require all corresponding authors to certify that they are aware of and that they agree with these policies as part of the online submission process.

Our new guidelines will come as no surprise to most authors, who follow these best-practice rules anyway. Would they have prevented some of the cases of misconduct in science? Probably not, but they may have helped co-authors become more aware of (and accountable for) their roles in a paper. For example, an official inquiry into the Jan Hendrik Schön misconduct affair a few years ago, regarding a string of high-profile papers in condensed matter physics with fabricated data, concluded that co-authors on the Schön had broadly met their responsibilities, but that the committee was unable to make specific judgments, as there was no clear consensus on the responsibilities of individual participants in collaborative research (http://publish.aps.org/reports/lucentrep.pdf). Establishing clear guidelines on responsibilities will hopefully help address some of these issues. We hope you will agree that these guidelines help to clarify the record and increase transparency and accountability in the reporting of scientific data. L

readers of Nature Neuroscience will have noticed a new section called Online Methods in our articles, technical reports and resources. We are combining our Methods and Supplementary

Methods sections and publishing them as a single Online Methods section. The Online Methods remains an integral part of the main paper, and a PDF download of the paper will automatically contain the Online Methods section. Readers of the print version will be directed to our website to access the methods.

A cornerstone of academic research is that the papers are sufficiently detailed to allow for the full assessment and reproduction of the data and the methods that the authors employed to reach their conclusions. A criticism that has been levied against many high-impact journals is that the methods sections are not sufficiently detailed to allow for this careful scrutiny. Most of our readers access our articles online; moving the methods online allows us to display more detailed methods in a cost-effective manner and helps us move to the ‘paper of the future’. As in our previous Methods sections, these Online Methods are copyedited and can contain references. Although these references are only included in the online version of the paper, they will be taken into account for impact factor calculations.

Despite this move, however, we will continue to ask that authors keep their methods sections to a limit of 2,000 words. We are aware that readers’ time is valuable, and this word limit is sufficient to accommodate the Methods and Supplementary Methods sections that we previously published in a typical Nature Neuroscience paper. We have already found that there may be a few exceptions that require a relaxing of this word limit and will continue to evaluate such papers on a case-by-case basis. The publication requirements that we previously established for methods must still be met (authors must, for example, still include information about institutional committees that approved the experiments and statements on informed consent or animal welfare as necessary).

We have also recently updated our guide to authors (http://www.nature.com/neuro/authors/index.html) to make the Author Contribution statements mandatory and to make author responsibilities clearer. In the past, at the acceptance of a paper in Nature Neuroscience, we suggested that the corresponding author add a statement detailing the author contributions in the published version of the paper. Most of our papers previously carried such author contribution statements. Recently, the Nature Publishing Group announced that all Nature-branded journals will require such statements to accompany all original research articles.

The purpose of the Author Contribution statement is to give co-authors the credit that they are due. Increasingly, research is carried out by collaborations from different laboratories, and it has become

Changes in house rulesNature Neuroscience will now only publish methods online. We have also amended our rules to clarify authors’ and co-authors’ responsibilities.

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature neuroscience volume 12 | number 8 | august 2009 957

co r r e s p o n d e n c e

PDLIM5 is not a neuronal CaV2.2 adaptor proteinTo the editor: PDLIM5 (postsynaptic density protein–95, discs-large, ZO1, Lin-11–Isl-1–Mec-3), formerly known as enigma homolog (ENH) is a ~63-kDa cytoplasmic protein composed of a PDZ domain at the N terminus and three consecutive LIM domains at the C terminus (Supplementary Fig. 1)1. Early studies showed that protein kinase C (PKC) binds to the LIM domains1, but did not identify a role for PDLIM5 in cell function. On the basis of a combination of electrophysiological and biochemical approaches, Maeno-Hikichi et al.2 reported that PDLIM5 localizes to presynaptic terminals and developed the hypothesis that the protein interacts with CaV2.2, a presynaptic calcium channel type that is known to be sensitive to enhancement by PKC. They proposed that PDLIM5 functions as an adaptor that links PKCε to the channel, facilitating the activity of the kinase on channel kinetics and, in a second study, that this association was enhanced by Ca2+ (ref. 3). Maeno-Hikichi et al.2 suggested that regulation of the channel via this complex may be important in the modulation of synaptic strength. Since these reports, however, there has been no further analysis of this putative calcium channel modulation mechanism by other laboratories. Recent studies have suggested an association between PDLIM5 with severe psychiatric disease, such as schizophrenia and depression4-6. In these studies, the CaV2.2-PKCε adaptor function of PDLIM5, as suggested by Maeno-Hikichi et al.2, is generally raised as a putative pathogenic mechanism.

Our original intent was to explore the interaction between PDLIM5/CaV2.2 at an intact presynaptic terminal. We used patch-clamp recording to test for an enhancement of PKC action on CaV2.2 current (see Supplementary Methods) in freshly dissociated rat dorsal root ganglion (DRG) neurons by infusing the cells with a GST–LIM1-3 fusion protein (Supplementary Fig. 1; for protocols, see Supplementary Methods). However, in contrast with what would be predicted from the earlier report, GST–LIM1-3 did not enhance but instead blocked the action of the PKC activator (Fig. 1).

Although we have not explored this action further, we speculate that this effect might be most simply attributed to binding of LIM1–3 to PKC and its removal as a potential CaV2.2 modulating agent. This finding prompted us to retest whether PDLIM5 is a component of a molecular complex with CaV2.2 by biochemical and immunocytochemical analysis. This was facilitated by the availability of a new commercially available selective polyclonal mouse antibody to PDLIM5, PDLIM5pM (Supplementary Table 1), which is effective for both western blotting and immunoprecipitation (Supplementary Fig. 2), which we used in combination with Ab571, our own well-characterized high-affinity antibody to CaV2.2 (refs. 7,8).

Immunostaining at the chick calyx presynaptic terminal and western blotting of solubilized rat brain synaptosomes (Supplementary Fig. 3) confirmed that PDLIM5 is present at presynaptic terminals of peripheral and central synapses and in two different vertebrate species. However, staining intensities for PDLIM5 and CaV2.2 did not covary at the calyx transmitter release face as assessed by quantitative staining intensity covariance analysis8 (intensity correlation quotient = 0.04 ± 0.02, n = 5, P > 0.05, t test; Supplementary Fig. 4), arguing against a transmitter release-site association. Furthermore, PDLIM5 and CaV2.2 failed to co-immunoprecipitate from purified synaptosome membrane (n = 4; Fig. 2a), whole purified synaptosome (n = 3; Fig. 2b) or whole brain lysates (data not shown) whether we immunoprecipitated with PDLIM5pM or Ab571 and under conditions previously demonstrated to retain a broad range of CaV2.2 binding partners7–9. Co-immunoprecipitation also failed when we used protocols that were similar to those of the previous studies, including ‘crude adult rat brain membrane fraction’ lysate, solubilizing this fraction in ‘immunoprecipitation buffer’ (as in ref. 2) (Supplementary Fig. 5) or with elevated external Ca2+ (Fig. 2c)3. As a control for our methods, we were able to reproduce PDLIM5 co-immunoprecipitation with PKCε (Supplementary Fig. 5), as described previously1,2.

The main remaining difference between our study and that of Maeno-Hikichi et al.2 was the use of different antibodies to PDLIM5. To explore this possibility further, we obtained anti-ENH (provided by J.F. Zhang, Jefferson Medical College), a polyclonal antibody that, as we now understand it, was created using the same PDLIM5 peptide antigen and has the same name as that used in Maeno-Hikichi et al.2, but was raised in a different rabbit. Rat brain lysate western blots probed with this antibody generated a ladder of protein bands (Fig. 2d), raising specificity concerns. Immunoprecipitation using anti-ENH from rat brain crude membrane lysate did precipitate both PDLIM5 and

a GST GST+PMA

LIMPMA

1 nA

10 ms

1.1 GST GST + PMA

**

0.9

1.0

t = 0

n = 6

n = 6

**

120 60 0 60 120 180 240 300 3600.7

0.8

I/I

Time (s)

PMAn = 6

__

Time (s)

GST–LIM1–3 + PMA

b

Figure 1 The PDLIM5 C-terminal region, LIM1-3, did not enhance PKC-dependent facilitation of CaV2.2 current. (a) Inward calcium currents from freshly dissociated and plated rat DRG neurons with either GST alone or GST-LIM1–3 in the patch electrode (solid line, t = 0 s; dotted line, t = 60 s). Phorbol myristate acetate (PMA) (or carrier alone) was added to the bath at time 0. (b) Plot of current amplitude (ICa) normalized to time 0 for each cell before averaging. Error bars represent ± s.e. (n is the number of cells pooled from two separate experiments). GST + PMA was significantly larger than GST–LIM1-3 + PMA at 60 and 120 s (** indicates P < 0.02, Student’s t-test).

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

958 volume 12 | number 8 | august 2009 nature neuroscience

co r r e s p o n d e n c e

Eur. J. Neurosci. 26, 547–559 (2007).8. Li, Q. et al. J. Neurosci. 24, 4070–4081 (2004).9. Khanna, R., Zougman, A. & Stanley, E.F. J. Biochem.

Mol. Biol. 40, 302–314 (2007).10. Khanna, R., Li, Q., Sun, L., Collins, T.J. & Stanley, E.F.

Neuroscience 140, 1201–1208 (2006).

Maeno-Hikichi et al. reply: Our original conclusion that the PKC binding protein, enigma homolog (ENH), interacts specifically with both PKCe and N-type Ca2+ channels, forming a PKCε-ENH–Ca2+ channel macromolecular complex was based on a polyclonal antibody that used the last 18 amino acids of ENH as an epitope. We do not have any more stock of this original antibody. The Abnova antibody used by Stanley and colleagues was raised against the entire ENH protein (a total of 597 amino acids) and this may explain some of the differences between our results. However, the antibody that we supplied to Stanley and colleagues was an antibody that was similar to the one that we used in our original study and was also raised against the last 18 amino acids of ENH. We are not sure why Stanley and colleagues found this antibody to be nonspecific. In addition, it is not clear to us why inclusion of a fusion protein containing ENH binding domains in their recording pipette did not facilitate modulation of N-type Ca2+ channel activity by PKC in the recorded rat DRG neurons.

Ji-fang Zhang

Department of Molecular Physiology & Biophysics, Jefferson Medical College, Philadelphia, Pennsylvania, USA.e-mail: ji-fang.zhang@jefferson.edu

that this protein serves as an adaptor to link PKCε to CaV2.2. It should be noted that an important tool in our study, the highly specific and avid antibody to PDLIM5, PDLIM5pM, was not available at the time of the original study2 and the capture of CaV2.2 using their antibody to ENH would have been compelling evidence in support of such a complex. Hence, the CaV2.2-to-PKCε adaptor role for PDLIM5 was not unreasonable at that time. However, the results presented here do not support this hypothesis and instead indicate a re-examination of the function of PDLIM5 in the brain and also specifically at the presynaptic terminal, a quest that becomes more important by the reported association of this protein with serious psychiatric disorders.

Sabiha R Gardezi, Alexander M Weber, Qi Li, Fiona K Wong & Elise F Stanley

Laboratory of Synaptic Transmission, Genes and Development Division, Toronto Western Research Institute, Toronto, Canada. e-mail: estanley@uhnres.utoronto.ca

Note: Supplementary information, including acknowledgements and author contributions, is available on the Nature Neuroscience website.

1. Kuroda, S. et al. J. Biol. Chem. 271, 31029–31032 (1996).

2. Maeno-Hikichi, Y. et al. Nat. Neurosci. 6, 468–475 (2003).

3. Chen, Y., Lai, M., Maeno-Hikichi, Y. & Zhang, J.F. Cell. Signal. 18, 215–224 (2006).

4. Iga, J. et al. Neurosci. Lett. 400, 203–207 (2006).5. Li, C. et al. Int. J. Neuropsychopharmacol. 11, 27–34

(2008).6. Horiuchi, Y. et al. Biol. Psychiatry 59, 434–439 (2006).7. Khanna, R., Li, Q., Bewersdorf, J. & Stanley, E.F.

CaV2.2, (when probed with PDLIM5pM; Fig. 2e and Supplementary Fig. 6). However, we had reservations as to whether this truly reflected co- immunoprecipitation with PDLIM5 because the reverse, Ab571 immunoprecipitation, probing with ENH did not recover PDLIM5 (Supplementary Fig. 6). This reservation was confirmed by repeating the immunoprecipitation experiment after depleting antibody to PDLIM5 clones from anti-ENH using LIM1-3 immobilized on beads (termed anti-ENH(–LIM); Fig. 2d). Depletion was confirmed by a very weak band corresponding to the molecular weight of PDLIM5 in western blots (Fig. 2d) and its absence when used to probe immunoprecipitated proteins (Fig. 2e and Supplementary Fig. 6). However, despite the gross reduction in the antibody to PDLIM5 fraction, ENH(–LIM) still immunoprecipitated CaV2.2 (Fig. 2e). There were two simple explanations for this finding. Either anti-ENH contains antibody clones that bind directly to CaV2.2 or it captures the channel as a binding partner of a target protein other than PDLIM5. We favor the latter interpretation, as neither anti-ENH nor anti-ENH(–LIM) identified a band with a molecular weight that corresponds to CaV2.2 in western blots of brain lysates (Fig. 2d).

Thus, although we confirmed that PDLIM5 is present in presynaptic terminals, our electrophysiology, immunocytochemistry and biochemistry results argue against the idea

Figure 2 PDLIM5 and CaV2.2 do not co-immunoprecipitate from rat brain lysates. (a,b) Ab571 or PDLIM5pM were used to immunoprecipitate proteins from purified synaptosome membrane fraction10 (P2”, a) or whole synaptosome lysates (b) solubilized in modified RIPA lysis buffer (Supplementary Methods). Proteins were separated by SDS-PAGE and immunoblots were probed with Ab571 (top) or PDLIM5pM (bottom) from the same experiment. In both a and b, Ab571 identified CaV2.2 protein (~260 kDa) in the Ab571, but not PDLIM5pM immunoprecipitation, lanes and vice versa. A western blot of the P2” fraction identifying CaV2.2 and PDLIM5 protein bands is shown in the right lane of a (see also Supplementary Fig. 2). MIgG, mouse IgG; RIgG, rabbit IgG. (c) The P2” fraction was solubilized in RIPA lysis buffer with or without 3 mM EDTA and 5 mM EGTA and with the [Ca2+] set to 10 nM or 10 µM. PDLIM5 failed to co-immunoprecipitate with CaV2.2 at all of the Ca2+ conditions tested. (d) Anti-ENH or anti-ENH(–LIM) were used to probe rat brain P2 western blots. (e) Ab571, PDLIM5pM, anti-ENH or anti-ENH(–LIM) were used for immunoprecipitation of proteins from adult rat brain crude membrane fraction and immunoblots were probed with Ab571 (top) or PDLIM5pM (bottom). PDLIM5 was seen in the PDLIM5pM and anti-ENH IP lanes, but not in the anti-ENH(–LIM) immunoprecipitation lane. CaV2.2 was detected in the Ab571 and anti-ENH lanes, as expected, and also in the anti-ENH(–LIM) immunoprecipitation lane. The white dashed box indicates the location of the PDLIM5 band. The closed arrow indicates CaV2.2 and the open arrow indicates PDLIM5. * indicates the antibody protein bands. Blots grouped together are all from the same experiment. Anti-ENH and anti-ENH(–LIM) are referred to as ENH and ENH(–LIM), respectively.

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature neuroscience volume 12 | number 8 | august 2009 959

n e w s a n d v i e w s

The authors are at the Department of Neurology &

Neurological Sciences, Stanford University School

of Medicine, Stanford, California, USA.

e-mail: john.huguenard@stanford.edu

Who let the spikes out?Chris G Dulla & John R Huguenard

Quantitative immunostaining, electrophysiology and modeling show that two sodium channel isoforms are asymmetrically distributed in the axon initial segment. Their polarized distribution explains many of the unique properties of the axon initial segment, including its ability to both initiate spikes and guarantee subsequent backpropagation.

Are you impulsive or do you tend to be more deliberate? Have you ever felt the need to be cautious only to be dragged into something by a reckless accomplice? In this issue, Hu et al.1 provide compelling evidence for molecular peer-pressure: two sodium channel (NaCh) isoforms with different demeanors located in the axon initial segment (AIS), one of which is a bit cautious and the other is more impetuous. Neurons are continuously barraged by synaptic input that opens neurotransmitter receptors and induces changes in membrane potential (Vm). Once Vm becomes sufficiently elevated, voltage-gated NaChs open and initiate an action potential, or spike, fulfilling the neuron’s role as an information integrator. Previous studies have shown that the AIS, a structure at the juncture between the soma and the axon, is rich in NaChs2 and initiates action potentials3. Once initiated, spikes propagate in two directions: forward down the axon to cause neurotransmitter release by depolarizing presynaptic terminals4 and backwards through the soma and then on to the dendrites. Although the forward-propagating action potential transmits information to downstream postsynaptic neurons, the backpropagating action potential enables forms of synaptic plasticity5,6. The unique characteristics of the AIS that allow it to both initiate spikes with relative ease and then guarantee subsequent backpropagation have remained elusive.

Here, Hu et al.1 show, using quantitative immunostaining, electrophysiology (including the method of axonal bleb recording developed by one of the authors, Y. Shu) and computer modeling, that two NaCh subtypes, the high-threshold Nav1.2 and the low-threshold

Nav1.6, are asymmetrically distributed in the AIS, precisely localizing these NaChs in the complex topography of the neuron. Nav1.2 is found mainly in the 25 µm of the AIS that is closest to the soma and requires substantial depolarization for activation. Nav1.6, on the other hand, is found in more distal portions of the AIS, 25–50 µm from the soma, and is activated by relatively little depolarization7. This polarized configuration, low-threshold NaChs in the distal AIS flanked by high-threshold NaChs closer to the soma, creates a new blueprint of AIS function that explains many of the unique properties

of the AIS, including the faithful generation of backpropagating action potentials (Fig. 1).

In this new model, action potentials are detonated by NaV1.6 channels because of their low threshold for activation and high channel density8. NaV1.6 channels sit in the perfect location to allow their easy initiation of action potentials: distal to the incoming dendritic excitation and insulated from it by somatic inhibitory neurotransmission and a reserve pool of timid NaV1.2 channels in the proximal AIS. If synaptic depolarization makes it as far as the distal AIS, the trigger-happy NaV1.6

Figure 1 A new blueprint for action potential initiation in the AIS. Excitatory neurotransmission onto the dendrites of a layer V pyramidal cell (green oval) causes depolarization of the postsynaptic Vm. This local depolarization moves electrotonically toward the soma (green arrow), where it can be shunted by inhibitory GABAergic neurotransmission (red oval). However, with sufficient synaptic input, depolarization will spread beyond the soma and into the AIS (inset). Once incoming depolarization (green arrow) reaches the AIS (1), it will first enter an area rich in NaV1.2 (blue). These channels are the ‘cautious’ high-threshold subtype, so depolarization will pass through without rapidly activating the NaV1.2 channels (green arrow, masked by blue). When the wave of depolarization reaches the trigger-happy low-threshold NaV1.6 channels (yellow), however, they will quickly open (2) and initiate an inward sodium current (red arrow). This will rapidly depolarize Vm in the distal AIS, activating other nearby NaV1.6 channels, causing a chain reaction of NaCh opening and initiating a forward-propagating action potential (AP, 3). Because NaV1.2 channels were bypassed by the initial synaptic depolarization, they are available for activation, rather than being in an inactivated state. When NaV1.6 channels open, they drive NaV1.2 channel activation (4), inducing a secondary wave of inward sodium currents and initiating a backpropagating action potential (5). Because NaV1.6 channels will be in their inactive state, NaV1.2 channel opening will not induce a secondary forward-propagating action potential.

Kim

Cae

sar

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

960 volume 12 | number 8 | august 2009 nature neuroscience

n e w s a n d v i e w s

the threshold for generating a backpropagating somatodendritic action potential is controlled by the hesitant NaV1.2. It will be exciting to see which other unique biophysical parameters of NaV1.2 and NaV1.6 are relevant to additional aspects of spike generation and neuronal excitability. Will the faster recovery from inactivation seen in NaV1.2 mean that they are more responsible for action potential generation during high-frequency firing? Or will the ability of NaV1.6 to maintain high current amplitude during repeated activation put it in the driver’s seat during high-frequency spiking7? Will the differential effects of drugs modulating NaCh properties (that is, phenytoin, carbamazepine, lamotrigine, etc.) be better understood now that we know more about the specific ion channels mediating action potential generation? With this detailed picture of the spike-generation machinery, we are much better equipped to answer these and other pressing questions.

Finally, our understanding of spike generation has truly paralleled our technical advances in electrophysiological and imaging techniques. From early intracellular recordings from motoneurons3 to our ability to make simultaneous patch clamp recordings from a single neuron at multiple locations to in vivo recording of action potential threshold, our knowledge of spike initiation continues to grow. Now techniques such as voltage and sodium imaging and bleb recording are rapidly advancing our ability to characterize excitability in specific neuronal substructures. The most intriguing question that Hu et al.1 leave unanswered is how is the NaCh distribution built and maintained. Which cytoskeletal components, signaling molecules and NaCh domains are responsible? Does inappropriate trafficking or anchoring of NaChs underlie pathological states? The ion channel trafficking and cytoskeletal interaction that have been so elegantly studied in the synapse now must be understood in the AIS.

1. Hu, W. et al. Nat. Neurosci. 12, 996–1002 (2009).2. Catterall, W.A. J. Neurosci. 1, 777–783 (1981).3. Coombs, J.S., Curtis, D.R. & Eccles, J.C. J. Physiol.

(Lond.) 139, 232–249 (1957).4. Katz, B. The Release of Neural Transmitter Substances

(Liverpool University Press, Liverpool, UK, 1969).5. Markram, H., Lubke, J., Frotscher, M. & Sakmann, B.

Science 275, 213–215 (1997).6. Magee, J.C. & Johnston, D. Science 275, 209–213

(1997).7. Rush, A.M., Dib-Hajj, S.D. & Waxman, S.G. J. Physiol.

(Lond.) 564, 803–815 (2005).8. Kole, M.H. & Stuart, G.J. Nat. Neurosci. 11, 1253–1255

(2008).9. Colbert, C.M. & Johnston, D. A. J. Neurosci. 16,

6676–6686 (1996).10. Kole, M.H. et al. Nat. Neurosci. 11, 178–186 (2008).11. Song, A.H. et al. Cell 136, 1148–1160 (2009).12. Zhou, D. et al. J. Cell Biol. 143, 1295–1304 (1998).13. Pan, Z. et al. J. Neurosci. 26, 2599–2613 (2006).14. Naundorf, B., Wolf, F. & Volgushev, M. Nature 440,

1060–1063 (2006).15. McCormick, D.A., Shu, Y. & Yu, Y. Nature 445, E1–E2

(2007).

of the AIS, a structure that is notorious for its dense cytoskeleton11. The AIS is rich in the adaptor protein ankyrin G, which helps cluster both NaChs12 and potassium channels13. It was demonstrated10 that disruption of the actin cytoskeleton, and presumably its ability to stabilize ankyrin G, caused a threefold increase in the sodium current that could be recorded in the AIS of layer V pyramidal neurons10. This suggests that the AIS is indeed highly enriched in NaChs, but rigid cytoskeletal scaffolding somehow prevents ideal attachment of a patch pipette. These results thus confirmed immunohistological and sodium-imaging findings and reconciled previous electrophysiological findings. Overall, these results highlight the high value neurons place on bidirectional spike propagation. They have evolved an anatomical distribution of NaChs at a location distinct from that of incoming synaptic input and developed an extensive cytoskeletal system to ensure its stability.

Hu et al.1 also address a recent controversy in the spike generation field: the possibility that NaCh activation is a cooperative process14,15. When action potentials are recorded from the soma of layer V cortical neurons, their onset is so rapid that some believe they cannot be described using classic Hodgkin-Huxley models, but can be recreated if NaCh gating is cooperative. According to the cooperative gating model, the statistical probability of any given channel opening in an environment rich with NaChs, such as the AIS, would not only be determined by Vm, but also by the open state of nearby NaChs. However, Hu et al.1 report that neither partial blockade of voltage-gated NaChs with tetrodotoxin nor decreasing NaCh currents with a low-sodium buffer alters the voltage dependence of channel activation. If NaCh activation were cooperative, one would expect that removing a subset of NaChs from the active pool of channels with tetrodotoxin would alter channel activation, whereas reducing the sodium driving force would not. This result should lay to rest the notion that unique, cooperative, NaCh gating occurs in the AIS to initiate action potentials and supports the idea that the rapid onset of action potentials in the soma results from recording distally from the site of action potential initiation.

Is there a new integrated view of spike initiation in pyramidal neurons? Hu et al.1 combined their electrophysiological and immunohistochemical findings with elegant modeling experiments to confirm the roles of NaV1.6 and NaV1.2. By altering the relative amounts of NaV1.2 and NaV1.6 in their model, they found that the forward-propagating action potential threshold is almost completely dependent on the impulsive NaV1.6, whereas

figures that the neuron deserves to spike. Once NaV1.6 channels are activated, they rapidly depolarize the nearby area, coercing the hesitant NaV1.2 channels in the proximal AIS to open and generate a backpropagating action potential. Having a reserve of high-threshold NaV1.2 channels proximal to the soma, the majority of which fail to open in response to the initial synaptic depolarization, provides a source of non-inactivated NaChs that are ready and waiting to initiate a backpropagating action potential. Furthermore, because NaV1.6 channels in the distal AIS have entered an inactive state by the time NaV1.2 channels open, a second forward-propagating action potential is prevented. Although elements of this scheme are not perfectly clear, this mechanism of spike initiation followed by faithful generation of a backpropagating action potential is both alluring and exciting.

The initiation of action potentials in the AIS is not a new concept. In fact the mechanism proposed by Hu et al.1 draws on years of work from groups dedicated to understanding the specific mechanism of spike generation. It was initially reported over 50 years ago that the action potential appears first in the AIS of motoneurons and is followed by a backpropagating somatodendritic action potential3. As electrophysiological and imaging techniques advanced, so did our understanding of spike initiation. Pioneering studies9 used simultaneous recording from the soma and AIS of subicular neurons to demonstrate that Vm rises more rapidly in the AIS during an action potential, which occurs presumably as a result of the NaV1.6 localization found by Hu et al.1. In the soma, a previous study9 showed that the onset of a spike occurs more slowly initially, resulting from what we now think is NaV1.6-mediated depolarization in the distal AIS, and is then followed by a rapid increase in Vm which now appears to be driven by NaV1.2 activation in the proximal AIS. This study9 also showed that somatic action potential threshold is established by sodium channels ≈50 µm from the soma, where Hu et al.1 have localized NaV1.6. Although the authors of that study did not know the identity of the NaChs subtypes driving action potentials, they proposed the idea of a ‘heminode’ beyond the AIS where action potentials originate, an idea that is conceptually validated by Hu et al.’s1 finding of a high concentration of NaV1.6 channels in the distal AIS.

Recently, a study10 unraveled the long– standing mystery of why previous recordings haven’t revealed a higher density of NaChs in the AIS than elsewhere in the neuron if AIS NaCh density explains spike initiation. Answering this question required a literal deconstruction

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature neuroscience volume 12 | number 8 | august 2009 961

n e w s a n d v i e w s

The authors are in the Department of Neurobiology,

David Geffen School of Medicine, University of

California Los Angeles, Los Angeles, California, USA.

e-mail: feldman@ucla.edu

Practice makes perfect, even for breathingJack L Feldman, Kaiwen Kam & Wiktor A Janczewski

Breathing relies on a respiratory rhythm generator. A study characterizes an early emerging oscillatory group of Phox2b-expressing parafacial cells that entrain and couple with the preBötzinger Complex at the onset of fetal breathing.

Successful team performance requires practice; stepping onto the field with new players who have just met is not a recipe for success. Even if all the players have been working out individually, practicing together is essential to coordinate plays and to develop a successful team rhythm. At birth, mammals are thrust into a game of survival and to play this game must reliably breathe and suckle. To accomplish these goals, mammals practice breathing in utero. These episodic rhythmic fetal breathing movements (FBMs) are required for proper lung development and assure that respiratory muscles and the neural system that drives them are functional and coordinated at birth. In rodents1, the onset of FBMs involving the diaphragm is coincident with the onset of rhythmicity in the preBötzinger Complex (preBötC), a medullary area that is essential for respiration2. In this issue, Thoby-Brisson et al. have identified a second rhythmogenic area, the embryonic parafacial nucleus (e-pF), as being important in the neurogenesis of respiratory rhythms. They present elegant and definitive experiments showing that the e-pF is the source of the earliest behaviorally relevant rhythm for FBMs, starting at embryonic day 14.5 (E14.5) in mouse3. Furthermore, they show that it contributes substantially to the subsequent onset and development of rhythmicity at E15.5 in the preBötC, the presumptive onset and maintenance of FBMs, and reliable breathing at birth.

Since its identification in 19904, the preBötC has increasingly assumed the mantle of the principal rhythm generator for breathing. The preBötC drives inspiratory muscle activity and is the only known group of neurons that, when silenced, promptly results in a complete arrest of breathing, sufficient to asphyxiate conscious, unanesthetized adult rodents5. The coincident onset of rhythmicity in the preBötC and FBMs in rodents suggests a causal relationship. In 1996, however, it was shown

that the preBötC is not sufficient to secure robust breathing during the perinatal period6. Mice with a deletion of the transcription factor Egr2, also known as Krox20, have alterations affecting rhombomeres 3 and 5 that remove an embryonic rhythmic source near the facial motor nucleus (nVII), resulting in markedly depressed breathing at birth. These results suggest that the e-pF, defined as the population of neurons flanking and partially capping the lateral aspect of nVII and extending approximately 200 µm caudal to nVII, is essential for driving breathing rhythm at birth when it acts as an ‘anti-apnea’ center7.

In this issue, Thoby-Brisson et al.3 examine the ontogeny of the e-pF and its relationship to the preBötC during prenatal development and reveal intriguing functional interactions in the respiratory rhythm generator. They treated blocks and slices of medulla from embryonic mice with a calcium indicator dye that fluoresces to reflect neuronal activity. Observing the ventral face of the embryonic brainstem, they found that the very first rhythmic neurons appeared at E14.5 (Fig. 1a). These bilateral neuronal populations each formed a cap over the ventrolateral and caudal part of nVII (that is, the e-pF). The e-pF oscillator on each side of the medulla is composed of about 260 Phox2b-positive glutamatergic neurons that are derived from Egr2-expressing progenitors. About 70% of these neurons express the neurokinin 1 receptor, which is also a marker for critical preBötC neurons2. Using pharmacology and knockout mice, Thoby-Brisson et al.3 found that rhythm generation in the e-pF network appears to be independent of gluta-matergic synaptic transmission and opioid modulation, relying instead on a riluzole- and carbenoxolone-sensitive mechanism. This suggests the involvement of a persistent Na+ current and functional gap junction coupling. However, glutamatergic synaptic transmission is necessary for synchrony across the midline between bilateral e-pF areas (Fig. 1b).

PreBötC neurons begin to oscillate 1 d later (E15.5) in synchrony with the e-pF. These two regions can oscillate independently; when the en bloc brainstem is completely transected between the preBötC and the e-pF, both

segments continue to oscillate endogenously. However, the frequency of oscillation is altered in both regions (Fig. 1a). These changes are probably the results of either the removal of interactions between the two oscillators and/or a modification in common, modulatory inputs, such as the raphe or locus coerelus. Although the e-pF contributes substantially to the establishment of a normal rhythmic activity in the preBötC, the e-pF does not appear to be essential for preBötC development. In mutants lacking the Egr2 gene, there is no rhythmic activity around nVII at E15.5, probably resulting from a loss of neurons that express Phox2b or neurokinin 1 receptor in the expected e-pF region. The respiratory rhythm measured in the hypoglossal nerve is still present, although it is slowed to half of the frequency of that observed in wild-type mice. Transection between the presumptive e-pF and the preBötC has no effect on this rhythm, suggesting that its origin is the preBötC, which presumably developed in the absence of the e-pF. At birth, breathing in these mutants is slow and variable, and most of the mice die shortly after birth. However, injection of an opioid receptor antagonist after birth can rescue the breathing defect and markedly improve mutant viability. The e-pF may therefore be essential for overcoming preBötC depression caused by the substantial opiate surge at birth2.

Although the e-pF is important during practice and in the earliest portion of the game, what happens after birth? Convergent data from many laboratories point to the parafacial region as a potential rhythmic source for breathing in postnatal rodents, and three main lines of supporting evidence are highlighted here. First, on the basis of its projections to the medullary respiratory network, we identified and named the retrotrapezoid nucleus (RTN)8, a small region that is ventral to nVII and is demarcated by neurons expressing the transcription factor Phox2b. We speculated that it was a site for central chemoreception8 and a potential respiratory oscillator4. Humans with mutations affecting Phox2b have congenital central hypoventilation syndrome, which is characterized by an inability to sustain robust breathing during sleep and a marked

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

962 volume 12 | number 8 | august 2009 nature neuroscience

n e w s a n d v i e w s

Figure 1 Development and properties of the respiratory rhythm generator. (a) Rhythmic activity appears earliest in the e-pF at E14.5 and only inconsistently drives FBMs in nVII and the hypoglossal nerve (XIIn). At E15.5, with the appearance of the preBötC, rhythmic FBMs are reliably generated. Transections do not eliminate rhythmic activity in either the e-pF or the preBötC, but the frequency is altered. After birth, the preBötC controls inspiratory motor activity, as recorded in the phrenic nerve (C4/phr), whereas a parafacial region, the pFRG/RTN, whose precursor is likely the e-pF, generates expiratory-modulated motor activity in abdominal muscles recorded in L1 and serves as a chemosensory area. (b) Schematic of the prenatal respiratory rhythm generator circuit. Arrows are schematic and may represent mono- or polysynaptic connections through intervening areas. Rhythmic activity in the e-pF is blocked by riluzole (RIL). Rhythmic activity in the preBötC can be silenced by DAMGO or CNQX.

may be changes in inspiratory-modulated Cl–-dependent inputs, presumably from the preBötC, which are depolarizing at E15.5, but are hyperpolarizing after birth (Fig. 1a). Second, it is not clear under what conditions the RTN/pFRG is rhythmic in the adult rat. Under resting conditions in anesthetized rats, there is little, if any, rhythmic activity in RTN15; there is also very little, if any, active expiration. Third, Egr2–/– mutants retain their responsiveness to CO2 after birth, whereas mutants lacking Phox2b-derived neurons do not. Perhaps there are two groups of Phox2b neurons, one of which is critical for chemoreception (RTN) and the other of which is rhythmogenic (e-pF/pFRG).

Thoby-Brisson et al.3 convincingly demonstrate that the e-pF is important in the ontogeny of rhythmic circuits for FBMs. They also provide tantalizing details regarding its function and its interactions with the preBötC. These findings contribute substantially to our understanding of how the neural network underlying the vital motor behavior of breathing prepares to perform for a lifetime.

1. Kobayashi, K., Lemke, R.P. & Greer, J.J. J. Appl. Physiol. 91, 316–320 (2001).

2. Feldman, J.L. & Del Negro, C.A. Nat. Rev. Neurosci. 7, 232–242 (2006).

3. Thoby-Brisson, M. et al. Nat. Neurosci. 12, 1028–1035 (2009).

4. Feldman, J.L., Connelly, C.A., Ellenberger, H.H. & Smith, J.C. Eur. J. Neurosci. 3 Suppl, 171 (1990).

5. Tan, W. et al. Nat. Neurosci. 11, 538–540 (2008).6. Jacquin, T.D. et al. Neuron 17, 747–758 (1996).7. Borday, C. et al. Prog. Biophys. Mol. Biol. 84, 89–106

(2004).

however, when O2 consumption and CO2 production rise substantially, the activity of RTN/pFRG neurons may become increasingly rhythmic (Fig. 1a).

To help understand the distinct prenatal development of the preBötC and e-pF, we considered their evolutionary origin, a perspective that, although speculative, suggests a basis for their functional roles postnatally. The e-pF, which develops first, represents the phylogenetically ancient rhythm generator that drove breathing in aquatic vertebrates, whereas the preBötC represents the newer oscillator that emerged with the evolution of the lung and its complement of muscles. The evolutionary appearance of the diaphragm in mammals enabled a highly efficient inspiratory-driven pattern at rest that is sufficient to support endothermy. This would have led to the dominance of the preBötC at rest, with RTN/pFRG quiescent at rest, but becoming rhythmic to produce active expiration necessary for higher levels of ventilation, such as during exercise.

The most parsimonious interpretation, then, is that the e-pF becomes the pFRG. However, several issues warrant further investigation. First, in neonatal en bloc preparations, the activity pattern of pFRG neurons is markedly different from patterns at E14.5/E15.5. The inspiratory- modulated pattern in the e-pF is transformed into a peri-inspiratory pattern consisting of pre-inspiratory, and sometimes post-inspiratory, activity, but is silent during inspiration. The cause of this transformation

insensitivity to CO2 stimulation of breathing9. In mice, similar mutations severely disrupt breathing at birth and typically result in early postnatal death10. Second, another study found two sources of respiratory-phased rhythm in neonatal brainstem: the preBötC and a region ventral to nVII that was called the parafacial respiratory group (pFRG)11. Some RTN/pFRG neurons project caudally to brainstem premotoneurons, which drive spinal expiratory motoneurons, suggesting that these neurons are involved in the generation of expiratory movements12. Third, after depressing preBötC neurons with opioids, an unusual breathing pattern, called quantal slowing, can develop in both en bloc preparations and in young rats in vivo13. In this pattern, inspiratory motor activity skips beats, but expiratory motor activity is unaffected13,14. Transecting the brainstem rostral to the RTN/pFRG in juvenile rats does not substantially affect inspiratory and expiratory motor activity, but transection between RTN/pFRG and preBötC completely abolishes active expiratory motor activity with only a modest effect on inspiratory pattern14. These data underlie our hypothesis that in older rodents, and presumably other mammals, the preBötC drives the inspiratory-dominated respiratory pattern, whereas the RTN/pFRG produces a CO2/state-dependent rhythmic drive to expiratory muscles. The preBötC-driven inspiratory breathing pattern dominates at rest, during which the RTN/pFRG may only have tonic activity2,14. During exertion,

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature neuroscience volume 12 | number 8 | august 2009 963

n e w s a n d v i e w s

13. Mellen, N.M., Janczewski, W.A., Bocchiaro, C.M. & Feldman, J.L. Neuron 37, 821–826 (2003).

14. Janczewski, W.A. & Feldman, J.L. J. Physiol. (Lond.) 570, 407–420 (2006).

15. Mulkey, D.K. et al. Nat. Neurosci. 7, 1360–1369 (2004).

10. Dubreuil, V. et al. Proc. Natl. Acad. Sci. USA 105, 1067–1072 (2008).

11. Onimaru, H. & Homma, I. J. Neurosci. 23, 1478–1486 (2003).

12. Janczewski, W.A., Onimaru, H., Homma, I. & Feldman, J.L. J. Physiol. (Lond.) 545, 1017–1026 (2002).

8. Smith, J.C., Morrison, D.E., Ellenberger, H.H., Otto, M.R. & Feldman, J.L. J. Comp. Neurol. 281, 69–96 (1989).

9. Gronli, J.O., Santucci, B.A., Leurgans, S.E., Berry-Kravis, E.M. & Weese-Mayer, D.E. Pediatr. Pulmonol. 43, 77–86 (2008).

The authors are in the Department of Experimental

Psychology, University of Oxford, Oxford, UK.

e-mail: jerome.sallet@psy.ox.ac.uk or matthew.

rushworth@psy.ox.ac.uk

Should I stay or should I go: genetic bases for uncertainty-driven explorationJérôme Sallet & Matthew F S Rushworth

In the face of uncertainty, how do we choose between maintaining our current strategy or trying new strategies? A study shows that a gene controlling prefrontal dopamine function is predictive of uncertainty-driven exploration.

The Clash’s Mick Jones’ lyrics “Should I stay or should I go?...This indecision’s bugging me” eloquently conveyed the experience of being caught on the horns of a dilemma. The song seems particularly apposite when the dilemma occurs in the context of cognitive control and when it concerns uncertainty about that most basic of questions — whether or not to act. His lyrics convey the singer’s uncertainty in identifying the best course of action. They also imply that action is prompted by the hope of desirable consequences, reward, while refraining from action is more naturally linked to avoidance of negative outcomes. Recent papers by Frank and colleagues1, 2, including one in this issue, investigate the neurogenetic underpinnings of these basic behaviors. Frank et al. have shown that action or “go” responses and inaction or “no-go” responses are under the control of different dopamine-related genes expressed primarily in the striatum. In contrast, a gene controlling prefrontal dopamine function is predictive of exploration when the value of alternative strategies is uncertain.

Dopamine is thought to be important in reward guided learning3. It is a key regulator of two neural pathways for action control that run from the basal ganglia to the thalamus and back to the cortex (Fig. 1a) and that are the subject of a computational model4 refined by Frank et al.1. The direct pathway entails two inhibitory connections: striatum to internal globus pallidus to thalamus and then back to cortex. Activating this pathway leads to disinhibition of the thalamus and excitation of the cortex, which is thought to promote a

go response. Dopamine is thought to facilitate the direct pathway via D1 receptors in the striatum. In contrast, inhibition of action is thought to occur via an indirect pathway running from the striatum through the external globus pallidus, subthalamic nucleus and internal globus pallidus that ultimately results in inhibiting thalamic excitation of the cortex. It is thought that this pathway is inhibited by dopamine via D2 striatal receptors. There is strong evidence linking this pathway to action inhibition5, although it is clear that it is not the only pathway to mediate suppression of an unwanted action in favor of a desired action6. By influencing both the direct and indirect pathways,

dopamine might promote actions that were associated with its release and reward delivery and it might promote restraint from action associated with dopamine absence and nondelivery of expected rewards.

Polymorphisms of genes that encode proteins involved in dopamine signaling pathways are associated with differences in learning2. This is presumably because of efficiency differences in the aspects of dopamine signaling in which the proteins are involved. The PPP1RIB gene codes for a protein phosphorylated by D1 stimulation, DARPP-32, and is thought to influence the direct pathway. The DRD2 gene is associ-ated with the distribution of D2 receptors,

Figure 1 The neural substrates of learning from rewards, errors and exploring. (a) This network summarizes the key network nodes suggested by Frank et al.1,4 and those discussed by others9,12,13,15. DL-PFC, dorso-lateral prefrontal cortex; GPe, globus pallidus external segment; GPi, globus pallidus internal segment; LC, locus coeruleus; OFC, orbito-frontal cortex; SNc, substantia nigra pars compacta; SNr, substantia nigra pars reticulata; VTA, ventral tegmental. (b) Subjects were asked to press a button to stop the clock presented on a computer screen. The picture approximates the situation in the DEV condition used in the experiment: the shorter the response time, the higher the magnitude of the points given as reward. (c) In contrast, a response made with the same latency in the IEV condition was associated with low reward levels. Instead, later responses were associated with high reward levels in IEV. (d) Discovering the correct solution is only achieved by exploring both modes of response and Frank et al.1 suggest that this might be done by exploring each option in proportion to the subject’s uncertainty about the likelihood of a positive prediction error when making that choice.

Kim

Cae

sar

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

964 volume 12 | number 8 | august 2009 nature neuroscience

n e w s a n d v i e w s

relation to error feedback, in the anterior cingulate cortex (ACC)10 a region that has already been linked with changes in learning rate8. It should be emphasized that the same ACC region is densely interconnected with the striatum11 and it is likely that they work together when learning rate is determined. Our emerging understanding of the influences of dopamine signaling genes in the context of the detailed physiology and anatomy of the basal ganglia may challenge some of the model’s simplifications.

Less is known about the neural mechanisms underlying exploration because it has been difficult to describe when and why a person is exploring. A previous study12 identified three regions in frontopolar (FPC), medial frontal (probably including ACC) and posterior parietal cortex with exploration, although discussions have often focused on just the first of these. Recent findings suggest that FPC is actually continually monitoring the evidence in favor of alternative actions even when no exploration or switching is taking place13. Changes in behavior are then associated with a change in functional connectivity between FPC and posterior parietal cortex. COMT polymorphisms are associated with reward-related signal differences in adjacent areas in FPC14. Notably, however, the same study also implicated COMT in reward-related signal changes in ventral striatum14.

An important future challenge will be to relate the current findings and model to other ways of thinking about exploration. An influential model has focused on a different neurotransmitter, norepinephrine, and suggested that norepinephrine levels, which have been proposed to be influenced by ACC activity, drive the switch from exploitation to exploration9. It will therefore be important to assess whether COMT polymorphisms are associated with changes in the levels of any other neurotransmitters, particularly other catecholamines such as norepinephrine. Again, the potential role of the ACC, and not just of the striatum, will need further consideration; both recordings from single units and neuroimaging results implicate ACC in the transition between exploration and exploitation15. There remains some uncertainty about the neural mechanisms underlying this fundamental aspect of behavior, but according to the argument advanced by Frank et al.1, it is just such uncertainty that is likely to lead us to explore further.

1. Frank, M.J., Doll, B.B., Oas-Terpstra, J. & Moreno, F. Nat. Neurosci. 12, 1062–1068 (2009).

2. Frank, M.J., Moustafa, A.A., Haughey, H.M., Curran, T. & Hutchison, K.E. Proc. Natl. Acad. Sci. USA 104, 16311–16316 (2007).

learning rate was higher (Fig. 1b,d). A similar pattern of results for both DARPP-32 and DRD2 was previously reported in a study of reward- and error-based learning2.

Learning rates can explain how some subjects quickly learn to exploit fast- or slow-action strategies. But how do naive subjects find the best strategy? Such uncertainty is obviously a feature of many real-world decision scenarios. In this study’s task, uncertainty about the prevailing action-reward contingencies, particularly at the start of each block, make it necessary to explore the merits of each alternative course of action. Frank et al.1 argue that the need for exploration can be quantified in their computational model and that its influence varies as a function of polymorphism in the third gene investigated, COMT (Fig. 1a,d). The model tries to estimate the likelihood of a reward prediction error given that a fast or slow response is made. It does not, however, simply represent a single estimate or belief about reward prediction error likelihood, given the response type. The Bayesian approach means that it represents the distribution or standard deviation of possible estimates of the reward prediction error for each type of response, the go response and the no-go response. Therefore, the index of uncertainty (represented by the width of the distribution) drives exploration. This means that exploration does not occur randomly. Rather, exploration is directed toward a particular behavior in proportion to the subject’s uncertainty that that behavior will be followed by a positive reward prediction error. The effect of uncertainty on the probability of making an exploratory response was scaled by a factor called ε. Subjects with the met allele of the COMT gene had higher ε values than those homozygous for the val allele of the COMT gene; they scaled their exploration rates more sharply as uncertainty about the true reward contingencies increased. The findings complement Frank and colleagues’ previous report of COMT polymorphism influence on flexible and adaptive behavior in another reward-guided learning task2. More generally, the results complement a growing body of work suggesting that representations capturing a person’s uncertainty about a critical aspect of the environment can be used to determine the speed of adaptation to the environment8,9.

Is it really plausible that the three genes studied here have such circumscribed physiological effects? The neural pathways mediating go and no-go responses are delineated in considerable detail4, but there is evidence that the DRD2 polymorphism is also associated with signal changes, in

a key component of the indirect pathway. The COMT gene, coding for the enzyme catechol-o-methyltransferase, which degrades extracellular dopamine, is thought to exert its influence via an alteration of frontal cortical dopamine levels.

In this study, Frank et al.1 asked human subjects, who were genotyped, to perform the simple task of deciding when to stop a clock1. Although the outcomes of their actions were determined by a set rule, subjects were not told about the rule and had to discover the contingency by trial and error. Across the blocks, action outcomes were modulated as a function of response time according to four different rules. In the decreasing expected value (DEV) condition (Fig. 1b), shorter response times yielded greater reward; whereas in the increasing expected value (IEV) condition, longer response times were associated with higher reward (Fig. 1c). Thus, although the DEV condition promotes the fast commission of an action, the IEV condition rewards restraint. Each individual’s go and no-go performances on the DEV and IEV tasks were compared with performance on a baseline control task in which reward value expectation was constant regardless of response time. In other words, the DEV condition probes the direct pathway and the IEV probes the indirect pathway.

A variant of the reinforcement learning model7, in which learning is driven by prediction errors, provided a framework for interpreting the results. The prediction error is the difference between the outcome that an action was expected to yield and the outcome that it actually did yield. Positive prediction errors promote go responses, whereas no-go responses are promoted by negative prediction errors. The effect that each positive prediction error has on the go response tendency is determined by the learning rate parameter. When the learning rate is high, each prediction error has a large effect on the tendency to make an action, and a subject with a high learning rate should therefore learn to respond quickly in the DEV condition. A separate learning rate parameter influences the effect of negative prediction errors on the no-go response tendency. Positive prediction error had a greater effect on go response tendencies in subjects with the T/T DARPP-32 genotype than in those with the G DARP-32 genotype; their go learning rate was higher (Fig. 1a,d). Negative prediction errors had a greater effect on no-go response tendencies in subjects with the DRD2 T/T genotype associated with greater D2 receptor density than it did in subjects carrying the C/C or C/T variants; their no-go

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

nature neuroscience volume 12 | number 8 | august 2009 965

n e w s a n d v i e w s

12. Daw, N.D., O’Doherty, J.P., Dayan, P., Seymour, B. & Dolan, R.J. Nature 441, 876–879 (2006).

13. Boorman, E.D., Behrens, T., Woolrich, M.W. & Rushworth, M. Neuron 62, 733–743 (2009).

14. Dreher, J.C., Kohn, P., Kolachana, B., Weinberger, D.R. & Berman, K.F. Proc. Natl. Acad. Sci. USA 106, 617–622 (2009).

15. Quilodran, R., Rothe, M. & Procyk, E. Neuron 57, 314–325 (2008).

1998).8. Behrens, T.E., Woolrich, M.W., Walton, M.E. &

Rushworth, M.F. Nat. Neurosci. 10, 1214–1221 (2007).

9. Cohen, J.D., McClure, S.M. & Yu, A.J. Phil. Trans. R. Soc. Lond. B 362, 933–942 (2007).

10. Klein, T.A. et al. Science 318, 1642–1645 (2007). 11. Calzavara, R., Mailly, P. & Haber, S.N. Eur. J. Neurosci.

26, 2005–2024 (2007).

3. Schultz, W., Dayan, P. & Montague, P.R. Science 275, 1593–1599 (1997).

4. Frank, M.J., Seeberger, L.C. & O’Reilly, R. Science 306, 1940–1943 (2004).

5. Isoda, M. & Hikosaka, O. J. Neurosci. 28, 7209–7218 (2008).

6. Mars, R.B. et al. J. Neurosci. 29, 6926–6931 (2009). 7. Sutton, R.S. & Barto, A.G. Reinforcement Learning: an

Introduction (MIT Press, Cambridge, Massachusetts,

The authors are in the Department of

Neurosciences, Medical University of South

Carolina, Charleston, South Carolina, USA.

e-mail: smithrac@musc.edu or astong@musc.edu

(a protein encoded by the c-fos gene, and a marker of neural stimulation) following this context-specific cocaine-induced locomotor sensitization. The authors hypothesized that these stimulated neurons might instantiate the context-related conditioning that resulted in locomotor sensitization to cocaine.

To test this hypothesis, the authors set out to selectively inactivate only those neurons that expressed Fos. To do this, they combined two previous developments: c-fos–lacZ transgenic rats8 and the anthracycline prodrug Daun02 (ref. 9). Daun02 is converted by β-galactosidase to daunorubicin, a cytotoxic antitumor agent that kills cells and has also been reported to electrically silence cells. In c-fos–lacZ transgenic rats, neuronal activity stimulates the c-fos promoter in the transgene, driving expression of the bacterial lacZ gene that encodes the protein β-galactosidase. Local microinjection of Daun02 then leads to production of daunorubicin in cells that contain β-galactosidase, leading to presumed long-lasting inactivation or death of Fos-expressing neurons.

In these Daun02 studies, c-fos–lacZ transgenic rats were conditioned with cocaine in a novel environment for 7 d (Fig. 1). These rats received an injection of cocaine or saline in the paired environment 7 d later (to induce both a sensitized response and Fos and β-galactosidase expression) and received intra-accumbal infusions of Daun02 90 min after that (induction day). Cocaine was again injected in the same paired environment 3 d later (test day). Rats that were given intra-accumbal Daun02 on the cocaine induction day had attenuated cocaine-induced locomotion on the subsequent test day. Locomotor levels

drug effects. Both of these phenomena can be influenced by the context in which the drug is administered2, as a result of attachment of motivational importance to previously neutral environmental stimuli following repeated associations with the drug3. For example, context-specific cocaine-induced psychomotor sensitization is a progressive increase in cocaine-induced locomotion in rats following repeated drug exposure in a particular environment4; rats previously exposed to cocaine in one environment do not show locomotor sensitization when cocaine is administered in an alternative environment.

Dopaminergic projections to nucleus accumbens have been implicated in cocaine-induced locomotor sensitization5; however, only a small subset (2–3%) of accumbal neurons shows activation related to context-specific sensitization6. Previous research using histochemistry and behavioral electrophysiology/electrochemistry indicates that this small population of neurons may represent a neuronal ensemble, a group of cells with a distinct set of functional properties that are activated by specific inputs6,7. This sparsely distributed subset of accumbal neurons is proposed to encode the learned association between drugs of abuse and the environment in which they are administered, but this hypothesis has not been directly testable with the available methods.

In the study by Koya et al.1, cocaine was administered to rats for 7 d in a novel environment. After 7 d of withdrawal, cocaine or saline was injected in the same paired environment and a sensitized locomotor response was observed for cocaine only when given in the paired environment, as expected. Examination of brain tissue revealed that a subset of accumbal neurons expressed Fos

In the carnival game ‘Whack-a-mole’, the aim is to knock down all of the moles that pop up from an array of holes in a console in front of you. Although you can’t reach all of the moles, the active ones are vulnerable to attack. In this issue, Koya et al.1 use a similar idea to develop a method to manipulate only the neurons that are critical for a particular behavior. Previous studies trying to prove the importance of a particular group of neurons have used the classical approaches of electrolytic/ neurochemical lesions, as well as local microinjections of pharmacologic agents and electrical/ chemical stimulation. Investigators can now also make use of targeted toxins for neurochemical-specific lesions, selective receptor agonists/ antagonists, transgenic animals and promoter- specific viruses capable of inserting select genes. However, for the most part, these manipulations target all of the neurons with a given profile in a brain region, regardless of their activity or involvement in a particular behavior. Koya et al.1 now present a technique that advances the field one step closer to the goal of activity-dependent manipulation. Using an approach termed the Daun02-inactivation method, the authors were able to selectively inactivate a subset of neurons that are activated during expression of cocaine-induced locomotor sensitization and to demonstrate that specific subsets of neurons in the nucleus accumbens are necessary for this context-dependent learned behavior.

Repeated administration of drugs can cause tolerance or sensitization to specific

Inactivating the activated: identifying functions of specific neural networksRachel J Smith & Gary Aston-Jones

Manipulation of the neurons required for a specific behavior can be difficult, but is required for proof of causality. A clever technique now allows inactivation of only the subset of neurons that have been recently active.

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

966 volume 12 | number 8 | august 2009 nature neuroscience

n e w s a n d v i e w s

Future studies will undoubtedly examine the long-lasting behavioral effects of inactivating a specific subset of accumbal neurons. It is unknown whether rats previously subjected to Daun02 inactivation for blockade of cocaine-induced locomotor sensitization would be able to exhibit future sensitization. It is also important to determine whether this technique can be used in other brain regions. Does Daun02 inactivation require specific neuronal qualities, such as Fos expression exclusively in a subset of neurons related to the function being studied (for example, behavioral association), expression of a substantial level of Fos in response to neural activity or low levels of basal Fos expression? Answers to questions such as these will help validate the usefulness of this potentially powerful technique.

1. Koya, E., et al. Nat. Neurosci. 12, 1069–1073 (2009).

2. Vezina, P. & Leyton, M. Neuropharmacology 56 Suppl 1, 160–168 (2009).

3. Everitt, B.J. & Robbins, T.W. Nat. Neurosci. 8, 1481–1489 (2005).

4. Caprioli, D., Celentano, M., Paolone, G. & Badiani, A. Prog. Neuropsychopharmacol. Biol. Psychiatry 31, 1639–1653 (2007).

5. Kalivas, P.W. & Stewart, J. Brain Res. Brain Res. Rev. 16, 223–244 (1991).

6. Mattson, B.J. et al. Eur. J. Neurosci. 27, 202–212 (2008).

7. Carelli, R.M. & Wightman, R.M. Curr. Opin. Neurobiol. 14, 763–768 (2004).

8. Kasof, G.M. et al. J. Neurosci. 15, 4238–4249 (1995).

9. Farquhar, D. et al. Cancer Chemother. Pharmacol. 50, 65–70 (2002).

are responsible for encoding other types of learned behavior and (if the Daun02 effect on activated neurons is permanent) whether new populations of neurons can be recruited for a given behavior if the first subset is inactivated or lesioned. The elegance of the technique is that the effect is specific for neurons that are activated during a behavior, and, therefore, pinpoints the subset of neurons that is thought to be involved in encoding the learned response. It can be potentially useful in a variety of fields that are interested in relating the activity of small populations of neurons to a larger function such as behavior and can be critical for moving past correlative studies that are based on Fos and electrophysiological activation and on to causal relationships between specific neurons and behaviors.

However, as with most new techniques, there are still several questions to be answered relating to the mechanism of Daun02 inactivation. These studies indicate that the effects of Daun02 last at least 3 d, but does daunorubicin cause reversible inactivation or cell death? In addition, how effective is Daun02 for inactivating Fos-expressing cells? β-galactoside expression was reduced by only 50%, even though cocaine-induced locomotor sensitization appeared to be completely blocked. Therefore, the potency of Daun02 to inactivate or lesion Fos-expressing neurons in this model is unknown, perhaps limiting a tight association between activated neurons and behavior with this method.

were comparable with those observed on the first day of cocaine exposure in the paired environment, indicating a complete loss of sensitization. In these same Daun02-treated rats, cocaine-induced β-galactosidase expression was reduced by 50% on later tissue analysis, consistent with a loss of activated neurons. These behavioral and neural responses were not observed in transgenic rats given Daun02 dorsal to the accumbens or given intra- accumbal vehicle injections, or in wild-type rats given intra-accumbal Daun02. Notably, Daun02 injections following saline administration on the induction day had no effect on subsequent cocaine-induced locomotion. This is consistent with their observations that exposure to the environment alone is not sufficient to induce Fos expression in the accumbal ensemble and that Daun02 does not cause nonspecific inactivation. These results lead the authors to speculate that Daun02 inactivated a subset of accumbal neurons that specifically encoded the learned association between cocaine and the paired environment, resulting in a loss of context-specific cocaine-induced locomotor sensitization.

If further validated with additional behavioral procedures, this new technique provides the opportunity to investigate the causal role of activated neurons (even if sparsely distributed) in specific behaviors. This approach can be used to determine whether neuronal ensembles such as these

Figure 1 Koya et al.1 developed the Daun02-inactivation method to demonstrate that a specific subset of activated neurons in nucleus accumbens is involved in context-specific cocaine-induced locomotor sensitization. c-fos–lacZ transgenic rats were repeatedly exposed to cocaine in a novel environment for 7 d (pairing days, left). Following 7 d of withdrawal, a cocaine injection in the paired environment resulted in a sensitized cocaine-induced locomotor response and neuronal activation in a specific subset (2–3%) of accumbal neurons (induction day, center). In activated neurons, the c-fos promoter transgene drove expression of the lacZ gene, resulting in β-galactosidase production. Local microinjection of the prodrug Daun02 into nucleus accumbens 90 min after the behavioral response resulted in production of daunorubicin (converted by β-galactosidase), which inactivated or lesioned the activated neurons. When cocaine was again administered in the paired environment 3 d later (test day, right), there was a complete loss of cocaine-induced locomotor sensitization and a 50% reduction in β-galactosidase–expressing neurons, indicating that Daun02 caused inactivation of the specific subset of accumbal neurons that is necessary for the learned association between cocaine and the environment.

Kim

Cae

sar

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Resolving single cone inputs tovisual receptive fieldsLawrence C Sincich1, Yuhua Zhang2,3, Pavan Tiruveedhula2,Jonathan C Horton1 & Austin Roorda2

With the current techniques available for mapping receptive

fields, it is impossible to resolve the contribution of single cone

photoreceptors to the response of central visual neurons. Using

adaptive optics to correct for ocular aberrations, we delivered

micron-scale spots of light to the receptive field centers

of neurons in the macaque lateral geniculate nucleus.

Parvocellular LGN neurons mapped in this manner responded

with high reliability to stimulation of single cones.

Human vision is subserved by three types of cone photoreceptorsthat differ in spectral sensitivity, in density with respect to distancefrom the fovea and in relative abundance across individuals1. Becausethe receptive field centers of most neurons in the early visual systemare composed of multiple cones, the question arises of whether thestimulation of just one cone is sufficient to activate retinal ganglioncells and, consequently, for stimuli to be perceived. Moreover, giventhe dispersal of photoreceptor signals through the retinal layers2, itis likely that different cones will vary in their efficacy at drivingdownstream neural activity. We found that thalamic responses couldbe reliably mapped by stimulating individual cones and that theprobability of evoking a spike with each stimulus flash varied, in partbecause of exquisite sensitivity to the position of stimuli relative toeach cone.

Ordinarily, a neuron’s response properties are characterized bypresenting stimuli on a screen while recording action potentials withan extracellular electrode3. This yields a receptive field that is delimitedin time and space, indicating which stimuli are effective at drivingthe cell. Such a method does not provide direct identification ofthe cones feeding the receptive field. This method is also limitedby optical aberration, diffraction, scatter, pre-retinal absorption andeye movement, all of which degrade visual stimuli. These limitationsalter the location and spectral intensity of light impinging on thephotoreceptors, making it difficult to map the cone field precisely,especially near the fovea, where cone spacing is only a few microns.To overcome these difficulties, we used an adaptive optics scanninglaser ophthalmoscope (AOSLO) to visualize and directly stimulate thecones in vivo in the macaque4,5 (details in Supplementary Methods).Experiments were conducted using procedures approved by theUniversity of California San Francisco Institutional Animal Care andUse Committee, in accordance with US National Institutes of Healthguidelines. Neurons were recorded in the lateral geniculate nucleus

(LGN), allowing us to explore receptive field properties that cannot

be examined with traditional mapping techniques.As expected from the retinotopic organization of the LGN, an

orderly sequence of receptive fields was recorded in the left eye duringan electrode penetration (Fig. 1a). Superimposed on the fundusphotograph is a montage of AOSLO cone images (each image is1.21 � 1.21) used to aid navigation in the retina when searching forresponsive cones. Once an LGN neuron was encountered, the first taskwas to find the retinal location where a flashed stimulus generated aresponse. We identified which cones produced the briskest firing bymoving a flickering spot across the retina, which was a straightforwardprocedure because the cones being stimulated could be seen in real time(Supplementary Video 1).

The diameter of the cone field was then determined by plotting theneural response profile obtained by flashing a smaller stimulus pseudo-randomly at locations spaced every 3 mm through the middle of thefield (Fig. 1b). These 3-mm square stimuli subtended 52 arcsec, which isapproximately equal to the diameter of one cone’s inner segment at3.71, the eccentricity of this parvocellular ON-center field. Because thestimuli were constructed from a 30-Hz raster scan, the neural responsewas phase-locked to the frame rate, with the flash duration at eachlocus being about 5 ms (see Supplementary Fig. 1 for details). Tocontrol for retinal motion, we used a video-based algorithm to trackcone positions (Supplementary Video 2 and Supplementary Fig. 2).Single stimulus flashes delivered to single cones reliably led to LGNspikes, with the likelihood of generating any spikes reaching 85% at thepeak response location. Responses diminished sharply at the edge of thecone field, where a shift in the stimulus position of just 3 mm droppedthe LGN spike probability to baseline, suggesting that light delivery wasnot affected seriously by optical blur or intra-retinal scatter. Toward thelimits of the tested area, the probability of firing began to dip below thebaseline level, presumably reflecting inhibition from the receptivefield surround.

The response profile of neuron 1 (Fig. 1b) has variation in the spikeprobability at each stimulus location, particularly in the receptive fieldcenter. There are several possible sources for this variation. First,although parvocellular neurons are dominated by a single cone type,it is unclear whether this property arises from only one cone type beingwired up to the field center or if one cone type simply outnumbers theothers6–8. With the narrowband light that we used for stimulation(centered at a 680-nm wavelength), in vitro measurements of macaquespectral sensitivities would predict that L cones are 14-fold moresensitive than M-cones and 105.8-fold more so than S cones9. However,the relationship between sensitivity differences and firing rate differ-ences in downstream neurons is unclear, as neural responses often sumnonlinearly. Second, the synaptic weighting of the input from eachcone to a ganglion cell, even from cones of the same type, may not beequal. Such variation in synaptic efficacy would be transmitted from

Received 10 March; accepted 20 May; published online 28 June 2009; doi:10.1038/nn.2352

1Beckman Vision Center, University of California, San Francisco, San Francisco, California, USA. 2School of Optometry, University of California, Berkeley, Berkeley,California, USA. 3Department of Ophthalmology, University of Alabama, Birmingham, Alabama, USA. Correspondence should be addressed to L.C.S.(sincichl@vision.ucsf.edu).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 967

BR I E F COMMUNICAT IONS

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ganglion cells to LGN neurons. Finally, a substantial source of variationcould be the sensitivity of cones to the exact position of the stimulus,an effect that is heightened by their light-guide properties10.

To examine the sources of variability in more detail, it was crucial tofirst measure how well stimuli were restricted to single cones. As a directtest of the spatial effect of a stimulus spot, we recorded responses 61from the fovea, where cones are more widely spaced. If intra-retinalscatter were minimal and the adaptive optics–corrected spots were smallenough, stimuli targeted between cones would be unlikely to driveresponses effectively. We flashed rectangularstimuli that had a narrow dimension that

was smaller than the gap between cones inthe adaptive optics images (Fig. 2a). In thisOFF-center cell, stimuli that impinged ona single cone (positions 8 and 9) generatedthe largest response, whereas flanking stimulithat landed between cones (for example, posi-tions 7 and 10) evoked significantly fewerspikes (P o 0.05, one-tailed Fisher’s exacttest). The response did not go to baselineat the flanks because a fraction of light still

landed on the adjacent cone profiles, as shown by the time-averagedenergy distribution of the stimulus. As with the other LGN neuronsthat we recorded, some cones generated significantly different responses,even when stimuli were positioned over them in a similar fashion(for example, responses at positions 8 and 14 in Fig. 2a; P o 0.05,one-tailed Fisher’s exact test), implying that this neuron eitherreceived input from both L and M cones or had mixed coneweights. It is worth noting that cone-to-cone electrical coupling ispresent11, which will reduce the apparent discreteness of the stimuli

Cone field

1

15

b

aPosition-time spike density plot

Spike probability per stimulus flash

0 0.1 0.2 0.3 0.4

5

15

25

35

45

Microns

0 38 spikes per s

Stimulus intensity 50 ms

StimulusArea =Centers =95% energy density =

Cone field Position-time spike density plotSpike probability per stimulus flash

0 0.2 0.4 0.6 0.8 1.0

5

15

25

Microns

134 spikes per s0

Stimulus intensity

StimulusArea =Centers =

Background spike probability

50 ms

Background spike probability

Gray level from cone field image

Significant transitions (P < 0.05)

Figure 2 Parvocellular LGN activity varies with

stimulus position and cone type. (a) Narrow

1.5- � 6-mm stimuli flashed at 15 positions

resulted in OFF responses in neuron 2 that

were highest when stimuli landed on a cone

(for example, position 8) and were significantly

reduced when stimuli fell between cones

(red transitions, P o 0.05, one-tailed Fisher’s

exact test). The green contour shows the region

where 95% of the time-averaged light energywas delivered for this stimulus, taking into

account the point-spread function and motion

remaining after stabilization. The targeted

stimulus area (white) contained 66% of

the delivered light. (b) Stimuli flashed at

15 overlapping positions yielded differential

responses in neuron 3 that depended on the

cone being stimulated. The lowest response

corresponded to one cone, indicated between

the red lines. For all panels, mean probability

per stimulus flash (± s.e.m.) was computed for

the blue outlined areas in the spike density plots.

Micron scale applies to each row of panels.

Cone fieldb

a

Position-time spike rate density plot

Spike probability per stimulus flash

Fovea

1

2 deg400 µm

3

2

0 0.2 0.4 0.6 0.8 1.0

10

30

Microns

50

700 110 spikes per s

Stimulus intensity 50 ms

StimulusArea =Centers =

Background spike probability

Figure 1 Localizing cone fields of LGN neurons.

(a) Left eye fundus photograph with AOSLO

images shown as a montage over the macula

where receptive fields (red circles) were recorded.

The neurons analyzed are numbered for reference.

(b) Stimuli flashed at 19 contiguous locations

across the cone field of neuron 1 (left) led to

an adapting ON response and an inhibitory OFFresponse (middle, each row in the spike density

plot represents the temporal response of one

position in the cone field). Response latency

was B45 ms. The cone field is rotated 901 for

display purposes. Residual light in the optical

path generated background activity at the frame

rate. Activity above the background rate occurred

over a region spanning four cones, indicated by

red lines. Mean spike probability (right, ± s.e.m.)

was measured in the blue outlined area of

the spike density plot. Micron scale applies

to all panels.

968 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

BR I E F COMMUNICAT IONS

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

actually delivered to the retina because coupled cones can funnelactivity through single midget bipolar cells.

Variable response levels for different cones were found in allparvocellular LGN neurons (n ¼ 6) and appeared more discretewhen stimuli were kept small and positioned in small increments.The cone field for neuron 3, a parvocellular ON-center cell (Fig. 2b),was probed with a 3-mm stimulus at twice the spatial resolution usedfor neuron 1 (Fig. 1b). As the stimulus was shifted from cone to cone,the spike activity stepped to several distinct levels. The lowest response,which also showed a temporal delay, coincided with the extent of onecone, but was still above the background firing rate. This difference inspatiotemporal firing pattern suggests that an M cone was at thislocation, whereas the flanking higher responses probably originatedfrom L cones. However, as mentioned earlier, it is also possible thatcolor-tuned neurons are supplied by cones of the same type that differin input strength. An instrument that allows stimulation with multiplewavelengths to facilitate direct assessment of the spectral tuning in eachcone could address this issue.

Because LGN neurons respond reliably when only one of their coneinputs is stimulated, it suggests that single cone activation is sufficientfor perception, even away from the fovea. This is supported by datashowing that frequency-of-seeing curves asymptote well below100% when small adaptive optics–corrected spots are flashed in ahuman subject missing a subpopulation of cones, as stimuli occasion-ally land in ‘holes’ in the photoreceptor mosaic12. In normal subjects,frequency-of-seeing curves are also liable to be affected when suchstimuli fall between cones13. With the ability to probe visual responsesin vivo at their elemental level, the single cone, it will be possible toinvestigate how variable cone weighting leads to the deviations fromcone ‘purity’ seen for color-sensitive neurons in the LGN14 and primaryvisual cortex15.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe thank Q. Yang, D.W. Arathorn and M. Feusner for help withsoftware development, and D. Williams and D. Copenhagen forinsightful manuscript comments. This work was supported by NationalEye Institute grants EY10217 (J.C.H.), EY014375 (A.R.) and EY02162(Beckman Vision Center). Support was also received from the NationalScience Foundation, through grant IIS-0712852 (L.C.S.), the Center forAdaptive Optics cooperative agreement AST-9876783, managed by Universityof California Santa Cruz, and Research to Prevent Blindness. The CaliforniaRegional Primate Research Center is supported by US National Institutes ofHealth Base Grant RR00169.

COMPETING INTERESTS STATEMENTThe authors declare competing financial interests: details accompany the full-textHTML version of the paper at http://www.nature.com/natureneuroscience/.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Hofer, H., Carroll, J., Neitz, J., Neitz, M. & Williams, D.R. J. Neurosci. 25, 9669–9679(2005).

2. Wassle, H. Nat. Rev. Neurosci. 5, 747–757 (2004).3. Hubel, D.H. & Wiesel, T.N. J. Physiol. (Lond.) 160, 106–154 (1962).4. Roorda, A. et al. Opt. Express 10, 405–412 (2002).5. Arathorn, D.W. et al. Opt. Express 15, 13731–13744 (2007).6. Wiesel, T.N. & Hubel, D.H. J. Neurophysiol. 29, 1115–1156 (1966).7. Diller, L. et al. J. Neurosci. 24, 1079–1088 (2004).8. Buzas, P., Blessing, E.M., Szmajda, B.A. & Martin, P.R. J. Neurosci. 26, 11148–11161

(2006).9. Baylor, D.A., Nunn, B.J. & Schnapf, J.L. J. Physiol. (Lond.) 390, 145–160 (1987).10. Roorda, A. & Williams, D.R. J. Vis. 2, 404–412 (2002).11. Hornstein, E.P., Verweij, J. & Schnapf, J.L. Nat. Neurosci. 7, 745–750 (2004).12. Makous, W. et al. Invest. Ophthalmol. Vis. Sci. 47, 4160–4167 (2006).13. Wesner, M.F., Pokorny, J., Shevell, S.K. & Smith, V.C. Vision Res. 31, 1021–1037 (1991).14. Reid, R.C. & Shapley, R.M. J. Neurosci. 22, 6158–6175 (2002).15. Johnson, E.N., Hawken, M.J. & Shapley, R. J. Neurophysiol. 91, 2501–2514 (2004).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 969

BR I E F COMMUNICAT IONS

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Representation of internal modelsof action in the autistic brainCourtney C Haswell1, Jun Izawa1, Lauren R Dowell2,Stewart H Mostofsky2,3 & Reza Shadmehr1

Children with autism spectrum disorder (ASD) have deficits

in motor control, imitation and social function. Does a

dysfunction in the neural basis of representing internal models

of action contribute to these problems? We measured patterns

of generalization as children learned to control a novel tool

and found that the autistic brain built a stronger than normal

association between self-generated motor commands and

proprioceptive feedback; furthermore, the greater the reliance

on proprioception, the greater the child’s impairments in

social function and imitation.

Theory suggests that when the brain learns to perform a movement, itbuilds an association between motor commands and sensory feedback.These internal models allow the brain to predict the sensory conse-quences of self-generated motor commands and to produce motorcommands that maximize expected rewards at a minimum effort1.Children with autism have impairments in motor control2 and imita-tion3. Is there a fundamental difference in how these children buildassociations between their motor commands and sensory feedback?

Generalization is a signature of the activation fields of neurons withwhich the brain forms an internal model4. To quantify the representa-tion of internal models in the autistic brain, we measured patterns ofgeneralization as autistic children learned to control a novel tool. Weasked 14 children with ASD (age, 10.5 ± 1.7 years) and 13 typicallydeveloping children (age, 10.4 ± 1.8 years) to play a game in which theyheld a robotic arm in their hand and reached with it to capture animalsthat had escaped from a zoo (see Supplementary Methods). The robotperturbed the children’s arm movements by producing a force field andthe children learned to control the tool so as to capture the animals. Inthis task, the typically developing brain builds an association betweenself-generated motor commands and the sensory consequences (visualand proprioceptive). The strength of each association can be inferred byhow the brain generalizes the learning from the trained movements tonovel movements. The training took place in the left workspace (target1; Fig. 1a) while a velocity-dependent field pushed their hand perpen-dicular to the direction of motion. We quantified generalization in theright workspace in the intrinsic coordinates of the arm (target 3,identical joint rotations as compared to target 1), and in the extrinsiccoordinates of the task (target 2, identical hand motion as comparedto target 1). Movements to targets 2 and 3 were always made in

‘error-clamp’ trials, in which the robot produced a channel thatartificially eliminated movement errors, but allowed us to measureforce output at the hand.

In the baseline period in which no perturbations were present, bothASD and typically developing groups produced straight reachingmovements (Fig. 1b). On presentation of the field, hand trajectorywas perturbed (Fig. 1b) and the lateral deviations declined withtraining (Fig. 1c), indicating comparable learning rates (F1,979 ¼ 1.8,P ¼ 0.20). In randomly selected trials, an error clamp was presented.We quantified the amount of adaptation/generalization on each error-clamp trial by computing the ratio of the peak lateral force produced bythe child and the ideal force required for compensation on that trial(Fig. 1d). The three targets were presented randomly. For target 1, 6 outof 96 trials were error clamp, whereas all trials were error clamp for theother targets. Therefore, for targets 2 and 3, the children were nevertrained in a force field and never experienced error. This design allowedus to simultaneously assay learning and generalization.

We plotted the adaptation index for each target direction during theerror-clamp trials (Fig. 1d). The average of the first five trials in the testblock was used as a measure of generalization (Fig. 1e). Superficially,learning appeared to be normal in children with ASD; the performancefor target 1 was indistinguishable from that of typically developingchildren on both the last trial of learning (P ¼ 0.18) and the test trials(P ¼ 0.94). However, the generalization patterns were markedlydifferent (F1,25 ¼ 15, P o 0.001, interaction between group and targetdirection). Typically developing children generalized to the rightworkspace both in intrinsic (P o 0.001) and extrinsic (P ¼ 0.003)coordinates, whereas children with ASD generalized in intrinsic coor-dinates (P o 0.0001), but not in extrinsic coordinates (P ¼ 0.30).Furthermore, children with ASD generalized about twice as strong astypically developing children in intrinsic coordinates (Bonferronipost hoc t test, P ¼ 0.0017), reflecting a much stronger than normalassociation between motor commands and proprioceptive feedback5.

In this task, the neurons that participate in representing the internalmodel include cells in the primary motor cortex (M1)6 and thepremotor cortex7. These cells have distinct activation fields and axonalconnectivity. The activation fields of M1 cells tend to be in the intrinsiccoordinates of joints and muscles8 and these cells are stronglyconnected to the adjacent somatosensory cortex. In contrast, theactivation fields of premotor cells tend to be in the extrinsic coordinatesof the task9 and the cells have dense, long-range connections to theposterior parietal cortex. In the brains of typically developing children,reach adaptation produced generalization in both coordinate systems,which is consistent with a representation that engaged both the short-range connections of the primary motor/somatosensory regions andthe long-range connections of the premotor/posterior parietal regions.In the brains of children with ASD, however, there is an overgrowth oflocalized cortical connections10 with increased white matter volume in

Received 17 April; accepted 2 June; published online 5 July 2009; doi:10.1038/nn.2356

1Department of Biomedical Engineering, Johns Hopkins University School of Medicine, Baltimore, Maryland, USA. 2Kennedy Krieger Institute, 3Departments of Neurologyand Psychiatry, Johns Hopkins University School of Medicine, Baltimore, Maryland, USA. Correspondence should be addressed to R.S. (shadmehr@jhu.edu).

970 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

BR I E F COMMUNICAT IONS

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

M1 that predicts motor impairment11. Our results here suggest thatone consequence of this anatomical miswiring in the brains of childrenwith ASD is a representation of internal models that place an unusuallystrong reliance on proprioception.

When we observe another person performing a movement, theinternal models to execute the same movement may also be activatedin our brain12. A strong prediction of this idea is that if the person thatwe are watching makes errors, those errors should help to teach ourown internal model. Indeed, after volunteers observe another personreach while holding a robot that is producing a force field, they performbetter than naive volunteers if they are tested on the same field13. This isconsistent with the hypothesis that observation of an action instantiatesthe same internal models that are required for production of thataction. However, because this instantiation relies on visual cues,

internal models that place a greater than normal reliance on proprio-ception, while discounting visual consequences, might place theobserver at a substantial disadvantage in understanding other people’sactions and imitating their movements. To test our hypothesis, welooked for correlations between how the children represented oursimple reaching task and clinical measures of motor, imitation andsocial function.

We found that the greater the proprioceptive-driven generalizationin our task, the greater the impairments in general motor function,social interaction and imitation/praxis. For example, the AutismDiagnostic Observation Schedule G (ADOS-G) Module 3 ReciprocalSocial Interaction score, a standardized interview/observational assess-ment of social, communicative and stereotyped behaviors in childrenwith ASD, showed that the greater the proprioceptive generalization,the greater the impairment in social function (R ¼ 0.572, P ¼ 0.032;Fig. 2a). The Total T Score from the Social Responsiveness survey, aquestionnaire that is administered to the parents and inquires about thechild’s social interactions in naturalistic settings, was similarly corre-lated with proprioceptive generalization (R ¼ 0.586, P ¼ 0.003;Fig. 2b). We also found that the greater the proprioceptive-drivengeneralization, the greater the impairment in clinical measures of basicmotor skill function (R¼ 0.577, P¼ 0.004), as measured using the totalscore from the Revised Physical and Neurological Examination ofSubtle Signs.

We next asked whether the patterns of generalization were relatedto the ability of the children to imitate movements (SupplementaryTable 1). Imitation was quantified by asking the children to reproducemovements of an examiner14, some of which were meaningful gestures(pretending to use a key in a lock) and others of which werenonmeaningful (tapping of right hand on the left forearm threetimes). The exam was videotaped and analyzed to score each trial ascorrect or incorrect. As expected, children with ASD were impaired inimitation as compared with typically developing children (P o 0.01).However, the greater the internal model’s relative reliance on theintrinsic coordinates of movements (generalization to target 3 minustarget 2), the greater the impairment in imitation (R ¼ �0.57, P ¼0.006; Fig. 2c).

Finally, we asked whether the patterns of generalization were alsorelated to the ability of the children to perform skilled movements in

Right workspace(generalize)

Left workspace(train)

1

2

45 deg

Trial number

100

Baseline

Target 1

Target 2

For

ce in

err

or c

lam

p tr

ials

(%

per

turb

atio

n)

For

ce (

% p

ertu

rbat

ion)

Target 3

1 5 10 15

Trial number (error clamp)

20 25 30

Autism

TD

Learning Test

TDAutism

Late

ral d

evia

tion

(cm

)

0

–1

–2

–3

80

0

80

0

80

0

80 TD Autism

Target 1 Target 2 Target 30

–4

150 200 250 300

1 st

BaselineLast

2 cm2 cm

3

a

c

d

e

b Figure 1 Learning and generalization of an internal model in typically

developing children and children with ASD. (a) Children held the handle of a

robotic arm and played a game in which the objective was to capture animals

that had escaped from a zoo. At the start of the trial, the robot moved the

child’s arm to a starting posture. Next, an animal would appear at the target

location (8 cm). If the child could reach the target in time (0.5 ± 0.05 s), the

animal would be captured and the child was given points that could later be

traded in for a prize. The robot produced a velocity-dependent curl force field.Learning took place in the left posture (1) and generalization was quantified

in the right posture (2, identical hand motion as 1; 3, identical joint motion

as 1). The target sequence was random. This study was approved by the

Johns Hopkins Medicine Institutional Review Board. Informed written

consent was obtained from a parent/guardian and written assent was obtained

from the children. (b) Across subject mean ± s.e.m. hand paths during the

last trial of the baseline block and the first and last trials of the learning

block. Red lines represent children with ASD. (c) Movement error mean ±

s.e.m. for target 1, as quantified by maximum lateral hand deviation; negative

values indicate hand deviations to the left. The filled circles indicate trials in

which the robot perturbed the hand and the unfilled circles indicate error-

clamp trials. TD, typically developing children. (d) In error-clamp trials, the

robot produced a channel from the start position to the target, essentially

eliminating movement errors. We measured the force that the child produced

against the channel walls. (e) The average of force in the first five error-clamp

trials in the test block.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 971

BR I E F COMMUNICAT IONS

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

response to verbal commands and with common tools (that is,praxis)14. Gestures to command were assessed by verbally asking thechild to perform transitive (‘‘Show me how you brush your teeth’’) andintransitive (‘‘Show me how you salute’’) actions. Tool use was assessedby giving the child a tool (for example, a comb) and asking her/him todemonstrate how to use it. Consistent with previous findings14,15,children with ASD were impaired in performance of gestures tocommand (P o 0.01) and tool use (P o 0.01). Furthermore, thegreater the internal model’s relative reliance on the intrinsic coordinatesof movements (generalization to target 3 minus target 2), the greaterthe impairment in the ability to perform gestures to command (R ¼�0.544, P ¼ 0.009) and to use common tools (R ¼�0.551, P ¼ 0.008).

Our findings demonstrate that when children with ASD learn amotor task, the internal models that they form create a stronger thannormal association between the self-generated motor commands andproprioception. This suggests a greater than normal dependence oncortical regions in which movements are represented in intrinsiccoordinates of motion (M1 and somatosensory cortex) and a lessthan normal dependence on regions in which movements are repre-sented in extrinsic coordinates (premotor and posterior parietal). Astronger than normal association between motor commands andproprioceptive feedback may be a consequence of the fact that M1and somatosensory cortex are nearby cortical regions and short-rangecortical connections are overexpressed in children with ASD10.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSThis research was funded by grants from the National Alliance for AutismResearch/Autism Speaks, the US National Institutes of Health (R01 NS037422,R01 NS048527 and K02 NS044850) and the Johns Hopkins University Schoolof Medicine Institute for Clinical and Translational Research, a US NationalInstitutes of Health/National Center for Research Resources Clinical andTransitional Science Award Program (UL1-RR025005).

1590 100

90

80

70

60

50

40

30

70

50

30

Social interaction score children with ASD

Social responsivenessall children

Imitation (percent correct)all children

Generalization in intrinsic coordinates(force, % perturbation)

Generalization in intrinsic coordinates(force, % perturbation)

15 25

Mor

e im

paire

d

Mor

e im

paire

d

Mor

e im

paire

d

35 45 55 65 75 0 10 20 30 40 50 60

ASDTD

70 –20Extrinsic Intrinsic

Relative generalization(force, % perturbation)

0 20 40 60 80

13

11

9

7

5

a b c Figure 2 Motor generalization patterns as a

predictor of social and imitation abilities.

(a) The ADOS-G is a standardized interview

and observational assessment of social,

communicative and stereotyped behaviors used

for diagnosis of autism. The x axis represents the

force produced for T3. (b) The Social Responsive-

ness Scale, a measure of social anxiety/avoidancein naturalistic settings, was scored for most of the

typically developing children (10 of 13) and

children with ASD (13 of 14). (c) Imitation was

measured by asking the child to reproduce a

sequence of 36 actions (performed one at a time),

some of which were meaningful and others of

which were meaningless14. The x axis represents

the force produced during the test of

generalization (T3 minus T2).

AUTHOR CONTRIBUTIONSC.C.H. and J.I. conducted the robot experiments, L.R.D. conducted the social,praxis and imitation experiments and R.S. and S.H.M. wrote the manuscript.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Shadmehr, R. & Krakauer, J.W. A computational neuroanatomy for motor control. Exp.Brain Res. 185, 359–381 (2008).

2. Jansiewicz, E.M. et al. Motor signs distinguish children with high functioning autism andAsperger’s syndrome from controls. J. Autism Dev. Disord. 36, 613–621 (2006).

3. Rogers, S.J., Bennetto, L., McEvoy, R. & Pennington, B.F. Imitation and pantomimein high-functioning adolescents with autism spectrum disorders. Child Dev. 67,2060–2073 (1996).

4. Shadmehr, R. Generalization as a behavioral window to the neural mechanisms oflearning internal models. Hum. Mov. Sci. 23, 543–568 (2004).

5. Hwang, E.J. & Shadmehr, R. Internal models of limb dynamics and the encoding of limbstate. J. Neural Eng. 2, S266–S278 (2005).

6. Li, C.S.R., Padoa-Schioppa, C. & Bizzi, E. Neuronal correlates of motor performance andmotor learning in the primary motor cortex of monkeys adapting to an external force field.Neuron 30, 593–607 (2001).

7. Xiao, J., Padoa-Schioppa, C. & Bizzi, E. Neuronal correlates of movement dynamics inthe dorsal and ventral premotor area in the monkey. Exp. Brain Res. 168, 106–119(2006).

8. Scott, S.H. & Kalaska, J.F. Reaching movements with similar hand paths but differentarm orientation. I. Activity of individual cells in motor cortex. J. Neurophysiol. 77,826–852 (1997).

9. Scott, S.H., Sergio, L.E. & Kalaska, J.F. Reaching movements with similar hand paths,but different arm orientations. II. Activity of individual cells in dorsal premotor cortex andparietal area 5. J. Neurophysiol. 78, 2413–2426 (1997).

10. Herbert, M.R. et al. Localization of white matter volume increase in autism anddevelopmental language disorder. Ann. Neurol. 55, 530–540 (2004).

11. Mostofsky, S.H., Burgess, M.P. & Gidley Larson, J.C. Increased motor cortex whitematter volume predicts motor impairment in autism. Brain 130, 2117–2122 (2007).

12. Oztop, E., Kawato, M. & Arbib, M. Mirror neurons and imitation: a computationallyguided review. Neural Netw. 19, 254–271 (2006).

13. Mattar, A.A. & Gribble, P.L. Motor learning by observing. Neuron 46, 153–160 (2005).14. Mostofsky, S.H. et al. Developmental dyspraxia is not limited to imitation in children

with autism spectrum disorders. J. Int. Neuropsychol. Soc. 12, 314–326 (2006).15. Dziuk, M.A. et al. Dyspraxia in autism: association with motor, social and communicative

deficits. Dev. Med. Child Neurol. 49, 734–739 (2007).

972 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

BR I E F COMMUNICAT IONS

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

The genesis of cerebellar interneurons and theprevention of neural DNA damage require XRCC1

Youngsoo Lee1, Sachin Katyal1, Yang Li1, Sherif F El-Khamisy2,3, Helen R Russell1, Keith W Caldecott2 &Peter J McKinnon1

Defective responses to DNA single strand breaks underlie various neurodegenerative diseases. However, the exact role of

this repair pathway during the development and maintenance of the nervous system is unclear. Using murine neural-specific

inactivation of Xrcc1, a factor that is critical for the repair of DNA single strand breaks, we found a profound neuropathology

that is characterized by the loss of cerebellar interneurons. This cell loss was linked to p53-dependent cell cycle arrest and

occurred as interneuron progenitors commenced differentiation. Loss of Xrcc1 also led to the persistence of DNA strand breaks

throughout the nervous system and abnormal hippocampal function. Collectively, these data detail the in vivo link between

DNA single strand break repair and neurogenesis and highlight the diverse consequences of specific types of genotoxic

stress in the nervous system.

The ability to respond to genotoxic stress is necessary for developmentof the nervous system. Defective responses to DNA damage can resultin a multitude of human syndromes that feature pronounced neuro-pathology1–3. For example, spinocerebellar ataxia with axonal neuro-pathy (SCAN1) and ataxia with oculomotor apraxia (AOA1) aresyndromes that are associated with single-strand break repair (SSBR)defects and feature ataxia linked to cerebellar degeneration and neuro-pathy4–7. SCAN1 results from disruption of tyrosyl-DNA phospho-diesterase 1 (TDP1), an enzyme that is required for the 3¢-endprocessing of certain DNA lesions5. In AOA1, mutations in aprataxin(APTX) cause defects in the removal of 5¢-adenylate–DNA intermedi-ates that can occur during DNA ligation reactions3,4. Although theseobservations indicate the importance of addressing SSBs to avoidneuropathology, the general requirements for SSBR during neurogen-esis are unknown.

XRCC1 is central to SSBR and interacts with multiple DNArepair factors, including APTX and DNA polymerase b, to process aDNA break for ligation involving DNA ligase III (Lig3)2. BecauseXrcc1 null mice die around embryonic day 7 (E7)8, the physio-logical role of this repair factor remains unclear. To address this, wegenerated a conditional Xrcc1 allele to assess the role of SSBR in themouse nervous system. We found that Xrcc1 is required forneurogenesis of cerebellar interneurons and for hippocampalhomeostasis. We also found that Xrcc1 deficiency leads to theprogressive and persistent accumulation of strand breaks inmature neuronal populations. These data are fundamentally impor-tant for understanding the etiology and neuropathology associatedwith neurodegenerative diseases arising from defective DNAdamage responses.

RESULTS

Conditional inactivation of Xrcc1 in the nervous system

Germline deletion of Xrcc1 results in early embryonic lethality, pre-cluding the analysis of Xrcc1 during development8. To circumvent this,we used a Cre/LoxP approach to generate a conditional Xrcc1 allele(Fig. 1a,b and Supplementary Fig. 1). We initially examined the effectsof Xrcc1 inactivation throughout the embryo by crossing mice contain-ing an Xrcc1loxP allele with Meox2-cre mice, which expresses the Crerecombinase in the epiblast layer9. This allowed us to determinewhether early lethality after germline deletion involves the placenta.Xrcc1loxP/loxP; Meox2-cre embryos were malformed by E10 and showeda high index of apoptosis throughout the embryo (SupplementaryFig. 1). We did not observe live Xrcc1loxP/loxP; Meox2-cre embryos afterE12.5, which is consistent with an essential embryonic function forXrcc1 (ref. 8).

To explore the link between defective SSBR repair and neurodegen-eration2,3,10, we used Nes-cre11,12 to inactivate Xrcc1 in the nervoussystem. Efficient deletion of Xrcc1 in neural tissues in Xrcc1loxP/loxP; Nes-cre mice (hereafter referred to as Xrcc1Nes-cre) was confirmed at the DNA,RNA and protein level (Fig. 1c,d and Supplementary Fig. 1). Repair ofDNA strand breaks via Xrcc1-dependent SSBR and base-excision repairis associated with Lig3 activity to reseal the DNA nick. Previous studieshave reported that Lig3 levels are linked to Xrcc1 (ref. 13). To determinewhether this occurs in vivo, we carried out western analysis usingantibody to Lig3. Compared with control tissue extracts, Xrcc1Nes-cre

cortex and cerebellum had markedly reduced Lig3 levels (Fig. 1d),indicating that a loss of Xrcc1 results in reduced Lig3 levels.

Aside from a growth delay, Xrcc1Nes-cre mice did not show any otherdiscernable defects, such as cancer, and the cause of premature death

Received 28 April; accepted 27 May; published online 26 July 2009; doi:10.1038/nn.2375

1Department of Genetics and Tumor Cell Biology, St. Jude Children’s Research Hospital, Memphis, Tennessee, USA. 2Genome Damage and Stability Unit, University of Sussex,Brighton, UK. 3Biochemistry Department, Faculty of Pharmacy, Ain Shams University, Cairo, Egypt. Correspondence should be addressed to P.J.M. (peter.mckinnon@stjude.org).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 973

ART IC L E S

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

remains uncertain. However, adult Xrcc1Nes-cre mice developed pro-nounced neurological dysfunction that was characterized by progres-sive mild ataxia accompanied by episodic spasms (SupplementaryMovie 1). Xrcc1Nes-cre mice generally survived to 4 months of age(Fig. 1e), had smaller brains and developed to only 75% of the size andbody weight of matched control animals (Fig. 1f). Magnetic resonanceimaging analysis of Xrcc1Nes-cre mice revealed a pronounced reductionin cerebellar size compared with the rest of the brain, although theoverall brain to body weight ratio was similar (Fig. 1g).

Xrcc1 loss results in DNA repair deficiency in the brain

To assess the role of Xrcc1 in repairing DNA damage in the nervoussystem, we subjected neurons that were isolated from postnatal day 15(P15) cerebella (Fig. 2a) to DNA repair assays using alkaline cometanalysis (ACA). We found a fourfold increase in global DNA strandbreaks in Xrcc1Nes-cre neurons over control neurons (Fig. 2b). To furthercharacterize neuronal repair, we analyzed cultured postmitotic cere-bellar granule cell neurons (Fig. 2c,d and Supplementary Fig. 2) andquiescent cortical astrocytes (Supplementary Fig. 2) after exposure toexogenous genotoxic agents. In both Xrcc1-deficient neural cell types,we observed that a substantial proportion of DNA strand breaksinduced by ionizing radiation or hydrogen peroxide (H2O2) remainunrepaired after prolonged recovery in genotoxic free conditions orafter methyl methanesulfonate exposure, whereas repair was robust incontrol cells (Fig. 2c and Supplementary Fig. 2). Knockdown ofXRCC1 also sensitizes human neurons to oxidative stress after mena-dione treatment14. Because Xrcc1 is critical for SSBR and ionizingradiation and H2O2 produce 20–2,000-fold more SSBs than DNAdouble-strand breaks (DSBs), our data indicate that Xrcc1 loss inneural tissue results in a pronounced SSBR deficiency.

To further determine the consequence of Xrcc1 loss, we analyzed thebrains of Xrcc1Nes-cre mice for the accumulation of DNA strand breaks.

We used gH2AX as a marker for DNA damage; gH2AX is thephosphorylated version of histone H2AX that forms at DNA breaksto enhance DNA repair efficiency15. Although gH2AX typically marksDSBs16, defective SSBR from Xrcc1 loss can result in DSBs that arisefrom the replication fork or transcriptional machinery colliding withSSBs, or from random damage accumulation leading to adjacentbreaks. During early neural development, gH2AX occurred inproliferating regions of Xrcc1Nes-cre embryos (Supplementary Fig. 3),whereas postnatal Xrcc1Nes-cre mouse brains showed a progressive, age-dependent accumulation of gH2AX foci (Fig. 3 and SupplementaryTable 1). These gH2AX foci overlapped with those of another DNAdamage marker, 53BP1 (Supplementary Fig. 3), confirming that therewas widespread accumulation of DNA damage throughout theXrcc1Nes-cre mouse brain.

Xrcc1 deficiency leads to loss of cerebellar interneurons

Analysis of the Xrcc1Nes-cre brain, although smaller, revealed that mostregions were generally similar to wild-type brains. A survey of theXrcc1Nes-cre central and peripheral nervous system using variousmarkers of differentiation, proliferation and apoptosis failed to findmajor histological consequences of Xrcc1 loss (SupplementaryFig. 4). However, the cerebellum and hippocampus (see below) wereexceptions to this.

Histological analysis of the Xrcc1Nes-cre cerebellum revealed a propor-tional reduction in size (Fig. 4a) and a profound and widespread loss ofinterneurons (Fig. 4b). These interneurons are critical for modulatingthe output of the cerebellum and function to attenuate granule andPurkinje cell electrical activity17,18. Five different types of interneuronsreside in the cerebellum: the stellate and basket cells in the molecularlayer, the Golgi cells in the granule cell layer (GCL), and the lessabundant unipolar brush and Lugaro cells, which are also found in theGCL17,19–24. Cerebellar interneurons are GABAergic, although unipolarbrush cells are glutamatergic. Accordingly, we found that glutamic aciddecarboxylase (the enzyme that catalyses synthesis of GABA) immuno-reactivity in the molecular layer was almost absent in the Xrcc1Nes-cre

cerebellum, indicating that there was a loss of basket and stellate cells(Fig. 4c). Furthermore, we found a substantial reduction of mGluR2, ametabotropic glutamate receptor present in Golgi cells, in the GCL,indicating an absence of this interneuron in the Xrcc1Nes-cre cerebellum(Fig. 4c). In contrast, the numbers of unipolar brush and Lugaro cellsin the mutant cerebellum were comparable with those of controls, asjudged by histology and immunohistochemistry (data not shown).

Xrcc1

Xrcc1Nes-cre

LoxP

a

c e

d

f g

b

LoxP

COOH631

WT

KOExon 3 Exon 4 Exon 11,12

Exon 3 Exon 11,12

LoxP LoxP

1 kb

NH2

Xrcc1 (2.5 kb)

Control

Xrcc1Nes-cre

Xrcc1Nes-cre

Control

Xrcc1Nes-cre

Xrcc1Nes-cre

Control

Genotype Cerebellum/brain Brain/body weight

Control 0.18 ± 0.001

0.13 ± 0.005

0.018 ± 0.005

0.017 ± 0.0002

3 months

Control

Actin (2 kb)

Xrcc1 (80 kDa)

100

80

60S

urvi

val (

%)

40

20

00 2 4 6

Months

Xrcc1+/+ (n = 91)

Lig3 (110 kDa)

Actin (40 kDa)

Cereb

Cortex

Spleen

Thym

us

Cereb

Cortex

Spleen

Thym

us

Cereb

Cortex

Spleen

Thym

us

Cereb

Cortex

Spleen

Thym

us

BRCT BRCT

1

Polβ Ape1 PNK LIG3PARP

Xrcc1loxP/loxP (n = 85)Nestin-cre (n = 106)

Xrcc1Nes-cre (n = 64)

Figure 1 Generation of an Xrcc1 conditional mouse. (a) Schematic diagram

of Xrcc1 indicating BRCT (BRCA1-related C-terminal) domains and the

regions associated with DNA repair protein interactions. (b) The Xrcc1

targeting construct was engineered with loxP sites flanking exons

4–10 (specific details are provided in Supplementary Fig. 1). KO, knockout.

(c) Northern blot analysis showed that control animals produced an Xrcc1

transcript (2.5 kb), whereas the Xrcc1 transcript was absent in the Xrcc1Nes-cre

cerebellum and cerebral cortex. Actin mRNA (2 kb) served as a loadingcontrol. (d) Western blot analysis indicated that Xrcc1 protein (B80 kDa) was

absent from the Xrcc1Nes-cre brain and this was associated with a decrease

in the levels of the Xrcc1-associated protein DNA Ligase 3 (B110 kDa).

In contrast, the spleen and thymus of the Xrcc1Nes-cre mice produced

comparable amounts of Xrcc1 and Ligase 3 to those of controls. Actin was

used as a loading control. (e) Kaplan-Meier curves showing Xrcc1Nes-cre mice

can survive up to 4 months of age. (f) Growth retardation and a smaller brain

size were found in adult Xrcc1Nes-cre mice compared with wild-type control

mice. (g) Magnetic resonance imaging analysis of 3-month-old control and

mutant brains showed that the cerebellum was markedly affected by Xrcc1

loss (white boxes). Three-dimensional volumetric analysis revealed a

B30% reduction in cerebellar size, expressed as relative units.

974 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Other cerebellar cell types, including Purkinje cells (Fig. 4c), granuleneurons and Bergmann glia (data not shown), appeared to be unaf-fected in the Xrcc1-deficient cerebellum. Furthermore, interneuronpopulations in other areas of the Xrcc1-deficient brain showed normalneuronal organization, maturation and morphology (SupplementaryFigs. 5 and 6). These data indicate that Xrcc1 is important during thegenesis of basket, stellate and Golgi interneurons.

Interneuron progenitors are susceptible to Xrcc1 loss

Despite the pronounced effect of Xrcc1 loss on cerebellar neurogenesis,examination of embryonic neural development showed normal indicesof proliferation (Ki67 and PCNA), differentiation (Tuj1) and matura-tion (NeuN) throughout the nervous system, including the embryoniccerebellum (Supplementary Figs. 7–9). However, although neuralprogenitors were similar between controls and mutants, as judged bySox2 immunoreactivity, we found increased apoptosis in the prolif-erative ventricular zone of Xrcc1Nes-cre mice (Supplementary Fig. 9),which could contribute to an overall reduction in the brain size ofXrcc1Nes-cre mice.

The normal development of the Xrcc1Nes-cre embryonic cerebellum(Supplementary Fig. 8) suggested that Xrcc1 becomes particularlyimportant as cerebellar interneuron progenitors begin to differentiatepostnatally. Cerebellar interneurons originate from progenitors thatreside in the white matter of the cerebellum and begin to differentiatearound birth19,20,23,25–28. To define the stage at which interneurons aresusceptible to Xrcc1 loss, we monitored interneuron progenitors using

Pax2 (refs. 20,29). We also used Pax3 as ahindbrain marker and as an adjunct to Pax2for identifying immature cerebellar interneu-rons. Compared with wild-type tissue, therewas a marked reduction of Pax2-positiveinterneuron progenitors in the Xrcc1Nes-cre

cerebellum from P0 onwards (Fig. 5a,b andSupplementary Fig. 10). We observed a simi-lar absence of Pax3-positive cells from thewhite matter of Xrcc1Nes-cre mice, whereas, incontrast, abundant Pax3-positive cells were

present in wild-type tissue (Fig. 5b,c). Pax3 also marks granule neuronprogenitors (GNPs) in the external germinal layer (EGL). Pax3- andPCNA-positive cells were also found in the EGL of both mutants andcontrols, indicating that this region is substantially less affected byXrcc1 loss (Fig. 5b,c). These data indicate that the interneuron defect in

Xrcc1 WT andcondition KO

cerebellaExtract

a

b

d

c

PercollGlia and debris

Culture neurons

Genotoxic insultand ± recovery

P7

Neuronsgradient

Isolate enriched neurons

Titurate

P15

Enriched P15 cerebellar neurons

9

Untreated

R0 R60

DNA damage

876

Control

30IR H2O225

20

15

Com

et ta

il m

omen

tC

omet

tail

mom

ent

10

5

0

Unt R0R60

R180

Unt R0R60

R180

Unt R0R60

R180

Unt R0R60

R180

543

Com

et ta

il m

omen

t

210

Control

Cell number

Cell number

Control Xrcc1Nes-Cre Control

Unt

reat

edR

0R

60R

180

Xrcc1Nes-Cre

P7 granule cell neuron culture

Tuj1 Actin

DAPI Merged

Xrcc1Nes-Cre

Xrcc1Nes-Cre

Figure 2 Xrcc1Nes-cre cerebellar granule neurons

are DNA repair deficient. (a) A single-cell

cerebella suspension was passed through a

35–60% percoll gradient and neurons at the

interface were isolated for experiments. (b) Freshly

isolated granule neurons were used to determine

endogenous DNA strand breaks in the Xrcc1Nes-cre

cerebellum (P15) via ACA. Comet data from 250individual neurons are plotted as the mean comet

tail moment (bar graph) and individual comet tail

moments from these neurons are shown (scatter

plot). (c) Xrcc1Nes-cre cerebellar granule neurons

(P7) were defective in DNA repair after ionizing

radiation (20Gy, IR) or 100 mM H2O2. Various

periods post-damage were examined (0 min (R0),

60 min (R60) and 180 min (R180)). Photo-

micrographs of representative damaged and

repaired neurons following ACA are shown for both

Xrcc1Nes-cre and control neurons. Mean comet tail

moments for each time point are represented

on a bar graph and tail moment data for 30

representative neurons are shown on a scatter

plot. (d) Cultured granule neurons were

immunopositive for Tuj1 (b-tubulin III), a marker

of postmitotic neurons (40� magnification).

CTX

Con

trol

Con

trol

Xrc

c1N

es-c

reX

rcc1

Nes

-cre

DG CA1 Cerebellum

PC

25 µmγH2AX/DAPI

P10

2 months

Figure 3 The Xrcc1-deficient brain accumulates DNA damage in mature

neurons. DNA damage accumulates in different Xrcc1Nes-cre brain regions

progressively with age as measured by gH2AX foci (arrows) at P10 and

2 months of age. CA1, the CA1 region of the hippocampus; CTX, cerebral

cortex; DG, the dentate gyrus of the hippocampus; PC, Purkinje cell layer.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 975

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

the cerebellum of Xrcc1Nes-cre mice is the result of a selective loss ofdifferentiating interneuron progenitors and that this occurs early inpostnatal cerebellar development. However, because defects are notice-able from P0 onwards, Xrcc1 loss probably affects the interneuronprogenitors before P0.

Interneuron progenitors undergo p53-dependent arrest

The loss of interneuron progenitor cells in the white matter ofXrcc1Nes-cre mice could result from either DNA damage–induced cellcycle arrest or apoptosis. We confirmed that interneuron progenitorcells in the Xrcc1Nes-cre mice, but not in matched controls, accumulatedgH2AX foci, indicative of DNA damage (Fig. 6a). We then determinedthe levels of cellular proliferation in various regions of the developingcerebellum using PCNA and BrdU and compared them with the levelsof apoptosis. The cerebellar white matter of Xrcc1Nes-cre mice showed amarked decrease in proliferating cells relative to control tissue (Fig. 6b).In contrast, Xrcc1Nes-cre white matter did not show increased apoptosis

beyond that of the control cerebellum (Fig. 6b), nor was increasedapoptosis seen between P0 through P10 or in the embryonic cerebellum(data not shown). However, apoptosis was high in the EGL ofXrcc1Nes-cre mice as determined by TUNEL and caspase-3 activation(Fig. 6c,d), highlighting the distinct outcomes after DNA damage indifferent cell types. Moreover, we found increased p53 levels andconcomitant increased immunoreactivity of the cyclin-dependentkinase inhibitors p21 and p27 in the cerebellar white matter(Fig. 6e), suggesting that cell cycle arrest, rather than apoptosis,accounts for the loss of interneurons in the Xrcc1Nes-cre cerebellum.

Because of the prominent role of p53-mediated signaling after DNAdamage in the developing nervous system30, we generated Xrcc1Nes-cre;p53�/� mice to determine the contribution of p53 toward interneuronloss. We found a complete rescue of interneurons in the molecular layerof the cerebella of Xrcc1Nes-cre; p53�/� mice, which was reflected by anormal distribution of basket and stellate cells (Fig. 7). Theseinterneurons retained many functional properties, as they were

Control

Con

trol

mG

luR

2G

AD

/DA

Pl

7 weeks

P5 P10

EGL

ML

2 weeks 7 weeks

1.5 mm

Xrcc1Nes-cre

Control

ML

PC

GCL 1.5 mm P5

Calbindin

P5

Xrcc1Nes-cre Control Xrcc1Nes-cre

Xrc

c1N

es-c

re

a

b

c

Pax3

WM EGL

Pax3/PCNA/DAPl

Pax2

Pax3

1510

GCL

GCLEGL

*

**

* *

WM

WM

DN

DN

450

900

Imm

unop

ositi

ve c

ells

per

µm

2

700

500

300

100

350

250

150

50

Xrcc1Nes-creControl

Xrc

c1N

es-c

reC

ontr

ol

c

bPax2

P0

P5

Xrcc1Nes-creControla

Figure 4 Interneurons are missing in the Xrcc1Nes-cre cerebellum.

(a) Comparative view of 7-week-old wild-type and Xrcc1Nes-cre cerebella

after calbindin immunostaining; although smaller, histology of the mutant

cerebellum was similar to that of the wild type. (b) Nissl staining at different

ages indicates the presence of interneurons in the molecular layer (ML) of

the control cerebellum (arrows) while there was little Nissl staining in the

molecular layer of Xrcc1Nes-cre mice. (c) GAD immunostaining identified

GABAergic neurons (upper panels), which were absent in Xrcc1Nes-cre

tissue (white arrow). mGluR2 staining (black arrows, lower panels) identified

Golgi cells in the granule layer of wild-type cerebellum that were largely

absent in the Xrcc1Nes-cre tissue. Calbindin staining of Purkinje cells

identified a similar organization of the Purkinje layer between the wild-type

and mutant cerebella (black arrows). Except where indicated, scale bars

represent 200 mm.

Figure 5 Differentiating interneurons are decreased in the Xrcc1Nes-cre

cerebellum. (a) Staining for Pax2 identified differentiating interneurons in

the white matter (WM) of the developing postnatal control cerebellum (black

arrows). The Xrcc1Nes-cre cerebellum showed a substantial reduction in the

number of Pax2-positive cells at birth (P0), whereas interneurons were

almost absent by P5. In contrast, the number of Pax2-positive cells

in the brainstem remained unaffected (red arrows, middle panels).

(b) Quantification of Pax immunostaining averaged from three different

animals in the cerebellar EGL, GCL, white matter (WM) and deep nucleus

(DN) at P7; * indicates P o 0.05. (c) Pax3-positive cells in the white matter

were also reduced in the Xrcc1Nes-cre cerebellum from birth and were absent

by P7. Many Pax3-positive (red) cells co-stained with PCNA (green) in thecontrol P7 cerebellum, but not in Xrcc1Nes-cre white matter. However,

the distribution of Pax3-positive cells was similar in both wild-type and

Xrcc1Nes-cre EGL. The white arrows indicate cells that are immunopositive

for both Pax3 and PCNA. Scale bar represents 200 mm.

976 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

immunopositive for parvalbumin, superoxidedismutase 2 and glutamic acid decarboxylase(GAD), and synapsed extensively with Pur-kinje dendrites (Fig. 7 and SupplementaryFigs. 6 and 11). However, Golgi cells in theGCL of the cerebellum were only partiallyrescued (Fig. 7), suggesting a particular sus-ceptibility to genotoxic stress, possibly reflect-ing the earlier genesis and unique identity ofthe Golgi interneurons20,23. It is noteworthythat only complete inactivation of p53 rescuedthe interneuron phenotype, and p53 hetero-zygosity had no effect (Fig. 7 and Supplementary Fig. 11), as it doeswhen apoptosis is the end point of DNA damage in the developingnervous system30.

Xrcc1Nes-cre;p53�/� mice also rapidly developed medulloblastomabrain tumors (Supplementary Fig. 12). Some Xrcc1Nes-cre; p53+/�

mice (12.5%, 3 of 24) also developed medulloblastoma after the lossof the wild-type p53 allele (Supplementary Fig. 12 and data notshown). Comparative analysis of Xrcc1-deficient medulloblastomawith other medulloblastoma models from our laboratory showed asimilar cytogenetic and expression array profile (data not shown),indicating that this tumor originated from the GNPs of the EGL11,31,32.This was further confirmed by breeding Xrcc1Nes-cre; p53+/� mice withmice engineered to express green fluorescent protein (GFP) from theMath1 (also known as Atoh1) promoter33 (Math1-gfp), which is highlyexpressed in the GNPs (Supplementary Fig. 12).

Thus, Xrcc1 loss results in two unique outcomes in the cere-bellum in response to genotoxic stress. In the white matter, Xrcc1 isnecessary in interneuron progenitors to suppress DNA damage–induced p53-mediated cell cycle checkpoint activation. In contrast,Xrcc1 loss in the EGL leads to DNA damage–induced p53-mediated apoptosis.

P0 P2

WM

γH2AX/DAPI

γH2AX/Tuj1/DAPI TINEL

ssDNA/DAPI

DN

25 µm

EGL

EGL

EGL

EGL

P5

P7

Xrc

c1N

es-c

reC

ontr

ol

Xrcc1Nes-creControl

PCNA

Active caspase 3

EGL WM

P7

BrdU/DAPI

Imm

unop

ositi

ve c

ells

per

mm

2

P0 w

hite matter

350

250

EGL IGL

p53 p21 p27

WM

* * *

**

*

DN

150

800

600

400

200

75

50

25

50

Xrc

c1N

es-c

reC

ontr

ol

Xrcc1Nes-creControl

Xrc

c1N

es-c

reC

ontr

olX

rcc1

Nes

-cre

Con

trol

a

b

e

d

c

p53–/–

ML

Nis

slP

arva

lbum

inm

Glu

R2

SO

D2/

DA

PI/G

AD

*

GCL

PC ML

Xrcc1Nes-cre p53+/– Xrcc1Nes-cre p53–/–

Figure 6 Cell cycle arrest in response to DNA

damage in the Xrcc1Nes-cre cerebellum. (a) DNA

damage accumulated in the Xrcc1Nes-cre white

matter and deep nucleus (lower panels), but not

in control tissue (upper panels), as shown by

gH2AX formation. Nuclei were stained with DAPI.

(b) Quantification of cells positive for cell cycle

(BrdU and PCNA) or apoptosis (ssDNA) markersin the EGL, IGL, white matter and deep nucleus

of the developing cerebellum at P7. * indicates

P o 0.05. Inset panels show immunohisto-

chemistry of the respective markers used

for quantification of each genotype. (c) Although

gH2AX also formed in the EGL, in contrast to the

white matter, robust apoptosis occurred in the

EGL as judged by TUNEL that coincides with

gH2AX foci (lower panels). (d) Caspase-3

activation occurs in the P7 Xrcc1Nes-cre EGL but

not in the Xrcc1Nes-cre white matter. (e) p53, p21

and p27 markers of cell cycle arrest were found

in the white matter of the Xrcc1Nes-cre cerebellum

during postnatal development (arrows). Except

where indicated, scale bars represent 200 mm.

Figure 7 Interneuron loss in the Xrcc1-null cerebellum is dependent on

p53. Loss of p53 rescued stellate and basket interneurons in the cerebellaof Xrcc1Nes-cre; p53�/� mice, as shown by Nissl and parvalbumin staining

(arrows). In contrast, p53 loss was less effective at restoring Golgi interneurons

in the GCL, as determined by mGluR2 staining (asterisk). GAD immunostaining

revealed restored synaptic density in the Xrcc1Nes-cre; p53�/� molecular layer,

as did superoxide dismutase 2 (SOD2) immunostaining (arrows). Scale bars

represent 200 mm.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 977

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Xrcc1 deficiency affects hippocampal function

Xrcc1Nes-cre adult mice displayed a behavioral phenotype that isconsistent with either seizure or episodic epilepsy (SupplementaryMovie 1). A brain region often associated with this type of behavioraldeficit is the hippocampus34. We found no differences in the generalcellularity between the hippocampi of control and mutant mice(Supplementary Figs. 5 and 7), although increased DNA damage(gH2AX foci) was present in Xrcc1Nes-cre tissue (Figs. 3 and 8a andSupplementary Fig. 3). A marked reduction in the levels of the Xrcc1-binding protein Parp1 was also found in the Xrcc1Nes-cre hippocampus(Fig. 8a), indicating that a loss of Xrcc1 in this tissue may furtherdestabilize components of the SSBR pathway.

Consistent with a seizure phenotype, we found features common totemporal lobe epilepsy35 in the Xrcc1Nes-cre hippocampus, such asabnormally high levels of neuropeptide Y (NPY), and evidence ofgliosis, such as increased levels of glial fibrillary acidic protein andvimentin (Fig. 8b). Xrcc1-deficient hippocampal neurons also showedabnormal activity as revealed by increased c-Fos immunoreactivity36

and a specific reduction of the oligodendrocyte marker 2,3-cyclicnucleotide 3-phosphodiesterase (CNPase). These results indicate thatXrcc1 is required for hippocampal homeostasis.

DISCUSSION

DNA repair is an essential feature of neuraldevelopment and defects in this process canresult in human neurological disease1–3. Forexample, defects in the DNA SSBR end-processing enzymes APTX and TDP1, whichare required for the repair of specific DNAlesions, are associated with the neurodegenera-tive syndromes AOA1 and SCAN1, respectively.Although these syndromes illustrate the needfor maintaining genomic integrity, they donot fully elaborate the requirements for theDNA SSBR pathway during neural develop-ment. Therefore, we inactivated the centralDNA SSBR factor Xrcc1 to assess the role ofthis repair pathway during neural development.

Although inactivation of Xrcc1 throughoutthe nervous system markedly affected thecerebellar interneurons and the hippocampus,there was also a progressive accumulation ofpersistent DNA strand breaks throughout thebrain. A general increase in apoptosis in theneuroepithelia was also observed throughout

development and probably accounts for the smaller brain size found inthe Xrcc1Nes-cre mice. Our findings illustrate the importance of thisDNA repair pathway during the genesis of the nervous system andhighlight cerebellar interneurons as a previously unrecognized target ofdefective DNA repair. This is particularly noteworthy, as the cerebellumis often affected in human neurological disease resulting from DNArepair deficiency37. It will be important to further investigate thesediseases to determine the extent to which interneurons are involved inthe resulting neuropathology.

The cerebellum is a laminar structure containing three main layers:the GCL, the Purkinje cell layer and the molecular layer18,38,39. Stellateand basket interneurons reside in the molecular layer, whereas the Golgicells and other interneurons (the less common unipolar brush cells andLugaro cells) reside in the GCL18,40. Purkinje cells are the primaryoutput from the cerebellum. Parallel fiber projections from granuleneurons make synaptic contact with Purkinje cell dendrites and provideexcitatory contact with the Golgi, stellate and basket cells. In turn, thestellate and basket cells make inhibitory contact with the Purkinje cellsto modulate signaling to the deep cerebellar nuclei17,18,39,40. Therefore,interneurons are integral to cerebellar function and are important formodulating Purkinje and granule cell output in the cerebellum.

Parp

DG

DG CA1

CA2

1 month

NP

Yc-

Fos

GFA

P/D

AP

l/Vim

entin

25 µm

Con

trol

Control Control

Xrc

c1N

es-c

re

Xrcc1Nes-cre Xrcc1Nes-cre

ParpParp/γH2AX/DAPl Parp/γH2AX/DAPl

CN

Pas

e/D

AP

l

a

b

Figure 8 Loss of Xrcc1 affects hippocampal

homeostasis. (a) Levels of the Xrcc1-binding

protein Parp1 are reduced in the mature

Xrcc1Nes-cre hippocampus (arrows) and many cells

accumulate DNA strand breaks in the dentate

gyrus and other hippocampal regions, such as the

CA2 region, indicated by gH2AX foci. (b) NPY

immunostaining was increased in the Xrcc1Nes-cre

hippocampus. Gliosis in the adult Xrcc1Nes-cre

hippocampus was shown by increased GFAP and

vimentin immunoreactivity (arrows). Increased

c-Fos immunoreactivity indicates altered gene

expression in the hippocampus after Xrcc1 loss.

Reduced CNPase immunostaining indicates

that oligodendrocytes were also affected in

the Xrcc1Nes-cre hippocampus. Except where

indicated, scale bars represent 200 mm.

978 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Notably, selective ablation of Golgi cells can result in profoundcerebellar dysfunction21.

Our data demonstrate for the first time, to the best of our knowledge,that Xrcc1 is critical for the formation of cerebellar interneurons. Theseinterneurons form postnatally from migrating progenitor cells in thewhite matter of the cerebellum23,26,28,41. To determine the mechanismof interneuron loss, we surveyed embryonic and postnatal cerebellardevelopment in Xrcc1Nes-cre mice to identify when interneuron pro-genitors were susceptible to Xrcc1 loss. We found that interneuron losscoincided with a failure of progenitor cell proliferation and differen-tiation from around P0 onwards. At this stage, DNA damageaccumulated in Xrcc1Nes-cre progenitors, leading to an increase in thecyclin-dependent kinase inhibitors p21 and p27, reflecting activation ofa p53-dependent cell cycle checkpoint.

It is well established that apoptosis and cell cycle arrest are alternativeoutcomes of DNA damage42, and our data indicate that specific celltypes in the cerebellum respond differently to DNA damage. We foundthat the granule neurons of the EGL undergo DNA damage–inducedapoptosis, whereas interneuron progenitors respond to DNA damagevia a p53-dependent cell cycle arrest. Xrcc1 loss affects this progenitorpopulation postnatally as they commence differentiation. Before thisstage, homologous recombination is available as a backup system forDNA repair5 and may partially compensate for Xrcc1 loss. Otherstudies indicate that interneurons can be affected by perturbations inproliferation control, as cyclin D2–null mice develop a cerebellum thatis essentially devoid of stellate interneurons, suggesting a selectiveregulation of cell fate as these progenitors differentiate43,44.

In addition to the cerebellum, the hippocampus was also affected byXrcc1 inactivation. Although there was no overt cell loss, hippocampalneurons from Xrcc1Nes-cre mice showed DNA damage and associatedneuropathology as indicated by gliosis and elevated c-Fos levels. Thesefeatures resemble the neuropathology of temporal lobe epilepsy, atype of epilepsy in human adults with seizure as a primary clinicalmanifestation35. Similarly, adult Xrcc1Nes-cre mice exhibited sporadicseizure-like activity. Perhaps the high levels of NPY found in theXrcc1Nes-cre hippocampus and its anti-epileptic functions reflect anattempt to counteract uncontrolled neuronal activity and seizure45,46.

Mature neurons throughout the Xrcc1-deficient brain accumulateDNA damage with age. Because persistent gH2AX and 53BP1 fociusually denote DNA DSBs and not SSBs, it is surprising to us that theseare found in the mature Xrcc1Nes-cre mouse brain. This may indicatethat gH2AX foci reflect the rapid in vivo accumulation of DNA SSBs,whereby adjacent breaks are sensed as DNA DSBs. Although the linkbetween Xrcc1 and DNA DSB repair is unclear47, we did not finddefects in DSBR in Xrcc1Nes-cre neural cells, as determined by gH2AXrecovery assays after ionizing radiation (data not shown). Furthermore,if Xrcc1 loss does lead to DNA DSBs, why does non-homologous end-joining not repair the damage12; particularly as recovery of ionizingradiation–induced gH2AX in the Xrcc1Nes-cre brain occurs with normalkinetics (data not shown)? Alternatively, gH2AX foci can reflectstructural changes in chromatin other than the formation of a DSB,which might occur by persistent DNA repair factor associationwith chromatin48. Perhaps then the progressive accumulation ofgH2AX in the Xrcc1Nes-cre brain reflects the slow repair of SSBs, andprolonged accumulation of repair factors at the break that modifychromatin structure.

Collectively, our data show that DNA SSBR is critical for neuraldevelopment and that this pathway is essential for the genesis ofcerebellar interneurons and hippocampal function. These data willcontribute to understanding the role of the DNA damage response inthe prevention of neurological disease.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe thank the Hartwell Center for biotech support, the Transgenic core facilityfor blastocyst injections, the Animal Imaging core for magnetic resonanceimaging analysis and J. Zhao for genotyping. P.J.M. is supported by the USNational Institutes of Health (NS-37956 and CA-21765), the Cancer CenterSupport Grant (P30 CA21765) and the American Lebanese and Syrian AssociatedCharities of St. Jude Children’s Research Hospital. K.W.C. is supported by theMedical Research Council (Grants G0600776 & G0400959) and by the EuropeanUnion Integrated Project on DNA Repair. S.K. is a Neoma Boadway AP EndowedFellow and S.F.E.-K. is supported by the Wellcome Trust (Grant 085284).

AUTHOR CONTRIBUTIONSY. Lee and S.K. performed all experiments characterizing the Xrcc1-deficientmouse and contributed to writing the manuscript. Y. Li and H.R.R. generated themouse model and were responsible for colony production and maintenance withassistance from S.K. and Y. Lee. S.F.E.-K. and K.W.C. designed and performedexperiments and contributed to preparation of the manuscript. P.J.M. was projectleader and produced the final version of the manuscript.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. McKinnon, P.J. DNA repair deficiency and neurological disease. Nat. Rev. Neurosci. 10,100–112 (2009).

2. McKinnon, P.J. & Caldecott, K.W. DNA strand break repair and human genetic disease.Annu. Rev. Genomics Hum. Genet. 8, 37–55 (2007).

3. Rass, U., Ahel, I. & West, S.C. Defective DNA repair and neurodegenerative disease. Cell130, 991–1004 (2007).

4. Ahel, I. et al. The neurodegenerative disease protein aprataxin resolves abortive DNAligation intermediates. Nature 443, 713–716 (2006).

5. El-Khamisy, S.F. et al. Defective DNA single-strand break repair in spinocerebellar ataxiawith axonal neuropathy-1. Nature 434, 108–113 (2005).

6. Moreira, M.C. et al. The gene mutated in ataxia-ocular apraxia 1 encodes the new HIT/Zn-finger protein aprataxin. Nat. Genet. 29, 189–193 (2001).

7. Takashima, H. et al. Mutation of TDP1, encoding a topoisomerase I–dependent DNAdamage repair enzyme, in spinocerebellar ataxia with axonal neuropathy. Nat. Genet. 32,267–272 (2002).

8. Tebbs, R.S. et al. Requirement for the Xrcc1 DNA base excision repair gene during earlymouse development. Dev. Biol. 208, 513–529 (1999).

9. Tallquist, M.D. & Soriano, P. Epiblast-restricted Cre expression in MORE mice: a tool todistinguish embryonic vs. extra-embryonic gene function. Genesis 26, 113–115(2000).

10. Caldecott, K.W. DNA single-strand break repair and spinocerebellar ataxia. Cell 112,7–10 (2003).

11. Frappart, P.O., Lee, Y., Lamont, J. & McKinnon, P.J. BRCA2 is required forneurogenesis and suppression of medulloblastoma. EMBO J. 26, 2732–2742(2007).

12. Shull, E.R. et al. Differential DNA damage signaling accounts for distinctneural apoptotic responses in ATLD and NBS. Genes Dev. 23, 171–180(2009).

13. Caldecott, K.W., McKeown, C.K., Tucker, J.D., Ljungquist, S. & Thompson, L.H. Aninteraction between the mammalian DNA repair protein XRCC1 and DNA ligase III. Mol.Cell. Biol. 14, 68–76 (1994).

14. Kulkarni, A., McNeill, D.R., Gleichmann, M., Mattson, M.P. & Wilson, D.M.,, III. XRCC1protects against the lethality of induced oxidative DNA damage in nondividing neuralcells. Nucleic Acids Res. 36, 5111–5121 (2008).

15. Rogakou, E.P., Pilch, D.R., Orr, A.H., Ivanova, V.S. & Bonner, W.M. DNA double-strandedbreaks induce histone H2AX phosphorylation on serine 139. J. Biol. Chem. 273,5858–5868 (1998).

16. Sedelnikova, O.A., Pilch, D.R., Redon, C. & Bonner, W.M. Histone H2AX in DNA damageand repair. Cancer Biol. Ther. 2, 233–235 (2003).

17. Barmack, N.H. & Yakhnitsa, V. Functions of interneurons in mouse cerebellum.J. Neurosci. 28, 1140–1152 (2008).

18. Sillitoe, R.V. & Joyner, A.L. Morphology, molecular codes and circuitry produce the three-dimensional complexity of the cerebellum. Annu. Rev. Cell Dev. Biol. 23, 549–577(2007).

19. Eccles, J.C. Neurogenesis and morphogenesis in the cerebellar cortex. Proc. Natl. Acad.Sci. USA 66, 294–301 (1970).

20. Weisheit, G. et al. Postnatal development of the murine cerebellar cortex: formation andearly dispersal of basket, stellate and Golgi neurons. Eur. J. Neurosci. 24, 466–478(2006).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 979

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

21. Watanabe, D. et al. Ablation of cerebellar Golgi cells disrupts synaptic integrationinvolving GABA inhibition and NMDA receptor activation in motor coordination. Cell 95,17–27 (1998).

22. Englund, C. et al. Unipolar brush cells of the cerebellum are produced in the rhombic lipand migrate through developing white matter. J. Neurosci. 26, 9184–9195 (2006).

23. Yamanaka, H., Yanagawa, Y. & Obata, K. Development of stellate and basket cells andtheir apoptosis in mouse cerebellar cortex. Neurosci. Res. 50, 13–22 (2004).

24. Weyer, A. & Schilling, K. Developmental and cell type–specific expression of theneuronal marker NeuN in the murine cerebellum. J. Neurosci. Res. 73, 400–409(2003).

25. Rakic, P. Kinetics of proliferation and latency between final cell division and onset ofdifferentiation of cerebellar stellate and basket neurons. J. Comp. Neurol. 147,523–546 (1973).

26. Lee, A. et al. Isolation of neural stem cells from the postnatal cerebellum. Nat. Neurosci.8, 723–729 (2005).

27. Kenney, A.M. & Segal, R.A. Subtracting the Math: prominin-positive cerebellar stemcells in white matter. Nat. Neurosci. 8, 699–701 (2005).

28. Zhang, L. & Goldman, J.E. Generation of cerebellar interneurons from dividing progeni-tors in white matter. Neuron 16, 47–54 (1996).

29. Maricich, S.M. & Herrup, K. Pax-2 expression defines a subset of GABAergic inter-neurons and their precursors in the developing murine cerebellum. J. Neurobiol. 41,281–294 (1999).

30. Lee, Y. & McKinnon, P.J. Responding to DNA double strand breaks in the nervoussystem. Neuroscience 145, 1365–1374 (2007).

31. Lee, Y. & McKinnon, P.J. DNA ligase IV suppresses medulloblastoma formation. CancerRes. 62, 6395–6399 (2002).

32. Orii, K.E., Lee, Y., Kondo, N. & McKinnon, P.J. Selective utilization of nonhomologousend-joining and homologous recombination DNA repair pathways during nervous systemdevelopment. Proc. Natl. Acad. Sci. USA 103, 10017–10022 (2006).

33. Lumpkin, E.A. et al. Math1-driven GFP expression in the developing nervous system oftransgenic mice. Gene Expr. Patterns 3, 389–395 (2003).

34. McNamara, J.O., Huang, Y.Z. & Leonard, A.S. Molecular signaling mechanisms under-lying epileptogenesis. Sci. STKE 2006, re12 (2006).

35. Aronica, E. & Gorter, J.A. Gene expression profile in temporal lobe epilepsy. Neuro-scientist 13, 100–108 (2007).

36. Morgan, J.I., Cohen, D.R., Hempstead, J.L. & Curran, T. Mapping patterns of c-fosexpression in the central nervous system after seizure. Science 237, 192–197 (1987).

37. Frappart, P.O. & McKinnon, P.J. Ataxia-telangiectasia and related diseases. Neuro-molecular Med. 8, 495–511 (2006).

38. Goldowitz, D. & Hamre, K. The cells and molecules that make a cerebellum. TrendsNeurosci. 21, 375–382 (1998).

39. Wang, V.Y. & Zoghbi, H.Y. Genetic regulation of cerebellar development. Nat. Rev.Neurosci. 2, 484–491 (2001).

40. Sotelo, C. Cellular and genetic regulation of the development of the cerebellar system.Prog. Neurobiol. 72, 295–339 (2004).

41. Leto, K., Carletti, B., Williams, I.M., Magrassi, L. & Rossi, F. Different types of cerebellarGABAergic interneurons originate from a common pool of multipotent progenitor cells.J. Neurosci. 26, 11682–11694 (2006).

42. Kastan, M.B. & Bartek, J. Cell-cycle checkpoints and cancer. Nature 432, 316–323(2004).

43. Glickstein, S.B. et al. Selective cortical interneuron and GABA deficits in cyclin D2–nullmice. Development 134, 4083–4093 (2007).

44. Huard, J.M., Forster, C.C., Carter, M.L., Sicinski, P. & Ross, M.E. Cerebellar histogenesisis disturbed in mice lacking cyclin D2. Development 126, 1927–1935 (1999).

45. Howell, O.W. et al. Neuropeptide Y is important for basal and seizure-induced precursorcell proliferation in the hippocampus. Neurobiol. Dis. 26, 174–188 (2007).

46. Baraban, S.C. Neuropeptide Y and epilepsy: recent progress, prospects and controver-sies. Neuropeptides 38, 261–265 (2004).

47. Caldecott, K.W. XRCC1 and DNA strand break repair. DNA Repair (Amst.) 2, 955–969(2003).

48. Soutoglou, E. & Misteli, T. Activation of the cellular DNA damage response in theabsence of DNA lesions. Science 320, 1507–1510 (2008).

980 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSGeneration of Xrcc1 conditional knockout animal model. To generate the

Xrcc1 conditional allele, we obtained a mouse BAC containing the Xrcc1 locus

(Invitrogen) and cloned a 9.8-kb NcoI genomic fragment containing exons

4–14 of the Xrcc1 gene into pBluescript II (Stratagene). An oligonucleotide

containing a loxP site was introduced into a BglII site located between exons

12 and 13. A neomycin-thymidine kinase cassette flanked by loxP sites was then

inserted into a SpeI site located upstream of exon 4. Targeted W9.5 embryonic

stem (ES) cells were selected using G418 followed by removal of a neomycin-

thymidine kinase cassette using pMC-cre and then selection in FIAU (1-(2-

deoxy-2-fluoro-1-D-arabinofuranosyl)-5-iodouracil)-containing media. ES cell

targeting was confirmed by Southern blot analysis of Nhe1- and Xba1-digested

genomic DNA using a 3¢ probe amplified by PCR using 5¢-TTA GCT ACT GCT

TAG CAC CAG GC-3¢ as a forward primer and 3¢-GCA GCT GCC TGT GAG

ACC TGT GC-5¢ as a reverse primer (Supplementary Fig. 1). Southern blot

analysis with this probe yielded a 5.8-kb fragment in wild-type ES cell genomic

DNA, whereas the fragments found in floxed and knockout genomic DNA were

3.4 kb and 4 kb, respectively (Supplementary Fig. 1). Male chimeras generated

after blastocyst injection of targeted ES cells were bred with C57BL/6 females to

identify germline transmission and then used to generate mice carrying the

floxed Xrcc1 gene (Xrcc1loxP/+).

To inactivate the Xrcc1 gene in the nervous system, we bred Xrcc1loxP/+ mice

with Nes-cre mice (B6.Cg-Tg(Nes-cre)1Kln/J, JAX #003771). Xrcc1loxP/loxP; Nes-

cre mice were obtained by crossing male Xrcc1loxP/loxP or Xrcc1loxP/+ mice and

female Xrcc1loxP/+; Nes-cre mice. We maintained Nes-cre maternally to restrict

ectopic Cre recombinase activity in the testis. Meox2-cre (B6.129S4-Meox2tm1

(cre)Sor/J, JAX # 003755) was used to delete Xrcc1 in the epiblast layer to

determine the consequence of gene deletion in the embryo without placental

gene inactivation. Notably, mice with floxed Xrcc1 alleles (Xrcc1loxP/loxP) were

identical to littermate controls (Xrcc1loxP/+ or Xrcc1+/+), indicating that the loxP

sites did not affect Xrcc1 function. Mice carrying only Meox2-cre or Nes-cre

were also indistinguishable from wild-type mice. Math1-GFP mice33 were

kindly provided by J. Johnson (University of Texas Southwestern).

The genotypes of the mice were determined by PCR. For the Xrcc1 gene, the

following three primers were used (Supplementary Fig. 1): PC1 (5¢-TAT GCT

TGC TGT ACA GGG ATT GGG C-3¢), PC2 (5¢-TGG ACC ATG AAA AAG

CTG TGT GC-3¢) and PC4 (5¢-GTC CTC ACT GCT GGA TCC AAG G-3¢).

PCR with a combination of primers PC1 and PC2 discriminated between the

wild-type (245 bp) and the floxed Xrcc1 gene (279 bp), whereas PCR using PC1

and PC4 produced a detectable fragment (274 bp) when the targeted portion of

the Xrcc1 gene was deleted. PCR products were amplified for 35 cycles of 94 1C

for 30 s, 59 1C for 45 s and 72 1C for 45 s. cre was detected using the following

primers and PCR conditions: Cre-3 (5¢-CTG CCA CGA CCA AT GAC AGC-3¢)and Cre-4 (5¢-ACC TGC GGT GCT AAC CAG CG-3¢) for 35 cycles of 94 1C for

30 s, 60 1C for 45 s and 72 1C for 45 s.

The presence of a vaginal plug was designated as E0.5 and the day of birth as

P0. All mice were housed in an Association for Assessment and Accreditation of

Laboratory Animal Care approved facility and were maintained in accordance

with the US National Institutes of Health Guide for the Care and

Use of Laboratory Animals. All procedures involving animals were approved

by the St. Jude Children’s Research Hospital Institutional Animal Care and

Use Committee.

Generation of Xrcc1 antibody and western blot analysis. A rabbit polyclonal

antibody to Xrcc1 was generated using a 15-residue peptide (NH2-AEDSGD-

TEDELRRVA) corresponding to amino acids 481–495 of murine Xrcc1.

Western blot analysis was carried out using 3-week-old neural (cortex and

cerebella) and extra-neural (spleen and thymus) whole tissue extracts from

control (Xrcc1loxP/+; Nes-cre) and conditional knockout mice (Xrcc1loxP/loxP;

Nes-cre). Protein extracts were prepared in lysis buffer (50 mM Tris-HCl,

200 mM NaCl, 0.2% NP-40, 1% Tween-20 (vol/vol), 1 mM NaF, 1 mM sodium

vanadate, 50 mM b-glycerophosphate, 2 mM PMSF, 1� Complete EDTA

(Roche)) and quantified by Bradford assay (Bio-Rad). Proteins (50 mg per lane)

were separated through a 4–12% (wt/vol) Bis-Tris SDS polyacrylamide gel

(Invitrogen) and transferred onto nitrocellulose membrane (Bio-Rad). Blots

were sequentially immunostained with affinity-purified rabbit polyclonal anti-

body to Xrcc1 (1:1,000) followed by horseradish peroxidase–conjugated donkey

antibody to rabbit (1:2,000, GE Healthcare) and detected using ECL Plus

chemiluminescence reagent (GE Healthcare). To assess Ligase 3 protein levels,

we stained immunoblots with mouse antibody to Lig3 (1:200, BD Biosciences)

and processed them as described above. Actin antibody (1:750, Santa Cruz

Biotech) served as a protein-loading control.

Histology. Embryos and brain tissues were removed at indicated ages after

transcardial perfusion with 4% (wt/vol) buffered paraformaldehyde, cryo-

protected in buffered 25% sucrose (wt/vol) solution and sectioned at 10 mm

using an HM500M cryostat (Microm). Alternatively, tissues were fixed in 10%

buffered formalin (vol/vol), embedded in paraffin and sectioned at 7 mm using

an HM325 microtome (Microm). Nissl staining was carried out with 1%

(wt/vol) thionin. Hematoxylin and eosin staining was done according to

standard procedures.

Immunohistochemical and immunocytochemical analysis of tissues and

cells, respectively, were performed using the antibodies listed below. For

colorimetric visualization of positive signals, sections were incubated with

antibodies overnight after quenching endogenous peroxidase using 0.6%

(vol/vol) H2O2 in methanol. After washing slides with phosphate-buffered

saline (PBS) several times, the tissues were treated with biotinylated secondary

antibody and avidin-biotin complex (Vectastain Elite kit, Vector Labs). Immu-

noreactivity was visualized with the VIP substrate kit (Vector Labs) according

to the manufacturer’s protocol. Sections were counterstained with 0.1%

(wt/vol) methyl green, dehydrated, and mounted in DPX (Fluka). For fluor-

escent signals of immunoreactivity, FITC- or Cy3-conjugated secondary anti-

bodies (Jackson Immunologicals) were used and counterstained with 4¢6-

diamidino-2-phenylindole (DAPI) or propidium iodide (Vector Laboratories).

For BrdU incorporation studies, P7 brains were removed 2 h after BrdU

injection (60 mg per g of body weight, intraperitoneal injection). Apoptosis was

measured by immunoreactivity to ssDNA antibody (IBL International GmbH)

and TUNEL assays using Apoptag (Chemicon) according to the manufacturer’s

protocol. All stained slides were examined and imaged using an Axio Imager A1

microscope (Zeiss) and captured and processed using Adobe Photoshop.

Antibodies. For immunohistochemistry and immunocytochemistry, we used

antibodies to b-tubulin III (clone Tuj1, mouse, 1:1,000, BabCo), calbindin-

D28K (mouse, 1:2,000, Sigma), calretinin (mouse, 1:1,000, Chemicon), GABA

receptor a6 (rabbit, 1:500, Chemicon), GFAP (mouse, 1:400, Sigma; rabbit,

1:200, Abcam), Pax2 (rabbit, 1:500, Zymed), Pax3 (rabbit, 1:500, a generous gift

from G. Grosveld, St. Jude Children’s Hospital), superoxide dismutase 2 (rabbit,

1:500, Stressgen), Somatostatin (rabbit, 1:2,000, ImmunoStar), synaptophysin

(mouse, 1:1,000, Chemicon), BrdU (rat, 1:200, Abcam), gH2AX (rabbit, 1:500,

Abcam), 53BP1 (rabbit, 1:500, Bethyl Laboratories), Ki67 (rabbit, 1:2,500,

Vector Laboratories), parvalbumin (guinea pig, 1:1,000, Chemicon), p21

(mouse, 1:100, Santa Cruz Biotechnology), p53 (CM5) (rabbit, 1:1,000, Vector

Laboratories), p27 (rabbit, 1:100, Santa Cruz Biotechnology), Parp1 (rabbit,

1:25, Abcam; mouse, 1:50, Novus) and PCNA (mouse, 1:500, Santa Cruz

Biotechnology). The antibodies listed above were used with a citrate buffer–

based antigen retrieval treatment.

The following antibodies were used without antigen retrieval: c-Fos (K-25,

rabbit, 1:1,000, Santa Cruz Biotechnology), CNPase (mouse, 1:500, Sigma),

GAD67 (rabbit, 1:500, Chemicon), mGluR2 (rabbit, 1:500, Upstate), NeuN

(mouse, 1:500, Chemicon), NPY (rabbit, 1:5,000, Novus), Vimentin (mouse,

clone RV202, 1:200, Abcam), ssDNA (rabbit, 1:300, IBL) and actin (goat,

1:1,000, Santa Cruz Biotechnology).

Isolation of primary mouse granule cell neurons and astrocytes. Cerebellar

granule cells were purified from P7 brains49; whole cerebella were treated with

low activity (1:250) trypsin (Gibco) DNAse I (Worthington) and triturated

with Pasteur pipettes into a single-cell suspension. The cell suspension was

applied to a Percoll (GE Healthcare) gradient (35%/60%) and separated by

centrifugation (1,200 g). Enriched granule cell neurons were grown in neural

basal medium (Gibco) supplemented with 1� glutamax, 100 U ml–1 penicillin,

100 mg ml–1 streptomycin, 2% (wt/vol) D-glucose (Sigma), 1� SPITE (Sigma),

1� Oleic acid albumin/linoleic acid (Sigma) and 16 mg ml–1 N-acetyl cysteine

(Sigma) on poly-D-lysine/matrigel-coated (Gibco and Becton Dickson, respec-

tively) glass-bottom multi-chamber slides (Falcon) at a density of 3 � 105 cells

doi:10.1038/nn.2375 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

per well. For in vivo comet assays, cells from P15 cerebella were subjected to

ACA directly on Percoll-mediated enrichment. Immunofluorescent staining of

granule cell neurons was performed as previously described50, using antibody

to Tuj1 (1:500, b-tubulin III) and phalloidin (1:500, Molecular Probes).

Primary murine astrocytes were prepared from P3 brains as described

previously50. Cortices were dissociated by passage through a 5-ml pipette

and cells were resuspended in Dulbecco’s modified Eagle’s medium and Ham’s

nutrient mixture F-12 (1:1 DMEM/F12, Gibco-BRL) supplemented with 10%

fetal bovine serum (vol/vol), 1� glutamax, 100 U ml–1 penicillin, 100 mg ml–1

streptomycin and 20 ng ml–1 epidermal growth factor (Sigma). Primary

astrocytes were established in Primeria T-25 tissue culture flasks (Falcon) at

37 1C in a humidified, oxygen-regulated (9%) incubator. Culture medium was

changed after 3 d and astrocyte monolayers reached confluence 3 d later.

ACA. Primary granule cell neurons or quiescent primary murine astrocytes

(3 � 105 cells per ml per sample) were treated with either H2O2 (100 mM,

neurons; 150 mM, astrocytes) for 10 min on ice, methyl methanesulfonate for

10 min at various concentrations or g-irradiation (20Gy, Cs137). Cells were then

incubated for various times in drug-free medium at 37 1C. Cells were

suspended in pre-chilled PBS, mixed with an equal volume of 1.2% low-

melting point agarose (Invitrogen) maintained at 42 1C, immediately layered

onto frosted glass slides (Fisher) pre-coated with 0.6% agarose and maintained

in the dark at 4 1C for all further steps. Slides were immersed in pre-chilled lysis

buffer (2.5 M NaCl, 10 mM Tris-HCl, 100 mM EDTA (pH 8.0), 1% (vol/vol)

Triton X-100, 3% (vol/vol) DMSO, pH10) for 1 h, washed with pre-chilled

distilled water (twice for 10 min each) and placed for into pre-chilled alkaline

electrophoresis buffer (50 mM NaOH, 1 mM EDTA, 1% DMSO) for 45 min.

Electrophoresis was carried out at 95 mA for 25 min, followed by neutralization

in 0.4 M Tris-HCl (pH 7.0). Comets were stained with SYBR Green (1:10,000 in

PBS, Sigma) for 10 min. A minimum of 50 comet tail moments were measured

using the Comet Assay IV system (Perceptive Instruments) coupled to an

Axioskop2 plus microscope (Zeiss) at 200� magnification.

Magnetic resonance imaging and statistical analysis. Adult mutant and

control mice (age and sex matched) were subject to magnetic resonance

imaging analysis. All subjects were scanned on a 7 Tesla Bruker Clinscan

animal magnetic resonance imaging scanner (Bruker BioSpin MRI GmbH)

using the Bruker 12S gradient BGA12S and the 4 channel mouse surface coil

(for T2 weighted images in two directions; in the sagittal scan, matrix ¼ 320 �320, field of view ¼ 25 � 25 mm, slices ¼ 17 contiguous, thickness ¼ 0.7 mm,

echo time ¼ 39 ms, repetition time ¼ 2,579 ms; in the axial scan, matrix ¼320 � 320, field of view ¼ 25 � 25 mm, slices ¼ 16 contiguous, thickness ¼0.5 mm, echo time ¼ 42 ms, repetition time ¼ 2620 ms). The volumetric

measurement of the brain was performed using ImageJ (v1.41n). For quanti-

fication analysis of cell cycle and cell death, immunopositive cells in the EGL,

IGL, white matter and deep nucleus within 1 mm2 of the lingual and lobulus

centralis in the cerebellum at P7 were counted in at least six representative

sections and mean values were calculated and statistically analyzed. These data

were collected from three different mice per each group. All statistical analyses

were performed using Prism (v4.0, Graphpad) and differences were considered

significant at P o 0.05.

49. Hatten, M.E. Neuronal regulation of astroglial morphology and proliferation in vitro.J. Cell Biol. 100, 384–396 (1985).

50. Katyal, S. et al. TDP1 facilitates chromosomal single-strand break repair in neurons andis neuroprotective in vivo. EMBO J. 26, 4720–4731 (2007).

NATURE NEUROSCIENCE doi:10.1038/nn.2375

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

A trophic role for Wnt-Ror kinase signaling duringdevelopmental pruning in Caenorhabditis elegans

Yu Hayashi1,4, Takaaki Hirotsu2,5, Ryo Iwata3,5, Eriko Kage-Nakadai1,4, Hirofumi Kunitomo3, Takeshi Ishihara2,Yuichi Iino3 & Takeo Kubo1

The molecular mechanism by which neurites are selected for elimination or incorporation into the mature circuit during

developmental pruning remains unknown. The trophic theory postulates that local cues provided by target or surrounding cells

act to inhibit neurite elimination. However, no widely conserved factor mediating this trophic function has been identified.

We found that the developmental survival of specific neurites in Caenorhabditis elegans largely depends on detection

of the morphogen Wnt by the Ror kinase CAM-1, which is a transmembrane tyrosine kinase with a Frizzled domain. Mutations

in Wnt genes or in cam-1 enhanced neurite elimination, whereas overexpression of cam-1 inhibited neurite elimination in a

Wnt-dependent manner. Moreover, mutations in these genes counteracted the effect of a mutation in mbr-1, which encodes

a transcription factor that promotes neurite elimination. These results reveal the trophic role of an atypical Wnt pathway

and reinforce the classical model of developmental pruning.

In the developing nervous system, many of the neurites that are initiallyformed are later eliminated, a phenomenon that is known as develop-mental pruning. Developmental pruning proceeds in a precisely regu-lated manner to establish elaborate connectional patterns, for example,during the transition from one target–multiple axons to one target–oneaxon in the developing cerebellum and neuromuscular junctions1.During developmental pruning, the manner in which specific neuritesare maintained while most others are eliminated remains largelyunknown. Findings that axons with the strongest neural connectionto target cells are exempted from elimination during competitiveaxon pruning2 and that the dysfunction of target or surroundingcells leads to axon regression3,4 have led to the proposal of the trophictheory of neurite survival. According to this theory, whether a neuritesurvives to be incorporated into the mature circuit depends on whetherthe neurite can capture trophic cues that are provided locally by thetarget and/or surrounding cells to inhibit neurite elimination2,5. How-ever, a lack of effective genetic models has hampered the identificationof such cues.

The simple nervous system and the genetic accessibility of the nema-tode Caenorhabditis elegans offer substantial advantages for examiningthe molecular mechanisms of axon guidance, synaptogenesis and otherdevelopmental events. We previously reported that developmental neu-rite pruning occurs in this organism in an interneuron subclass termedAIM6 (Fig. 1a,b). Developmental pruning in AIM is regulated by thehighly conserved transcription factor MBR-1, which promotes theelimination of neurites6. The stochastic nature of this developmentalevent, however, implies that opposing unknown signals that promote

the survival of neurites also have a role in determining whether neuritesare developmentally eliminated or maintained.

To identify factors other than MBR-1 that are involved in develop-mental pruning, we examined an array of genes expressed in AIM andfound that the rate of neurite elimination was altered in loss-of-function mutants of cam-1, which encodes the sole C. elegans Rorkinase7,8. Ror kinases comprise a family of highly conserved trans-membrane tyrosine kinases that are expressed in the developingnervous system of both invertebrates and vertebrates9. By analyzingthe mechanisms of CAM-1–mediated regulation of developmentalpruning, we found a role for the well-known morphogen Wnt inpromoting neurite survival. Our results indicate that a noncanonicalWnt pathway that is not widely recognized at present is important as atrophic signal during developmental pruning.

RESULTS

The Ror kinase CAM-1 inhibits neurite elimination in AIM

Each of the left and right AIM neurons of a newly hatched larvatypically project two neurites: one anteriorly to the nerve ring and onemedially to the other AIM neuron. During subsequent periods of thefirst larval stage (L1), the medially projected neurite is eliminated ina large proportion of individuals6 (Fig. 1a–c). In cam-1 mutant worms,neurite elimination was enhanced (that is, the percentage of adultswith AIM interconnections was decreased; Fig. 1c). In contrast, thecam-1 mutation did not affect the percentage of individuals withAIM interconnections in early L1 (Fig. 1c), indicating that cam-1 isnot involved in the initial formation of the AIM interconnections.

Received 3 April; accepted 5 May; published online 28 June 2009; doi:10.1038/nn.2347

1Department of Biological Sciences, Graduate School of Science, The University of Tokyo, Bunkyo-ku, Tokyo, Japan. 2Department of Biology, Graduate School of Science,Kyushu University, Fukuoka, Japan. 3Deparment of Biophysics and Biochemistry, Graduate School of Science, The University of Tokyo, Bunkyo-ku, Tokyo, Japan. 4Presentaddress: Laboratory for Behavioral Genetics, RIKEN Brain Science Institute, Saitama, Japan (Y.H.), and Department of Physiology, Tokyo Women’s Medical University Schoolof Medicine, Shinjuku-ku, Tokyo, Japan (E.K.-N.). 5These authors contributed equally to this work. Correspondence should be addressed to Y.H. (yuhayashi@brain.riken.jp).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 981

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

The overall morphology of AIM neurons also appeared to be normal(Supplementary Fig. 1). The two cam-1 mutant alleles that we used,cam-1(ak37) and cam-1(gm122), are thought to be null alleles, as a largeproportion of the extracellular domain and the whole transmembraneand intracellular domains of the encoded protein are deleted in bothalleles8,10 (Fig. 2a). Therefore, these results indicate that cam-1 is aninhibitor of neurite pruning.

We previously found that the transcription factor MBR-1 cell-autonomously promotes neurite elimination in AIM6, a functionthat would oppose that of CAM-1. Therefore, we next examined thegenetic interaction between cam-1 and mbr-1. The mbr-1 mutant allelembr-1(qa5901) lacks much of the coding region and the upstreamsequence and therefore probably represents a strong loss-of-functionallele6. In this mbr-1 mutant, the rate of neurite elimination was largelyreduced (Fig. 1c). When we compared the percentage of individualswith AIM interconnections between mbr-1 single and mbr-1; cam-1double mutants, we found that the cam-1 mutation suppressed thereduced rate of neurite elimination in the mbr-1 mutant (Fig. 1c).Conversely, a comparison between cam-1 single and mbr-1; cam-1double mutants revealed that the mbr-1 mutation also suppressed thecam-1 mutant phenotype (Fig. 1c). These results suggest that the twogenes function at least partially via parallel mechanisms. Consistentwith this view, expression of mbr-1 appeared to not be affected by

the cam-1 mutation (Fig. 1b and Supplementary Fig. 1) and viceversa (Supplementary Fig. 2), as assessed by expression analysesof mbr-1 or cam-1 Pmbr-1Hgfp and Pcam-1Hgfp (green fluorescentprotein) fusion DNA constructs.

We confirmed the inhibitory role of cam-1 during developmentalpruning by overexpressing a CAM-1–GFP fusion protein in AIMneurons of wild-type and mbr-1; cam-1 double mutant strains usingthe mbr-1 promoter. GFP fusion does not interfere with CAM-1function10,11. As expected, the percentage of adults with AIM inter-connections was considerably increased in both strains (Fig. 2b).A similar effect was also observed when we overexpressed CAM-1–GFP using the zig-3 promoter (Fig. 2b), which overlaps with thembr-1 promoter only in AIM neurons12.

In vivo analyses of Ror kinases in vertebrates and invertebrateshave revealed their common role in neuronal migration8,13. CAM-1in C. elegans is also involved in synaptic transmission10. The entireintracellular region of Ror kinases, including the tyrosine kinasedomain (Fig. 2a), is dispensable for these functions10,11,13. We testedwhether this was also the case for the function of CAM-1 in develop-mental pruning. Overexpression of a mutant form of CAM-1 lackingthe intracellular domain, CAM-1(Dintra), did not inhibit, but ratherenhanced, neurite elimination (Fig. 2b), indicating that the functionof CAM-1 in developmental pruning requires the intracellular domain.

Early L1 Early L1

Early L1

Wild typecam-1(ak37)cam-1(gm122)

mbr-1(qa5901)mbr-1(qa5901); cam-1(ak37)mbr-1(qa5901); cam-1(gm122)

Late L1 L2 Adult

Left AIM

Adult Adult

AIM

inte

rcon

nect

ed in

divi

dual

s

80

0.9990.760

0.4360.7490.000

0.004

0.0070.000 0.000

0.000

0.000

0.000

0.000

0.018

0.001

0.000

0.002

0.093

0.061

(%)

60

40

20

0

a b cRight AIM

Figure 1 The Ror kinase CAM-1 and the transcription factor MBR-1 have opposite roles in

developmental neurite pruning. (a,b) Schematic (a) and images (b) of developmental neurite

pruning in C. elegans AIM neurons observed from the dorsal aspect. The somas of AIM and

the neurites that are to be eliminated are each emphasized by dotted circles and arrowheads

in b. Scale bars represent 5 mm. (c) Percentage of individuals with AIM interconnections during

early L1 (n ¼ 41–97), late L1 (n ¼ 43–64), L2 (n ¼ 44–79) and the adult stage (n ¼ 102–159) in wild-type, cam-1 and mbr-1 mutants, and mbr-1;cam-1

double mutants. The numbers on the graph indicate P values in the population rate test (see Supplementary Table 1 for numerical data).

Ig Frizzled Kri mRFP

CAM-1–GFP

TM Kin

Deleted in ak37

gm122 (Q253→→Stop)

S/T(%)

80

0.000

0.000

0.000

0.000

0.046

0.128

60

40

20

P mbr

-1 ::ca

m-1

::gfp

P mbr

-1 ::ca

m-1

(∆In

tra)::

gfp

P mbr

-1 ::ca

m-1

(∆Fr

izzled

)::gf

p

P mbr

-1 ::ca

m-1

(kina

se d

ead)

::gfp

P mbr

-1 ::ca

m-1

::gfp

P zig-3

::ca

m-1

::gfp– –

0

AIM

inte

rcon

nect

ed in

divi

dual

s

Host: Wild type mbr-1(qa5901);cam-1(ak37)

a b c

Figure 2 CAM-1 functions cell-autonomously

to inhibit neurite elimination. (a) CAM-1 protein

structure. Frizzled, Frizzled domain (also known

as cysteine-rich domain); Ig, immunoglobulin

domain; Kin, kinase domain; Kri, kringle

domain; S/T, serine/threonine-rich domain;

TM, transmembrane domain. The amino terminus,

which is exposed to the extracellular space, is

to the left. The resultant products of the loss-of-

function mutant alleles that we used are also

shown. (b) Percentage of adult individuals

with AIM interconnections in transgenic worms

expressing full length or partial length cam-1

complementary DNA (cDNA) fused to gfp in AIM in a wild-type background or an mbr-1; cam-1 double mutant background (n ¼ 50–132). As indicated,

all transgenes were expressed using the mbr-1 promoter except for Pzig-3Hcam-1–gfp. Numbers in the graph indicate P values in the population rate test

(see Supplementary Table 1 for numerical data). (c) Localization of CAM-1–GFP in AIM during early L1. The morphology of AIM is visualized by mRFP.CAM-1–GFP–derived signals at the proximal regions of neurites and the nerve ring are indicated by arrowheads and brackets, respectively. Scale bars

represent 5 mm.

982 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

We further tested whether the kinase activity is required by over-expressing a mutant form of CAM-1 in which two conserved lysines inthe ATP-binding motif of the tyrosine kinase domain were replaced byarginine, which is predicted to eliminate kinase activity8. This mutantform of CAM-1 (CAM-1 kinase dead) had an effect that was compar-able to that of the wild-type form (Fig. 2b), suggesting that the kinaseactivity is dispensable for CAM-1 function in developmental pruning.Besides the tyrosine kinase domain, the serine/threonine-rich domainis the only highly conserved intracellular region and thus might beimportant. These results suggest that CAM-1 exerts its effect on develop-mental pruning via mechanisms that are distinct from its function inneuronal migration and synaptic transmission.

We next determined the intracellular area in which CAM-1 func-tions by coexpressing CAM-1–GFP with monomeric red fluorescentprotein (mRFP). Although punctate signals in the nerve ring wereconsistent with previous reports of CAM-1 localization at synapticsites10, we detected CAM-1–GFP localization mainly at the proximalregions in neurites that were destined to be eliminated (Fig. 2c).CAM-1–GFP–derived signals were weak or not detected in otherregions of the neurites. These results suggest that CAM-1 mightfunction at particular proximal segments of neurites to interfere withtheir developmental regression.

Inhibition of neurite elimination by CAM-1 requires Wnt

We next focused on the roles of the extracellular region of CAM-1,which is highly conserved in Ror kinases across phylogeny (Fig. 2a).The Frizzled domain of Ror kinases, also known as the cysteine-richdomain, is known to interact directly with the morphogen Wnt14–16.Moreover, recent genetic analyses in Xenopus and C. elegans haverevealed that Ror kinases can actually transmit Wnt signals in vivoduring morphogenesis, thus comprising a previously unknown form of

Wnt signaling17,18. We therefore investigated the possibility that Wntfunctions as an extracellular cue that regulates neurite pruning viaCAM-1. In the C. elegans genome, five genes (cwn-1, cwn-2, egl-20,lin-44 and mom-2) encode members of the Wnt family and functionredundantly in several developmental events19. We initially examinedAIM neurons in several double mutants of Wnt genes, with theexception of mom-2, which is embryonic lethal19; neurite eliminationwas enhanced in the cwn-1; cwn-2 double mutant (Fig. 3a), suggestingthat the two Wnt genes cooperate to inhibit neurite elimination.

To test our hypothesis that the function of CAM-1 in developmentalpruning requires Wnt regulation, we overexpressed CAM-1–GFP in thecwn-1; cwn-2 double mutant. CAM-1–GFP overexpression had nosignificant effect on developmental pruning in this mutant background(P ¼ 0.210, population rate test; Fig. 3b), strongly suggesting that theWnt genes and cam-1 function in the same pathway. We also examinedthe possibility that Wnt regulates neurite pruning by directly binding tothe Frizzled domain of CAM-1. When we overexpressed a mutant formof CAM-1 that lacked the Frizzled domain, CAM-1(DFrizzled), in AIMneurons, we did not detect a significant effect on neurite elimination(P ¼ 0.128, population rate test; Fig. 2b), supporting the possibilitythat a direct interaction between Wnt and CAM-1 is required to inhibitneurite elimination.

To confirm that the Wnt genes cwn-1 and cwn-2 function in the samepathway with cam-1, we examined whether the cwn-1; cwn-2 mutationcan suppress the mbr-1 mutant phenotype and vice versa. As predicted,neurite elimination in the mbr-1; cwn-1; cwn-2 triple mutant wassignificantly enhanced compared with the mbr-1 mutant (P o 0.01,population rate test; Fig. 3c) and significantly reduced compared withthe cwn-1; cwn-2 double mutant (P ¼ 0.029, population rate test;Fig. 3a,c), similar to the case in the cam-1 mutant. As with cam-1,mutations in cwn-1 and cwn-2 did not significantly affect the initial

20 80

60

40

20

0

0.454

0.000

0.210

0.7940.000

0.0000.000

0.000

0.002

0.602

0.0000.000

0.000 0.1890.765

0.372

Early L1 Adult

0.008

Wild

type – – – ––

cwn-

1(ok

546)

; egl-

20(n

585)

P mbr

-1 ::ca

m-1

::gfp

P mbr

-1 ::ca

m-1

::gfp

cwn-

1(ok

546)

; cwn-

2(ok

895)

cwn-

2(ok

895)

; lin-

44(n

1792

)

cwn-

1(ok

546)

; cwn-

2(ok

895)

cwn-

1(ok

546)

Wild

type

Wild

type

cwn-

1 an

d cw

n-2

cwn-

1

cwn-

2

cwn-

2(ok

895)

cam

-1(g

m12

2)

cam

-1(g

m12

2); c

wn-1(

ok54

6)

cam

-1(g

m12

2); c

wn-2(

ok89

5)

vang

-1(o

k114

2)

vang

-1(o

k114

2)

mbr

-1(q

a590

1)

mbr

-1(q

a590

1)

mbr

-1(q

a590

1);

vang

-1(o

k114

2)

mbr

-1(q

a590

1);

vang

-1(o

k114

2)

cwn-

1(ok

546)

;

cwn-

2(ok

895)

0.723

(%) (%)80

60

40

20

0

(%)80

60

40

20

0

(%)80

60

40

20

0

(%)

AIM

inte

rcon

nect

ed in

divi

dual

s

AIM

inte

rcon

nect

ed in

divi

dual

s

AIM

inte

rcon

nect

ed in

divi

dual

s

AIM

inte

rcon

nect

ed in

divi

dual

s

AIM

inte

rcon

nect

ed in

divi

dual

s

15

10

5

0

mbr-1(qa5901) background

Host: Host:

AdultEarly L1

0.601 0.0000.000

0.012

0.334

0.603

Wildtype

cwn-1(ok546);cwn-2(ok895)

mbr-1(qa5901);cwn-1(ok546);cwn-2(ok895)

80

60

40

20

0

(%)80

60

40

20

0

(%)

a b c d

e

Figure 3 Wnt is required for CAM-1–mediated inhibition of neurite elimination.

(a–d) Percentages of individuals with AIM interconnections in wild-type and Wnt

double mutant adults (n ¼ 74–117, a), wild-type and cwn-1; cwn-2 double mutant

adults overexpressing CAM-1–GFP in AIM under the mbr-1 promoter (n ¼ 67–117, b),

the mbr-1 mutant, mbr-1; cam-1, mbr-1; cwn-1 and mbr-1; cwn-2 double mutants,

mbr-1; cwn-1; cwn-2, mbr-1; cam-1; cwn-1 and mbr-1; cam-1; cwn-2 triple mutants

(n ¼ 51–174, c), and the mbr-1; cwn-1; cwn-2 triple mutant transfected with cwn-1

and cwn-2 genomic fragments (n ¼ 50–70, d). (e) Percentages of individuals with AIM

interconnections in wild-type, mbr-1 and vang-1 mutants and the mbr-1; vang-1 double

mutant (n ¼ 50–132). The numbers in the graph indicate P values in the population

rate test (see Supplementary Table 1 for numerical data).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 983

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

formation of the AIM interconnections (P ¼ 0.794, population ratetest; Fig. 3c) nor substantially affect the overall morphology of AIM(Supplementary Fig. 1).

We also generated mbr-1; cwn-1 and mbr-1; cwn-2 double mutants toelucidate the individual roles of each Wnt gene. Mutations in eithercwn-1 or cwn-2 alone moderately suppressed the mbr-1 mutant pheno-type in comparison with their simultaneous mutation (Fig. 3c). Thisresult implies that cwn-1 and cwn-2 have different functions, at least inpart, or that high total expression levels of cwn-1 and cwn-2 are critical.To discriminate between these two possibilities, we compared the effectsof overexpressing cwn-1 and cwn-2 alone or together in the mbr-1;

cwn-1; cwn-2 triple mutant. Overexpression of single and double genesinhibited neurite elimination to similar extents (Fig. 3d), suggestingthat cwn-1 and cwn-2 function via similar mechanisms, thus supportingthe latter possibility. In addition, mutations in cwn-1 or in cwn-2 did notfurther enhance the increased rate of neurite elimination in the cam-1mutant background (Fig. 3c). These results suggest that both genesact in the same pathway with cam-1, although it remains possible thatother receptors also function in parallel with CAM-1 to transduce theWnt signal.

The mechanisms by which Wnt-activated CAM-1 inhibits neuriteelimination cannot be easily predicted by analogy to other functions of

1 kb

1 kb

K10B4.1

T04B2.5

ATG

ATG W01B6.3

LGIV

cwn-1

cwn-2

gfpgfp

gfpgfp

Pcwn-1(0.7 kb)::gfp

Pcwn-2(4.0 kb)::cwn-2Pcwn-2(2.1 kb)::cwn-2

Pcwn-2(0.8 kb)::gfpPcwn-2(160 bp)::gfp

Pcwn-2(0.8 kb)::gfpPcwn-1(0.7 kb)::gfp

Pcwn-1(170 bp)::gfp

Pcwn-2(0.8 kb)::cwn-2Pcwn-2(160 bp)::cwn-2

Pcwn-1(1.5 kb)::cwn-1cwn-1 fragment

cwn-2 fragment

Pcwn-1(0.7 kb)::cwn-1Pcwn-1(170 bp)::cwn-1

Pcwn-1(170 bp)::gfp

srd-3

LGll

Pcwn-2(160 bp)::gfp

AIM

inte

rcon

nect

ed in

divi

dual

s

AIM

inte

rcon

nect

ed in

divi

dual

s

600.000

0.000

0.0000.000

0.000

0.158 0.121

Promoter

40

40

(%)

0.003

0.043

0.370

30

20

10

0

20

1.5

kb

Neuro

ns

Phary

ngea

l

mus

cle AIM–

0.7

kb

4.0

kb

2.1

kb

0.8

kb

160

bp

170

bp

Trans-gene: cwn-1 cwn-2

–0

(%)a

cd

e f

b

DIC/GFP

Figure 4 Wnt is provided nonautonomously by nearby neurons. (a) Genomic positions of cwn-1 relative to neighboring genes on linkage group II. The

genomic fragments used in cwn-1 rescue experiments (Figs. 3d and 4b) and the gfp fusion constructs used for expression analyses (e) are also indicated.

(b) Percentages of adults with AIM interconnections in the mbr-1; cwn-1; cwn-2 triple mutant transfected with cwn-1 and cwn-2 genomic fragments

containing various lengths of upstream sequence (n ¼ 40–65). The numbers in the graph indicate P values in the population rate test when compared with

the leftmost control strain (see Supplementary Table 1 for numerical data). (c) Genomic positions of cwn-2 relative to neighboring genes on linkage group IV.

The genomic fragments used in cwn-2 rescue experiments (Figs. 3d and 4b) and the gfp fusion constructs used for expression analyses (f) are also indicated.

(d) Percentages of adult individuals with AIM interconnections in the mbr-1; cwn-1; cwn-2 triple mutant transfected with cwn-1 and cwn-2 cDNAs linked

to the following promoters for expression in the indicated cell types: H20 (neurons), myo-2 (pharyngeal muscle) and mbr-1 (AIM, n ¼ 46–70). The numbers

in the graph indicate P values in the population rate test when compared with the leftmost control strain (see Supplementary Table 1 for numerical data).

(e,f) Expression of Pcwn-1(0.7 kb)Hgfp and Pcwn-1(170 bp)Hgfp (e) or Pcwn-2(0.8 kb)Hgfp and Pcwn-2(160 bp)Hgfp (f) observed from the lateral aspect. Differential

interference contrast (DIC) images and fluorescent signals are each shown in gray and green. White arrowheads indicate cells in which GFP-derived signals

were only detected in Pcwn-1(0.7 kb)Hgfp (ventral nerve cord neurons, e) or Pcwn-2(0.8 kb)Hgfp (pharyngeal neurons, f) transfected worms. Magenta arrowheads

indicate the putative positions of AIM somas. Asterisks indicate the anterior and posterior ends of the intestine. In all strains, we detected fluorescent signals

in the intestine and in a few head neurons. In e, the white arrow indicates neuronal processes that extend from the ventral nerve cord neurons expressing

Pcwn-1(0.7 kb)Hgfp to near AIM. In f, the right images show an enlarged view of the head region enclosed in the left images and the dashed lines indicate

the outline of the pharynx. Scale bars represent 50 mm.

984 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Ror kinases, as most of them do not require the intracellular region ofRor kinases, whereas regulation of developmental pruning in AIMdoes. In one case, during the development of the C. elegans vulva,CAM-1 regulates certain aspects of cell polarity in an intracellularregion–dependent manner18. During this event, CAM-1 functions inthe same pathway with the Wnt factor EGL-20 and the Van Goghortholog VANG-1, which is the core component of the planar cellpolarity (PCP) pathway18. We tested whether developmental pruningin AIM neurons was affected in the vang-1 mutant. Similar to thecam-1 and cwn-1; cwn-2 mutants, neurite elimination was enhanced inthe vang-1 mutant in both wild-type and mbr-1 mutant backgrounds,although to a smaller extent (Fig. 3e). Thus, Wnt and CAM-1 mightregulate developmental pruning at least partly via the PCP pathway(see Discussion).

Wnt is provided non-autonomously by nearby neurons

To identify the cellular sources of CWN-1 and CWN-2, we searched forcwn-1 and cwn-2 promoter regions that are essential for the function ofthese genes in developmental pruning and then addressed the cells inwhich the promoter regions drive expression (Fig. 4). We made a seriesof genomic fragments of cwn-1 and cwn-2, including various lengths ofDNA upstream of the initiation codon, and introduced them into thembr-1; cwn-1; cwn-2 triple mutant (Fig. 4a–c).

In the case of cwn-1, the percentage of adult individuals withAIM interconnections significantly increased when we introducedgenomic fragments with upstream sequences of 1.4 or 0.7 kilobases(kb, P o 0.001, population rate test; Fig. 4b), but not 170 base pairs(bp, P o 0.158, population rate test; Fig. 4b), suggesting that theregulatory sequences that are important for cwn-1 function in develop-mental pruning are located within 170 bp to 0.7 kb upstream of theinitiation codon. Therefore, we compared the expression of GFP linkedto 0.7 kb or 170 bp of cwn-1 upstream sequence (Pcwn-1(0.7 kb)Hgfpand Pcwn-1(170 bp)Hgfp, respectively; Fig. 4a); cells expressing GFP fromPcwn-1(0.7 kb)Hgfp, but not Pcwn-1(170 bp)Hgfp, are strong candidates asthe major source of CWN-1. Although both Pcwn-1(0.7 kb)Hgfp andPcwn-1(170 bp)Hgfp induced GFP expression in a small number of headneurons and the intestine, only Pcwn-1(0.7 kb)Hgfp induced GFP expres-sion in the ventral nerve cord neurons (Fig. 4e). Some of these neuronsprojected their processes immediately ventral to AIM neurons (Fig. 4e),and CWN-1 might be secreted from such structures.

On the other hand, in the case of cwn-2, the number of adult indi-viduals with AIM interconnections significantly increased when we

introduced genomic fragments with upstream sequences of 4.0, 2.1or 0.8 kb (P o 0.001, population rate test; Fig. 4b), but not 160 bp(P ¼ 0.121, population rate test; Fig. 4b), suggesting that importantregulatory sequences are located within 160 bp to 0.8 kb upstream.Although the introduction of gfp linked to 0.8 kb or 160 bp of thecwn-2 upstream sequence (Pcwn-2(0.8 kb)Hgfp and Pcwn-2(160 bp)Hgfp,respectively; Fig. 4c) both induced GFP expression in a small numberof head neurons and in the intestine, only Pcwn-2(0.8 kb)Hgfp inducedGFP expression in pharyngeal neurons (Fig. 4f). These neurons arelocated on the dorsal side of the AIM neurons and are probably themajor source of CWN-2.

In sum, these rescue experiments and expression analyses suggestthat, during developmental pruning of AIM, the two Wnt factorsCWN-1 and CWN-2 are each provided by neurons located on theventral and dorsal side of AIM neurons to modulate the rate of neuriteelimination. Consistent with this, expressing cwn-1 and cwn-2 usingthe pan-neural promoter H20 (ref. 20) also rescued the reduction ofAIM connection in the adults of mbr-1; cwn-1; cwn-2 mutant (Fig. 4d).Expressing cwn-1 and cwn-2 in the pharyngeal muscle using the myo-2promoter21 also had a subtle, yet significant, rescuing effect (P ¼ 0.043,population rate test; Fig. 4d), supporting the idea that Wnt functions ina nonautonomous manner. In contrast, cwn-1 and cwn-2 expression inAIM itself and several other neurons using the mbr-1 promoter had nosignificant effect (P ¼ 0.370, population rate test; Fig. 4d). This mightbe a result of a lack of ability of these neurons to properly secrete Wnt;cells are probably not homogeneous in their ability to secrete Wnt, asthe transmembrane protein MIG-14/Wntless, which is required forWnt secretion, is not expressed in all cells22.

Behavioral studies reveal that pruning also occurs in RIF

So far, we have examined the opposing roles of Wnt-Ror kinasesignaling and MBR-1 in the developmental pruning of AIM. Thegenerality of this mechanism should be established if the develop-mental program is also used in other neuronal subclasses. AlthoughAIM is the only subclass in C. elegans previously reported to undergodevelopmental pruning, behavioral analyses of the mbr-1 mutantindicates that this event also occurs in another subclass.

When we compared the behavioral responses of wild-type and mbr-1mutant adults to various stimulants in an olfactory plasticity assay,the mutants showed an altered response to the odorant benzaldehyde.Although wild-type worms showed reduced chemotaxis to benzaldehydeafter 5 min pre-exposure to the same odorant23, such plasticity was

1.0 0.0000.000

0.001

0.013

Pre-exposure toGFP INX-19/NSY-5–mRFP

Water

Benzaldehyde

Che

mot

axis

inde

x

mbr

-1(q

a590

1)W

ildty

pe

RIF

AIM

0.8

0.6

0.4

0.2

–0.0

–0.2

Host:Wildtype

– – Pmbr-1

mbr-1(qa5901)

Podr-2 Plin-11–0.4

a b c

Figure 5 Developmental neurite pruning occurs in multiple neuronal subclasses, as revealed by behavioral

analyses of the mbr-1 mutant. (a) Results of olfactory plasticity assays using the odorant benzaldehyde of

wild-type, the mbr-1 mutant and the mbr-1 mutant transfected with rescue constructs in which mbr-1 was expressed under the indicated promoters. The three

promoters were coexpressed in RIF (Supplementary Fig. 3). Error bars represent s.e.m. (n ¼ 5–21). The numbers in the graph indicate P values in the Welch’s

test. Chemotaxis following pre-exposure to water produced no significant difference between experimental groups. (b) Images of typical wild-type and mbr-1

mutant adult RIF visualized by Pmbr-1Hgfp. The left and right RIF neurons of mbr-1 mutant adults were frequently interconnected by ectopic neurites (arrowhead).

(c) Localization of the gap junction protein INX-19/NSY-5–mRFP during early L1 in wild-type RIF and AIM. The morphologies of the neurons were visualized by

coexpression with GFP. Punctate signals derived from INX-19/NSY-5–mRFP were detected at the interconnections (arrowheads). Scale bars represent 5 mm.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 985

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

considerably reduced in the mbr-1 mutant (Fig. 5a). To determine thecell type in which mbr-1 functions, we performed rescue experimentsusing different promoters, and the lin-11, odr-2 and mbr-1 promoterswere found to rescue the behavioral phenotype (Fig. 5a). Thesepromoters were coexpressed only in the interneuron subclass RIF(Supplementary Fig. 3). To further investigate the involvement ofRIF in olfactory plasticity, we analyzed lin-11, which encodes a LIMhomeodomain transcription factor24 and is expressed in RIF (Supple-mentary Fig. 3). The lin-11 loss-of-function mutant also showedreduced olfactory plasticity to benzaldehyde (Supplementary Fig. 4)and mbr-1 expression was specifically reduced in RIF in these mutants(Supplementary Fig. 4). Together, these results suggest that mbr-1expression in RIF is important for olfactory plasticity. When weobserved the morphology of RIF in wild-type and mbr-1 mutantadults, the left and right neurons of the mbr-1 mutant were typicallyconnected by neurites that were rarely seen in wild-type adults(Fig. 5b). During early L1, it was difficult to judge whether wild-typeRIF neurons were also interconnected as a result of the close loca-tions of the somas. Nonetheless, localization of a gap junction proteinINX-19/NSY-5 (refs. 25,26) at the point of contact implied theexistence of a neuronal connection (Fig. 5c and SupplementaryFig. 5). We therefore conclude that, aside from AIM, RIF also under-goes developmental neurite pruning in an MBR-1–dependent manner.

Given the finding that developmental pruning also occurs in RIF, weexamined whether Wnt-Ror kinase signaling is involved in the pruningof this neural subclass. As in AIM, mutations in cam-1 or in cwn-1 andcwn-2 decreased the percentage of adults with RIF interconnections(Fig. 6). Moreover, mutations in these genes and mbr-1 counteractedeach other’s effects (Fig. 6), indicating the similarity in the mechanismsof neurite pruning between RIF and AIM. Our results support thenotion that Wnt-Ror kinase signaling has a general role in C. elegansdevelopmental neurite pruning, acting as a trophic signal that counter-acts the regressing effect of the transcription factor MBR-1 on neurites(Supplementary Fig. 6).

DISCUSSION

The trophic theory of neurite survival has long been proposed asa mechanism to explain developmental pruning2. Over the years,researchers have sought the molecules that assume the trophic func-tion. One candidate is the family of neurotrophins, which regulates thesurvival of neurites and of the neurons themselves1. In at least somewell-studied in vivo systems, however, neurotrophins seem to not exertany trophic effects on neurites during developmental pruning27,28.Moreover, the invertebrate animal models Drosophila melanogasterand C. elegans seem to lack obvious homologs of neurotrophins29,whereas developmental pruning is a widely conserved phenomenon.To the best of our knowledge, this is the first report to identify secretedfactors that are widely conserved in both vertebrate and invertebrateanimal models that can exert a trophic effect during developmentalpruning. In mammals and D. melanogaster, Wnt secretion is regu-lated in an activity-dependent manner30–32. In the future, it will beinteresting to test whether Wnt-Ror kinase signaling is involved in theactivity-dependent competitive pruning of neurites found in manydeveloping systems.

Here, we also found that developmental pruning is not restricted to asingle neural subclass, but instead may be a general event in C. elegans.In addition to the previously identified AIM, we found that RIF issubject to developmental pruning, and expression of mbr-1, whichpromotes neurite elimination, in this subclass affected olfactorybehavior. In RIF, the left and right neurons form synapses with differentneuronal subclasses and the two neurons are not connected to eachother at the nerve ring, unlike most other neurons33. These anatomicalfeatures might be associated with functional asymmetry; the high rateof interconnections in the mbr-1 mutant adults might be one factorunderlying altered behavior, although direct evidence for this islacking. We expect that our findings will provide a start point forunderstanding the physiological importance of developmental pruningin C. elegans.

The mechanisms by which Wnt–CAM-1 signaling and MBR-1inhibit or promote neurite elimination, respectively, remain unsolved.A hallmark of developmental pruning is the local fragmentation ofthe cytoskeleton in neurites that are to be eliminated5 and the twoopposing pathways might function by affecting this process. TheD. melanogaster ortholog of MBR-1, E93, induces expression of theapoptotic protease Dronc during metamorphosis to cause large-scaletissue destruction34,35. Dronc is also known to act locally to direct thedevelopmental pruning of dendrites during metamorphosis36,37 andMBR-1 might promote neurite elimination in C. elegans by regulatingsimilar factors. On the other hand, Wnt-activated CAM-1 mightcounteract this process by stabilizing the cytoskeleton. During thedevelopment of the C. elegans vulva, CAM-1 and the Wnt factorEGL-20 cooperate with the PCP pathway component VANG-1 in amanner that is independent of the major Wnt receptor Frizzled18. Itwas thus proposed that CAM-1 and EGL-20 comprise a previouslyunknown Wnt pathway affecting the PCP pathway. The PCP pathwayhas various roles in the regulation of cell morphology by locallystabilizing the cytoskeleton via small GTPases38. During developmentalpruning, neurites in which CAM-1 is activated by Wnt might tolerateelimination through a PCP pathway-dependent stabilization of thecytoskeleton. Consistent with this, we found that the vang-1 mutantalso showed an increased rate of developmental pruning. These modelsmight be tested by further screening for mutants with altered rates ofdevelopmental pruning, which should contribute to the dissection of aWnt pathway.

The identification of a nonautonomous pathway inhibitingneurite elimination is also important from the perspective of medical

800.000 0.000

0.000

0.0120.001

0.025

0.046

Wild

type

cam

-1(a

k37)

cwn-

1(ok

546)

; cwn-

2(ok

895)

mbr

-1(q

a590

1)

mbr

-1(q

a590

1); c

am-1

(ak3

7)

mbr

-1(q

a590

1);

cwn-

1(ok

546)

; cwn-

2(ok

895)

(%)

RIF

inte

rcon

nect

ed in

divi

dual

s

60

40

20

0

Figure 6 Wnt, CAM-1 and MBR-1 also function in developmental neurite

pruning of neurons involved in olfactory processing. The percentages of adults

with RIF interconnections in wild-type, mbr-1 and cam-1 mutants, mbr-1;

cam-1 and cwn-1; cwn-2 double mutants, and the mbr-1; cwn-1; cwn-2 triple

mutant (n ¼ 50–72) are shown. The numbers in the graph indicate P values

in the population rate test (see Supplementary Table 1 for numerical data).

986 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

application. Many neurodegenerative disorders, including Alzheimer’s,Parkinson’s and motoneuron diseases, are accompanied by neuriteatrophy3. Notably, in animal models, genetically delaying neurite elimi-nation alleviates the symptoms of some of these diseases39,40. Thediscovery of administrable drugs that prohibit pathologic eliminationof neurites is desired and the Wnt-Ror kinase pathway is a strongcandidate target for this purpose.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe thank the Caenorhabditis Genetics Center funded by the US NationalInstitutes of Health National Center for Research Resources for nematode strains;M. Tomioka, K. Yamada and M. Matsuki for constructs and technical advice;A. Fire for vectors; C. Bargmann, A. Kuhara and I. Mori for information onconstructs; O. Hobert for sharing unpublished results on neurite pruning; andE. Matsuzaka for technical assistance. This work was supported by the Programfor Promotion of Basic Research Activities for Innovative Bioscience. Y.H.was the recipient of a Grant-in-Aid for Japan Society for the Promotion ofScience Fellows.

AUTHOR CONTRIBUTIONSY.H. designed and Y.H., R.I. and H.K. conducted the experiments on anatomicalstudies. Y.H., T.H. and E.K.-N. designed and conducted the experiments onbehavioral studies. Y.H. wrote the manuscript. T.I., Y.I. and T.K. supervisedthe project.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Low, L.K. & Cheng, H.J. Axon pruning: an essential step underlying the develop-mental plasticity of neuronal connections. Philos. Trans. R. Soc. Lond. B Biol. Sci.361, 1531–1544 (2006).

2. Lichtman, J.W. & Colman, H. Synapse elimination and indelible memory. Neuron 25,269–278 (2000).

3. Raff, M.C., Whitmore, A.V. & Finn, J.T. Axonal self-destruction and neurodegeneration.Science 296, 868–871 (2002).

4. Nave, K.A. & Trapp, B.D. Axon-glial signaling and the glial support of axon function.Annu. Rev. Neurosci. 31, 535–561 (2008).

5. Luo, L. & O’Leary, D.D. Axon retraction and degeneration in development and disease.Annu. Rev. Neurosci. 28, 127–156 (2005).

6. Kage, E. et al. MBR-1, a novel helix-turn-helix transcription factor, is required for pruningexcessive neurites in Caenorhabditis elegans. Curr. Biol. 15, 1554–1559 (2005).

7. Koga, M., Take-uchi, M., Tameishi, T. & Ohshima, Y. Control of DAF-7 TGF-b expressionand neuronal process development by a receptor tyrosine kinase KIN-8 in Caenorhabditiselegans. Development 126, 5387–5398 (1999).

8. Forrester, W.C., Dell, M., Perens, E. & Garriga, G.A. C. elegans Ror receptor tyrosinekinase regulates cell motility and asymmetric cell division. Nature 400, 881–885(1999).

9. Oishi, I. et al. Spatio-temporally regulated expression of receptor tyrosine kinases,mRor1, mRor2, during mouse development: implications in development and functionof the nervous system. Genes Cells 4, 41–56 (1999).

10. Francis, M.M. et al. The Ror receptor tyrosine kinase CAM-1 is required for ACR-16-mediated synaptic transmission at the C. elegans neuromuscular junction. Neuron 46,581–594 (2005).

11. Kim, C. & Forrester, W.C. Functional analysis of the domains of the C. elegans Rorreceptor tyrosine kinase CAM-1. Dev. Biol. 264, 376–390 (2003).

12. Aurelio, O., Hall, D.H. & Hobert, O. Immunoglobulin-domain proteins required formaintenance of ventral nerve cord organization. Science 295, 686–690 (2002).

13. Hikasa, H., Shibata, M., Hiratani, I. & Taira, M. The Xenopus receptor tyrosine kinaseXror2 modulates morphogenetic movements of the axial mesoderm and neuroectodermvia Wnt signaling. Development 129, 5227–5239 (2002).

14. Oishi, I. et al. The receptor tyrosine kinase Ror2 is involved in noncanonical Wnt5a/JNKsignalling pathway. Genes Cells 8, 645–654 (2003).

15. Mikels, A.J. & Nusse, R. Purified Wnt5a protein activates or inhibits beta-catenin-TCFsignaling depending on receptor context. PLoS Biol. 4, e115 (2006).

16. Green, J.L., Inoue, T. & Sternberg, P.W. The C. elegans ROR receptor tyrosine kinase,CAM-1, nonautonomously inhibits the Wnt pathway. Development 134, 4053–4062(2007).

17. Schambony, A. & Wedlich, D. Wnt-5A/Ror2 regulate expression of XPAPC through analternative noncanonical signaling pathway. Dev. Cell 12, 779–792 (2007).

18. Green, J.L., Inoue, T. & Sternberg, P.W. Opposing Wnt pathways orient cell polarityduring organogenesis. Cell 134, 646–656 (2008).

19. Eisenmann, D.M. Wnt signaling. WormBook ohttp://www.wormbook.org/chapters/www_wntsignaling/wntsignaling.html4 (2005).

20. Shioi, G. et al. Mutations affecting nerve attachment of Caenorhabditis elegans.Genetics 157, 1611–1622 (2001).

21. Gaudet, J. & Mango, S.E. Regulation of organogenesis by the Caenorhabditis elegansFoxA protein PHA-4. Science 295, 821–825 (2002).

22. Yang, P.T. et al. Wnt signaling requires retromer-dependent recycling of MIG-14/Wntlessin Wnt-producing cells. Dev. Cell 14, 140–147 (2008).

23. Hirotsu, T. & Iino, Y. Neural circuit–dependent odor adaptation in C. elegans is regulatedby the Ras-MAPK pathway. Genes Cells 10, 517–530 (2005).

24. Freyd, G., Kim, S.K. & Horvitz, H.R. Novel cysteine-rich motif and homeodomain in theproduct of the Caenorhabditis elegans cell lineage gene lin-11. Nature 344, 876–879(1990).

25. Barnes, T.M. & Hekimi, S. The Caenorhabditis elegans avermectin resistance andanesthetic response gene unc-9 encodes a member of a protein family implicated inelectrical coupling of excitable cells. J. Neurochem. 69, 2251–2260 (1997).

26. Chuang, C.F., Vanhoven, M.K., Fetter, R.D., Verselis, V.K. & Bargmann, C.I. An innexin-dependent cell network establishes left-right neuronal asymmetry in C. elegans. Cell129, 787–799 (2007).

27. Nguyen, Q.T., Parsadanian, A.S., Snider, W.D. & Lichtman, J.W. Hyperinnervation ofneuromuscular junctions caused by GDNF overexpression in muscle. Science 279,1725–1729 (1998).

28. Woolley, A.G., Sheard, P.W. & Duxson, M.J. Neurotrophin-3 null mutant mice display apostnatal motor neuropathy. Eur. J. Neurosci. 21, 2100–2110 (2005).

29. Jaaro, H., Beck, G., Conticello, S.G. & Fainzilber, M. Evolving better brains: a need forneurotrophins? Trends Neurosci. 24, 79–85 (2001).

30. Chen, J., Park, C.S. & Tang, S.J. Activity-dependent synaptic Wnt release regulateshippocampal long term potentiation. J. Biol. Chem. 281, 11910–11916 (2006).

31. Wayman, G.A. et al. Activity-dependent dendritic arborization mediated by CaM-kinase Iactivation and enhanced CREB-dependent transcription of Wnt-2. Neuron 50, 897–909(2006).

32. Ataman, B. et al. Rapid activity-dependent modifications in synaptic structure andfunction require bidirectional Wnt signaling. Neuron 57, 705–718 (2008).

33. White, J.G., Southgate, E., Thomson, J.N. & Brenner, S. The structure of the nervoussystem of the nematode Caenorhabditis elegans. Philos. Trans. R. Soc. Lond. B Biol. Sci.314, 1–340 (1986).

34. Lee, C.Y., Cooksey, B.A. & Baehrecke, E.H. Steroid regulation of midgut cell deathduring Drosophila development. Dev. Biol. 250, 101–111 (2002).

35. Daish, T.J., Cakouros, D. & Kumar, S. Distinct promoter regions regulate spatial andtemporal expression of the Drosophila caspase dronc. Cell Death Differ. 10, 1348–1356(2003).

36. Williams, D.W., Kondo, S., Krzyzanowska, A., Hiromi, Y. & Truman, J.W. Local caspaseactivity directs engulfment of dendrites during pruning. Nat. Neurosci. 9, 1234–1236(2006).

37. Kuo, C.T., Zhu, S., Younger, S., Jan, L.Y. & Jan, Y.N. Identification of E2/E3 ubiquitinat-ing enzymes and caspase activity regulating Drosophila sensory neuron dendritepruning. Neuron 51, 283–290 (2006).

38. Montcouquiol, M., Crenshaw, E.B. III & Kelley, M.W. Noncanonical Wnt signaling andneural polarity. Annu. Rev. Neurosci. 29, 363–386 (2006).

39. Ferri, A., Sanes, J.R., Coleman, M.P., Cunningham, J.M. & Kato, A.C. Inhibiting axondegeneration and synapse loss attenuates apoptosis and disease progression in a mousemodel of motoneuron disease. Curr. Biol. 13, 669–673 (2003).

40. Sajadi, A., Schneider, B.L. & Aebischer, P. Wlds-mediated protection of dopaminergicfibers in an animal model of Parkinson disease. Curr. Biol. 14, 326–330 (2004).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 987

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSStrains. We obtained the wild-type strain N2 and strains carrying the

following mutations from the Caenorhabditis Genetics Center: cam-1(ak37),

cam-1(gm122), cwn-1(ok546), cwn-2(ok895), egl-20(n585), him-5(e1467), lin-

11(n389), lin-44(n1792) and vang-1(ok1142). We generated mbr-1(qa5901) in a

previous study6.

Transgenic lines. To generate transgenic lines that were applied to behavioral

assays, we injected constructs at 5 ng ml–1 with Pmyo-3Hgfp as an injection

marker (25 ng ml–1, only the injection marker was injected in the case of control

lines). We injected transcriptional reporter constructs at 30–50 ng ml–1 with

pRF-4 as an injection marker (50 ng ml–1). We injected other constructs at 5 ng

ml–1 with pRF-4 as an injection marker (50 ng ml–1).

Fluorescence imaging. We obtained images using an Olympus IX71 confocal

microscope controlled by Olympus Fluoview. We examined worms within 4 h

of hatching at the early L1 and worms within 10–12 h of hatching at the late L1.

Constructs. To generate Pmbr-1Hcam-1Hgfp worms, we inserted a 5-kb

sequence upstream of the mbr-1 initiation codon into pDONR201 (Invitrogen)

by BP-recombinase reaction (Gateway System, Invitrogen) to generate pENTR-

Pmbr-1. Next, we linked together cam-1 cDNA, which was amplified by reverse

transcription (RT)-PCR, and the GFP coding sequence with the unc-54 3¢ UTR

sequence, which was amplified by PCR from pPD95_75 (a gift from A. Fire,

Stanford University), in order by splicing by overlapping extension (SOE)-PCR

to generate cam-1Hgfp. We inserted cam-1Hgfp into the multi-cloning site

of the plasmid pPD-DEST3 (a gift from M. Matsuki, the University of Tokyo)

to generate pDEST-cam-1Hgfp. Finally, we carried out the LR-recombinase

reaction (Gateway System, Invitrogen) between pENTR-Pmbr-1 and pDEST-

cam-1Hgfp to generate Pmbr-1Hcam-1Hgfp.

To generate Pcam-1Hgfp, we linked together a 5.8-kb sequence upstream of

the cam-1 initiation codon and the GFP coding sequence with the unc-54

3¢ UTR sequence by SOE-PCR to generate Pcam-1Hgfp. We generated mutant

versions of cam-1, Pmbr-1Hcam-1(DFrizzled)Hgfp, Pmbr-1Hcam-1(DIntra)Hgfp

and Pmbr-1Hcam-1(kinase dead)Hgfp, in accordance with previous reports8,10,11.

For Pmbr-1Hcam-1(DFrizzled)Hgfp, we amplified cam-1 cDNA sequences cor-

responding to CAM-1(1–178) and CAM-1(304–902) by RT-PCR and linked

them together in order with the GFP coding sequence with the unc-54 3¢ UTR

sequence by SOE-PCR to generate cam-1(DFrizzled)Hgfp. For Pmbr-1Hcam-

1(DIntra)Hgfp, we amplified the cam-1 cDNA sequence corresponding to

CAM-1(1–469) by RT-PCR and linked it with the GFP coding sequence

and the unc-54 3¢ UTR sequence in order by SOE-PCR to generate cam-

1(DIntra)Hgfp. For Pmbr-1Hcam-1(kinase dead)Hgfp, we amplified cam-1

cDNA sequences corresponding to CAM-1(1–599) and CAM-1(600–902) by

RT-PCR and linked them together in order with the GFP coding sequence and

the unc-54 3¢ UTR sequence by SOE-PCR. During this PCR, we replaced the

cDNA sequence AAGAAA, which corresponds to the two lysine residues in the

putative ATP-binding motif (CAM-1(598–599)), with CGTCGT, resulting in a

lysine to arginine alteration (cam-1(kinase dead)Hgfp). Subsequent methods

were the same for Pmbr-1Hcam-1Hgfp.

To generate Pzig-3Hcam-1Hgfp, we inserted a 4.3-kb sequence upstream of the

zig-3 initiation codon into pDONR201 by BP-recombinase reaction to generate

pENTR-Pzig-3. Subsequent methods were the same for Pmbr-1Hcam-1Hgfp. To

generate Plin-11Hmbr-1, we inserted the mbr-1 cDNA into the multi-cloning site

of the plasmid pPD-DEST3 to generate pDEST-mbr-1, and the third intron

of lin-11, which drives expression in a few neurons41, into pDONR201 by

BP-recombinase reaction to generate pENTR-Plin-11. Then we carried out an

LR-recombinase reaction between pDEST-mbr-1 and pENTR-Plin-11 to generate

Plin-11Hmbr-1. We made Podr-2Hmbr-1 similarly using pENTR-Podr-2 (a gift from

M. Tomioka, University of Tokyo), in which a 3-kb sequence upstream of the

odr-2 2b exon4 (ref. 42) is inserted into pDONR201 by BP-recombinase reaction.

Pmbr-1Hmbr-1 is a 7.5-kb genomic fragment containing the 5-kb upstream

sequence, the 2.3-kb coding sequence and a 0.2-kb 3¢-UTR region of mbr-1.

We deleted the GFP coding sequence from pPD95_75 and instead inserted

the mRFP coding sequence to generate pPD95_75-mrfp. Next, we inserted

nsy-5 cDNA, which we amplified by RT-PCR, into the multi-cloning site

of pPD95_75-mrfp in frame with mrfp to generate pPD95_75–nsy-5Hmrfp.

Finally, we amplified a 5-kb sequence upstream of the mbr-1 initiation

codon, nsy-5Hmrfp, and an 1.2-kb 3¢-UTR region of nsy-5 and linked them

together in order by SOE-PCR to generate Pmbr-1Hnsy-5Hmrfp. Podr-2Hgfp,

Plin-11Hmrfp, Podr-2Hmrfp, Pmbr-1Hmrfp were generated by LR-recombinase

reaction between pENTR-Plin-11, pENTR-Podr-2 or pENTR-Pmbr-1 and pDEST-

gfp or pDEST-mrfp.

Pcwn-1(1.8 kb)Hcwn-1, Pcwn-1(1.5 kb)Hcwn-1, Pcwn-1(0.7 kb)Hcwn-1, Pcwn-1(170 bp)Hcwn-1 are genomic fragments containing the 2.4-kb coding and 0.3-kb 3¢-UTR

regions (2.4-kb coding and 2.4-kb 3¢-UTR regions for Pcwn-1(1.8 kb)Hcwn-1) of

cwn-1 plus the indicated length of sequence upstream of the initiation codon.

We refer to Pcwn-1(1.8 kb)Hcwn-1 simply as cwn-1 (Fig. 3d). Pcwn-2(5.8 kb)Hcwn-2,

Pcwn-2(4.0 kb)Hcwn-2, Pcwn-2(2.1 kb)Hcwn-2, Pcwn-2(0.8 kb)Hcwn-2, Pcwn-2(160 bp)Hcwn-2 are genomic fragments containing the 2.1-kb coding and 0.8-kb 3¢-UTR

regions of cwn-2 plus the indicated length of sequence upstream of the

initiation codon. We refer to Pcwn-2(5.8 kb)Hcwn-2 simply as cwn-2 (Fig. 3d).

For Pcwn-1(0.7 kb)Hgfp, Pcwn-1(170 bp)Hgfp, Pcwn-2(0.8 kb)Hgfp, Pcwn-1(160 bp)Hgfp, we

linked the indicated length of genomic fragments upstream of the cwn-1 or

cwn-2 initiation codon with the GFP coding sequence plus unc-54 3¢ UTR

sequence by SOE-PCR.

For Pmbr-1Hcwn-1, Pmbr-1Hcwn-2, Pmyo-2Hcwn-1, Pmyo-2Hcwn-2, PH20Hcwn-1,

PH20Hcwn-2, we amplified cwn-1 and cwn-2 cDNAs by RT-PCR and inserted

them into the multi-cloning site of the plasmid pPD-DEST3 to generate pDEST-

cwn-1 and pDEST-cwn-2. We inserted a 1-kb sequence upstream of the myo-2

initiation codon, which drives expression in the pharyngeal muscle21, into

pDONR201 to generate pENTR-Pmyo-2. pENTR-PH20 (a gift from M. Matsuki,

University of Tokyo), was generated by inserting a 2.5-kb sequence amplified

from genomic DNA using the primers CTC CTT GAA GCT CAT CCA G and

TGG GCG CCT GCA GGA ATT TTT, which drives expression in all neurons20,

into pDONR201. We carried out an LR-recombinase reaction between pENTR-

Pmbr-1, pENTR-Pmyo-2 or pENTR-PH20 and pDEST-cwn-1 or pDEST-cwn-2.

Construction of Pmbr-1Hgfp and mbr-1Hgfp was described previously6.

Olfactory plasticity assay. We carried out olfactory plasticity assays as

described previously23. We collected worms grown on Escherichia coli strain

NA22–pasted NGM plates in microfuge tubes and washed them three times

with basal buffer (5 mM K3PO4, 1 mM CaCl2, 1 mM MgSO4 and 0.5 g L–1

gelatin) and then we added 100 ml of 10�4 dilutions of benzaldehyde in water

or 100 ml of water (control group). After 5 min, we removed 70 ml of the

supernatant and added 1ml of basal buffer and centrifuged them for 5 s at 100g.

We took the worms that settled at the bottom and spotted about 50 of them at

the center of 9-cm assay plates. We removed excess liquid with paper towels,

at the same time dispersing the animals along the midline of the plates. We

prepared assay plates by spotting 1 ml of a 10�2 dilution of benzaldehyde and

0.5 ml of 1M sodium azide as an anesthetic on two points separated by 2.5 cm at

one end of the 9-cm diameter plates. Sodium azide was solely spotted on the

other side. We counted the number of animals 30 min after placing them at

the center of the plate and calculated the chemotaxis index. The chemotaxis

index ¼ (A � B)/(A + B), where A is the number of worms on the odorant-

spotted side of the plate and B is the number of the worms on the opposite side,

whereas worms that remained within 0.5 cm of the midline were not adopted

in the chemotaxis index to exclude immotile worms from consideration.

Statistical analyses. We did comparisons of the percentage of worms with

neuronal interconnections between strains using the population rate test. We

carried out comparisons of the chemotaxis index between strains using the

Welch’s test. All tests were two tailed. We considered the results of all tests to be

significant when P o 0.05.

41. Hobert, O., D’Alberti, T., Liu, Y. & Ruvkun, G. Control of neural development andfunction in a thermoregulatory network by the LIM homeobox gene lin-11. J. Neurosci.18, 2084–2096 (1998).

42. Chou, J.H., Bargmann, C.I. & Sengupta, P. The Caenorhabditis elegans odr-2 geneencodes a novel Ly-6–related protein required for olfaction. Genetics 157, 211–224(2001).

NATURE NEUROSCIENCE doi:10.1038/nn.2347

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

A discrete alcohol pocket involved in GIRKchannel activation

Prafulla Aryal1,2, Hay Dvir3, Senyon Choe2,3 & Paul A Slesinger1,2

Ethanol modifies neural activity in the brain by modulating ion channels. Ethanol activates G protein–gated inwardly rectifying

K1 channels, but the molecular mechanism is not well understood. Here, we used a crystal structure of a mouse inward

rectifier containing a bound alcohol and structure-based mutagenesis to probe a putative alcohol-binding pocket located in

the cytoplasmic domains of GIRK channels. Substitutions with bulkier side-chains in the alcohol-binding pocket reduced or

eliminated activation by alcohols. By contrast, alcohols inhibited constitutively open channels, such as IRK1 or GIRK2 engineered

to strongly bind PIP2. Mutations in the hydrophobic alcohol-binding pocket of these channels had no effect on alcohol-dependent

inhibition, suggesting an alternate site is involved in inhibition. Comparison of high-resolution structures of inwardly rectifying

K1 channels suggests a model for activation of GIRK channels using this hydrophobic alcohol-binding pocket. These results

provide a tool for developing therapeutic compounds that could mitigate the effects of alcohol.

Many ligand-gated ion channels, such as those gated by GABA, NMDA,glycine, acetylcholine and serotonin, are responsive to ethanol and otheralcohols1–4. Initially, alcohol was hypothesized to indirectly alter thefunction of channels by changing the fluidity of the lipid bilayer5. Morerecent studies, however, suggest that alcohol acts directly through aphysical binding pocket located in the channel protein1,6. In addition toligand-gated channels, alcohols also modulate potassium channels7–9.For example, ethanol activates G protein–gated inwardly rectifyingpotassium (GIRK or Kir3) channels7,8. Behavioral studies have shownthat mice lacking GIRK2 channels have diminished ethanol-dependentanalgesia10 and consume more ethanol than wild-type mice11, suggest-ing a functional role for GIRK channels in response to alcohols in vivo.

GIRK channels are also activated following stimulation of G protein–coupled receptors (GPCRs) such as m2 muscarinic receptors (m2Rs).The mechanism of G protein activation has been extensively studied.Agonist binding to the GPCR leads to activation of the pertussis toxin–sensitive G protein heterotrimer (Gabg), allowing the Gbg subunits toassociate directly with the channel and induce channel activation12,13.Mutagenesis and chimeric studies have identified several regions in thecytoplasmic domains of GIRK channels that are involved in Gbg bindingand activation14–19. Notably, pertussis toxin treatment, which preventsGPCR-mediated G protein activation of GIRK channels, does not preventalcohol activation8. These experiments suggest that alcohol activationoccurs through a mechanism that is distinct from G protein activation.

Similar to GABA-gated ion channels, a physical pocket in the channelwith a defined cutoff is postulated to mediate alcohol activation of GIRKchannels. Alcohols with a carbon chain length of up to four carbons (thatis, methanol, ethanol, 1-propanol and 1-butanol) activate GIRK1/2

heteromeric channels, whereas longer alcohols inhibit the channels7,8.This cutoff effect suggests that there are physical constraints, possiblylinked to the length or hydrophobicity of the alcohol, that determine thesensitivity to alcohol modulation1,8. However, the molecular mechanismunderlying alcohol activation of GIRK channels is not known. Mutagen-esis studies of GIRK2 channels have implicated the distal C-terminalcytoplasmic domain in activation by alcohol7,20, but these studies did notreveal a physical alcohol-binding pocket in the channel.

Recently, we described a high-resolution structure of the cytoplasmicdomains of a G protein–insensitive inwardly rectifying potassiumchannel (IRK1 or Kir2.1) that contained bound alcohols21. The alcohol,2-methyl-2,4-pentanediol (MPD), is bound to four similar solvent-accessible hydrophobic pockets, each formed by two adjacent subunitsof the tetramer. This IRK1-bound pocket has features that are similar tothe structure of an odorant alcohol-binding protein, LUSH, that wascrystallized with ethanol22. In both structures, the alcohol pocket isformed by hydrophobic amino acids and hydrogen-bonding polargroups. Thus, the hydrophobic alcohol-bound pocket in IRK1 is aputative site for modulation by alcohols. Because the crystal structureof the cytoplasmic domain of GIRK1 or GIRK2 channels is very similarto that of IRK1 (refs. 23–25), we hypothesize that GIRK channels alsopossess cytoplasmic hydrophobic alcohol-binding pockets that areinvolved in alcohol-dependent activation.

RESULTS

Conservation of MPD-bound pocket in IRK1 and GIRK2

Recently, we found that a high-resolution structure of the IRK1cytoplasmic domains contains bound alcohols (which we refer to

Received 7 April; accepted 3 June; published online 28 June 2009; doi:10.1038/nn.2358

1Peptide Biology Laboratory, The Salk Institute for Biological Studies, La Jolla, California, USA. 2Graduate Program in Biology, Division of Biology, University of California,San Diego, La Jolla, California, USA. 3Structural Biology Laboratory, The Salk Institute for Biological Studies, La Jolla, California, USA. Correspondence should be addressedto P.A.S. (slesinger@salk.edu) or S.C. (choe@salk.edu).

988 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

here as IRK1-MPD)21. The alcohol-binding pocket in the IRK1-MPDcomplex is formed by hydrophobic amino acid side chains from threedifferent domains: the N terminus, the bD-bE ribbon and the bL-bMribbon (Fig. 1a,b)21. There are seven amino acids that interact withMPD21. Of these, the hydrophobic side chains of F47, L232, L245 andL330, and Y337 point toward the pocket. In addition to the hydro-phobic environment of the pocket, hydrogen bonds may form betweenone of the hydroxyl groups of MPD and a hydrogen bonding trianglebetween the backbone carbonyl of P244 and the hydroxyl group ofY242 via a water and between the second hydroxyl group of MPD andthe hydroxyl group of Y337 (ref. 21).

We compared the high-resolution structure of IRK1 with that ofGIRK2 (ref. 25) and identified the putative alcohol-binding pocket in

GIRK2. The hydrophobic pocket in GIRK2 has substantial conserva-tion with the amino acids that line the pocket in IRK1. The hydro-phobic pocket in GIRK2 is large enough to accommodate MPD, similarto IRK1 (Fig. 1c,d).

MPD activates GIRK2 channels similar to primary alcohols

To test whether the hydrophobic pocket in GIRK2 is the site foralcohol-mediated activation, we first investigated whether GIRK2channels are sensitive to MPD modulation. Although primary alcoholsup to the size of butanol (1-butanol, four carbons) activate GIRK1/2channels7,8, the effect of MPD with five backbone carbons is unknown.GIRK2 channels expressed in HEK-293T cells produced a smallinwardly rectifying basal K+ current that was inhibited by extracellularBa2+ (Fig. 1e). Bath application of MPD (100 mM) increased theamplitude of the inwardly rectifying current (Fig. 1e), indicating thatMPD activates GIRK channels. In addition, MPD appeared to inhibitan endogenous voltage-gated outward current at positive potentials(Fig. 1e), which is probably a voltage-gated K+ channel9. All threealcohols activated GIRK2 channels at 10 mM and showed a steepincrease in activation around 100 mM (Fig. 2a). The activation curvefor MPD fell between that of ethanol and 1-propanol (Fig. 2b) and didnot reach a maximum, similar to the findings of previous studies7,8.

Figure 1 A conserved alcohol-binding pocket

in IRK1 and GIRK2 channels. (a) Space-filling

model of the cytoplasmic domains from two

subunits of IRK1 in complex with an alcohol,

MPD. The pocket for MPD consists of three

structural elements: the N-terminal domain

(blue), the bL-bM ribbon (orange) from one

subunit and the bD-bE ribbon (green) from anadjacent subunit. Inset, schematic of IRK1 (red)

showing the major structural elements of the

subunit, including the pore loop and helix,

two transmembrane domains, and the N- and

C-terminals used in the structure (dashed box).

(b,c) Detailed structural views of amino acids

forming the hydrophobic alcohol pocket of IRK1

with MPD (b) and a putative hydrophobic alcohol

pocket in GIRK2 (c). Amino acid residues shown

in stick format are colored according to the

domain they originate from; MPD is shown in

ball-and-stick format. The putative position of

MPD in GIRK2 (dashed circle) was obtained by

superposition of two adjacent cytoplasmic domains

from IRK1 structure and corresponding subunits from GIRK2 structure. (d) Sequence alignment for the three domains comprising the hydrophobic alcohol pocket

in IRK1 and GIRK2 channels. Boxes indicate amino acids that form hydrophobic and hydrogen-bond interactions in IRK1-MPD and are conserved in GIRK2. HG

in the N-terminal domain of IRK1 indicates the polypeptide linker in the IRK1-MPD structure. (e) Current-voltage plots for GIRK2 channels recorded in the

presence of 20K solution (blue), (see Online Methods), 20K and 1 mM Ba2+ (black) or 20K and 100 mM MPD (red). Currents were elicited by voltage ramps

from �100 mV to +50 mV. MPD-induced current was 246% ± 27% (n ¼ 5, mean and s.e.m.) of basal K+ current (Ba2+ sensitive).

aIRK1 IRK1

GIRK2

V (mV)

0.5

–25

–0.5

–1.0 I (nA

)

25 50Ba2+

Basal

MPD

N terminus

Y337 Y58 Y349

L342P256

F254

I244L257

L330P244

Y242

L232L245

H42 F47 βL-βM

βD-βE

GIRK2

N--C MPD

b c

d e

–100

IRK1 42 5667

245257

341353

N-terminal domain

βD-βE domain

βL-βM domain

55

230242

326338

GIRK2

IRK1GIRK2

IRK1GIRK2

a

c

d

e

bGIRK2

0.1MPD

20 s

m2R + GIRK2

m2R + GIRK2 + m-Phos

Ethanol

Ethanol

MPD

MPD

1-propanol

1-propanol

Carb

0.5

nA

20 s

0.5

nA

20 s

Carb

0.5

nA

Ba2+

40

30

20

10Fol

d in

crea

se

00.1 1 10

Alcohol (mM)

m-Phos

Ethanol

1-propanol

MPD

Carb

0 0

200

Alc

ohol

res

pons

e (%

)

Carbachol response (%

)

400

600

800

1,000

100

300

500

– + – + – + – +

*

100 1,000

1-propanol

MPD

Ethanol

10 2550

100150

200250300 mM

Figure 2 MPD activates GIRK2 in a manner similar to other alcohols.

(a) The inward current through GIRK2 channels plotted as a function of time

(at –100 mV) shows the response to the increasing concentrations of MPD

and to 1 mM Ba2+. Dashed line shows zero current level. (b) Dose-response

curves are shown for MPD (n ¼ 6), 1-propanol (n ¼ 6) and ethanol (n ¼ 6).

The fold-increase was calculated by normalizing to the basal K+ current

(Ba2+ sensitive). (c,d) Chelating Gbg with m-Phos attenuated m2R-mediated,

but not alcohol-mediated, activation of GIRK2. Current responses recorded

at –100 mV are shown for m2R and GIRK2 (c) or m2R, GIRK2 and m-Phos

(d) in response to 100 mM 1-propanol, 100 mM MPD, 100 mM ethanol or

5 mM carbachol (Carb). (e) Bar graphs show the mean percentage alcohol

and carbachol responses (± s.e.m.), normalized to the Ba2+-sensitive basal

current, in the absence (solid, n ¼ 4) or presence of m-Phos (gray, n ¼ 7).

Asterisk indicates statistical significant difference from wild type (P o 0.05).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 989

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

These results indicate that MPD activates GIRK2 in a similar manner toother small n-alcohols. Notably, 1-pentanol, which has five backbonecarbons, similar to MPD, predominantly inhibited GIRK2 channels(Supplementary Fig. 1). Therefore, a large diol, such as MPD, activatesGIRK2 channels in a similar manner to small primary alcohols, such asethanol, but is different from 1-pentanol (see Discussion).

Pertussis toxin treatment does not prevent ethanol activation of GIRKchannels, indicating that GPCR coupling to G proteins is not involved8.To rule out the possibility that alcohols activate GIRK channels bydirectly stimulating G protein heterotrimers and liberating Gbg subunits,we measured the alcohol response of GIRK2 channels in cells coexpres-sing a myristoylated form of phosducin (m-Phos) that chelates Gbgfollowing stimulation of GPCRs26. Compared with controls, carbacholapplication led to smaller and rapidly desensitizing m2R-evoked GIRK2

currents in cells coexpressing m-Phos (Fig. 2c–e). All three alcohols, onthe other hand, activated GIRK2 channels to the same extent in thepresence of m-Phos (Fig. 2d,e). Thus, alcohol-dependent activation ofGIRK2 channels does not appear to require free Gbg subunits. Together,these results support the interpretation that alcohols directly activateGIRK channels through a physical alcohol-bound pocket.

Role for hydrophobic pocket in alcohol activation

To determine whether the hydrophobic pocket in GIRK2 mediatesalcohol activation, we examined the effects of the side-chain volumeby substituting an amino acid with a small (alanine) or large (trypto-phan) side chain (Fig. 3a,b). Six mutants did not express basal K+

currents (o–1 pA pF–1; Fig. 3b). In mutant channels engineered withan extracellular hemagglutinin (HA) tag, we investigated whether thelack of basal current was the result of a trafficking defect using confocalmicroscopy. Four mutant channels, HA-GIRK2-Y58W, HA-GIRK2-Y58A, HA-GIRK2-L342W and HA-GIRK2-Y349A, were expressed onthe plasma membrane, but did not conduct currents (Fig. 3b andSupplementary Fig. 2). Mutations at GIRK2-I244 impaired expressionon the membrane surface (Fig. 3b). These findings suggest the hydro-phobic pocket in GIRK2 is important for channel gating and/or assemblyin the absence of alcohol. Four other mutants, GIRK2-L257A, GIRK2-L257W, GIRK2-L342A and GIRK2-Y349W, produced functional chan-nels that were activated by ethanol (Fig. 3c). However, GIRK2-L257Wshowed significantly smaller ethanol-activated currents (P o 0.05 versuswild type; Fig. 3c), suggesting that the leucine at position 257 in thebD-bE ribbon is an important residue that is required for alcohol-dependent activation of GIRK2 channels.

a b

c

GIRK2

GIRK2N terminus

βL-βM

HA-GIRK2

MutationBas

al

curre

nt

Surfac

e

n

Y58 Ala

Trp

Trp

Trp

Trp

Trp

Ala

Ala

Ala

Ala

I244

L257

L342

Y349

L257Wild-typeGIRK2

0

50

100

150

200

Eth

anol

res

pons

e (%

)

A W A W A W

L342 Y349

βD-βE

Y349Y58

I244L257

L342

P256

++

++

++

+

+

+

+

+

+

0

0

0

0

0

0

00

*

+

+

+

–

–

++

34

7

7

7

7

7

7

9

9

10

10

Figure 3 Alanine/tryptophan scan of the hydrophobic alcohol-binding pocket

in GIRK2. (a) Ribbon structure shows amino acids that line the hydrophobic

alcohol pocket in GIRK2. (b) Summary table of alanine (Ala)/tryptophan (Trp)

mutagenesis. Basal K+ currents (Ba2+ sensitive) were divided into three

groups: o–1 pA pF–1 (0), –1 to –5 pA pF–1 (+) and 4–5 pA pF–1 (++)

(n ¼ number of recordings). Surface expression on the plasma membrane

was assessed in separate experiments with HA-tagged channels; detected on

the surface (+) or detected only in cytoplasm (–) (see Supplementary Fig. 2).Schematic shows the location of the HA tag in GIRK2 (gray). (c) Bar graph

shows the mean ethanol percentage response, normalized to the basal

K+ current, for different mutant channels (± s.e.m.). L257W showed a

significant statistical decrease in ethanol response (*P o 0.05 versus

wild type).

a b

c d e

–25

**

**

* * * *

*

*

–20

–15

–10

Bas

al c

urre

nt (

pA p

F–1

)

–5

0G

GIRK2-L257A

GIRK2-L257 (wt)

Ba2+

Ba2+

Ba2+ Ba2+

Carb

Carb

Carb Carb

1-propanol

1-propanol

1-propanol

1-propanol

20 s

0.5

nA

20 s

0.1

nA

20 s 20 s

0.1

nA

0.1

nA

MPD

MPD

MPDMPD

Ethanol

Ethanol

EthanolEthanol

GIRK2-L257Y GIRK2-L257W

A S C D N I L K M F Y W

Figure 4 Comprehensive mutagenesis of GIRK2-

L257 in the bD-bE ribbon of hydrophobic alcohol-

binding pocket reveals changes in alcohol- and

Gbg-activated currents. (a) Bar graph shows the

mean (± s.e.m.) amplitude of basal K+ current

(Ba2+ sensitive) for substitutions of increasing

molecular side-chain volume at GIRK2-L257:

glycine (n ¼ 7, G), alanine (n ¼ 9, A), serine

(n ¼ 7, S), cysteine (n ¼ 8, C), aspartic acid

(n ¼ 7, D), asparagine (n ¼ 6, N), isoleucine

(n ¼ 7, I), leucine (wt; n ¼ 34, L, gray bar), lysine(n ¼ 7, K), methionine (n ¼ 8, M), phenylalanine

(n ¼ 7, F), tyrosine (n ¼ 9, Y) and tryptophan

(n ¼ 9, W). Asterisks indicate statistical

difference (P o 0.05 versus leucine).

(b–e) Inward K+ currents for wild-type GIRK2 (b)

and the indicated GIRK2-L257 mutants (c–e) in

response to 100 mM 1-propanol, 100 mM MPD,

100 mM ethanol, 5 mM carbachol or 1 mM Ba2+.

Inset shows the approximate position of the

C-terminal mutation.

990 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

In the IRK1-MPD structure, L245, which is homologous to L257, ispositioned at the base of the pocket and interacts intimately with MPD.The decrease in ethanol-activated current of GIRK2-L257W raised thepossibility that amino acids with bulky side chains might generallyinterfere with alcohol activation. We systematically evaluated the effectof substituting 12 different amino acids of increasing molecular side-chain volume in GIRK2-L257. Of the 12 mutant channels, 5 wereexpressed (4–1 pA pF–1) and could be examined for possible changesin alcohol-mediated activation (Fig. 4a). The magnitude and rankorder (1-propanol 4 MPD 4 ethanol) for alcohol activation withsmaller molecular volume substitutions, such as alanine, cysteine andmethionine, were indistinguishable from wild-type leucine in GIRK2channels (Fig. 4b,c). On the other hand, GIRK2-L257Y reduced1-propanol and MPD, but not ethanol, activation, whereas GIRK2-L257Waffected ethanol-, 1-propanol– and MPD-dependent activation(Figs. 4d,e and 5a). Notably, 100 mM MPD no longer activated andinstead inhibited the basal currents for GIRK2-L257W (Figs. 4e and5a). For GIRK2-L257Y and GIRK2-L257W, the decrease in alcoholactivation was observed at a full range of concentrations (25, 125 and250 mM) for 1-propanol or MPD (Fig. 5b,c), indicating a substantialimpairment in alcohol sensitivity. In addition to the change inalcohol response, mutations at L257 also reduced the m2R-mediatedcurrents (Figs. 4c–e and 5a), indicating that L257 is involved in both

alcohol-mediated and Gbg-mediated activation (see Discussion).Taken together, these results demonstrate that increasing the side-chain volume at L257 leads to a progressive loss in alcohol-mediatedactivation (Fig. 5a). A switch in alcohol activation occurred with anincrease in volume from leucine (101 A3) in the wild type to tyrosine(133 A3) or tryptophan (168 A3). In addition, bulky substitutions atL257 affected larger alcohols (MPD) more than smaller alcohols(ethanol), suggesting the molecular volume of the pocket is animportant determinant of alcohol specificity.

Because alcohol activation is a property of most types of GIRKchannels7,8, we reasoned that a homologous mutation in a relatedGIRK channel would also alter the response to alcohols. To test thisidea, we investigated the effects of mutating L252 in GIRK4*. GIRK4*contains a mutation in the pore helix (S143T) that enhances channelactivity without affecting surface expression27. Substituting alanine(26 A3), tyrosine (133 A3) or tryptophan (168 A3) at L252 in GIRK4*channels did not change the basal K+ currents (Fig. 6a). Similar tomutations of L257 in GIRK2, tryptophan and tyrosine substitutions inGIRK4* decreased ethanol, 1-propanol and MPD activation, as com-pared with L252A, with 1-propanol now inhibiting GIRK4*-L252W(Fig. 6b–f). In contrast with GIRK2, however, MPD activationof GIRK4*-L252W was not significantly different from wild type(P 4 0.05; Fig. 6e,f). Mutating GIRK4*-L252 also significantly reducedm2R-activated GIRK currents (P o 0.05; Fig. 6c–f). Thus, the putative

a

b c

–100 Ethanol 1-propanol MPD Carb

0

200 Carbachol response (%

)

400

600

800

1,000

100

300

Alc

ohol

res

pons

e (%

)

500

700

A C LGIRK2-257

GIRK2-257

Leu

Tyr

Trp

Leu

TyrTrp

1-propanol (mM) MPD (mM)

Fol

d re

spon

se

0

10 100 1,000 10 100 1,000

2468

1012

Fol

d re

spon

se

02468

1012

M Y W0

100Å3

0

* *

*

*

* *

**

**

Figure 5 Reduced alcohol activation with increasing bulkiness of amino acid

substitutions at GIRK2-L257. (a) Bar graph shows the mean percentage

response to different alcohols and carbachol (± s.e.m.), normalized to the

basal K+ current (Ba2+ sensitive). Upward response indicates inhibition.

Amino acid substitutions are arranged by increasing side-chain volume

(A3, see inset). Asterisks indicate significant statistical difference (P o 0.05

versus leucine). (b,c) Dose-response curves are shown for GIRK2-L257,

GIRK2-L257Y and GIRK2-L257W channels for 1-propanol (b) and MPD (c).Note the suppression of alcohol activation over a range of concentrations.

Leu, leucine; Trp, tryptophan; Tyr, tyrosine.

–250

a

dc

e f

b

Bas

al c

urre

nt (

pA p

F–1

)

–200

–150

–100

GlRK4*-L252A

GlRK4*-L252W

GlRK4*-L252Y

GlRK4*-L252 (wt)

A L Y W20 s

1-propanolMPD

Ethanol

Carb

Ba2+

1 nA

20 s

1-propanol

MPD

Ethanol

Carb

Ba2+

1 nA

20 s

1-propanol

MPD

EthanolCarb

Ba2+

1 nA

20 s

1-propanol

1-propanol

MPD

MPDEthanol

Ethanol

* *

*

**

**

*

*

Carb

Carb

0

100

Carbachol response (%

)

200

Ba2+

1 nA

Alc

ohol

res

pons

e (%

)

–50

–50

100

200

GIRK4*-252

Å3

1000

A L YW

0

0

Figure 6 Mutations in the hydrophobic alcohol-binding pocket of GIRK4*

alter alcohol-activated currents. (a) Mean basal K+ currents (Ba2+ sensitive)

measured for alanine (n ¼ 8), leucine (wt; gray bar, n ¼ 8), tyrosine

(n ¼ 8) and tryptophan (n ¼ 8) substitutions at GIRK4*-L252. There were no

statistical differences in basal currents (P 4 0.05 versus Leu). (b–e) Inward

K+ currents for GIRK4* (b) and different GIRK4*-L252 mutants (c–e) in

response to 100 mM 1-propanol, 100 mM MPD, 100 mM ethanol, 5 mMcarbachol or 1 mM Ba2+. (f) Bar graphs show the mean percentage responses

to different alcohols and carbachol, normalized to the basal K+ current (Ba2+

sensitive). Amino acid substitutions are arranged by increasing side-chain

volume (A3) (see inset). Asterisks indicate significant difference (P o 0.05

versus Leu). Channel schematics show the approximate position of the

pore-helix (white ellipse) mutation for making GIRK4* and the C-terminal

mutation (black circle). All values are mean ± s.e.m.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 991

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

hydrophobic alcohol-binding pocket in GIRK4* is important formediating alcohol activation, but GIRK4* may accommodate MPDdifferently than GIRK2 (see Discussion).

Mutations of MPD pocket do not alter alcohol inhibition

MPD is bound to a hydrophobic pocket in the crystal structure ofIRK1, suggesting that MPD might inhibit IRK1 channels, similar toother alcohols7,8. Bath applying 100 mM MPD inhibited nearly 50%the basal inwardly rectifying K+ current through IRK1 channels(Fig. 7a). The MPD inhibition was dose dependent and had an IC50

(the concentration required to inhibit 50% of the current) of 104 ±23 mM and a Hill coefficient of 0.93 ± 0.02 (n ¼ 8; Fig. 7b). We nextinvestigated whether tryptophan substitutions in the hydrophobicalcohol-binding pocket of IRK1 altered alcohol-dependent inhibition(Fig. 7c). IRK1-F47W, IRK1-L232W, IRK1-L245W and IRK1-L330W,but not IRK1-Y337W, produced a substantial basal K+ current (datanot shown). Similar to wild-type IRK1, MPD inhibited the basalcurrents of mutant channels in a dose-dependent manner. The IC50

for MPD inhibition was indistinguishable among the different IRK1mutants (Fig. 7d). Furthermore, IRK1-L245W mutation did not alterinhibition by ethanol, 1-propanol or 1-butanol (data not shown).Thus, mutations at the hydrophobic pocket of IRK1 channels do notappear to alter the sensitivity to inhibition by alcohols.

IRK1 channels are constitutively open, producing a large basal K+

current. We speculated that alcohols might therefore inhibit GIRKchannels that are engineered to be constitutively open. We introduceda high-affinity PIP2 site that has previously been shown to producea large basal current28,29 (GIRK2-PIP2). In contrast with wild-typeGIRK2, GIRK2-PIP2 showed large basal currents, as expected(–530 ± 197 pA pF–1, n ¼ 5). Application of 100 mM MPD inhibitedthe basal K+ current of GIRK2-PIP2 (Fig. 7e). Similar to IRK1, wehypothesized that mutating L257 to tryptophan in GIRK2-PIP2

would have no effect in alcohol-mediated inhibition. Accordingly,

GIRK2-PIP2-L257W produced large basal K+ currents (–363 ±182 pA pF–1, n ¼ 6) that were inhibited by MPD, similar to GIRK2-PIP2 channels (Fig. 7f). We conclude that the hydrophobic alcohol-binding pocket in IRK1 or GIRK2 is not involved in alcohol-dependentinhibition. Furthermore, these results indicate that constitutively openinwardly rectifying K+ channels are not activated further by alcohols.

While investigating alcohol modulation of GIRK4*, we found that1-butanol activated GIRK4*, in contrast with its inhibition of wild-typeGIRK4 (Supplementary Fig. 3) or GIRK1/4 heterotetramers7,8. GIRK4and GIRK4* differ by a point mutation (S143T) in the pore helix27,suggesting that S143T may regulate sensitivity to alcohol inhibition.To assess whether this site is generally involved in alcohol-mediatedinhibition of GIRK channels, we introduced a threonine at theequivalent serine (S148T) in GIRK2-PIP2 channels. As predicted,GIRK2-PIP2-S148T significantly shifted the IC50 for MPD-dependentinhibition compared with GIRK2-PIP2 (P o 0.05; Fig. 7f). Takentogether, these experiments implicate amino acids in the pore helix inregulating the extent of alcohol-dependent inhibition and support theconclusion that the cytoplasmic alcohol-binding pocket mediatesalcohol-dependent activation, but not inhibition, of GIRK channels.

DISCUSSION

On the basis of high-resolution channel structures and functionalmutagenesis, we have identified a physical site for alcohol-mediatedactivation of GIRK channels. Amino acid substitutions that increasedthe molecular side-chain volume at a conserved leucine in the bD-bEribbon of the hydrophobic pocket of GIRK2 decreased alcohol-mediated activation of GIRK channels (Fig. 8a). In particular, twosubstitutions, tryptophan and tyrosine, at the leucine in the hydro-phobic pocket of GIRK2 (L257) and GIRK4* (L252) channels pro-duced a progressive loss in alcohol activation. For GIRK4*, alaninesubstitution increased the amplitude of alcohol-activated currents.Thus, increasing or decreasing the volume of the pocket by alteringthe amino acid side-chain produced changes in alcohol activation.Similarly, the size of the putative alcohol-binding pocket in GABAA a1and glycine receptors is important for determining modulation byalcohol and other small anesthetics. Increasing the bulkiness of aminoacids in the putative alcohol-binding pocket of these channels elimi-nates modulation by ethanol1 or isofluorane30. In contrast, decreasingthe size of amino acids in the same region of the decanol-insensitive

IRK1

–100

a b

d

f

c

e

Ba2+

MPD

Basal

Ba2+

MPD

Basal

IRK1

GIRK2-PIP2

N terminus

βL-βM

βD-βE

L232L245

L330

P244

Y337

F47

–75–25

2 100

IRK1

Cur

rent

rem

aini

ng (

%)

50

0

150

100

IC50

(m

M)

500

0

100

Cur

rent

rem

aini

ng (

%)

50

GIRK2-PIP2-L257W

GIRK2-PIP2-S148T

GIRK2-PIP2

0.1 1 10 100MPD (mM)

1,0000

IRK1

F47W

L232

W

L245

W

L330

W

Y337W

1 10 100MPD (mM)

1,000

–2

–4

–6

–8

2

–2

–4

–6

25

V (mV)

V (mV)

I (nA

)I (

nA)

50

–100 –75–25 25 50

Figure 7 Mutations in the hydrophobic alcohol-binding pocket of IRK1 have

no effect on alcohol-dependent inhibition. (a) Current-voltage plots for IRK1

channels are shown for 20K (blue), 20K and 1 mM Ba2+ (black) or 20K and

100 mM MPD (red). MPD inhibited the basal K+ current (Ba2+ sensitive)

by 53.1% ± 4.1% (n ¼ 8). (b) Dose-response curve for MPD inhibition of

IRK1 channel. The smooth curve shows the best fit using the Hill equation,

with an IC50 of 104 ± 23 mM and Hill coefficient of 0.93 ± 0.03 (n ¼ 8).

(c) Structural view of amino acids that line the hydrophobic alcohol pocketin IRK1. (d) Bar graph shows mean IC50s for MPD-dependent inhibition of

IRK1 (n ¼ 8), IRK1-F47W (n ¼ 7), IRK1-L232W (n ¼ 7), IRK1-L245W

(n ¼ 6) and IRK1-L330W (n ¼ 6). There was no statistically significant

difference compared with wild-type IRK1 (P 4 0.05). (e) Current-voltage

plots are shown for GIRK2-PIP2 (GIRK2 engineered with high-affinity PIP2

binding domain from IRK1) channels recorded in the presence of 20K (blue),

20K and 1 mM Ba2+ (black) or 20K and 100 mM MPD (red). (f) Dose-

response curves for MPD-dependent inhibition of GIRK2-PIP2 (solid circle),

GIRK2-PIP2-L257W (open circle) and GIRK2-PIP2-S148T (solid triangle).

The smooth curves show the best fit using the Hill equation and have IC50s

and Hill coefficients of 7.7 ± 1.0 mM and 0.66 ± 0.03 (n ¼ 5) for GIRK2-

PIP2, 5.2 ± 1.0 mM and 0.77 ± 0.04 (n ¼ 5) for GIRK2-PIP2-L257W,

and 147.0 ± 31.5 mM and 0.67 ± 0.05 (n ¼ 6) for GIRK2-PIP2–S148T.

All values are mean ± s.e.m.

992 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

GABA r1 receptors enables potentiation by decanol1,6. Together with ourfindings, these studies suggest that physical pockets of defined dimen-sions can be probed with mutations that change the dimension of thealcohol-binding sites. Using the IRK1-MPD structure as a guide, weestimated the volume of the hydrophobic alcohol-binding pocket ofGIRK channels to be B250 A3, which is large enough to accommodatebulky alcohols such as MPD (B130 A3; Fig. 8b). A tryptophan mutationin the pocket would decrease the volume that could potentially occludelarger alcohols (Fig. 8b). In addition, the sensitivity to activation may bedifferent between GIRK2 and GIRK4* channels. For example, tryptophansubstitution in GIRK2 eliminated MPD activation, revealing inhibition ofcurrent. In GIRK4*, tryptophan substitution eliminated 1-propanolactivation, but showed only a substantial decrease in MPD activationwhen comparing alanine with tryptophan substitutions. One possibleexplanation is that MPD fits differently in the alcohol pocket of GIRK4*,which could be revealed in a GIRK4 structure in complex with MPD.

The observations that mutations at the hydrophobic pocket didnot alter alcohol-dependent inhibition of IRK1 and that S148T muta-tion (but not L257W) in GIRK2-PIP2 decreased alcohol-dependentinhibition suggest that GIRK channels possess two different sites foralcohol modulation. In the Kirbac1.3 structure, the serine is located inthe pore helix of Kirbac1.3 where there is no space for alcohol(Supplementary Fig. 4), suggesting that S148 regulates sensitivity toinhibition, but does not form the binding site. Alcohols might interferewith ion permeation or possibly with gating at the transmembranedomains, similar to voltage-gated K channels9. Another possibility isthat alcohols inhibit the channel by altering the fluidity of the lipidmembrane5 and/or decreasing interactions with PIP2, which is requiredfor channel function31.

Notably, although 1-octanol inhibits GIRK channels, coapplicationof 1-octanol with ethanol has no effect on ethanol-mediated activation,also raising the possibility of a second site for inhibition7. The net effectof alcohol modulation in GIRK channels would therefore be deter-mined by the relative potencies of activation and inhibition. In supportof this, we found that bath application of 1-pentanol inhibited GIRK2channels, but induced a large current immediately after washout(Supplementary Fig. 1), revealing two components of alcohol

modulation. It is notable that MPD predominantly activates GIRKchannels, in contrast with large primary alcohols of similar size. Afunctional difference between diols and primary alcohols has beenreported previously for NMDA channels32. The addition of a hydroxylgroup may lead to decreased sensitivity to inhibition for GIRKchannels. The presence of two sites for alcohol modulation alsosuggests that ascribing a cutoff number for alcohol activation ofGIRK channels would not be accurate, in contrast with the determina-tion of the cutoff number for GABAA channels1,6,33.

Two different types of alcohol-bound protein structures have beensolved previously, the enzymatic/catalytic alcohol dehydrogenase and thenoncatalytic Drosophila odorant-binding protein LUSH. In alcoholdehydrogenase, primary alcohol is coordinated with Zn2+ in a hydro-phobic pocket, where it catalyzes the oxidation of alcohol to alde-hyde34,35. Mutagenesis studies in the pocket indicated that thebulkiness of the side chains in the hydrophobic pocket determines thealcohol specificity36. The high-resolution structure of LUSH in complexwith small alcohols showed that, in addition to hydrophobic interactions,a network of hydrogen bonds help to stabilize alcohol in the alcohol-binding pocket22,37. The hydrophobic pocket in IRK1-MPD has many ofthe same features of these alcohol-binding pockets. First, hydrophobicamino acids form the pocket and interact intimately with hydrocarbonsof the alcohol (Fig. 1b). In GIRK2, mutations of L257 to bulkier aminoacids substantially reduced or eliminated alcohol-mediated activation.This finding is consistent with a role for hydrophobic side chains indetermining the size of the alcohol-binding pocket. Second, the IRK1-MPD structure indicates that hydrogen-bonds form between MPD andthe backbone carbonyl of P244, Y242 via a water and a hydroxyl of Y337(Fig. 1)21. At the homologous position for IRK1-Y242, GIRK2 contains aphenylalanine (F254), which indicates that this hydrogen-bond trianglemay not be essential. In addition, we found that MPD-mediatedactivation was not affected by a Y349W mutation at the homologousposition of IRK1-Y337 (Supplementary Fig. 5). Therefore, it is possiblethat the carbonyl group of proline in the bD-bE ribbon is the linchpinthat stabilizes alcohol in the pocket via hydrogen bonding. Unnaturalamino acid mutagenesis38 would be needed to further establish theimportance of this hydrogen-bond interaction in stabilizing alcohol.

100%

a b

c

50% N terminus

Leu Leu + MPD Trp

βL-βM

βD-βE

**

Eth

anol

res

pons

e

0%

IRK1-MPDIRK1-MPD

βL-βM

N terminus

F46 L333

F338L333

Y337

F47

L245

Y242

P244

βD-βE

GIRK1 open

GIRK1 closedD

D A

A

GIRK1 closed

GIRK2-

L257

W

GIRK4*-

L252

W

IRK1-

L245

W

GIRK2-

PIP2

-L25

7W

Figure 8 Model for alcohol-dependent activation of GIRK channels. (a) Bar

graph shows the mean percentage ethanol response (activation or inhibition

normalized to wild type) for a tryptophan mutation in four different channels:

GIRK2-L257W (n ¼ 9), GIRK4-L252W (n ¼ 8), IRK1-L245W (n ¼ 8) and

GIRK2-PIP2-L257W (n ¼ 5). Mutation in the alcohol-binding pocket affected

activation (* P o 0.05 versus wild type), but not inhibition, of GIRK

channels. (b) Top, schematic of inward rectifier shows the location of the

alcohol-binding pocket in the cytoplasmic domains, two gates (G loop andM2 transmembrane, black triangles) and pore-helix region (red ellipse). PIP2

is enriched in the lower leaflet of bilayer (orange). Bottom, molecular surface

representations of the alcohol pocket without (Leu), with MPD (Leu + MPD)

and modeled with L257W (Trp), using the IRK1-MPD structure as a guide.

(c) Left, alignment of the putative closed state of GIRK1 chimeric channel

(GIRK1 closed, green) with the IRK1-MPD structure (gray). Spaghetti

structures show two adjacent cytoplasmic subunits (subunits D and A) and

the hydrophobic alcohol pocket at the cytoplasmic subunit interface. Right,

zoom shows alignment of the N-terminal domain, bD-bE and bL-bM ribbons

from the IRK1-MPD (gray), GIRK1 open (orange) and GIRK1 closed (green)

structures. IRK1-MPD aligns better with the putative open state of GIRK1.

Note the substantial displacement in the bL-bM ribbon (arrow) and the side

chains of hydrophobic amino acids in the two structures. GIRK1 closed, but

not GIRK1 open, has a collapsed alcohol-binding pocket, as a result of the

interaction and rotation of F46 (IRK1-F47), L246 (IRK1-L245) and F338

(IRK1-Y337). GIRK1-L333 in the bL-bM domain, implicated previously in

Gbg gating of GIRK channels17–19, is shown for reference.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 993

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Although specific alcohol-induced conformational changes in thechannel protein remain unknown, our structural analysis between theIRK1-MPD structure and that of the chimeric KirBac1.3-GIRK1provides some new clues into channel gating21,39. Two differentconformational states of GIRK have been described: a putative openstate, resulting from the open position of the G loops in the cytoplasmicgate24,39 (GIRK1 open; Fig. 8c), and a putative closed state (GIRK1closed; Fig. 8c). We aligned the IRK1-MPD structure with these twodifferent Kirbac1.3-GIRK1 structures and discovered that IRK1-MPDstructure aligns better with the GIRK1 open in the hydrophobicalcohol-binding pocket (Fig. 8c). In contrast, the alignment with theGIRK1 closed structure showed marked differences in the hydro-phobic alcohol pocket. In particular, the side chains from F46 in theN-terminal domain, L246 in the bD-bE ribbon and L333 in the bL-bMribbon, fill the hydrophobic pocket of the putative closed state ofGIRK1. In the open state, structural rearrangements of F46, L246, L333and F338 occur, which would enable MPD to fit in the pocket.

On the basis of our mutagenesis data and structural analyses, wepropose a tenable model for alcohol activation of GIRK channels. Atrest, GIRK2 channels undergo infrequent structural rearrangements inthe pocket that correlate with the open and closed positions of thechannel’s cytoplasmic gates, the G loops24,39 and M2 transmembranedomains40–42 (Fig. 8b,c). Alcohol entering the pocket could thenstabilize the open conformation, leading to alcohol-activated currents.Bulky substitutions at L257/L252 of GIRK channels, located at the baseof the alcohol pocket, would hinder alcohols from filling the pocket.Previous studies have shown that gating of GIRK channels requiresPIP2

28,31. Similarly, alcohol-dependent activation may increase PIP2

affinity for the channel, stabilizing an open confirmation. Thus, a PIP2-activated GIRK channel would not be activated further by alcohol, aswas observed with GIRK2-PIP2 (Fig. 7). Future studies will need toinvestigate the molecular relationships between movement of theN-terminal domain, the bD-bE and bL-bM ribbons in the hydro-phobic pocket, PIP2 interactions, and the channel gates.

The alcohol-binding pocket may also be involved in Gbg-dependentactivation. Mutation of a conserved leucine (GIRK2-L344,GIRK4*L339, GIRK1-L333; Fig. 8c) to glutamic acid in the bL-bMribbon of GIRK channels attenuates Gbg activation17–19. We found thatmutations in the bD-bE (L257) ribbon that altered alcohol-dependentactivation also reduced Gbg-dependent activation (Figs. 4 and 5).Together, these results suggest that conformational changes in thebD-bE and bL-bM structural elements, along with the N-terminaldomain43,44, may be central to both alcohol- and Gbg-dependentactivation. Notably, hydrophobic amino acids in the G protein Gbsubunit have been implicated in GIRK channel activation45, whichperhaps interact directly with the alcohol-binding pocket in GIRK.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Accession codes. We used protein structures from the ProteinDatabase, accession codes 2GIX (IRK1-MPD), 2E4F (GIRK2) and2QKS (Kirbac1.3-GIRK1).

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe would like to thank Y. Kurachi for GIRK2 coordinates, M. Lazdunsky forGIRK2 cDNA, D. Clapham for GIRK4 cDNA, N. Dascal for m-Phos cDNA,S. Pegan for initial discussions on structure of IRK1-MPD and members of theSlesinger laboratory for helpful comments. This work was funded, in part, by a

pre-doctoral National Research Service Award from the National Institute onAlcohol Abuse and Alcoholism (F31AA017042, P.A.), by the National Institute onGeneral Medical Sciences (R01GM056653, S.C.), and by the H.N. & Frances C.Berger Foundation and the Salk Institute for Biological Studies (P.A.S.). Thecontent is solely the responsibility of the authors and does not necessarilyrepresent the official views of the National Institute on Alcohol Abuse andAlcoholism or the National Institute on General Medical Sciences.

AUTHOR CONTRIBUTIONSP.A.S. and P.A. designed the experiments and analyzed the data. P.A. conductedthe molecular cloning, electrophysiology and imaging experiments. H.D. andP.A. collaborated on structural analysis and figure production. H.D. conductedmodeling experiments. P.A., H.D. and P.A.S co-wrote and revised the manuscript.P.A.S and S.C. supervised the project.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Mihic, S.J. et al. Sites of alcohol and volatile anaesthetic action on GABAA and glycinereceptors. Nature 389, 385–389 (1997).

2. Lovinger, D.M., White, G. & Weight, F.F. Ethanol inhibits NMDA-activated ion current inhippocampal neurons. Science 243, 1721–1724 (1989).

3. Cardoso, R.A. et al. Effects of ethanol on recombinant human neuronal nicotinicacetylcholine receptors expressed in Xenopus oocytes. J. Pharmacol. Exp. Ther. 289,774–780 (1999).

4. Zhou, Q. & Lovinger, D.M. Pharmacologic characteristics of potentiation of 5-HT3receptors by alcohols and diethyl ether in NCB-20 neuroblastoma cells. J. Pharmacol.Exp. Ther. 278, 732–740 (1996).

5. Harris, R.A., Trudell, J.R. & Mihic, S.J. Ethanol’s molecular targets. Sci. Signal. 1, re7(2008).

6. Wick, M.J. et al. Mutations of gamma-aminobutyric acid and glycine receptorschange alcohol cutoff: evidence for an alcohol receptor? Proc. Natl. Acad. Sci. USA95, 6504–6509 (1998).

7. Lewohl, J.M. et al. G protein–coupled inwardly rectifying potassium channels are targetsof alcohol action. Nat. Neurosci. 2, 1084–1090 (1999).

8. Kobayashi, T. et al. Ethanol opens G protein–activated inwardly rectifying K+ channels.Nat. Neurosci. 2, 1091–1097 (1999).

9. Covarrubias, M., Vyas, T.B., Escobar, L. & Wei, A. Alcohols inhibit a cloned potassiumchannel at a discrete saturable site. Insights into the molecular basis of generalanesthesia. J. Biol. Chem. 270, 19408–19416 (1995).

10. Blednov, Y.A., Stoffel, M., Alva, H. & Harris, R.A. A pervasive mechanism for analgesia:activation of GIRK2 channels. Proc. Natl. Acad. Sci. USA 100, 277–282 (2003).

11. Blednov, Y.A., Stoffel, M., Chang, S.R. & Harris, R.A. Potassium channels as targets forethanol: studies of G protein–coupled inwardly rectifying potassium channel 2 (GIRK2)null mutant mice. J. Pharmacol. Exp. Ther. 298, 521–530 (2001).

12. Reuveny, E. et al. Activation of the cloned muscarinic potassium channel by G proteinbg-subunits. Nature 370, 143–146 (1994).

13. Wickman, K.D. et al. Recombinant G protein bg-subunits activate the muscarinic-gatedatrial potassium channel. Nature 368, 255–257 (1994).

14. Huang, C.L., Slesinger, P.A., Casey, P.J., Jan, Y.N. & Jan, L.Y. Evidence that directbinding of G bg to the GIRK1 G protein–gated inwardly rectifying K+ channel is importantfor channel activation. Neuron 15, 1133–1143 (1995).

15. Kunkel, M.T. & Peralta, E.G. Identification of domains conferring G protein regulation oninward rectifier potassium channels. Cell 83, 443–449 (1995).

16. Krapivinsky, G. bg binding to GIRK4 subunit is critical for G protein–gated K+ channelactivation. J. Biol. Chem. 273, 16946–16952 (1998).

17. He, C., Zhang, H., Mirshahi, T. & Logothetis, D.E. Identification of a potassium channelsite that interacts with G protein bg subunits to mediate agonist-induced signaling.J. Biol. Chem. 274, 12517–12524 (1999).

18. Ivanina, T. et al. Mapping the Gbg-binding sites in GIRK1 and GIRK2 subunits of theG protein–activated K+ channel. J. Biol. Chem. 278, 29174–29183 (2003).

19. Finley, M., Arrabit, C., Fowler, C., Suen, K.F. & Slesinger, P.A. bL-bM loop in theC-terminal domain of G protein–activated inwardly rectifying K+ channels is importantfor G(bg) subunit activation. J. Physiol. (Lond.) 555, 643–657 (2004).

20. Hara, K., Lewohl, J.M., Yamakura, T. & Harris, R.A. Mutational analysis of ethanolinteractions with G protein–coupled inwardly rectifying potassium channels. Alcohol 24,5–8 (2001).

21. Pegan, S., Arrabit, C., Slesinger, P.A. & Choe, S. Andersen’s syndrome mutation effectson the structure and assembly of the cytoplasmic domains of Kir2.1. Biochemistry45, 8599–8606 (2006).

22. Kruse, S.W., Zhao, R., Smith, D.P. & Jones, D.N. Structure of a specific alcohol-bindingsite defined by the odorant binding protein LUSH from Drosophila melanogaster.Nat. Struct. Biol. 10, 694–700 (2003).

23. Nishida, M. & MacKinnon, R. Structural basis of inward rectification: cytoplasmic poreof the G protein–gated inward rectifier GIRK1 at 1.8 A resolution. Cell 111, 957–965(2002).

24. Pegan, S. et al. Cytoplasmic domain structures of Kir2.1 and Kir3.1 show sites formodulating gating and rectification. Nat. Neurosci. 8, 279–287 (2005).

994 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

25. Inanobe, A., Matsuura, T., Nakagawa, A. & Kurachi, Y. Structural diversity in thecytoplasmic region of G protein–gated inward rectifier K+ channels. Channels (Austin)1, 39–45 (2007).

26. Rishal, I., Porozov, Y., Yakubovich, D., Varon, D. & Dascal, N. Gbg-dependent andGbg -independent basal activity of G protein–activated K+ channels. J. Biol. Chem. 280,16685–16694 (2005).

27. Vivaudou, M. et al. Probing the G-protein regulation of GIRK1 and GIRK4, the twosubunits of the KACh channel, using functional homomeric mutants. J. Biol. Chem.272, 31553–31560 (1997).

28. Zhang, H., He, C., Yan, X., Mirshahi, T. & Logothetis, D.E. Activation of inwardlyrectifying K+ channels by distinct PtdIns(4,5)P2 interactions. Nat. Cell Biol. 1,183–188 (1999).

29. Zhou, W., Arrabit, C., Choe, S. & Slesinger, P.A. Mechanism underlying bupivacaineinhibition of G protein–gated inwardly rectifying K+ channels. Proc. Natl. Acad. Sci. USA98, 6482–6487 (2001).

30. Jenkins, A. et al. Evidence for a common binding cavity for three general anestheticswithin the GABAA receptor. J. Neurosci. 21, RC136 (2001).

31. Huang, C.L., Feng, S. & Hilgemann, D.W. Direct activation of inward rectifierpotassium channels by PIP2 and its stabilization by Gbg. Nature 391, 803–806 (1998).

32. Peoples, R.W. & Ren, H. Inhibition of N-methyl-d-aspartate receptors by straight-chaindiols: implications for the mechanism of the alcohol cutoff effect. Mol. Pharmacol. 61,169–176 (2002).

33. Dildy-Mayfield, J.E., Mihic, S.J., Liu, Y., Deitrich, R.A. & Harris, R.A. Actions oflong chain alcohols on GABAA and glutamate receptors: relation to in vivo effects.Br. J. Pharmacol. 118, 378–384 (1996).

34. Ramaswamy, S., Eklund, H. & Plapp, B.V. Structures of horse liver alcohol dehydro-genase complexed with NAD+ and substituted benzyl alcohols. Biochemistry 33,5230–5237 (1994).

35. Svensson, S., Hoog, J.O., Schneider, G. & Sandalova, T. Crystal structures of mouseclass II alcohol dehydrogenase reveal determinants of substrate specificity and catalyticefficiency. J. Mol. Biol. 302, 441–453 (2000).

36. Weinhold, E.G. & Benner, S.A. Engineering yeast alcohol dehydrogenase.Replacing Trp54 by Leu broadens substrate specificity. Protein Eng. 8, 457–461(1995).

37. Thode, A.B., Kruse, S.W., Nix, J.C. & Jones, D.N. The role of multiple hydrogen-bondinggroups in specific alcohol binding sites in proteins: insights from structural studies ofLUSH. J. Mol. Biol. 376, 1360–1376 (2008).

38. Lu, T. et al. Probing ion permeation and gating in a K+ channel with backbone mutationsin the selectivity filter. Nat. Neurosci. 4, 239–246 (2001).

39. Nishida, M., Cadene, M., Chait, B.T. & MacKinnon, R. Crystal structure of aKir3.1-prokaryotic Kir channel chimera. EMBO J. 26, 4005–4015 (2007).

40. Yi, B.A., Lin, Y.F., Jan, Y.N. & Jan, L.Y. Yeast screen for constitutivelyactive mutant G protein–activated potassium channels. Neuron 29, 657–667(2001).

41. Sadja, R., Smadja, K., Alagem, N. & Reuveny, E. Coupling Gbg -dependent activation tochannel opening via pore elements in inwardly rectifying potassium channels. Neuron29, 669–680 (2001).

42. Jin, T. et al. The bg subunits of G proteins gate a K+ channel by pivoted bending of atransmembrane segment. Mol. Cell 10, 469–481 (2002).

43. Sarac, R. et al. Mutation of critical GIRK subunit residues disrupts N- and C-terminiassociation and channel function. J. Neurosci. 25, 1836–1846 (2005).

44. Riven, I., Kalmanzon, E., Segev, L. & Reuveny, E. Conformational rearrangementsassociated with the gating of the G protein–coupled potassium channel revealed byFRET microscopy. Neuron 38, 225–235 (2003).

45. Ford, C.E. et al. Molecular basis for interactions of G protein bg subunits with effectors.Science 280, 1271–1274 (1998).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 995

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSMolecular biology and cell culture. cDNAs for mouse GIRK2c46 (GIRK2c is

referred to as GIRK2 in this study for clarity), rat GIRK4 (ref. 16), mouse IRK1

(ref. 24)24, human m2R46 and bovine m-Phos26 were subcloned into pcDNA3.1

vector (Invitrogen). GIRK4* contains a S143T mutation in the pore-helix27.

Point mutations were introduced by a Quickchange site-directed mutagenesis

kit (Stratagene). GIRK2-PIP2 mutant was created by the overlap-PCR meth-

od47. Briefly, a region of GIRK2-D228-L234 was replaced with the homologous

region of IRK1-N216-L222; this region contains seven amino acids in the bC-

bD region that have previously been implicated in PIP2 binding28,29 (we refer to

this mutant as GIRK2-PIP2). All mutations were confirmed by DNA sequen-

cing. HEK-293T cells were cultured in DMEM supplemented with 10% fetal

bovine serum (vol/vol) and 1� Glutamax (Invitrogen) in a humidified 37 1C

incubator with 5% CO2. Cells were plated in 12-well dish and transiently

transfected with DNA using Lipofectamine 2000 (Invitrogen). Cells were

replated to 12-mm glass coverslips coated with poly-D-lysine (20 mg ml–1)

12–24 h after transfection.

Detection of channels expressed on membrane surface. GIRK2c and mutant

channels were engineered with an extracellular HA epitope inserted between

I126 and E127 for immunohistochemical detection with antibodies to HA46.

HEK-293T cells were transfected with 0.5 mg of channel cDNA and examined

24–48 h after transfection. Cells were washed with 1� Dulbecco’s phosphate

buffer saline (DPBS, Invitrogen), fixed with 2% paraformaldehyde (wt/vol) in

1� DPBS for 10 min and rinsed with 1� DPBS (at 22–25 1C). To label surface

channels, we incubated cells with blocking buffer (3% BSA in 1� DPBS) for 1 h

and then with mouse antibody to HA (1:400 in blocking buffer, Covance) for

2 h at 22 1C. To label cytoplasmic channels, cells were rinsed with 1� DPBS,

permeabilized with 0.25% Triton X-100 (Sigma) in blocking buffer for 10 min

at 22 1C and incubated with blocking buffer for 1 h. Cells were then incubated

with rabbit antibody to GIRK2 (1:200 in blocking buffer, Alomone) for 2 h.

Following rinses in 1� DPBS, cells were incubated with fluorescent secondary

antibodies, antibody to mouse Alexa-647 and antibody to rabbit Alexa-488

(1:300, Invitrogen), for 1 h in the dark. Cells were rinsed with 1� DPBS,

mounted on microscope slides using Progold anti-fading reagent (Invitrogen)

and both fluorophores were imaged with a Leica TSC SP2 AOBS laser

confocal microscope.

Whole-cell patch-clamp electrophysiology. HEK-293T cells were transfected

with 0.2 mg of channel cDNA and 0.04 mg of enhanced yellow fluorescent

protein cDNA to identify transfected cells. For some experiments, 0.8 mg of

m2R and 0.8 mg of m-Phos cDNA were also transfected. Whole-cell patch-

clamp recordings were performed 24–72 h after transfection. Borosilicate glass

electrodes (Warner Instruments) of 5–7 mO were filled with intracellular

solution (130 mM KCl, 20 mM NaCl, 5 mM EGTA, 2.56 mM K2ATP, 5.46 mM

MgCl2, 0.30 mM Li2GTP and 10 mM HEPES, pH 7.4). Extracellular 20K

solution contained 20 mM KCl, 140 mM NaCl, 0.5 mM CaCl2, 2 mM MgCl2and 10 mM HEPES (pH 7.4). Alcohols (0.1–300 mM), carbachol (5 mM) or

BaCl2 (1 mM) were diluted into the 20K solution and applied directly to the

cell with a rapid, valve-controlled perfusion system (Warner Instruments, VC6,

MM-6 manifold). All of the chemicals that we used for our electrophysiology

experiments were purchased from Sigma-Aldrich. Whole-cell patch-clamp

currents were recorded using an Axopatch 200B (Molecular Devices, Axon

Instruments) amplifier. Currents were adjusted electronically for cell capaci-

tance and series resistance (80–100%), filtered at 1 kHz with an 8-pole Bessel

filter and digitized at 5 kHz with a Digidata 1200 interface (Molecular Devices,

Axon Instruments). Currents were elicited with voltage ramp protocol, from

–100 mV to +50 mV, delivered at 0.5 Hz. Currents were measured at –100 mV

and converted to current density (pA pF–1) by dividing with the membrane

capacitance. Basal K+ currents (Ba2+ sensitive) were quantified at –100 mV by

applying 1 mM BaCl2 in 20K and measuring the amplitude of the Ba2+-

inhibited current. Alcohol- and carbachol-modulated currents were measured

at �100 mV by averaging current from two consecutive sweeps on reaching

steady state and subtracting the mean basal current before and after the

application of the modulator. Pooled data are presented as mean ± s.e.m.

and evaluated for statistical significance (P o 0.05) using a one-way ANOVA,

followed by Bonferroni multiple comparison post hoc test. To determine the

IC50, we fitted the inhibition curves with the Hill equation (y ¼ 1

1+ðxcÞ

b), where

y is fraction of current remaining, x is the concentration of alcohol, b is the

Hill coefficient and c is the IC50 (the concentration of alcohol that produces

50% inhibition).

Structural analysis. Molecular representations were made using PyMOL

(DeLano Scientific) with PDB files 2GIX (IRK1-MPD), 2E4F (GIRK2) and

2QKS (Kirbac1.3-GIRK1). The cavity of the IRK1-MPD pocket was calculated

using CASTp server48 with a 1.4-A probe radius. L245W was modeled in the

IRK1-MPD structure by optimizing the best rotamer position for Trp using

PyMOL software. Molecular volume estimates for amino acid side-chains were

based on reported values49.

46. Clancy, S.M. et al. Pertussis toxin–sensitive Galpha subunits selectively bind toC-terminal domain of neuronal GIRK channels: evidence for a heterotrimericG protein–channel complex. Mol. Cell. Neurosci. 28, 375–389 (2005).

47. Ho, S.N., Hunt, H.D., Horton, R.M., Pullen, J.K. & Pease, L.R. Site-directed mutagen-esis by overlap extension using the polymerase chain reaction. Gene 77, 51–59 (1989).

48. Dundas, J. et al. CASTp: computed atlas of surface topography of proteins with structuraland topographical mapping of functionally annotated residues. Nucleic Acids Res. 34,W116–118 (2006).

49. Harpaz, Y., Gerstein, M. & Chothia, C. Volume changes on protein folding. Structure 2,641–649 (1994).

NATURE NEUROSCIENCE doi:10.1038/nn.2358

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Distinct contributions of Nav1.6 and Nav1.2 inaction potential initiation and backpropagation

Wenqin Hu, Cuiping Tian, Tun Li, Mingpo Yang, Han Hou & Yousheng Shu

The distal end of the axon initial segment (AIS) is the preferred site for action potential initiation in cortical pyramidal neurons

because of its high Na1 channel density. However, it is not clear why action potentials are not initiated at the proximal AIS,

which has a similarly high Na1 channel density. We found that low-threshold Nav1.6 and high-threshold Nav1.2 channels

preferentially accumulate at the distal and proximal AIS, respectively, and have distinct functions in action potential initiation

and backpropagation. Patch-clamp recording from the axon cut end of pyramidal neurons in the rat prefrontal cortex revealed

a high density of Na1 current and a progressive reduction in the half-activation voltage (up to 14 mV) with increasing distance

from the soma at the AIS. Further modeling studies and simultaneous somatic and axonal recordings showed that distal Nav1.6

promotes action potential initiation, whereas proximal Nav1.2 promotes its backpropagation to the soma.

Although action potentials can be independently initiated at many sitesin a neuron1–7, pioneering studies in spinal motorneurons8–11 andrecent works in cortical pyramidal neurons6,7,12–14 have demonstratedthat the AIS has the lowest threshold for action potential initiation, upto 15 mV lower than that at the soma8–10,14, as a result of a highconcentration of Na+ channels4,15–17. Further studies in layer 5 pyra-midal neurons have shown that the action potential initiation site islocated in the distal region of the AIS, a specialized subcellular domaintargeted by certain types of K+ channels18–21 and innervated byaxo-axonic GABAergic interneurons19,22.

Because immunostaining studies and patch-clamp recordings haveindicated a similar density of Na+ channels along the length of theAIS3,23,24, it has been unclear why action potentials are preferentiallyinitiated at the distal rather than at the proximal region of the AIS. Oneexplanation could be that the biophysical property of Na+ channelsmay differ in the proximal versus distal compartments. Indeed, initialcell-attached and outside-out patch recordings have revealed a B7-mVhyperpolarizing shift in the activation curve of Na+ channels at the AISrelative to that at the soma3. Recent immunostaining studies ondifferent neuronal types have demonstrated that the Na+ channelsubtypes may be specifically targeted to different AIS regions18,24–27.Because Na+ channel subtypes have different activation thresholds28,their subcellular distribution and density at the AIS may contribute tothe action potential initiation and regulation25,29–31. Thus, a fullresolution of the role of density versus voltage-dependent propertiesof Na+ channels at the AIS requires the mapping of the distribution ofNa+ channel subtypes and direct recording from the axonal membraneat both the proximal and the distal regions of the AIS.

We carried out immunostaining for Na+ channel subtypes andpatch-clamp recordings from giant patches (somatic nucleated patches

and isolated axon blebs), as well as regular outside-out patches, toinvestigate the distribution and biophysical properties of Na+ channelsin the somatic and AIS/axonal membrane. Our results provide aquantitative description of the selective accumulation of Nav1.6 andNav1.2 channels at the distal and proximal AIS, respectively. Furthercomputational modeling and simultaneous recordings from the somaand the axon indicated that this selective accumulation of Nav1.6 at thedistal AIS determines the action potential initiation site, whereasaccumulation of Nav1.2 at the proximal AIS promotes action potentialpropagation to the soma and sets the action potential threshold of thesomatodendritic region of the neuron.

RESULTS

Differential distribution of Na+ channel subtypes at AIS

We first performed immunostaining of layer 5 pyramidal neurons ofthe rat prefrontal cortex to map the distribution of Na+ channelsubtypes and the Na+ channel–associated protein ankyrin-G (AnkG),which is known to accumulate at the AIS (Fig. 1). We found adifferential distribution of Nav1.2 and Nav1.6 at the proximal anddistal AIS, respectively (Fig. 1a,b). Quantitative measurements of theimmunofluorescence intensity for a large population of neurons in16-mm-thick cortical sections showed a polarized distribution of Nav1.2and Nav1.6 at the AIS (Fig. 1a,b). Further immunostaining of100-mm-thick sections showed a peak in the density of Nav1.6 channelsbetween 30 and 50 mm from the soma (Supplementary Fig. 1),corresponding to the action potential initiation zone that was pre-viously identified6,7,12,13. The length of the AIS, as reflected by AnkGstaining, extended up to 70 mm from the soma.

To examine the density of total Na+ channels, we stained the AISwith a pan-alpha Na+ channel antibody (Pan-Nav), which recognizes

Received 5 December 2008; accepted 27 May 2009; published online 26 July 2009; doi:10.1038/nn.2359

Institute of Neuroscience, State Key Laboratory of Neuroscience, Shanghai Institutes for Biological Sciences, Chinese Academy of Sciences, Shanghai, China.Correspondence should be addressed to Y.S. (shu@ion.ac.cn).

996 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

both Nav1.2 and Nav1.6 channels. Consistent with previous findings,we observed strong Pan-Nav immunoreactivity at the AIS, with a peakdistribution between 10 and 20 mm (Fig. 1c), corresponding to theregion of Nav1.2 accumulation. Although immunoreactivity was unde-tectable in the somatic membrane, electrophysiological recordings inthe following experiments indicated that high-threshold Na+ channelswere present in the soma (see below).

Na+ current density is highest at the AIS

We next carried out voltage-clamp recordings in layer 5 pyramidalneurons to examine the density and voltage-dependent property of Na+

currents along the somato-axonal axis. These neurons had an initialunmyelinated axon segment that was up to 300 mm long13. We obtainedoutside-out patches from the soma and the blebs (3–6 mm in diameter)formed at the axon cut end during slicing procedures13,21,32.

In contrast with previous findings2,3, we found that the amplitude ofthe transient Na+ currents progressively increased from 0 to 30 mmfrom the soma, reaching an average value of 923 ± 85 pA (n ¼ 23) at30–70 mm (Fig. 2a–c), the distal AIS region corresponding to the actionpotential initiation zone6,7,13,33. Furthermore, the Na+ current ampli-tude at the unmyelinated axonal regions distal to the AIS (470 mm),although lower than that at the distal AIS, was still much higher thanthat found at the soma (Fig. 2a–c). Consistent with previousreports2,3,23, the somatic Na+ currents recorded in the regular out-side-out patches were small (26.9 ± 4.7 pA, n ¼ 17; Fig. 2a–c). Insomatic nucleated patches (Fig. 2d), we observed much larger Na+

currents (670 ± 54 pA, n ¼ 34). In addition to recording from excisedoutside-out patch, we also recorded currents from isolated axon blebsand compared the data with those obtained from somatic nucleatedpatches. The peak Na+ currents from isolated blebs varied from 0.5 to3.5 nA (Fig. 2d,e). Normalization by the bleb surface area yielded acurrent density of 25.4 ± 2.3 pA mm�2 (n ¼ 11; Fig. 2e), which wasB19-fold higher than that of somatic nucleated patches (1.5 ±0.1 pA mm�2, n ¼ 34). Together, these results provide direct electro-physiological evidence for the highly clustered distribution of Na+

channels at the AIS, as predicted previously by computationalmodeling4,17 and shown by immunostaining studies12,18,19,25,34.

Distal axonal Na+ channels have the lowest threshold

Examination of the voltage-dependent properties of Na+ currentsrevealed that Na+ channels in the distal AIS and axonal membrane(50–500 mm from soma) had an average activation threshold of�55.5 ± 0.8 mV (n ¼ 14, average distance 77.8 ± 10.1 mm) andcomplete activation at around �20 mV, whereas somatic Na+ channelshad an activation threshold of �42.7 ± 1.1 mV (n ¼ 11) and completeactivation at �10 to �5 mV (Fig. 3a,b). Fitting the activation curvesusing Boltzmann equation yielded the half-activation voltages (V1/2)and slopes (axon: V1/2 ¼ �43.9 ± 1.3 mV, k ¼ 5.7 ± 0.2; soma:V1/2 ¼ �29.7 ± 1.0 mV, k ¼ 5.8 ± 0.2). There was no significantdifference between the slopes (P ¼ 0.12), but the activation curve for

AnkG

Merge Merge Merge

Nav1.2

Nav1.2

Soma 10 20 30 40 Soma

Distance from soma (µm)

10 20 30 40 Soma 10 20 30 40

Nav1.6 Pan-NavNav1.20.8

0.6

0.4

Nor

mal

ized

fluor

esce

nce

inte

nsity

0.2

0.8

0.6

0.4

0.2

0.8

0.6

0.4

0.2

Nav1.6 Pan-Nav

AnkG AnkG

a b c

(n = 55)(n = 55) (n = 43)

Figure 1 Polarized distribution of Na+ channel subtypes. (a) Antibody

staining for AnkG (red) and Nav1.2 (green) in the rat prefrontal cortex. Note

that the proximal AIS had strong staining for Nav1.2. (b) Double staining for

AnkG and Nav1.6 (green). Note that the distal region of the AIS was heavily

stained. (c) Double staining for AnkG and Pan-Nav (green). Plots of the

averaged (± s.e.m.) fluorescence intensity (see Online Methods) as a

function of distance from the soma at the AIS are shown for a–c. Images

are projections of confocal z stacks. Scale bars represent 10 mm. Error barsrepresent s.e.m.

Figure 2 Estimates of Na+ channel density at the

soma and the axon with regular and giant outside-

out patch recording. (a) Top, schematic diagram

of the outside-out recording from patches

excised from the soma and axon blebs. Bottom,

examples of peak Na+ current evoked by step

depolarizations (30 ms) from a holding potential

of �100 to +20 mV in outside-out patches

obtained from the soma (black), AIS (orange,

distance (d) ¼ 39 mm) and axon (red,d ¼ 265 mm). (b) Plot of peak Na+ current in

somatic and axonal outside-out patches with varying distances from the soma, indicating a peak distribution of Na+ currents at the distal AIS. (c) Average

amplitude (± s.e.m.) of the peak Na+ current obtained from the soma and different compartments of the axon. Error bars represent s.e.m. *** indicates

P o 0.001. (d) Top, schematic diagram of the giant outside-out patch recordings: nucleated patch and isolated bleb recording. Bottom, examples of peak Na+

current in nucleated patch (black) and isolated bleb (red, 450 mm). (e) Plot of peak Na+ current as a function of bleb surface area. The dashed line represents

the linear regression fit.

2.0

Soma

a b c d e

50 pA

1 ms

500 pA500 pA

AIS Axon

I Id

Pea

k N

a+ c

urre

nt (

nA)

Pea

k N

a+ c

urre

nt (

nA)

Pea

k N

a+ c

urre

nt (

nA)

1.5

1.0

1.0

Nucleatedpatch

Isolatedbleb

0.8

0.6

500 pABleb area (µm2)

3.0

0 100

2.0

1.0

0

1 ms

0.4

0.2

0

0.5

Soma 100 200Distance from soma (µm)

300

***

Soma

<30

µm

>70

µm

30–7

0 µm

0

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 997

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

axonal currents showed a hyperpolarizing shift of B14 mV (Fig. 3b), avalue twice that reported previously3,23. The half-inactivation voltagesshowed a similar shift (axon: �80.0 ± 1.0 mV, n ¼ 13; soma: �67.0 ±1.7 mV, n ¼ 11), whereas the slopes were 5.4 ± 0.2 and 7.1 ± 0.3 foraxonal and somatic Na+ currents, respectively (Fig. 3b).

We next examined the properties of Na+ channels in outside-outpatches excised from axon blebs at various distances from the soma(8–310 mm; Fig. 3c). V1/2 showed a gradual reduction (hyperpolariza-tion) with increasing distance from the soma (8–50 mm) and reached arelatively steady value in the axon regions beyond the AIS (Fig. 3c). Onthe basis of these data and our immunostaining results, we concludedthat the Na+ channels of the lowest threshold found at the distal AISand the regions distal to AIS corresponded to the low-threshold Nav1.6channels and those with higher threshold at the proximal AIS werelargely the high-threshold Nav1.2 channels28 (Figs. 1 and 3c).

Modeling studies35 have suggested that Na+ channels may havecooperative activation, which may lead to a hyperpolarizing shift intheir activation threshold. We examined this possibility by comparingthe activation curves of Na+ currents recorded in isolated axon blebsbefore and after partial blockade of Na+ channels with tetrodotoxin(TTX), which decreases Na+ currents by reducing the density offunctional channels, a treatment that should diminish cooperativeactivation. Alternatively, we compared the activation of Na+ currentsin normal and low-Na+ artificial cerebrospinal fluid (ACSF), whichreduces the current amplitude without affecting the density of func-tional channels. As expected, we found that the peak amplitude of Na+

currents was significantly reduced after puffing TTX (control, 1,389 ±265 pA; TTX, 432 ± 166 pA; P o 0.01, n ¼ 8) or in the presence of low-Na+ ACSF (577 ± 60 pA, P o 0.01, n ¼ 9) (Fig. 4a,b). We observed aslope change with TTX and in low-Na+ ACSF (control, 3.6 ± 1.3; TTX,5.3 ± 0.3; P o 0.01; low Na+, 5.4 ± 0.2, P o 0.01), presumably as aresult of better voltage clamp of small currents (Supplementary Fig. 2).However, there was no significant difference in the V1/2 for both cases(control, �42.0 ± 0.7 mV; TTX, �42.1 ± 0.5 mV, P ¼ 0.95; low Na+,�41.8 ± 0.8 mV, P ¼ 0.82; Fig. 4c,d), suggesting that cooperativity ofchannel opening does not contribute to the low threshold of axonalNa+ channels that we observed under our experimental conditions.

Nav1.6 controls action potential initiation

We next used modeling analysis to investigate the role of Na+ channeldensity and selective distribution of channel subtypes in actionpotential initiation and propagation. On the basis of our estimates ofchannel density and our immunostaining results (see Online Methods,Fig. 5a and Supplementary Fig. 3), our modeling analysis indicatedthat action potential initiation site is located at 40–45 mm from thesoma, which is consistent with previous experimental findings6,13,33.Stepwise decreases in the total Na+ channel density at the AIS from avalue 40-fold to fivefold greater than that of soma resulted in a shift ofthe initiation site to a more distal location in the axon and an increasein the threshold current (or voltage, see Online Methods), with a rate of0.023 nA (or 5.5 mV) per tenfold density decrease. Thus, the total Na+

channel density at the AIS is critical for determining the location andthreshold of action potential initiation.

To investigate the respective contribution of Nav1.6 and Nav1.2channel density in determining the action potential initiation thresh-old, we constructed a model of the axon with uniform diameter andelectrophysiological properties. Setting various combinations of Nav1.2and Nav1.6 densities (from 80 to 3,000 pS mm�2) resulted in actionpotential threshold changes. The contour lines of the action potentialthreshold changes largely paralleled the Nav1.2 axis, but were perpen-dicular to the Nav1.6 density axis, indicating a predominant role ofNav1.6, but not Nav1.2, in determining the voltage threshold for actionpotential initiation (Fig. 5b). Plotting different combinations ofchannel subtype densities along the length of the AIS revealed that

Figure 3 Comparison of voltage dependence

of somatic and axonal Na+ currents.

(a) Representative families of Na+ currents

evoked by step depolarizations (inset) at soma

(nucleated patches), AIS and axon (regular

outside-out patches). (b) Activation and

availability curves for somatic and axonal

(450 mm from the soma) Na+ current. Theactivation curve for proximal-AIS Na+ current

(o30 mm) was inserted for comparison. The

values were normalized to the peak current and

then averaged (± s.e.m.) between different neurons. These data could be well fitted with Boltzmann equations (dashed lines). Soma, proximal AIS and axon

indicate the results obtained from somatic nucleated patches, proximal AIS and axonal outside-out patches, respectively. (c) The V1/2 of activation was plotted

as a function of recording distances from the soma. Single exponential fit (dashed line) revealed a steep reduction in V1/2 with a length constant of 18.4 ± 5.1

mm (n ¼ 53). The averaged V1/2 (± s.e.m.) for different compartments are shown. Error bars represent s.e.m.

Soma

Soma 50 100 150Distance from soma (µm)

250 300200

AIS(18 µm)

Axon(265 µm)

–100 mV–70 mV

+20 mV

200 pA

200 pA

200 pA

a b c100

Proximal AIS

AxonAxon

–25 SomaProximal AISDistal AISAxon

–30

–40

–45

–50

–3580

Frac

tion

of p

eak

(%)

V1/

2 of

act

ivat

ion

(mV

)

60

40

20

–100 –80

Soma Soma

–60Voltage (mV)

–40 –20 0

0

0.2 ms

Control

a

c d

b

250 pA

Pea

k N

a+ c

urre

nt (

nA)

V1/

2 of

act

ivat

ion

(mV

)

1.5

**

1.0

0.5

Contro

lTTX

Low N

a+

Contro

lTTX

Low N

a+

0

0.2 ms

100 ControlTTXLow Na+

–43

P = 0.82

P = 0.95

–42

–41

–40

–39

–38

80

60

40

20

–80Membrane potential (mV)

Frac

tion

of p

eak

(%)

–60 –40 –20

0

TTX(partial block)

Figure 4 The low activation threshold of axonal Na+ channels is not

attributed to cooperative activation. (a) Representative families of currents

recorded from an isolated bleb before and after puffing TTX (partially blockingNa+ channels). (b) Average peak Na+ currents in controls (open, n ¼ 8),

with TTX (gray, n ¼ 8) and in low-Na+ ACSF (black, n ¼ 9). (c) Activation

curves for Na+ currents under these conditions. (d) Bar plot of V1/2 in

the three conditions (mean ± s.e.m.). ** indicates P o 0.01. Error bars

represent s.e.m.

998 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

the distal AIS had the lowest threshold (approximately �58 mV;Fig. 5b). Further calculation of the ratio of Nav1.6-mediated conduc-tance versus the total Na+ conductance at the time of action potentialinitiation revealed a ratio close to 1 at the AIS beyond the first 10 mm(Fig. 5c). Together, these results indicate that the Nav1.6 channelsdetermine the lowest threshold of action potential initiation at thedistal AIS. In contrast, the Nav1.2 channels have a rather limited role inaction potential initiation.

Nav1.2 promotes action potential propagation to the soma

To examine how high-threshold perisomatic Na+ channels affect thebackpropagation of action potentials into the somatodendritic compart-ment, we performed modeling analysis (Fig. 5d,e) as well as simulta-neous whole-cell recording from the soma and the axonal bleb (Fig. 6a).The generation of full somatic action potentials in response to axonalcurrent injection depended on the level of somatic membrane potential(Vm; Fig. 6b). Close examination of the rising phase of the somaticaction potentials revealed two components: the AIS and somatodendriticpotential. Somatic spikelets (resulting from the arrival of AIS potentials)occurred frequently in response to axonal stimulation, reflecting thefailure of somatodendritic potential (Fig. 6b,c). Plotting the probabilityof somatic action potential as a function of Vm revealed a steep voltagedependence, with an average Vm for the failure in action potentialbackpropagation (failure threshold) of�75.1 ± 2.4 mV (n¼ 14; Fig. 6d).The variability of this threshold may be attributed to the density varia-tion of AIS Nav1.2 channels among these neurons (see Fig. 5).

Consistent with these results, our simulations indicated that AISNav1.2 channels are important in action potential backpropagation

(Fig. 5). A decrease in the density of the Nav1.2 channels at the AIS, butnot of that in the somatodendritic compartment, resulted in adepolarizing shift in the failure threshold of action potential back-propagation to the soma, with a 2.8-mV shift per 10% density change(Fig. 5d). Complete removal of the AIS Nav1.2 channels resulted in adecrease in the amplitude of somatic spikelets and caused a completefailure of action potential backpropagation (Supplementary Fig. 4).Setting hyperpolarizing V1/2 of Nav1.2 channels at the AIS (or soma-todendritic compartments) from �30 to �43 mV resulted in ahyperpolarizing shift in the failure threshold with a rate of 1.8 mVper mV (or 0.3 mV per mV for somatodendritic compartment)(Fig. 5e). In contrast, the failure threshold was independent of theV1/2 and the density of the AIS Nav1.6 channels (Fig. 5d,e). Takentogether, these results indicate that Nav1.2 channels accumulated at theproximal AIS promote action potential backpropagation.

AIS and somatodendritic potential threshold

Our results suggest that both distal Nav1.6 and proximal Nav1.2channels at the AIS are involved in generating full somatic action

2,000

Den

sity

(pS

µm

–2)

Den

sity

of N

a v1.

6 (p

S µ

m–2

)

Failu

re th

resh

old

(mV

)

1,000

–60

–70

–80

–90

–60

–70

–80

–90

–70Somatodendritic

–80

–90

–100–45 –35 –30–40 –45 –35 –30–40

–70

–80

–90

–1000.5 1.0

Density (fold)

Failu

re th

resh

old

(mV

)

1.5 0.5 1.0

Density (fold) V1/2 of activation (mV) V1/2 of activation (mV)

1.5

2,000 3,0001,000

2,000

3,000–61

Threshold voltage (mV)

Proximal AISDistal AIS

Proximal AISDistal AIS

Ratio of Nav1.6 contribution

–38

1,000

2,000

3,0000.3 1.0

1,000

Soma 20Distance from soma (µm)

Somatodendritic

Density of Nav1.2 (pS µm–2)

2,000 3,0001,000

Density of Nav1.2 (pS µm–2)

40 60

Nav1.2

a

d e

b c

Nav1.2

AIS Nav1.2

AIS Nav1.6

AIS Nav1.2

AIS Nav1.6

Nav1.2

Nav1.6

0

Figure 5 Simulations indicate distinct functions

of AIS Nav1.6 and Nav1.2 in action potential

initiation and backpropagation to the soma.

(a) Density settings for Nav1.6 and Nav1.2 at

the AIS based on experimental observations.

(b) Contour map of the action potential threshold

with various density combinations of the two

channel subtypes. Dots indicate the local densityfor individual channel subtypes at the AIS (2-mm

interval). (c) Contour map of the ratio of Nav1.6-

mediated conductance versus total Na+ conduc-

tance at the action potential threshold. Dots

indicate the local density for individual channel

subtypes at the AIS (2-mm interval). (d) Plots of

the failure threshold for somatic action potentials

as a function of different density settings of the

somatodendritic or AIS Nav1.2 (left) and AIS

Nav1.6 (right). Backpropagating action potentials

were evoked by current injection in axon regions

distal to the AIS (Supplementary Fig. 8).

(e) Dependence of the failure threshold on the

voltage-dependent activation of the somatoden-

dritic or AIS Nav1.2 (left) and AIS Nav1.6 (right).

200

dV/d

t (V

s–1

)

Pro

babi

lity

of s

omat

icac

tion

pote

ntia

l

1.0

0.8

0.6

0.4

0.2

0.0

150

100

AIS

Somatodendritic

Somatodendritic

Soma

a

c d

b

–71 mV

Spikelets AIS

AIS

1 nA

50 mV

20 msAction potentials

–100–80 –60 –40 –50–60–70–80–90–100

Membrane potential (mV) Membrane potential (mV)

–20 0 20

50

–50

0

Figure 6 Vm dependence of action potential backpropagation. (a) Morphology

of a recorded layer 5 pyramidal neuron in the prefrontal cortex. The two

electrodes indicate the somatic and axonal recording sites. Scale bar

represents 25 mm. (b) Simultaneous somatic and axonal recordings revealed

that axonal action potentials evoked by current injection at the axon failed to

generate full action potentials at hyperpolarizing somatic Vm. Note that only

spikelets were detected at the soma when failure occurred. Inset, overlay of

the somatic action potentials and one of the spikelets revealed twocomponents, the AIS and somatodendritic potential. (c) Phase plots of the

somatic action potentials and spikelets. (d) Plots of somatic action potential

probability as a function of somatic Vm. Colored lines represent recordings

from different neurons (n ¼ 14).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 999

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

potentials; activation of Nav1.6 initiated the AIS potential, whereasactivation of Nav1.2 triggered the somatodendritic potential (Fig. 6b,c).Indeed, local application of TTX at the distal AIS markedly reduced therising slope of the AIS potential, whereas similar application of TTX atthe perisomatic region mostly affected the somatodendritic potential(n ¼ 12; Supplementary Fig. 5). Further analysis showed that the AISpotential threshold measured at the soma (the traditional actionpotential threshold) correlated linearly with the somatic Vm (slope ¼1.03 ± 0.04 mV per mV, n ¼ 14, r ¼ 0.99; Fig. 7a,b), suggesting thatsomatic Vm regulates the action potential threshold. This was furthersupported by the observation that action potentials generated at theactive brain state, or the ‘Up’ state36,37, had a threshold (�47.5 ±0.1 mV, n ¼ 2254 action potentials in seven regular spiking cells;Fig. 7c,d) that linearly correlated with Vm (slope ¼ 0.93 ± 0.01 mV permV; r ¼ 0.93; Fig. 7d), suggesting that the highly variable thresholdnormally observed at the soma35,38,39 may result from Vm fluctuations.

Because the somatodendritic potential is generated at the peri-somatic region as a result of the activation of Nav1.2 channels, somaticrecording may faithfully reveal somatodendritic potential threshold, assuggested by a recent study14, rather than AIS potential threshold. Tofurther test this idea, we defined the somatodendritic potential thresh-old as the voltage at which the time derivative of Vm (dV

dt ) is 20 V s�1

above the maximum curvature of the trough between the AIS andsomatodendritic potentials in the phase plot (Fig. 8a). Consistent withrecent estimates of the somatodendritic potential threshold14, wefound that the average threshold was �19.6 ± 4.8 mV (n ¼ 13 cells),a value that was independent of the preceding Vm (r ¼ 0.12),

whereas AIS potential threshold showed a strong correlation withthe Vm (r ¼ 0.96, n ¼ 13 cells; Fig. 8b).

These experimental data, together with simulation results, indicatethat activation of Nav1.6 channels at the distal AIS contributes to thedetermination of the AIS potential threshold, whereas activation ofNav1.2 channels contributes to the determination of the somatoden-dritic potential threshold and thereby sets the threshold for thegeneration of full somatic action potentials.

DISCUSSION

In this study, we found that Nav1.2 and Nav1.6 are selectively targetedto the proximal and distal AIS, respectively. The distribution of Nav1.6channel peaked at the distal AIS, corresponding to the action potentialinitiation zone. Consistently, our electrophysiological experimentsindicated that Na+ channels at the distal AIS and the adjacent axonhad a lower activation threshold than those at the proximal AIS and thesoma. Moreover, our regular and giant patch recordings provide directevidence that AIS has a higher Na+ current density than the soma andnearby axon regions, with a ratio of 1 to 34 to 27 (soma to AIS toadjacent axon). Our simulation results further support the notion thatNav1.6 channels accumulated at the distal AIS determine the lowestthreshold for action potential initiation. The results of our simulta-neous somatic and axonal recordings, together with simulations,indicate that the highly clustered Nav1.2 channels at the proximalAIS strongly promote, if not guarantee, action potential backpropaga-tion into the soma.

Density and activation threshold of axonal Na+ channels

Previous outside-out patch recordings from AIS of pyramidal neuronshave suggested that the Na+ channel density is up to threefold greaterthan that of the soma3. However, a recent study23 has demonstratedthat the channel density may be underestimated as a result of tightchannel anchoring to the intracellular cytoskeleton. We found that thepeak amplitudes of Na+ currents recorded from the axonal patcheswere significantly higher than those of the somatic ones (P o 0.001;Fig. 2), consistent with previous immunostaining results12,18,19,25,34

and computational predictions4,17. Although we obtained recordingsfrom injured axons, several lines of evidence indicate that these densityestimates are likely to reflect the actual density in the axon. First, ourimmunostaining revealed similar signal levels at the axon trunk versusterminal bleb (Supplementary Fig. 6). Second, axonal action potentialsrecorded at the bleb had a profile similar to those recorded at the AIS,

50a

c d

b

–50

–40

In vivo

In vitro

–30 –38

–40

–42

–44

–46

–48

–50 –48 –46Membrane potential (mV)

–44 –42 –40

–40

–50–0.7 –0.3

0.5 1.0

Time (ms)–60

–80

–70

–60

–80

–60

–80 –70Membrane potential (mV)

–60

Mem

bran

e po

tent

ial (

mV

)M

embr

ane

pote

ntia

l (m

V)

Thr

esho

ld p

oten

tial (

mV

)T

hres

hold

pot

entia

l (m

V)

0

0

0 2.5 5.0Time (s)

7.5

20 40 60Time (ms)

80 100 120

400

a b

20 V s–1

dV/d

t (V

s–1

)

Thr

esho

ld v

olta

ge (

mV

)

Thr

esho

ld v

olta

ge (

mV

)

300

200

100

Membrane potential (mV)

–100

–60

–40

–20

–80

–60

–40

–20

–80

–100 –90 –80–70

–80 –70

Membrane potential (mV)–60

r = 0.12

r = 0.96

–5050–50 0

0

Figure 8 Voltage threshold of somatodendritic potential is independent of

the preceding Vm. (a) Example phase plots of somatic action potentials (gray)and spikelets (black) evoked by current injection in the axon. Gray and black

circles indicate the thresholds of the AIS potentials and the somatodendritic

potentials, respectively. Inset, plot of the threshold voltage as a function

of Vm. (b) Group results (n ¼ 13). Gray, thresholds for AIS potentials; black,

thresholds for somatodendritic potentials. Dashed lines indicate the best

linear regression fits.

Figure 7 The threshold of somatic action potential is dependent on the

preceding Vm. (a) Example sweeps from an in vitro recording (similar

recording as in Fig. 6). Gray, full action potentials; black, spikelets (AIS

potentials). Horizontal bars indicate the action potential thresholds. Inset,

expansion of the action potential onset for clarity. (b) Threshold potentials of

the action potentials (gray) and spikelets (black) in a are plotted as a function

of Vm. The dashed line represents the linear regression fit. (c) An example

trace of intracellular recording from a regular spiking neuron in vivo.Action potentials are truncated for clarity. Bars indicate the action potential

thresholds. Inset, overlay of the onsets of all action potentials. Three

individual traces are highlighted to show the correlation of action potential

threshold and the level of Vm (vertical bars) at 0.4 ms before the threshold.

(d) A plot of the threshold potential (from c) as a function of Vm (n ¼ 593).

The dashed line represents the linear regression fit.

1000 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

with a rapid rising phase reflecting the high availability of Na+

channels39. Third, axonal Na+ channels may be freed from the tightassociation with cytoskeleton during the slicing procedure. Immuno-reactivity for AnkG was undetectable in the axon blebs (n ¼ 25 of 25;Supplementary Fig. 6). Finally, our giant patch recording made fromthe isolated blebs as a whole had minimized the influence of tightchannel anchoring to the cytoskeleton and yielded results similar tothose from outside-out patches (Fig. 2). Together, our results provide,to the best of our knowledge, the first direct electrophysiologicalevidence of a high Na+ channel density at the AIS and the regionsbeyond the AIS.

Differential biophysical properties of Na+ channels along thesomato-axonal axis may result from a differential distribution ofchannel subtypes18, different levels of phosphorylation40 and density-dependent cooperative activation35. Previous work has revealed thatNav1.6 channels are activated and inactivated at more negative Vm thanNav1.2 channels28. Here, we found that the selective accumulation ofNav1.2 and Nav1.6 channels at the proximal and distal AIS correlateddirectly with the reduction in V1/2 (Fig. 3). A recent study24 found agradual increase in Nav1.6 immunostaining intensity at the AIS in layer2/3, but not in layer 5 pyramidal cells. In contrast, our experimentsrevealed a nonuniform distribution of Nav1.6 at the AIS of layer 5pyramidal neurons in both the prefrontal cortex and the primarysomatosensory cortex (Supplementary Fig. 7). Whether these differ-ences in immunostaining results could be attributed to regionaldifferences in the brain remains unknown.

The notion of cooperative channel activation35, whereby the openingof a Na+ channel lowers the activation threshold of its neighboringchannels in a distance-dependent manner, was excluded by partialTTX blockade and low-Na+ experiments. Application of TTX shoulddiminish the cooperative activation of Na+ channels by increasingthe distance between functional channels, whereas application of low-Na+ ACSF would not change the channel cooperativity because thedensity of functional channels were not affected. However, we found nosubstantial change in the V1/2 with either treatment (Fig. 4). Therefore,we conclude that preferential distribution and clustering of Na+

channel subtypes represent the predominant factors in determiningthe low threshold of action potential initiation at the AIS. Whetherdifferential channel phosphorylation in subcellular compartments mayalso contribute to the activation threshold remains to be examined.

Distinct functions of Nav1.6 and Nav1.2

Previous computational studies have yielded conflicting conclusionsabout the role of Na+ channel density and biophysical properties inaction potential initiation as a result of inconsistent estimates ofchannel density3,4,17,23. Our modeling studies integrated the immu-nostaining and electrophysiological results and showed that the lowestthreshold for action potential initiation at the distal AIS was largelydetermined by the density of low-threshold Nav1.6 channels (Fig. 5).Consistent with these modeling results, axonal Nav1.6 channels havebeen shown to lower the threshold voltage and participate in actionpotential initiation in different neuronal types1,2,26.

Distinct from the function of Nav1.6 channel, the Nav1.2 channelmay control action potential backpropagation because of its highdensity at the proximal AIS and high threshold. Our simulationsrevealed a depolarizing shift of the failure threshold for action potentialbackpropagation with a decreasing density of the AIS Nav1.2 channels(Fig. 5d). Removal of these channels resulted in a decrease in the spikeletamplitude and a complete failure of action potential backpropagation(Supplementary Fig. 4). Experimental (Fig. 6) and modeling results(Fig. 5e) indicate that somatic invasion of backpropagating action

potentials is tightly linked to the somatic Vm and voltage-dependentactivation of AIS Nav1.2 channels. Consistent with a recent study14,our experiments revealed a threshold of approximately �20 mV forsomatodendritic potentials, and further demonstrated that this thresh-old, but not that of AIS potential recorded at the soma, remainedconstant in accordance with somatic Vm fluctuations. Because success-ful invasion of the backpropagating action potential into the dendritictree is critical for synaptic plasticity7,41–45, Nav1.2 channels at AIS mayparticipate in neuronal plasticity.

In conclusion, distal AIS accumulation of Nav1.6 channels deter-mines the low threshold for action potential initiation; whereasproximal AIS accumulation of Nav1.2 channels sets the threshold forthe generation of somatodendritic potentials and ensures actionpotential backpropagation to the soma and dendrites. Thus, Nav1.6and Nav1.2 channels serve distinct functions in action potentialinitiation and backpropagation.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe thank M.M. Poo, D.A. McCormick and M.H. Kole for their valuablecomments on this work. We are also grateful to Y. Yu for his help in computermodeling. This work was supported by the 973 Program (2006CB806600),a Shanghai Commission of Science and Technology grant (06DJ14010), theShanghai Pujiang Program (07PJ14108), the Hundreds of Talents Program andKnowledge Innovation Project from Chinese Academy of Sciences (KSCX2-YW-R-102), and Projects of the Scientific Research Foundation of the State HumanResource Ministry and the Education Ministry.

AUTHOR CONTRIBUTIONSW.H. performed the patch-clamp and whole-cell recording experiments,simulations, and data analysis. C.T. carried out the immunostaining experiments.T.L. performed the sharp electrode recordings. M.Y. and H.H. helped with dataanalysis and simulations. Y.S. designed the experiments and wrote the paper.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Clark, B.A., Monsivais, P., Branco, T., London, M. & Hausser, M. The site of actionpotential initiation in cerebellar Purkinje neurons. Nat. Neurosci. 8, 137–139 (2005).

2. Colbert, C.M. & Johnston, D. Axonal action-potential initiation and Na+ channeldensities in the soma and axon initial segment of subicular pyramidal neurons.J. Neurosci. 16, 6676–6686 (1996).

3. Colbert, C.M. & Pan, E. Ion channel properties underlying axonal action potentialinitiation in pyramidal neurons. Nat. Neurosci. 5, 533–538 (2002).

4. Mainen, Z.F., Joerges, J., Huguenard, J.R. & Sejnowski, T.J. A model of spike initiation inneocortical pyramidal neurons. Neuron 15, 1427–1439 (1995).

5. Milojkovic, B.A., Wuskell, J.P., Loew, L.M. & Antic, S.D. Initiation of sodium spikelets inbasal dendrites of neocortical pyramidal neurons. J. Membr. Biol. 208, 155–169 (2005).

6. Stuart, G., Schiller, J. & Sakmann, B. Action potential initiation and propagation in ratneocortical pyramidal neurons. J. Physiol. (Lond.) 505, 617–632 (1997).

7. Stuart, G., Spruston, N., Sakmann, B. & Hausser, M. Action potential initiation andbackpropagation in neurons of the mammalian CNS. Trends Neurosci. 20, 125–131(1997).

8. Coombs, J.S., Curtis, D.R. & Eccles, J.C. The interpretation of spike potentials ofmotoneurones. J. Physiol. (Lond.) 139, 198–231 (1957).

9. Eccles, J.C. The Physiology of Nerve Cells (The Johns Hopkins Press, Baltimore, 1957).10. Fatt, P. Sequence of events in synaptic activation of a motoneurone. J. Neurophysiol. 20,

61–80 (1957).11. Fuortes, M.G., Frank, K. & Becker, M.C. Steps in the production of motoneuron spikes.

J. Gen. Physiol. 40, 735–752 (1957).12. Meeks, J.P. & Mennerick, S. Action potential initiation and propagation in CA3

pyramidal axons. J. Neurophysiol. 97, 3460–3472 (2007).13. Shu, Y., Duque, A., Yu, Y., Haider, B. & McCormick, D.A. Properties of action-potential

initiation in neocortical pyramidal cells: evidence from whole cell axon recordings.J. Neurophysiol. 97, 746–760 (2007).

14. Kole, M.H. & Stuart, G.J. Is action potential threshold lowest in the axon? Nat. Neurosci.11, 1253–1255 (2008).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1001

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

15. Dodge, F.A. & Cooley, J.W. Action potential of the motoneuron. IBM J. Res. Develop. 17,219–229 (1973).

16. Moore, J.W., Stockbridge, N. & Westerfield, M. On the site of impulse initiation in aneurone. J. Physiol. (Lond.) 336, 301–311 (1983).

17. Rapp, M., Yarom, Y. & Segev, I. Modeling back propagating action potential in weaklyexcitable dendrites of neocortical pyramidal cells. Proc. Natl. Acad. Sci. USA 93,11985–11990 (1996).

18. Van Wart, A., Trimmer, J.S. & Matthews, G. Polarized distribution of ion channels withinmicrodomains of the axon initial segment. J. Comp. Neurol. 500, 339–352 (2007).

19. Inda, M.C., DeFelipe, J. & Munoz, A. Voltage-gated ion channels in the axon initialsegment of human cortical pyramidal cells and their relationship with chandelier cells.Proc. Natl. Acad. Sci. USA 103, 2920–2925 (2006).

20. Kole, M.H., Letzkus, J.J. & Stuart, G.J. Axon initial segment Kv1 channels control axonalaction potential waveform and synaptic efficacy. Neuron 55, 633–647 (2007).

21. Shu, Y., Yu, Y., Yang, J. & McCormick, D.A. Selective control of cortical axonal spikes by aslowly inactivating K+ current. Proc. Natl. Acad. Sci. USA 104, 11453–11458(2007).

22. Howard, A., Tamas, G. & Soltesz, I. Lighting the chandelier: new vistas for axo-axoniccells. Trends Neurosci. 28, 310–316 (2005).

23. Kole, M.H. et al. Action potential generation requires a high sodium channel density inthe axon initial segment. Nat. Neurosci. 11, 178–186 (2008).

24. Lorincz, A. & Nusser, Z. Cell type–dependent molecular composition of the axon initialsegment. J. Neurosci. 28, 14329–14340 (2008).

25. Boiko, T. et al. Functional specialization of the axon initial segment by isoform-specificsodium channel targeting. J. Neurosci. 23, 2306–2313 (2003).

26. Royeck, M. et al. Role of axonal Nav1.6 sodium channels in action potential initiation ofCA1 pyramidal neurons. J. Neurophysiol. 100, 2361–2380 (2008).

27. Duflocq, A., Le Bras, B., Bullier, E., Couraud, F. & Davenne, M. Nav1.1 is predominantlyexpressed in nodes of Ranvier and axon initial segments. Mol. Cell. Neurosci. 39,180–192 (2008).

28. Rush, A.M., Dib-Hajj, S.D. & Waxman, S.G. Electrophysiological properties of two axonalsodium channels, Nav1.2 and Nav1.6, expressed in mouse spinal sensory neurones. J.Physiol. (Lond.) 564, 803–815 (2005).

29. Boiko, T. et al. Compact myelin dictates the differential targeting of two sodium channelisoforms in the same axon. Neuron 30, 91–104 (2001).

30. Kaplan, M.R. et al. Differential control of clustering of the sodium channelsNa(v)1.2 and Na(v)1.6 at developing CNS nodes of Ranvier. Neuron 30, 105–119(2001).

31. Komai, S. et al. Postsynaptic excitability is necessary for strengthening of corticalsensory responses during experience-dependent development. Nat. Neurosci. 9,1125–1133 (2006).

32. Shu, Y., Hasenstaub, A., Duque, A., Yu, Y. & McCormick, D.A. Modulation of intracorticalsynaptic potentials by presynaptic somatic membrane potential. Nature 441, 761–765(2006).

33. Palmer, L.M. & Stuart, G.J. Site of action potential initiation in layer 5 pyramidalneurons. J. Neurosci. 26, 1854–1863 (2006).

34. Wollner, D.A. & Catterall, W.A. Localization of sodium channels in axon hillocks andinitial segments of retinal ganglion cells. Proc. Natl. Acad. Sci. USA 83, 8424–8428(1986).

35. Naundorf, B., Wolf, F. & Volgushev, M. Unique features of action potential initiation incortical neurons. Nature 440, 1060–1063 (2006).

36. Shu, Y., Hasenstaub, A. & McCormick, D.A. Turning on and off recurrent balancedcortical activity. Nature 423, 288–293 (2003).

37. Steriade, M., Nunez, A. & Amzica, F. A novel slow (o1 Hz) oscillation of neocorticalneurons in vivo: depolarizing and hyperpolarizing components. J. Neurosci. 13,3252–3265 (1993).

38. Yu, Y., Shu, Y. & McCormick, D.A. Cortical action potential backpropagation explains spikethreshold variability and rapid-onset kinetics. J. Neurosci. 28, 7260–7272 (2008).

39. McCormick, D.A., Shu, Y. & Yu, Y. Neurophysiology: Hodgkin and Huxley model—stillstanding? Nature 445, E1–2; discussion E2–3 (2007).

40. Colbert, C.M. & Johnston, D. Protein kinase C activation decreases activity-dependentattenuation of dendritic Na+ current in hippocampal CA1 pyramidal neurons.J. Neurophysiol. 79, 491–495 (1998).

41. 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).

42. Kampa, B.M., Clements, J., Jonas, P. & Stuart, G.J. Kinetics of Mg2+ unblock of NMDAreceptors: implications for spike timing–dependent synaptic plasticity. J. Physiol.(Lond.) 556, 337–345 (2004).

43. Kampa, B.M., Letzkus, J.J. & Stuart, G.J. Requirement of dendritic calcium spikes forinduction of spike timing–dependent synaptic plasticity. J. Physiol. (Lond.) 574,283–290 (2006).

44. Markram, H., Lubke, J., Frotscher, M. & Sakmann, B. Regulation of synaptic efficacy bycoincidence of postsynaptic APs and EPSPs. Science 275, 213–215 (1997).

45. Sjostrom, P.J., Turrigiano, G.G. & Nelson, S.B. Rate, timing, and cooperativity jointlydetermine cortical synaptic plasticity. Neuron 32, 1149–1164 (2001).

1002 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSSlice preparation. We obtained coronal slices of the prefrontal cortex from

P16–20 Sprague-Dawley rats. The use and care of the rats complied with the

guidelines of the Animal Advisory Committee at the Shanghai Institutes for

Biological Sciences. We anesthetized the rats deeply with sodium pentobarbital

(30 mg per kg of body weight) and then quickly dissected out the brain and

placed it in ice-cold, oxygenated (95% O2 and 5% CO2), sucrose-substituted

ACSF in which equiosmolar sucrose was used as a substitute for NaCl. In this

solution, we cut the slices (300 mm) with a microtome (VT1000S, Leica) and

incubated them in aerated normal ACSF containing 125 mM NaCl, 2.5 mM

KCl, 2 mM MgSO4, 2 mM CaCl2, 26 mM NaHCO3, 1.25 mM NaH2PO4 and

25 mM dextrose (315 mOsm, pH 7.4, 35 1C). After B45 min of incubation, we

transferred the slices to a chamber bathed with aerated ACSF (36–36.5 1C) and

visualized the cortical neurons with an upright infrared differential interference

contrast microscope (BX51WI, Olympus).

Electrophysiological recordings. In this study, we used a new method of patch

recording from the cut end of the axon to probe the ion channels at the AIS and

the axon13,21,32. We performed patch-clamp and whole-cell recordings from

layer 5 pyramidal neurons with a Multiclamp 700B or an Axopatch 200B

amplifier (Molecular Devices). Patch pipettes had an impedance of 5–7 MOwith an internal solution containing 145 mM CsCl, 2 mM MgCl2, 2 mM

Na2ATP, 10 mM HEPES, 0.2 mM EGTA and 2 mM TEA (tetraethylammo-

nium) (pH 7.2 with CsOH, 287 mOsm). To trace and label the recorded

neurons, we added Alexa Fluor 488 (100 mM) and biocytin (0.2%) to the

pipette solution. The pipettes that we used for regular somatic and axonal

outside-out patch recordings were made with the same protocol in the puller

and had similar impedance. To isolate the Na+ currents, we added CdCl2(200 mM) to the bath solution and TEA to the Cs+-based internal solution.

We recorded Na+ currents from regular outside-out patches excised from

the soma and the axon blebs, and also from giant outside-out patches: the

somatic nucleated patches46 and the isolated axon blebs. We obtained the axon

blebs by sweeping a sharp electrode underneath the slice surface at the border

of layer 6 and white matter (to disconnect the bleb and the main axon). Bath

application of 1 mM TTX could completely block the somatic and axonal

inward currents.

We recorded families of Na+ currents in response to serials of depolarizing

voltage steps (30 ms) from a preceding prepulse of �100 (50 ms) to +20 mV

and generated the activation curves on the basis of the peak currents at each

step. Activation threshold of the Na+ currents was the voltage at which the

evoked peak current reached 10% of the maximum value. We plotted the

inactivation curves using the peak currents elicited by a test pulse (30 ms) to

�5 mV following a range of voltage steps (50 ms) from �115 to �5 mV. We

calculated the current density of a certain giant patch by dividing the peak

current (evoked during the activation protocols) by membrane area (obtained

by carefully measuring the diameter of the nucleated patches and isolated axon

blebs). The current density was measured at �5 mV and �20 mV for somatic

and axonal patches, respectively. To test the voltage dependence of axonal Na+

channels with reduced density, we put a glass pipette filled with TTX (50 mM)

upstream of the perfusion flow to obtain a peak current less than one-third of

the original amplitude (Fig. 4). In contrast, we bathed the slice with a low-Na+

ACSF (67% NaCl was replaced with NMDG) to reduce the peak current, but

leave all channels available for activation. We obtained simultaneous somatic

and axonal recording in normal ACSF with patch pipettes filled with K+-based

internal solution (140 mM potassium gluconate, 3 mM KCl, 2 mM MgCl2,

2 mM Na2ATP, 10 mM HEPES and 0.2 mM EGTA, pH 7.2 with KOH,

288 mOsm).

For in vivo recordings, we initially anesthetized the rats (280–350 g) with

urethane (1.0 g per kg, intraperitoneal) and supplemented with ketamine and

xylazine hourly (35 mg per kg and 7 mg per kg, respectively, intramuscular).

We performed intracellular recordings from cortical regular spiking neurons

with an AxoClamp 2B amplifier (Molecular Devices). Sharp electrodes had an

impedance of 60–90 MO filled with 2 M potassium acetate.

Our somatic and axonal recordings had an access resistance of less than 20

and 25 MO, respectively. We digitally subtracted the leakage currents using an

on-line P/5 procedure (in which the currents evoked by five hyperpolarizing

pulses with one-fifth amplitude of the test pulse P were summed and added to

the current trace of interest) and low-pass filtered the currents at 10 kHz and

sampled at 100 kHz using pClamp 10 software and a Digidata 1440A interface

(Molecular Devices). The Vm values shown in the text and figures were

corrected with calculated liquid junctions for different internal solutions.

Statistical analysis. Data are presented as mean ± s.e.m. Statistical tests were

performed using two-tailed Student’s t test.

Immunostaining. We killed the rats by perfusion with 1% paraformaldehyde

and 1% sucrose (wt/vol) in 0.1 M phosphate buffer (pH 7.4) after deep

anesthesia with sodium pentobarbital. The brain was removed and post-fixed in

the same fixative for 2 h and then immersed in 30% sucrose in 0.1 M phosphate

buffer for 48 h. We obtained cryostat coronal sections (16 or 100 mm) using a

freezing microtome. In some experiments, we also used slices (300 mm)

prepared for electrophysiological recordings.

We rinsed the sections in 0.01 M phosphate-buffered saline (PBS, pH 7.4)

and incubated them in a blocking solution (5% normal goat serum, 0.3%

Triton X-100 in PBS, vol/vol) at 20–25 1C for 2 h. Sections were incubated

overnight at 4 1C with primary antibody to AnkG (1:200, Santa Cruz) and

either Nav1.2, Nav1.6 or Pan-Nav (1:200, Alomone) in 0.1% Triton (vol/vol).

After a complete wash in PBS, these sections were incubated in Alexa 488–

conjugated goat antibody to rabbit IgG and Alexa 594 goat antibody to mouse

IgG in 0.1% Triton (1:1,000; Molecular Probes) at 20–25 1C for 2 h. We then

washed and mounted the sections with Vecta shield mounting media (Vector

Laboratories). We took images in the linear range of the photomultiplier with a

laser scanning confocal microscope (ZEISS LSM 510 META NLO) and used the

projection of z stack images (0.4 mm per image) in the figures. We did not

detect labeling in the controls (not treated with primary antibodies).

We then used Autoquant X2 software (Media Cybernetics) to deconvolve the

images and employed Amira software (Mercury Computer Systems) to extract

axon signals and measure the fluorescence intensity. Briefly, we subtracted the

background (mean fluorescence intensity of the stack) from the deconvolved

image and obtained an image named C1, and then generated two masks from

this image (for Na+ channel staining and AnkG staining) by defining an

intensity threshold of mean ± 3 s.d. We applied an OR operation of the masks

and obtained a mask image called C2, and then applied an AND operation of

C1 and C2 to obtain the image C3, in which only axonal profiles were kept. In

C2, we traced the axons (total volume 45,000 voxels) and got an axon line set.

Short axons and those with truncated axon hillock were discarded. We

measured the axon diameter in C2 and sampled the fluorescence values in

C3 on the basis of the line set. The fluorescence intensities were normalized to

the peak value at the AIS and averaged between different axons.

Model construction. We implemented two new voltage-gated Na+ channels,

nasoma and naaxon, whose properties were described by Hodgkin-Huxley–

style equations4,47, to simulate the channels at the soma and the distal AIS,

respectively. The parameters of nasoma and naaxon were initially set to the

experimentally observed values (Fig. 3). We inserted the two channels (gbar ¼100 pS mm�2) into a single compartment model (length, 10 mm; diameter,

10 mm) to examine their voltage-dependent properties (Supplementary Fig. 3).

The computational model was implemented using NEURON 6 (ref. 48) and

a previously published multicompartmental model of the full dendritic and

somatic structure of a layer 5 cortical pyramidal cell49 coupled with a cylindrical

AIS and axon segments32,38. The main axon started with an axon hillock

(length, 10 mm) that connected to the soma and tapered from a diameter of

3.8 to 2.4 mm, followed by an AIS (length, 50 mm; diameter, 1.22 mm). The AIS

was attached by an unmyelinated axon (length, 400 mm; diameter, 1.02 mm)

and then a myelinated axon (400 mm in length) with internode distances of

100 mm. The electrical properties Rm, Cm and Ri were set to 30,000 O cm2,

1 mF cm�2 and 150 O cm, respectively, and distributed uniformly throughout

the cell. Myelination was simulated by reducing Cm to 0.02 mF cm�2. The

resting Vm at soma was set to �70 mV. All simulations were run with 20-ms

time steps and the nominal temperature of simulation was 37 1C.

The transient Na+ current was present in all parts of the model cell, but the

channel subtypes were nonuniformly distributed according to our experimental

observations. Specifically, nasoma was present in soma, dendrites and proximal

AIS, whereas naaxon was present in proximal AIS, but dominant in distal AIS

and the axon regions distal to AIS (Fig. 5a). From the recordings of nucleated

doi:10.1038/nn.2359 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

patches, we obtained an average channel density of 45.6 ± 3.7 pS mm�2

(16–107 pS mm�2, corrected with a maximum open probability of 0.5,

n ¼ 34)50. Considering that this density might be underestimated as a result

of membrane stretching and that a density of 80 pS mm�2 (which was in the

range of our estimates) in the dendrite and soma has been found by previous

studies50, we therefore set a density of 80 pS mm�2 for the somatodendritic

compartments as a starting point of the simulations. Reducing channel density

to 60 and 45 pS mm�2 for somatic and dendritic compartments, respectively,

yielded similar results. Unless otherwise stated, channel density was set to

300 pS mm�2 in the axon regions distal to the AIS (460 mm) and 1,600 pS mm�2

in the nodes of Ranvier. At the AIS, nasoma and naaxon contributed to the

total channel density according to immunostaining results with a maximum

density of 3,200 pS mm�2 (40-fold greater than that in soma). The Na+

equilibrium potential was set to +60 mV.

The fast voltage-gated K+ current, IKv, was present in the soma (20 pS mm�2),

dendrites (10 pS mm�2) and unmyelinated axon (1,500 pS mm�2). It increased

linearly with distance in the AIS to a maximum value of 1,000 pS mm�2.

The K+ equilibrium potential was set to �90 mV. The slow non-inactivating

potassium current (M current; Ikm), high-voltage activated Ca2+ current (ICa)

and Ca2+-dependent K+ current (IkCa) were distributed throughout the somatic

and dendritic compartments and had conductances of 0.3, 0.3 and 3 pS mm�2,

respectively. Background leak conductance (gleak) was distributed through-

out the cell with a density of 0.33 pS mm�2, except for the nodes of Ranvier

(200 pS mm�2). The reversal potential of the leak current was �70 mV.

The action potential initiation site was defined as the location where the dVdt

first reached 20 V s�1. The threshold voltage was defined as the somatic Vm at

which dVdt reached 20 V s�1, whereas the threshold current was the minimum

amount of current (duration, 100 ms) injected into the soma to evoke a single

action potential.

46. Sather, W., Dieudonne, S., MacDonald, J.F. & Ascher, P. Activation and desensitizationof N-methyl-D-aspartate receptors in nucleated outside-out patches from mouse neu-rones. J. Physiol. (Lond.) 450, 643–672 (1992).

47. Hodgkin, A.L. & Huxley, A.F. A quantitative description of membrane current and itsapplication to conduction and excitation in nerve. J. Physiol. (Lond.) 117, 500–544(1952).

48. Hines, M.L. & Carnevale, N.T. The NEURON simulation environment. Neural Comput. 9,1179–1209 (1997).

49. Mainen, Z.F. & Sejnowski, T.J. Influence of dendritic structure on firing pattern in modelneocortical neurons. Nature 382, 363–366 (1996).

50. Engel, D. & Jonas, P. Presynaptic action potential amplification by voltage-gated Na+

channels in hippocampal mossy fiber boutons. Neuron 45, 405–417 (2005).

NATURE NEUROSCIENCE doi:10.1038/nn.2359

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Ca2+ and calmodulin initiate all forms of endocytosisduring depolarization at a nerve terminal

Xin-Sheng Wu1,3, Benjamin D McNeil1,3, Jianhua Xu1, Junmei Fan1, Lei Xue1, Ernestina Melicoff2,Roberto Adachi2, Li Bai1 & Ling-Gang Wu1

Although endocytosis maintains synaptic transmission, how endocytosis is initiated is unclear. We found that calcium influx

initiated all forms of endocytosis at a single nerve terminal in rodents, including clathrin-dependent slow endocytosis, bulk

endocytosis, rapid endocytosis and endocytosis overshoot (excess endocytosis), with each being evoked with a correspondingly

higher calcium threshold. As calcium influx increased, endocytosis gradually switched from very slow endocytosis to slow

endocytosis to bulk endocytosis to rapid endocytosis and to endocytosis overshoot. The calcium-induced endocytosis rate increase

was a result of the speeding up of membrane invagination and fission. Pharmacological experiments suggested that the calcium

sensor mediating these forms of endocytosis is calmodulin. In addition to its role in recycling vesicles, calcium/calmodulin-

initiated endocytosis facilitated vesicle mobilization to the readily releasable pool, probably by clearing fused vesicle membrane

at release sites. Our findings provide a unifying mechanism for the initiation of various forms of endocytosis that are critical in

maintaining exocytosis.

After vesicle exocytosis, endocytosis generates new vesicles, whichreplenishes the vesicle pool and maintains exocytosis1. Five differentkinetic forms of endocytosis have been reported. First, slow endo-cytosis, which takes tens of seconds, has been observed widely atsynapses and non-neuronal secretory cells2–7. It is mediated by aclassical, clathrin-dependent mechanism5,8–10. Second, rapid endo-cytosis, which takes a few seconds, has been observed at ribbon-typeand calyx-type synapses2,6, but its existence is debated at small synapses1.Rapid endocytosis may be clathrin independent in chromaffin cellsand goldfish ribbon-type synapses8,10. Third, bulk endocytosis, whichforms large endosome-like structures by retrieving large pieces ofmembrane, has been seen at many synapses after strong stimulation(for examples, see refs. 11,12). Fourth, endocytosis overshoot, whichretrieves more membrane than was exocytosed by the stimulation,has been observed in non-neuronal secretory cells13 and calyx-typesynapses14. The mechanisms mediating bulk endocytosis and endo-cytosis overshoot are largely unclear. Finally, very slow or absentendocytosis has been observed15,16.

Because of these various forms, endocytosis in different conditionsdiffers markedly in speed, amount and vesicle size. The mechanismsthat initiate these different forms and rates of endocytosis are unclear. Ithas been proposed that the same mechanism underlies rapid and slowrates, with individual events occurring stochastically17. Another view isthat different rates depend on the ratio between simultaneouslyoccurring rapid and slow endocytosis pathways2,4,6,18, although whatdetermines this ratio is somewhat controversial. For example, calciumfacilitates rapid endocytosis at calyceal synapses6, but was reported to

facilitate or inhibit rapid endocytosis in different studies at ribbonsynapses2,4,18. It has been suggested that calcium regulates the endo-cytic rate at many nerve terminals3,19–25. Whether calcium initiatesendocytosis remains unclear, as quantitative measurements did notreveal a complete block of endocytosis when calcium influx wasreduced3,25. Because of the difficulty in identifying the endocytictrigger, it is often assumed that endocytosis follows exocytosis auto-matically, perhaps through molecular coupling.

Wee studied the mechanisms that initiate various forms of endo-cytosis with capacitance measurements at a large nerve terminal, thecalyx of Held. We found that calcium influx initiated slow endocytosis,bulk endocytosis, rapid endocytosis and endocytosis overshoot withincreasingly higher thresholds. Pharmacological experiments suggestedthat the calcium sensor mediating these forms of endocytosis wascalmodulin. Our results may explain how exocytosis in single nerveterminals is maintained by various forms of endocytosis in variousphysiological conditions.

RESULTS

Calcium influx initiates slow endocytosis

Similar to trains of action potential-like stimuli6,14, a 20-ms depolar-ization (from �80 to +10 mV, unless indicated otherwise) at the calyxinduced slow endocytosis with a time constant (t) of 15.6 ± 2.1 s and aninitial endocytosis rate (rateendo) of 39.2 ± 7.1 fF s�1 (n¼ 12; Fig. 1a) incontrol conditions, in which the extracellular calcium concentration([Ca2+]o) was 2 mM and the pipette contained 50 mM BAPTA.Decreasing the [Ca2+]o from 5.5 to 0.5 mM decreased the calcium

Received 24 April; accepted 1 June; published online 26 July 2009; doi:10.1038/nn.2355

1National Institute of Neurological Disorders and Stroke, Bethesda, Maryland, USA. 2Department of Pulmonary Medicine, The University of Texas M.D. Anderson CancerCenter, Houston, Texas, USA. 3These authors contributed equally to this work. Correspondence should be addressed to L.-G.W. (wul@ninds.nih.gov).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1003

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

current charge (QICa) and the rateendo induced by the 20-ms depolar-ization (Fig. 1a). At a [Ca2+]o of 0.25 mM, the capacitance jumpinduced by a 20-ms depolarization was often too small to observe.However, 5–10 pulses of 10–20-ms depolarization at 0.5–1 Hz induceda clear capacitance jump (363 ± 114 fF, n ¼ 4), which was followed by arateendo of 2.0 ± 0.8 fF s�1 (n ¼ 4; Fig. 1b). This rate was B54-fold lessthan that induced after the same stimulus at a [Ca2+]o of 2 mM (107 ±17 fF s�1, n ¼ 4, P o 0.01; Fig. 1b) and B20–49-fold less than thatafter a 20-ms depolarization at a [Ca2+]o of 2–5.5 mM (Fig. 1a). Asimilarly low rateendo was observed after a prolonged depolarizationthat induced no detectable Ca2+ current (Supplementary Fig. 1).

The slow calcium chelator EGTA (10 mM in the pipette) reduced therateendo after a 20-ms depolarization B4.9-fold (EGTA: 8.0 ± 1.2 fF s�1,n ¼ 18; control: 39.2 ± 7.1 fF s�1, n ¼ 12; P o 0.01; Fig. 1c). With thefast calcium buffer BAPTA (10 mM in the pipette), the 20-msdepolarization induced no detectable capacitance jump. However,10–20 pulses of 20-ms depolarization at 10 Hz induced a jump of288 ± 32 fF (n ¼ 10), followed by a rateendo of 0.2 ± 0.5 fF s�1 (n ¼ 10;Fig. 1d). This rate (mean) wasB1,440-fold less than that after the samestimulus in control (50 mM BAPTA: 288 ± 31 fF s�1, n ¼ 14; P o 0.01;Fig. 1d), and B196-fold less than that induced by a 20 ms depolariza-tion in control (39.2 ± 7.1 fF s�1, n ¼ 12). Thus, 10 mM BAPTA

a 2 mM Ca

ICa

Cm

200 fF10 s

2 nA20 ms

5.5 mM Ca 0.5 mM Ca

b 2 mM Ca

500 fF

10 s

Control

200 fF

10 s

10 mM EGTA

Control

500 fF

10 s

10 mM BAPTA

100 fF

10 s

0.25 mM Ca

200 fF

10 s

DNFDPF

200 fF

10 s

c

e

150

100

50

0Rat

e endo

(fF

s–1

)

9060300QICa (pC)

1218

80.5 mM2 mM5.5 mM

d

Figure 1 Calcium influx triggers endocytosis.

(a) The calcium channel current (ICa, upper) and

the capacitance change (Cm, lower) induced by a

20-ms depolarization at an [Ca2+]o of 2, 5.5 or

0.5 mM. In all figures, the stimulus was applied

before the capacitance jump; sampled traces are

mostly single traces and occasionally an average

of 2–3 traces. A summary is also shown (right,mean ± s.e., n in labels applies to other similar

figures). (b) Sampled Cm induced by 5–10 pulses

of 10–20-ms depolarization at 0.5–1 Hz with

[Ca2+]o ¼ 2 or 0.25 mM. (c) Sampled Cm induced

by a 20-ms depolarization in control (50 mM

BAPTA) or with 10 mM EGTA. (d) Sampled Cm

induced by 20 pulses of 20-ms depolarization at

10 Hz in control or with 10 mM BAPTA. To reduce

the capacitance noise during prolonged

recordings, we used the software X-chart (HEKA)

to average the capacitance value every B0.3 s

(applies only to this panel). (e) Sampled Cm

induced by a 20-ms depolarization in the

presence of DPF (1 mM) or DNF (1 mM).

Figure 2 Very slow endocytosis at a [Ca2+]i of

0.5–0.75 mM. (a) Sampled mEPSCs (left) and the

corresponding presynaptic Cm (right) from three

synapses (red, blue and black) at 0.5 mM

presynaptic [Ca2+]i. (b) The accumulated number

of mEPSCs (NmEPSC, left, circles) and Cm (right,

triangles) plotted versus the time after presynaptic

break in (time 0) with a [Ca2+]i of 0.5 mM. Data

were divided into three groups corresponding to

mEPSC frequencies of 0–5 Hz (32 synapses,black), 5–50 Hz (89 synapses, blue) and 450 Hz

(14 synapses, red). The Cm in calyces dialyzed

with 0 calcium and 0.5 mM BoNT/C is also shown

(eight synapses, brown triangles). Color codes

apply to a–c. (c) The numbers of vesicles (Nves),

calculated from the mEPSC and Cm traces in b,

are plotted versus the time after presynaptic break

in (NmEPSC, circles; NCm, triangles). To obtain

NCm, we corrected the Cm traces in b for the

baseline drift (subtracting the mean of the Cm

trace obtained in 0.5 mM BoNT/C (brown triangles

in b) from each Cm trace and divided by the mean

vesicle’s capacitance, which is 65 aF29). The

black curve is the predicted NCm based on the

mean NmEPSC with an endocytosis t of B600 s

(see Supplementary Data 1). (d) Sampled

mEPSC (left) and Cm (right) from two synapses,

one with 0.75 mM calcium (red) and the otherwith 0.75 mM calcium and 0.5 mM BoNT/C in the

calyx (black). (e) NmEPSC (left, circles) and Cm (right, triangles) plotted versus the time after presynaptic break-in with a pipette containing 0.75 mM calcium

(n ¼ 14, red) or 0.75 mM calcium with 0.5 mM BoNT/C (n ¼ 15, black). (f) Nves, calculated from NmEPSC and NCm, is plotted in the presence of 0.75 mM

calcium. To obtain NCm, we corrected the Cm traces in e for the baseline drift (brown triangles in b) and divided by 65 aF. Color codes in e apply to f. The

predicted NCm (black curve) was obtained as described in c with an endocytosis t of 600 s.

d

BoNT/C

mEPSC Cm

30 pA

200 ms

500 fF

40 s

60

30

0

6003000Time (s)

Ca2+ + BoNT/C

Ca2+ 2

1

0

Cm

(pF

)

6003000Time (s)

60

40

20

0

6003000Time after break-in (s)

–

e

f

CmmEPSC

30 pA

200 ms

500 fF

40 s

[Ca2+]i = 0.5 M [Ca2+]i = 0.75 M

2

1

0

Cm

(pF

)

3001500Time (s)

30

15

0

3001500Time after break-in (s)

NmEPSCNCm

– NCm predicted

NmEPSCNCm

NCm predicted

a

b

c

30

15

0

Nm

EP

SC (

×1,0

00)

Nve

s (×

1,00

0)

Nve

s (×

1,00

0)N

mE

PS

C (

×1,0

00)

3001500Time (s)

>50 Hz 5–50 Hz<5 Hz BoNT/C

1004 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

essentially abolished endocytosis. The same result was observed in P13–14 rats with more mature calyces (n ¼ 7, data not shown)26. Theseresults suggest that calcium influx is required for endocytosis.

Two pieces of evidence suggest that the slow endocytosis that weobserved is clathrin dependent, which would be consistent with studiesat other synapses5,9,10. First, the proline-rich domain peptide pp11,which interrupts the interaction between amphiphysin and dynamin,an important step in clathrin-dependent endocytosis9,10, blocked slowendocytosis27,28. Second, a 12mer containing the DNF motif ofamphiphysin I (1 mM), which disrupts the interaction betweenamphiphysin and the AP2 adaptor complex10, reduced rateendo aftera 20-ms depolarization to 43 ± 12% (n ¼ 7, P o 0.01), as comparedwith its control peptide (DPF, 1 mM, 100 ± 23%, n ¼ 8, rateendo

normalized to the mean of control; Fig. 1e).Our results (Fig. 1) predicted a low rateendo at low intracellular

calcium concentrations ([Ca2+]i). We tested this prediction by dialyzing0.5 mM calcium into the calyx via the whole-cell pipette and simulta-neously monitoring the calyx membrane capacitance and the miniatureexcitatory postsynaptic currents (mEPSCs) from the same synapses(Fig. 2a)29. We divided synapses into three groups corresponding tomEPSC frequencies of o5 Hz, 5–50 Hz and 450 Hz (mean ¼ 22.9 ±1.9 Hz, n ¼ 135 synapses) and found that higher mEPSC frequencieswere accompanied by larger capacitance increases (Fig. 2a–c).

Considering that a single vesicle’s capacitance is 0.065 fF29, a mEPSCfrequency of o5 Hz (mean ¼ 2.3 ± 1.4 Hz, n ¼ 32 synapses) predicteda negligible capacitance increase in 330 s of recordings (2.3 Hz � 330 s� 0.065 fF ¼ 49 fF). Thus, the capacitance change for the group witho5 Hz of mEPSCs largely reflected baseline drift (Fig. 2b). Consistentwith this, a similar drift occurred in calyces dialyzed with pipette

solution containing 0 calcium and with botu-linum neurotoxin C (BoNT/C, 0.5 mM;Fig. 2b), which blocked exocytosis6. Thisdrift implies that whole-cell break-in influ-ences the normal membrane equilibrium on avery slow time scale. By correcting the capa-citance drift and converting the capacitanceincrease into vesicle number (NCm; seeFig. 2c), we found that NCm approximately

matched the accumulated mEPSC number (NmEPSC), particularlyduring the first 200 s of calcium dialysis (P 4 0.2; Fig. 2c).

Compared with dialysis of 0.5 mM calcium, dialysis of 0.75 mM calciumcaused a higher capacitance increase and mEPSC frequency (Fig. 2d–f),both of which were blocked by 0.5 mM BoNT/C (n ¼ 15; Fig. 2d,e).Similar to 0.5 mM calcium, baseline-corrected NCm mostly overlappedwith NmEPSC during the first 200 s of calcium dialysis (n ¼ 14 synapses;Fig. 2f). This close overlap suggests that endocytosis was very slow (seeSupplementary Data 1 for additional discussion). After 200 s of calciumdialysis, the NmEPSC continued to increase, whereas the NCm approacheda plateau level at B600 s of dialysis of 0.75 mM calcium (Fig. 2f),suggesting that very slow endocytosis was able to balance further exo-cytosis after the initial rise in capacitance. The measured mean NmEPSC

with a mean endocytosis t ofB600 s (B400–800 s) could well predict theobserved NCm and thus account for the difference between NCm andNmEPSC (Fig. 2c,f and Supplementary Fig. 2; see Supplementary Data 1for details). Simulations predict that a very slow endocytosis with a t of600 s will balance out continuous exocytosis in 1,200–1,800 s during awide range of exocytosis frequencies (Supplementary Fig. 3, see Supple-mentary Data 1 for details). The rateendo decrease that we observed wasindependent of the amount of exocytosis (for detail, see SupplementaryData 1). Given that the amount of calcium influx determined whetherefficient endocytosis occurred or not, we concluded that calcium influx isrequired to initiate endocytosis (see the Discussion).

Calcium influx regulates the endocytosis rate and amount

We previously showed that ten pulses of 20-ms depolarization deliveredat 10 Hz activates rapid endocytosis and thus increased the rateendo

6. Tosystematically study how increasing calcium influx affects endocytosis,

50 ms(5.5 mM Ca)

3n A300 ms

500 fF10 s

50 ms20 ms5 ms

ICa

Cm

2 ms

30

20

10

06003000

QICa (pC)

4

2

06003000

QICa (pC)

150

100

50

0

Am

p endo

(%)

Am

p endo

(%)

6003000QICa (pC)

10 mM EGTA50 ms(5.5 mM Ca)

Cm

500 fF10 s

70 mM EGTA10 mM EGTA

500 fF10 s

Control

Cm

500 pulses

10 mM EGTA

500 pulses200 pulses20 pulses

Cm

500 fF10 s

10

5

09004500

1.0

0.5

0.09004500

QICa (pC) QICa (pC)

150

100

50

09004500

QICa (pC)

a

b

c

e

f

d

Rat

e endo

(fF

s–1

)R

ate en

do (f

F s

–1)

500

06003000

QICa (pC)

SumRapidSlow

1,000

09004500

QICa (pC)

SumRapidSlow

2,000

1,000

� rap

id (

s)

� slo

w (s

)

� rap

id (

s)

� slo

w (s

)

Figure 3 Calcium influx increases rateendo and

triggers rapid endocytosis and endocytosis

overshoot. (a) ICa (upper) and Cm (lower) induced

by ten pulses of 2-, 5-, 20- and 50-ms

depolarization at 10 Hz in a [Ca2+]o of 2 mM and

ten pulses of 50-ms depolarization at 10 Hz in a

[Ca2+]o of 5.5 mM (n ¼ 7–17). (b) Rateendo, the

endocytosis amplitude (ampendo, normalized to thecapacitance jump) and t are plotted against QICa

induced by the stimuli listed in a (QICa increased

in the order of stimuli listed). The rateendo (left)

and ampendo (middle left) of the rapid and slow

component of endocytosis and their sum are

shown. For t, the rapid (trapid, middle right) and

the slow (tslow, right) components are shown

separately. (c) Sampled Cm induced by ten pulses

of 50-ms depolarization in control in 10 or

70 mM EGTA. (d) Sampled Cm induced by ten

pulses of 50-ms depolarization in 10 mM EGTA

with a [Ca2+]o of 5.5 mM. (e) At 34–36 1C, Cm

was induced by 20, 200 and 500 pulses of

0.5-ms depolarization at 200 Hz and by 500

pulses with 10 mM EGTA. (f) Similar to b but with

20, 200 and 500 pulses of 0.5-ms depolarization

at 100–300 Hz at 34–36 1C (n ¼ 5–6).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1005

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

we applied ten pulses of 2–50-ms depolarization at 10 Hz at a [Ca2+]o

of 2 mM and ten pulses of 50-ms depolarization at a [Ca2+]o of5.5 mM. Except for ten pulses of 2-ms depolarization, which inducedonly slow endocytosis, endocytosis after ten pulse stimuli wasbi-exponential with a rapid (t, B1–4 s) and a slow component(t, B12–23 s) (Fig. 3a,b)6. Notably, we found that as QICa increased,rateendo and endocytosis overshoot gradually increased (Fig. 3a,b),with 480% of the rateendo increase being contributed by the rapidcomponent, as a result of an increase in its amplitude and a decrease inits t, from B4 s to B1.5 s (Fig. 3a,b). The rate of the slow componentwas also increased by calcium influx, though its contribution to theincreased rateendo was minimal (Fig. 3b and Supplementary Fig. 4 andSupplementary Data 2). Rateendo could increase to 1,138 ± 172 fF s�1

(n ¼ 7; Fig. 3a,b), corresponding to B32 vesicles per s per activezone, considering that a calyx contains B550active zones and a vesicle’s capacitance is65 aF29,30. Thus, differing from the generalassumption, rapid endocytosis does not have

a fixed speed or time constant but can be regulated in speed, timeconstant and relative contribution by calcium influx.

The addition of EGTA decreased the rateendo after ten pulses of50-ms depolarization in a concentration-dependent manner (control:830 ± 177 fF s�1, n ¼ 14; 10 mM EGTA: 184 ± 40 fF s�1, n ¼ 5; 70 mMEGTA: 74 ± 25 fF s�1, n ¼ 3; Fig. 3c). The decrease was not a result ofchanges in the capacitance jump (for details, see ref. 6). EGTA (10–70mM) also abolished the overshoot induced by ten pulses of 50-msdepolarization at a [Ca2+]o of 5.5 mM (n ¼ 18, P o 0.01; Fig. 3d),which induced an overshoot with an amplitude 60% of the capacitancejump in control (50 mM BAPTA; Fig. 3a,b). We obtained similar resultsat near body temperature (34–36 1C) with trains of 0.5-ms depolari-zation to +10 mV that mimicked action potential trains in vivo31

(Fig. 3e,f). We concluded that increasing calcium influx increasedrateendo by initiating rapid endocytosis and endocytosis overshoot.

Calcium triggers bulk endocytosis and speeds up fission

We previously found that bulk endocytosis is reflected as a briefdownward capacitance shift (DCS) of B20–500 fF32, the frequency ofwhich increases with stronger stimulation (Supplementary Data 2)32.Consistent with these results, ten pulses of 50-ms depolarization at10 Hz at a [Ca2+]o of 5.5 mM significantly increased the DCS frequencywithin 10 s of stimulation (0.14 ± 0.05 Hz, n ¼ 12 calyces, P o 0.01;Fig. 4), indicating that rapid bulk endocytosis was occurring32. Thisfrequency increase was much higher than that (o0.03 Hz)32 observedafter ten pulses of 20-ms depolarization with a [Ca2+]o of 2 mM, andwas reduced to near baseline level by adding 70 mM EGTA in the pipette(P ¼ 0.02, n ¼ 15 calyces; Fig. 4a–c). The DCS frequency change was

300 fF

1 s

Control

Cm

70 mM EGTA

Cm 300 fF

1 s

Cm

Gp

100 fF

500 ms

1 nS

100 fF

1 nS

50 ms

Cm

Gp

200

100

0

Gp

rate

(nS

s–1

)

*

Before After

*

ControlEGTA

0.2

0.1

0.0

DC

S fr

eq (

Hz)

40200Time after stim (s)

ControlEGTA

a

c d

b

–20

Figure 4 Calcium influx initiates bulk endocytosis and speeds up the

fission pore closure rate. (a,b) Sampled Cm induced by ten pulses of 50-ms

depolarization at 10 Hz in a [Ca2+]o of 5.5 mM in control conditions (a) and

in 70 mM EGTA (b). For each trace, the DCSs (arrows) are enlarged and

shown with the calculated Gp (right). (c) Frequency of DCSs, binned every

10 s, plotted in control conditions (n ¼ 12 calyces) and in 70 mM EGTA

(n ¼ 15 calyces). At time 0, ten pulses of 50-ms depolarization at 10 Hz

were applied. The [Ca2+]o was 5.5 mM. (d) The rate of the Gp decreaseduring DCSs before and within 10 s of stimulation in control conditions and

in 70 mM EGTA. The stimulation was the same as that described in a–c.

CBD

200 fF

10 s

CBD-c

MLCKMLCK-c

500 fF

10 s

CBD-c CBD

MLCKMLCK-c

a

CalmControl CalmControl

Control

500 fF

10 s

Calmc200 fF

50 ms

0.2

0.1

0.0

DC

S fr

eq (

Hz)

40200Time after stim (s)

ControlCalm

b

d

–20

Figure 5 Calmodulin blockers inhibit endocytosis.

(a) Top, sampled Cm induced by a 20-ms

depolarization in the presence of scrambled CBD

(CBD-c, 500 mM, left) or CBD (500 mM, right).

Middle, mutated MLCK peptide (MLCK-c, 20 mM)or MLCK (20 mM). Bottom, 0.1% DMSO (control)

or calmidazolium (Calm, 10 mM with 0.1%

DMSO). (b) Data are presented as in a but with

Cm induced by ten pulses of 20-ms depolarization

at 10 Hz. (c) Sampled Cm induced by ten pulses

of 50-ms depolarization at 10 Hz at a [Ca2+]o of

5.5 mM in control conditions (0.1% DMSO, left)

or in the presence of 20 mM calmidazolium

(with 0.1% DMSO, right). The arrow denotes

the DCS from the left sampled trace on a

different scale (middle). (d) The frequency of

DCSs (right) induced by ten pulses of 50-ms

depolarization at 10 Hz at a [Ca2+]o of 5.5 mM

in control conditions (0.1% DMSO) or with

20 mM calmidazolium.

1006 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

not accompanied by significant changes in the DCS size (P 4 0.6, seealso ref. 32). Although 70 mM EGTA reduced the capacitance jumpto B57% of the control value (EGTA: 0.82 ± 0.09 pF, n ¼ 15; control:1.45 ± 0.15 pF, n ¼ 12), the drop could not account for the nearlyfourfold decrease in the DCS frequency (Fig. 4c). These results suggestthat calcium influx initiates bulk endocytosis. Two pieces of evidencefurther support this suggestion. First, at a [Ca2+]i of 0.5–0.75 mM, atwhich endocytosis was very slow (Fig. 2), the DCS was rarely observed(n ¼ 17, data not shown). Second, calmidazolium, an inhibitor of thecalcium-binding protein calmodulin, essentially abolished the stimula-tion-induced increase of the DCS frequency but decreased the capaci-tance jump by less than 30%. Given that a 20-ms depolarization couldinduce DCSs32, but not rapid endocytosis (Fig. 1a), the QICa requiredto initiate bulk endocytosis must be lower than that required forrapid endocytosis.

During DCSs within 10 s of stimulation, the rate of the fission poreconductance (Gp) decrease (20–80%) was 182 ± 45 nS s�1 (n ¼ 21DCSs, Fig. 4a, see also Supplementary Fig. 5), corresponding to adecrease of the fission pore diameter at B105 nm s�1. This rate wasmuch higher than that before (P o 0.05; Fig. 4d) or 420 s afterstimulation (data not shown) and was significantly reduced by 70 mMEGTA (P o 0.01; Fig. 4a,b,d), suggesting that calcium influx increasesthe fission pore closure rate.

Calmodulin involvement in endocytosis

It has been postulated that the calcium-binding proteins calmodulinand synaptotagmin (Syt) regulate endocytosis5,33. To test this, wemeasured endocytosis at B4–10 min after break in with a pipettecontaining calmodulin inhibitors: the calmodulin binding–domainpeptide (CBD, consisting of CAM kinase II residues 290–309,300–500 mM), the myosin light chain kinase peptide (MLCK, 20 mM)or the organic small molecule inhibitor calmidazolium (10 mM in0.1% DMSO)34. Compared with controls (scrambled CBD, mutatedMLCK or 0.1% DMSO, respectively), these three inhibitors reducedthe rateendo by 475% (P o 0.01; Fig. 5a), but did not significantlyreduce the capacitance jump (P 4 0.05)after a 20-ms depolarization (scrambledCBD: 100 ± 6%, n ¼ 12; CBD: 13 ± 9%,n ¼ 14; mutated MLCK: 100 ± 15%, n ¼ 7;MLCK: 22 ± 9%, n ¼ 11; 0.1% DMSO:100 ± 15%, n ¼ 12; calmidazolium: 10 ±5%, n ¼ 8; rateendo were normalized to themean of each control group, respectively).A similar large reduction of rateendo wasfound in the presence of these inhibitorsafter ten pulses of 20-ms depolarizationat 10 Hz (scrambled CBD: 100 ± 23%,n ¼ 10; CBD: 20 ± 4%, n ¼ 11; mutatedMLCK: 100 ± 14%, n ¼ 7; MLCK: 18 ± 5%,n ¼ 12; 0.1% DMSO: 100 ± 19%, n ¼ 12;calmidazolium: 21 ± 7%, n ¼ 8; all P o 0.01;

Fig. 5b). These inhibitors reduced the QICa by B0–22% (Supple-mentary Fig. 6) and the capacitance jump induced by the ten pulsestimulus by o30%, which could not account for the observed 475%decrease in rateendo (see also Supplementary Data 1 and 2). Further-more, calmidazolium essentially blocked overshoot (0.1% DMSO:960 ± 90 fF overshoot, n ¼ 20; calmidazolium: 280 ± 110 fF under-shoot, n ¼ 19; P o 0.01; endocytosis was not complete in the pre-sence of calmidazolium) and significantly reduced the DCS frequencyafter ten pulses of 50-ms depolarization at a [Ca2+]o of 5.5 mM(P o 0.01; Fig. 5c,d). These results suggest that calmodulin maymediate calcium-triggered slow and rapid endocytosis, endocytosisovershoot and bulk endocytosis.

The calmodulin proteins involved in endocytosis are probablyimmobile and membrane-bound, as endocytosis persisted duringwhole-cell dialysis of a control solution containing no calmodulin forlonger than 10 min28. Furthermore, whole-cell dialysis of wild-typecalmodulin protein (100–200 mM) did not significantly change therateendo induced by ten pulses of 20-ms depolarization, as measured5–10 min after whole-cell break in (calmodulin: 458 ± 111 fF s�1, n¼ 6;control: 394 ± 58 fF s�1, n ¼ 10; P ¼ 0.62).

Calyces do not contain Syt1, but do express the closely related Syt2,and genetic deletion of Syt2 abolishes synchronized release in calyces35.However, we were able to induce less synchronized release by a 20-msdepolarization (Fig. 6a) and ten pulses of 20-ms depolarization at10 Hz (Fig. 6b). The rateendo after these two stimuli in Syt2�/� micewere similar to those from Syt2+/+ littermates (Fig. 6), suggesting thatSyt2 is not critical for calcium-triggered endocytosis (20-ms pulse:Syt2+/+, 100 ± 10%, n ¼ 6; Syt2�/�, 85 ± 15%, n ¼ 7; P ¼ 0.48; tenpulse train: Syt2+/+, 100 ± 17%, n ¼ 10; Syt2�/�, 109 ± 25%, n ¼ 11;P ¼ 0.75; data were normalized to the mean in Syt2+/+ group).

Endocytosis speeds up vesicle mobilization

Calcium/calmodulin speeded up both the rateendo (Figs. 1–5) andvesicle recruitment to the readily releasable pool (RRP)34, whichimplies that faster endocytosis may facilitate RRP replenishment. Totest this possibility, we monitored the recovery of the RRP during tenpulses of 10–20 ms depolarization at 10 Hz, where each pulse depletedthe RRP36. Consistent with previous observations34, the capacitancejump evoked by the second to the tenth depolarizing pulse, whichreflected the RRP replenishment rate, was reduced by including CBD(300–500 mM), MLCK (20 mM, data not shown) or calmidazolium(10 mM) in the pipette (P o 0.01; Fig. 7a,b). We observed a similarreduction with dynasore (100 mM, P o 0.01; Fig. 7a,b), a dynamininhibitor that inhibits rapid and slow endocytosis at calyces28. Thedecrease in the capacitance jump in the presence of CBD, calmidazolium

Syt2–/–

200 fF

10 s

Syt2+/+ mice Syt2–/–Syt2+/+ mice

500 fF

10 s

a b

Figure 6 Syt2 is not critical in mediating endocytosis. (a) Sampled Cm

induced by a 20-ms depolarization in a Syt2�/� mouse or a Syt2+/+

littermate. (b) Sampled Cm induced by ten pulses of 20-ms depolarizationat 10 Hz in a Syt2�/� mouse or a Syt2+/+ littermate.

200 msCm

ControlCBDCalmDyn

400

300

200

100

01050

100

50

0

∆Cm

(%

)

Σ∆C

m (

%)

1050Stim number Stim number

a b

Figure 7 CBD, calmidazolium and dynasore slow down the RRP replenishment. (a) Sampled Cm induced

by ten pulses of 20-ms depolarization at 10 Hz in control conditions (0.1% DMSO), in the presenceof 500 mM CBD, 10 mM calmidazolium or 100 mM dynasore. The Cm jump induced by the first 20-ms

depolarization was normalized. (b) The capacitance jump (DCm) and the accumulated capacitance jump

(P

DCm) induced by each 20-ms depolarization during 10 such pulses at 10 Hz in the conditions described

in a (with same color code as in a). Data were normalized to the capacitance jump induced by the first

20-ms depolarization (control (0.1% DMSO), n ¼ 12; CBD, n ¼ 11; Calm, n ¼ 8; dynasore, n ¼ 5).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1007

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

or dynasore was not a result of a change in the QICa (SupplementaryData 2). Thus, as with dynasore, CBD and calmidazolium may inhibitthe RRP replenishment by inhibiting endocytosis (see also Supplemen-tary Data 2).

We observed an inhibition of RRP replenishment 100 ms after thefirst pulse; that is, at the second pulse during the ten pulse train. Thecapacitance jumps induced by the second pulse were 24.6 ± 3.4%(n ¼ 8), 28.4 ± 2.1% (n ¼ 11) and 30.2 ± 2.4% (n ¼ 5) in the presenceof calmidazolium, CBD and dynasore, respectively, all of which weresignificantly lower than that of the control (45.3 ± 3.4%, n ¼ 12,P o 0.01; Fig. 7a,b). Vesicle depletion arising from the block inendocytosis could not account for this difference, as endocytosisretrieves only 0.6% of newly exocytosed vesicles within 100 ms incontrol conditions (calculated from the measured mean endocytosis tof B15.6 s after a 20-ms depolarization). Thus, inhibition of the RRPreplenishment was a result of a block of an endocytic mechanism afterfusion but before vesicle fission (see Discussion).

Endocytosis overshoot retrieves fused vesicle membrane

Vesicles stranded in low calcium conditions (Figs. 1d and 2) mayprovide the substrate for endocytosis overshoot. This possibility isdifficult to test directly, as low calcium conditions were achieved byincluding higher concentrations of calcium buffers that precluded theendocytosis overshoot (Fig. 3a,d). To overcome this difficulty, weaccumulated vesicles at the plasma membrane by blocking endocytosiswith GDPbS (0.3 mM in the pipette instead of GTP), which inhibitsdynamin function28. A 20-ms depolarization applied every 20 s for10–20 repetitions increased the membrane capacitance by 3,278 ± 612fF (n ¼ 5), after which depolarization-induced endocytosis recoveredas a result of activation of GTP-independent endocytosis (Fig. 8a, seeref. 28 for details). When applied after this treatment, ten pulses of50-ms depolarization at 10 Hz induced an overshoot (2,492 ± 650 fF,n ¼ 5 calyces; Fig. 8a) that was much larger than that induced incontrols (254 ± 61 fF, n ¼ 32, P o 0.03; Fig. 3b) but was similar to themembrane accumulation (3,278 ± 612 fF, n ¼ 5) observed before theten pulse stimulus.

The overshoot amplitude decreased to nearly 0 as the ten pulse trainwas repeated 2–4 times (n ¼ 3; Fig. 8a), indicating that the membranepool retrievable by the overshoot was limited. A similar decrease in theovershoot was observed when the ten pulse train was repeated in calycescontaining no GDPbS but showing an overshoot in response to thefirst ten pulse train (n ¼ 12; Fig. 8b). In the presence of GDPbS, therewas no overshoot when the ten pulse train was applied withoutconditioning pulses, as a result of the initial endocytic block byGDPbS (n ¼ 6, data not shown)28. These results suggest that endo-cytosis overshoot retrieves a limited pool of vesicles stranded at theplasma membrane.

DISCUSSION

High calcium concentration transients initiate endocytosis

Two pieces of evidence suggest that the [Ca2+]i that initiates efficientendocytosis is high, probably a few or tens of micromolar. First, at a[Ca2+]i of r0.75 mM, endocytosis was very slow, with a t of B600 s(Fig. 2). Second, a fast (BAPTA), but not a slow (EGTA), calciumchelator abolished endocytosis (Fig. 1c,d), suggesting that a highcalcium concentration transient at the release site during depolariza-tion34 initiates endocytosis. EGTA probably reduced this calciumtransient, as it partially inhibited release (Fig. 1c)6,37, which explainswhy EGTA partially inhibited endocytosis (Figs. 1c and 3c–e). Ourresults suggest that the calcium threshold for initiating endocytosis ishigher than for exocytosis, as much less endocytosis than exocytosisoccurred in the presence of 10 mM BAPTA or a [Ca2+]i of 0.5–0.75 mM(Figs. 1d and 2). Strong calcium influx increased the rateendo by severalthousand-fold over the lowest rate (Figs. 1d and 3b,f), resulting in arateendo of up to B32 vesicles per s per active zone immediately afterstimulation (Fig. 3a,b,e). Thus, the same calcium influx must triggerboth exocytosis and endocytosis. Considering that endocytosis iscomposed of two steps, membrane invagination and fission1, thecontrol of rateendo (Figs. 1–3) and the fission pore closure (Fig. 4) bycalcium suggests that calcium influx initiates endocytosis and speedsup both membrane invagination and fission.

Our data suggest that calcium binding with calmodulin initiatesendocytosis. Calmodulin forms a physical complex with voltage-gatedcalcium channels38 and its N and C lobes may sense different calciumconcentrations38, allowing calmodulin to mediate multiple forms ofendocytosis. Because no endocytosis occurs when exocytosis isabolished by botulinum neurotoxins6,27, we conclude that fusedvesicle membrane is the substrate used by calcium/calmodulin toinitiate endocytosis.

Calcium influx initiates all forms of vesicle endocytosis

Our results are consistent with those of earlier studies supportingregulation of endocytosis by extracellular calcium in extremely intenseor nonphysiological conditions, such as prolonged high potassiumapplication in synaptosomes or a-latrotoxin application (see alsoSupplementary Discussion)19–24. Whether this finding can be extra-polated to regulation of endocytosis by calcium influx in physiologicalconditions is unclear. Furthermore, in these studies, endocytosis wasmeasured indirectly using dye uptake and/or electron microscopy witha single time point, which may not distinguish between endocytosisand recycling, between different endocytosis forms or between partialand full inhibition, and thus between regulation and initiation ofendocytosis. Our results resolve these issues at calyces.

We found that endocytosis was mostly abolished at a [Ca2+]o of0.25 mM, but not 0.5 mM (Fig. 1a,b), which may explain why

rateendo is only moderately decreased whenthe [Ca2+]o is reduced to B0.5–0.75 mMat hippocampal synapses3,25. Our findingthat high [Ca2+]i initiates endocytosis couldexplain the lack of correlation between asubmicromolar residual calcium increase andthe rateendo at neuromuscular junctions7.

Our observation that calcium/calmodulininitiated bulk endocytosis during more intensestimulation (Fig. 4, see also ref. 32) mayexplain why endosome-like structures areobserved after more intense stimulation atmany synapses (for example, refs. 11,12). Atribbon-type synapses, calcium influx, which

500 fF

50 sCm

GDPβS

30 s

500 fFCm

GTPa b

Figure 8 Endocytosis overshoot retrieves vesicles stranded at the plasma membrane. (a) The capacitance

responses to a 20-ms depolarization repeated 11 times every 20 s, followed by a ten pulse train (ten pulses

of 50-ms depolarization at 10 Hz, arrows) repeated every 1–2 min three times in a calyx dialyzed with

0.3 mM GDPbS. (b) The capacitance response to a ten pulse train (ten pulses of 50-ms depolarization

at 10 Hz, arrows) repeated every 1–2 min three times in a control calyx (0.3 mM GTP).

1008 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

is probably above tens of micromolar18, facilitates the occurrence ofrapid endocytosis4 (but see ref. 2). At calyces, rapid endocytosis occurson an increase of calcium influx6. Here, we found that calcium influxspeeded up endocytosis by increasing the contribution of rapid endo-cytosis and decreasing the t of rapid and slow endocytosis (Fig. 3b,f).Contrary to the general assumption, rapid endocytosis did not have afixed t or speed. Notably, the lowest detectable rateendo for rapid endo-cytosis approached the fastest rateendo for slow endocytosis (Fig. 3b,f).Although rapid and slow endocytosis are suggested to be clathrinindependent and clathrin dependent, respectively8,10, both are initiatedby calcium/calmodulin (Figs. 1–5) and require GTP hydrolysis anddynamin10,27,28. Thus, rapid and slow endocytosis may share at leastsome of the same endocytic machinery.

Endocytosis overshoot has been observed in endocrine cells13,16 andcalyces14. Although calcium influx is considered to be a candidate formediating overshoot, it has not been tested thoroughly with calciumbuffers. Furthermore, it is unclear whether the overshoot retrievesvesicle membrane. We found that large calcium influx during intensestimulation, including high-frequency action potential–like trains(Fig. 3e,f) that may occur in vivo26 triggered endocytosis overshoot,which retrieved vesicles stranded at the plasma membrane (Fig. 8).

Although very slow endocytosis in near resting conditions may resultin stranded vesicles at the plasma membrane, the rate of fusion andretrieval may reach a balance in a duration of B2–3-fold longer thanthe time constant of very slow endocytosis (Fig. 2f, see also Supple-mentary Data 1). Thus, very slow endocytosis may recycle vesicles andparticipate in maintaining the nerve terminal morphology. Strandedvesicles in near resting conditions (Fig. 2f) might explain whyvesicle proteins, such as synaptobrevin and synaptophysin, are presentat the plasma membrane of hippocampal boutons3,9,15. Retrieval ofstranded vesicles by endocytosis overshoot may prevent nerveterminal expansion and increase the vesicle cycling capacity, whichis needed to maintain exocytosis during intense firing. Thesefunctions are physiologically relevant, as some neurons, includingcalyces, may fire from 0 to hundreds of hertz spontaneously andduring stimulation39.

Vesicle recycling in resting conditions has been observed at hippo-campal synapses40–42. However, observation requires prolonged incu-bation (B15 min to 10 h) of fluorescent materials for vesicle uptake40

and endocytosis has been estimated at a vesicle per bouton per100 s41,42. These results are consistent with a very slow rate ofendocytosis, as we observed here, and/or a low rate of exocytosis,which was not estimated previously40–42.

Taken together, we found that calcium influx initiated all knownforms of endocytosis at calyces, including slow, clathrin-dependentendocytosis, bulk endocytosis, rapid endocytosis and endocytosisovershoot, each of which had a correspondingly higher calcium thresh-old. Increasing calcium influx increased rateendo by increasing theamplitude of rapid endocytosis (Figs. 1 and 3) and the rate of fissionpore closure (Fig. 4) and by decreasing the slow and rapid endocytosistime constants. These findings may explain various forms, rates andamounts of endocytosis reported under various conditions at manysynapses and non-neuronal secretory cells over the last three decades.We suggest that calcium influx is the unifying signal that initiates allforms of endocytosis in secretory cells, including nerve terminals.

Vesicle uptake of FM dyes has been observed after sucrose applica-tion, which presumably induces calcium-independent exocytosis43,44,implying that calcium-independent endocytosis follows calcium-independent exocytosis. This observation does not counter our con-clusion that calcium-initiated endocytosis follows calcium-triggeredexocytosis, as the existence of sucrose-induced calcium-independent

exocytosis does not overthrow the principle that calcium triggersexocytosis during depolarization.

Calmodulin is the calcium sensor of endocytosis

The results of our pharmacological experiments (Fig. 5) suggest thatcalmodulin is the calcium sensor for calcium-triggered endocytosis.This finding may apply beyond calyces because, as discussed above,calcium influx may initiate endocytosis at many synapses. In support ofthis possibility, calmodulin blockers inhibit rapid endocytosis atchromaffin cells33 (but see ref. 45) and calcineurin, a phosphataseactivated by calcium/calmodulin, may regulate bulk endocytosis insynaptosomes and cerebellar cultures22,23,46.

The incomplete block of endocytosis by calmodulin blockers (Fig. 5)probably reflects the inefficiency of these inhibitors in vivo, althoughadditional calcium sensors cannot be excluded. Calmodulin mightinitiate endocytosis by activating the phosphatase calcineurin, whichdephosphorylates endocytic proteins22,23,46. This possibility, althoughsupported by our preliminary results, is beyond the scope of the presentwork. Placing calcium/calmodulin at the initial step opens thepossibility of putting together the sequence of protein interactionsmediating endocytosis.

Our observation that Syt2 deletion did not affect endocytosis afterprolonged (20 ms) depolarization is somewhat surprising to us, as Syt1deletion slows endocytosis after action potential trains at hippocampalsynapses and Drosophila neuromuscular junctions5,47. The difference inthe stimulation protocol might explain this discrepancy. In Syt1 andSyt2 knockout animals, synchronous release is abolished, but actionpotential–evoked release still occurs asynchronously at a few to tens ofmilliseconds after calcium influx35,48. Thus, asynchronously releasedvesicles may be exposed to much lower calcium levels during fusionthan synchronously released vesicles. This would result in a slowedendocytosis, as the calcium influx controls the rateendo (Figs. 1–4). Inour study, most asynchronous release may have occurred duringprolonged 20-ms depolarization, during which calcium influx initiatesefficient endocytosis.

Endocytosis before fission facilitates RRP replenishment

Calcium/calmodulin facilitates the RRP replenishment34, which iscritical in maintaining release during repetitive firing26. Our resultssuggest that calcium/calmodulin facilitates the RRP replenishment viaan endocytic mechanism before fission (Fig. 7).

We propose that vesicle fusion disrupts the release site structure,interfering with subsequent vesicle priming and/or fusion. Calcium/calmodulin may facilitate the structural recovery, and thus the RRPreplenishment, by rapid transfer of vesicle membrane and proteinsfrom the release site to the endocytic zone.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSThis work was supported by the National Institute of Neurological Disorders andStroke Intramural Research Program of the US National Institutes of Health.

AUTHOR CONTRIBUTIONSX.-S.W. conducted the double-patch experiments for Figure 2 and many of thecapacitance recordings for other figures. B.D.M. initiated the project, designedand conducted capacitance experiments and helped write the paper. J.X. and J.F.conducted capacitance experiments. L.X. performed the simulations. E.M. and

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1009

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

R.A. provided the Syt2�/� mice. L.B. maintained the animal colonies and L.-G.W.supervised the project, designed experiments and wrote the paper.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Wu, L.G., Ryan, T.A. & Lagnado, L. Modes of vesicle retrieval at ribbon synapses, calyx-type synapses and small central synapses. J. Neurosci. 27, 11793–11802 (2007).

2. von Gersdorff, H. & Matthews, G. Inhibition of endocytosis by elevated internal calciumin a synaptic terminal. Nature 370, 652–655 (1994).

3. Sankaranarayanan, S. & Ryan, T.A. Calcium accelerates endocytosis of vSNAREs athippocampal synapses. Nat. Neurosci. 4, 129–136 (2001).

4. Neves, G., Gomis, A. & Lagnado, L. Calcium influx selects the fast mode of endocytosisin the synaptic terminal of retinal bipolar cells. Proc. Natl. Acad. Sci. USA 98,15282–15287 (2001).

5. Poskanzer, K.E., Fetter, R.D. & Davis, G.W. Discrete residues in the c(2)b domain ofsynaptotagmin I independently specify endocytic rate and synaptic vesicle size. Neuron50, 49–62 (2006).

6. Wu, W., Xu, J., Wu, X.S. & Wu, L.G. Activity-dependent acceleration of endocytosis at acentral synapse. J. Neurosci. 25, 11676–11683 (2005).

7. Wu, L.G. & Betz, W.J. Nerve activity, but not intracellular calcium, determines the timecourse of endocytosis at the frog neuromuscular junction. Neuron 17, 769–779 (1996).

8. Artalejo, C.R., Henley, J.R., McNiven, M.A. & Palfrey, H.C. Rapic endocytosis coupled toexocytosis in adrenal chromaffin cells involves Ca2+, GTP, and dynamin, but not clathrin.Proc. Natl. Acad. Sci. USA 92, 8328–8332 (1995).

9. Granseth, B., Odermatt, B., Royle, S.J. & Lagnado, L. Clathrin-mediated endocytosis isthe dominant mechanism of vesicle retrieval at hippocampal synapses. Neuron 51,773–786 (2006).

10. Jockusch, W.J., Praefcke, G.J., McMahon, H.T. & Lagnado, L. Clathrin-dependent andclathrin-independent retrieval of synaptic vesicles in retinal bipolar cells. Neuron 46,869–878 (2005).

11. Richards, D.A., Guatimosim, C. & Betz, W.J. Two endocytic recycling routes selectivelyfill two vesicle pools in frog motor nerve terminals. Neuron 27, 551–559 (2000).

12. Holt, M., Cooke, A., Wu, M.M. & Lagnado, L. Bulk membrane retrieval in the synapticterminal of retinal bipolar cells. J. Neurosci. 23, 1329–1339 (2003).

13. Thomas, P., Lee, A.K., Wong, J.G. & Almers, W. A triggered mechanism retrievesmembrane in seconds after Ca2+-stimulated exocytosis in single pituitary cells. J. CellBiol. 124, 667–675 (1994).

14. Renden, R. & von Gersdorff, H. Synaptic vesicle endocytosis at a CNS nerve terminal:faster kinetics at physiological temperatures and increased endocytotic capacity duringmaturation. J. Neurophysiol. 98, 3349–3359 (2007).

15. Gandhi, S.P. & Stevens, C.F. Three modes of synaptic vesicular recycling revealed bysingle-vesicle imaging. Nature 423, 607–613 (2003).

16. Smith, C. & Neher, E. Multiple forms of endocytosis in bovine adrenal chromaffin cells.J. Cell Biol. 139, 885–894 (1997).

17. Balaji, J. & Ryan, T.A. Single-vesicle imaging reveals that synaptic vesicle exocytosis andendocytosis are coupled by a single stochastic mode. Proc. Natl. Acad. Sci. USA 104,20576–20581 (2007).

18. Beutner, D., Voets, T., Neher, E. & Moser, T. Calcium dependence of exocytosis andendocytosis at the cochlear inner hair cell afferent synapse. Neuron 29, 681–690(2001).

19. Ceccarelli, B. & Hurlbut, W.P. Ca2+-dependent recycling of synaptic vesicles at the frogneuromuscular junction. J. Cell Biol. 87, 297–303 (1980).

20. Henkel, A.W. & Betz, W.J. Monitoring of black widow spider venom (BWSV) induced exo-and endocytosis in living frog motor nerve terminals with FM1–43. Neuropharmacology34, 1397–1406 (1995).

21. Ramaswami, M., Krishnan, K.S. & Kelly, R.B. Intermediates in synaptic vesicle recyclingrevealed by optical imaging of Drosophila neuromuscular junctions. Neuron 13,363–375 (1994).

22. Marks, B. & McMahon, H.T. Calcium triggers calcineurin-dependent synaptic vesiclerecycling in mammalian nerve terminals. Curr. Biol. 8, 740–749 (1998).

23. Cousin, M.A. & Robinson, P.J. Ba2+ does not support synaptic vesicle retrieval in ratcerebrocortical synaptosomes. Neurosci. Lett. 253, 1–4 (1998).

24. Gad, H., Low, P., Zotova, E., Brodin, L. & Shupliakov, O. Dissociation between Ca2+-triggered synaptic vesicle exocytosis and clathrin-mediated endocytosis at a centralsynapse. Neuron 21, 607–616 (1998).

25. Balaji, J., Armbruster, M. & Ryan, T.A. Calcium control of endocytic capacity at a CNSsynapse. J. Neurosci. 28, 6742–6749 (2008).

26. von Gersdorff, H. & Borst, J.G. Short-term plasticity at the calyx of held. Nat. Rev.Neurosci. 3, 53–64 (2002).

27. Yamashita, T., Hige, T. & Takahashi, T. Vesicle endocytosis requires dynamin-dependentGTP hydrolysis at a fast CNS synapse. Science 307, 124–127 (2005).

28. Xu, J. et al. GTP-independent rapid and slow endocytosis at a central synapse. Nat.Neurosci. 11, 45–53 (2008).

29. Wu, X.S. et al. The origin of quantal size variation: vesicular glutamate concentrationplays a significant role. J. Neurosci. 27, 3046–3056 (2007).

30. Satzler, K. et al. Three-dimensional reconstruction of a calyx of Held and its postsynapticprincipal neuron in the medial nucleus of the trapezoid body. J. Neurosci. 22,10567–10579 (2002).

31. Kushmerick, C., Renden, R. & von Gersdorff, H. Physiological temperatures reduce therate of vesicle pool depletion and short-term depression via an acceleration of vesiclerecruitment. J. Neurosci. 26, 1366–1377 (2006).

32. Wu, W. & Wu, L.G. Rapid bulk endocytosis and its kinetics of fission pore closure at acentral synapse. Proc. Natl. Acad. Sci. USA 104, 10234–10239 (2007).

33. Artalejo, C.R., Elhamdani, A. & Palfrey, H.C. Calmodulin is the divalent cation receptorfor rapid endocytosis, but not exocytosis, in adrenal chromaffin cells. Neuron 16,195–205 (1996).

34. Neher, E. & Sakaba, T. Multiple roles of calcium ions in the regulation of neurotrans-mitter release. Neuron 59, 861–872 (2008).

35. Sun, J. et al. A dual-Ca2+-sensor model for neurotransmitter release in a centralsynapse. Nature 450, 676–682 (2007).

36. Xu, J. & Wu, L.G. The decrease in the presynaptic calcium current is a major cause ofshort-term depression at a calyx-type synapse. Neuron 46, 633–645 (2005).

37. Borst, J.G.G. & Sakmann, B. Calcium influx and transmitter release in a fast CNSsynapse. Nature 383, 431–434 (1996).

38. Tadross, M.R., Dick, I.E. & Yue, D.T. Mechanism of local and global Ca2+ sensing bycalmodulin in complex with a Ca2+ channel. Cell 133, 1228–1240 (2008).

39. Kopp-Scheinpflug, C., Tolnai, S., Malmierca, M.S. & Rubsamen, R. The medial nucleusof the trapezoid body: comparative physiology. Neuroscience 154, 160–170 (2008).

40. Malgaroli, A. et al. Presynaptic component of long-term potentiation visualized atindividual hippocampal synapses. Science 268, 1624–1628 (1995).

41. Murthy, V.N. & Stevens, C.F. Reversal of synaptic vesicle docking at central synapses.Nat. Neurosci. 2, 503–507 (1999).

42. Ryan, T.A., Reuter, H. & Smith, S.J. Optical detection of a quantal presynapticmembrane turnover. Nature 388, 478–482 (1997).

43. Pyle, J.L., Kavalali, E.T., Piedras-Renteria, E.S. & Tsien, R.W. Rapid reuse ofreadily releasable pool vesicles at hippocampal synapses. Neuron 28, 221–231(2000).

44. Rizzoli, S.O. & Betz, W.J. The structural organization of the readily releasable pool ofsynaptic vesicles. Science 303, 2037–2039 (2004).

45. Nucifora, P.G. & Fox, A.P. Barium triggers rapid endocytosis in calf adrenal chromaffincells. J. Physiol. (Lond.) 508, 483–494 (1998).

46. Clayton, E.L., Evans, G.J. & Cousin, M.A. Activity-dependent control of bulk endocytosisby protein dephosphorylation in central nerve terminals. J. Physiol. (Lond.) 585,687–691 (2007).

47. Nicholson-Tomishima, K. & Ryan, T.A. Kinetic efficiency of endocytosis at mammalianCNS synapses requires synaptotagmin I. Proc. Natl. Acad. Sci. USA 101,16648–16652 (2004).

48. Xu, J., Mashimo, T. & Sudhof, T.C. Synaptotagmin-1, -2 and -9: Ca2+ sensors for fastrelease that specify distinct presynaptic properties in subsets of neurons. Neuron 54,567–581 (2007).

1010 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSSlice preparation, capacitance and mEPSC recordings and solutions. All

experiments were conducted in accordance with the guidelines established by

the National Institute of Neurological Disorders and Stroke Intramural Animal

Care and Use Committee. Parasagittal brainstem slices (200 mm thick) contain-

ing the medial nucleus of the trapezoid body were prepared from Wistar rats or

mice using a vibratome36. Unless mentioned otherwise, the animals were

7–10 d old, experiments were carried out in rats at 22–24 1C, the [Ca2+]o was

2 mM, the presynaptic pipette contained 50 mM BAPTA and t tests were used to

determine statistical significance. Means are presented as ± standard errors.

Whole-cell capacitance measurements were made with the EPC-9 amplifier

and the software lock-in amplifier (PULSE, HEKA) to implement the Lindau-

Neher technique (for details, see ref. 29). The frequency of the sinusoidal

stimulus was 1,000 Hz, and the peak-to-peak voltage of the sine wave was

r60 mV. The mEPSCs were recorded with an Axopatch 200B amplifier (Axon

Instruments). Data were low-pass filtered at 5 kHz and sampled at 20 kHz (for

details, see ref. 29). The holding potential for both pre- and postsynaptic

recordings was �80 mV. To reduce the noise, we did not compensate for the

postsynaptic series resistance (o15 MO) in mEPSC recordings.

For capacitance recordings, we pharmacologically isolated presynaptic Ca2+

currents and postsynaptic AMPA receptor–mediated EPSCs with a bath

solution containing 105 mM NaCl, 20 mM TEA-Cl, 2.5 mM KCl, 1 mM

MgCl2, 2 mM CaCl2, 25 mM NaHCO3, 1.25 mM NaH2PO4, 25 mM dextrose,

0.4 mM ascorbic acid, 3 mM myo-inositol, 2 mM sodium pyruvate, 0.001 mM

tetrodotoxin, 0.1 mM 3,4-diaminopyridine and 0.05 mM D(–)-2-amino-

5-phosphonovaleric acid(D(-)-AP-5), at pH 7.4 when bubbled with 95% O2

and 5% CO2. The presynaptic pipette contained 125 mM cesium gluconate,

20 mM CsCl, 4 mM MgATP, 10 mM sodium phosphocreatine, 0.3 mM GTP,

10 mM HEPES and 0.05 mM BAPTA, pH 7.2, adjusted with CsOH. When a

high concentration of EGTA (10–70 mM) or BAPTA (10 mM) was added to

the pipette, the cesium gluconate was reduced to keep the same osmolarity

(310–320 mOsm).

For simultaneous recordings of the capacitance and the AMPA receptor–

mediated mEPSC (Fig. 2), the bath solution contained 105 mM NaCl, 20 mM

TEA-Cl, 2.5 mM KCl, 1 mM MgCl2, 2 mM CaCl2, 25 mM NaHCO3, 1.25 mM

NaH2PO4, 25 mM dextrose, 0.4 mM ascorbic acid, 3 mM myo-inositol, 2 mM

sodium pyruvate, 0.001 mM tetrodotoxin, 0.1 mM 3,4-diaminopyridine,

0.05 mM D(-)-AP-5, 0.01 mM bicuculline and 0.01 mM strychnine, at

pH 7.4 when bubbled with 95% O2 and 5% CO2. The presynaptic pipette

(2.5–4.5 MO) solution contained 105 mM cesium gluconate, 20 mM CsCl,

4 mM MgATP, 10 mM sodium phosphocreatine, 0.3 mM GTP and 10 mM

HEPES, pH 7.2, adjusted with CsOH. To this solution, we added 10 mM EGTA

and 7.50 mM CaCl2 to clamp the free calcium concentration at 0.5 mM or

10 mM EGTA and 8.55 mM CaCl2 to clamp the free calcium concentration at

0.75 mM. The osmolarity was 310–320 Osm. The free calcium concentration

was calculated by assuming a calcium dissociation constant of 0.15 mM (for the

binding between EGTA and calcium)49. This was confirmed by the intracellular

calcium measurement29. The measured value was 0.48 ± 0.03 mM (n ¼ 3) when

free calcium was clamped at 0.5 mM29.

The myosin light chain kinase peptide (MLCK), the MLCK control peptide

(MLCK-c, a point mutation) and the CBD were obtained from EMD

Chemicals. The scrambled CBD peptide with a sequence of KLRLARLK-

ATKNTFKMLGIA, used as a control, was purchased from 21st Century

Biochemicals. DNF and DPF were also obtained from 21st Century Bio-

chemicals. Calmidazolium and the wild-type calmodulin protein were

purchased from Sigma.

Measurements of the rate and amount of endocytosis. The rateendo was

measured as the capacitance jump divided by the endocytosis time constant,

which was similar to the rate of the capacitance decay in the first 2–3 s after

stimulation. Similarly, the rateendo for the rapid and slow component of

endocytosis (Figs. 3 and 5) was measured as the amplitude of the rapid and

the slow component divided by their corresponding endocytosis time constant.

When endocytosis was slow or blocked (Figs. 1b,d and 4, and Supplementary

Fig. 1), the capacitance decay was approximately linear in the first 10–60 s after

stimulation. In these conditions, the rateendo was measured as the mean

capacitance decay within 10–60 s of stimulation.

We used rateendo instead of t to evaluate the rate of endocytosis50 because

many manipulations substantially inhibited endocytosis, making it difficult to

measure t. Furthermore, t may not necessarily reflect the rate of endocytosis.

Saturation of the endocytic machinery could result in an increase of t with a

constant rateendo50 (see also Supplementary Data 1). Thus, rateendo is a more

accurate measurement of the speed of endocytosis25,50.

The endocytosis overshoot was measured as the difference between the

baseline and the capacitance value at B40–60 s after stimulation. When

comparing two groups of data (for example, Figs. 1e, 5a,b and 6), we often

normalized our rateendo data in the experimental condition to the respective

control group. We did this because the control values for each experiment were

somewhat different, as the data for the various manipulations were obtained by

several experimenters. However, for each manipulation (such as dialysis of the

CBD peptide), the control and experimental data were obtained by one

experimenter. By normalizing the data, we avoided meaningless comparisons

of the slight differences in control values obtained by different experimenters.

Measurements of bulk endocytosis. The method for measuring large DCSs,

which reflect bulk endocytosis, the fission pore conductance change during

DCSs and the fission pore diameter change during DCSs were described

previously32. To detect DCSs, we low-pass filtered capacitance traces at 30 Hz

and differentiated. The differentiation was calculated as the difference between

capacitance values of two neighboring samples with an interval of 1 ms. A DCS

was identified when the rate of the capacitance decay was more than 50 fF per

100 ms in the differentiated capacitance trace, the size of a DCS was more than

20 fF in the filtered capacitance trace, and the measured series conductance and

membrane conductance did not change in parallel with the DCS. A DCS was

often superimposed on the smooth capacitance decay. To correct the baseline

decay, we fit the baseline decay in 100–500 ms before a DCS with a linear

regression line and subtracted from the measured capacitance trace. With a

baseline-corrected DCS, we calculated the fission pore conductance and

diameter as described previously32.

Synaptotagmin 2 knockout mice. Mice lacking synaptotagmin 2 (Syt2) have

been described previously35. In short, they were generated by replacing part of

exon 2 through exon 7 of the Syt2 gene with a lacZ cDNA construct using

homologous recombination. Mice for these experiments were obtained by

heterozygous breeding using standard mouse husbandry procedures. Geno-

typing was performed by PCR, which was previously confirmed to correlate

with Syt2 expression by immunoblotting. In some experiments, we confirmed

by fluorescence imaging that Syt2�/� mice did not show any immunoreactivity

to antibodies raised against Syt2 (data not shown, but see ref. 35).

49. Augustine, G.J. & Neher, E. Calcium requirements for secretion in bovine chromaffincells. J. Physiol. (Lond.) 450, 247–271 (1992).

50. Sankaranarayanan, S. & Ryan, T.A. Real-time measurements of vesicle-SNARE recy-cling in synapses of the central nervous system. Nat. Cell Biol. 2, 197–204 (2000).

doi:10.1038/nn.2355 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

SAP97 and CASK mediate sorting of NMDA receptorsthrough a previously unknown secretory pathway

Okunola Jeyifous1, Clarissa LWaites1, Christian G Specht1, Sho Fujisawa2, Manja Schubert3, Eric I Lin4,John Marshall5, Chiye Aoki2, Tharani de Silva3, Johanna M Montgomery3, Craig C Garner1,6 &William N Green4,6

Synaptic plasticity is dependent on the differential sorting, delivery and retention of neurotransmitter receptors, but the

mechanisms underlying these processes are poorly understood. We found that differential sorting of glutamate receptor subtypes

began in the endoplasmic reticulum of rat hippocampal neurons. As AMPA receptors (AMPARs) were trafficked to the plasma

membrane via the conventional somatic Golgi network, NMDA receptors (NMDARs) were diverted from the somatic endoplasmic

reticulum into a specialized endoplasmic reticulum subcompartment that bypasses somatic Golgi, merging instead with dendritic

Golgi outposts. This endoplasmic reticulum subcompartment was composed of highly mobile vesicles containing the NMDAR

subunits NR1 and NR2B, the microtubule-dependent motor protein KIF17, and the postsynaptic adaptor proteins CASK and

SAP97. Our data demonstrate that the retention and trafficking of NMDARs in this endoplasmic reticulum subcompartment

requires both CASK and SAP97. These findings indicate that NMDARs are sorted away from AMPARs via a non-conventional

secretory pathway that utilizes dendritic Golgi outposts.

NMDARs regulate synaptic plasticity by functioning as coincidencedetectors that integrate synapse-specific information with the over-all excitability of neuronal cells1. NMDAR activation is essential foreliciting changes in synaptic strength, primarily by regulating thelevels of postsynaptic AMPARs2,3. Numerous studies have demon-strated that NMDARs and AMPARs are independently deliveredto nascent synapses during synaptogenesis4–6 and to maturesynapses during synaptic plasticity2,3. Mechanistically, the sortingof these receptors into distinct vesicles4,6 and/or their differentialretention at the postsynaptic density (PSD)7 may underlie thesedifferences, although the relative contribution of each mechanismremains unclear.

Studies of AMPAR biogenesis, transport and synaptic deliverysuggest a three step mechanism, wherein receptor subunits aresynthesized in the somatic endoplasmic reticulum and Golgi,transported to the plasma membrane by constitutive membraneflow8,9 and subsequently internalized into recycling endosomes10

before their activity-dependent reinsertion at extrasynapticsites11,12 and capture by scaffold proteins in the PSD7,13. It isunclear whether NMDARs follow a similar biosynthetic pathwayand/or at what stage they are sorted from AMPARs. Studies suggestthat members of the membrane-associated guanylate kinase(MAGUK) family of synaptic scaffold proteins (for example,SAP97, PSD-95, SAP102 and CASK)7, the protein GRIP/ABP, and

subclasses of microtubule and actin-dependent motor proteinscontribute to the differential sorting and trafficking of NMDAand AMPARs14–18. For example, SAP97 associates with AMPARsubunits during their biosynthesis and transport to the plasmamembrane15 and GRIP1 and KIF5 participate in the synapticdelivery of these receptors19. Similarly, subunits of the NMDARform complexes with CASK, Velis/MALS, Mint and KIF17 onvesicles that move rapidly along dendritic microtubules14,19 atrates that are distinct from vesicles carrying AMPARs6,20. Addi-tional reports indicate that NMDAR subunits also form complexeswith SAP102, sec8 and mPins21,22, as well as SAP97 (refs. 23–25).However, it remains unclear when and where each complex formsor how each contributes to individual steps in the biogenesis,transport and recycling of NMDARs.

We explored the mechanisms that underlie sorting of AMPARsversus NMDARs. We found that NMDARs were trafficked via aSAP97- and CASK-dependent pathway from somatic endoplasmicreticulum to a dendritic endoplasmic reticulum subcompartment,and subsequently to Golgi outposts. In contrast, AMPARs followedthe conventional route from somatic endoplasmic reticulum and Golgito reach the plasma membrane. These data not only provide newinsights into the cellular mechanisms underlying glutamate receptorsorting, but also indicate that dendritic Golgi outposts26 are part of afunctionally distinct secretory pathway.

Received 25 March; accepted 8 June; published online 20 July 2009; doi:10.1038/nn.2362

1Department of Psychiatry and Behavioral Science, Stanford University, Palo Alto, California, USA. 2Center for Neural Science and Department of Biology, New YorkUniversity, New York, New York, USA. 3Department of Physiology, University of Auckland, Auckland, New Zealand. 4Department of Neurobiology, University of Chicago,Chicago, Illinois, USA. 5Department of Molecular Pharmacology, Physiology and Biotechnology, Brown University, Providence, Rhode Island, USA. 6These authorscontributed equally to this work. Correspondence should be addressed to C.C.G. (cgarner@stanford.edu) or W.N.G. (wgreen@midway.uchicago.edu).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1011

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

RESULTS

NMDARs bypass somatic Golgi

Sorting of membrane proteins typicallyoccurs at the trans-Golgi network (TGN)27.To determine whether glutamate receptorstraffic through somatic TGN, we transfectedenhanced green fluorescent protein (EGFP)-tagged AMPAR (GFP-GluR1) or NMDAR(NR1-GFP) subunits with a constitutivelyactive mutant of the GTPase ARF1 (ARF1-Q71I), which causes the accumulation ofcargo in the Golgi by preventing COPI-coatedvesicle formation28.

When GFP-GluR1 was transfected intocultured hippocampal neurons with hemag-glutinin (HA)-tagged ARF1-Q71I (ARF1-Q71I–HA), both proteins accumulated inthe cell soma and colocalized with antibodiesto the Golgi marker GM130 (Fig. 1a). Quan-tifying the percent overlap with GM130 revealed a 45% increase in thelevels of Golgi-localized GFP-GluR1 in neurons coexpressing ARF1-Q71I–HA (P ¼ 0.02, t test). In contrast, we saw no accumulation ofNR1-GFP in neurons coexpressing ARF1-Q71I–HA (Fig. 1a), indicat-ing that NMDARs may bypass somatic Golgi. Similar results wereobtained with two other conditions that block protein traffickingthrough the Golgi (Supplementary Fig. 1).

NMDARs traffic through dendritic Golgi outposts

The absence of NR1-GFP in somatic Golgi following coexpression withARF1-Q71I was puzzling to us, as membrane proteins are usually

post-translationally processed in the Golgi before their insertion atthe plasma membrane. We therefore tested whether NMDARs insteaduse dendritic Golgi membranes. These Golgi outposts are considered tobe an extension of the somatic Golgi29 or a site of biosynthesis forintegral membrane proteins translated from dendritically localizedmRNAs26,30. To assess a role for Golgi outposts in NMDAR traffickingand processing, we again used ARF1-Q71I–HA to block Golgi traffick-ing. As with somatic Golgi, GM130 immunoreactivity identifieddendritic Golgi outposts (Fig. 1b). In neurons expressing NR1-GFPalone, we observed modest colocalization with GM130-positive puncta.However, when ARF1-Q71I–HA was transfected with NR1-GFP, we

NR1GM130

ARF1-Q71I–HA

Galtase-GFP NR1

ARF1-Q71I–HA

NR1 intensity

c

d

NR2B intensityGaltase-GFP NR2B

ARF1-Q71I–HA

e

a

b

GM130

GM130

GM130

GFP-GluR1

NR1-GFP

ARF1-Q71I–HA

ARF1-Q71I–HA

ARF1-Q71I–HA

NR1-GFP

Figure 1 NMDARs exit the soma via a somatic

Golgi–independent pathway and insert into

dendritic Golgi outposts. (a,b) Hippocampal

neurons were transfected with GFP-GluR1 or

NR1-GFP alone or with one of the subunits and

ARF1-Q71I–HA. Cultures were fixed at 12–15 h

post-transfection. Co-expression with ARF1-Q71I-

HA significantly increased the Golgi localization ofGFP-GluR1 (P = 0.02, arrows, second row), but

had no effect on NR1-GFP distribution or Golgi

localization (fourth row) (a). In dendrites,

coexpression with ARF1-Q71I–HA resulted in a

significant accumulation of NR1-GFP at Golgi

outposts (P = 0.029, bottom panel, b), as

compared with dendrites in which the subunit was

expressed alone (top panel, b). (c–e) Accumu-

lation of endogenous NMDARs at Golgi outposts

following mutant ARF1 overexpression. Expressing

the ARF1 mutant had no effect on NR1

localization in somatic Golgi (arrows, bottom row,

c). We transfected cultures with the Golgi marker

Galtase-GFP alone (top rows) or with ARF1-Q71I–

HA (bottom rows) (d,e). We observed minimal

colocalization of endogenous NR1 (d) and NR2B

(e) with outposts in neurons only expressing

Galtase-GFP (arrows in top rows, merged panels).

Coexpression of the ARF1 mutant resulted in anincrease in the colocalization of both NR1 and

NR2B with Golgi outposts (arrows in bottom rows,

merged panels), as well as in the fluorescence

intensity at these structures (pseudo-color

intensity panels; increasing intensity is reflected in

progression from blue-to-red-to-yellow). Scale bars

represent 10 mm.

1012 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

detected a significant increase in Golgi outpost size and the intensityof colocalizing NR1-GFP puncta (Bthreefold, P ¼ 0.029), which isconsistent with NMDARs trafficking through dendritic, rather thansomatic, Golgi. No accumulation of AMPAR subunits was observed inGolgi outposts following ARF1-Q71I block, as almost all of the GFP-GluR1 remained in the soma and failed to traffic beyond proximaldendritic segments (Supplementary Fig. 2).

To rule out the possibility that NR1-GFP overexpression wasresponsible for these findings, we examined the effects of ARF1-Q71I–HA on endogenous NR1 and NR2B subunits. Again, ARF1-Q71I-HA did not trigger the accumulation of NR1 in somatic Golgi(Fig. 1c), but did significantly increase the percent overlap betweendendritic NR1- and NR2B-immunoreactive puncta (60% and 43%,P ¼ 0.016 and 0.007, respectively) and Golgi outposts (labeled with theEGFP-tagged Golgi marker galactosyltransferase, Galtase-GFP). Thisaccumulation was not a result of alterations in synapse stability orformation by ARF1-Q71I–HA expression (Supplementary Fig. 3).These data indicate that NMDARs traffic from the soma via a pathwaythat bypasses somatic Golgi, but utilizes dendritic Golgi, in contrastwith the conventional sorting pathway of AMPARs.

NMDAR transport in an endoplasmic reticulum subcompartment

If NMDARs bypass the somatic Golgi by budding directly from theendoplasmic reticulum, then NMDAR-containing transport vesiclesin dendrites may have features in common with the endoplasmicreticulum. Recent studies have demonstrated the existence of anendoplasmic reticulum subcompartment throughout the dendritesof hippocampal neurons29. The endoplasmic reticulum subcompart-ment contains endoplasmic reticulum–resident proteins, proteinsinvolved in Ca2+ storage and release31,32, and associated vesicles that

move via kinesin motors in both the antero-grade and retrograde direction at velocitiesof 0.2–0.3 mm s–1 at 36–37 1C31,32. To assesswhether NMDAR-containing vesicles are part

of this compartment, we examined the expression of three endoplasmicreticulum markers, KDEL, IP3 receptor and DsRed-ER (a DsRed fusionprotein containing endoplasmic reticulum–retention signals from bothcalreticulin and KDEL). All three markers colocalized with each otherand with newly synthesized NR1-GFP in small punctate structures indendrites (Fig. 2a).

To further characterize these NMDAR-containing vesicles, we carriedout live imaging on 13-d-old neurons coexpressing newly synthesizedNR1-GFP and DsRed-ER. We routinely observed small, discrete, highlymobile NR1-GFP puncta that also contained DsRed-ER (Fig. 2b).These puncta moved rapidly along dendrites in both anterograde andretrograde directions at a mean rate of 0.26 ± 0.03 mm s–1 at 36–37 1C(for 11 measured events). This rate is similar to those reported previ-ously for endoplasmic reticulum–like vesicles in dendrites31,32 and forNMDAR-containing vesicles6,17.

We next tested whether other NMDAR or AMPAR subunits werealso present on endoplasmic reticulum subcompartment vesicles.Brefeldin A (BFA) was used to enrich for endoplasmic reticulumvesicles trafficking from the soma. Cultures were transfected withDsRed-ER, treated with BFA (2 h post-transfection) and assayed forendogenous NR2B and GluR1 subunits on endoplasmic reticulumsubcompartment puncta (Fig. 2c). We observed a high degree ofcolocalization between DsRed-ER and NR2B (60%), but not betweenDsRed-ER and GluR1 (19%), which is consistent with NR1- andNR2B-containing NMDARs, but not with GluR1-containing AMPARs,in endoplasmic reticulum–derived vesicles that traffic to dendrites.

It was previously shown that dendritic transport of NMDARs ismediated by a complex composed of three multi-domain scaffoldproteins, CASK (Lin2), Velis/MALS (Lin7) and Mint (Lin10), thattether NMDARs to the kinesin KIF17 (refs. 14,33). We used

IP3RKDELDsRed-ER

KDEL NR1-GFP

IP3R NR1-GFP

DsRed-ER NR1

DsRed-ER NR1-GFP

NR1-GFP

DsRed-ER

0 s 20 s 60 s

a

b

DsRed-ER NR2B GluR1

Proximal

Distal

BFA added at 2 h post-transfection

c

Figure 2 NMDARs traffic in a mobile, vesicular,

endoplasmic reticulum subcompartment.

Hippocampal neurons were transfected with

DsRed-ER and/or NR1-GFP, fixed at 1 d post-

transfection, and stained with antibodies to KDEL,

NR1 or IP3 receptor (IP3R). (a) Co-distribution of

previously characterized markers for the mobile

endoplasmic reticulum subcompartment indendrites (top row). Newly synthesized, trafficking

puncta of NR1-GFP colocalized with KDEL, IP3R

and DsRed-ER (arrows; second, third and fifth row,

respectively). Dendritic DsRed-ER puncta

colocalized with endogenous NR1 subunits

(arrows, fourth row). Scale bars represent 5 mm.

(b) Inverted monochrome images from time-lapse

imaging of neurons coexpressing NR1-GFP (top

row) and DsRed-ER (bottom row). NR1-GFP and

DsRed-ER–containing puncta (arrows) were highly

mobile (arrowhead indicates position at time ¼ 0

s). Scale bar represents 5 mm. (c) Cultures were

transfected with DsRed-ER, treated with BFA at

2 h post-transfection for a period of 10 h, and

then fixed and stained with antibodies to NR2B

and GluR1. DsRed-ER–containing puncta that

trafficked out of the soma despite Golgi membrane

disruption colocalized highly with NR2B (arrows),

but not GluR1 (arrows), in both proximal (toppanels) and distal (bottom panels) dendritic

processes. Scale bar represents 10 mm.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1013

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

immunostaining to evaluate whether the observed NMDAR-contain-ing vesicles contained CASK and KIF17 (Supplementary Fig. 4). Wefound that clusters of CASK and KIF17 had a high degree of overlapwith NR1-GFP and DsRed-ER puncta (Supplementary Fig. 4), sug-gesting that mobile NMDAR/endoplasmic reticulum vesicles sharefeatures with previously characterized NMDAR transport vesicles.

SAP97 and CASK alter NMDAR trafficking

CASK not only forms complexes with Velis/MALS and Mints33,34, butalso with another MAGUK, SAP97 (refs. 35,36). SAP97 interacts via itsN-terminal PDZ domains with the C-terminal tails of NR2A or NR2Bsubunits of the NMDAR23 and is known to do so in the endoplasmicreticulum of transfected cells25,37. Taken together, these features led usto hypothesize that a complex containing SAP97 and CASK may trafficNMDARs into the dendritic endoplasmic reticulum subcompartment.

This theory was initially tested in HEK293 cells by coexpressing HA-NR1 (HA-tagged on its extracellular domain) and NR2B subunits(Fig. 3). We detected NMDARs at the cell surface, which is consistentwith assembled NMDARs having transited through the endoplasmicreticulum and Golgi (Fig. 3a,e). However, when we transfected cellswith EGFP-tagged SAP97 and HA-NR1, surface NMDAR expressionwas lost (Fig. 3b,e) and nearly all of the NMDARs were sequesteredin large structures colocalizing precisely with EGFP-SAP97 (Fig. 3b).These clusters also colocalized with ERp57 (Fig. 3b), a resident

endoplasmic reticulum protein, suggesting that SAP97 sequesteredNMDARs in an endoplasmic reticulum–derived compartment. AlthoughSAP97 can also bind AMPAR GluR1 subunits38, it had no effect on thesurface expression or endoplasmic reticulum distribution of these recep-tors when a HA-tagged version of GluR1 (HA-GluR1) was coexpressedwith EGFP-SAP97 (Fig. 3a).

We also examined whether CASK affected the distribution ofNMDARs in HEK293 cells. In the presence of Myc-tagged CASK andEGFP-SAP97, NMDARs trafficked to the cell surface and were nolonger retained in the endoplasmic reticulum (Fig. 3c,e). This effectrequired a direct interaction between CASK and SAP97, as a CASKdeletion mutant lacking its N-terminal L27 domain (CASK-DL27N),which binds the N-terminal L27 domain of SAP97 (ref. 35), had noeffect when it was co-transfected with GFP-SAP97 (Fig. 3d,e). Thesedata indicate that SAP97 can mediate the endoplasmic reticulumretention of NMDARs and, together with CASK, may regulate thesorting and trafficking of NMDARs through the endoplasmic reticu-lum and Golgi.

We next examined whether SAP97 and CASK formed complexeswith NMDARs in hippocampal neurons. To begin, we immunostainedcultured hippocampal neurons with antibodies to NR1, SAP97 and/orCASK. We observed strong colocalization of NR1 with SAP97 or CASKin somata and punctate structures along the dendrites of these neurons(Fig. 4a). To distinguish between synaptic puncta and NMDARtransport vesicles, we first immunostained cultures with antibodiesto SAP97 and synapsin. Endogenous SAP97 was present not only atsynapses, as defined by the colocalization of SAP97 and synapsinpuncta along dendrites, but also in a distinct population of smallerpuncta lacking synapsin (Fig. 4a). The latter finding suggests thatnonsynaptic SAP97 might be associated with dendritic vesicles contain-ing NMDARs. As another test, we transfected neurons with yellowfluorescent protein (YFP)-tagged SAP97 and GFP-tagged NR1, andthen immunostained them with antibodies to synapsin and CASK.Using a filter set that separated YFP and GFP spectra, we foundtwo populations of YFP-SAP97 puncta: larger synaptic puncta thatcolocalized with NR1-GFP, CASK and synapsin, and smaller punctathat colocalized only with CASK and NR1-GFP (Fig. 4b).

a b

c

d

Intracellular Surface

e104

103

104

103

104

103

104

103

5K 10K 15K 20K 25KSurface NMDAR expression

(AFU)

NMDAR

NMDAR + SAP97

NMDAR + SAP97+ CASK

NMDAR + SAP97+ CASK∆L27

Sur

face

NM

DA

R e

xpre

ssio

n(n

orm

aliz

ed)

**

GFP-SAP97 HA-NMDAR HA-GluR1

GFP-SAP97

GFP-SAP97 HA-NMDAR GFP-SAP97 HA-NMDARMyc-CASK

GFP-SAP97 HA-NMDAR GFP-SAP97

1.2

1.0

0.8

0.6

0.4

0.2

0

NMDAR

NMDAR +

SAP97

NMDAR +

SAP97

+ C

ASK

NMDAR +

SAP97

+ C

ASK∆L27N

HA-NMDARMyc-CASK∆L27N

+HA-GluR1 +HA-GluR1GFP-SAP97

GFP-SAP97+HA-NMDAR +HA-NMDAR

GFP-SAP97

GFP-SAP97+HA-NMDAR

GFP-SAP97 ERP57+HA-NMDAR

+HA-NMDARGFP-SAP97

Figure 3 SAP97 and CASK alter NMDAR trafficking in HEK293 cells.

(a) Top, we transiently transfected cells with GFP-SAP97, HA-tagged

NMDARs containing NR1 and HA-NR2B, or HA-GluR1, respectively. NR2B

and GluR1 subunits contained N-terminal, HA epitope tags to permit surface

HA staining. Bottom, GFP-SAP97 and HA-GluR1 colocalized at the cell

surface. (b) Top, loss of GFP-SAP97 and HA-tagged NMDAR trafficking to the

plasma membrane. Middle, permeabilization revealed the intracellular

retention of both proteins. Bottom, intracellular clusters were specificallylabeled with antibodies to ERp57. (c,d) Cells expressing GFP-SAP97, HA-

NMDAR, (HA-NR1/wt-NR2B) and Myc-CASK (c) or Myc-CASK-DL27N (d).

Single z planes of intact cells revealed that coexpression of GFP-SAP97 with

Myc-CASK (c), but not with Myc-CASK-DL27N (d), restored surface

expression of SAP97 and NMDARs (surface panels). Permeabilized cells

expressing Myc-CASK (c) lacked intracellular clusters of NMDAR/SAP97

complexes. Expressing Myc–CASK-DL27N (d) did not relieve the clustering of

NMDARs with SAP97. (e) Live cells were labeled with antibodies to HA and

flow cytometry was used to quantitatively assay for GFP and surface HA

expression. Surface NMDAR expression of GFP-expressing cell populations is

shown on the x axis (arbitrary fluorescence units) and cell granularity (side

scatter) is shown on the y axis. The total fluorescence intensities of surface

NMDARs were quantified (right). Expressing GFP-SAP97 caused a reduction

in surface NMDARs (54.93% ± 6.80%, P ¼ 0.000882). This reduction was

rescued by Myc-CASK (93.34% ± 15.94%), but not by Myc–CASK-DL27N

(69.03% ± 7.83%, P ¼ 0.006717). Data are expressed as mean ± s.e.m.,

n ¼ 3 independent experiments assaying 50,000 cells per experiment. Scale

bars represent 5 mm.

1014 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

To test whether SAP97 and NMDARs were found together in mobileendoplasmic reticulum vesicles, we transfected cells with red fluores-cent protein (RFP)-SAP97 and NR1-GFP or EGFP-SAP97 and DsRed-ER before time-lapse imaging (Fig. 4c). We detected highly mobileRFP-SAP97 and EGFP-SAP97 puncta that comigrated along dendriteswith NR1-GFP (average velocity in both the retrograde and antero-grade directions of 0.13 ± 0.07 mm s–1, mean ± s.d., n ¼ 13; seeSupplementary Video 1) or DsRed-ER puncta (average velocity in

both the retrograde and anterograde direc-tions of 0.34 ± 0.05 mm s–1, mean ± s.d.,n ¼ 15), respectively. These data indicatethat SAP97 and NMDARs associate withmotile, endoplasmic reticulum–like vesiclesin dendrites.

We next assessed whether SAP97 exists in a complex with NMDARsand CASK by immunoprecipitating SAP97 from postnatal day 2 (P2)rat brain microsomal membranes (Fig. 5). Using western blot analysis,we found that CASK and the NR2B subunits of the NMDAR co-immunoprecipitated with SAP97, which is consistent with the forma-tion of a complex containing SAP97, CASK and NMDARs (Fig. 5a).The formation of this complex was confirmed in HEK293 cells. HAantibodies immunoprecipitated HA-tagged NMDAR subunits as well

NR1SAP97

CASK

b

a

YFP-SAP97 Synapsin CASK NR1-GFP

DsRed-ER

NR1-GFP

0 s

0 s

25 s

12 s

75 s

23 s

RFP-SAP97

GFP-SAP97

c

CASK

SAP97 Synapsin

NR1 NR1

SAP97 NR1Figure 4 SAP97 co-distributes with NR1

and CASK in nonsynaptic, mobile puncta.

(a) Hippocampal neurons were fixed at 14 d

in vitro (14 DIV), permeabilized and double-

labeled with combinations of antibodies to

SAP97, NR1, CASK and synapsin. SAP97 (top

row, scale bar represents 10 mm) and CASK

(second row, scale bar represents 5 mm)colocalized with NR1 in the soma and in dendritic

puncta. Bottom, SAP97 was found at synaptic

and nonsynaptic dendritic puncta (arrows, scale

bar represents 10 mm). (b) Cultures were

transfected for 1 d with NR1-GFP and YFP-

SAP97. Neurons were then fixed, permeabilized

and double labeled with antibodies to CASK and

synapsin. NR1-GFP, YFP-SAP97 and CASK

colocalized in synaptic (arrowheads) and

nonsynaptic (arrows) dendritic puncta (scale bar

represents 10 mm). (c) Time-lapse imaging of

neurons transfected for 1 d with NR1-GFP and

RFP-SAP97 (top panels, scale bar represents

5 mm) or with GFP-SAP97 and DsRed-ER (bottom

panels, scale bar represents 5 mm). Comigrating

puncta for both transfection combinations

trafficked down dendrites at similar velocities

(0.13 ± 0.07 mm s–1 (n ¼ 13) and 0.34 ± 0.05

(n ¼ 15), respectively).

a b c

dSAP97

Cask

NR2B

Input

HA-NMDAR/GFP-SAP97+ rbrb

P R

rERrER

GFP, Myc

IgGLysate SAP97 IP

Sham

GFP-SAP97

Myc

–CASK-F

L

Myc

–CASK-

∆L27

IgG

Myc-CASK FL Myc-CASK ∆L27

IP IgG IP

Myc

GFP

Figure 5 SAP97, NMDARs and CASK form a complex in brain. (a) Co-immunoprecipitation of SAP97, CASK and NR2B from P2 rat brain lysates. (b) HEK293

cells were transfected with cDNAs encoding GFP-SAP97, HA-NR1/wt-NR2B, and Myc-CASK–FL (full length) or Myc–CASK-DL27. Lysates were immuno-

precipitated with antibody to HA and the precipitated proteins were immunoblotted with antibodies to GFP and c-Myc. (c,d) Ultrastructural colocalization

of SAP97 and NMDARs. A photograph showing pyramidal cell (P) and radiatum (R) layers of CA1 in P7 rat hippocampus is presented in c. The tissue was

labeled for SAP97 by SIG using a polyclonal rabbit antibody (rb). The micrograph shows SAP97 labeling in the somata as well as in the neuropil. A high-

magnification electron micrograph of the dendrites of a CA1 pyramidal neuron is shown in d. SAP97 was immunolabeled with SIG, the irregularly shapedand sized black spots. They were observed in the cytoplasm and in association with membranous organelles, such as the rough endoplasmic reticulum (rER).

NR2B subunits were immunolabeled with 10-nm colloidal gold, examples of which are indicated by arrows. The circles represent the 60-nm radius area

around SAP97 labeling; if NR2B labeling fell in this area, the two were considered to be colocalized. Black circles are examples of SAP97 labeling without

any NR2B colocalization and red circles are examples of those with NR2B colocalization. Scale bars represent 25 mm for the light micrograph and 200 nm

for the electron micrograph.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1015

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

as co-transfected GFP-SAP97 and Myc-CASK (Fig. 5b). Wefailed to detect any Myc-CASK in this complex when the L27NCASK deletion mutant was used (Fig. 5b), indicating that its associa-tion with the SAP97/NMDAR complex requires a L27 domain–mediated interaction.

To test whether SAP97 and NMDARs colocalized on cellularmembranes, we also performed dual-label immunogold electronmicroscopy on dendrites of CA1 pyramidal cells. SAP97 immuno-reactivity was detected at the pre-embed stage and enhanced by silver-intensified gold (SIG), and NR2B immunoreactivity was detected at thepost-embed stage and enhanced with 10 nm colloidal gold. We readilydetected 10-nm gold particles in the vicinity of larger SIG particles(Fig. 5d), many of which were associated with dendritic membranousstructures. Our limit of resolution for colocalization was 60 nm, the

estimated maximal distance that could betraversed by four IgG molecules linked to asingle SAP97-NR2B complex. Altogether, ourfindings of a SAP97-CASK-NMDAR complexin mobile endoplasmic reticulum–like vesiclesare consistent with the idea that SAP97 andCASK participate in the sorting of NMDARsinto the endoplasmic reticulum subcompart-ment found in dendrites.

SAP97 and CASK regulate NMDAR sorting

to Golgi outposts

To further examine the roles of SAP97 andCASK in NMDAR trafficking, we generated

short-hairpin RNAs (shRNAs) that were specific for SAP97 or CASKtranscripts. On the basis of western blot analysis, these shRNAsmarkedly suppressed the expression of SAP97 (490%) and CASK(470%), whereas a scrambled shRNA had no effect (SupplementaryFig. 5). To determine whether SAP97 or CASK knockdown affectedNMDAR trafficking, we examined the colocalization of NR1 subunitswith the Golgi marker GM130 in the somas of shRNA-expressingneurons (Fig. 6a). We observed a substantial increase in the perinucleardistribution of NR1 in neurons expressing shRNAs for SAP97 or CASKversus control cells (Fig. 6a) and substantial colocalization withGM130, suggesting that NR1 now trafficked via the somatic Golgi.Quantitatively, SAP97 knockdown led to a 33% increase in NR1 at thesomatic Golgi and CASK knockdown led to a 48% increase comparedwith neurons expressing scrambled shRNAs (Fig. 6b). These results

Rel

ativ

e G

olgi

syp

hin

tens

ity

*

Rel

ativ

e G

olgi

Glu

R1

inte

nsity

GM130 NR1

SAP97shRNA

Rel

ativ

e G

olgi

NR

1in

tens

ity

a b

CASKshRNA

**

c d

TGN38 Syph

SAP97shRNA

CASKshRNA

ScrambledshRNA

TGN 38 GluR1

SAP97shRNA

CASKshRNA

ScrambledshRNA

e f

ARF1- Q71I–HA NR1

SAP97shRNA

ScrambledshRNA

g hGM130 NR1 ARF1-Q71I–HA

ScrambledshRNA

SAP97shRNA R

elat

ive

Gol

gi N

R1

inte

nsity

*

i

2.0

1.6

1.2

0.8

0.4

0

2.5

2.0

1.6

1.2

0.8

0.4

0

2.0

1.5

1.0

0.5

0

1.6

1.2

0.8

0.4

0

SAP97 sh

RNA

Scram

bled

shRNA

CASK shRNA

SAP97 sh

RNA

SAP97 sh

RNA

Scram

bled

shRNA

CASK shRNA

Scram

bled

shRNA

ARF1-Q71

I

+Scr

amble

d sh

RNA

+SAP97

shRNA

CASK shRNA

Scram

bled

shRNA

Figure 6 Knockdown of SAP97 or CASK

reroutes NMDARs to the somatic Golgi and

away from dendritic Golgi outposts. (a–f)

Cultures were infected at 0 DIV with lentivirus

encoding hairpin sequences specific to SAP97,

CASK or a randomized sequence (scrambled).

Cultures were fixed at 16 DIV and stained with

antibodies to NR1 and GM130 (a), GluR1 (c),or synaptophysin (e) and TGN38. shRNA

knockdown of SAP97 (a) or CASK (b) resulted

in a significant buildup of NR1 in the Golgi

versus control cultures. Quantitative analysis

revealed increases in the amount of NR1 in the

Golgi of 33% and 48% (mean ± s.e.m. from ten

neurons per group, P ¼ 0.004 and P ¼ 0.019,

respectively; b). Although SAP97 knockdown

yielded a significant accumulation of GluR1 in

the Golgi (mean ± s.e.m. from ten neurons per

group, P ¼ 0.008; d), CASK knockdown had no

effect on its distribution (c). Knockdown of

SAP97 and CASK had no qualitative (e) or

quantitative (f) effect on the Golgi localization

of synaptophysin, another integral membrane

protein. (g) Overexpression of ARF1-Q71I–HA

had no effect on the somatic distribution of

NR1 in uninfected (top row) and scrambled

control (second row) cultures. Neurons infectedwith SAP97 shRNA showed an increased

somatic Golgi concentration of NR1 (bottom

row). (h) Quantitative analysis of the effect of

ARF1-Q71I–HA overexpression in somatic Golgi

membranes (mean ± s.e.m. from ten neurons

per group, P ¼ 0.03). (i) Knockdown of SAP97

yielded a loss of NR1 at ARF1-Q71I–HA puncta

in the dendrites of transfected neurons. Scale

bars represent 10 mm.

1016 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

suggest that SAP97 and CASK are required forthe selective sorting of NMDARs away fromsomatic Golgi.

We also tested the effects of SAP97 andCASK knockdown on AMPAR trafficking.We observed an accumulation of endogenousGluR1 in the somatic Golgi of neuronsinfected with the SAP97 shRNA. No accu-mulation was seen in uninfected neurons,neurons infected with a scrambled shRNAor neurons infected with shRNA specificto CASK (Fig. 6c,d). Reducing the levelsof SAP97 or CASK had no effect on thetrafficking of another synaptic protein, synap-tophysin, through somatic Golgi (Fig. 6e,f).Together, these observations indicate thatSAP97, but not CASK, has a specific rolein the forward trafficking of GluR1 throughsomatic Golgi. This result is consistent withthose of previous studies demonstrating thata SAP97-GluR1 interaction has a role inforward trafficking of AMPARs to the plasmamembrane15. No such involvement of CASKin AMPAR trafficking has been described.

The accumulation of NR1 in somatic Golgiwith knockdown of SAP97 or CASK suggeststhat these MAGUKs are required for theforward trafficking of NMDARs throughGolgi membranes. Alternatively, NMDARsmay still have the capacity to transit throughsomatic Golgi in the absence of SAP97 orCASK, but may accumulate as a result of aslowed transport rate. To distinguish betweenthese possibilities, we examined whetherblocking Golgi trafficking with ARF1-Q71Ifurther enhanced NMDAR levels in thesomatic Golgi of neurons lacking SAP97.Neurons expressing both SAP97 shRNAs andARF1-Q71I–HA showed a marked increase inNR1 immunofluorescence in somatic Golgi(Fig. 6g,h) compared with neurons withSAP97 shRNAs alone (Fig. 6a). The averageoverlap between NR1 and GM130 inten-sities in somata was increased by 80% inneurons with ARF1-Q71I–HA and SAP97knockdown (Fig. 6h) compared with 33% in neurons withoutARF1-Q71I–HA (Fig. 6b).

Finally, we examined whether NMDARs still accumulated in Golgioutposts when Golgi trafficking was blocked by ARF1-Q71I in neuronslacking SAP97. Endogenous NR1 accumulated at ARF1-Q71I–HA–positive puncta in the dendrites of neurons expressing a scrambledshRNA, but we observed a 65% (P ¼ 0.00004) reduction in the amountof NR1 present at these same sites in neurons expressing the SAP97shRNA (Fig. 6i).

To address whether the alternative secretory pathway taken byNMDARs contributes to their synaptic delivery, we assessed whetherSAP97 or CASK knockdown altered synaptic levels of NMDARs. Thiswas tested by measuring the average intensity of NR1 immunoreactivepuncta colocalizing with synaptophysin in neurons infected with theshRNAs (Fig. 7). The normalized ratio of NR1 to synaptophysinintensity was 30% less in neurons expressing SAP97 shRNA and

46% less in those with CASK shRNA when compared with un-infected neurons or neurons expressing a scrambled shRNA sequence(Fig. 7). These results are consistent with the 45% decrease in NMDAREPSCs that was observed previously with SAP97 knockdown inhippocampal neurons39. Furthermore, they correlate with the33% and 48% increases in NMDAR somatic Golgi accumulationthat we observed using the same SAP97 and CASK knock-downs (Fig. 6a,b). These effects were not the results of alterations insynapse formation or stability, as neither SAP97 nor CASK knock-down yielded a change in synaptic density versus scrambled shRNAcontrols (Supplementary Fig. 6). Taken together, these data indicatethat NMDARs traffic to synapses via the conventional secretorypathway in the absence of SAP97 or CASK, but are more efficientlydelivered to synapses when they utilize an atypical pathwaythrough the endoplasmic reticulum subcompartment and dendriticGolgi outposts.

NR1Synaptophysin NR1:synaptophysin

SAP97 shRNA

a

c

Rel

ativ

e N

R1:

syna

ptop

hysi

n in

tens

ity

Uninfe

cted

Scram

bled

shRNA

SAP97 sh

RNA

Uninfe

cted

Scram

bled

shRNA

CASK shRNA

*

NR1Synaptophysin NR1:synaptophysin

CASK shRNA

b

*

1.2

1.0

0.8

0.6

0.4

0.2

0

Figure 7 Knockdown of SAP97 or CASK reduces synaptic NMDARs. Hippocampal neurons were infected

shortly after plating, at 0 DIV, with lentivirus encoding shRNAs specific to SAP97 or CASK, respectively,

or a lentivirus with a randomized shRNA sequence (scrambled). (a) Representative images from anuninfected sister culture (top row) and cultures infected with a high titer of SAP97 shRNA lentivirus

(bottom row). Cultures were fixed at 16 DIV and stained with antibodies to NR1 and synaptophysin.

The relative intensities of the synaptic NR1 puncta in each field are shown in the third panels

(NR1:synaptophysin, pseudo-color intensity image). These were obtained by thresholding and dividing

the synaptic NR1 image by the synaptophysin image (see Online Methods). The resultant pixels were

displayed in a pseudo-color channel (progression from blue to red to yellow pixels reflects increasing

intensities). (b) Representative images from an uninfected sister culture (top row) and those infected

with a high titer of CASK shRNA lentivirus (bottom row). (c) Quantification of the decrease in synaptic

NMDAR levels following SAP97 or CASK knockdown. Fluorescence intensities of overlapping NR1 and

synaptophysin puncta were measured. The normalized ratios of NR1 to synaptophysin average intensities

are shown (mean ± s.e.m. from ten fields per group, n ¼ 1,103, 1,108 and 881 puncta from

uninfected, scrambled and SAP97 shRNA cultures, respectively, P ¼ 0.001; n ¼ 1,524, 1,294 and

840 puncta from uninfected, scrambled and CASK shRNA cultures, respectively, P ¼ 0.0000003).

Scale bar represents 10 mm.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1017

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

DISCUSSION

NMDARs utilize an atypical secretory pathway

With their long processes and polarized structures, neurons faceunparalleled demands to correctly sort and target proteins to theappropriate destinations. As with other cells, neurons have partiallysolved this problem by sorting proteins into vesicles at the TGN beforetransporting them to axons or dendrites40–42. Sorting of postsynapticglutamate receptors was initially thought to only occur at theTGN. However, recent studies have demonstrated that sorting ofAMPARs utilizes an endosomal pathway after their insertion at ornear the PSD10.

Here, we found yet another pathway for glutamate receptor traffick-ing, which involves a dendritic endoplasmic reticulum subcompart-ment and Golgi outposts and is used by NMDARs, but not AMPARs.This trafficking pathway requires SAP97 and CASK and appears to beessential for the efficient synaptic delivery of NMDARs. Our data revealthree previously unknown aspects of NMDAR trafficking. First, themajority of NMDARs use this alternative pathway, as no increase inNMDAR immunoreactivity or NR1-GFP fluorescence was seen insomatic Golgi following a block of Golgi trafficking by ARF1-Q71I.Second, NMDAR trafficking from the soma to dendrites was indepen-dent of COPI vesicle formation and NMDARs remained in anendoplasmic reticulum–like compartment during their traffickinginto dendrites. Finally, NMDAR transit through dendritic Golgi out-posts was dependent on COPI vesicle formation, suggesting thatGolgi outposts function for NMDARs similar to somatic TGN func-tions for other proteins; that is, as a transit and processing station. Thismay enable neurons to more tightly and locally regulate the delivery ofsynaptic NMDARs.

SAP97 and CASK direct NMDARs to Golgi outposts

Our data indicate that NMDARs exit the conventional secretory path-way just after their exit from the somatic endoplasmic reticulumcompartment. This prediction is supported by our 15 1C temperatureshift experiments, which led to the somatic accumulation of newlysynthesized NMDARs (Supplementary Fig. 1). Furthermore, wefound that a substantial fraction of dendritically localized, extrasynap-tic NMDARs colocalized with endoplasmic reticulum markers, includ-ing KDEL, IP3 receptor and DsRed-ER (Fig. 2). Notably, theseNMDAR-associated, endoplasmic reticulum–derived vesicles werefound to be highly dynamic, moving at speeds consistent with theirtransport along dendritic microtubules31, and to share features withpreviously characterized transport vesicles14 containing NMDARs,CASK and KIF17.

By further characterizing endoplasmic reticulum–derived NMDARtransport vesicles, we found that SAP97 associated with NMDAR-CASK-KIF17 complexes. SAP97 directly binds CASK and NMDARs23,35,36 andcauses endoplasmic reticulum retention of voltage-gated ion channelsin HEK293 cells37. These data suggest that SAP97 is necessary forsorting NMDARs into the alternative secretory pathway. Three linesof evidence support this hypothesis. First, SAP97 caused the endo-plasmic reticulum retention of NMDARs when it was transfected withNMDAR subunits into HEK293 cells (Fig. 3). Second, SAP97 knock-down in neurons caused NMDAR accumulation in somatic Golgi(Fig. 6). Finally, endogenous NMDAR localization in dendritic Golgioutposts was lost with SAP97 knockdown and ARF1-Q71I–HAblock of Golgi trafficking (Fig. 6). These data suggest that SAP97associates with NMDARs during/after their assembly in the endo-plasmic reticulum and that this interaction prevents their ability totransit from endoplasmic reticulum exit sites to the somatic Golgi(Supplementary Fig. 7).

Our data also indicate that SAP97 does not act alone in divertingNMDARs from somatic Golgi, although it has the ability to directlybind NR2 subunits of the NMDAR23. CASK, a second component ofthe NMDAR transport complex, is also required for this process,although it does not directly bind NMDAR subunits. Two sets ofexperiments support this conclusion. First, CASK coexpression enabledNMDAR transport to the cell surface and prevented SAP97-dependentendoplasmic reticulum retention of NMDARs (Fig. 3). This activityrequired the L27N domain of CASK, which is necessary for complexformation with SAP97 (ref. 35). Second, CASK knockdown in culturedhippocampal neurons re-routed NMDARs through somatic Golgi(Fig. 6). These data indicate that a CASK-SAP97 complex functionsin concert to sort NMDARs through this atypical secretory pathway.

Functional importance of the alternative secretory pathway

We evaluated whether levels of synaptic NMDARs were altered byknockdown of SAP97 or CASK and found that knockdown of eitherreduced the amount of synaptic NMDARs by 30–40% (Fig. 7). Thesevalues are consistent with the 45% decrease in NMDAR EPSCs that wasobserved previously with SAP97 knockdown39. A similar reduction insynaptic NMDARs was reported for neurons lacking KIF17, themicrotubule-dependent motor that is thought to be responsible forthe anterograde transport of vesicles carrying NMDARs/CASK17.When this alternative pathway is unavailable, NMDARs can traffic tosynapses via the conventional secretory pathway, as evidenced by theaccumulation of NMDARs in somatic Golgi during the knockdown ofSAP97 or CASK (33% and 48%, respectively; Fig. 6), as well as by the80% increase in the number of NMDARs present on somatic Golgiwhen SAP97 is knocked down or ARF1-Q71I is used to block Golgitrafficking (Fig. 6). We also detected a reciprocal relationship betweenthe increase in NMDARs at the Golgi (33% and 48%; Fig. 6) and theirreduction at synapses (30% and 46%; Fig. 7) during the SAP97 orCASK knockdown, respectively. Our data suggest that loss of synapticNMDARs is caused by a reduced rate of trafficking as they movethrough the conventional secretory pathway and is not by reducedsynaptic anchoring of the receptors by SAP97 or CASK.

Another question raised by our studies is whether this alternativesecretory pathway is specific for NMDARs or whether other integralmembrane and/or secreted proteins traffic through this pathway.Clearly, proteins that have previously been shown to reside in thedendritic endoplasmic reticulum subcompartment, such as the integralmembrane IP3 receptor and the calcium ATPase-2a, are trafficked usingthis pathway31. Our studies on GluR1 subunits of the AMPAR indicatethat this pathway is not available to all proteins and that there is somedegree of specificity. Other studies have found that recombinant brain-derived neurotrophic factor, a secreted protein, and vesicular stomatitisvirus G protein, an integral membrane protein, traffic though dendriticGolgi outposts, although both proteins also appear to use somaticGolgi, perhaps as a result of overexpression29. Inwardly rectifyingpotassium channels (Kir2) also complex with SAP97, CASK, Mintand Velis/MALS in fibroblast and neuronal cells36. Although notexplored here, the ability of SAP97 and CASK to form complexeswith channels other than NMDARs suggests that this alternativesorting pathway may be used by multiple dendritically sorted ionchannels and receptors.

In summary, we found that the differential sorting of NMDARs andAMPARs begins in the somatic endoplasmic reticulum, causing amajority of NMDARs to bypass somatic Golgi in favor of dendriticGolgi outposts. Our data indicate that the MAGUKs SAP97 and CASKare required for the sorting of NMDARs from AMPARs. Finally, ourfindings suggest that this alternative pathway allows NMDARs to be

1018 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

more efficiently delivered to synapses and may provide a platform forlocal control of NMDAR insertion near synapses.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe would like to thank S. Leal-Ortiz for assistance with the generation of lentivirusand shRNA, L.A. Needleman and J. Marks. Also, thanks to B. Margolis for theCASK constructs, J. Lippincott-Schwartz for the Galtase-GFP construct andV. Bindokas for help and advice with imaging protocols. This work was supportedby the Nancy Pritzker Family and US National Institutes of Health grant DA016758to C.C.G., a Deutsche Forschungsgemeinschaft postdoctoral fellowship to C.G.S.,The Marsden Fund (Royal Society of New Zealand) to J.M.M., US NationalInstitutes of Health grants NS043782, DA13602 and DA019695 to W.N.G.,and the Albert & Ellen Grass Faculty Award to W.N.G. and J.M.

AUTHOR CONTRIBUTIONSO.J. conducted the majority of the experiments and data analysis and co-wrotethe manuscript. C.L.W. contributed to the shRNA experiments and helped in theediting of the manuscript. C.G.S. designed, constructed and aided in the testingof the shRNAs. S.F. performed the electron microscopy studies. M.S. carriedout temperature block, BFA and some of the live-imaging experiments. E.L.performed flow cytometry experiments. J.M. provided advice and aided in theinterpretation of data. C.A. supervised the electron microscopy studies. T.d.S.assisted with the temperature block experiments. J.M.M. supervised and designedstudies and helped with the editing of the manuscript. C.C.G. and W.N.G.supervised the project and co-wrote the manuscript.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Dumas, T.C. Developmental regulation of cognitive abilities: modified composition of amolecular switch turns on associative learning. Prog. Neurobiol 76, 189–211 (2005).

2. Malinow, R. & Malenka, R.C. AMPA receptor trafficking and synaptic plasticity. Annu.Rev. Neurosci. 25, 103–126 (2002).

3. Malenka, R.C. & Bear, M.F. LTP and LTD: an embarrassment of riches. Neuron 44, 5–21(2004).

4. Rao, A., Kim, E., Sheng, M. & Craig, A.M. Heterogeneity in the molecular composition ofexcitatory postsynaptic sites during development of hippocampal neurons in culture.J. Neurosci. 18, 1217–1229 (1998).

5. Friedman, H.V., Bresler, T., Garner, C.C. & Ziv, N.E. Assembly of new individual excitatorysynapses: time course and temporal order of synaptic molecule recruitment. Neuron 27,57–69 (2000).

6. Washbourne, P., Bennett, J.E. & McAllister, A.K. Rapid recruitment of NMDA receptortransport packets to nascent synapses. Nat. Neurosci. 5, 751–759 (2002).

7. Montgomery, J.M., Zamorano, P.L. & Garner, C.C. MAGUKs in synapse assembly andfunction: an emerging view. Cell. Mol. Life Sci. 61, 911–929 (2004).

8. Kennedy, M.J. & Ehlers, M.D. Organelles and trafficking machinery for postsynapticplasticity. Annu. Rev. Neurosci. 29, 325–362 (2006).

9. Greger, I.H. & Esteban, J.A. AMPA receptor biogenesis and trafficking. Curr. Opin.Neurobiol. 17, 289–297 (2007).

10. Park, M., Penick, E.C., Edwards, J.G., Kauer, J.A. & Ehlers, M.D. Recycling endosomessupply AMPA receptors for LTP. Science 305, 1972–1975 (2004).

11. Yudowski, G.A. et al. Real-time imaging of discrete exocytic events mediating surfacedelivery of AMPA receptors. J. Neurosci. 27, 11112–11121 (2007).

12. Cognet, L., Groc, L., Lounis, B. & Choquet, D. Multiple routes for glutamate receptortrafficking: surface diffusion and membrane traffic cooperate to bring receptors tosynapses. Sci. STKE 2006, pe13 (2006).

13. Sheng, M. & Hoogenraad, C.C. The postsynaptic architecture of excitatory synapses: amore quantitative view. Annu. Rev. Biochem. 76, 823–847 (2006).

14. Setou, M., Nakagawa, T., Seog, D.H. & Hirokawa, N. Kinesin superfamily motor proteinKIF17 and mLin-10 in NMDA receptor-containing vesicle transport. Science 288,1796–1802 (2000).

15. Sans, N. et al. Synapse-associated protein 97 selectively associates with a subset ofAMPA receptors early in their biosynthetic pathway. J. Neurosci. 21, 7506–7516(2001).

16. Wyszynski, M. et al. Interaction between GRIP and liprin-alpha/SYD2 is required forAMPA receptor targeting. Neuron 34, 39–52 (2002).

17. Guillaud, L., Setou, M. & Hirokawa, N. KIF17 dynamics and regulation of NR2Btrafficking in hippocampal neurons. J. Neurosci. 23, 131–140 (2003).

18. Lise, M.F. et al. Involvement of myosin Vb in glutamate receptor trafficking. J. Biol.Chem. 281, 3669–3678 (2006).

19. Setou, M. et al. Glutamate receptor-interacting protein GRIP1 directly steers kinesin todendrites. Nature 417, 83–87 (2002).

20. Perestenko, P.V. & Henley, J.M. Characterization of the intracellular transport of GluR1and GluR2 alpha-amino-3-hydroxy-5-methyl-4-isoxazole propionic acid receptor sub-units in hippocampal neurons. J. Biol. Chem. 278, 43525–43532 (2003).

21. Sans, N. et al. NMDA receptor trafficking through an interaction between PDZ proteinsand the exocyst complex. Nat. Cell Biol. 5, 520–530 (2003).

22. Sans, N. et al. mPins modulates PSD-95 and SAP102 trafficking and influences NMDAreceptor surface expression. Nat. Cell Biol. 7, 1179–1190 (2005).

23. Bassand, P., Bernard, A., Rafiki, A., Gayet, D. & Khrestchatisky, M. Differentialinteraction of the tSXV motifs of the NR1 and NR2A NMDA receptor subunits withPSD-95 and SAP97. Eur. J. Neurosci. 11, 2031–2043 (1999).

24. Gardoni, F. et al. CaMKII-dependent phosphorylation regulates SAP97/NR2A inter-action. J. Biol. Chem. 278, 44745–44752 (2003).

25. Mauceri, D., Gardoni, F., Marcello, E. & Di Luca, M. Dual role of CaMKII-dependentSAP97 phosphorylation in mediating trafficking and insertion of NMDA receptor subunitNR2A. J. Neurochem. 100, 1032–1046 (2007).

26. Pierce, J.P., Mayer, T. & McCarthy, J.B. Evidence for a satellite secretory pathway inneuronal dendritic spines. Curr. Biol. 11, 351–355 (2001).

27. Pfeffer, S. Membrane domains in the secretory and endocytic pathways. Cell 112,507–517 (2003).

28. Dascher, C. & Balch, W.E. Dominant inhibitory mutants of ARF1 block endoplasmicreticulum to Golgi transport and trigger disassembly of the Golgi apparatus. J. Biol.Chem. 269, 1437–1448 (1994).

29. Horton, A.C. & Ehlers, M.D. Dual modes of endoplasmic reticulum-to-Golgi transport indendrites revealed by live-cell imaging. J. Neurosci. 23, 6188–6199 (2003).

30. Schuman, E.M., Dynes, J.L. & Steward, O. Synaptic regulation of translation of dendriticmRNAs. J. Neurosci. 26, 7143–7146 (2006).

31. Bannai, H., Inoue, T., Nakayama, T., Hattori, M. & Mikoshiba, K. Kinesin-dependent,rapid, bi-directional transport of ER subcompartment in dendrites of hippocampalneurons. J. Cell Sci. 117, 163–175 (2004).

32. Mironov, S.L. & Symonchuk, N. ER vesicles and mitochondria move and communicateat synapses. J. Cell Sci. 119, 4926–4934 (2006).

33. Jo, K., Derin, R., Li, M. & Bredt, D.S. Characterization of MALS/Velis-1, -2, and -3: afamily of mammalian LIN-7 homologs enriched at brain synapses in association withthe postsynaptic density–95/NMDA receptor postsynaptic complex. J. Neurosci. 19,4189–4199 (1999).

34. Butz, S., Okamoto, M. & Sudhof, T.C. A tripartite protein complex with the potential tocouple synaptic vesicle exocytosis to cell adhesion in brain. Cell 94, 773–782 (1998).

35. Lee, S., Fan, S., Makarova, O., Straight, S. & Margolis, B. A novel and conserved protein-protein interaction domain of mammalian Lin-2/CASK binds and recruits SAP97 to thelateral surface of epithelia. Mol. Cell. Biol. 22, 1778–1791 (2002).

36. Leonoudakis, D., Conti, L.R., Radeke, C.M., McGuire, L.M. & Vandenberg, C.A. Amultiprotein trafficking complex composed of SAP97, CASK, Veli and Mint1 isassociated with inward rectifier Kir2 potassium channels. J. Biol. Chem. 279,19051–19063 (2004).

37. Tiffany, A.M. et al. PSD-95 and SAP97 exhibit distinct mechanisms for regulating K+

channel surface expression and clustering. J. Cell Biol. 148, 147–158 (2000).38. Leonard, A.S., Davare, M.A., Horne, M.C., Garner, C.C. & Hell, J.W. SAP97 is associated

with the alpha-amino-3-hydroxy-5-methylisoxazole-4-propionic acid receptor GluR1subunit. J. Biol. Chem. 273, 19518–19524 (1998).

39. Nakagawa, T. et al. Quaternary structure, protein dynamics and synaptic function ofSAP97 controlled by L27 domain interactions. Neuron 44, 453–467 (2004).

40. Dotti, C.G. & Poo, M.M. Neuronal polarization: building fences for molecular segrega-tion. Nat. Cell Biol. 5, 591–594 (2003).

41. Dresbach, T. et al. Assembly of active zone precursor vesicles: obligatory trafficking ofpresynaptic cytomatrix proteins Bassoon and Piccolo via a trans-Golgi compartment.J. Biol. Chem. 281, 6038–6047 (2006).

42. Tang, B.L. Protein trafficking mechanisms associated with neurite outgrowth andpolarized sorting in neurons. J. Neurochem. 79, 923–930 (2001).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1019

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSAntibodies. In our immunofluorescence and western blot experiments, we

used primary antibodies to HA (monoclonal antibody HA.11, Covance), c-Myc

(9E10 monoclonal antibody and a polyclonal antibody, Santa Cruz), GFP

(polyclonal, Invitrogen), ERp57 (polyclonal, StressGen), NR2B (monoclonal,

BD Biosciences), chicken and rabbit antibodies to HA (polyclonal, Bethyl),

KDEL (monoclonal, StressGen), IP3R (polyclonal, Affinity Bioreagents), NR1

(monoclonal antibody 363, Chemicon), synapsin (monoclonal, Chemicon),

SAP97 (polyclonal, Affinity Bioreagents), CASK (polyclonal, Zymed; mono-

clonal, BD Biosciences), KIF17 (polyclonal, Sigma), GluR1 (polyclonal, Cal-

biochem), tubulin (monoclonal, Sigma), TGN38 (monoclonal, BD

Biosciences), GM130 (polyclonal, Calbiochem), MAP2 (polyclonal, Sigma),

PSD95 (monoclonal, Neuromab), and synaptophysin (polyclonal, Santa Cruz).

We obtained secondary antibodies from Molecular Probes (goat antibody to

mouse and rabbit Alexa Fluors 488, 568 and 647, and goat antibody to chicken

DSB-X biotin and streptavidin–Marina Blue) and Aurion (10-nm and 0.8-nm

colloidal gold–conjugated goat antibody to rabbit IgG and 0.8-nm colloidal

gold–conjugated goat antibody to mouse IgG).

Cell culture and transfection. HEK-293 cells (a gift from J. Kyle, University of

Chicago) were maintained in DMEM supplemented with 10% calf serum

(Hyclone). Cells were transiently transfected with cDNA using a calcium

phosphate protocol43. Cells transfected with HA-GluR1 or HA-NMDAR were

maintained in media containing 1 mM kynurenic acid (Sigma) or 100 mM

D(-)-2-amino-5-phosphonovaleric acid (Sigma) and 10 mM MK-801 (RBI),

respectively, to prevent excitotoxicity.

Hippocampal cultures were prepared using a modified Banker culture

protocol44. Briefly, hippocampi from embryonic (E18–19) Sprague-Dawley

rats were dissected, dissociated in 0.05% trypsin (vol/vol, Invitrogen), and cells

were plated at a density of 165 cells per mm2 on poly-L-lysine–coated coverslips

(Carolina Biological). Coverslips were transferred in pairs to 60-mm dishes

containing a glial feeder layer 1 h after plating, where they were inverted (to

maximize neuronal contact with secreted glial factors) and maintained in

Neurobasal medium containing B27 and GlutaMAX (all from Invitrogen).

Neuronal cultures were transfected at 12–14 DIV with the Lipofectamine

2000 transfection reagent (Invitrogen) according to manufacturer’s recommen-

dations, with the exception that 1–2.5 mg of each cDNA in 62.5 ml of

Neurobasal media and 2.5 ml of Lipofectamine 2000 in 62.5 ml of Neurobasal

media were mixed and added to coverslips in 6-well plates.

cDNA constructs. Rat NR1 and NR2B were obtained from J. Boulter (Uni-

versity of California, Los Angeles). All subunits were tagged at the NH2

terminus with the HA (YPYDVPDYA) epitope using the extension overlap

PCR method45. These cDNAs were then subcloned into a pCB6 mammalian

expression vector. Myc-CASK and Myc-CASKDL27 were gifts from B. Margolis

(University of Michigan). GluR1-GFP was a gift from R. Malinow (Cold Spring

Harbor). Construction of NR1-GFP and GFP-SAP97 has been previously

described46,47. YFP-SAP97 and RFP-SAP97 were generated by excising SAP97

from pEGFP-C1 (Clontech) with EcoRI and KpnI and subcloning into pEYFP-

C1 and pDsRed1-C1 (Clontech), respectively. ARF1-Q71I–HA was a gift from

A. El-Husseini (University of British Columbia). Galactosyltransferase-GFP

(Galtase-GFP) was a gift from J. Lippincott-Schwartz (US National Institutes of

Health), and pDsRed-ER is from Clontech Laboratories.

Immunofluorescence staining. HEK-293 cells (18–24 h post-transfection) and

neuronal cultures were washed twice in phosphate-buffered saline (PBS) at

22–25 1C, fixed in 4% paraformaldehyde/sucrose (vol/vol, 4 1C, 15 min) and

washed three times in PBS (5–10 min). For permeabilization, cells were

incubated in 0.1% Triton-X in TBS (10 min), incubated in blocking solution

(2% glycine (wt/vol), 2% BSA (wt/vol), 0.2% gelatin (wt/vol) and 50 mM

NH4Cl in 1� PBS; 10 min) and then incubated with the indicated primary

antibody diluted in blocking solution (1 h). Following primary antibody

incubation, cells were washed three times in blocking solution (5–10 min)

and overlaid with an appropriate secondary antibody diluted in blocking

solution (1 h). Cells were then washed three times in PBS (5–10 min) and

the slips mounted in Vectashield (Vector Laboratories) or Prolong Gold

(Invitrogen). Fluorescence images were acquired using the Leica SP2 AOBS

spectral laser scanning confocal microscope (Leica Microsystems) or with a

Yokogawa spinning disc confocal head (Perkin Elmer), fitted on a Zeiss

Axiovert 200M microscope. Images were processed using Image J (US National

Institutes of Health) and Adobe Photoshop software.

Image analysis. Quantification of fluorescence data was performed using

MetaMorph (Universal Imaging) and ImageJ software (US National Institutes

of Health). Somatic Golgi localization was assayed by setting the z plane limits

for acquisition (B0.5–1.0 mm per slice, 5–10 slices) on the basis of the

fluorescence signal for the Golgi marker (GM130 or TGN38). The signal was

then thresholded and used to create a binary mask, allowing for the measure-

ment of subunit pixel intensities in each plane. Averaged pixel intensities for the

entire z stack were calculated, background subtracted and normalized to

control values. Expression levels in exogenous subunit experiments were

controlled for by measuring the average pixel intensity in a small box drawn

outside of the Golgi, where subunits showed an even distribution. A Golgi-to-

soma (non-Golgi) ratio was calculated for each cell and normalized to control

ratio values. Experiments were conducted from a minimum of two indepen-

dent culture preparations, with ten neurons per experimental group.

Analysis of Golgi outpost localization was performed as described above,

with the exception that averaged z projections for all channels were used for

quantification. A combined Golgi and mutant ARF1 binary mask was created

(Golgi alone for control fields), ensuring the analysis of outposts with impaired

COPI function. Averaged NR1 and NR2B pixel intensities in the thresholded

regions were measured, background subtracted and normalized to control

values. For exogenous NR1-GFP experiments, differences in protein expression

levels across cells were controlled for by measuring average pixel intensities in a

small region of the immediate proximal dendrite where subunit distribution

was uniform, calculating an outpost to dendrite ratio for each cell analyzed

and normalizing to control ratio values. Experiments were conducted from

a minimum of two independent culture preparations, with five fields

per condition.

Synaptic NR1 expression was quantified by setting z plane limits using the

synaptophysin signal, then acquiring both channels. Averaged z projections of

image stacks were created, thresholded and scored for colocalization and signal

intensity using the cell-scoring feature in MetaMorph. Averaged intensities in

each field were background-subtracted and normalized, and the ratio of NR1 to

synaptophysin intensities was analyzed in Excel. Experiments were conducted

from a minimum of two independent culture preparations, with ten fields

per condition.

Analysis of the colocalization of DsRed-ER–containing vesicles with endo-

genous NR2B or GluR1 were carried out by background-subtracting and

thresholding image fields so that only puncta that were twofold greater than

background were selected. Colocalizing puncta were evaluated using the

Analyze Particles function in ImageJ.

Analysis of the effects of mutant ARF1 expression on GFP-GluR1 dendritic

distribution (Supplementary Fig. 2) was performed on Z stacks of dendrites.

Averaged pixel intensities along the first 20 mm were calculated, background-

subtracted and normalized to control values. Expression levels in individual

cells were controlled for as described above. Experiments were conducted from

a minimum of two independent culture preparations, with ten neurons per

experimental group.

Synaptic density experiments (Supplementary Figs. 3 and 6) used coloca-

lization of pre- and postsynaptic markers to denote synaptic sites. Thresholded

puncta were analyzed using the Analyze Particle function in ImageJ, and the

number of colocalizing puncta were divided by the length of dendrite analyzed.

Experiments were conducted from a minimum of two independent culture

preparations, with 5–10 neurons per experimental group.

Statistics. Statistical comparisons for all confocal analyses were made using

two-tailed Student’s t tests. Flow cytometric comparisons were made using

ANOVA/Tukey post hoc analysis.

Spectral separation of GFP and YFP. Samples were sequentially excited in

spectral scan mode (Leica) at reduced levels, using the 488- and 514-nm laser

lines, and emission windows were visually optimized. GFP bleed-through into

the YFP channel was assessed by scanning neurons that only expressed the GFP

NATURE NEUROSCIENCE doi:10.1038/nn.2362

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

fusion protein, excited by the 514-nm laser line, and collecting data with the

emission window set at 535–565 nm. Under these conditions, GFP fluorescence

was negligible. YFP cross-talk into the GFP channel was assessed and controlled

for by scanning neurons that only expressed the YFP fusion protein, excited by

the 488-nm laser line, and setting the emission window at 490–515 nm. At the

laser intensities used, YFP cross-talk was negligible when the emission window

did not extend beyond 515 nm. The effectiveness of this protocol was

confirmed by the presence of mutually exclusive puncta in overlay images

from co-transfected neurons.

Time-lapse imaging. For the time-lapse imaging experiments, the culture

medium was replaced with a 10 mM HEPES buffer solution (pH 7.3). Neurons

were visualized under an inverted microscope (IX81, Olympus) and a 60�objective (NA 1.4, Olympus) using standard filter sets and an Hg lamp. Culture

slips were maintained at B37 1C by placing dishes in a heating chamber affixed

to the microscope stage. Sequential images were acquired, 100–1,000-ms

exposures, on a Retiga EXi chilled charged-coupled device (QImaging) under

the control of MetaMorph software.

Flow cytometry. We examined surface HA expression by labeling live cells 24 h

post-transfection with an antibody to HA, washing them twice in PBS and

labeling them with a fluorescence-conjugated secondary antibody. Fluorescence

quantification was carried out using an LSRII Flow Cytometer (488- and

647-nm laser, BD Biosciences) and analysis of flow data was performed using

FlowJo 7.2.4 software.

Electron and light microscopy. Two Sprague-Dawley rats, P7, were deeply

anesthetized with Nembutal (50 mg per kg). Brains were fixed by transcardial

perfusion with a mixture of aldehydes (4% paraformaldehyde and 1%

glutaraldehyde in PBS, vol/vol). Vibratome sections from these brains were

incubated in 1% hydrogen peroxide (vol/vol) and then 1% albumin-BSA

(wt/vol) in PBS for 30 min each to minimize background staining. Sections

were incubated in polyclonal rabbit antibody to SAP97 (1:500) for 2 d at

22–25 1C. Antigen-antibody complexes were visualized by applying 0.8-nm

gold-conjugated secondary antibodies (diluted 1:100). The gold particles were

silver intensified (Silver IntensEM kit, Amersham). Digital images were captured

at a magnification of 40� for light microscopy.

The SAP97-immunolabeled vibratome sections were subjected to osmium-

free tissue processing48, dehydrated and embedded in resin (Embed 812, EMS).

The CA1 field of hippocampus was sectioned at a thickness of 80 nm and

immunolabeled for the NR2B subunit of NMDARs using the post-embed gold

procedure49. Briefly, the grids holding the sections were incubated in a 1:20

dilution of rabbit antibody to NR2B (Upstate) overnight, followed by 1 h

incubation at a 1:40 dilution of 10-nm gold-conjugated secondary antibody

(Aurion). The grids were counterstained with lead citrate to enhance contrast.

The pyramidal cell layer of CA1 was examined by a electron microscope

(JEOL) along the tissue-Epon interface. Micrographs containing somata and

proximal dendrites were captured digitally (Hamamatsu CCD Camera) at a

magnification of 40,000�. To determine whether SAP97 colocalized with the

NR2B-subunit at a frequency above chance level, we drew a circle with a radius

of 60 nm (the maximum distance between two primary antibody–secondary

antibody complexes associated with a single antigenic site) around every SIG

label (SAP97; Fig. 5) that represented the region that could be ascribed to the

coexistence of NR2Bs with SAP97.

Immunoprecipitation and western blot. Cells were pelleted by brief centrifu-

gation, resuspended, washed once with PBS and solubilized in lysis buffer (150

mM NaCl, 5 mM EDTA pH 7.4, 50 mM Tris pH 7.4, 0.02% NaN3) containing

1% Triton X-100 + NEM (2 mM), phenylmethanesulphonylfluoride (2 mM),

leupeptin (10 mg ml–1), Na-Tosyl-Lys-chloromethylketone�HCl (10 mg ml–1),

chymotrypsin (10 mg ml–1) and pepstatin (10 mg ml–1). Following a 1 h

solubilization (4 1C), samples were centrifuged at 14,000 g for 30 min at 4 1C.

Subsequent analyses were performed using the Triton X-100–soluble fraction.

Immunoprecipitations were performed by overnight antibody incubation at

4 1C. Protein-antibody complexes were isolated by incubation with Protein

G–Sepharose for 3 h at 4 1C.

For immunoblotting, proteins separated by SDS-PAGE were transferred to

nitrocellulose membranes. After transfer, the nitrocellulose was blocked with

3% milk in wash buffer (10 mM Tris (pH 7.4), 0.05% Tween 20 (wt/vol) and

150 mM NaCl). Membranes were washed briefly in wash buffer and then

incubated for 1 h with primary antibodies. The blots were washed and

incubated with secondary antibody (goat antibody to mouse or rabbit horse-

radish peroxidase) at the appropriate dilution for 1 h. After washing,

membranes were treated with an enhanced chemiluminescent reagent (ECL,

Amersham) according to the manufacturer’s protocol and exposed to film.

Lentivirus and shRNA. The shRNA target sequences for SAP97, 5¢-GCA AGA

TAC CCA GAG AGC A-3¢ (rat, N terminal), and CASK, 5¢-ATC CAT GAG

CAG GGG CTG A -3¢ (rat) were subcloned into the lentiviral vector

FUGWH1(+). This is a modified version of the FUGW vector containing an

H1 promoter that drives expression of the hairpin sequences50.

43. Claudio, T. Stable expression of heterologous multisubunit protein complexes estab-lished by calcium phosphate– or lipid-mediated cotransfection. Methods Enzymol. 207,391–408 (1992).

44. Banker, G. & Goslin, K. Developments in neuronal cell culture. Nature 336, 185–186(1988).

45. Ho, S.N., Hunt, H.D., Horton, R.M., Pullen, J.K. & Pease, L.R. Site-directed mutagen-esis by overlap extension using the polymerase chain reaction. Gene 77, 51–59 (1989).

46. Wu, H., Reuver, S.M., Kuhlendahl, S., Chung, W.J. & Garner, C.C. Subcellular targetingand cytoskeletal attachment of SAP97 to the epithelial lateral membrane. J. Cell Sci.111, 2365–2376 (1998).

47. Bresler, T. et al. Postsynaptic density assembly is fundamentally different frompresynaptic active zone assembly. J. Neurosci. 24, 1507–1520 (2004).

48. Phend, K.D., Rustioni, A. & Weinberg, R.J. An osmium-free method of epon embedmentthat preserves both ultrastructure and antigenicity for post-embedding immunocyto-chemistry. J. Histochem. Cytochem. 43, 283–292 (1995).

49. Fujisawa, S. & Aoki, C. In vivo blockade of N-methyl-D-aspartate receptors induces rapidtrafficking of NR2B subunits away from synapses and out of spines and terminals inadult cortex. Neuroscience 121, 51–63 (2003).

50. Leal-Ortiz, S. et al. Piccolo modulation of Synapsin1a dynamics regulates synapticvesicle exocytosis. J. Cell Biol. 181, 831–846 (2008).

doi:10.1038/nn.2362 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Balanced gene regulation by an embryonic brain ncRNAis critical for adult hippocampal GABA circuitry

Allison M Bond1, Michael J W VanGompel1,3, Evgeny A Sametsky2, Mary F Clark1, Julie C Savage1,3,John F Disterhoft2 & Jhumku D Kohtz1

Genomic studies demonstrate that, although the majority of the mammalian genome is transcribed, only about 2% of these

transcripts are code for proteins. We investigated how the long, polyadenylated Evf2 noncoding RNA regulates transcription of

the homeodomain transcription factors DLX5 and DLX6 in the developing mouse forebrain. We found that, in developing ventral

forebrain, Evf2 recruited DLX and MECP2 transcription factors to important DNA regulatory elements in the Dlx5/6 intergenic

region and controlled Dlx5, Dlx6 and Gad1 expression through trans and cis-acting mechanisms. Evf2 mouse mutants had

reduced numbers of GABAergic interneurons in early postnatal hippocampus and dentate gyrus. Although the numbers of

GABAergic interneurons and Gad1 RNA levels returned to normal in Evf2 mutant adult hippocampus, reduced synaptic inhibition

occurred. These results suggest that noncoding RNA–dependent balanced gene regulation in embryonic brain is critical for proper

formation of GABA-dependent neuronal circuitry in adult brain.

The potential of the genome to code for functional noncoding RNAs(ncRNAs) is only beginning to be uncovered1,2. Although many ncRNAsbelong to classes of small regulatory RNAs, one subset, long, polyade-nylated ncRNAs (lpncRNAs), act cooperatively with protein partners3.We previously found that Evf2, a lpncRNA target of sonic hedgehog(SHH) signaling in the developing telencephalon, exhibits trans-actingtranscriptional cooperativity with DLX homeodomain proteins, increas-ing Dlx5/6 enhancer activity in a neural stem cell line4. Identification ofan ultraconserved Dlx5/6 intergenic DNA regulatory element5 has led tothe discovery of more than 1,000 ultraconserved DNA sequences nearimportant developmental regulators or transcription factors6–8. Thedomain of Evf2 that is necessary and sufficient for its transcription-regulating activity lies in this ultraconserved sequence at the 5¢ end ofEvf2 RNA4. The finding that Evf2 has transcription-regulating activity4

raised the possibility that subsets of ultraconserved DNA sequences aretranscribed and functional. Recently, additional ultraconserved brainlpncRNAs have been identified9, supporting the possibility that ultra-conserved ncRNAs constitute a new class of transcription-regulatingultraconserved ncRNAs (trucRNAs). In this study, we found, to the bestof our knowledge, for first time that the Evf2 trucRNA is important forgene regulation and the development of interneurons that produceGABA, the major inhibitory neurotransmitter in the brain.

The balance between excitation and inhibition in the brain is criticalfor proper function and is maintained by two major classes of neurons:excitatory projection neurons and inhibitory local circuit interneurons.Although excitation is primarily mediated by the neurotransmitterglutamate, GABA primarily mediates inhibition. Recently, multiple

regulatory roles of GABAergic interneurons have been identified10. Thedysfunction of GABA-regulated circuits has been implicated in differ-ent psychiatric disorders, such as schizophrenia, autism and Tourette’ssyndrome, as well as epilepsy. In methyl CpG–binding protein (Mecp2)mutant mice11–13, a model for the human autism spectrum disorder(ASD) Rett syndrome, GABA-dependent inhibitory cortical activitydecreases14. In dorsal lateral prefrontal cortex of schizophrenic patients,one of the most consistent findings is a reduction in GAD67, theenzyme responsible for GABA synthesis15. Therefore, multiple lines ofevidence implicate alterations of GABAergic function in a variety ofneurological diseases. Here, we found that a single noncoding RNAcontrols development of GABAergic interneurons and adult braincircuitry, making this mechanism a target for studying both develop-ment and disease.

RESULTS

Trans and cis -gene regulation by Evf2 in vivo

We designed Evf2 loss-of-function mice to determine the role of Evf2in vivo. Because of the overlap of Evf2 with key Dlx5/6 DNA regulatoryelements, we inserted transcription termination sites rather thanremoving DNA fragments (Fig. 1a). We subcloned a 19.4-kb fragmentspanning the Dlx5, Dlx6 and Evf genes from a mouse BAC. Wethen introduced a triple polyadenylation signal16 into exon 1 (Evf2TS;Fig. 1a). Southern analysis verified correct targeting in embryonic stem(ES) cell lines (Fig. 1b) and mice (data not shown). Evf2TS/TS mutantmice were fertile, lived for at least 1 year and were morphologicallyindistinguishable from wild-type littermate controls (data not shown).

Received 31 March; accepted 19 June; published online 20 July 2009; doi:10.1038/nn.2371

1Developmental Biology and Department of Pediatrics, Children’s Memorial Hospital and Feinberg School of Medicine, Northwestern University, Chicago, Illinois, USA.2Department of Physiology, Feinberg School of Medicine, Northwestern University, Chicago, Illinois, USA. 3Present address: Division of Reproductive Biology Research,Feinberg School of Medicine, Northwestern University, Chicago, Illinois, USA (M.J.W.V.); Neurosciences Department, Case Western Reserve University, Cleveland, Ohio, USA(J.C.S.). Correspondence should be addressed to J.D.K. (j-kohtz@northwestern.edu).

1020 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

In situ hybridization analysis showed that transcription stop insertioninto Evf exon1 eliminated Evf2, but not Evf1 or Dlx5 expression inembryonic day 13.5 (E13.5) ventral telencephalon (Fig. 1c–h). Real-time quantitative reverse-transcribed RNA PCR (qRT-PCR) of E13.5medial ganglionic eminence (MGE) tissue from Evf2TS/TS mice showedthat Dlx6 and 5 transcripts increased by eight and twofold, respectively(Fig. 1i). Despite the fact that Evf1 and Dlx5 transcription start sites areapproximately equidistant from the triple polyadenylation signal inser-tion site (Fig. 1a), Dlx5, but not Evf1, transcription was affected in Evf2mutants (Fig. 1i). Selective transcriptional effects supported the ideathat Evf2 loss, rather than insertion of the triple polyadenylation signalsequence, was responsible. If the triple polyadenylation signal insertionwas causing the observed transcriptional effects, all Dlx5/6 enhanceractivities would be expected to change, including Evf1.

To distinguish between trans and cis-dependent Evf2 RNA regulatoryeffects, we performed Evf2 electroporation into E12.5 Evf2TS/TS brainsat two different concentrations (Fig. 1j). At a lower Evf2 concentration(1 mg) Dlx5 expression decreased, whereas Dlx6 and Evf1 concentra-tions remain unchanged. At a higher Evf2 concentration (2 mg), theconcentrations of both Dlx5 and Dlx6 increased, whereas that of Evf1did not change. The ability of Evf2 to partially rescue Dlx5 increasesuggested that Evf2 trans-regulatory mechanisms were involved in Dlx5transcriptional control. The inability of ectopically expressed Evf2 torescue Dlx6 increase in Evf2TS/TS mutants supported the idea that Evf2reduced Dlx6 expression through anti-sense competition in cis, ratherthan by trans, mechanisms. At higher concentrations of Evf2 (2 mg), theconcentrations of both Dlx5 and Dlx6 increased, supporting previouslypublished results that Evf2 RNA can function as a transcriptionalactivator of Dlx5 and Dlx6 ei and eii (conserved intergenic DNAregulatory elements) when ectopically expressed4. Electroporation ofan Evf2siRNA construct into E12.5 brains also increased the levels of

Dlx5 transcripts (Supplementary Fig. 1), further supporting the ideathat an Evf2 trans-acting mechanism regulates Dlx5 expression.Together, these data suggested that Evf2-mediated transcriptionalcontrol was concentration-dependent, using both trans and cis regula-tory mechanisms in vivo.

Evf2 recruitment of DLX and MECP2 to intergenic enhancers

We recently showed that Evf2 forms a complex with DLX proteinsin vivo and acts as a transcriptional coactivator of DLX activity withboth target and homeodomain specificity in C17 cells4. In addition, weproposed a model in which Evf2 recruits DLX proteins to Dlx5/6intergenic enhancers. Here, however, we found that qRT-PCR analysisof mice lacking Evf2 (Fig. 1i) indicated that the levels of Dlx5 and Dlx6transcripts increased, suggesting that Evf2 has a negative, rather thanpositive, transcription-regulating role in vivo. Rescue experiments(Fig. 1j) suggested that increased levels of Dlx6 in Evf2TS/TS mutantsresulted from a loss of anti-sense interference in cis, whereas moresubtle repressive effects on Dlx5 transcription occurred in trans. Tofurther investigate the mechanism of Evf2-dependent transcriptionalcontrol, we used chromatin immunoprecipitation (ChIP) followed byquantitative PCR (ChIP-PCR) on wild-type and Evf2TS/TS mutantE13.5 MGE chromatin to determine whether Evf2 affects DLX bindingto Dlx5 and Dlx6 ei and/or eii. Qualitative ChIP assays using antibodiesto DLX1 and 2 previously identified Dlx5/6 ei and eii as specific bindingsites (Fig. 2a)17. In wild-type E13.5 MGE, pan-antibodies to DLXproteins4,18–21 recognized DLX protein–Dlx 5/6 enhancer (ei and eii)complexes (Fig. 2b). In Evf2TS/TS mutant chromatin, DLX did not bindto ei and eii (Fig. 2b), indicating that Evf2 is required for DLX proteinrecruitment to Dlx regulatory enhancers ei and eii.

Given previous reports that mice lacking both Dlx1 and 2 showreduced expression of Dlx5 and Dlx6 (ref. 22), failure of DLX protein

Dlx6Dlx5eii

1

ei

2 3 4 4*

Triple polyAtranscription stop

Evf1

Evf2TS/+

ApaL1

Evf2 Dlx5 Dlx6 Evf1 Nrp2

WT

Evf2TS/TS

Rea

l-tim

e qR

T-P

CR

E13

.5 M

GE

(fol

d in

crea

se)

0

2

4

6

8

10

12

Dlx2

WT Evf2TS/TS WT Evf2TS/TS

Evf2 Evf1

LGE

MGE

Dlx5

WT

a

c d e f g h

b

Dlx5 Dlx6 Evf1

Rea

l-tim

e qR

T-P

CR

E13

.5 M

GE

(fol

d in

crea

se)

0

0.5

1.0

1.5

2.0

2.5 Evf2TS/TS

Control1 µg Evf22 µg Evf2

i j

*

*

*

*

*

*

Evf2TS

~10.2 kb ~10.4 kb

TargetedWT

WT

Evf2TS/TS

Figure 1 Evf2TS/TS mice have increased Dlx5 and

Dlx6 expression in the embryonic brain. (a)

Schematic of targeting a triple polyadenylation

transcription stop site into Evf exon 1, truncating

Evf2 transcripts from 3.7 kb to 101 bp (Evf2TS),

but not Evf1, which is transcribed starting from

exon 3. Truncated Evf2TS transcript (101 bp)

completely lacks the ultraconserved ei region.Only the distance from the triple polyadenylation

signal insertion to the Dlx5 (B10.2 kb) and Evf1

(B10.4 kb) transcription start sites are shown to

scale. (b) ES cells that were used for making

Evf2TS/TS mice contained correctly targeted

transcription stop into Evf exon 1, as verified by

Southern analysis. (c–h) RNA in situ hybridization

of E13.5 coronal sections of wild-type (WT) and

Evf2TS/TS mutant telencephalon, probed with anti-

sense Evf2 (c,d), anti-sense Evf1 (e,f) or anti-

sense Dlx5 (g,h). (i) Quantitative real-time RT-PCR

analysis of E13.5 MGE from wild-type and Evf2TS/

TS mutants. Error bars represent s.e.m. Evf2

cDNA, P = 0.0055; Dlx5 cDNA, P = 0.044;

Dlx6 cDNA, P = 0.005; Mann Whitney U test.

(j) Quantitative real-time RT-PCR analysis of

E12.5 Evf2TS/TS mutant MGE electroporated with

2 mg pcDNA (control), 1 mg pcDNA-Evf2 and 1 mg

pcDNA), and 2 mg pcDNA-Evf2. Error barsrepresent s.e.m. Dlx5 cDNA: 1 mg electroporated

Evf2, P = 0.05; 2 mg electroporated Evf2,

P = 0.001; ANOVA Dunnett’s two-sided test.

Dlx6 cDNA, 2 mg electroporated Evf2, P = 0.041,

n = 4, ANOVA Tukey test. * P o 0.05, t test,

Mann Whitney or ANOVA.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1021

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

recruitment to Dlx5/6 ei and eii would be expected to decrease Dlx5/6transcription. However, the number of Dlx5/6 transcripts increased andDlx2 expression did not change (Fig. 1i). The report that a twofoldincrease in Dlx5 can occur in the adult prefrontal cortex on loss of thetranscriptional repressor Mecp2 (ref. 23) led us to investigate whetherloss of Evf2 in the embryonic brain can affect MECP2 binding in theDlx5/6 region. In the absence of Evf2, MECP2 (ref. 11), a DNA methyl-binding protein associated with repressed chromatin, did not bind ei oreii (Fig. 2c). One mechanism that was proposed for MECP2-mediatedtranscriptional repression is its ability to recruit HDAC, a histonedeacetylase responsible for chromatin inactivation24. However, loss ofMECP2 binding to ei in Evf2 mutants did not alter HDAC binding to ei(Fig. 2d), suggesting that MECP2 binding functions through analternate mechanism at this site and stage of development. Loss ofMECP2 binding at eii did reduce HDAC binding, suggesting that ei andeii differ in how MECP2 function affects transcriptional regulatoryactivity by these sites. Although site 2 was previously defined as anMECP2- and HDAC-binding site in the adult prefrontal cortex23,MECP2 bound minimally to site 2 at this time during development.In addition, HDAC bound to site 2 in wild-type embryonic MGE, butreduction of HDAC binding to site 2 in Evf2TS/TS mutants was not

statistically significant (P = 0.5; Fig. 2d). Therefore, increased Dlx5expression in Evf2 mutants may result from a loss of MECP2 binding toei and eii and subsequent loss of HDAC from eii.

Evf2 recruited DLX and MECP2 to ei and eii, affecting Dlx5/6enhancer activities in trans on specific nearby genes, such as Dlx5 butnot Evf1, and regulating Dlx6 transcription through anti-sense inhibi-tion in cis (Figs. 1 and 2). Several possibilities may explain how Dlx5,Dlx6 and Evf1 are transcribed in the absence of DLX binding to ei andeii. First, DLX1/2 may only be required for initial activation of Dlx5/6expression in an Evf2-independent manner; subsequent regulation ofDlx5/6 levels by DLX and MECP2 may then require Evf2. Second, otherDLX-binding sites may compensate in the absence of DLX ei/eiiinteractions. Third, the major role of DLX1 and 2 may be to preventMECP2 from binding ei/eii, acting through inhibition rather than asdirect activators.

Evf2 did not control DLX or MECP2 nuclear localization

The absence of DLX and MECP2 protein binding to ei and eii inEvf2TS/TS mice raised the possibility that Evf2 influences proteinstability and/or nuclear localization. In a direct test of the effect ofEvf2 on DLX2 protein stability in C17 neural cells, DLX2 protein levelsresulting from co-transfection with Evf2 did not change (Fig. 3a). Inaddition, DLX2 protein levels in Evf2TS/TS mutant embryonic gang-lionic eminences did not change compared with those of wild type(Fig. 3b). DLX protein distribution in wild type (Fig. 3c–e) wasindistinguishable from that in Evf2 mutant nuclei (Fig. 3f–h). Inaddition, transcript levels of neuropilin 2 (Nrp2), a direct target ofDLX1/2 (ref. 25), did not change in Evf2TS/TS mutants (Fig. 1i). If Evf2were controlling DLX protein stability, all DLX1/2 activities would be

20 kb10 3 mB

Dlx6 Dlx5

0

4

6

10

2

61 3 52

Rel

ativ

e en

richm

ent

(fol

d in

crea

se)

Rel

ativ

e en

richm

ent

(fol

d in

crea

se)

Rel

ativ

e en

richm

ent

(fol

d in

crea

se)

DLX

Mouse chromosome 6qA1

ei eii

0

4

Evf1 Evf2

0

2

3

1

61 3 52

HD

AC

1

4

0

4

6

8

2

61 3 52

ME

CP

2

4

10

WTEvf2TS/TS

* *

* * *

*

8

a

b

c

d

Figure 2 Loss of Evf2 affects DLX and MECP2 binding to Dlx5/6 intergenic

enhancers in E13.5 MGE. (a) Schematic showing the region in mouse

chromosome 6qA1 surrounding Dlx5/6 and Evf1/2. (b–d) Quantitative ChIP

of E13.5 MGE from wild types (black bars) and Evf2TS/TS mutants (gray bars)

using primers 1–6 across the Dlx5/6 region and pan-antibody to DLX

(primers 3 and 5, P = 0.025; b), antibody to MECP2 (primers 2, 3 and 5,

P = 0.025; c) or antibody to HDAC1 (primers 5 and 6, P = 0.025; d). qChIP-

PCR was performed on optimized primer sets using previously defined primersets (1 ¼ 15, 2 ¼ 24, 4 ¼ 27)23 and newly defined sets for ei (3, red), eii

(5, red) and external primer (6). Primer 2 (green) was identified as a MECP2/

HDAC-binding site for transcriptional repression in adult cortex23. Error bars

represent s.e.m. * P o 0.05, Mann Whitney U test. The schematic in a is

reversed from that shown for Evf2 transcription-stop insertion (Fig. 1) to align

with previously published primers23.

DLX DLX/DAPI

E13.5 mouse MGE

WT

Evf2TS/TS

PanDlx

β-actin

Evf2 +

Dlx2 Dlx2 Evf2a

c d e

f g h

WT

Evf2TS/T

S

Dlx2

β-actin

E13.5GE

b

i j

ME

CP

2/D

AP

I

MGE

MGE

VZSVZ

VZSVZ

Mantle

WT Evf2TS/TS

Figure 3 Evf2 does not affect DLX or MECP2 nuclear localization.

(a) Western blot analysis of DLX2 protein after transfection of a construct

containing Dlx2 cDNA and/or Evf2 cDNA. Transfection was performed in C17

mouse neural stem cells showing equal levels of DLX2 protein with or without

Evf2 RNA. We determined that Evf2 RNA was stable after transfection by

RT-PCR (data not shown). (b) Western analysis of Evf2TS/TS and wild-type

E13.5 ganglionic eminence extracts probed with antibody to DLX2 (gift of

D. Eisenstat, University of Manitoba) showed that loss of Evf2 did not affect

DLX2 protein levels. (c–h) Pan-antibody to DLX (green) stained nuclei that

were counterstained with DAPI (blue) in both wild-type and Evf2TS/TS E13.5

MGE. SVZ, subventricular zone; VZ, ventricular zone. Scale bars represent200 mm (c,f), 20 mm (d,g) and 10 mm (e,h). (i,j) Immunohistochemistry

using antibodies to MECP2 (green) and DAPI (blue nuclei) showed similar

nuclear localization in both wild-type and Evf2TS/TS MGE cells. Scale bar

represents 20 mm.

1022 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

affected, including Nrp2, which is not. These data support a mechan-ism in which Evf2 recruits DLX to ei and eii, rather than stabilizing DLXprotein or directing DLX nuclear localization.

We next asked whether the loss of MECP2 binding to ei/eii in Evf2mutants resulted from its effects on nuclear localization. Analysis ofMECP2 localization in developing brain shows a gradient of increasingexpression in neurons as they mature, with very little expression inimmature neurons26. However, expression in E13.5 MGE has not beenreported. Analysis of MECP2 in E13.5 MGE nuclei in which Evf2 andDLX are normally expressed showed nonhomogeneous, speckledMECP2 localization (Fig. 3i). In Evf2TS/TS mutants, MECP2 distribu-tion is indistinguishable from that in wild type (Fig. 3i,j), indicatingthat failure of MECP2 recruitment to ei and eii in Evf2 mutants did notresult from altered subcellular localization.

Reduced numbers of Evf2TS/TS hippocampal interneurons

The DLX homeodomain protein family is related to the DrosophilaDistal-less gene (Dll)27. Mice lacking Dlx1 and 2 have a substantial lossof GABAergic interneurons in cortex (B75%) and hippocampus(B90%) as a result of defective migration from embryonicMGE28–30. Loss of Evf2 affected Dlx5/6 expression and DLXfunction (Figs. 1 and 2) in E13.5 MGE, the source of the majorityof GABAergic interneuron precursors that will migrate to thehippocampus and dentate gyrus29,31. We therefore asked whether

GABAergic interneuron development in P2 hippocampus ordentate gyrus was affected in mice lacking Evf2.

Expression of Evf2 in developing ventral telencephalon/subventri-cular zone persisted until birth in P2 subventricular zone and cells thatappeared to be migrating to the hippocampus (Fig. 4a). However, Evf2was undetectable in wild-type postnatal day 2 (P2) hippocampus anddentate gyrus, as well as in adult subventricular zone or hippocampus(data not shown). In situ hybridization using a probe against Gad1, anenzyme necessary for converting glutamate to GABA, showed that thenumber of GABAergic interneurons in Evf2 mutant dentate gyrus andhippocampal CA1 and CA3 layers were reduced by 40–65%(Fig. 4b–d). Reduced Gad1 (the mouse gene coding for Gad1)expression in Evf2 mutants was accompanied by a reduction inGABA, as shown by GABA immunohistochemistry (Fig. 4e,f). In situhybridization against vesicular glutamate transporter 1 (vGlut1, alsoknown as Slc17a7), a gene expressed specifically in glutamatergicneurons showed that vGlut1 expression in Evf2TS/TS and wild-typehippocampus were similar (Fig. 4g,h). In addition, TUNEL staining ofEvf2TS/TS and wild-type hippocampus and dentate gyrus showedsimilar levels of cell death (Fig. 4i–k). Together, these data indicatethat Evf2 was required for proper GABAergic interneuron developmentand that fate transformation or increased cell death were not respon-sible for this reduction.

Evf2 regulated Gad1 in embryonic, but not adult brain

To further understand the role of Evf2 in GABAergic interneurondevelopment, we analyzed Gad1 RNA levels in E13.5 MGE, whereGABAergic interneuron precursors first arise. The level of Gad1transcripts decreased by B30% in Evf2TS/TS compared with wild-typeE13.5 MGE (Fig. 5a). Electroporation of Evf2 into Evf2TS/TS E12.5 MGErestored Gad1 levels by B30% compared with a pcDNA control(Fig. 5b). The ability of Evf2 to rescue decreased Gad1 levels inEvf2TS/TS mutants, suggesting that Evf2 activates Gad1 transcriptionthrough trans-acting mechanisms.

In the forebrain, Evf2 expression was limited to embryonic and earlypostnatal times of development (Figs. 1c and 4a) and was undetectablein adult subventricular zone or hippocampus (data not shown). Wenext asked whether reduced embryonic expression of Gad1 andreduction of GABAergic interneurons seen in early postnatal hippo-campus persisted into adulthood. At P60, Evf2TS/TS Gad1 transcript

Evf2TS/TS

Evf2TS /TS

WT

Gad

1

Evf2

Or

RadLMol

MolDGC

DG LVCA3

CA1

a

c d

e

i

j

f

g h

WT

Evf 2TS/TS

Gad

1 ne

uron

sno

rmal

ized

1.2

1.0

0.8

0.6

0.4

0.2

0DG CA3 CA1

GA

BA

b

DG

vGlu

t1

WT

TU

NE

L-po

sitiv

e ce

lls

3025

2015105

0DG CA3 CA1

35

k

TU

NE

L/D

AP

I WT

Evf

2T

S/T

S

**

*

Py

Figure 4 GABAergic interneuron loss in the P2 hippocampus and dentate

gyrus of Evf2TS/TS mutant mice. (a) RNA in situ hybridization analysis of Evf2

expression in the subventricular zone and cells lining the lateral ventricle as

they migrate to the hippocampus and dentate gyrus. Scale bar represents

130 mm. DG, dentate gyrus; DGC, dentate granule cell layer; LMol,

lacunosum molecular; LV, lateral ventricle; Mol, molecular layer of the

dentate gyrus; Or, oriens; Py, pyramidal; Rad, radiatum. (b) Quantification of

the number of Gad1-expressing GABAergic interneurons in the dentate gyrusand hippocampal CA1 and CA3 regions. Error bars represent s.e.m.

P ¼ 0.025 for dentate gyrus, CA3 and CA1, Mann Whitney U test (n ¼ 3

for wild type, n ¼ 3 for Evf2TS/TS). (c,d) RNA in situ hybridization of Gad1-

expressing interneurons in the hippocampus. Scale bar represents 135 mm.

(e,f) Immunohistochemistry with antibodies to GABA (green) and DAPI (blue

nuclei). Scale bar represents 200 mm. (g,h) RNA in situ hybridization of

vGlut1-expressing neurons in the hippocampal CA3 region showed similar

expression in wild type and Evf2TS/TS. Scale bar represents 80 mm.

(i,j) Immunohistochemistry using the apoptosis TUNEL assay (green) and

DAPI (blue nuclei) in the hippocampal CA3 region. Scale bar represents

10 mm. (k) Quantification of the number of TUNEL-positive cells in the

dentate gyrus and hippocampal CA1 and CA3 regions showed no difference

in cell death between wild types and Evf2TS/TS mutants. Error bars represent

s.e.m. P 4 0.05 for dentate gyrus, CA3 and CA1, Student’s t test (n ¼ 3 for

wild types, n ¼ 3 for Evf2TS/TS). * P o 0.05 (t test or Mann Whitney U test).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1023

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

levels were comparable with those in wild type (Fig. 5c). In addition,the number of GABAergic interneurons in 8-month-old Evf2TS/TS andwild-type hippocampus and dentate gyrus were similar (Fig. 5d). Thissuggests that Gad1 expression and GABAergic interneuron number inEvf2 mutants recovered to normal levels sometime between earlypostnatal and 2-month-old hippocampal development, temporallycorrelating with the timing of Evf2 downregulation.

Reduced synaptic inhibition in Evf2TS/TS pyramidal neurons

Dlx1�/� mice show a loss of a specific population of hippocampalGABAergic interneurons, which subsequently results in reduced synap-tic inhibition32. Analysis of Evf2TS/TS hippocampus not only showed agreater percentage loss of total GABAergic interneurons in earlypostnatal hippocampus (Fig. 4b) than Dlx1�/� mice32, but also showedthat GABAergic defects appeared earlier. In addition, unlike Dlx1�/�

mice, Evf2TS/TS mice showed recovery of GABAergic interneurons inolder animals (Fig. 5c,d). This led us to ask whether lower levels ofembryonic and perinatal Gad1 can cause long-lasting effects on adultsynaptic activity, despite apparent transcriptional recovery.

To answer this question, we compared inhibitory postsynapticcurrents (IPSCs) in CA1 pyramidal cells from Evf2TS/TS and wild-type littermates at ages older than P60, when Gad1 mRNA levels haverecovered to normal levels. First, we analyzed spontaneous inhibitorypostsynaptic currents (sIPSCs) and minimal inhibitory postsynapticcurrents (mIPSCs). To maximize outward inhibitory current in artifi-cial cerebrospinal fluid (ACSF) containing glutamate receptor antago-nists (20 mM 6-ciano-7-dinitroquinoxaline-2,3-dione (CNQX) and50 mM D(–)-2-amino-5-phosphonovaleric acid (D-AP5)), we recordedat +20 mV. We added tetrodotoxin (2 mM) to isolate mIPSCs(Fig. 6a,b). We divided all mice into two age groups: adult (3–5months old) and old (o12 months old) mice. The two age groupsallowed us to distinguish any persistent differences between mutantand wild-type mice from those that might be attributable to specificdevelopmental stages.

Both Evf2TS/TS age groups showed a significant reduction insIPSCs event frequency in CA1 pyramidal cells (adult, P = 0.018; old,P = 0.022). sIPSC mean frequency in the adult group was lower inmutant mice by 42% compared with those in wild-type mice. In oldmice, sIPSCs event frequency in mutant mice was lower by 38% than inwild types. We observed similar significant reductions in mIPSCs eventfrequencies in adult and old Evf2TS/TS groups compared with wild types(adult, P = 0.024; old, P = 0.039). Cumulative probability plots andcorresponding Kolmogorov-Smirnov statistical analysis further con-firmed the reduction in IPSC frequency in Evf2TS/TS mice (Fig. 6c,d). Incontrast, we did not find significant differences in sIPSP amplitude ormIPSP amplitude between wild-type and mutant mice (sIPSC: adult(P = 0.217), old (P = 0.417); mIPSC: adult (P = 0.675), old (P = 0.373);Fig. 6e,f). This suggested that Evf2 did not control the properties ofsynaptic contacts or GABA receptors formed by normally differentiatedinterneurons, but instead reduced the number of GABAergic synapsesthat formed on CA1 pyramidal neurons.

Notably, the frequency of sIPSCs and mIPSCs were lower in old micethan in adult mice, but the comparable age-dependent changesoccurred in both Evf2TS/TS and wild-type mice. Thus, the differencebetween mutant and wild-type mice persisted. Likewise, the amplitudeof sIPSCs increased in old mice. These age differences are consistentwith those reported previously33,34.

To further determine whether inhibitory input in CA1 pyramidalneurons were altered in the mutants, we measured the evoked IPSCs inCA1 pyramidal neurons from the adult mice group on stimulation ofthe stratum radiatum. First, we recorded evoked excitatory postsynap-tic currents (EPSCs) with amplitude 0.4–0.2 nA at �70mV, adjustingthe stimulus intensity as necessary. Glutamatergic antagonists blockedEPSCs (see above), and a series of evoked IPSCs were recorded atdifferent holding potentials (�70 mV to +20 mV). The I-V plotconstructed from these recordings was normalized to the amplitudeof evoked EPSCs at �70mV (Fig. 6g). We included only cells withsimilar biophysical characteristics in the analysis (wild type, n ¼ 5;mutant, n ¼ 4). Evoked IPSCs were significantly smaller in mutantmice (P o 0.01, paired t test), further suggesting that CA1 pyramidalneurons received less GABAergic innervation.

Finally, we did not find significant changes in the rise or decay timesfor spontaneous and evoked IPSCs (sIPSCs rise time, P = 0.363;mIPSCs decay time, P = 0.921; evoked IPSCs rise time, P = 0.594;evoked IPSCs decay time, P = 0.240). When added in the recordingACSF, 0.25 mM picrotoxin abolished both spontaneous and evokedIPSCs in all our recordings, confirming the evolvement of GABAA

receptors. Therefore, GABAergic interneuron recovery in older micedid not result in the recovery of normal synaptic inhibition inthe hippocampus.

Gad

1 qR

T-P

CR

E13

.5 M

GE

(fol

d in

crea

se)

0

0.2

0.4

0.6

0.8

1.0

Evf2TS/TS

1.2

WT

Gad

1 qR

T-P

CR

P60

hipp

ocam

pus

0

0.5

1.0

1.5

2.0

2.5

Evf2TS/TS

3.0

WT

Gad

1-ex

pres

sing

cel

ls

8-m

onth

-old

hip

poca

mpu

s no

rmal

ized

0

0.2

0.4

0.6

0.8

1.0

Evf2TS/TS

1.2

WT

DG CA3 CA1

a

c

Gad

1 qR

T-P

CR

E13

.5 M

GE

(fol

d in

crea

se)

0

0.2

0.4

0.6

1.2

1.4

Evf2TS/TS

pcDNA Evf2

0.8

1.0

1.6

Electroporatedconstruct

b

d

*

*

Figure 5 Evf2 trans-positively regulates Gad1 expression in E13.5 MGE, but

not adult hippocampus. (a) Quantitative real-time RT-PCR analysis of Gad1

in E13.5 MGE from wild types and Evf2TS/TS mutants. Error bars represent

s.e.m. P ¼ 0.031, Student’s t-test (n ¼ 3 for wild types, n ¼ 5 for Evf2TS/TS).

(b) Quantitative real-time RT-PCR analysis of Gad1 in E12.5 Evf2TS/TS

mutant MGE electroporated with control (pcDNA, 2 mg) and 2 mg pcDNA-

Evf2. Error bars represent s.e.m. P ¼ 0.0375, Student’s t-test (n ¼ 4).

(c) Quantitative real-time RT-PCR analysis of Gad1 in P60 hippocampus fromwild types and Evf2TS/TS mutants. Error bars represent s.e.m. P ¼ 0.031,

Student’s t test (n ¼ 2 for wt, n ¼ 2 for Evf2TS/TS). (d) Quantification of the

number of Gad1-expressing GABAergic interneurons in 8-month-old dentate

gyrus and hippocampal CA3 and CA1 regions. Error bars represent s.e.m.

P 4 0.05 for dentate gyrus, CA3 and CA1, Student’s t test (n ¼ 3

for wild types, n ¼ 3 for Evf2TS/TS). * P o 0.05, t test.

1024 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

DISCUSSION

Noncoding RNA–dependent balanced gene regulation

Our results indicate that the Evf2 noncoding RNA is required forbalanced gene regulation in developing ventral forebrain. In Evf2TS/TS

mutants, Dlx6 transcripts increased by about eightfold. The inability torescue increased Dlx6 in Evf2TS/TS mutants supports the idea that Evf2anti-sense transcriptional inhibition of Dlx6 occurs in cis throughopposite-strand transcription rather than anti-sense annealing. Datashowing that higher Evf2 levels caused an increase rather than adecrease in Dlx6 (Fig. 1j) support opposite-strand transcriptionalinhibition, as would be expected with an anti-sense annealing mechan-ism. Although we cannot rule out that the triple polyadenylation signalinsertion, rather than the loss of Evf2 anti-sense interference, resulted inan increase in Dlx6 transcription, it is unlikely that the triplepolyadenylation signal insertion disrupted Dlx5/6 ei or eii, given thatEvf1 transcription remained unaffected in Evf2TS/TS mutants. Further-more, the triple polyadenylation signal insertion in the 5¢ end of the Airnoncoding RNA35 does not affect adjacent transcription. Together,these data suggest that in vivo Evf2 transcription, rather than Evf2 RNA,negatively regulated Dlx6 transcription through competitive anti-senseinhibition in cis.

In contrast with its effects on Dlx6, the ability of Evf2 to rescueincreased Dlx5 and decreased Gad1 indicated that Evf2 RNA negativelyregulated Dlx5 and positively regulated Gad1 in trans. Therefore,Evf2-mediated trans-acting transcription-regulating effects were targetand concentration dependent; low levels of Evf2 repressed Dlx5 andhigher levels activated Dlx5, Dlx6 and Gad1. These results also supportprevious data that identified trans-acting transcriptional activation ofDlx5/6 ei and eii by Evf2 (ref. 4). Further support for the idea that Evf2acts in trans to repress Dlx5 stems from our knockdown studies(Supplementary Fig. 1).

Studies have shown that SHH activates Dlx and Evf genes and anembryonic form of Gad1 in embryonic forebrain4,19. In addition,ectopic expression of Dlx2 and Dlx5 activates Gad1 in embryonicforebrain slices36. Together, these data identify Evf2 and DLX ascomponents of a signaling cascade that activates Gad1, supportingthe idea that reduced Gad1 levels in Evf2TS/TS mutants may result frominterference with Evf2/DLX regulation of Gad1. However, both ourresults and previous results do not distinguish whether the loss of Evf2directly or indirectly affects Gad1 transcription. Of the two knowndirect targets of DLX1/2 signaling, Evf2 loss affected Dlx5/6 (ref. 5), butnot Nrp2 (ref. 25) (Fig. 1i), indicating that Evf2 did not affect all Dlx1/2activities. Therefore, it is possible that unidentified targets of Evf2 maybe responsible for Gad1 regulation. An important question raised bythese experiments is whether Evf2/DLX directly or indirectly regulatesGad1 expression.

Transcription factor recruitment by noncoding RNAs

Our results indicate that the Evf2 trucRNA is required for positive(DLX) and negative (MECP2) transcription factor recruitment toultraconserved DNA regulatory elements Dlx5/6 ei and Dlx5/6 eii(Fig. 2) in developing ventral forebrain. The loss of Evf2 preventedthe recruitment of a known transcriptional activator (DLX) to posi-tively acting DNA regulatory elements (Dlx5/6 ei and eii), with anunexpected increase, rather than decrease, in Dlx5 transcription. Wefound that Evf2-mediated recruitment of MECP2, a known transcrip-tional repressor, was lost from both ei and eii and propose that MECP2loss may explain Dlx5 deregulation. Despite differences betweenembryonic and adult MECP2-binding sites in the Dlx5/6 region,

a b

WT Evf2TS /TS

010203040

mIP

SC

sam

plitu

de (

pA)

0 50 100 150 2000

0.2

0.4

0.6

0.8

1.0

Cum

ulat

ive

prob

abili

ty

0 200 400 600 8000

0.2

0.4

0.6

0.8

1.0

Cum

ulat

ive

prob

abili

ty

048

1216

mIP

SC

sfr

eque

ncy

(Hz)

*

c

e

g

f

d

400 ms80 pA

WTEvf2TS/TS

1

2

3

-80 -60 -40 -20 20-0.5

0.5

1.5

2.5 WT

Evf2TS/TS

nA

250 ms

0.5

nA

0

5

10

15

20

*

0 200 400 600 8000

0.2

0.4

0.6

0.8

1.0

Cum

ulat

ive

prob

abili

ty

1

2

0

15

30

45

sIP

SC

sam

plitu

de (

pA)

0 100 200 300 4000

0.2

0.4

0.6

0.8

1.0

Cum

ulat

ive

prob

abili

ty

Amplitude (pA) Amplitude (pA)

–80 –60 –40 –20–0.5

sIP

SC

sfr

eque

ncy

(Hz)

Inter-event interval (ms) Inter-event interval (ms)

Figure 6 GABAergic synaptic inhibition is reduced in CA1 layer of the adult

hippocampus of Evf2TS/TS mutant mice. (a,b) Representative traces of sIPSCs

(1) and mIPSCs (2) from CA1 pyramidal cells of wild-type and Evf2TS/TS

mice. Picrotoxin 0.25 mM blocked all IPSC activity in Evf2TS/TS and wild-

type mice (3). (c,d) Averaged cumulative probability plots of sIPSC and

mIPSC inter-event intervals from CA1 pyramidal cells (Evf2TS/TS, red; wild

type, black; Kolmogorov-Smirnov test, P o 0.001). Insert, sIPSCs and

mIPSCs event frequency in adult and old Evf2TS/TS mice was less (sIPSCadult, 11.75 ± 1.73 Hz; sIPSC old, 9.17 ± 0.71; mIPSCs adult, 10.65 ±

1.05 Hz; mIPSCs old, 6.5 ± 0.83 Hz) than in control wild-type mice (sIPSC

adult, 16.66 ± 1.26 Hz; sIPSC old, 12.62 ± 0.92 Hz; mIPSCs adult, 14.16

± 1.67 Hz; mIPSCs old, 9.06 ± 0.77 Hz). (e,f) Averaged cumulative

probability plots of sIPSCs and mIPSCs amplitude (Evf2TS/TS, red; wild type,

black’ sIPSC adult, 33.6 ± 2.7 pA; sIPSC old, 48.50 ± 3.06; mIPSCs adult,

35.26 ± 1.41 pA; mIPSCs old, 34.87 ± 3.76) and wild-type mice (sIPSC

adult, 37.4 ± 3.2 pA; sIPSC old, 53.78 ± 3.98; mIPSCs adult, 34.54 ±

1.64 pA; mIPSCs old, 37.03 ± 4.63). (g) The I-V plot of evoked IPSCs of

GABAergic inhibition in Evf2TS/TS mice (P o 0.01, paired t test). IPSC

current was normalized to the amplitude of evoked EPSC in ACSF containing

no drugs to suppress excitatory synaptic activity. Trace 1 shows representative

IPSCs recordings at +20 mV in Evf2TS/TS (red) and wild-type (black) mice

(average n ¼ 5). Picrotoxin 0.25 mM blocked evoked IPSCs in CA1 neurons

from both wild-type and Evf2TS/TS mice (trace 2). Error bars represent s.e.m.

* P o 0.05, t test.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1025

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

a twofold increase in Dlx5 in adult brain occurred on loss of MECP2(ref. 23) and supports a repressive role for MECP2 on Dlx5/6 ei activity.At eii, MECP2 loss was accompanied by HDAC1 loss, suggesting thatchromatin in eii was more active in Evf2 mutants, which in turn couldincrease Dlx5 transcription. However, HDAC1 binding to ei did notchange, suggesting that an alternative regulation mechanism wasemployed at this site. Recent evidence suggests that MECP2 can actas a transcriptional activator in some cases, associating with CREB1 atactivated targets37. What determines whether MECP2 acts as a tran-scriptional activator or repressor? Our data indicate that MECP2recruitment is dependent on Evf2, supporting the possibility thattrans-acting RNAs may influence the choice between activator orrepressor activities, either through co-recruitment of additional factorsor by an unknown mechanism.

Another question regarding the mechanism of Evf2-mediatedrecruitment of MECP2 and DLX proteins is whether recruitmentoccurs competitively, equally on both alleles or in a mutually exclusivemanner (Supplementary Fig. 2). Allelic imbalance causes both subtle(1.5-fold) and marked (ninefold) changes in gene regulation andoccurs in 20–50% of tested genes38–40. Similarly, it is possible thatthe level of Evf2 localized at specific alleles may determine theamount or identity of the factor(s) recruited. Previous fluorescentRNA in situ hybridization data support the idea that E13.5 MGE nucleihave heterogeneous distribution of Evf2 between alleles; in anE13.5 MGE section, Evf2 transcripts may be distributed on single,equal double and unequal double alleles4. We found that DLX andMECP2 were also distributed nonhomogenously among different cellsand in MGE nuclei (Fig. 3c–j). These data raise the possibility thatunequal distribution of Evf2, DLX and MECP2 may have a role inregulating allelic imbalance in different populations of neuronalprecursors, leading to neuronal diversity or phenotypic variation.Further studies to test the possible role of RNA-dependentcontrol of allelic imbalance will be important to understand howcellular diversity arises.

In humans, 15–25% of genes participate in anti-sense transcrip-tion41; however, there are only a few reports of an in vivo mechanismthat employs specific anti-sense transcriptional regulation1. Amongthese, in vivo roles of anti-sense transcripts have been described forAir35 and Kcnq1ot1 (ref. 42) in imprinting control. The data from ourEvf2 mutant analysis, combined with that from our rescue experiments,suggests that Evf2, an anti-sense transcript of Dlx6, negatively regulatedDlx6 transcription in cis through opposite-strand transcriptionalcompetition. However, the ability of Evf2 to regulate Dlx5 and Gad1in trans raises the possibility that anti-sense transcripts may not belimited to cis-acting mechanisms or even to local regulation. Given thenumber of known anti-sense transcripts in nonimprinted regions, itwill be important to determine how often anti-sense transcripts havetrans-regulatory effects and the role of this regulation in variousbiological processes.

Noncoding RNA–dependent GABAergic interneuron development

The embryonic MGE is regulated by Dlx-dependent mechanisms andproduces GABAergic interneuron precursors that will later populatethe hippocampus29. We found that loss of Evf2 resulted in imbalancedgene expression in the embryonic MGE, leading to decreasedGABAergic interneurons in early postnatal (P2) hippocampus anddentate gyrus. Furthermore, this decrease in GABAergic interneuronsin P2 Evf2 mutants did not result from increased cell death or a cell-fatetransformation in the hippocampus (Fig. 4g–k).

Why then are GABAergic interneurons decreased in P2 Evf2 mutanthippocampus? One possibility is that decreased Gad1 in E13.5 Evf2

mutant MGE reduced GABA levels in interneuron precursors, alteringtheir tangential ventral to dorsal migration and reducing the number ofGABAergic interneurons that reach their destination in the hippocam-pus. This is supported by reports that GABA affects neuronal migrationin multiple contexts43, including tangential migration from MGE tocortex38,44,45. In addition, the interference with DLX1/2 activity thatwas seen in Evf2 mutants would be expected to impair GABAergicinterneuron migration from E13.5 MGE28,29. However, if the majorityof GABAergic interneurons derive from embryonic and early postnatalventral sources that are defective in Evf2 mutants, how then didGABAergic interneuron numbers recover in the adult Evf2 mutanthippocampus? One possibility is that the Evf2 mutant adult compen-sates by neurogenesis of GABAergic interneurons in either the hippo-campal subgranular zone or rostral subventricular zone. If so, it is clearfrom our electrophysiology experiments that these neurons were notfunctionally equivalent at the synaptic level to their embryonicallygenerated counterparts. An alternate possibility is that GABAergicprecursors in Evf2 mutants migrated at the appropriate time and tothe proper destination, but produced lower levels of Gad1 in theabsence of Evf2. If migration defects are found, future experimentswill be needed to investigate why the numbers of GABAergic inter-neurons decrease at P2, how they eventually recover and what is thebasis for their synaptic defects.

SHH signaling in the embryonic ventral forebrain initiates a tran-scriptional cascade that requires DLX proteins, Evf2, MECP2 and Gad1for proper GABAergic interneuron development (SupplementaryFig. 3). Although a large body of literature suggests that the postnataleffects of MECP2 are likely to be critical for Rett syndrome46,47, ourresults raise the possibility that Evf2 loss and MECP2 loss share acommon mechanism, in which transcriptional effects in embryonicMGE cause adult GABAergic defects at the synaptic level. Efforts toidentify single or multiple targets of MECP2 in the etiology of Rettsyndrome have been inconclusive. Controversial evidence has raisedthe possibility that deregulation of Dlx5 in adult Mecp2�/� brains maybe responsible for GABAergic defects23,48. However, analysis of Evf2mutants reveals that the correlation between MECP2 loss, increasedDlx5 and GABAergic defects may be indirect.

Given that DLX5 is a known activator of Gad1 (ref. 36), GABAergicinterneuron loss would not be a predicted effect of the twofold Dlx5increase in Evf2 mutants. In fact, Evf2 rescue experiments arguedagainst Dlx5 upregulation as a cause for GABAergic defects in Evf2mutants and suggested that the mechanism involved in Evf2 regulationof Dlx genes is separate from Evf2 regulation of Gad1. Our rescueexperiments suggested that Evf2 controls Gad1, Dlx5 and Dlx6 throughdistinct trans and cis mechanisms. In addition, at the 2 mg Evf2 rescueconcentration, Dlx5 increased while Gad1 was rescued. If a twofoldDlx5 increase were the cause of GABAergic defects, Gad1 would not beexpected to increase. These experiments suggest that Gad1 reduction,rather than increased Dlx5, is more likely to cause GABAergic inter-neuron defects.

To the best of our knowledge, these findings describe the firstdemonstration that an ncRNA-dependent mechanism critical forearly GABAergic interneuron development can determine GABA-dependent connectivity in the adult brain. The inability to recoverproper connectivity, regardless of restored Gad1 and GABAergicinterneuron number in adult brain, reinforces the idea thatcritical factors in the developing embryo influence GABAergic inter-neuron function in adult. This is especially important given thelong-standing question of whether mental disorders, in the absenceof apparent physiological adult deficits, can result from alteredembryonic development.

1026 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe thank K. Campbell for providing early passage W4 embryonic stem cells,and training and guidance in ES cell manipulation, M. Ekker for Dlx5/6 BAC,P. Lui and N. Copeland for PL253, PL452, bacterial strains and protocols forBAC recombineering, G. Taborn and P. Iannaccone for blastocyst injections,D. Eisentstat for antibody specific to DLX2, A. Joyner for the triplepolyadenylation construct, J. Rubenstein for Gad1 probe, and Q. Ma forv-Glut1 probe. We thank K. Jones (Northwestern University) for establishing theconditions for embryonic tissue ChIP. This work was funded by NationalInstitute of Child Health and Human Development grants RO1 HD044745 andR21 HD049875, the Illinois Regenerative Medicine Institute, and an IllinoisExcellence in Academic Medicine grant to J.D.K.

AUTHOR CONTRIBUTIONSA.M.B. performed the experiments in Figures 1–5. M.J.W.V. carried out BACrecombineering, ES cell homologous targeting and screening for generating Evf2TS mice. J.C.S. performed the experiments shown in Figures 1, 3 and 5.M.F.C. carried out the experiments shown in Figures 1, 5 and SupplementaryFigure 1. E.A.S. and J.F.D. contributed electrophysiology experiments and resultsfor Figure 6. J.D.K. conceived of and directed experiments and wrotethe manuscript.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Prasanth, K.V. & Spector, D.L. Eukaryotic regulatory RNAs: an answer to the ‘genomecomplexity’ conundrum. Genes Dev. 21, 11–42 (2007).

2. Mattick, J.S. A new paradigm for developmental biology. J. Exp. Biol. 210, 1526–1547(2007).

3. Shamovsky, I. & Nudler, E. Gene control by large noncoding RNAs. Sci. STKE 2006,pe40 (2006).

4. Feng, J. et al. The Evf-2 noncoding RNA is transcribed from the Dlx-5/6 ultraconservedregion and functions as a Dlx-2 transcriptional coactivator. Genes Dev. 20, 1470–1484(2006).

5. Zerucha, T. et al. A highly conserved enhancer in the Dlx5/Dlx6 intergenic region is thesite of cross-regulatory interactions between Dlx genes in the embryonic forebrain.J. Neurosci. 20, 709–721 (2000).

6. Gilligan, P., Brenner, S. & Venkatesh, B. Fugu and human sequence comparisonidentifies novel human genes and conserved non-coding sequences. Gene 294,35–44 (2002).

7. Santini, S., Boore, J.L. & Meyer, A. Evolutionary conservation of regulatory elements invertebrate Hox gene clusters. Genome Res. 13, 1111–1122 (2003).

8. Bejerano, G. et al. Ultraconserved elements in the human genome. Science 304,1321–1325 (2004).

9. Mercer, T.R., Dinger, M.E., Sunkin, S.M., Mehler, M.F. & Mattick, J.S. Specificexpression of long noncoding RNAs in the mouse brain. Proc. Natl. Acad. Sci. USA105, 716–721 (2008).

10. Di Cristo, G. Development of cortical GABAergic circuits and its implications forneurodevelopmental disorders. Clin. Genet. 72, 1–8 (2007).

11. Bienvenu, T. & Chelly, J. Molecular genetics of Rett syndrome: when DNA methylationgoes unrecognized. Nat. Rev. Genet. 7, 415–426 (2006).

12. Moretti, P. & Zoghbi, H.Y. MeCP2 dysfunction in Rett syndrome and related disorders.Curr. Opin. Genet. Dev. 16, 276–281 (2006).

13. Chahrour, M. & Zoghbi, H.Y. The story of Rett syndrome: from clinic to neurobiology.Neuron 56, 422–437 (2007).

14. Dani, V.S. et al. Reduced cortical activity due to a shift in the balance between excitationand inhibition in a mouse model of Rett syndrome. Proc. Natl. Acad. Sci. USA 102,12560–12565 (2005).

15. Lewis, D.A., Hashimoto, T. & Volk, D.W. Cortical inhibitory neurons and schizophrenia.Nat. Rev. Neurosci. 6, 312–324 (2005).

16. Soriano, P. Generalized lacZ expression with the ROSA26 Cre reporter strain. Nat.Genet. 21, 70–71 (1999).

17. Zhou, Q.P. et al. Identification of a direct Dlx homeodomain target in the developingmouse forebrain and retina by optimization of chromatin immunoprecipitation. NucleicAcids Res. 32, 884–892 (2004).

18. Panganiban, G., Sebring, A., Nagy, L. & Carroll, S. The development of crustacean limbsand the evolution of arthropods. Science 270, 1363–1366 (1995).

19. Kohtz, J.D., Baker, D.P., Corte, G. & Fishell, G. Regionalization within the mammaliantelencephalon is mediated by changes in responsiveness to Sonic Hedgehog. Develop-ment 125, 5079–5089 (1998).

20. Kohtz, J.D. et al. N-terminal fatty-acylation of sonic hedgehog enhances the induction ofrodent ventral forebrain neurons. Development 128, 2351–2363 (2001).

21. Feng, J. et al. Synergistic and antagonistic roles of the Sonic hedgehog N- andC-terminal lipids. Development 131, 4357–4370 (2004).

22. Anderson, S.A. et al. Mutations of the homeobox genes Dlx-1 and Dlx-2 disrupt thestriatal subventricular zone and differentiation of late born striatal neurons. Neuron 19,27–37 (1997).

23. Horike, S., Cai, S., Miyano, M., Cheng, J.F. & Kohwi-Shigematsu, T. Loss of silent-chromatin looping and impaired imprinting of DLX5 in Rett syndrome. Nat. Genet. 37,31–40 (2005).

24. Nan, X., Campoy, F.J. & Bird, A. MeCP2 is a transcriptional repressor with abundantbinding sites in genomic chromatin. Cell 88, 471–481 (1997).

25. Le, T.N. et al. Dlx homeobox genes promote cortical interneuron migration from the basalforebrain by direct repression of the semaphorin receptor neuropilin-2. J. Biol. Chem.282, 19071–19081 (2007).

26. Kishi, N. & Macklis, J.D. MECP2 is progressively expressed in post-migratory neuronsand is involved in neuronal maturation rather than cell fate decisions. Mol. Cell.Neurosci. 27, 306–321 (2004).

27. Panganiban, G. & Rubenstein, J.L. Developmental functions of the Distal-less/Dlxhomeobox genes. Development 129, 4371–4386 (2002).

28. Anderson, S.A., Eisenstat, D.D., Shi, L. & Rubenstein, J.L. Interneuron migration frombasal forebrain to neocortex: dependence on Dlx genes. Science 278, 474–476(1997).

29. Pleasure, S.J. et al. Cell migration from the ganglionic eminences is required for thedevelopment of hippocampal GABAergic interneurons. Neuron 28, 727–740(2000).

30. Marın, O. & Rubenstein, J.L. Cell migration in the forebrain. Annu. Rev. Neurosci. 26,441–483 (2003).

31. Wichterle, H., Turnbull, D.H., Nery, S., Fishell, G. & Alvarez-Buylla, A. In utero fatemapping reveals distinct migratory pathways and fates of neurons born in the mamma-lian basal forebrain. Development 128, 3759–3771 (2001).

32. Cobos, I. et al. Mice lacking Dlx1 show subtype-specific loss of interneurons, reducedinhibition and epilepsy. Nat. Neurosci. 8, 1059–1068 (2005).

33. Potier, B., Jouvenceau, A., Epelbaum, J. & Dutar, P. Age-related alterations of GABAer-gic input to CA1 pyramidal neurons and its control by nicotinic acetylcholine receptors inrat hippocampus. Neuroscience 142, 187–201 (2006).

34. Xu, C., Cui, C. & Alkon, D.L. Age-dependent enhancement of inhibitory synaptictransmission in CA1 pyramidal neurons via GluR5 kainate receptors. Hippocampus(2009).

35. Sleutels, F., Zwart, R. & Barlow, D.P. The non-coding Air RNA is required for silencingautosomal imprinted genes. Nature 415, 810–813 (2002).

36. Stuhmer, T., Anderson, S.A., Ekker, M. & Rubenstein, J.L. Ectopic expression of the Dlxgenes induces glutamic acid decarboxylase and Dlx expression. Development 129,245–252 (2002).

37. Chahrour, M. et al. MeCP2, a key contributor to neurological disease, activates andrepresses transcription. Science 320, 1224–1229 (2008).

38. Cowles, C.R., Hirschhorn, J.N., Altshuler, D. & Lander, E.S. Detection of regulatoryvariation in mouse genes. Nat. Genet. 32, 432–437 (2002).

39. Yan, H., Yuan, W., Velculescu, V.E., Vogelstein, B. & Kinzler, K.W. Allelic variation inhuman gene expression. Science 297, 1143 (2002).

40. Doss, S., Schadt, E.E., Drake, T.A. & Lusis, A.J. Cis-acting expression quantitative traitloci in mice. Genome Res. 15, 681–691 (2005).

41. Yelin, R. et al. Widespread occurrence of antisense transcription in the human genome.Nat. Biotechnol. 21, 379–386 (2003).

42. Mancini-DiNardo, D., Steele, S.J., Levorse, J.M., Ingram, R.S. & Tilghman, S.M.Elongation of the Kcnq1ot1 transcript is required for genomic imprinting of neighboringgenes. Genes Dev. 20, 1268–1282 (2006).

43. Heng, J.I., Moonen, G. & Nguyen, L. Neurotransmitters regulate cell migration in thetelencephalon. Eur. J. Neurosci. 26, 537–546 (2007).

44. Cuzon, V.C., Yeh, P.W., Cheng, Q. & Yeh, H.H. Ambient GABA promotes cortical entry oftangentially migrating cells derived from the medial ganglionic eminence. Cereb. Cortex16, 1377–1388 (2006).

45. Lopez-Bendito, G. et al. Blockade of GABA(B) receptors alters the tangential migrationof cortical neurons. Cereb. Cortex 13, 932–942 (2003).

46. Guy, J., Gan, J., Selfridge, J., Cobb, S. & Bird, A. Reversal of neurological defects in amouse model of Rett syndrome. Science 315, 1143–1147 (2007).

47. Giacometti, E., Luikenhuis, S., Beard, C. & Jaenisch, R. Partial rescue of MeCP2deficiency by postnatal activation of MeCP2. Proc. Natl. Acad. Sci. USA 104,1931–1936 (2007).

48. Schule, B., Li, H.H., Fisch-Kohl, C., Purmann, C. & Francke, U. DLX5 and DLX6expression is biallelic and not modulated by MeCP2 deficiency. Am. J. Hum. Genet. 81,492–506 (2007).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1027

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSPrimer sequences. For gene targeting, we used the following primers for

retrieval homology arms: Cla I (5¢-GAT GCG AAT CGA TCG GCT TAG GCC

TCC AGG TTT C-3¢), HindIII (5¢-AAA CCC TAA GCT TGA CTA GCG TGG

CCC AAA GGT-3¢), HindIII* (5¢-GAT GCG AAA GCT TCT GTC AGT GCC

AAA ATG GAA GGA CAT-3¢), NheI (5¢-GAT GCG AGC TAG CGG GGT TGG

GAC CTG GTT TTA GG-3¢). For targeting arms, we used Sac II (5¢-TTA GTT

CCG CGG CCT GGT CCT TTC TTC GTC TCA AGT C-3¢), NotI (5¢-ATT

TGC GGC CGC CTT AAG AGA TAT TCA CCG GGG TAA GTT TTT ATT-3¢),

ClaI (5¢-GAT TTT ATC GAT CAA TGA TCA GGG TCT AGA AAT CTA TAC

TGA G-3¢), Kpn (5¢-GAT TTT GGT ACC TTC AGG GTT TGA TTT GAT

CGC TAC TG-3¢), 5¢ ES Southern probes mEvf5¢.1 (5¢-TGG TGA AGC

TGG AGG AAG GAC-3¢) and mEvf5¢.2 (5¢-CAC ACT GAC TTC TGA

ACA CCC CTG-3¢), and 3¢ Southern probes mEvf3¢.1 (5¢-GGG GTG AAG

GAT GGT GAT TAA AGA GC-3¢) and mEvf3¢.1 (5¢-GTG GCT GGC TGT

CCT TTG GT-3¢).

For quantitative reverse transcription PCR, we used the following primers

with SYBR Green: Evf2-F (0.2 mM, 5¢-CTC CCT CCG CTC AGT ATA GAT

TTC-3¢), Evf2-R (0.2 mM, 5¢-CCT CCC CGG TGA ATA TCT CTT-3¢), Dlx2-F

(0.15 mM, 5¢-CCC TAC GGC ACC AGT TCG T-3¢), Dlx2-R (0.15 mM, 5¢-TCG

GAT TTC AGG CTC AAG GT-3¢), Nrp2-F (0.5 mM, 5¢-ACT TTT CAG GAC

ACG AAG TGA GAA-3¢), Nrp2-R (0.5 mM, 5¢-GCC AGC ATC TTT GGA ATT

CAG-3¢), Gad1-F (0.4 mM, 5¢-ACT CCT TCG CCT GCA ACC T-3¢), Gad1-R

(0.4 mM, 5¢-CGC CAC ACC AAG TAT CAT ACG T-3¢), b-actin–F (0.3 mM, 5¢-GCG AGC ACA GCT TCT TTG C-3¢) and b-actin–R (0.3 mM, 5¢-TCG TCA

TCC ATG GCG AAC T-3¢). For TaqMan PCR, we used the following primers:

Dlx5-probe (0.1 mM, 5¢-CAA GCA TCC GAT CCG GCG ACT TC-3¢), Dlx5-F

(0.1 mM, 5¢-TAT GAC AGG AGT GTT TGA CAG AAG AGT-3¢), Dlx5-R

(0.1 mM, 5¢-ACG TCG GGA ACG GAG CTT-3¢), Dlx6-probe (0.1 mM, 5¢-AAC

GCC TAC GGA GCT TCT GAA GGA GAC A-3¢), Dlx6-F (0.1 mM, 5¢-GAG

ACC ACA GAT GAT GTG ACT TCT CT-3¢), Dlx6-R (0.1 mM, 5¢-CTG CCA

TGT TTG TGC AGA TTC T-3¢), Evf1-probe (0.1 mM, 5¢-AGA GCT ATG CGA

CTG TAG GCA AGC CAT-3¢), Evf1-F (0.1 mM, 5¢-GCA TGG AAA CTT TGA

TAC CTT GGT-3¢), Evf1-R (0.1 mM, 5¢-GCC TTT CAG AAC TAG AAG GGA

TTT AAA-3¢), b-actin–probe (0.1 mM, 5¢-CAA CGA GCG GTT CCG ATG CCC

T-3¢), b-actin–F (0.1 mM, 5¢-ACG GCC AGG TCA TCA CTA TTG-3¢) and

b-actin–R (0.1 mM, 5¢-CAA GAA GGA AGG CTG GAA AAG A-3¢).

For quantitative PCR, we used the following primers with the SYBR Green

PCR Core Reagents Kit: 1-F (0.25 mM, 5¢-TAT GAA AAG CCC AGG ATT GC-

3¢), 1-R (0.25 mM, 5¢-TGT CCC AGC TTC CTA TCA CC-3¢), 2-F (0.25 mM, 5¢-TGG TTT GAA AGA GGG GAA TG-3¢), 2-R (0.25 mM, 5¢-AGA GCG CTT ATT

CTG AAA CCA-3¢), 3-F (0.12 mM, 5¢-CCC AGG ATC AAT TCT

GAA CAA AG-3¢), 3-R (0.50 mM, 5¢-TCC CCA ATG TCT GCT TCA AAT-

3¢), 4-F (0.10 mM, 5¢-TGG ATT CCC TGA ACT CCA AG-3¢). 4-R (0.10 mM, 5¢-AGG GCT TGG GAA CTC AAA CT-3¢), 5-F (0.24 mM, 5¢-GGC GCA TCT

TTG CAA ATT ACA-3¢), 5-R (0.50 mM, 5¢-GCA GGC TGG ATT AGG ATG

CTA-3¢), 6-F (1.0 mM, 5¢-TCG AAA GTA TTG CGT GGA TG-3¢), 6-R (1.0 mM,

5¢-GTG TGT ACC AAG CGC ATG TC-3¢), 7-F (0.25 mM, 5¢-GGC GTG TCA

GCA CCT GAT TT-3¢) and 7-R (0.25 mM, 5¢-GCC AAG TCA CTG CCC

ATC TC-3¢).

Generation of Evf2TS/TS mice. The Evf2 targeting construct was generated using

lambda phage–based recombineering in E. coli as described previously49. Using

high-fidelity Taq (Roche), homology arms of approximately 500 bp were PCR

amplified (with restriction sites added) from BAC DNA. Using a three-

fragment ligation, homology arms were cloned into Cla I and Nhe I sites of

PL253, with a HindIII site engineered between them. A 16.1-kb region

(corresponding to position 6,809,651–6,825,742 on mouse chromosome 6,

National Center for Biotechnology Information assembly) was retrieved from

pBAC e3.6 M8 (M. Ekker, University of Ottawa) into the retrieval plasmid

using recombination-induced EL250 cells49. Further targeting was performed

on the retrieved plasmid. The polyadenylation targeting vector was constructed

in PL452, a floxed-Neo–containing plasmid. The triple polyadenylation signal

was cloned into EcoRI and SalI sites of PL452. Approximately 500 bp of

targeting homology arms were cloned sequentially on either side of the polyA–

floxed-Neo insert. Briefly, fragments were PCR amplified as described above

and cloned into either Cla I and KpnI sites or Not I and SacII sites. This triple

polyA–floxed-Neo cassette was targeted into the retrieved 16.1-kb region using

recombination-induced EL250 cells. Successful targeting was confirmed by

Southern blot analysis of the completed construct using internal probes

(NEBlot kit, NEB).

Mouse ES cells were targeted by homologous recombination using standard

procedures. Successful targeting in ES cells was confirmed by Southern blot,

verifying proper recombination at both the 5¢ and 3¢ ends. Probes were

generated outside the 16.1-kb homologous region. The 5¢ probe was 499 bp

(chromosome 6 bases 6,808,430–6,808,928) and the 3¢ probe was 991 bp

(chromosome 6 bases 6,825,821–6,826,811). Wild-type ApaLI sites are at

chromosome 6 bases 6,828,765 and 6,817,913, yielding a 10.8-kb fragment

that hybridizes with the 3¢ probe. EL250 cells and recombineering plasmids

PL253 and PL452 were provided by N. Copeland (National Cancer Institute).

Evf2TS (floxed-Neo) heterozygotes were verified by Southern blot and

crossed to EIIAcre mice (Jackson Labs) for two generations. Neo removal

was verified by PCR (data not shown). Mice are kept on a mixed 129/FVB/

C57Bl6 background and housed according to guidelines approved by the

Institutional Animal Care and Use Committee of the Children’s Memorial

Research Center.

ChIP. MGE tissue was collected from E13.5 mouse Evf2+/+ or Evf2TS/TS

embryos (10–15 embryos per group). Tissue was broken into single cells by

pipetting and spun through a 70-mm filter. The DNA was crosslinked with 1%

paraformaldehyde (wt/vol) for 90 min on a rotator and then resuspended in

SDS lysis buffer (1% SDS (wt/vol), 50 mM Tris-HCl (pH 8.1) and 10 mM

EDTA) and the protease inhibitors PMSF (170 mg ml�1), pepstatin (0.7 mg

ml�1), leupeptin (10 mg ml�1) and aprotinin (10 mg ml�1). The crosslinked

DNA was sonicated on a Microson Ultrasonic Cell Disruptor at power 6 for six

pulses of 10 s each. The sonicated mixture was spun down and the supernatant

was used for ChIP.

For each immunoprecipitation condition, 20 mg of chromatin was used in a

total volume of 1,000 ml TES buffer (50 mM Tris-HCl (pH 8.1), 1 mM EDTA

and 150 mM NaCl) plus protease inhibitors. The chromatin was precleared on

a rotator at 4 1C. We added 75 ml of washed Protein G–Agarose beads for 1 h

and incubated the supernatant with 10 ml of rabbit pre-immune serum for 1 h.

Afterwards, 75 ml of Protein G–Agarose were added for 1 h and the supernatant

was split in half for an antibody condition and a rabbit pre-immune condition.

The chromatin was immunoprecipitated at 4 1C. Antibody (5 mg) or rabbit

prei-mmune serum (2 ml) were added for 4 h and then 100 ml of Protein

G–Agarose was added overnight. The Protein G-Agarose beads were washed

twice with low salt buffer (0.1% SDS, 1% Triton X-100 (vol/vol), 2 mM EDTA,

20 mM Tris-HCl (pH 9.1) and 150 mM NaCl), once with 500 mM NaCl, twice

with LiCl buffer (0.25M LiCl, 1% sodium deoxycholate (wt/vol), 1 mM EDTA,

10 mM Tris-HCl (pH 8.1) and 1% tert-octylphenoxy poly(oxyethylene)ethanol

(vol/vol, Sigma)), and twice with TE buffer (10 mM Tris-HCl (pH 8.1) and

1 mM EDTA). The chromatin was then eluted off the Protein G–Agarose beads

by incubation with 100 ml of elution buffer (1% SDS, and 0.1 M NaHCO3)

twice. The DNA crosslinking was reversed by incubation in 0.5 M NaCl for 5 h

at 65 1C. Then the DNA was proteinase K treated by adding half a volume of

15 mM EDTA/30 mM Tris-HCl (pH 8.1) and 750 mg ml�1 proteinase K for 1 h

at 65 1C. The uncrosslinked DNA was phenol extracted and ethanol precipi-

tated using glycogen. Rabbit pan-antibody to Dll was produced in the

laboratory. Pan-antibodies to Dlx, as originally described18, were made in

our laboratory and extensively characterized4. Pan-antibodies to Dlx were

raised in rabbits against the 62 amino acid butterfly Dll homeodomain–

containing fragment, affinity purified and tested by western blot (Dll fragment

and zebrafish Dlx1, 2, 4 and 6, and mouse Dlx2) and immunohistochemistry

(neural explants, embryonic forebrain sections). Pan-antibody to Dlx recog-

nizes zebrafish Dlx family members, including Dlx1, 2, 4 and 6 (ref. 4).

Monoclonal anti-Mecp2 antibodies were obtained commercially (Affinity

Bioreagents) and characterized previously26. We verified that these antibodies

to Mecp2 bound only two bands in adult olfactory bulb extracts by western

analysis (data not shown). Antibodies to HDAC1 were obtained from Sigma.

Primers were optimized to concentrations at which they were B100% efficient

with a standard curve slope of B�3.32.

The enrichment of the antibodies and rabbit pre-immune serum were

determined by comparing the Ct values (threshold cycle number) of the

NATURE NEUROSCIENCE doi:10.1038/nn.2371

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

immunoprecipitation condition to the Ct value of input chromatin using the

following formula: 2�ðCtðIPÞ�CtðInputÞÞ. Once the enrichments were calculated, the

relative enrichment of each antibody was calculated for Evf2+/+ and Evf2TS/TS

E13.5 MGE tissue by dividing antibody enrichment by pre-immune enrichment.

Immunohistochemistry. Brain tissue was fixed in 4% paraformaldehyde over-

night and then sunk consecutively in 15% and then 30% sucrose (wt/vol) in

phosphate-buffered saline (PBS) overnight. Tissue was cryostat sectioned at 18

mm and air dried before staining. In situ hybridization was carried out as

previously published4. Immunohistochemistry was performed using tyramide

amplification (TSA Kit #12, Invitrogen). Tissue was treated with 50% forma-

mide/50% 2� SSC (300 mM sodium chloride, 30 mM sodium citrate, pH 7.0)

for 2 h at 65 1C and then washed in 2� SSC. Cells were permeabilized with 0.5%

Triton X-100 in PBS and endogenous peroxidase activity was quenched by

incubating the tissue in 1% hydrogen peroxide (vol/vol) in PBS for 30 min.

Tissue was blocked in 1% blocking solution (vol/vol, Component D) and

labeled overnight with primary antibody, rabbit antibody to GABA (Sigma,

1:500) or pan-antibody to Dll (0.01 mg ml�1, purified in the laboratory) in

blocking solution. Tissue was washed in PBS and then incubated in secondary

antibody, goat antibody to rabbit IgG–horseradish peroxidase conjugate (Com-

ponent C, 1:100) for 1 h. Tissue was washed in PBS and a working tyramide

solution was applied for 15 min: Alexa Fluor 488 dye (Component A, 1:100) in

amplification buffer (Component E) with 0.0015% hydrogen peroxide (vol/vol,

Component F). Finally DAPI stain was applied for nuclear localization.

Western blot. Evf2+/+ and Evf2TS/TS E13.5 MGE tissue was homogenized in

SDS-PAGE sample buffer, separated on a 12.5% SDS-PAGE gel, transferred to

nitrocellulose and probed with primary rabbit antibody to Dlx2 (1:2,000) from

D. Eisenstat followed by secondary antibody to rabbit peroxidase (1:5,000,

Sigma). Blots were reprobed with mouse antibody to b-actin (1:20,000, AC-15,

Sigma). Bands were visualized using a chemiluminescence kit (Perkin Elmer).

Electroporation. DNA was dissolved to a concentration of 0.1 mg ml�1 in H2O

and 0.04% trypan blue (wt/vol) for visualization during injections. E12.5

Evf2TS/TS embryos were obtained from mating Evf2TS/TS males and females,

and dissected in cold L-15 media. The heads were placed in cold PBS and 1.5 ml

of DNA was injected into the lateral ventricles of the forebrain under a

dissection microscope. Electroporation using the square wave protocol (BioRad

Gene Pulser XCell) delivered 5 � 50-ms pulses of 30 V with 1-s intervals. Each

embryo was electroporated twice using electroporation paddles (Protech) that

were placed laterally on each side and then reversed to electroporate both left

and right sides of the brain. The MGE was dissected and cultured in neurobasal

medium (DMEM/F12 containing B-27 (Gibco), N2 supplement (Gibco),

200 mM L-glutamine, 0.1 mg ml�1 penicillin/streptomycin and 1 mg/ml

mitomycin C on 0.02-mm filters (Nunc)). After 24 h in culture, RNA was

isolated from pools of four MGE explants for qRT-PCR analysis.

RNA isolation and reverse transcription. Total RNA from pairs of E13.5 MGE

tissue was isolated50 and treated with RNASE-free DNASE I (NEB) and

reverse-transcribed using random hexamers (NEB) and murine Moloney

leukemia virus reverse-transcriptase (Invitrogen).

Quantitative PCR. The SYBR Green PCR Core Reagents Kit (Applied Biosys-

tems, 4304886) was used for all quantitative ChIP PCR on the 7500 Fast Real-

Time PCR System (Applied Biosystems). The following PCR program was

used: AMPErase UNG treatment at 50 1C for 2 min, AmpliTaq Gold activation

at 95 1C for 10 min, 40 cycles of denaturation at 95 1C for 10 s, anneal at 58 1C

for 5 s, and elongate at 72 1C for 32 s. Each 25-ml reaction included 5 ml of a

1:50 immunoprecipitation dilution, 2.5 ml of 10� PCR buffer, 3 ml of 25 mM

MgCl2, 2.0 ml of dNTP blend, 0.25 ml of AmpErase UNG and 0.25 ml of

AmpliTaq Gold DNA polymerase. Primers were optimized to concentrations at

which they were B100% efficient or had a standard curve slope of B�3.32.

The TaqMan PCR Core Reagents Kit (Applied Biosystems, N8080228)

was used for the rest of the quantitative RT-PCR primers on the 7500 Fast

Real-Time PCR System (Applied Biosystems). The following PCR program was

used: AMPErase UNG treatment at 50 1C for 2 min, AmpliTaq Gold activation

at 95 1C for 15 min, 40 cycles of denaturation at 95 1C for 30 s and anneal/

elongate at 59 1C for 1 min. Each 25-ml reaction included 50 ng of cDNA, 2.5 ml

of 10� PCR Buffer, 5.0 ml of 25 mM MgCl2, 0.75 ml of dATP, 0.75 ml of dGTP,

0.75 ml of dCTP, 0.75 ml of dUTP, 0.25 ml of AmpErase UNG and 0.25 ml of

AmpliTaq Gold DNA polymerase.

The fold gene expression of Evf2+/+ and Evf2TS/TS E13.5 MGE tissue was

determined by comparing the Ct values of the target gene to the Ct value of the

control gene (b-actin) using the following formula: 2�ðCtðIPÞ�CtðInputÞÞ.

Electrophysiology. Transverse hippocampal slices (300 mm) were prepared

from 3–5-month-old Evf2TS/TS mice. Slices were cut in ice-cold (B4 1C)

oxygenated ACSF and then were allowed to recover for 1–2 h before being

transferred to the recording chamber where they were continuously superfused

with solution heated to 32–34 1C and saturated with 95% O2/5% CO2. The

standard extracellular solution contained 124 mM NaCl, 3 mM KCl, 1.25 mM

KH2PO4, 2.0 mM CaCl2, 1.3 mM MgCl2, 26 mM NaHCO3 and 10 mM

glucose. The concentration of MgCl2 was raised to 2.0 mM and that of CaCl2decreased to 1.4 mM in the slicing solution. To isolate GABAergic current,

20 mM CNQX and 50 mM D-AP5 were added to the standard ACSF.

Tetrodotoxin (2 mM) was added to the drug solution to isolate minimal IPSCs.

Somatic whole-cell recordings were obtained from CA1 pyramidal neurons

under visual guidance using infrared differential interference contrast micro-

scopy. Patch pipettes were pulled from borosilicate glass capillary tubing (A-M

Systems). Internal solution for the patch pipettes contained 120 mM cesium

methanesulphonate, 10 mM CsCl, 5 mM NaCl, 10 mM HEPES, 10 mM EGTA,

5 mM TEA-Cl, 4 mM Mg-ATP, 0.3 mM GTP and 5 mM QX-314 (pH 7.3–7.4),

adjusted with CsOH, osmolarity 290 ± 10 mOsm.

Voltage-clamp recordings were performed using Axopatch-200B (Axon

Instruments). In all of our recordings, series resistance was o22 MO and

was compensated at 80%. The recorded signal was low-pass filtered at 5 kHz

and digitized at 10 kHz with a PCI-MIO-16E-4 board (National Instruments).

All data were stored on a PC computer with custom software using C++

Builder 5.0 (Borland) and a NI DAQ 6.5 driver (National Instruments).

Evoked IPSCs were elicited at 0.1–0.07 Hz with a bipolar Pt/Ir electrode (2 �25 mm) positioned on the stratum radiatum within 150 mm of the cell being

recorded (Schaffer collateral–commisural fibers location site). At the beginning

of each experiment, stimulus intensity was adjusted to evoke EPSCs with

amplitude of 0.4–0.2 nA. Evoked IPSCs were recorded at holding potential

from �70 mV to +20 mV with 10-mV voltage steps after blocking EPSCs with

a solution containing CNQX and D-AP5.

Spontaneous IPSCs were analyzed using MiniAnalysis 6.0.3 software (Synap-

tosoft). All events were manually selected on the basis of their kinetics. Between

250 and 450 individual events were analyzed for each cell. Post-recording

analyses of evoked IPSPs were carried out using Clampfit 10 (Axon Instru-

ments). All statistical values were evaluated with Origin 8 (MicroCal Software).

Values are presented as mean ± s.e.m. Statistical differences were established at

P o 0.05 using the Student t test, unless otherwise indicated.

Animal Use. Mice were handled under according to the guidelines approved by

the Institutional Animal Care and Use Committee of the Children’s Memorial

Research Center.

Statistical Analysis. Statistical significance was determined using a variety of

different tests, as indicated in individual figure legends, with *P o 0.05. Error

bars on graphs correspond to s.e.m.

49. Liu, P., Jenkins, N.A. & Copeland, N.G. A highly efficient recombineering-basedmethod for generating conditional knockout mutations. Genome Res. 13, 476–484(2003).

50. Chomczynski, P. & Sacchi, N. Single-step method of RNA isolation by acid guani-dinium thiocyanate-phenol-chloroform extraction. Anal. Biochem. 162, 156–159(1987).

doi:10.1038/nn.2371 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Genetic identification of an embryonic parafacialoscillator coupling to the preBotzinger complex

Muriel Thoby-Brisson1, Mattias Karlen2, Ning Wu1, Patrick Charnay3, Jean Champagnat1 & Gilles Fortin1

The hindbrain transcription factors Phox2b and Egr2 (also known as Krox20) are linked to the development of the autonomic

nervous system and rhombomere-related regulation of breathing, respectively. Mutations in these proteins can lead to abnormal

breathing behavior as a result of an alteration in an unidentified neuronal system. We characterized a bilateral embryonic

parafacial (e-pF) population of rhythmically bursting neurons at embryonic day (E) 14.5 in mice. These cells expressed Phox2b,

were derived from Egr2-expressing precursors and their development was dependent on the integrity of the Egr2 gene. Silencing

or eliminating the e-pF oscillator, but not the putative inspiratory oscillator (preBotzinger complex, preBotC), led to an abnormally

slow rhythm, demonstrating that the e-pF controls the respiratory rhythm. The e-pF oscillator, the only one active at E14.5,

entrained and then coupled with the preBotC, which emerged independently at E15.5. These data establish the dual organization

of the respiratory rhythm generator at the time of its inception, when it begins to drive fetal breathing.

In mammals, breathing is one of the earliest motor behaviors of thefetus and it relies on the activity of a brainstem respiratory rhythmgenerator (RRG). Recent advances in the neurobiology of breathing inneonatal mammals suggest that the RRG is located in two prominentrhythmogenic sites, the preBotC1 and the para-facial respiratory group(pFRG)2. The genetic regulation of progenitor cell fate and differentia-tion in the hindbrain, and of the plasticity of the RRG, remainspoorly understood.

The developmental origin and functional nature of the respiratoryrhythm-generating circuits involved in fetal and neonatal breathinghave been studied using mutant mice in which developmental genesencoding transcription factors have been inactivated, including Egr2,Phox2b, Hoxa1, Tlx3 and Mafb3–7. In particular, inactivation of Egr2,the gene for the zinc finger transcription factor Egr2, which controls theformation of hindbrain rhombomeric segments 3 and 5 (refs. 8,9),results in defective breathing. This is attributed to a reorganization ofneuronal circuits in the caudal pontine reticular formation thatconsequently leads to poor survival at birth6. Further evidence fromchicks shows that Egr2 is involved in the specification of a centralrhythm generator at the level of the facial motor nucleus10. Althoughthese studies have shown that the importance of hindbrain segmenta-tion extends beyond modular anatomical organization to thelevel of network assembly and function, they have failed to identify acandidate Egr2-derived cell group that would explain the mutantrespiratory deficit.

Phox2b is a transcription factor that is specifically expressed andrequired in neurons that form the visceral reflex circuits controllingdigestive, cardiovascular and respiratory functions11,12. In humans, aheterozygous mutation in PHOX2B is the main cause of congenital

central hypoventilation syndrome (CCHS)13,14, a genetic disease thattypically manifests itself at birth by respiratory distress during sleep15.Newborn mice that are heterozygous for the most common humanmutation have a slowed and often irregular breathing pattern, do notrespond to hypercapnia, and die at birth from respiratory failure16.Anatomically, these mutants selectively lack a group of glutamatergic,Phox2b-expressing interneurons in the ventro-lateral medulla in thevicinity of the facial motor nucleus16 called the retrotrapezoid nucleus(RTN)17. The RTN has previously been identified as a CO2/pH sensorsystem in the adult rat18,19. Notably, the rhythmogenic pFRG over-laps anatomically with the RTN and was recently shown to hostCO2-sensitive Phox2b-positive interneurones in rat neonates20, thussupporting the view that the pFRG and RTN may correspond toneonatal and adult forms of the same neuronal population19.

We investigated embryonic stages using knock-in alleles of Egr2 andidentified a population of Phox2b-positive interneurons deriving fromEgr2-expressing cells that forms an e-pF oscillator. The e-pF oscillator,which is first active at E14.5, couples with the developing preBotCoscillator within 24 h, thus establishing the dual organization of theRRG at the time at which it begins to pace fetal breathing.

RESULTS

The embryonic parafacial oscillator

We used mice carrying a knock-in of the Cre recombinase gene into theEgr2 locus (Egr2cre)21 and a Cre-responsive indicator allele (R26R-EYFP)22 to trace the derivatives of Egr2-expressing cells. At E14.5, cellsderived from Egr2-positive progenitors formed two transverse stripeswith little cell dispersal along the anterior-posterior axis; these stripescorrespond to the location of and are likely to be derived from the

Received 14 April; accepted 29 May; published online 5 July 2009; doi:10.1038/nn.2354

1Institut de Neurobiologie Alfred Fessard, Centre National de la Recherche Scientifique UPR2216, Gif sur Yvette, France. 2Department of Cell and Molecular Biology,Karolinska Institute, Stockholm, Sweden. 3Institut National de la Sante et de la Recherche Medicale, U784, Ecole Normale Superieure, Paris, France. Correspondenceshould be addressed to G.F. (gilles.fortin@inaf.cnrs-gif.fr).

1028 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

rhombomeric 3 and 5 segmental domains that have been observedduring hindbrain development (see ref. 9, the r5 stripe is shown inFig. 1a). Notably, inspecting the ventral surface of the hindbrain, thisrestriction along the anterior-posterior axis was clearly disruptedlaterally, resulting in a parafacial stripe of yellow fluorescent protein(YFP)-expressing cells caudal to r5, flanking and partially capping thelateral aspect of facial motor nucleus (nVII) and continuing approxi-mately 200 mm caudal to the nVII (Fig. 1a). Using optical recordingsfrom brainstem en bloc preparations23 after Calcium Green 1AMloading, we identified a rhythmic cell population that matched thisparafacial YFP+ territory (frequency, 14.3 ± 3.9 bursts min�1; range,8–18 bursts min�1; n ¼ 19 preparations; Fig. 1b–f). Occasionally, aburst of activity in the neighboring nVII (Fig. 1d) was associated with aburst of activity in the parafacial domain (see below). Low-resolutionimaging revealed bilateral coactivation of parafacial domains and theabsence of activity in other regions of the hindbrain. In transverse facialslices (n ¼ 8), a comparable rhythm was maintained in cells of theparafacial region (Supplementary Fig. 1). These data demonstrate thatthe parafacial domain forms an intrinsically active rhythmic network,which we refer to as the e-pF oscillator.

Imaging at cellular resolution showed synchronized activity ofindividual e-pF cells (Fig. 1e,f) and synchronized rhythmic changesin fluorescence were detected in 95% of the YFP+ cells (98 of 103 cellsfrom three en bloc preparations) that were sampled along the rostro-caudal extent (B700 mm) of the e-pF (Fig. 1g,h). Because theproportion of YFP� calcium-loaded cells in the e-pF area was lessthan 10%, this suggests that almost all of the e-pF cells expressed YFPand were therefore presumably derived from rhombomeric segments 3

or 5. Active cells were not detected before E14.5 (data not shown; E12.5,n ¼ 3; E13.5, n ¼ 5); thus, E14.5 marks the onset of the e-pF.

On the basis of previous work showing that respiratory-relatedPhox2b-positive and neurokinin-1 receptor (NK1R)–positive neuronsare located in an area similar to the e-pF at E15.5 (ref. 16), we examinedthose markers in the e-pF. We selected rhythmically active e-pF neuronsby loose patch recording, filled them with biocytin and determinedwhether the Phox2b and NK1R proteins were present by immuno-labeling. We found that 15 of 16 e-pF cells expressed Phox2b (Fig. 1i)and 7 of 10 expressed NK1R (Fig. 1j); thus, a large fraction of e-pF cellsexpressed both markers at E14.5. The NK1Rs were functional becausethe e-pF rhythm frequency increased twofold in the presence of substanceP, the endogenous ligand for NK1R (0.1 mM, data not shown). InEgr2cre/+; R26R-EYFP embryos, most, if not all, of the YFP+ cells in theparafacial area expressed Phox2b and could be easily distinguishedfrom the Phox2b+, YFP� and Islet1/2+ nVII cells (Fig. 1k). All Phox2b+

cells in the parafacial region express the type 2 vesicular glutamatetransporter (VGlut2) at birth16, indicating the probable glutamatergicnature of e-pF cells. The number of e-pF YFP+, Phox2b+, Islet1/2�

neurons on one side (280 ± 36, n ¼ 3) was similar to the number ofrhythmic cells detected in calcium-loaded preparations (259 ± 54 cells,n ¼ 5), indicating that the e-pF oscillator is composed of Phox2b+

neurons, which are probably glutamatergic, derived from Egr2-expres-sing precursors.

Operating principles of the e-pF oscillator

We then investigated the cellular properties underlying rhythm gen-eration in the e-pF oscillator. In whole-cell recordings performed onE14.5 whole hindbrain preparations, rhythmic burst discharges ofaction potentials in e-pF cells (43 of 43; Fig. 2a) appeared as all-or-none voltage-dependent events. Spontaneous bursts were curtailedby short negative-current pulses (Fig. 2b) that were applied during theburst, whereas burst discharges were evoked in between spontaneousbursts by comparable current pulses of opposite polarity (Fig. 2c).Slowly ramping the somatic potential of e-pF cells from �80 to+40 mV revealed a tetrodotoxin-sensitive (n ¼ 3, data not shown)and riluzole-sensitive (n¼ 10) persistent sodium current (INaP; Fig. 2d).In the presence of riluzole (20 mM), rhythmic activity of the e-pF cell

∆F/F

∆F/F10%

∆F/F4%

5 s

5 s

∆F/F 5010

1

0

12

10

1

e-pF

A

r5

YFP IsI1/2

YFP IsI1/2 Phox2b

12

345

679

811 10

12

YFP Ca Green1 Merge

Phox2b Biocytin Merge

NK1R Biocytin Merge

1

10

Direct F

M

D

M

a b c

e

g

i

j

k

h

f

d

Figure 1 A parafacial oscillator emerges at E14.5 and is derived from Egr2-

positive territories in the mouse embryo hindbrain. (a) Partial ventral view of

a hindbrain from an Egr2cre/+; R26R-EYFP embryo showing the respective

positions of cells derived from rhombomeric segments 3 and 5 (YFP positive,

green) and nVII motoneuronal populations (Islet1/2 positive, red). Note the

stripe of YFP-expressing cells caudal to rhombomeric segment 5, flanking the

lateral aspect of the nVII (blue outline). (b–d) A Calcium Green 1AM–loaded

whole hindbrain preparation (b) showing fluorescence changes, which weregenerally restricted to the e-pF oscillator (red outline) in c, and were

sometimes concomitant to activity of the nVII (d). (e,f) Photomicrograph of

e-pF cells (numbered 1 to 12) during a burst of activity (e) and corresponding

individual (black) and average (red) relative fluorescence changes traces (f).

(g) Same field of an Egr2cre/+; R26R-EYFP preparation showing YFP-

expressing cells (red), Calcium Green 1AM–loaded cells (green) and the

merged image used to derive the individual rhythmic activities of ten double-

labeled (yellow) cells. (h) Fluorescence changes of individual cells (black)

and averaged signal (red) from g. (i,j) Immunolabeling for Phox2b (i) and for

NK1R protein (j) in two biocytin-filled e-pF neurons. (k) Single transverse

section from an E14.5 Egr2cre/+; R26R-EYFP embryo that were triple

immunolabeled with antibodies specific to Islet1/2 (blue), Phox2b (red) and

YFP (green). Cells of the e-pF expressing both Phox2b and YFP are shown in

yellow and nVII motoneurons expressing both Phox2b and Islet1/2 appear in

purple. Scale bars represent 20 mm (e,g,i,j) and 200 mm (a–d,k). A, anterior;

D, dorsal; M, median.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1029

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

population was abolished and the all-or-none bursts of action poten-tials that were evoked by depolarizing current pulses or that occurredspontaneously were replaced by single spikes (Fig. 2e). Hyperpolariza-tion-activated and cyclic nucleotide–gated cation channel (HCN)-mediated currents (Ih), which are often expressed together with INaP

in pacemaking neurons24, were present in 25 of 26 e-pF neurons(Fig. 2f). Although the hyperpolarization level required to elicit asomatic Ih current was typically 20 mV below the resting membranepotential of e-pF cells (Vm ¼ �47 ± 1.5 mV, n ¼ 20), application ofZD7288 (100 mM) systematically blocked the Ih current present in e-pFcells and reduced the frequency of the e-pF oscillator from 13.7 ± 1.3bursts min�1 to 5.4 ± 0.7 bursts min�1 (n ¼ 11; Fig. 2g), suggesting arole for possibly distally located HCN channels in the modulationof the rhythm.

Rhythm generation and intercellular synchrony of the e-pF oscillatorwere preserved in the presence of 10 mM CNQX, which blocks gluta-matergic transmission mediated by AMPA/kainate receptors (n ¼ 5;Fig. 2h), as well as in the presence of either the m-opioid agonistD-Ala2-N-Me-Phe4-glycol5-enkephalin (DAMGO, 0.3 mM, n ¼ 5) or acocktail of bicuculline (10 mM) and strychnine (5 mM), which blockGABAA and glycinergic receptors, respectively (n ¼ 5, data not shown).Furthermore, the e-pF oscillator rhythm was preserved (althoughslower) in mice that do not express VGlut2 (Vglut2f/f;PCre; exons 4–6of Vglut2 are flanked with loxP sites and cre is driven by the Pgk

promoter)25 (n ¼ 3; Fig. 2i). Moreover, thee-pF oscillator activity was spared in the pre-sence of the NMDA-R antagonist D-2-amino-

5-phosphonovaleric acid (AP5, 5 mM, n ¼ 6; Fig. 2j). These datasuggest that glutamatergic synaptic transmission is not essential forrhythm generation or for intercellular synchrony of the e-pF oscillator.Application of lanthanum (La3+, 100 mM, n ¼ 4), which can efficientlyblock hemichannels, but not gap junctions permeabilities26,27, failed toalter the e-pF collective synchronous activity (Fig. 2k). However,carbenoxolone (CBX, 50 mM, n ¼ 11) blocking gap junctions, inaddition to hemichannels, abolished intercellular synchrony (Fig. 2l).Under CBX, e-pF cells that maintained rhythmic fluorescence changes(96 of 229 e-pF cells from ten preparations) were readily silenced byfurther application of riluzole (n ¼ 5; Fig. 2l). We observed similareffects of CBX (n ¼ 2), riluzole (n ¼ 3), CBX and riluzole(n ¼ 2), and ZD7288 (n ¼ 3) on the activity of e-pF cells in transversefacial slices (five slices, data not shown), further establishing the e-pFas the source of rhythmic activities.

Low-resolution imaging on E14.5 whole hindbrain preparationsindicated that application of CNQX (n ¼ 3) spared rhythmic activitiesand intercellular synchrony in parafacial domains on either side ofthe midline, but caused their left/right de-synchronization. This de-synchronization was not observed in response to the bicuculline/strychnine cocktail (n ¼ 3). Independent rhythms in left and rightparafacial domains were also observed in transverse slices from the axiallevel of the nVII (n¼ 4; Supplementary Fig. 2). Thus, on the one hand,e-pF cells collectively function as an oscillator in a manner that is

Neuron

e-pF int

lh current lh current l/V curve

Vm (mV)

e-pF int

e-pF

12

1

1

1

1

1

e-pF

12

e-pFe-pF

e-pF

e-pF

11

9

10

1

1

10

9

–60 +20–40 –20 0 –60 +20–40 –20

–120 –100

CTL (n = 14)ZD (n = 5)

–80 –60

10

–10

0

–20

–30

–40

0

Ril 50 pA

Voltage (mV)

0

20

40

CTL–Ril Current (pA

)

lh amplitude (pA

)

Control

–60 mV2 s

iV

–50 mV

i

i

20 mV

1 s

1 s

–100 V–50 mV

50 pA

–0.1 nA

–60 mV

∆F/F10%

∆F/F10%

∆F/F2%

∆F/F4%

∆F/F4%

∆F/F4%

∆F/F4%

–45 mV

Riluzole

CTL

V

V

CTLZD

13 mV s

–1

Neuron

i

Neuron

Control

ZD7288

WT in CNQXWT

Control

Control

AP5

VGlut2f/f;PCre

CBX + riluzoleCBX

La3+

Control Control

e-pF int

20 mV

20 mV

–50 mV0.4 nA

20 mV

–50 mV

5 s

20 mV0.1 nA

20 mV0.1 nA

1 s

5 s

5 s

5 s

5 s

5 s

5 s

2 sVi

a

d

f

h i

j k

l

g

e

b c Figure 2 Operating principles of the e-pF

oscillator at E14.5. (a) Spontaneous rhythmic

depolarization and firing (top trace) of an e-pF cell

synchronous with e-pF population integrated

activity (bottom trace). (b,c) Spontaneous bursts

(top, gray control trace) in phase with the

population activity (bottom traces, e-pF int) could

be prematurely terminated (b, black traces) orfully evoked (c, black traces) by negative and

positive current pulse injections (i, middle traces),

respectively. (d,e) Bursting activity in e-pF cells

relied on the INaP current. Riluzole (Ril) blocked

the control (CTL) inward current activated in e-pF

cells by slow depolarizing voltage ramps (d, left).

The INaP current-voltage relationship was obtained

by subtracting the riluzole-resistant current from

the control current (d, right, CTL – Ril). Riluzole

also blocked spontaneous and evoked burst of

activity in the e-pF (e). (f,g) The hyperpolarization-

activated current Ih present in e-pF cells was

blocked by ZD7288 (ZD, f) and modulated

the e-pF rhythm (g). Error bars represent s.e.m.

(h–j) Glutamatergic transmission was not required

for rhythmic activity of the e-pF oscillator. The

e-pF rhythm was maintained in 10 mM CNQX (h),

in Vglut2f/f;PCre mutant mice25 preparations (i) and

in the presence of 5 mM AP5 (j). (k,l) Intercellularcoactivation in the e-pF oscillator relied on gap

junctions. Intercellular coactivation was maintained

after blockade of hemichannel permeabilities by

100 mM lanthanum (La3+, control not shown, k),

but was lost after CBX (50 mM) treatment and

active cells were silenced on further application

of riluzole (l). Black traces in g–l represent

fluorescence changes in individual e-pF cells

captured in the frame and gray traces represent

the average fluorescence changes over the regions

encompassing these cells.

1030 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

largely independent of glutamatergic transmission; on the other hand,e-pF bilateral coactivation appears to rely on glutamatergic commis-sural projections that are already established at E14.5.

The e-pF oscillator couples with the preBotC oscillator

In mice, the respiratory preBotC oscillator, located at the vagal level ofthe hindbrain, is first active at E15.5 and can maintain the rhythmicactivity of hypoglossal motoneurons in transverse slices28. Moreover, atE15.5, fluorescence changes in the e-pF area and the nVII occurredtogether with bouts of hypoglossal nerve (XIIn) activity in a conti-nuous rhythmic and synchronized manner, indicating that there is acommon source of rhythmic activity, which possibly involves the e-pFoscillator (Fig. 3a–c). To investigate a possible interaction between thee-pF and the preBotC, we carried out simultaneous optical recordingsof the e-pF oscillator and nVII and electrophysiological recordings ofXIIn on E15.5 preparations.

We first investigated the consequences on nVII and XIIn activities ofa transection that was just posterior to the facial motor nucleus, inbetween the e-pF oscillator and the preBotC oscillators (n ¼ 4).Notably, rhythmic activities were maintained on both sides of thesection. Caudal to the section, the XIIn had a lower frequency thanintact preparations (Fig. 3a,d), reflecting the presence of an activepreBotC oscillator. Indeed, this activity was abolished by bath applica-tion of DAMGO29 (n ¼ 3, data not shown), but was unaffected byriluzole (n ¼ 3), which also failed to modify the rhythm observed inE15.5 preBotC transverse slices (n ¼ 6; Fig. 3d). Rostral to the section,the e-pF oscillated faster than in intact preparations, whereas theadjacent nVII was completely silenced, indicating that the e-pF prob-ably does not entrain it directly (Fig. 3a). The increased frequency ofthe e-pF was consistent with that observed in facial transverse slices,suggesting that partial e-pF cellular loss and/or severed commissuralfibers are causal factors. At any rate, caudal to the section, the resultingslower rhythm suggested the elimination of a descending excitatoryinput, indicating that the two oscillators interact at E15.5.

Second, in intact E15.5 preparations, bath applications of riluzolethat selectively suppress the e-pF activity (n ¼ 8) led to maintainedrhythmic activities in the nVII and XIIn, although at about half of the

frequency observed in control conditions (Fig. 3b,d). These dataindicate that functional impairment of the e-pF results in a slowerrhythm that is driven by the sole preBotC oscillator. Finally, silencingthe preBotC oscillator by application of CNQX (n ¼ 5, Fig. 3c,d) orDAMGO (n ¼ 7, Fig. 3d) led to a selective cessation of rhythmicactivities of the nVII and XIIn, demonstrating the pre-motor status ofthe preBotC oscillator. Our data support the view that the e-pFoscillator increases the frequency of rhythmic motor bursts generatedby the preBotC. Altogether, these data suggest that the dual organiza-tion of the RRG described at later stages2,30 is already achieved by thetime of its inception at E15.5.

We then investigated, during the E14.5–15.5 period, the establish-ment of the continuous rhythmic motor output. We confirmed thatspontaneous collective cellular rhythmic behavior was absent at E14.5in transverse preBotC slices (n ¼ 5), denoting its immature status28. Insome of the E15.5 and in all of the E14.5 whole hindbrain preparations,we observed discrete failings of the nVII and XIIn bursts, despite thecontinuous generation of e-pF rhythmic bursts (Fig. 4a). This wasreminiscent of skipped respiratory cycles, as opioid-induced reductionof excitability in one (preBotC) of the two coupled oscillators causestransmission failures of the rhythmic drive from the other coupledoscillator in the neonatal RRG31. Therefore, we investigated, duringE14.5–15.5, the possibility that excitability build-up in the preBotC inthe presence of an already active e-pF rhythmic drive may converselylead to discrete occurrences of motor bursts in phase with e-pF bursts.To explore this, we first compared the frequency distribution of motorbursts with that of e-pF bursts between E14.5 and E15.5 prepara-tions. At E14.5, the frequencies of motor bursts (3.4 ± 1.5 bursts min�1,n ¼ 11) and e-pF bursts (10.6 ± 1.7 bursts min�1, n ¼ 11) weresignificantly different (P o 0.001), whereas the distributions werefound to partially overlap at E15.5 (Fig. 4b). This resulted from anincreased occurrence of motor bursts (8.6 ± 2.1 bursts min�1, n ¼ 9),whereas the frequency of the e-pF oscillator was unchanged (10.2 ±2.4 bursts min�1, n ¼ 9). At E14.5, each individual burst of motoractivity was generated with a delay after the onset of an e-pF burst,which was reduced or non-existent by E15.5 (Fig. 4c,d). These dataindicate that the continuous generation of rhythmic motor bursts

Figure 3 Coupled oscillators control the motor

activity at E15.5. Simultaneous optical

recordings of the e-pF oscillator and nVII

activities and electrophysiological recordings of

XIIn were performed at E15.5 in en bloc

preparations. Note that all activities are

synchronized in control conditions (top set of

traces in a–c). (a) A transverse section below thenVII led to spared independent rhythmic

activities of the e-pF oscillator (top trace) and

the XIIn (bottom trace) and to complete

suppression of activity in the nVII (middle

trace), suggesting that the preBotC oscillator

drove motoneuronal pools before the section.

(b) Selective silencing of the e-pF oscillator

by 20 mM riluzole resulted in a slower motor

rhythm. (c) CNQX application (preserving

left/right de-synchronized activities of the e-pF)

disrupted the activity of the preBotC oscillator

and abolished motor activities. (d) Graph

representing the percentage change (mean ±

s.e.m.) of the frequency of rhythmic activities of

the e-pF (black bars), nVII or XIIn (Motor, gray

bars) and preBotC (white bar) after the transverse section (n ¼ 4), the applications of 10 mM CNQX (n ¼ 11), 0.3 mM DAMGO (n ¼ 7) and riluzole (Ril,

n ¼ 8) in whole-hindbrain preparations (WHB) or transverse slice (slice) preparations (n ¼ 6). * P o 0.05.

Control

Control

CNQX

Section 2% 2%

2%

10 s 10 s

10 s

e-pF

e-pF

e-pF

nVII

nVII

nVII

Xlln

Control

Riluzole

e-pF

nVII

Xlln

e-pF

e-pFMotorPreBötC

nVII

Xlln

200

Per

cent

age

of c

ontr

ol

150

Section CNQX DAMGO Ril Ril DAMGO

SliceWHB

Xlln

*

*

* * *

*

*

100

50

0

Xlln

e-pFnVII

Xlln

Xlln

a b

dc

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1031

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

does not appear abruptly at E15.5 together with the preBotC oscillator,but instead results from a dynamic process in which the e-pF oscillatorseems to be important. In fact, the frequency of motor activitiesincreased linearly with the value of transmission ratios; that is, thenumber of motor bursts to the number of e-pF bursts (E14.5, n ¼ 11;E15.5, n ¼ 19) to reach unity at E15.5 in 12 of 19 preparations (twocases shown in Fig. 4e). Inspection of activities in individual prepara-tions clearly illustrated the quantal nature of motor burst incorpora-tion, eventually giving way to a continuous rhythm (Fig. 4c). Thisdynamic process reflected an increasing excitability at pre-motor/motor synapses during E14.5, as the time lags separating the onsetsof e-pF and motor bursts decreased with increasing transmission ratiosand eventually approached zero when the rhythm became continuous(12 of 19 preparations at E15.5 and 5 of 5 preparations at E16.5;Fig. 4d,f). Therefore, the motor rhythm became continuous duringthe E14.5–15.5 period, as a result of increased efficiency of the e-pF intransmitting activity to motoneuronal pools via a maturing preBotC,which acquires autonomous rhythm generation at the end of the process.

We next investigated the signature of this coupling among oscillatorsat the single-cell level by recording the membrane potential of indivi-dual e-pF cells and the activity of the preBotC oscillator in E15.5 enbloc preparations (Fig. 5). We found, using low chloride concentration

(6 mM) pipette solution for whole-cell recordings, that the e-pF oscillatorcomprised two distinct sets of neurons. During a burst of activity inthe preBotC, 16 of 30 e-pF cells discharged a burst of action potentials(Fig. 5a). In contrast, the remaining 14 e-pF cells showed a pause infiring that was associated with a barrage of chloride-mediated synapticpotentials (Fig. 5b). Because these differences could be the results of avariable effectiveness in changing the intracellular chloride concentra-tion through whole-cell recordings, we performed voltage-clampexperiments to check the polarity and kinetics of underlying synapticcurrents. Working at a holding potential of �40 mV, excited (n ¼ 7)and pausing (n ¼ 10) e-pF cells featured both inward and outwardbackground synaptic events (Fig. 5b,e), and during preBotC burstingactivity, they featured barrages of prominently fast (decay, B1.5 ms)inward (Fig. 5b,c) and slow (decay, B8.0 ms) outward (Fig. 5e,f)synaptic currents, respectively. Fast inward synaptic currents wereblocked by CNQX (data not shown) and slow outward synapticcurrents by bicuculline (Fig. 5g), indicating their respective mediationby AMPA/kainate and GABAA receptors. Thus, pausing cells mayeventually show active inhibition during preBotC inspiratorybursts, whereas those discharging in phase with the preBotC arecandidate neurons through which the e-pF oscillator could contributeto increase the frequency of the RRG. These data suggest that synapticinteractions are established at E15.5 between e-pF and preBotCoscillators. In sum, these experiments show that the e-pF oscillator isfated to couple with the preBotC, the essential oscillator involved in thecontrol of respiration.

Elimination of the e-pF oscillator in Egr2 null mutants

Our data indicate that the e-pF may be important in increasingthe frequency of fetal breathing. In Egr2lacZ/lacZ mice, which are nullmutants for Egr2, the elimination of populations derived fromrhombomeric segments 3 and 5 leads to abnormally slow breathingat birth and poor survival6. Calcium imaging in E15.5 Egr2lacZ/lacZ

preparations (Fig. 6) showed rhythmic fluorescence changes ofthe nVII in phase with the XIIn activity, although at half of thefrequency of Egr2lacZ/+ or wild-type embryos (pooled Egr2lacZ/+/WT,f ¼ 8.9 ± 0.7 bursts min�1, n ¼ 7; Fig. 6a–c; Egr2lacZ/lacZ, f ¼ 3.8 ± 0.3bursts min�1, n ¼ 9; Fig. 6h–j).

Notably, there was a complete absence of rhythmic fluorescencechanges in the e-pF area in the homozygous mutants (Fig. 6i,j).Anatomically, the longitudinal stripe of NK1R expression lateral to thenVII (Fig. 6d,e) was absent (Fig. 6k,l). In transverse and parasagittal(data not shown) sections, NK1R expression (Fig. 6f) and Phox2b+/Islet1/2� cells (Fig. 6g) were lacking at the location of the e-pF

e-pF

e-pF

2%

10 s

10 s 1 s

∆F/F

nVII

nVII

e-pF

nVII

e-pF

15

1

0

3

3

1

1

2 2

10

0 0.2 0.4 0.6

Transmission ratio

0.8 1.0 0 0.2 0.4 0.6

Transmission ratio

0.8 1.0

5

0

nVII

Xlln

E14.5

E14.5

1%

E14.5

E15.5

E15.5

Freq. (min–1)

Mot. Mot.e-pF e-pF

100

0 5 10 15 0 5 10 15

Nb

of p

repa

ratio

n (%

)M

ot. f

req.

(m

in–1

)

Lag

e-pF

-nV

II (s

)

50

0

a

b

e f

c d

Figure 4 Coupling between the e-pF oscillator and motor activity during

development. (a) Spontaneous calcium changes in the e-pF (red trace) and

nVII (blue trace) with electrophysiological recording of XIIn (black trace)

in an E14.5 preparation. Note the absence of motor bursts with continuous

rhythmic e-pF bursts. (b) Distributions of the frequencies of motor (empty

bars, blue fit line) and e-pF (black bars, red fit line) activities obtained from

E14.5 (left graph, n ¼ 11) and E15.5 (right graph, n ¼ 9) preparations. Grey

bars indicate overlapping bins. (c,d) Calcium changes in the e-pF (red tracesand triangles) and nVII (blue traces and triangles) in two (top and middle)

E14.5 and one (bottom) E15.5 preparation (c). Note the increased

occurrence of motor bursts coupled to the e-pF bursts, associated with the

reduction of the time lag (distance between downward red and blue arrows

and vertical lines) between the sequential onsets of the e-pF and nVII bursts (d).

(e,f) Developmental trends among E14.5 (filled circles) and E15.5 (empty

circles) preparations; motor frequency increased (e) and the e-pF–nVII time

lag decreased (f) with augmenting transmission ratios. Data points 1, 2

and 3 correspond to the top, middle and bottom traces in c.

1032 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

oscillator (Fig. 6g,n). Moreover, sectioning the preparation caudal to thenVII reduced the frequency of rhythmic activity by half in Egr2lacZ/+

preparations, as shown above for wild-type embryos (f ¼ 5.1 ± 0.2 burstsmin�1, n ¼ 16 pooled genotypes), but had no significant effect onthe rhythm in Egr2lacZ/lacZ mutants (f ¼ 4.5 ± 0.4 bursts min�1, n ¼ 5,P ¼ 0.3). In fact, the slow rhythm of the homozygous mutants wascomparable with that produced in preBotC transverse slices prepared

from embryos of either genotypes (wild type, 5.1 ± 0.8 bursts min�1,n ¼ 6; Egr2lacZ/lacZ, 5.6 ± 1.5 bursts min�1, n ¼ 5; P ¼ 0.9). Egr2 is thusrequired for the development of the NK1R- and Phox2b-expressingcells that form the e-pF oscillator. This result suggests that the reducedbreathing frequency observed in newborn Egr2 homozygous mutantpups6 may be a result of a lack of entrainment of the preBotC bythe e-pF.

Figure 6 The e-pF oscillator is absent in Egr2 null

mutant embryos. (a,b) Ventral view of an E15.5

wild-type (WT) hindbrain (a) over the facial area(inset) imaged at higher magnification (b) during a

burst of activity of the e-pF (red outline) and the

nVII (blue outline). (c) Traces showing the fast

and synchronized rhythmic fluorescence changes

of the e-pF (red) and nVII (blue) with electrical

activity of the XIIn (black). (d,e) Whole-mount

double immunolabeling for NK1R (green) and

Islet1/2 (red) over one facial area (d, NK1R only).

The inset (white rectangle) shows the lateral

aspect of the nVII at a higher magnification (e).

(f,g) Single transverse sections, taken at the level

indicated by the arrow in d, were double immuno-

labeled for NK1R (green) and Islet1/2 (red) in f,

and for Phox2b (green) and islet1/2 (red) in g.

(h–n) Egr2 null mutant (data are presented as

in a–g) showing the more rostral position of the

facial area, owing to the absence of rhombomeric

segment 5 (h), and the lacking activity of thee-pF (i,j) associated with maintained, but slower,

synchronous rhythmic motor activities of the

nVII and XIIn (j). The absence of the e-pF in the

mutant was associated with a loss (arrowheads) of NK1R+ and Phox2b+ neurons lateral and ventral to the nVII (k,l,n). Dotted lines in a and h indicate the

caudal limit of migrating pontine neurons. Scale bars represent 200 mm.

e-pF cell

20 mV

0.5 s

20 mV

0.5 s 0.2 s

0.2 s

50 pA

50 pA

40 pA

Control

Bic

10 ms

40 pA

10 ms

� = 1.72 ± 0.06 ms

� = 8.85 ± 0.76 ms

� (ms)

� = 1.78 ± 0.23 ms

� = 7.39 ± 0.23 ms

Vh = –40 mV

Vh = –40 mV

e-pF cell

–50 mV

–50 mV

PreBötCint

PreBötCint

25

20

15

Freq

uenc

y (%

)

10

5

0

25

20

15

Freq

uenc

y (%

)

10

5

0

0 2 4 6 8 10 12 14 16

� (ms)0 2 4 6 8 10 12 14 16

a b c g

d e f

Figure 5 The e-pF oscillator at E15.5 comprises two types of neurons. (a,d) Membrane potential trajectories (top traces) of two representative e-pF cells at

E15.5 and preBotC population activity (bottom traces) reveal that e-pF neurons were either excited (a), discharged a burst of action potential or were inhibited

(d), showing a pause in firing, during preBotC rhythmic bursts. Red traces show the average membrane potential trajectory and integrated activity of the

preBotC corresponding to ten superimposed cycles (black traces) locked on the onset of the preBotC bursts (vertical gray lines and arrowheads). (b,e) In voltage

clamp mode, during preBotC bursts of activity (gray background), excited (b) and pausing (e) e-pF cells showed a barrage of prominently inward (black

triangles) and outward (white triangles) synaptic currents, respectively. (c) In excited e-pF cells, inward synaptic currents had fast kinetics. The top set of

traces shows five superimposed inward currents (black triangle, black traces) and their average (red trace), which we used to derive the decay time constant (t)after single exponential fit. The bottom set of traces shows the same analysis for slow outward events collected in between bursts (white triangle). (f) Data for a

pausing e-pF cell are presented as in c. (g) Histograms showing that the bimodal frequency distribution of t’s in one pausing e-pF cell (n ¼ 118 events) in

control (top histogram) transformed under bicuculline into a modal distribution (bottom histogram) owing to preservation of fast (n ¼ 132 events), but not

slow, synaptic events. Whole-cell recordings were performed using a low (6 mM) chloride pipette solution.

∆F/F

A

M

D

M

∆F/F

∆F/F 2%

∆F/F 4%

10 s

10 s

0

4

0

2

Wild

type

Egr

2–/

–

e-pF

Xlln

nVll

e-pF

Xlln

nVll

c g

h

j

k l m

n

i

a b d e f

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1033

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

DISCUSSION

We have identified a neuronal population in the mouse hindbrain thatshows spontaneous rhythmic activity as early as E14.5. The e-pFoscillator, composed of a small number of neurons (B300 per side),constitutes, to the best of our knowledge, the earliest hindbrainneuronal population showing a continuous spontaneous rhythmicactivity (with a period in the second range) that affects the develop-ment of the RRG. Cells of the e-pF oscillator are confined to the surfaceof the hindbrain and are derived from Egr2-expressing cells. We alsofound that the e-pF oscillator cells expressed the homeodomain factorPhox2b, a transcription factor that is expressed and required inneuronal types that maintain bodily homeostasis through the reflexcontrol of digestive, cardiovascular and respiratory functions32. In therat, Phox2b is expressed by the chemosensitive cells of the pFRG inneonates20 and of the RTN in adults18. Our data now extend the role ofPhox2b in the specification of visceral reflex circuit11, from first andsecond order sensory neurons (including the RTN19), noradrenergiccenters or efferent elements, to a set of interneurons linked to theinception of the fetal respiratory rhythm.

Recurrent glutamatergic synaptic inputs are essential for establish-ment of the rhythmic inspiratory drive in preBotC neurons1,25,33,34.This is not the case in the e-pF, where we found that rhythm generationwas maintained in several conditions that impaired glutamatergictransmission. We provide pharmacological evidence that the e-pFrhythm at E14.5 relies on INaP and is modulated by Ih, although thecellular distribution of HCN channels and the allosteric regulations35

ensuring their activation in e-pF cells remain unknown. Furthermore,intercellular synchronization in the parafacial region appears to requirecommunication through gap junctions (CBX sensitive) and probablynot through hemichannels (La3+ insensitive). This contrasts withsynchronization of parafacial regions across the midline that relies onglutamatergic synaptic transmission mediated by AMPA/kainate recep-tors. Because bilateral coactivation is absent in facial transverse slices,the commissural system involved probably resides beyond the slicelimits. One may argue that the preBotC, which is sensitive to CNQX1,has built-in commissural connectivity36,37, and is entrained by andcouples to the e-pF, could take part in setting up this synchrony.However, no functional commissural connectivity was found for thepreBotC at E14.5, when bilateral coactivation of the e-pF is overt28.Hence, the possibility remains that there is another unknown com-missural system or, more simply, that some of the glutamatergic e-pFcells bear commissural axons coursing outside the anterior-posteriorlimits of the slice.

The preBotC oscillator is spared in Egr2 null mutants, in which thee-pF oscillator does not form, indicating that the latter is not requiredfor the emergence of the former. Thus, together with previous evidencethat the preBotC can develop in an isolated context38, our data indicatethat rhythmogenic circuits in the vicinity of branchiomotor nuclei(e-pF/nVII, preBotC/nucleus ambiguous) are deployed independentlyat pre- and post-otic levels and connect to form the RRG. The indepen-dent emergence of the e-pF and preBotC is consistent with speculationsthat they have distinct evolutionary origins; the pFRG would haveappeared first during the evolution of vertebrates, possibly co-optedwhen abdominal expiration was combined with buccal pumping inamphibians39, whereas the preBotC would have emerged later, with themammalian aspiration pump31,40–42.

Three independent lines of evidence (acute riluzole application,transverse sections and Egr2 invalidation) argue that the inactiva-tion of the e-pF oscillator slows down the E15.5 respiratory-likerhythm. We previously proposed that the abnormally slow breathingphenotype of Egr2 null mutants at birth was a result of the

suppression of a rhythm-promoting system that we tentativelylocated in the caudal pontine reticular formation6. The presentdata indicate that the e-pF is a prime contender for setting theneonatal breathing pace at a normal frequency. Hence, at E15.5 inthe mouse, a RRG with a dual organization emerges, in which thee-pF can be considered to contribute in that it increases the rhythmfrequency2, whereas the preBotC can be considered essential in thatit entrains the motor output30.

In addition to its anatomical layout of and its expression of Phox2b,several functional properties indicate that the e-pF may be consideredas a forerunner of the neonatal pFRG. First, similar to the pFRG atneonatal stages, the e-pF upregulates the respiratory-like rhythm2 anddrives the incorporation of motor bursts in a quantal manner31,41 atE15.5. Second, the e-pF at E15.5 includes cells receiving a barrage ofchloride-mediated synaptic currents in phase with the RRG rhythmicbursts. These inputs, excitatory during the E14.5–E15.5 period28,43,may transiently contribute to the phasing of the e-pF to the preBotC,but will cause inhibition during inspiration and post-inhibitoryrebound excitation on later maturation of the chloride gradient43,two features associated with pFRG neurons2,44. Third, our preliminaryresults indicate that the frequency of the e-pF (but not that of thepreBotC) oscillator is increased by a low pH challenge at E14.5(Supplementary Fig. 3), suggesting a role in chemosensitivity that isconsistent with the demonstration that Phox2b expressing neurons ofthe pFRG are CO2-sensitive in the neonatal rat20. Neurons of the pFRG,unlike those of the e-pF, show a pre-inspiratory pattern of discharge(from E19–20 onward in the rat)44. Even though the e-pF cells can alsobe described as being pre-active at the earliest stages (E14.5), when theirassociated calcium changes precede those in facial motoneurons, thisdelay progressively disappears within 24 h as the e-pF and preBotCoscillators couple with one another. Thereafter (at E15.5–16.5), therhythmic activities of the e-pF, the preBotC and the motor nerves aresynchronous. Thus, by E15.5 (that is, the time when it begins pacingfetal breathing), the RRG produces a respiratory-like rhythm charac-terized by a single inspiratory-like phase28,45,46. How pre-inspiratorydepolarization later arises in e-pF cells, thereby transforming them intopFRG neurons, remains obscure, as does the location of the presynapticinhibitory interneurons. Nonetheless, these data argue that the e-pFoscillator is the forerunner of the neonatal pFRG.

A reasonable assumption is that the neonatal pFRG may evolve intothe adult RTN20,47. In this context, the e-pF oscillator could representyet another, earlier developmental stage of this same entity. Pacemakingmechanisms relying on INaP and modulation by Ih, present in the e-pF,may exist only transiently during the postnatal maturation of rhythmicneural circuits24. Their partial downregulation in cells destined to formthe adult RTN may account for the advent of their characteristic tonicfiring mode18,19. Egr2 null mutants were reported to preserve aventilatory response to hypercapnia at birth6. In contrast, the Phox2-b27Ala/+ mutant, a mouse model for CCHS, has both a severe andselective loss of Phox2b/NK1R double-positive neurons at E15.5 in aregion encompassing the e-pF oscillator and a complete lack ofsensitivity to hypercapnia16. These findings, together with ours,would be best explained if the Egr2-dependent, rhythm-promotinge-pF represented a subset of a larger chemosensitive Phox2b-positivepopulation. Examining the contribution to the RTN of progenitordomains producing Phox2b interneurons48,49 outside of Krox20-expressing rhombomeres should allow for a better understanding ofthis potential heterogeneity.

In conclusion, we propose that a two-step developmental processestablishes fetal breathing in mice. In the first step, an Egr2-dependente-pF neuronal population clusters at the ventral surface of the

1034 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

hindbrain and functions as an oscillator. In a second step, at aroundE15.5, a second oscillator, the preBotC, emerges independently and thee-pF couples with and entrains it. In this manner, the dual organizationof the RRG is established at the time of inception of fetal breathing.Early e-pF rhythms preceding the first breathing movements appearto be of clinical, developmental and evolutionary relevance, and arerequired for the respiratory rhythm generator to pace breathing at anormal frequency during the perinatal period. The potential implica-tion of this early stage of respiratory development in breathingdisorders, particularly in CCHS, should be considered.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe thank J.F. Brunet and C. Goridis for comments on the manuscript and thePhox2b antibody, and K. Kullander for VGlut2 mutant mice. This work benefitedfrom the facilities and expertise of the Imagif Cell Biology Unit of the Gifcampus. This work was supported by the Centre National de la RechercheScientifique, Institut National de la Sante et de la Recherche Medicale (M.T.-B.)and ANR grant ANR-07-Neuro-007-01 (G.F.).

AUTHOR CONTRIBUTIONSM.T.-B. and G.F. designed, performed research and analyzed data, M.K. carriedout immunostaining on Egr2 mutants and N.W. performed pharmacologicaltreatments at E15.5. J.C. and P.C. helped with interpreting results and G.F. wrotethe manuscript.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Smith, J.C., Ellenberger, H.H., Ballanyi, K., Richter, D.W. & Feldman, J.L. Pre-Botzingercomplex: a brainstem region that may generate respiratory rhythm in mammals. Science254, 726–729 (1991).

2. Onimaru, H. & Homma, I. A novel functional neuron group for respiratory rhythmgeneration in the ventral medulla. J. Neurosci. 23, 1478–1486 (2003).

3. Blanchi, B. et al. MafB deficiency causes defective respiratory rhythmogenesis and fatalcentral apnea at birth. Nat. Neurosci. 6, 1091–1100 (2003).

4. Cheng, L. et al. Tlx3 and Tlx1 are post-mitotic selector genes determining glutamatergicover GABAergic cell fates. Nat. Neurosci. 7, 510–517 (2004).

5. del Toro, E.D. et al. Generation of a novel functional neuronal circuit in Hoxa1 mutantmice. J. Neurosci. 21, 5637–5642 (2001).

6. Jacquin, T.D. et al. Reorganization of pontine rhythmogenic neuronal networks in Krox-20 knock-out mice. Neuron 17, 747–758 (1996).

7. Shirasawa, S. et al. Rnx deficiency results in congenital central hypoventilation. Nat.Genet. 24, 287–290 (2000).

8. Schneider-Maunoury, S. et al. Disruption of Krox-20 results in alteration of rhombomeres3 and 5 in the developing hindbrain. Cell 75, 1199–1214 (1993).

9. Wilkinson, D.G., Bhatt, S., Chavrier, P., Bravo, R. & Charnay, P. Segment-specificexpression of a zinc-finger gene in the developing nervous system of the mouse. Nature337, 461–464 (1989).

10. Coutinho, A.P. et al. Induction of a parafacial rhythm generator by rhombomere 3 in thechick embryo. J. Neurosci. 24, 9383–9390 (2004).

11. Brunet, J.F. a.G.C. Phox2b and the homeostatic brain. in Genetic Basis for RespiratoryControl Disorders (Gauthier, C., ed.) 25 (Springer, New York, 2008).

12. Dauger, S. et al. Phox2b controls the development of peripheral chemoreceptors andafferent visceral pathways. Development 130, 6635–6642 (2003).

13. Amiel, J. et al. Polyalanine expansion and frameshift mutations of the paired-likehomeobox gene PHOX2B in congenital central hypoventilation syndrome. Nat. Genet.33, 459–461 (2003).

14. Weese-Mayer, D.E., Berry-Kravis, E.M. & Marazita, M.L. In pursuit (and discovery) of agenetic basis for congenital central hypoventilation syndrome. Respir. Physiol. Neuro-biol. 149, 73–82 (2005).

15. Spengler, C.M., Gozal, D. & Shea, S.A. Chemoreceptive mechanisms elucidated bystudies of congenital central hypoventilation syndrome. Respir. Physiol. 129, 247–255(2001).

16. Dubreuil, V. et al. A human mutation in Phox2b causes lack of CO2 chemosensitivity,fatal central apnea and specific loss of parafacial neurons. Proc. Natl. Acad. Sci. USA105, 1067–1072 (2008).

17. Smith, J.C., Morrison, D.E., Ellenberger, H.H., Otto, M.R. & Feldman, J.L. Brainstemprojections to the major respiratory neuron populations in the medulla of the cat.J. Comp. Neurol. 281, 69–96 (1989).

18. Mulkey, D.K. et al. Respiratory control by ventral surface chemoreceptor neurons in rats.Nat. Neurosci. 7, 1360–1369 (2004).

19. Stornetta, R.L. et al. Expression of Phox2b by brainstem neurons involved in chemo-sensory integration in the adult rat. J. Neurosci. 26, 10305–10314 (2006).

20. Onimaru, H., Ikeda, K. & Kawakami, K. CO2-sensitive preinspiratory neurons of theparafacial respiratory group express Phox2b in the neonatal rat. J. Neurosci. 28,12845–12850 (2008).

21. Voiculescu, O., Charnay, P. & Schneider-Maunoury, S. Expression pattern of a Krox-20/Cre knock-in allele in the developing hindbrain, bones, and peripheral nervous system.Genesis 26, 123–126 (2000).

22. Srinivas, S. et al. Cre reporter strains produced by targeted insertion of EYFP and ECFPinto the ROSA26 locus. BMC Dev. Biol. 1, 4 (2001).

23. Suzue, T. Respiratory rhythm generation in the in vitro brain stem-spinal cord prepara-tion of the neonatal rat. J. Physiol. (Lond.) 354, 173–183 (1984).

24. Chan, C.S. et al. ‘Rejuvenation’ protects neurons in mouse models of Parkinson’sdisease. Nature 447, 1081–1086 (2007).

25. Wallen-Mackenzie, A. et al. Vesicular glutamate transporter 2 is required for centralrespiratory rhythm generation, but not for locomotor central pattern generation.J. Neurosci. 26, 12294–12307 (2006).

26. Anselmi, F. et al. ATP release through connexin hemichannels and gap junction transferof second messengers propagate Ca2+ signals across the inner ear. Proc. Natl. Acad. Sci.USA 105, 18770–18775 (2008).

27. Contreras, J.E., Saez, J.C., Bukauskas, F.F. & Bennett, M.V. Gating and regulation ofconnexin 43 (Cx43) hemichannels. Proc. Natl. Acad. Sci. USA 100, 11388–11393(2003).

28. Thoby-Brisson, M., Trinh, J.B., Champagnat, J. & Fortin, G. Emergence of the pre-Botzinger respiratory rhythm generator in the mouse embryo. J. Neurosci. 25,4307–4318 (2005).

29. Gray, P.A., Rekling, J.C., Bocchiaro, C.M. & Feldman, J.L. Modulation of respiratoryfrequency by peptidergic input to rhythmogenic neurons in the preBotzinger complex.Science 286, 1566–1568 (1999).

30. Feldman, J.L. & Del Negro, C.A. Looking for inspiration: new perspectives on respiratoryrhythm. Nat. Rev. Neurosci. 7, 232–242 (2006).

31. Mellen, N.M., Janczewski, W.A., Bocchiaro, C.M. & Feldman, J.L. Opioid-inducedquantal slowing reveals dual networks for respiratory rhythm generation. Neuron 37,821–826 (2003).

32. Brunet, J.F. & Goridis, C. Phox2b and the homeostatic brain. in Genetic Basis forRespiratory Control Disorders (Gauthier, C., ed.) 25–44 (Springer, New York, 2008).

33. Greer, J.J., Smith, J.C. & Feldman, J.L. Role of excitatory amino acids in the generationand transmission of respiratory drive in neonatal rat. J. Physiol. (Lond.) 437, 727–749(1991).

34. Rubin, J.E.H.J., Mendenhall, J.L. & Del Negro, C.A. Calcium-activated nonspecificcation current and synaptic depression promote network-dependent burst oscillations.Proc. Natl. Acad. Sci. USA 106, 2939–2944 (2009).

35. Frere, S.G., Kuisle, M. & Luthi, A. Regulation of recombinant and native hyperpolariza-tion-activated cation channels. Mol. Neurobiol. 30, 279–305 (2004).

36. Koizumi, H. et al. Functional imaging, spatial reconstruction, and biophysical analysis ofa respiratory motor circuit isolated in vitro. J. Neurosci. 28, 2353–2365 (2008).

37. Koshiya, N. & Smith, J.C. Neuronal pacemaker for breathing visualized in vitro. Nature400, 360–363 (1999).

38. Borday, C., Coutinho, A., Germon, I., Champagnat, J. & Fortin, G. Pre-/post-oticrhombomeric interactions control the emergence of a fetal-like respiratory rhythm inthe mouse embryo. J. Neurobiol. 66, 1285–1301 (2006).

39. Vasilakos, K., Wilson, R.J., Kimura, N. & Remmers, J.E. Ancient gill and lung oscillatorsmay generate the respiratory rhythm of frogs and rats. J. Neurobiol. 62, 369–385(2005).

40. Feldman, J.L., Mitchell, G.S. & Nattie, E.E. Breathing: rhythmicity, plasticity, chemo-sensitivity. Annu. Rev. Neurosci. 26, 239–266 (2003).

41. Janczewski, W.A. & Feldman, J.L. Distinct rhythm generators for inspiration andexpiration in the juvenile rat. J. Physiol. (Lond.) 570, 407–420 (2006).

42. Milsom, W.K. Evolutionary trends in respiratory mechanisms. Adv. Exp. Med. Biol. 605,293–298 (2008).

43. Ren, J. & Greer, J.J. Modulation of respiratory rhythmogenesis by chloride-mediatedconductances during the perinatal period. J. Neurosci. 26, 3721–3730 (2006).

44. Onimaru, H. & Homma, I. Developmental changes in the spatio-temporal pattern ofrespiratory neuron activity in the medulla of late fetal rat. Neuroscience 131, 969–977(2005).

45. Kobayashi, K., Lemke, R.P. & Greer, J.J. Ultrasound measurements of fetal breathingmovements in the rat. J. Appl. Physiol. 91, 316–320 (2001).

46. Pagliardini, S., Ren, J. & Greer, J.J. Ontogeny of the pre-Botzinger complex in perinatalrats. J. Neurosci. 23, 9575–9584 (2003).

47. Guyenet, P.G. The 2008 Carl Ludwig Lecture: retrotrapezoid nucleus, CO2 homeostasis,and breathing automaticity. J. Appl. Physiol. 105, 404–416 (2008).

48. Pattyn, A., Morin, X., Cremer, H., Goridis, C. & Brunet, J.-F. Expression and interactionsof the two closely related homeobox genes Phox2a and Phox2b during neurogenesis.Development 124, 4065–4075 (1997).

49. Pagliardini, S. et al. Central respiratory rhythmogenesis is abnormal in lbx1-deficientmice. J. Neurosci. 28, 11030–11041 (2008).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1035

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSMouse lines. All of the mouse lines used in this study were maintained in a

mixed C57Bl6/DBA2 background. The Egr2lacZ allele carries an in-frame inser-

tion of the lacZ coding sequence in the second exon of Egr2 (ref. 50). In the

Egr2cre/+ allele, the Egr2 coding sequence was substituted by the Cre recombi-

nase coding sequence. The R26R-EYFP mouse line22, which allows Cre-

mediated activation of EYFP expression, was kindly provided by F. Costantini

(Columbia University). The day of the vaginal plug was considered E0.5. All

experiments were carried out in accordance with National (JO 87–848) and

European legislation (86/609/CEE) on animal experimentation.

Hindbrain electrophysiology. Pregnant mice were killed by cervical dislocation

at E14.5–E16.5. Embryos were excised from the uterus and kept in oxygenated

artificial cerebrospinal fluid (aCSF) at 20 1C until they were used in electro-

physiological and optical recordings sessions. aCSF was composed of 120 mM

NaCl, 8 mM KCl, 1.26 mM CaCl2, 1.5 mM MgCl2, 21 mM NaHCO3, 0.58 mM

Na2HPO4 and 30 mM glucose (pH ¼ 7.4). For low pH aCSF, the NaHCO3

concentration was decreased to 10.5 mM and the NaCl concentration was

increased to 130.5 mM. The pH of the superfusate was measured continuously

in the recording chamber with a microelectrode (MI-410, Microelectrodes)

calibrated with standard buffers. A high external [K+] was purposefully used

to ensure maintenance of the functional mode of the preBotC, as previously

described28, at the time of its emergence to optimally detect early network

interactions modulating the preBotC activity, although this may have generally

increased neuronal baseline excitability.

En bloc hindbrain preparations were prepared as described previously28 and

were positioned with the ventral side up in the recording chamber. Transverse

450-mm-thick slices were obtained at the level of the nVII or at the level of

the preBotC28 using a vibratome and were transferred into a 1-ml recording

chamber that was continuously superfused at 2 ml min�1 with oxygenated

aCSF at 30 1C. The caudal limit of the facial motor nucleus was used as an

anatomical landmark to generate transverse facial slices and preBotC slices.

To obtain transverse facial slices, we generated a series of 150-mm-thick slices

progressing from caudal to rostral up to the first planes where the presence of the

facial nucleus could be unambiguously distinguished as a result of its higher

optical refringence compared with that of the neighboring tissue. At this point, a

single 450-mm-transverse slice was cut that had an anterior limit corresponding

approximately to the equatorial transverse plan of the facial motor nucleus.

Facial slices were transferred to the recording chamber and positioned with

the anterior side up. For preBotC slices, the slicing was carried out as described

above, but progressing from rostral to caudal, up to the last plane of the facial

nucleus. At this point, a 250-mm-thick slice was made to reach the anterior limit

of the preBotC, and a single 450-mm-thick transverse preBotc slice was produced.

Hypoglossal nerve root activity and population activity (e-pF or preBotC)

in whole hindbrain preparations were recorded using glass micropipettes

suction electrodes (150-mm tip diameter). For e-pF recordings in whole hind-

brain preparations, the electrodes were positioned on the brain surface, and

for the recording of the preBotC, the electrodes were progressively inserted at

depths of about 100–150 mm below the hindbrain surface, sufficient to record

the activity of the preBotC. The micropipettes filled with aCSF were connected

through silver wires to a high-gain alternating current amplifier (Grass, 7P511),

filtered (bandwidth, 3 Hz through 3 kHz), integrated using an electronic filter

(Neurolog System, time constant of 100 ms), recorded on a computer via a

digitizing interface (Digidata 1322A, Molecular Devices) and analyzed with the

pClamp9 software (Molecular Devices). Whole-cell patch-clamp neuronal

recordings were performed under visual control using differential interference

contrast (DIC) and infrared video microscopy, an Axoclamp2A amplifier

(Molecular Devices), a digitizing interface (Digidata 1322A, Molecular Devices)

and the software program pClamp9 (Molecular Devices). Patch electrodes

(resistance of 4–6 MO) were pulled from borosilicate glass tubes (Clark GC

150TF) and filled with a solution containing 140 mM potassium gluconic acid,

1 mM CaCl2, 6 mM H2O, 10 mM EGTA, 2 mM MgCl2, 4 mM Na2ATP and

10 mM HEPES (pH 7.2). In voltage-clamp mode, we analyzed the persistent

sodium current INaP and the hyperpolarization-activated current Ih. INaP was

activated using a slow depolarizing ramp from �60 mV to +10 mVand blocked

by 5–20 mM riluzole (Fig. 2d). The Ih current was evoked by applying

hyperpolarizing voltage steps (from �50 mV to �120 mV) and was blocked

by 100 mM ZD7288. The Ih current/voltage curve was built by measuring the

difference between current amplitudes values measured at the beginning and

the end of each voltage step (Fig. 2f). In current clamp, the Ih current activation

led to depolarizing sags in response to hyperpolarizing current pulses (Fig. 2f).

Drugs were obtained from Sigma, dissolved in aCSF and bath-applied for

10–15 min a final concentration of 0.1 mM Substance P (SP), 0.3 mM DAMGO,

10 mM CNQX, 10 mM AP5, 10 mM bicuculline, 5 mM strychnine, 5–20 mM

Riluzole), 50 mM CBX, 100 mM ZD7288 and 100 mM lanthanum. To minimize

the risk of nonspecific effects resulting from long-term exposure to riluzole,

ZD7288, La3+ and CBX, we measured their effects during a 2-min period

beginning 5 min after switching on the perfusion source containing the tested

compounds, a delay that is approximately tenfold larger than the time constant

of concentration change kinetics of the chamber. In addition, when examining

rhythm generation in the e-pF at E14.5, nonspecific effects of these compounds

that were linked to interactions with glutamatergic transmission, being dis-

pensable, were probably minimal. We examined whether (e-pF cell recordings,

n ¼ 4, data not shown) 50 mM CBX had an effect on the amplitude and kinetics

of action potential and on the kinetics of synaptic currents. Values are given as

means ± s.e.m. Differences were regarded as significant at P o 0.05.

Calcium imaging. Whole hindbrain and slices were incubated for 40 min in

oxygenated aCSF containing the cell-permeable calcium indicator dye Calcium-

Green 1AM (10 mM, Molecular Probes). Whole hindbrain preparations were

positioned in the recording chamber with the ventral side up. After a 30-min

recovery period in the recording chamber to wash out the dye excess, a

standard epi-fluorescent illumination system on an E-600-FN upright micro-

scope (Nikon) equipped with a fluorescein filter block was used to excite the

dye and capture the emitted light. Fluorescence images were captured with a

cooled CCD camera (Coolsnap HQ, Photometrics) using an exposure time of

100 ms in overlapping mode (simultaneous exposure and readout) during

periods of 60–180 s and analyzed using Metamorph software (Universal

Imaging). To perform calcium imaging of YFP-expressing cells, we first

acquired images of EYFP cells with corresponding DIC images. After dye

loading, a careful positioning over the same cellular field was ascertained

through visualizing cellular profiles using DIC images and adjusting their

alignments. The EYFP-labeled cell false-colored red image and the calcium-

loaded cell false-colored green image were overlaid (Fig. 1g) to determine

double-labeled somas and to position regions of interest for measurements of

fluorescence changes (Fig. 1g). In all cases, the average intensity in a region of

interest was calculated for each frame and the changes in fluorescence were

normalized to their initial value by expression as the ratio of changes in

fluorescence to initial fluorescence (F/F).

Immunofluorescence. Mouse embryos were fixed for 2–3 h in 4% para-

formaldehyde (wt/vol) in phosphate-saline buffer, cryoprotected in 25% sucrose

(wt/vol), embedded in O.C.T. Compound (Tissue-Tek) and sectioned at 14 or

20 mm. Immunohistochemistry was performed on frozen sections as previously

described28. We used antibodies to rabbit NK1R (Sigma, 1:5,000), guinea pig

Islet1/2 (gift from J. Ericson, Karolinska Institute, 1:1500), chick GFP (Aves Lab,

1:2,000) and rabbit Phox2b (gift from C. Goridis, ENS, 1:1,500). Species

specific antibodies conjugated to Alexa 488, (Molecular Probes), Cy3 and

Cy5 (Jackson ImmunoResearch) were used. Biocytin-filled neurons were

labeled using Extra-Avidin-FITC (Sigma, 1:400). Rocking incubation with

primary antibodies was carried out overnight at 4 1C and secondary antibodies

were incubated for 3 h. The material was mounted in Vectashield medium

(Vector Labs). Counts of e-pF cells were made in an area delimited ventrally by

the medullary surface, dorsally by the nVII, rostrally by the rostral end of the

nVII and extending 100 mm caudal to the nVII. Cells were counted in the

transverse plan using all consecutive 20-mm sections (n ¼ 2) or every other

section (n ¼ 1) and in the sagittal plane on every other section (n ¼ 1). In all

cases, cells were counted on both sides; estimation of the total number of cells

included a multiplication factor of 2 when appropriate. Fluorescent labeling

was visualized on a Leica SP2 confocal microscope. All figures were color

corrected and assembled using Adobe Photoshop and Illustrator.

50. Voiculescu, O. et al. Hindbrain patterning: Krox20 couples segmentation and specifica-tion of regional identity. Development 128, 4967–4978 (2001).

NATURE NEUROSCIENCE doi:10.1038/nn.2354

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Cocaine-evoked synaptic plasticity: persistence inthe VTA triggers adaptations in the NAc

Manuel Mameli1, Briac Halbout2, Cyril Creton1, David Engblom3, Jan Rodriguez Parkitna3,Rainer Spanagel2 & Christian Luscher1,4,5

Addictive drugs hijack mechanisms of learning and memory that normally underlie reinforcement of natural rewards and induce

synaptic plasticity of glutamatergic transmission in the mesolimbic dopamine (DA) system. In the ventral tegmental area (VTA),

a single exposure to cocaine efficiently triggers NMDA receptor–dependent synaptic plasticity in DA neurons, whereas plasticity

in the nucleus accumbens (NAc) occurs only after repeated injections. Whether these two forms of plasticity are independent or

hierarchically organized remains unknown. We combined ex vivo electrophysiology in acute brain slices with behavioral assays

modeling drug relapse in mice and found that the duration of the cocaine-evoked synaptic plasticity in the VTA is gated by

mGluR1. Overriding mGluR1 in vivo made the potentiation in the VTA persistent. This led to synaptic plasticity in the NAc, which

contributes to cocaine-seeking behavior after protracted withdrawal. Impaired mGluR1 function in vulnerable individuals could

represent a first step in the recruitment of the neuronal network that underlies drug addiction.

Cocaine, one of the most addictive drugs of abuse, can induce synapticplasticity of glutamatergic transmission in the mesolimbic DA systemof rodents1–3. Within hours of a single cocaine injection, excitatoryinputs onto DA neurons of the VTA are strengthened, which can bemonitored by an increased AMPA/NMDA ratio4. This drug-evokedpotentiation is in part mediated through an exchange of GluR2-containing and GluR2-lacking AMPA receptors, leading to excitatorypostsynaptic currents (EPSCs) that are sensitive to polyamines andhave a rectifying current-voltage relationship5,6. This plasticity istriggered by all of the addictive drugs tested so far and lasts about5 d4,7,8. If cocaine is self-administered repetitively for 2 weeks, plasticityin the VTA becomes persistent, and can be detected even monthsafter withdrawal9.

In the NAc, cocaine-evoked plasticity also occurs, but on a slowertimescale and with a steeper induction threshold. A single cocaineinjection is not sufficient to trigger changes in synaptic transmission.However, after 5 d of consecutive injections, AMPA/NMDA ratios aredepressed10. During withdrawal from both passive exposure and self-administration, this long-term depression (LTD)-like plasticity in theNAc transforms into a potentiation through an insertion of AMPAreceptors (AMPARs)10,11. Biochemical and electrophysiological inves-tigations suggest that the inserted receptors are GluR1 homomericchannels12,13. Recently, it has been observed that the in vivo inhibitionof these channels by a polyamine toxin substantially reduces cue-induced cocaine seeking after withdrawal11. This behavioral phenom-enon, termed incubation of craving, becomes apparent followingprotracted withdrawal and is believed to mimic relapse in humans14.

Similarly, cocaine-seeking was blocked by injections of antisense oligo-nucleotides of GluR1 mRNA into the NAc15. GluR2-lacking AMPARsseem to be involved in a receptor redistribution that contributes to theremodeling of neuronal networks underling addictive behaviors16.

Taken together, a single injection of cocaine causes a switch-like,rapid, but transient, potentiation of excitatory inputs in the VTA,whereas several injections are required to induce plasticity in the NAc.We manipulated the persistence of the plasticity in the VTA via anmGluR1-dependent mechanism and used genetically modified micelacking NMDA receptor (NMDAR) selectively in DA neurons of themidbrain to test the effects of cocaine on the enduring forms ofplasticity in the NAc and on the incubation of craving.

RESULTS

mGluR1-dependent reversal of plasticity in vivo

In vitro, cocaine-evoked plasticity in the VTA can be reversed rapidlyvia mGluR1 activation by either synaptic glutamate or exogenousagonists5. Mechanistically, such mGluR-LTD exchanges the GluR2-lacking AMPARs that are inserted during cocaine-evoked plasticitywith newly synthesized GluR2-containing receptors17. We thereforeasked whether functional mGluR1 in the VTA are required for theendogenous reversal of cocaine-evoked plasticity in vivo. To this end,we interfered with mGluR1 function by using a dominant-negativepeptide that precludes the binding of mGluR1 and Homer1b/c(TAT-mGluRct)18. This peptide specifically inhibits mGluR-dependentsynaptic plasticity in the hippocampus19. The peptide was TAT- andfluorescein-conjugated and stereotactically delivered bilaterally to the

Received 26 May; accepted 15 June; published online 13 July 2009; doi:10.1038/nn.2367

1Department of Basic Neurosciences, Medical Faculty, University of Geneva, Geneva, Switzerland. 2Department of Psychopharmacology, Central Institute of MentalHealth, Mannheim, Germany. 3Division of Molecular Biology of the Cell I, German Cancer Research Center, Heidelberg, Germany. 4Clinic of Neurology, Departmentof Clinical Neurosciences, Geneva University Hospital, Geneva, Switzerland. 5Geneva Neuroscience Center, Geneva, Switzerland. Correspondence should be addressedto C.L. (christian.luscher@unige.ch).

1036 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

VTA of mice. To ensure that the peptide did not spread beyondthe VTA, but remained present in DA neurons for the duration ofthe experiment (Fig. 1a), we visualized the fluorescence 8 d afterthe injection (Fig. 1b) and made recordings from fluorescent-labeledDA neurons. Although the TAT peptide was probably also loadedinto GABA neurons, the overwhelming majority of the fluorescentneurons were DA neurons, as judged by electrophysiological criteria(see Online Methods). We also tested whether the dominant-negativepeptide reduced currents evoked by the potent agonist of mGluRsS-3,5-dihydroxyphenylglycine (DHPG) and whether it blocked mGluR-LTD in vitro, which was indeed the case (Supplementary Fig. 1). We thenmeasured EPSCs at –70, 0 and +40 mV to calculate the rectificationindex (EPSC–70mV/EPSC+40mV). In mice in which TAT-mGluRctwas delivered, rectification indices measured 7 d after a singlecocaine injection were significantly higher than those of controlmice or after the stereotactic injection of a TAT control peptide(TAT peptide delivery� intraperitoneal injection interaction: F1,31¼ 14.14,P ¼ 0.0007; Fig. 1c,d). In the presence of the dominant-negativepeptide, cocaine-evoked plasticity was intact, which demonstratesthat mGluR-Homer binding is not required for induction (for example,NMDA dependence4) or expression (for example, PICK1-dependentAMPAR redistribution5; Supplementary Fig. 1). Taken together, thesefindings suggest that the disruption of Homer 1b/c–mGluR interactionin the VTA renders the cocaine-evoked plasticity persistent.

To confirm this result pharmacologically, we blocked mGluR1with daily systemic (intraperitoneal) injections of the antagonist1-aminoindan-1,5-dicarboxylic acid (AIDA)following a single injection of cocaine(Fig. 2a). At a dose of 0.25 mg per kgof body weight, AIDA selectively inhibitsmGluR1 receptors20. Because we establishedthat 7 d is normally enough for a full reversalof cocaine-evoked plasticity, we cut slices onday 8 and found that the current-voltagerelationship (I-V) was rectifying (Fig. 2b;for corresponding AMPA/NMDA ratio seeSupplementary Fig. 2), which was not thecase if one injection of cocaine was followedby daily injections of saline (Fig. 2b). Asfurther controls, we ensured that saline orAIDA alone had no effect on the rectifica-tion index. In contrast, with seven injectionsof cocaine the EPSCs were rectifying and

the rectification index was significantly elevated compared withcontrol values (P o 0.001). Notably, the increased rectification indexfollowing seven injections of cocaine was similar to what we observedpreviously 24 h after a single injection, arguing that the plasticity mayalready be saturated by a single injection of cocaine8. To determinethe temporal requirement of mGluR1 function, we broke down theAIDA treatment and compared an immediate treatment (days 2 and 3)with a late treatment (days 5 and 6) and prepared slices at day 8. Onlythe latter was efficient in maintaining synaptic plasticity, arguing thatmGluR1 must be activated during a narrow time window to reversethe cocaine-evoked plasticity (Supplementary Fig. 3). Taken together,one injection of cocaine leads to persistent plasticity when followedby daily injections of AIDA, suggesting that the endogenous reversalof cocaine-evoked plasticity depends on functional mGluR1 receptorsin vivo.

We have previously shown that the positive mGluR1-modulatorRo 67-7674 leads to the disappearance of cocaine-evoked plasticity by24 h after one intraperitoneal injection5. We therefore tested whethera positive modulation of mGluR1 receptors could reverse cocaine-evoked plasticity with a more robust induction by seven injections ofcocaine, each given 1 h after Ro 67-7476 (Fig. 2a). Indeed seveninjections of cocaine when paired with Ro 67-7476 yielded a linear I-Vcurve (Fig. 2c). As a control, Ro 67-7476 injected together with salinehad no effect on rectification, whereas seven injections of cocainealong with saline, similar to the result above, led to significant rectifica-tion (P ¼ 0.031). Thus, positive modulation of mGluR1 causes rapid

SNc

MT

Mid

line

Caudal

Rostral

SNr

Pipette VTADay

VT

A s

lice

81 2 7

mGluRct Sal

mGluRct Coc

Control Sal

0

VTA IP

Control Coc Rec

tific

atio

n in

dex 5

0

1

2

3

4

(9) (12)

***

(9)(7)5 ms

50 pA

ca b dInjections

20 µm

500 µm

Figure 1 Disruption of Homer 1b/c–mGluR interaction in the VTA renders cocaine-evoked plasticity persistent. (a) Saline (blue and black circle) or cocaine (red and

gray circle) were injected intraperitoneally (IP) 24 h after stereotactic delivery of TAT-mGluRct or TAT control into the VTA. Acute midbrain slices were then prepared

at day 8. (b) Confocal image obtained 8 d after injection of TAT-mGluRct (0.6 ml at 1 mM). The fluorescence signal is superimposed on the transmitted light image.

The image is shown at 40� magnification. MT, medial terminal nucleus of the accessory optical tract. SNc, substantia nigra compacta; SNr, substantia nigrareticulata. (c) Examples of AMPAR-EPSCs obtained at �70, 0 and +40 mV. (d) Bar graph of averaged rectification indexes with superimposed scatter plot

(symbols as above). Dashed line in the bar graph indicates the value of a linear I-V curve (70/40 ¼ 1.75). *** indicates P o 0.001.

RI

5

0

1

2

3

4

VT

A s

lice

SalCocCocCocAIDA *** ***

Day 81 2 7

SalSalCocAIDAAIDA

IP injections

Day

VT

A s

lice

81 7

(15) (16) (12) (20) (11)0

1

2

3

4

5 *

(6) (9) (6)(7)

10 ms

50 pA

a b c

Sal + CocRo 67 + Coc

Ro 67 + SalSal + Sal

Figure 2 Bi-directional modulation of mGluR1 controls the persistency of cocaine-evoked plasticity

in VTA. (a) Experimental protocols used for intraperitoneal injections of AIDA (top) and Ro 67-7476

(Ro 67, bottom). (b) AMPAR-EPSCs obtained at �70, 0 and +40 mV and respective mean rectification

index for the experimental group treated with AIDA (F4, 69 ¼ 33.6, *** indicates P o 0.001).

(c) Corresponding graphs using Ro 67-7476, the positive modulator of mGluR1 (saline/cocaine �saline/ Ro 67-7476 interaction: F1, 23 ¼ 5.2, P ¼ 0.031).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1037

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

reversal of cocaine-evoked plasticity in the VTA in vivo, even inresponse to repeated cocaine injections.

Bi-directional control of plasticity in the NAc

We next used the modulation of mGluR1 as a tool to test the linkbetween cocaine-evoked potentiation in the VTA and cocaine-evokeddepression in the NAc. We started by testing the effect of the mGluR1antagonist AIDA. Similar to the experiments described above, a firstdose of cocaine was followed by six daily injections of either saline,cocaine or AIDA, and on day 8 we prepared coronal brain slicescontaining the NAc (Fig. 3a). In medium spiny neurons of the NAcshell, a day after the last injection, the I-V curves were linear in allconditions (Fig. 3b,c). A single injection of cocaine was not sufficientto induce changes in the AMPA/NMDA ratio, but transmission wassignificantly decreased with seven injections, consistent with pre-vious reports10 (Fig. 3b,d). Also consistent with previous reports, thedecrease of AMPA/NMDA ratio elicited by cocaine occluded theinduction of low-frequency stimulation (LFS)-induced LTD, suggestingthat the two phenomena share underlying mechanisms21 (Fig. 3e,f).

AIDA injections had no effect on synaptic transmission (Fig. 3b,d),but lowered the threshold for cocaine-evoked plasticity in the NAc; asingle injection of cocaine was sufficient to significantly decrease theAMPA/NMDA ratio and occlude the synaptically induced LTD ifthat injection of cocaine was followed by six injections of AIDA(Fig. 3b,d,g). Conversely, when we applied seven injections of cocainewith Ro 67-7476 (Fig. 3h), the decrease of AMPA/NMDA ratio in theNAc was abolished (Fig. 3i,k).

To ensure that effects on plasticity in the NAc were the results of alocal intervention at the level of the VTA, we carried out a series ofexperiments (Fig. 4a) using stereotactic injections in the VTA of theTAT-conjugated dominant-negative peptide (Fig. 1a). We observedthat with this selective, local disruption of mGluR1 function in neuronsof the VTA, a single injection of cocaine was sufficient to trigger thedepression in the NAc. The control peptide followed by one injection ofcocaine or saline had no effect on the AMPA/NMDA ratios (Fig. 4b–d).Similarly one saline injection after TAT-mGluRct delivery also had noeffect (Fig. 4b,d). We next tested whether overriding mGluR1 in theVTA had an effect on enduring forms of plasticity, such as the insertionof GluR2-lacking AMPARs that can be observed a month after with-drawal. To this end, we measured the rectification index in mediumspiny neurons of the NAc 35 d after the last cocaine injection (Fig. 4e).Ten daily injections followed by this protracted withdrawal period ledto strongly rectifying EPSCs compared with controls (t14 ¼ 2.19, P o0.05; Fig. 4f,h). A similarly high rectification index was observedafter only one cocaine injection if the mouse was pre-treated witha stereotactic injection of TAT-mGluRct (t16 ¼ 3.4, P o 0.01;Fig. 4f,h). Thus, our data suggest that persistent plasticity in theVTA triggers a synaptic depression in the NAc and that a swift reversalof this plasticity may prevent synaptic alterations in the NAc. Takentogether, interfering selectively with mGluR function in neurons of theVTA controls early forms and enduring forms of cocaine-evokedplasticity in the NAc.

VTA plasticity modulates cocaine seeking after withdrawal

It has been suggested that insertion of GluR2-lacking AMPARs con-tributes to the development of incubation of craving11. We hypothe-sized that preventing cocaine-evoked plasticity in the NAc wouldimpair drug-seeking behavior after protracted withdrawal. We there-fore tested this behavioral phenomenon in Grin1loxP/loxP; Slc6a3-creERT2 mice (Grin1 is also known as NR1 and Slc6a3 is also knownas DAT ; we refer to these mice as NR1DAT-CreERT2 mice). In this mouse,NMDARs are ablated in DA neurons during adulthood after tamoxifen-triggered recombination and we recently found that a single injection ofcocaine no longer induces a synaptic potentiation in the VTA22. Follow-ing food shaping, NR1DAT-CreERT2 and control mice were trained for8 d in 4-h sessions to lever press for intravenous cocaine self-administra-tion that was associated with a light cue. NR1DAT-CreERT2 mice acquiredstable lever pressing for cocaine similarly to controls (Fig. 5a; for leverpresses on the inactive lever see Supplementary Fig. 4). After 35 d of

50

–100

–80 40

Nor

m. E

PS

C (

%)

V (mV)

Day

NA

c sl

iceCoc

CocCoc

81

AIDA

2 7

Sal

CocAIDAAIDA

AM

PA

/NM

DA

+40

mV

0

0.5

1.0

1.5

2.0

2.5

(15) (6) (8) (7)

***

Day

NA

c sl

ice

81 7

Coc + SalRo 67 + Coc

Ro 67 + Sal

Sal + Sal

20 ms

50 pA

20 ms

50 pA

a b

c d

Time (min)

0 10 20 30 40

0

50

100

150

Nor

m. E

PS

C (

%)

50 0 10 20 30 40 50 0 10 20 30 40 50

n = 11 n = 6 n = 8

LFS LFS LFS20 ms

50 pA

e f g

h i

50

–100

–80 40

Nor

m. E

PS

C (

%)

V (mV)

j

AM

PA

/NM

DA

+40

mV

0

0.5

1.0

1.5

2.0

2.5

*

(6) (8) (11)(8)

k

IP injections

IP injections

Figure 3 Modulation of mGluR1 controls cocaine-evoked plasticity in the

NAc. (a) Experimental protocol. (b) AMPAR-EPSCs obtained at �70, 0 and

+40 mV and NMDAR-EPSCs (shaded trace) obtained at +40 mV in medium

spiny neurons of the NAc shell. (c) I-V plot for AMPAR-EPSCs. (d) Averaged

AMPA/NMDA ratios obtained at +40 mV for each experimental group (F3, 31

¼ 9.23, * indicates P o 0.05, ** indicates P o 0.01). (e–g) LFS-LTD in the

different experimental groups. (h) Experimental protocol. (i) AMPAR-EPSCs

and NMDAR-EPSCs obtained as above in slices from mice treated with themGluR1 enhancer Ro 67-7476 as indicated in h. (j). I-V plots of AMPAR-

EPSCs. (k) Averaged AMPA/NMDA ratios obtained at +40 mV (saline/cocaine

� saline/ Ro 67-7476 interaction: F1, 33 ¼ 4.34, P ¼ 0.045).

1038 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

withdrawal, the mice were tested for cue-induced cocaine seeking. Thistest session lasted 90 min, during which presses on the active leverunder a fixed ratio schedule 1 triggered conditioned stimulus pre-sentation without cocaine delivery. The incubation of cocaine-seekingbehavior was significantly reduced in NR1DAT-CreERT2 mice (factor geno-type: F1,14 ¼ 5.126, P o 0.05; factor lever: F1,14 ¼ 61.855, P o 0.001;genotype � lever interaction: F1,14 ¼ 2.30, P 4 0.05; Fig. 5b).

Slices of the midbrain and the ventral striatum were then preparedfrom the very same mice within 48 h to measure AMPA/NMDA ratiosand rectification indices. In the VTA of NR1DAT-CreERT2 mice, NMDAEPSCs were absent and AMPAR EPSCs showed a linear rectificationindex value. In control mice, the AMPA/NMDA ratio and rectification

index were high (t11 ¼ 2.61, P o 0.05; Fig. 5c,d), confirming thepresence of VTA plasticity following cocaine self-administration evenafter a protracted withdrawal period, a finding that was made previ-ously in rats9. Moreover, in slices of the NAc, rectification was presentin control mice, but not in NR1DAT-CreERT2 mice (t27 ¼ 5.68, P o 0.001;Fig. 5c,d). Plotting the rectification index as a function of the behav-ioral score revealed a significant correlation (weighted regression:F1,14 ¼ 13.7, P ¼ 0.003, r2 ¼ 0.533; Fig. 5e). Taken together, preventingthe induction of cocaine-evoked plasticity in DA neurons of theVTA abolished early and enduring plasticity in the NAc and attenuatedcue-induced cocaine seeking after prolonged withdrawal.

DISCUSSION

We found that mGluR1 receptors on DA neurons limit the persistenceof the cocaine-evoked potentiation in the VTA. Moreover, overridingmGluR1 function caused enduring adaptations in the VTA that set thestage for the synaptic plasticity in the NAc, eventually leading to theinsertion of GluR1 homomeric AMPARs and shaping cue-inducedcocaine seeking.

Although these observations are compatible with a hierarchicalorganization of cocaine-evoked synaptic plasticity in between theVTA and the NAc, we did not identify the nature of the signal thattransfers the message. It is possible that enhanced excitation ofprojection neurons in the VTA may facilitate the coincident releaseof DA and glutamate in the NAc through a continuous enhancedrelease of DA. This may then shift the threshold for the induction oflocal plasticity in the NAc by affecting circuit excitability or byintegrating biochemical signals such as intracellular calcium orCaMKII signaling10,23.

If such hierarchical organization of cocaine-evoked plasticity isconfirmed downstream of the NAc, it may be of relevance in thecontext of the anatomical organization of the striatum24,25. Thesetracing studies show that VTA and NAc are part of a reciprocal spiral

50

–100

–80 40

Nor

m. E

PS

C (

%)

V (mV)

Day

NA

c sl

ice

ControlmGluRct

451 2 10

SalCocCoc

0

Coc

SalCocSal

Sal

5

0

1

2

3

4

Rec

tific

atio

n in

dex

(5) (11) (9) (9)

* **

Day

NA

c sl

ice

81 2 7

mGluRct Sal

mGluRct Coc

Control Sal

0

Control Coc

AM

PA

/NM

DA

+40

mV

0

0.5

1.0

1.5

2.0

2.5

(7) (10)

*

(8)(7)

20 ms

50 pA

20 ms

50 pA

a b

c d

e f

g h50

–100

–80 40

Nor

m. E

PS

C (

%)

V (mV)

VTA IPInjections

VTA IPInjections

Leve

r pr

esse

s (9

0 m

in)

0

40

80

120

2 4 6 8

Infu

sion

s (4

h)

Days

Control

SA Withdrawal TestDay 1 8 43

Leve

r pr

esse

s (9

0 m

in)

AM

PA

/NM

DA

+40

mV

VTA NAcVTA

VTA NAc

Rec

tific

atio

n in

dex

00.40.81.2

012345

(6) (6) (7) (17) (13)

***

*

3)

20 ms

50 pA

10 ms

50 pA

a b c

e

d

NR1DAT-CreERT2

Rectification index0

0

100

200

400

300

1 2 3 4 5

Control

NR1DAT-CreERT2

0

50

100

250

150

200

Control NR1DAT-CreERT2

Active

Inactive

*

Lever

Figure 5 Disruption of NMDARs in midbrain DA neurons abolishes enduring plasticity in the NAc and

reduces incubation of craving. (a) Timeline of experimental protocol and number of cocaine infusions during

self-administration in the two groups. (b) Cue-induced lever pressing at day 35 in the NR1DAT-CreERT2 mice.

(c) AMPA- and NMDA-EPSCs (gray) obtained at �70, 0 and +40 mV for NR1DAT-CreERT2 (red) and control mice

(black) in VTA and NAc. (d) Averaged AMPA/NMDA ratio and rectification index in the VTA and NAc for the two

genotypes. (e) Correlation between rectification index and lever presses for each neuron recorded (filled symbols

represent the average rectification index for a given mouse).

Figure 4 Early and enduring synaptic plasticity in the NAc after a single

injection of cocaine. (a) Disruption of mGluR1 function selectively in the VTA

through stereotactic injection of TAT-mGluRct (see Fig. 1a) and preparation of

NAc slices at day 8. (b) AMPA- and NMDA-EPSCs obtained at –70, 0 and

+40 mV. (c) I-V plots of AMPA-EPSCs. (d) Averaged AMPA/NMDA ratio

obtained at +40 mV (TAT peptide delivery � intraperitoneal injection

interaction: F1, 26, 4.24, P ¼ 0.05). (e) Experimental protocol as described

in a, but preparation of NAc slices was performed after 35 d of withdrawal.(f) AMPA- and NMDA-EPSCs obtained at –70, 0 and +40 mV. (g) I-V

relationship of AMPA-EPSCs. (h) Averaged rectification indices.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1039

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

in connectivity between the midbrain and the striatum. These findingsimplicate the dorsal components of the DA system in cocaine-seekinghabits that are observed in addiction, which implies that the early drugeffects on the VTA need to be transferred to the dorsal striatum via theNAc2,26. The hierarchical link of cocaine-evoked plasticity betweenVTA and NAc suggested here may represent the first leg of the spiralingconnectivity underlying compulsive habits27. However, our work alsoclearly demonstrates that core components of addiction such as relapsecan be modeled in rodents and do not require the recruitment of thedorsal striatum. It is possible that synaptic plasticity may emerge inmore dorsal structures at later phases of the process.

Notably, recent studies suggest that the early cocaine-evoked plasti-city in DA neurons of the VTA does not mediate concurrent short-termbehavioral effects of the drug22,28, although this idea was supported byearlier experiments. For example, local application of NMDAR antago-nists in the VTA abolishes behavioral sensitization29 and conditionedplace preference (CPP)30, as does genetic knockout of the AMPAreceptor subunit GluR1 (ref. 31, but see ref. 32). Moreover, the viraloverexpression of GluR1 enhances cocaine-induced behavioral sensiti-zation in rats that have never been exposed to cocaine33. However, theseapproaches affect all cell types in the VTA. A selective deletion of theNMDAR subunit NR1 in the DA neurons of adult mice abolishescocaine-evoked plasticity onto dopamine neurons of the VTA, but doesnot affect behavioral sensitization and the development of a CPPresponse22. In contrast, in the same mice, later drug-associated beha-viors such as reinstatement of CPP are blocked. This finding remainscontroversial in the light of another study using a similar geneticapproach, although differences in the CPP protocol may explain thisdiscrepancy28. Altogether, these results raise the possibility thatcocaine-evoked plasticity in the VTA may be important behaviorallyfor the late-stage drug-seeking behaviors, such as we observed. Becausea recent report suggests that DA neurons projecting to the prefrontalcortex and the amygdala have reduced DAT expression34, which maypreclude recombination in the mutant mice, NAc projecting neuronsare likely to mediate the bulk of the behavioral phenotype.

Our results may shed light on the mechanism underlying theprogression from recreational use to compulsive abuse and relapse indrug addicts. They are consistent with the observation that self-administration of cocaine causes a more persistent plasticity in theVTA than passive injections9. Extending the observation that mGluR1-LTD rapidly reverses cocaine-evoked plasticity in vitro, we found thatmGluR1-LTD in the VTA is required for the endogenous reversal ofearly cocaine-evoked plasticity in the VTA, as well as for adaptive gatingof later cocaine-evoked plasticity in the NAc. Therefore, recruitment ofmGluR1 functions as a protective mechanism to counteract drugexposure. However, there is a critical time window over whichmGluR1 might control cocaine-evoked plasticity. Therefore, we thinkthat it is unlikely that mGluR enhancement will be useful as a means toreverse previously established addiction in cocaine abusers, wherepersistent synaptic plasticity has already been relayed to other struc-tures. However, our results do raise the possibility that individuals withdeficient mGluR1-dependent LTD mechanisms may be particularly atrisk of addiction. We suggest that the screening of genes controllingmGluR1 function may improve clinical efforts to assess individualvulnerability to drug addiction.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe thank members of the Luscher laboratory, as well as B.J. Everitt, M. Frerking,A. Luthi, M. Serafin, M. Carta and C.F. Valenzuela for helpful discussionsand suggestions regarding the manuscript. R. Sprengel (Max Planck InstituteHeidelberg) generated the NR12loxP/loxP mouse line and G. Schutz (DeutschesKrebsforschungszentrum Heidelberg) provided the NR1DAT-CreERT2 mouse line.C.L. is supported by grants from the Swiss National Science Foundation andthe Swiss initiative in system biology (SystemsX: neurochoice). R.S. is supportedby Nationales Genomforschungsnetz and Deutsche Forschungsgemeinschaft.

AUTHOR CONTRIBUTIONSM.M. carried out the electrophysiology experiments. B.H. performed thebehavioral experiments. D.E. generated the mutant mice. J.R.P. bred the mice forthe behavioral experiments and injected them with tamoxifen. C.L. designed thestudy with M.M. and C.C. and wrote the manuscript with the help of M.M.and R.S.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Kalivas, P.W. Glutamate systems in cocaine addiction. Curr. Opin. Pharmacol. 4, 23–29(2004).

2. Kauer, J.A. & Malenka, R.C. Synaptic plasticity and addiction. Nat. Rev. Neurosci. 8,844–858 (2007).

3. Thomas, M.J., Kalivas, P.W. & Shaham, Y. Neuroplasticity in the mesolimbic dopaminesystem and cocaine addiction. Br. J. Pharmacol. 154, 327–342 (2008).

4. Ungless, M.A., Whistler, J.L., Malenka, R.C. & Bonci, A. Single cocaine exposure in vivoinduces long-term potentiation in dopamine neurons. Nature 411, 583–587 (2001).

5. Bellone, C. & Luscher, C. Cocaine triggered AMPA receptor redistribution is reversedin vivo by mGluR-dependent long-term depression. Nat. Neurosci. 9, 636–641 (2006).

6. Argilli, E., Sibley, D.R., Malenka, R.C., England, P.M. & Bonci, A. Mechanism and timecourse of cocaine-induced long-term potentiation in the ventral tegmental area.J. Neurosci. 28, 9092–9100 (2008).

7. Saal, D., Dong, Y., Bonci, A. & Malenka, R.C. Drugs of abuse and stress trigger a commonsynaptic adaptation in dopamine neurons. Neuron 37, 577–582 (2003).

8. Borgland, S.L., Malenka, R.C. & Bonci, A. Acute and chronic cocaine-induced potentia-tion of synaptic strength in the ventral tegmental area: electrophysiological andbehavioral correlates in individual rats. J. Neurosci. 24, 7482–7490 (2004).

9. Chen, B.T. et al. Cocaine but not natural reward self-administration nor passive cocaineinfusion produces persistent LTP in the VTA. Neuron 59, 288–297 (2008).

10. Kourrich, S., Rothwell, P.E., Klug, J.R. & Thomas, M.J. Cocaine experience controlsbidirectional synaptic plasticity in the nucleus accumbens. J. Neurosci. 27, 7921–7928(2007).

11. Conrad, K.L. et al. Formation of accumbens GluR2-lacking AMPA receptors mediatesincubation of cocaine craving. Nature 454, 118–121 (2008).

12. Churchill, L., Swanson, C.J., Urbina, M. & Kalivas, P.W. Repeated cocaine altersglutamate receptor subunit levels in the nucleus accumbens and ventral tegmentalarea of rats that develop behavioral sensitization. J. Neurochem. 72, 2397–2403(1999).

13. Boudreau, A.C. & Wolf, M.E. Behavioral sensitization to cocaine is associated withincreased AMPA receptor surface expression in the nucleus accumbens. J. Neurosci. 25,9144–9151 (2005).

14. Grimm, J.W., Hope, B.T., Wise, R.A. & Shaham, Y. Neuroadaptation. Incubation ofcocaine craving after withdrawal. Nature 412, 141–142 (2001).

15. Ping, A., Xi, J., Prasad, B.M., Wang, M.H. & Kruzich, P.J. Contributions of nucleusaccumbens core and shell GluR1 containing AMPA receptors in AMPA- and cocaine-primed reinstatement of cocaine-seeking behavior. Brain Res. 1215, 173–182 (2008).

16. Kessels, H.W. & Malinow, R. Synaptic AMPA receptor plasticity and behavior. Neuron61, 340–350 (2009).

17. Mameli, M., Balland, B., Lujan, R. & Luscher, C. Rapid synthesis and synaptic insertionof GluR2 for mGluR-LTD in the ventral tegmental area. Science 317, 530–533 (2007).

18. Mao, L. et al. The scaffold protein Homer1b/c links metabotropic glutamate receptor 5 toextracellular signal–regulated protein kinase cascades in neurons. J. Neurosci. 25,2741–2752 (2005).

19. Ronesi, J.A. & Huber, K.M. Homer interactions are necessary for metabotropic gluta-mate receptor–induced long-term depression and translational activation. J. Neurosci.28, 543–547 (2008).

20. Moroni, F. et al. Pharmacological characterization of 1-aminoindan-1,5-dicarboxylicacid, a potent mGluR1 antagonist. J. Pharmacol. Exp. Ther. 281, 721–729 (1997).

21. Thomas, M.J., Beurrier, C., Bonci, A. & Malenka, R.C. Long-term depression in thenucleus accumbens: a neural correlate of behavioral sensitization to cocaine. Nat.Neurosci. 4, 1217–1223 (2001).

22. Engblom, D. et al. Glutamate receptors on dopamine neurons control the persistence ofcocaine seeking. Neuron 59, 497–508 (2008).

23. Anderson, S.M. et al. CaMKII: a biochemical bridge linking accumbens dopamineand glutamate systems in cocaine seeking. Nat. Neurosci. 11, 344–353 (2008).

24. Haber, S.N., Fudge, J.L. & McFarland, N.R. Striatonigrostriatal pathways in primatesform an ascending spiral from the shell to the dorsolateral striatum. J. Neurosci. 20,2369–2382 (2000).

1040 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

25. Ikemoto, S. Dopamine reward circuitry: two projection systems from the ventralmidbrain to the nucleus accumbens-olfactory tubercle complex. Brain Res. Rev. 56,27–78 (2007).

26. Everitt, B.J. & Robbins, T.W. Neural systems of reinforcement for drug addiction: fromactions to habits to compulsion. Nat. Neurosci. 8, 1481–1489 (2005).

27. Luscher, C. & Bellone, C. Cocaine-evoked synaptic plasticity: a key to addiction? Nat.Neurosci. 11, 737–738 (2008).

28. Zweifel, L.S., Argilli, E., Bonci, A. & Plamiter, R.D. Role of NMDA receptors indopamine neurons for plasticity and addictive behaviors. Neuron 59, 486–496(2008).

29. Vezina, P. & Queen, A.L. Induction of locomotor sensitization by amphetamine requiresthe activation of NMDA receptors in the rat ventral tegmental area. Psychopharmacology(Berl.) 151, 184–191 (2000).

30. Harris, G.C., Wimmer, M., Byrne, R. & Aston-Jones, G. Glutamate-associated plasticityin the ventral tegmental area is necessary for conditioning environmental stimuli withmorphine. Neuroscience 129, 841–847 (2004).

31. Dong, Y. et al. Cocaine-induced potentiation of synaptic strength in dopamine neurons:behavioral correlates in GluRA�/� mice. Proc. Natl. Acad. Sci. USA 101, 14282–14287 (2004).

32. Mead, A.N., Brown, G., Le Merrer, J. & Stephens, D.N. Effects of deletion of gria1 orgria2 genes encoding glutamatergic AMPA-receptor subunits on place preferenceconditioning in mice. Psychopharmacology (Berl.) 179, 164–171 (2005).

33. Carlezon, W.A. Jr & Nestler, E.J. Elevated levels of GluR1 in the midbrain: a trigger forsensitization to drugs of abuse? Trends Neurosci. 25, 610–615 (2002).

34. Lammel, S. et al. Unique properties of mesoprefrontal neurons within a dualmesocorticolimbic dopamine system. Neuron 57, 760–773 (2008).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1041

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSAnimals. NR1DAT-CreERT2 mice were generated by crossing mice carrying an

inducible Cre recombinase under the Slc6a3 (Dat) promoter with mice carrying

floxed alleles for NR1.1. The mutation was induced in 5-month-old mice by

repeatedly intraperitoneally injecting 1 mg per kg of tamoxifen twice a day for

5 d. To ensure downregulation of NMDARs and a sufficient wash out period

after tamoxifen treatment, we used two control and NR1DAT-CreERT2 mice that

were between 7 and 8 months old when the behavioral experiments started.

They weighed 29.8 ± 1.4 g and 31.1 ± 0.8 g, respectively, on the day of surgery.

Mice were single housed under constant temperature (21 ± 2 1C) and humidity

conditions (50 ± 5%). All experiments took place during the light phase of the

dark-light cycle, between 07:00 and 19:00. On three occasions during the food

training phase, mice received limited access to food. They were otherwise given

unrestricted access to food and water throughout the whole set of experiments.

The mice’s weight was monitored daily and kept above 85% of their initial

weight. The experiments were conducted in accordance with the ethical European

Union guidelines for the care and use of laboratory animals and were approved

by the Committee on Animal Care and Use (Regierungsprasidium Karlsruhe) and

the Institutional Animal Care and Use Committee of the University of Geneva.

Drug treatment. C57BL/6 mice and Pitx3–green fluorescent protein (GFP)

mice35 (postnatal day 16–35) were injected intraperitoneally with 15 mg per kg

cocaine, 0.9% saline, 0.25 mg per kg AIDA or 4 mg per kg Ro 67-7476 (injected

1 h prior to saline or cocaine injections) using a 26 gauge hypodermic needle

(injection volume of 50–100 ml) to minimize stress. Unless specified, drugs were

obtained from Tocris, spermine and picrotoxin were obtained from Sigma, and

cocaine was obtained from the pharmacy of the Hopitaux Universitaires de

Geneve. Ro 67-7476 was a gift from F. Knoflach (F. Hoffman–La Roche). For

the behavioral experiments, cocaine hydrochloride (Sigma-Aldrich, Chemie

GmbH) was dissolved in saline. Ketamine and xylazine solutions were obtained

from Pharmanovo GmbH. All solutions injected intravenously were first

filtered through sterile filters (0.2 mm).

Stereotactic TAT-fused peptide delivery. Mice were anesthetized with keta-

mine (100 mg per kg) and xylazine (10 mg per kg) and placed in a stereotactic

frame (myNeuroLab). The TAT sequence (YGRKKRRQRRR) was fused

with N-carboxyfluorescein and a homologous sequence of the carboxy-

terminal of Gp I mGluRs to prevent Homer 1b/c binding (ALTPPSPFR,

Primm). As a control, the TAT sequence was fused only with N-carboxyfluor-

escein. TAT-fused peptides were injected (0.6 ml at the concentration of 1 mM)

with a glass pipette (Drummond Scientific Company) bilaterally (�2.4 mm

antero-posterior and 0.8 mm lateral from Bregma, and �4.4 mm from

the surface).

Electrophysiology in acute brain slices. Horizontal slices from midbrain

(250 mm thick) and coronal slices containing NAc (300 mm thick) were

prepared following the experimental injections protocols described in the text.

Slices were kept in artificial cerebrospinal fluid containing 119 mM NaCl,

2.5 mM KCl, 1.3 mM MgCl2, 2.5 mM CaCl2, 1.0 mM NaH2PO4, 26.2 mM

NaHCO3 and 11 mM glucose, bubbled with 95% O2 and 5% CO2. Whole-cell

voltage-clamp recording techniques were used (30–32 1C, 2–3 ml min–1,

submerged slices) to measure the holding currents and synaptic responses of

DA neurons of the VTA and of medium spiny neurons of the NAc shell. The

VTA is defined as the region medial to the MT (medial terminal nucleus of the

accessory optical tract). DA neurons were identified either by the presence of a

large hyperpolarization-activated (Ih) current immediately after obtaining a

whole-cell configuration (80%) or by using slices form a Pitx3-GFP mouse

strain (20%) that express GFP only in cells that express Pitx3, a transcription

factor required for the development of DA neurons of the midbrain. The

internal solution contained 130 mM CsCl, 4 mM NaCl, 2 mM MgCl2, 1.1 mM

EGTA, 5 mM HEPES, 2 mM Na2ATP, 5 mM sodium creatine phosphate,

0.6 mM Na3GTP and 0.1 mM spermine. Currents were amplified, filtered at

5 kHz and digitized at 20 kHz. The liquid junction potential was small (–3 mV)

and traces were therefore not corrected. All experiments were carried out in the

presence of picrotoxin (100 mM) and AMPAR EPSCs were pharmacologically

isolated by application of the NMDA antagonist D,L(–)-2-amino-5-phospho-

novaleric acid (D,L-AP5, 100 mM). The NMDAR component was calculated as

the difference between the EPSCs measured in the absence and presence of

D,L-AP5. The AMPAR to NMDAR ratio was calculated by dividing the peak

amplitudes. The rectification index was calculated by dividing the amplitude

of the AMPAR EPSC measured at �70 mV by the amplitude at +40 mV.

The holding potential was �60 mV or �70 mV, and the access resistance

was monitored by a hyperpolarizing step of �10 mV with each sweep, every

10 s. Experiments were discarded if the access resistance varied by more

than 20%. Synaptic currents were evoked by stimuli (0.05–0.1 ms) at 0.1 Hz

through bipolar stainless steel electrodes placed rostral to the VTA or, when

recordings were performed in the NAc, at the prelimbic cortex�NAc border to

stimulate preferentially cortical afferences. Where indicated, mGluR LTD in DA

neurons of the VTA was induced by application of the Gp I mGluR agonist

DHPG (20 mM for 5 min.), whereas NMDA-dependent LTD in medium spiny

neurons of the NAc shell was induced by LFS (1 Hz at �40 mV for 10 min).

Representative example traces are shown as the average of 10–20 consecutive

EPSCs typically obtained at each potential or, in the case of plasticity protocols,

during the last 5 min of the baseline and at least 20 min after the induction

of plasticity.

Mouse behavior. All operant experiments were performed in mouse operant

chambers model ENV-307W enclosed in light- and sound-attenuating cubicles

(Med-Associates). Each chamber was equipped with two ultrasensitive retract-

able levers located on each side of a food pellet dispenser during the food-

shaping procedure. During intravenous self-administration, the drug delivery

PVC tubing was attached to a swivel (Instech Solomon) and connected to an

infusion pump (PHM-100, Med-Associates) located outside the cubicle.

Stimulus lights were located above each lever.

Food training. Sessions lasted 90 min and started with the presentation of the

two levers. To guarantee unbiased lever training, the side of the initially active

lever was alternated between each session. Active lever presses were reinforced

by the delivery of a 16-mg sweetened pellet (Bio-Serv) under the following

schedule: fixed ratio 1 (FR1) for eight reinforcements, FR2 for four reinforce-

ments and FR4 for at least ten reinforcements with 80% accuracy on the active

lever, which completed a cycle.

When such a cycle was achieved, the side of the active lever was switched. All

mice underwent 16 daily food-training sessions. On three occasions, sessions

were separated by a day of inactivity, during which mice had limited access to

food. This happened between sessions 2 and 3, 9 and 10, and between sessions

14 and 15.

Intravenous self-administration. Following food shaping, mice were allowed

24 h before undergoing surgery. Mice were anesthetized with a ketamine

(160 mg per kg) and xylazine (38 mg per kg) solutions and implanted with a

catheter in the right jugular vein. On one end catheters were introduced for

circa 1 cm toward the heart, while the other extremity was passed subcuta-

neously to an exit in the mid-scapular region. Catheters were (MIVSA,

CamCaths) made of silicone elastomer tubing (outside diameter 0.63 mm �inside diameter 0.30 mm) attached to 26 gauge stainless steel tubing secured

to a Bard mesh pad. Mice were given a minimum of 48 h recovery before

intravenous self-administration sessions were initiated. Catheters were flushed

daily before and after self-administration sessions with 30 ml of a heparine

solution in saline (20 international units). At the beginning of the cocaine self-

administration sessions, mice were placed into the operant chambers and the two

levers were presented. Presses on the active lever under FR1 resulted in the

infusion of 0.5 mg per kg per infusion of cocaine solution by activation of the

infusion pump for 1.2 s. The side of the active lever was randomly assigned across

mice on the first day of intravenous self-administration and remained constant

thereafter. Each drug infusion was associated with a conditioned stimulus

consisting of a 10-s flashing of stimulus lights located above the levers. To avoid

accidental overdose, we gave the mice a 40-s time-out period following each

infusion, during which active levers were recorded, but had no consequence.

Each session lasted 4 h and mice received a total of eight sessions.

Incubation effect and cue-induced cocaine-seeking. At the end of the

intravenous self-administration phase, mice were returned to their home cage,

where they remained for 35 d. After this time, they were re-introduced to the

operant chambers, where they were tested for cue-induced cocaine-seeking. The

NATURE NEUROSCIENCE doi:10.1038/nn.2367

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

session lasted 90 min, during which, as in the self-administration condition,

presses on the active lever under an FR1 triggered conditioned stimulus

presentation and activation of the infusion pump. However, cocaine was no

longer available and conditioned stimulus presentation lasted for 5 s only. Such

a short conditioned stimulus presentation was preferred to prevent as much as

possible extinction toward its repeated exposure.

Statistical analysis. Compiled data are expressed as mean ± s.e.m. The level

of significance was taken at P ¼ 0.05, as determined by the non-parametric

Mann-Whitney or Wilcoxon tests. Data from food training and cocaine self-

administration experiments were analyzed using repeated measures ANOVA

to investigate the effect of day, genotype and lever when appropriate. Data from

cue-induced cocaine seeking were analyzed using factorial ANOVA to investi-

gate the effect of genotype and lever. Analyses were followed by Newman-Keuls

tests when necessary.

35. Zhao, S. et al. Generation of embryonic stem cells and transgenic mice expressinggreen fluorescence protein in midbrain dopaminergic neurons. Eur. J. Neurosci. 19,1133–1140 (2004).

doi:10.1038/nn.2367 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Synaptic inhibition of Purkinje cells mediatesconsolidation of vestibulo-cerebellar motor learning

Peer Wulff1,7, Martijn Schonewille2,7, Massimiliano Renzi3, Laura Viltono4, Marco Sassoe-Pognetto4,Aleksandra Badura2, Zhenyu Gao2, Freek E Hoebeek2, Stijn van Dorp5, William Wisden1,6, Mark Farrant3 &Chris I De Zeeuw2,5

Although feedforward inhibition onto Purkinje cells was first documented 40 years ago, we understand little of how inhibitory

interneurons contribute to cerebellar function in behaving animals. Using a mouse line (PC-Dc2) in which GABAA receptor–

mediated synaptic inhibition is selectively removed from Purkinje cells, we examined how feedforward inhibition from molecular

layer interneurons regulates adaptation of the vestibulo-ocular reflex. Although impairment of baseline motor performance was

relatively mild, the ability to adapt the phase of the vestibulo-ocular reflex and to consolidate gain adaptations was strongly

compromised. Purkinje cells showed abnormal patterns of simple spikes, both during and in the absence of evoked compensatory

eye movements. On the basis of modeling our experimental data, we propose that feedforward inhibition, by controlling the fine-

scale patterns of Purkinje cell activity, enables the induction of plasticity in neurons of the cerebellar and vestibular nuclei.

Feedforward inhibitory microcircuits, in which interneurons and theirtarget principal cells receive common excitatory input, enhancenetwork performance in many brain regions1,2. In the hippocampus,feedforward inhibition, by reducing the time window of synapticintegration, increases the precision of spike timing in CA1 pyramidalneurons3, and plasticity of feedforward inhibition is required tomaintain the fidelity of information processing4. In the cerebellum,molecular layer interneurons (stellate and basket cells) control Purkinjecells by powerful feedforward inhibition5–9 (Supplementary Fig. 1). Inaddition, subsets of Purkinje cells sparsely inhibit each other via axoncollaterals10. Purkinje cells provide the only output of the cerebellarcortex and project to the cerebellar and vestibular nuclei. They firecomplex spikes in response to climbing fiber activity11 and simplespikes that reflect the integration of intrinsic pacemaker activitywith excitatory and inhibitory synaptic inputs from parallel fibersand molecular layer interneurons8,12–15.

Although feedforward inhibition onto Purkinje cells was documen-ted more than four decades ago5, we still know little about how itcontributes to cerebellar function in behaving animals. Fast synapticinhibition at molecular layer interneuron to Purkinje cell synapses ismediated by a1b2/3g2-type GABAA receptors16. The g2 subunit isrequired to target the receptors to the postsynaptic membrane17. Toinvestigate the role of GABAA receptor–mediated feedforward inhibi-tion, we selectively ablated the g2 subunit, and thereby synaptic GABAA

receptors, from Purkinje cells (PC-Dg2 mice). The resulting changesthat we observed in Purkinje cell simple-spike activity and motor

behavior indicate that molecular layer interneurons are essentialregulators of cerebellar signal coding and memory formation.

RESULTS

Purkinje cell–specific removal of synaptic GABAA receptors

To remove GABAA receptor–mediated feedforward inhibition ontoPurkinje cells, we selectively deleted the GABAA receptor g2 subunitusing the Cre/loxP system (Online Methods). Cre recombinase, underthe control of the L7 promoter, induced a Purkinje cell–specific deletionof the loxP-flanked Gabrg2 (g2 subunit) gene in the second postnatalweek16,18. Ablation of synaptic GABAA receptors from Purkinje cellscaused no anatomical alterations of the cerebellar circuitry (Fig. 1).

Patch-clamp recordings in acute slices of cerebellar vermis fromadult mice showed spontaneous fast inhibitory postsynaptic currents(sIPSCs) at high frequency in all Purkinje cells (n ¼ 21) from controlmice (Fig. 2a), which could be blocked by the GABAA receptorantagonist SR-95531 (20 mM, data not shown). In contrast, sIPSCswere absent from all Purkinje cells (n ¼ 19) of PC-Dg2 mice (Fig. 2b).In some PC-Dg2 cells (12 of 19), small, slow-rising currents remained.However, these produced on average less than 2% of the controlsynaptic charge (Fig. 2) and probably reflect spillover of synapticallyreleased GABA onto extrasynaptic a and b subunit–containing recep-tors19,20 (Supplementary Fig. 2). Consistent with a complete loss ofsynaptic GABAA receptors, recordings from PC-Dg2 mice in thepresence of tetrodotoxin (TTX) confirmed the absence of miniatureIPSCs (mIPSCs; Fig. 2c,d). The loss of synaptic GABAA receptors was

Received 4 November 2008; accepted 12 May 2009; published online 5 July 2009; doi:10.1038/nn.2348

1Institute of Medical Sciences, Foresterhill, University of Aberdeen, Aberdeen, UK. 2Department of Neuroscience, Erasmus MC, Rotterdam, The Netherlands. 3Departmentof Neuroscience, Physiology and Pharmacology, University College London, London, UK. 4Department of Anatomy, Pharmacology and Forensic Medicine, University of Turinand National Institute of Neuroscience-Italy, Turin, Italy. 5Netherlands Institute for Neuroscience, Royal Academy of Sciences, Amsterdam, The Netherlands. 6Division ofCell and Molecular Biology, Imperial College, London, UK. 7These authors contributed equally to this work. Correspondence should be addressed to C.I.D.Z.(c.dezeeuw@erasmusmc.nl) or M.F. (m.farrant@ucl.ac.uk) or W.W. (w.wisden@imperial.ac.uk).

1042 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

restricted to Purkinje cells; mIPSCs in molecular layer interneuronswere unaltered in PC-Dg2 mice (Supplementary Fig. 3).

PC-Dc2 mice show altered simple spike patterning

Feedforward inhibition via molecular layer interneurons is rapidly(B1 ms) recruited by parallel fiber activation and curtails the parallel

fiber–evoked excitatory postsynaptic potential (EPSP) in Purkinjecells7,21. To determine how the absence of synaptic GABAA receptorsaffected Purkinje cell responses to parallel fiber stimulation, weanalyzed the temporal dispersion (jitter) of evoked Purkinje cell simplespikes (Fig. 3a). The jitter, quantified as the s.d. of spike latency in a10-ms window following stimulation (10 V, 100 ms), was strongly

Figure 1 PC-Dg2 mice show normal cerebellar

morphology and synaptic organization. (a,b) Nissl

stains of sections through vermis (sagittal) and

flocculus (coronal) revealed no differences

between control (a) and PC-Dg2 (b) mice, and the

number of Purkinje cells (24.5 ± 2.0 versus

23.9 ± 2.5 cells per 1,000 mm, P ¼ 0.75) and

molecular layer interneurons (2.36 ± 0.19 versus2.28 ± 0.18 cells per 1,000 mm2, P ¼ 0.53)

were similar in both groups. C, cochlear nucleus;

Cb1–10, lobules 1–10; Gr, granule cell layer;

Mol, molecular layer; PC, Purkinje cell layer.

(c–e) Immunofluorescence labeling in the

flocculus showed no differences in the

distribution of GABAergic terminals (vesicular

g-aminobutyric acid transporter, VGAT, c),

climbing fiber terminals (vesicular glutamate

transporter 2, VGLUT2, d) and parallel fiber

terminals (VGLUT1, e). Quantification of puncta

per 1,000 mm2 revealed no difference (P ¼ 0.27,

0.62 and 0.68, respectively, n ¼ 4).

(f–h) Electron microscopy showed no obvious

morphological changes in parallel and climbing

fiber synapses. (f) Asymmetric synapses between parallel fibers and Purkinje cell spines (asterisks). The density of parallel fiber to Purkinje cell synapses was

unchanged (33.0 versus 32.9 synapses per100 mm2 in PC-Dg2 and control, see Online Methods). (g) Asymmetric synapses made by climbing fibers (CF).

(h) Symmetric synapses (arrowheads) made by basket cells (BC) onto the cell body of Purkinje cells. Scale bars represent 450 mm (a) and 150 mm for upper

and lower panels, respectively (b), 20 mm (c,d), 5 mm (e), 500 nm (f), 360 nm (g), and 440 nm (h).

Control PC-∆γ2

CF

* *

*

**

*

PC

BC

PC

BC

CF

*

*

*

*

* *

*

MolPC

Gr

C

Gr

MolPC

C

Cb2

Cb3

Cb4&5 Cb6a

b

f

g

h

Cb7

Cb8

Cb9Cb10Cb1

Control PC-∆γ2Control

PC-∆γ2

CalbVGLUT2

CalbVGAT

VGLUT1

d

e

c

Control

0.5 s

500 pA

0.5 s

100 pA

PC-∆γ2

Control + TTX PC-∆γ2 + TTX

160

120

80

40

05004003002001000

Charge = 25.9 pC

800

600

400

200

05004003002001000

Charge = 2.2 pC

1,600

1,200

800

400

0100806040200

Charge = 1.2 pC

2,000

1,500

1,000

500

0100806040200

Charge = 0 pC

Current (–pA)

Current (–pA)

Sam

ples

Sam

ples

a

200 pA100 pA

b

dc

Current (–pA)

Current (–pA)

Sam

ples

Sam

ples

Figure 2 Loss of fast synaptic inhibition from

Purkinje cells in PC-Dg2 mice. (a) Representative

contiguous segments of a whole-cell recording

(�70 mV) from a Purkinje cell of a control mouse.

Ionotropic glutamate receptors were blocked with

6-cyano-7-nitroquinoxaline-2,3-dione (CNQX) and

D(-)-2-amino-5-phosphonopentanoic acid (D-AP5).

Bottom, quantification of mean synaptic charge in

a different Purkinje cell, with a 2.5-s recording of

sIPSCs and corresponding all-point amplitude

histogram. The left-hand peak (most-positive

current values), corresponding to the baseline

current noise, is fitted with a single-sided Gaussian

(white). The peak of the histogram is taken as

the zero current value (dotted line in inset). The

filled gray area corresponds to all sample points

other than those in the baseline noise and thusrepresents the current produced by phasic synaptic

events. In this cell, the mean synaptic charge was

25.9 pC. (b) Corresponding data from two PC-Dg2mice. sIPSCs were seen in all cells from control

mice, but were not observed in cells from PC-Dg2mice. Slow SR-95531–sensitive currents were seen

in B60% of PC-Dg2 cells. For the cell shown in

the lower panel, the phasic charge transfer was

2.2 pC. On average, the charge transfer was

reduced from 59.8 ± 18.4 pC in control (n ¼ 8)

to 1.0 ± 0.5 pC in PC-Dg2 cells (n ¼ 15, P o0.0002, Mann-Whitney U test). (c,d) Corresponding

data recorded in the presence of TTX. Note the

different scaling of the current record and the

abscissa of the all-point histogram and the

complete absence of mIPSCs in PC-Dg2 cells.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1043

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

increased in PC-Dg2 Purkinje cells (control, 0.81 ± 0.14 ms, n ¼ 12;PC-Dg2, 1.80 ± 0.10 ms, n ¼ 11; P o 0.0001). Acute blockade ofGABAA receptors with SR-95531 significantly increased spike jitter incells from control mice (to 1.45 ± 0.14 ms, P ¼ 0.0011; see also ref. 7),but, as expected, had no effect in PC-Dg2 cells (1.76 ± 0.10 ms,P¼ 0.605). We also determined the number of spikes evoked by parallelfiber stimulation (Fig. 3a; see Online Methods). On average, 0.60 ±0.04 spikes were evoked in the 60 ms following each stimulus in controlcells and 0.41 ± 0.05 spikes in PC-Dg2 cells (n¼ 17 and 13, respectively;P ¼ 0.0069). This smaller evoked response is consistent with a reducedparallel fiber excitatory input (see Supplementary Fig. 4 and Discus-sion). Consistent with the complete loss of GABAA receptor–mediatedinhibition in PC-Dg2 cells, SR-95531 increased the number of evokedspikes only in control cells (control, 0.61 ± 0.05 to 0.76 ± 0.08, n ¼ 11,P ¼ 0.0248; PC-Dg2, 0.41 ± 0.05 to 0.45 ± 0.06, n ¼ 13, P ¼ 0.3199).Thus, a loss of molecular layer interneuron–mediated feedforwardinhibition in PC-Dg2 mice results in altered simple-spike responses toparallel fiber inputs.

Purkinje cells in cerebellar slices from PC-Dg2 mice showed an increase in simple-spikefiring regularity compared with controls(Fig. 3b). The mean firing rate at 23–26 1Cwas not different between groups (12.3 ± 1.6(control) versus 13.7 ± 0.6 Hz (PC-Dg2),n ¼ 26 and 9, P ¼ 0.062, Mann-Whitney

U test), but the coefficient of variation (s.d. divided by the mean)of the interspike interval (ISI) was reduced in PC-Dg2 mice (0.20 ±0.03 in control versus 0.10 ± 0.01 in PC-Dg2; P ¼ 0.018, Mann-Whitney U test). The coefficient of variation of adjacent intervals(mean value of 2� ISIn+1 � ISInjj

ðISIn+1 + ISInÞ, a measure for the regularity of firing onsmall timescales22) also differed. The coefficient of variation of adjacentintervals was 0.19 ± 0.02 in control versus 0.10 ± 0.01 in PC-Dg2 mice(P ¼ 0.018, Mann-Whitney U test). Blockade of GABAA receptors withSR-95531 in control Purkinje cells decreased the coefficient of variationof the ISI (0.20 ± 0.04 in control versus 0.13 ± 0.02 in SR-95531 treated,P ¼ 0.024, n ¼ 8) to a value comparable to that found in PC-Dg2 mice(see also refs. 12,15,23). As expected, SR-95531 failed to alter thecoefficient of variation of the ISI in cells from PC-Dg2 mice (0.13 ± 0.02versus 0.13 ± 0.04, n ¼ 3). Notably, similar results were obtained atnear-physiological temperature (34–35 1C), with no change in the meanrate (51.3 ± 9.1 in control versus 50.0 ± 3.5 Hz in PC-Dg2,n ¼ 9 and 7, P ¼ 0.61, Mann-Whitney U test), but a significant decreasein the coefficient of variation (0.14 ± 0.01 in control versus 0.06 ± 0.01 in

80

60

40

20

0

Spi

kes

per

bin

2

1

0Late

ncy

SD

(m

s)

+SR-95531

*2

1

0+SR-95531La

tenc

y S

D (

ms) **

80

60

40

20

0

Spi

kes

per

bin

Controla PC-∆γ2

1.0

0.8

0.6

0.4

0.2

0

6040200–20Time (ms)Time (ms)

6040200–20

∆ S

pike

s pe

r st

im +SR-955311.0

0.8

0.6

0.4

0.2

0

∆ S

pike

s pe

r st

im

500 ms

75

50

25

0

0

20

40

Frequency (Hz)

0

0.05

0.10

0.15

0.20

0.25

Control

Control

Mean ISI = 86.1 msCV = 0.23

b

Mean ISI = 83.6 msCV = 0.12

Num

ber ***

80

60

40

20

0

Num

ber

200150100500ISI (ms)

***

PC-∆γ2

ISI CV2ISI CV

PC-∆γ2

c d

Control EP

SC

am

plitu

de (

%)

80

90

100

110

120

130

140

150

Time (min)–10 –5 0 5 10 15 20 25 30 35

50

60

70

80

90

100

110

120

PC-∆γ2

EP

SC

am

plitu

de (

%)

Time (min)–10 –5 0 5 10 15 20 25 30 35

Figure 3 PC-Dg2 mice show altered parallel

fiber–evoked and spontaneous simple-spike firing

in vitro and unaltered parallel fiber–Purkinje cell

LTP and LTD. (a) Simple spikes evoked by parallel

fiber activation during cell-attached recording. Top

and middle panels show raster plots (400 sweeps

at 0.5 Hz) and corresponding peristimulus-time

histograms (PSTHs, 0.5-ms bin width). Arrowsand dashed lines denote stimulation. Insets show

s.d. of spike latency in a 10-ms window (red bars,

12 control and 11 PC-Dg2 cells). Error bars

denote s.e.m. Jitter was greater in PC-Dg2 (red)

than in control (blue) cells (*P o 0.0001). SR-

95531 increased jitter in control cells (**P ¼0.0011), but not in PC-Dg2 cells (P ¼ 0.605).

Bottom, global averages of baseline-corrected

cumulative spike probability (see Online

Methods). Shaded areas denote s.e.m. (17 control

and 13 PC-Dg2 cells). Stimulation evoked fewer

spikes in PC-Dg2 cells (averaged between 0 and

60 ms, P ¼ 0.0069). SR-95531 (40 mM, black

line, gray shading) increased spikes in control

cells (n ¼ 11, P ¼ 0.0248), but not in PC-Dg2cells (n ¼ 13, P ¼ 0.3199). (b) Representative

simple spikes (23–26 1C) and corresponding ISI

histograms. Right panels show pooled data (26

control and 9 PC-Dg2 cells). Although the meanfiring rates were not significantly different (P 40.05), the coefficient of variation (CV) and the

coefficient of variation of adjacent intervals (CV2)

of ISIs differed significantly (***P o 0.05) (see

text for details). (c) Pooled data showing parallel

fiber-Purkinje cell LTP in control (n ¼ 7, black)

and PC-Dg2 cells (n ¼ 4, gray); EPSC amplitude

was similarly increased in the strains (both

P o 0.005; control versus PC-Dg2, P ¼ 0.257).

(d) Parallel fiber–Purkinje cell LTD was similar in

PC-Dg2 (n ¼ 5) and control (n ¼ 4) cells (both

P o 0.05; control versus PC-Dg2, P ¼ 0.624).

1044 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

PC-Dg2, P ¼ 0.001; Mann-Whitney U test) and the coefficient of vari-ation of adjacent intervals (0.15 ± 0.02 versus 0.06 ± 0.01, P ¼ 0.0099).

Finally, we examined whether a loss of inhibition onto Purkinje cellsmodified long-term plasticity at parallel fiber to Purkinje cell synapses.Neither parallel fiber long-term depression (LTD) nor long-termpotentiation (LTP) (see Online Methods) were significantly impairedin PC-Dg2 mice compared with controls (P ¼ 0.624 and P ¼ 0.257,respectively; Fig. 3c,d).

Learning and consolidation deficits in PC-Dc2 mice

PC-Dg2 mice show no obvious neurological abnormality16. To assesscerebellar performance, we analyzed compensatory eye movements inmale PC-Dg2 mice (n ¼ 9) and littermate controls (n ¼ 8). Mice wereexposed to whole-field visual stimuli to determine the amplitude (gain)and timing (phase) of their optokinetic reflex (OKR) and/or testedwith turntable stimulation to investigate the same parameters forthe vestibulo-ocular reflex in the dark (VOR) and light (visual VOR).PC-Dg2 mice showed a relatively small, but significant, deficit in theirOKR, evident as a reduction in gain and a lag in phase compared withcontrols (P ¼ 0.018 and P ¼ 0.012, respectively, two-way repeated-measures ANOVA; Supplementary Fig. 5). The VOR gain values andphase leads of PC-Dg2 mice were larger and smaller, respectively, thanthose of controls (P ¼ 0.012 and P ¼ 0.030, two-way repeated-measures ANOVA; Supplementary Fig. 5). In contrast, no significantdifferences were observed during visual VOR (P ¼ 0.43 and P ¼ 0.63for gain and phase values, respectively; Supplementary Fig. 5). Thus,PC-Dg2 mice show small, but significant, abnormalities in motorperformance when visual and vestibular systems are investigated

separately, but not when they operate together, as under naturalconditions or during visuo-vestibular training.

Loss of inhibition onto Purkinje cells had more profound effects oncerebellar motor learning. We studied gain and phase learning byapplying a protocol aimed at reducing the gain of the VOR on day 1(five 10-min sinusoidal, in phase drum and table rotations at 0.6 Hz,both with an amplitude of 51) and subsequently shifting its phase ondays 2, 3 and 4 (five 10-min sinusoidal, in phase drum and tablerotations at 0.6 Hz, but with drum amplitudes of 7.51 on day 2 and 101on days 3 and 4, while the table amplitude remained 51). Mice werekept in the dark between the recording days.

Gain-decrease learning of PC-Dg2 (n ¼ 9) and control mice (n¼ 10)on day 1 was similar (P ¼ 0.11, two-way repeated-measures ANOVA;Fig. 4a,b). However, when the measurements were resumed the nextday, the degree of gain reduction carried forward from the previousday’s learning was significantly smaller in PC-Dg2 mice than in controls(P ¼ 0.001; Fig. 4b). This consolidation deficit was apparent at a widerange of frequencies (Fig. 4c). To exclude nonspecific effects (habitua-tion during gain-decrease learning), we tested PC-Dg2 and controlmice in nonadapting VOR protocols; notably, mice of both genotypesshowed no significant decreases in VOR over consecutive days (last gainvalue of session 1 versus first value of session 2, P ¼ 0.610 for controlsand 0.551 for PC-Dg2 mice; Supplementary Fig. 6).

Deficits in gain consolidation were also seen when the drum rotationamplitude was kept constant (three 10-min sinusoidal, in phase drumand table rotations at 0.6 Hz, both with an amplitude of 51; Fig. 4d).Here too, the initial level of learning was not significantly affected(PC-Dg2 mice versus controls, P ¼ 0.61; n ¼ 6 and 5, respectively),whereas the level of consolidation was significantly reduced (gain day 1- 2, P ¼ 0.034; gain day 2 - 3, P ¼ 0.046). Moreover, gainconsolidation deficits in PC-Dg2 mice did not depend on the directionof learning. With a gain-increase protocol (five 10-min sinusoidal, outof phase drum and table rotation at 1.0 Hz, both with an amplitude of1.61), no significant consolidation was present in the PC-Dg2 mice(gain day 1 - 2, P ¼ 0.744, n ¼ 6, one-Sample t test). In contrast,consolidation in control mice was present and was significantlystronger than in PC-Dg2 mice (P ¼ 0.002; Fig. 4d). Notably, thelevel of gain-increase learning in PC-Dg2 mice was not significantlydifferent from that in controls (P ¼ 0.800). Thus, deficits in consolida-tion of learned gain changes during both gain-decrease andgain-increase training procedures were not the results of differencesin baseline performance.

Day 1 and 2 gain decrease

0

0.2

0.4

0.6

0.8

1.0

0 10 20 30 40 50Time (min)

Nor

mal

ized

gai

n

ControlPC-∆γ2

PC-∆γ2

0 d

24 h

Con

solid

atio

n (%

)

b

0

60

120

180

Pha

se (

°)

Day 2 Day 3 Day 4

c

0

20

40

60

80

100

120

Gainday 1→2

Phaseday 2→3

d** **

0

0.2

0.4

0.6

0.8

1.0

1.00.4 0.6 0.8

0

30

60

90

150

Day 1

befo

re

Day 1

afte

r

24 h

afte

r

Frequency (Hz)

Day 2

befo

re

Day 2

afte

r

24 h

afte

r

Gai

nP

hase

(°)

Table 5°, drum 5°

Table 5°, drum 7.5°

Table 5°, drum 10°

Table 5°, drum 10°

Day 1

Day 2

Day 3

Day 4

a

*

Phaseday 3→4

Day 2, 3 and 4 phase reversal

120

1.00.4 0.6 0.8

Con

solid

atio

n (%

)

0

20

40

60

80

100

120

Gainday 1→2

Gainday 2→3

* *

Gainday 1→2

Control

2 s

*

Decrease Increase

Figure 4 Motor learning is severely affected in PC-Dg2 mice. (a) Illustrations

of table and drum rotations (left) during the training procedure and traces

(right) of sinusoidal table rotation (top trace) and of VOR eye movements after

the training (control, black traces; PC-Dg2, gray traces). Gain and phase

parameters were evaluated five times at 10-min intervals. (b) On day 1, PC-

Dg2 and control mice showed similar gain reduction (P ¼ 0.11), but the first

test on day 2 revealed clear differences (P ¼ 0.001) (top). During phase

reversal training, control mice learned better than PC-Dg2 mice (day 4, P o0.00001) (bottom). (c) Differences in gain consolidation and phase reversal

occurred over a wide range of frequencies. Day 1 before and Day 1 after

indicate values before and after training on day 1, and 24 h after indicates

the value on the next day, before the next training. (d) Top, differences in

consolidation (percentage change carried forward from the previous day) for

gain decrease (days 1 to 2) and phase reversal (days 2 to 3 and days 3 to 4).

Bottom, differences in gain consolidation were also seen with constant in

phase drum and table rotation (gain decrease, two histograms on the left) and

with constant out of phase drum and table rotation (gain increase, histogram

on the right). For the lower panel of d, data are from five control and six PC-

Dg2 mice. For all other panels, data are from ten control and nine PC-Dg2mice. Error bars denote s.e.m. * P o 0.05 and ** P o 0.01.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1045

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

The adaptation procedure provided on days 2, 3 and 4 immediatelyrevealed significant deficits in phase learning in PC-Dg2 mice, starting10–20 min after the initiation of visuo-vestibular training (for example,at 20 min, P ¼ 0.009; Fig. 4b). These deficits in phase changeacquisition were followed by clear differences in consolidation (forexample, from day 2 to day 3, P ¼ 0.0008; Fig. 4d). Phase adaptationdeficits also occurred at a wide range of frequencies (Fig. 4c) and werenot caused by visual problems in PC-Dg2 mice, as eye movementrecordings during the adaptation sessions showed that PC-Dg2 micewere capable of full phase reversal (Supplementary Fig. 7). In short,PC-Dg2 mice showed a relatively normal capacity for acquisitionduring gain-decrease and gain-increase motor learning, but a profounddeficit in acquisition during phase adaptation learning and a generaldeficit in consolidation of gain and phase adaptation.

Abnormal temporal patterns of Purkinje cell simple spikes

Because the flocculus controls the adaptation of compensatory eyemovements24–26 and Purkinje cells provide the sole output of thecerebellar cortex (Supplementary Fig. 1), we analyzed floccular Pur-kinje cell activity during optokinetic stimulation (Fig. 5a). Single unitsof Purkinje cells that responded optimally to stimulation around thevertical axis were identified by creating tuning curves of their complexspike responses and by identifying a clean climbing fiber pause26,27. Theaverage climbing fiber pause in PC-Dg2 mice and controls was 15.3 ±0.8 and 18.6 ± 1.3 ms, respectively (55 and 60 PC-Dg2 and controlcells, respectively, P¼ 0.029). The average simple-spike firing frequencyamplitude, phase relative to stimulus and modulation amplitude weresimilar in PC-Dg2 and control mice (Fig. 5b). However, as predicted byour in vitro recordings, the regularity of Purkinje cell firing was affected.

For floccular simple-spike activities, the coef-ficient of variation of the ISIs was significantlyreduced during visual stimulation in PC-Dg2mice (P ¼ 0.008; PC-Dg2, n ¼ 55; controls,

n ¼ 60; Fig. 5c). This difference reflected specific changes in temporalpatterning, as the coefficient of variation of adjacent intervals wassignificantly lower in PC-Dg2 mice (P o 0.0001; Fig. 5c, see alsoSupplementary Fig. 8).

If differences in Purkinje cell firing patterns contribute to consolida-tion deficits in PC-Dg2 mice, we would also expect to find them outsideof periods of optokinetic stimulation. Indeed, both the coefficient ofvariation and the coefficient of variation of adjacent intervals of ISIswere significantly reduced in the absence of stimulation (P ¼ 0.022 andP o 0.0001, respectively; PC-Dg2, n ¼ 41; controls, n ¼ 43; Fig. 5c). Incontrast, the patterns of complex spike activities of Purkinje cells didnot differ between PC-Dg2 and control mice (Fig. 5a,c, see alsoSupplementary Fig. 9). Also, the antiphasic modulation of complexand simple spikes was unchanged (Fig. 5a), arguing against a criticalinvolvement of molecular layer interneurons in this phenomenon8.

Model and simulations

We interpreted the experimental data from the 4-d gain-decrease, phaseadaptation routine using a ‘distributed memory’ model (Fig. 6). Short-term adaptation is assumed to take place in the cerebellar cortex and isexpressed as adaptation of the phase and gain of modulation ofPurkinje cell simple spikes, which in turn modulate the activity oftarget neurons in the vestibular nucleus; this process underlies the rapidVOR gain adaptation observed in both PC-Dg2 and control mice. On alonger timescale, the learned Purkinje cell activity guides plasticity at thetarget neurons in the vestibular nuclei28,29, the polarity of which ispresumably regulated by the precise timing of simple spikes relative toinput from mossy fiber collaterals30. Simultaneously, a partial extinctionof the previously learned changes at the level of the Purkinje cells takes

Visual stimulus (°)

Eye position (°)

Simple spikeresponses(Count – sweep)

Complex spikeresponses(Count – sweep)

Control

–30

0

30

60

90

120

0

20

40

60

Mod

. am

plitu

de (

%)

0

10

15

20

0.1 0.2 0.4 0.8 1.6Frequency (log Hz)

0.1 0.2 0.4 0.8 1.6Frequency (log Hz)

0.1 0.2 0.4 0.8 1.6Frequency (log Hz)

Complex spikesSimple spikes

a

c

b

5

PC-∆γ2

PC-∆γ2

PC-∆γ2

FF

am

plitu

de (

Hz)

Pha

se (

°)

Without visual

stimulus

0

0.2

0.4

0.6

0.8

CV

0

0.1

0.2

0.3

0.4

0.5

0.6

CV CV2

****

FF

0

0.2

0.4

0.8

1.2

1.0

0.6

Firi

ng fr

eque

ncy

(Hz)

010203040506070

FF

Firi

ng fr

eque

ncy

(Hz)

Control

Control

0

0.2

0.4

0.6

0.8

With visual

stimulus

***

0

0.4

0.8

0

10

20

30

40

50

60 1.2

0

0.2

0.4

0.6

0.8

1.0

0.2

0.6

1.0

Firi

ng fr

eque

ncy

(Hz)

Firi

ng fr

eque

ncy

(Hz) **

–202

–202

03060

0.5 1.0 1.5 2.0

048

01020

0 2.5Time (s)

01020

–202

–202

03060

048

Time (s)

01020

01020

0.5 1.0 1.5 2.00 2.5

Figure 5 Temporal patterns of simple spike

activities of floccular Purkinje cells are

specifically affected in PC-Dg2 mice during both

compensatory eye movement behavior and

spontaneous behavior. (a) Representative single-

unit activity recorded from Purkinje cells in the

flocculus of a control and a PC-Dg2 mouse during

fixed velocity (81 s�1, 0.2 Hz) OKR stimulation.The visual stimulus and eye position are shown

together with histograms of simple spike and

complex spike frequencies and corresponding

raster plots. Count (frequency) is the number of

spikes per bin divided by bin size, and sweep

is the cycle number. (b) Firing frequency (FF)

amplitude, phase relative to stimulus and

amplitude of modulation (see Online Methods) of

floccular simple-spike activities during optokinetic

stimulation (81 s�1, 0.1–1.6 Hz) were not

significantly different between PC-Dg2 and control

mice (P 4 0.3). (c) Although the average firing

frequency of simple and complex spike activity did

not differ between PC-Dg2 and control mice, the

coefficient of variation of simple spikes in PC-Dg2mice was significantly reduced in recordings both

with and without visual stimuli (P ¼ 0.008 and

P ¼ 0.022, respectively). Also, the coefficient of

variation of adjacent intervals values of simplespikes were significantly lower than those of

controls in both conditions. Error bars denote

s.e.m. * P o 0.05, ** P o 0.01 and

*** P o 0.0001.

1046 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

place31. The memory is thus partially transferred to the target nuclei,potentially underlying long-term consolidation29,32. After several daysof training, this form of ‘systems consolidation’ ensures that thecerebellar cortex is no longer responsible for the expression of thelearned behavior, but mainly regulates the precise timing (phase). InPC-Dg2 mice, the altered temporal patterns of Purkinje cell simplespikes could impair the induction of plasticity in the nuclei andthus consolidation.

Given this working hypothesis, we examined whether deficits inVOR gain consolidation and phase adaptation in PC-Dg2 micecould be replicated in a conceptual model of the idealized VORcircuit (Supplementary Data). We modeled the modulation of Pur-kinje cell simple-spike firing during head movement resulting fromlinear summation of sinusoidal excitatory (parallel fiber) and inhibi-tory (interneuron) inputs6 (Fig. 6a,b). We assumed the gain and phaseof such modulation to be regulated through bidirectional plasticity ofthe inputs33 (Fig. 6c) and that Purkinje cells and mossy fiber collateralssubsequently modulate, by linear summation of their activity, the firingof cells in the vestibular nuclei (Fig. 6d,e), which in turn control eyemovement (Fig. 6f). As a result of the absence of inhibition in PC-Dg2mice, the simple spike activation required for adequate modulation ofthe vestibular nucleus was out of range of the normal plasticitymechanisms in the cerebellar cortex (Fig. 6c). In addition, impairedplasticity of the inputs to the vestibular nucleus (Fig. 6e) bothabolished consolidation and excluded the possibility of extremephase adaptations (Fig. 6f).

Data from four training sessions, each followed by an overnightperiod (Fig. 6g), were simulated using the upper bounds on sinusoidalmodulation (Fig. 6a–f). Adaptation of modulation (Fig. 6h) wassimulated as an exponential decay from the start position in thepolar plot (defined by the initial gain and phase) toward a new positiondetermined by the experimental procedure (Fig. 6g). Simulationparameters were chosen to mimic the rate of adaptation observedexperimentally (Supplementary Data). Under these conditions, bothcontrol and PC-Dg2 Purkinje cells rapidly reached the requiredmodulation during short-term VOR adaptation (Fig. 6h). However,impairments in both VOR gain consolidation and phase adaptation canbe generated if we assume that disrupted plasticity in the vestibularnucleus is caused by poor timing of simple spikes (Figs. 3, 5 and 6h).

DISCUSSION

Signal coding and plasticity in cerebellar learning

Although inhibitory interneurons in the molecular layer of the cere-bellum have been studied extensively5–8, their behavioral relevance hasremained enigmatic. We found that these interneurons shape thetemporal patterns of Purkinje cell simple spikes and suggest that thisprocess could be essential for plasticity and consolidation in thecerebellar and vestibular nuclei.

Floccular Purkinje cells control the adaptation of compensatory eyemovements by modulating the activity of vestibular nucleus neurons(Supplementary Fig. 1). To adapt the VOR, two things should happenin the framework of a distributed memory model. First, Purkinje cells

Figure 6 Interpretation of VOR adaptation data

using a distributed memory model. (a) Modeled

activation of parallel fibers and interneurons

plotted in polar coordinates. Most parallel fibers

modulate (increase activation) in phase with

ipsilateral head movement (01), whereas a

fraction responds to input from the contralateral

horizontal canal (1801). Interneurons are modeledsimilarly, but with an opposite sign representing

their inhibitory nature. (b) Data are presented as

in a, but for PC-Dg2 mice lacking inhibition.

(c) Maximum simple-spike modulation attainable

by appropriate depression and potentiation of the

excitatory and inhibitory inputs shown in a and b

for control (blue) and PC-Dg2 (red) mice (based

on linear input summation). (d) Modulation of

target vestibular nucleus neurons attainable by

linear summation of mossy fiber inputs (blue

arrow, in phase with head movement) and

Purkinje cell inputs (c, blue curve) in control

mice. The black arrow represents the efficacy of

mossy fiber input before training. (e) Data are

presented as in d, but for PC-Dg2 mice, where the

efficacy of plasticity at the mossy fiber synapses

is presumably impaired. (f) Limited simple-spike

modulation and mossy fiber plasticity restrained

eye movements in PC-Dg2 mice (red) comparedwith control mice (blue). The control curve also

covers the area of VOR phase reversal, from

out-of-phase with head movement (1801 in this

figure) to in-phase (01). (g) Experimental data;

squares represent VOR gain and dashed lines

represent VOR phase relative to the head (shifted

by 1801, for ease of illustration; see also Fig. 4). For each session, after initial adaptation, the learned Purkinje cell signal determines the new ‘desired’ phase

and gain state for the vestibular nucleus neurons (dashed and solid black bars, respectively). The superimposed blue (control) and red (PC-Dg2) arrows

indicate the direction of change. (h) Simulation of the training procedure shown in g. During training, Purkinje cells rapidly approached their target modulation (i),

reflecting short-term VOR adaptation. Purkinje cell–guided plasticity of mossy fiber input to vestibular nuclei (ii) allowed control mice to gradually adapt the

phase of their VOR during prolonged training (iii). In PC-Dg2 mice, loss of vestibular nucleus consolidation impaired phase adaptation. For simplicity,

adaptation in the vestibular nuclei and partial extinction of cortical memory were simulated to occur between training sessions (gray bars).

0

900.5

0

900.5

180

Phase (°)

180

Gai

n

ii

0

0

i1.0 Purkinje cell

Mossy fiber1.0

0

90

180

0.5

1.0

0

iii Eye movement

Day 1 Day 2 Day 3 Day 4

0

0.25

0.5

0.75

1.0

0

60

120

180

Gai

n

Day 1 Day 2 Day 3 Day 4

Phase (°)

g

h

0

90

180

180 0

90

0

90

180

270

270

270

180 0

90

270

180 0

90

270

Simple spikemodulation

ControlVestibular nucleusmodulation

PC-∆γ2Vestibular nucleusmodulation

Parallel fibers

Parallel fibersControl

PC-∆γ2

a

b

c

d

e

Eyemovement

f

0

90

180

270

Interneurons

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1047

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

should ‘learn’ the correct simple-spike modulation and express it atsufficient gain to modulate the vestibular nuclei. Second, the inputfrom the direct vestibular pathway to the vestibular nuclei should besuppressed, as it only allows modulation in phase with ipsilateral headmovement. The first of these processes is thought to reflect comple-mentary inhibitory and excitatory actions, with plasticity at parallelfiber to Purkinje cell and parallel fiber to interneuron synapses, bothunder climbing fiber control13,24,25. The second is thought to occurthrough plasticity at mossy fiber to vestibular nuclei synapses25,28. InPC-Dg2 mice, the temporal fidelity of Purkinje cell firing was disruptedand consolidation of learned VOR adaptations was severely compro-mised (Figs. 4 and 6g). As induction of various forms of plasticity inthe vestibular and cerebellar nuclei could depend on the precise timingof inhibitory and excitatory input from Purkinje cells and mossyfiber collaterals (Supplementary Data), disruption of this timingwould impair transfer of plasticity to the nuclei and thus impairsystems consolidation.

Simple spike trains in Purkinje cells show substantially more tem-poral patterns than expected from random activation and these patternsare influenced by natural stimuli22. Both the electrical coupling amonginterneurons and the sagittal orientation of their axons34,35 (Supple-mentary Fig. 1) could enhance the effects of feedforward inhibition bypromoting common firing patterns in ensembles of Purkinje cells inindividual zones that are known to project to the same nucleus26. Theactivity patterns of individual Purkinje cells in an ensemble might thusinteract with each other and/or with those of mossy fiber and/orclimbing fiber collaterals to facilitate the induction of plasticity in thecerebellar and vestibular nuclei25,28,36,37. We therefore propose that thevestibular nuclei are the locus for consolidation (see also refs. 29,32).Alternatively, both initial learning and consolidation could occur in thecerebellar cortex and the consolidation signal could be preserved in theaverage simple-spike frequency of a particular Purkinje cell. In fact,changes in simple-spike frequencies in the flocculus of monkeys aresufficient to drive changes in eye velocity during trial-by-trial motorlearning38. To determine the extent to which spatiotemporal patterns ofsimple spikes contribute to consolidation and whether this consolida-tion occurs in the nuclei, future experiments will require simultaneousmulti-unit recording from ensembles of Purkinje cells and cerebellar orvestibular nuclei neurons during learning.

Previous studies have identified long-term changes at the parallelfiber to Purkinje cell synapse as a potential plasticity mechanism duringcerebellar learning, and some mouse lines with disrupted LTD induc-tion at this synapse show impaired motor learning39–41. However,neither parallel fiber LTD nor LTP were impaired in PC-Dg2 mice.Notably, the motor learning deficits in PC-Dg2 mice differed fromthose seen in mouse lines in which LTD was impaired by blocking PKC,PKG or aCaMKII activity in Purkinje cells. Furthermore, in thelatter mouse lines, acute learning was affected more severely than inPC-Dg2 mice, whereas learning over multiple days of training wasless affected39–42.

GABAergic interneurons in the cerebellar cortex have ample possibi-lities to induce and express plasticity at both the synaptic input andoutput level1,9,13,43–45. Simultaneous induction of LTP at molecular layerinterneuron and parallel fiber to Purkinje cell synapses is required forassociative fear conditioning9. In this scenario, the potentiation ofGABAergic synapses may balance the LTP of excitatory inputs in aform of scaling to preserve coincidence detection of parallel fiberinputs7,9, which mirrors the balanced LTD seen at parallel fiber andinterneuron synapses with the same conjunctive climbing fiber pair-ing45. Loss of this scaling mechanism in PC-Dg2 mice might contributeto the observed phenotype.

Inhibition is essential for spike patterning and learning

Although PC-Dg2 mice showed marked deficits in cerebellar motorlearning, baseline motor performance was only moderately affected.Despite the lack of synaptic GABAA receptors on PC-Dg2 Purkinjecells, we found their average simple-spike frequency to be normal. Thiscould reflect enhancement of another inhibitory input (for example,GABAB receptors) and/or reduced parallel fiber excitatory input.Although GABAB receptor–mediated inhibition of PC-Dg2 Purkinjecells was unchanged (Supplementary Fig. 10), we found a significantdecrease in AMPA receptor–mediated EPSC charge transfer afterparallel fiber stimulation (P o 0.0025; Supplementary Fig. 4). Thismight allow Purkinje cells to maintain their excitability in a normaloperational range in the absence of fast inhibition. In contrast, the lossof temporal fidelity in Purkinje cell responses to parallel fiber stimula-tion and the increase in simple-spike regularity in PC-Dg2 mice werecomparable to the changes seen after acute pharmacological blockadeof GABAA receptors7,12 (Figs. 3 and 5). The cerebellum may thuscompensate for the loss of certain functions of molecular layer inter-neurons, but these interneurons are essential for the temporal controlof Purkinje cell activity and for both phase adaptation learning andconsolidation of gain adaptations.

By deleting synaptic GABAA receptors from Purkinje cells in PC-Dg2mice, we also disrupted any inhibition mediated by recurrent collateralsof Purkinje cell axons10,46. However, because GABAergic terminals frombasket and stellate cells onto Purkinje cells vastly outnumber those fromrecurrent collaterals and because Purkinje-Purkinje contacts in micetend to be restricted to young animals46, the phenotype that we observedis most likely caused by the loss of inhibition from molecular layerinterneurons. Moreover, although it has been proposed that Purkinjecell axon collaterals contribute to fast cerebellar oscillations in adultrats47, such oscillations have not been recorded in wild-type mice10.

General functional implications

Studies on learning and memory have focused largely on the role ofplasticity at excitatory synapses onto projecting neurons. However,GABAergic interneurons also express plasticity, which increases thecomputational capacity of their microcircuit1,2,9. We examined the roleof fast synaptic inhibition in cerebellar motor learning using geneticdissection of the circuit and our results suggest that feedforwardinhibition is essential for specific aspects of procedural learning.

Can our findings be extrapolated to other brain regions? Feedfor-ward inhibition is a common motif throughout the CNS. In the amyg-dala, it mediates extinction learning of conditioned fear responses48. Incortical circuits, some interneuron types may serve functions similar tothose that we identified in the cerebellum. For example, feedforwardinhibitory interneurons in the hippocampus may promote the tem-poral fidelity of synaptic integration and action potential generation inpyramidal cells necessary for encoding declarative memories2,3. Thus,feedforward inhibition might be an operational necessity for memoryformation in different brain circuits.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe thank L. Cheyne, D. Massie, M. Rutteman, R. Avila Freire, E. Dalm and J.v.d.Burg for their excellent technical assistance, and D. Andersson and L. Kelly fortheir participation in the initial electrophysiological studies. This work wassupported by the J. Ernest Tait Estate Aberdeen (W.W.), a Medical Research

1048 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Council program grant G0800399 (P.W. and W.W.), the Royal Society (P.W.), theInstitute Pasteur-Fondazione Cenci Bolognetti (M.R.), a Wellcome Trust programgrant (M.F.), Regione Piemonte (Ricerca Scientifica Applicata A218 and RicercaSanitaria Finalizzata 2006) and Compagnia di San Paolo (M.S.-P.), the DutchOrganization for Medical Sciences (C.I.D.Z.), Life Sciences (C.I.D.Z.), Senter(Neuro-Bsik, C.I.D.Z.), Prinses Beatrix Fonds (C.I.D.Z.), and the SENSOPAC(SENSOrimotor structuring of Perception and Action for Emerging Cognition)program of the European Community (C.I.D.Z.).

AUTHOR CONTRIBUTIONSP.W. developed the mouse model and helped coordinate the project. M.S. andA.B. designed and performed VOR experiments and in vivo electrophysiology.M.F. and M.R. designed, performed and analyzed the in vitro electrophysiologyexperiments. P.W., L.V. and M.S.-P. performed quantitative anatomical studies.Z.G. and F.E.H. performed LTP and LTD experiments. S.v.D. designed andimplemented the model. W.W. initiated the project and coordinated collaborationsbetween groups. C.I.D.Z. designed experiments and guided the project.P.W., M.S., M.R., S.v.D., W.W., M.F. and C.I.D.Z. co-wrote the manuscript.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Smith, S.L. & Otis, T.S. Pattern-dependent, simultaneous plasticity differentially trans-forms the input-output relationship of a feedforward circuit. Proc. Natl. Acad. Sci. USA102, 14901–14906 (2005).

2. Kullmann, D.M. & Lamsa, K.P. Long-term synaptic plasticity in hippocampal interneur-ons. Nat. Rev. Neurosci. 8, 687–699 (2007).

3. Pouille, F. & Scanziani, M. Enforcement of temporal fidelity in pyramidal cells by somaticfeed-forward inhibition. Science 293, 1159–1163 (2001).

4. Lamsa, K., Heeroma, J.H. & Kullmann, D.M. Hebbian LTP in feed-forward inhibitoryinterneurons and the temporal fidelity of input discrimination. Nat. Neurosci. 8,916–924 (2005).

5. Eccles, J.C., Ito, M. & Szentagothai, J. The Cerebellum as a Neuronal Machine (Springer-Verlag, New York, 1967).

6. Miyashita, Y. & Nagao, S. Contribution of cerebellar intracortical inhibition to Purkinjecell response during vestibulo-ocular reflex of alert rabbits. J. Physiol. (Lond.) 351,251–262 (1984).

7. Mittmann, W., Koch, U. & Hausser, M. Feed-forward inhibition shapes the spike outputof cerebellar Purkinje cells. J. Physiol. (Lond.) 563, 369–378 (2005).

8. Barmack, N.H. & Yakhnitsa, V. Functions of interneurons in mouse cerebellum.J. Neurosci. 28, 1140–1152 (2008).

9. Scelfo, B., Sacchetti, B. & Strata, P. Learning-related long-term potentiation ofinhibitory synapses in the cerebellar cortex. Proc. Natl. Acad. Sci. USA 105,769–774 (2008).

10. Orduz, D. & Llano, I. Recurrent axon collaterals underlie facilitating synapsesbetween cerebellar Purkinje cells. Proc. Natl. Acad. Sci. USA 104, 17831–17836(2007).

11. Davie, J.T., Clark, B.A. & Hausser, M. The origin of the complex spike in cerebellarPurkinje cells. J. Neurosci. 28, 7599–7609 (2008).

12. Hausser, M. & Clark, B.A. Tonic synaptic inhibition modulates neuronal output patternand spatiotemporal synaptic integration. Neuron 19, 665–678 (1997).

13. Jorntell, H. & Ekerot, C.F. Reciprocal bidirectional plasticity of parallel fiber receptivefields in cerebellar Purkinje cells and their afferent interneurons. Neuron 34, 797–806(2002).

14. Santamaria, F., Tripp, P.G. & Bower, J.M. Feedforward inhibition controls the spread ofgranule cell–induced Purkinje cell activity in the cerebellar cortex. J. Neurophysiol. 97,248–263 (2007).

15. Raman, I.M. & Bean, B.P. Resurgent sodium current and action potential formation indissociated cerebellar Purkinje neurons. J. Neurosci. 17, 4517–4526 (1997).

16. Wulff, P. et al. From synapse to behavior: rapid modulation of defined neuronal typeswith engineered GABAA receptors. Nat. Neurosci. 10, 923–929 (2007).

17. Schweizer, C. et al. The gamma 2 subunit of GABA(A) receptors is required formaintenance of receptors at mature synapses. Mol. Cell. Neurosci. 24, 442–450(2003).

18. Barski, J.J., Dethleffsen, K. & Meyer, M. Cre recombinase expression in cerebellarPurkinje cells. Genesis 28, 93–98 (2000).

19. Brickley, S.G., Cull-Candy, S.G. & Farrant, M. Single-channel properties of synaptic andextrasynaptic GABAA receptors suggest differential targeting of receptor subtypes.J. Neurosci. 19, 2960–2973 (1999).

20. Lorez, M., Benke, D., Luscher, B., Mohler, H. & Benson, J.A. Single-channel propertiesof neuronal GABAA receptors from mice lacking the gamma2 subunit. J. Physiol. (Lond.)527, 11–31 (2000).

21. Brunel, N., Hakim, V., Isope, P., Nadal, J.P. & Barbour, B. Optimal information storageand the distribution of synaptic weights: perceptron versus Purkinje cell. Neuron 43,745–757 (2004).

22. Shin, S.L. et al. Regular patterns in cerebellar Purkinje cell simple spike trains. PLoSOne 2, e485 (2007).

23. Walter, J.T., Alvina, K., Womack, M.D., Chevez, C. & Khodakhah, K. Decreases in theprecision of Purkinje cell pacemaking cause cerebellar dysfunction and ataxia. Nat.Neurosci. 9, 389–397 (2006).

24. Ito, M. Cerebellar flocculus hypothesis. Nature 363, 24–25 (1993).25. Lisberger, S.G. Cerebellar LTD: a molecular mechanism of behavioral learning? Cell 92,

701–704 (1998).26. Schonewille, M. et al. Zonal organization of the mouse flocculus: physiology, input and

output. J. Comp. Neurol. 497, 670–682 (2006).27. Hoebeek, F.E. et al. Increased noise level of Purkinje cell activities minimizes impact of

their modulation during sensorimotor control. Neuron 45, 953–965 (2005).28. Gittis, A.H. & du Lac, S. Intrinsic and synaptic plasticity in the vestibular system.

Curr. Opin. Neurobiol. 16, 385–390 (2006).29. Kassardjian, C.D. et al. The site of a motor memory shifts with consolidation. J. Neurosci.

25, 7979–7985 (2005).30. Medina, J.F. & Mauk, M.D. Computer simulation of cerebellar information processing.

Nat. Neurosci. 3 Suppl: 1205–1211 (2000).31. Medina, J.F., Nores, W.L. & Mauk, M.D. Inhibition of climbing fibers is a signal for the

extinction of conditioned eyelid responses. Nature 416, 330–333 (2002).32. Shutoh, F., Ohki, M., Kitazawa, H., Itohara, S. & Nagao, S. Memory trace of motor

learning shifts transsynaptically from cerebellar cortex to nuclei for consolidation.Neuroscience 139, 767–777 (2006).

33. Coesmans, M., Weber, J.T., De Zeeuw, C.I. & Hansel, C. Bidirectional parallel fiberplasticity in the cerebellum under climbing fiber control. Neuron 44, 691–700 (2004).

34. Mann-Metzer, P. & Yarom, Y. Electrotonic coupling interacts with intrinsic properties togenerate synchronized activity in cerebellar networks of inhibitory interneurons.J. Neurosci. 19, 3298–3306 (1999).

35. Van Der Giessen, R.S., Maxeiner, S., French, P.J., Willecke, K. & De Zeeuw, C.I.Spatiotemporal distribution of Connexin45 in the olivocerebellar system. J. Comp.Neurol. 495, 173–184 (2006).

36. Steuber, V. et al. Cerebellar LTD and pattern recognition by Purkinje cells. Neuron 54,121–136 (2007).

37. Blazquez, P.M., Hirata, Y. & Highstein, S.M. Chronic changes in inputs to dorsal Yneurons accompany VOR motor learning. J. Neurophysiol. 95, 1812–1825 (2006).

38. Medina, J.F. & Lisberger, S.G. Links from complex spikes to local plasticity and motorlearning in the cerebellum of awake-behaving monkeys. Nat. Neurosci. 11, 1185–1192(2008).

39. De Zeeuw, C.I. et al. Expression of a protein kinase C inhibitor in Purkinje cells blockscerebellar LTD and adaptation of the vestibulo-ocular reflex. Neuron 20, 495–508(1998).

40. Feil, R. et al. Impairment of LTD and cerebellar learning by Purkinje cell–specificablation of cGMP-dependent protein kinase I. J. Cell Biol. 163, 295–302 (2003).

41. Hansel, C. et al. alphaCaMKII Is essential for cerebellar LTD and motor learning. Neuron51, 835–843 (2006).

42. Boyden, E.S. et al. Selective engagement of plasticity mechanisms for motor memorystorage. Neuron 51, 823–834 (2006).

43. Kano, M., Rexhausen, U., Dreessen, J. & Konnerth, A. Synaptic excitation produces along-lasting rebound potentiation of inhibitory synaptic signals in cerebellar Purkinjecells. Nature 356, 601–604 (1992).

44. Duguid, I.C. & Smart, T.G. Retrograde activation of presynaptic NMDA receptorsenhances GABA release at cerebellar interneuron–Purkinje cell synapses. Nat. Neurosci.7, 525–533 (2004).

45. Mittmann, W. & Hausser, M. Linking synaptic plasticity and spike output at excitatoryand inhibitory synapses onto cerebellar Purkinje cells. J. Neurosci. 27, 5559–5570(2007).

46. Watt, A.J. et al. Traveling waves in developing cerebellar cortex mediated by asymme-trical Purkinje cell connectivity. Nat. Neurosci. 12, 463–473 (2009).

47. de Solages, C. et al. High-frequency organization and synchrony of activity in thePurkinje cell layer of the cerebellum. Neuron 58, 775–788 (2008).

48. Likhtik, E., Popa, D., Apergis-Schoute, J., Fidacaro, G.A. & Pare, D. Amygdala inter-calated neurons are required for expression of fear extinction. Nature 454, 642–645(2008).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1049

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSProcedures involving mice were performed in accordance with regulations of

the United Kingdom Animals (Scientific Procedures) Act 1986, the Animal

Care and Use Committee of Turin University and the Dutch Ethical Committee

for animal experiments.

Generation of PC-Dc2 mice. We generated g2I77lox mice by flanking exon 4

of the Gabrg2 gene with loxP sites and changing the codon encoding F77 in

exon 4 to I, which resulted in a neutral amino acid substitution16. Homozygous

g2I77lox mice were crossed with mice heterozygous for g2I77lox and hemi-

zygous for an L7-cre transgene16,18. g2I77lox/g2I77lox; L7-cre (PC-Dg2) and

g2I77lox/g2I77lox (controls) littermates were used. Mice were genotyped by

PCR analysis of genomic DNA using primers to test for the g2I77lox allele (213-

bp control, 250-bp g2I77lox) (g2lx5¢_s (5¢-GTCATGCTAAATATCCTA

CAGTGG-3¢) and g2lx5¢_as (5¢-GGATAGTGCATCAGCAGACAATAG-3¢))

and primers to test for the cre transgene (250-bp L7Cre) (Cre1 (5¢-GAC

CAGGTTCGTTCACTCATGG-3¢) and Cre2 (5¢-AGGCTAAGTGCCTTCTCTA

CAC-3¢)).

Morphology. Adult mice were anaesthetized by intraperitoneal injection of

ketamine/xylazine and perfused with 4% paraformaldehyde (wt/vol) in phos-

phate-buffered saline (PBS, pH7.4). Cerebella were cryoprotected in sucrose

(10%, 20% and 30% in PBS, wt/vol) and cut into 16-mm coronal sections with

a cryostat. Following blocking in normal goat serum (10% in PBS, vol/vol with

0.5% Triton X-100, vol/vol), sections were incubated with antibodies to

calbindin (1:10,000, Swant), VGAT (1:3,000), VGLUT1 (1:1,000) or VGLUT2

(1:500, all Synaptic Systems). Sections were rinsed and incubated with

secondary antibodies conjugated to Alexa 488 (Molecular Probes) or Cy3

(Jackson Immunoresearch). Sections were examined with a laser-scanning

confocal microscope (Zeiss LSM5 Pascal). Stacks of 5–15 sections spaced

250–350 nm apart were acquired (pinhole, 1 Airy unit). Quantification of

VGLUT2-positive puncta was carried out on segmented images spanning the

molecular layer; eight confocal fields (13,225 mm2 per field) were counted per

mouse (n ¼ 4). For VGLUT1 and VGAT, images acquired at a magnification of

8.1 � 10�3 mm2 per pixel (512 � 512 pixels) were segmented using a threshold

that maximized the selection of immunofluorescent puncta over background.

The number and density of puncta were calculated with ImageJ software

(http://rsbweb.nih.gov/ij/). For VGLUT1 and VGAT, six and eight fields (2,125

mm2 per field) were counted per mouse (n ¼ 3 and 4, respectively). To quantify

the number of Purkinje cells, we placed a line through the Purkinje cell layer

and counted all of the calbindin-positive cells on the line (four sections

per mouse, n ¼ 4). The density of molecular layer interneurons was calculated

in 3–6 fields (5,000 mm2 per field) of 3–5 Nissl-stained sections per mouse

(n ¼ 4).

For electron microscopy, adult mice (n ¼ 2 per genotype) were perfused

with 4% paraformaldehyde and 2.5% glutaraldehyde (vol/vol) in phosphate

buffer (0.1 M, pH 7.4). Cerebella were postfixed in the same solution overnight.

Blocks of tissue were postfixed in 1% osmium tetroxide (vol/vol, in 0.1 M

cacodylate buffer), dehydrated in ethanol and embedded in Epon-Araldite.

Ultrathin sections were stained with uranyl acetate and lead citrate and

analyzed with a JEM-1010 transmission electron microscope (Jeol) equipped

with a side-mounted CCD camera (Mega View III, Soft Imaging System). In

each mouse, 90 electron micrographs were taken randomly from the neuropil

of the molecular layer at a magnification of 30,000� (15.7 mm2 per micro-

graph) to compare the density of parallel fiber to Purkinje cell synapses.

In vitro electrophysiology. Mice (10–25-weeks-old) were anaesthetized with

isoflurane (IVAX Pharmaceuticals) and parasagittal slices (250–300 mm) were

cut from the cerebellar vermis/paravermis (HM 650V, Microm International

GmbH) as described previously16. Slices were transferred to a submerged

recording chamber and perfused (1.5–2.5 ml min�1) with an external solution

containing 125 mM NaCl, 2.5 mM KCl, 2 mM CaCl2, 1 mM MgCl2, 25 mM

NaHCO3, 1.25 mM NaH2PO4 and 25 mM D-glucose, pH 7.4 when bubbled

with 95% O2 and 5% CO2. Patch-clamp recordings were made with Axopatch

200A or 200B amplifiers (Molecular Devices) from Purkinje cells visualized

under infrared differential interference contrast optics (Zeiss Axioscop or

Olympus BX51 WI). Whole-cell and single-channel currents were recorded

at 23–26 1C. Simple spike activity was recorded at both 23–26 1C and 34 ± 2 1C

in loose cell-attached mode with external solution in the recording pipette.

Firing was recorded in voltage- or current-clamp with the pipette current

set to zero12.

For whole-cell and loose cell-attached recordings, pipettes were pulled from

thin-walled borosilicate glass tubing (1.5 mm outer diameter, 1.17 mm inner

diameter, G150TF-3, Warner Instruments). For patch recordings, thick-walled

borosilicate glass tubing (1.5 mm outer diameter, 0.86 mm inner diameter, GC-

150F, Harvard Apparatus) was used. Pipettes were coated with Sylgard resin

(Dow Corning 184) and fire polished to give a final resistance of 2–6 MO(whole cell and loose cell attached) or 10–15 MO (single channel). The internal

solution contained 140 mM CsCl, 4 mM NaCl, 0.5 mM CaCl2, 10 mM N-2-

hydroxyethylpiperazine-N¢-2-ethanesulphonic acid (HEPES), 5 mM ethylene-

glycol-bis (b-aminoethylether)-N,N,N¢,N¢-tetraacetic acid (EGTA) and 2 mM

Mg-ATP, pH 7.3 with CsOH. Ionotropic glutamate receptors were blocked with

10 mM D-AP5 and 5 mM CNQX. mIPSCs were recorded in the presence of 0.5–1

mM TTX.

Parallel fiber–evoked responses were recorded in loose cell-attached mode

during molecular layer stimulation (a glass pipette containing external solution

was placed in fixed position B100–150 mm from the recorded Purkinje cell

soma). Stimuli of 5–10 V and 100-ms duration were delivered at 0.5 Hz

(Digitimer DS2 isolated stimulator). Recordings were made in the absence of

drugs. The effect of GABAA receptor blockade was tested using SR-95531

(40 mM). In all cases tested, 2,3-dioxo-6-nitro-1,2,3,4-tetrahydrobenzo[f]qui-

noxaline-7-sulfonamide (NBQX) completely blocked evoked responses (data

not shown).

Long-term plasticity at parallel fiber–Purkinje cell synapses was examined as

previously described41. We cut 200-mm-thick parasagittal slices in ice-cold

artificial cerebrospinal fluid (containing 124 mM NaCl, 5 mM KCl, 1.25 mM

Na2PO4, 2 mM MgSO4, 2 mM CaCl2, 26 mM NaHCO3 and 15 mM D-glucose,

bubbled with 95% O2 and 5% CO2). Experiments were carried out at 23–26 1C

with GABAA receptors blocked (100 mM picrotoxin). Whole-cell patch-clamp

recordings of Purkinje cells were performed using an EPC-10 amplifier (HEKA

Electronics). Pipette resistance was 4–5 MO when filled with intracellular

solution containing 120 mM potassium gluconate, 9 mM KCl,

10 mM KOH, 3.48 mM MgCl2, 4 mM NaCl, 10 mM HEPES, 4 mM Na2ATP,

0.4 mM Na3GTP and 17.5 mM sucrose (pH 7.25). LTP was induced by parallel

fiber stimulation at 1 Hz for 5 min in current-clamp mode and measured

by test responses recorded in voltage-clamp mode. LTD was induced using

combined parallel and climbing fiber stimulation41. All drugs were obtained

from Tocris Bioscience, Ascent Scientific or Sigma.

Eye-movement recordings. Mice (12–30-weeks-old) were surgically prepared

under general anesthesia with isoflurane. A construct with two nuts was

attached to the frontal and parietal bones using Optibond (Kerr) and Charisma

(Heraeus Kulzer). After 5 d of recovery, mice were placed in a restrainer, with

their heads bolted to a bar. The restrainer was fixed onto the turntable. A

cylindrical screen (diameter 63 cm) with a random-dotted pattern (each

element 21) surrounded the turntable (diameter 60 cm). OKR and visual

VOR were evoked by independently rotating the screen and turntable

(51 amplitude at different frequencies, AC servo-motors, Harmonic Drive

AG). The table and drum position were measured by potentiometers and the

signal was digitized (CED) and stored for off-line analysis. Eye movements were

recorded using an infrared CCD camera fixed to the turntable (240 Hz,

ISCAN). Video calibrations and eye movement computations were carried

out as described previously27,49.

In vivo electrophysiology. Mice (15–40-weeks-old) were surgically prepared

under general anesthesia by mounting a pedestal as described above27. A

recording chamber was built around craniotomies in left and right occipital

bones. Extracellular Purkinje cell activity was recorded using borosilicate glass

electrodes (outer diameter ¼ 2.0 mm, inner diameter ¼ 1.16 mm, 2 M NaCl,

4–8 MO). Electrodes were advanced by a hydraulic micro-drive (Narishige).

Recordings were made from left and right Crus I and II, paramedian lobule,

and (para)flocculus (recordings during optokinetic stimulation were from

floccular Purkinje cells). Purkinje cells were identified by the brief pause in

simple-spike activity following each complex spike. The raw signal was

NATURE NEUROSCIENCE doi:10.1038/nn.2348

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

amplified, filtered (CyberAmp, CED), digitized (CED) and stored for off-line

analysis. Following each recording session, the brain was covered with grami-

cidin-containing ointment and the chamber was sealed with bone wax.

Data analysis. During in vitro experiments, signals were recorded onto digital

audiotape (DTR-1204, BioLogic, DC to 20 kHz); for analysis, replayed signals

were filtered at 2 or 5 kHz (whole-cell and single-channel or loose cell-attached

recordings, respectively, –3 dB, 8-pole lowpass Bessel) and digitized at 10 kHz

(Digidata 1200, Axotape, Molecular Devices). mIPSCs were detected using a

scaled template detection method50 implemented in IGOR Pro 5.0 (Wave-

metrics) with NeuroMatic 1.91 (http://www.neuromatic.thinkrandom.com)16.

The decay of synaptic currents is best described by the sum of two exponential

functions according to

y ¼ A1 exp�ðx � x0Þ

t1

� �+ A2 exp

�ðx � x0Þt2

� �

where x0 is the decay onset, t1 and t2 are the decay time constants of the fast

and slow components, and A1 and A2 are their respective amplitudes. The

weighted time constant of decay (tw) was calculated according to

tw ¼ t1A1

A1 + A2

� �+ t2

A2

A1 + A2

� �

Purkinje cell simple spikes were detected by threshold crossing. ISIs and PSTHs

were generated using NeuroMatic and IGOR Pro 5.0 for spontaneous and

parallel fiber–evoked responses, respectively. To determine the number of extra

spikes evoked by stimulation, we integrated PSTHs. A linear fit to the

prestimulus section was extrapolated over the full duration of the integral

and subtracted to yield the cumulative spike probability corrected for baseline

firing7. The number of additional spikes evoked by each stimulus was

determined by averaging over a 0–60-ms period.

Off-line analysis of eye movements and in vivo recordings was performed in

Matlab (MathWorks)27. Gain and phase of eye movements were determined by

fitting sine functions to the slow-phase eye velocity traces. Gain was computed

as the ratio of eye velocity to stimulus velocity, whereas phase was expressed as

the difference (in degrees) between eye velocity and stimulus velocity traces.

In vivo simple spikes and complex spikes were discriminated using custom-

made routines based on principal component analysis. Simple-spike PSTHs

(100 bins per cycle) were compiled at each stimulus frequency and fitted by a

sine function. Epochs containing quick phases were deleted from the trace

(�50 to +150 ms). Modulation was calculated by dividing the amplitude of the

fitted sine wave by its offset. Phase difference was calculated as the difference

between the phase of the fitted sine wave and the optokinetic stimulus.

Models and simulations. A phenomenological model of an idealized VOR

circuit was created to elucidate the potential role of vestibular nuclei in a phase

adaptation procedure. Elements in the VOR circuit were characterized by the

gain and phase of their sinusoidal modulation, which define the coordinates of

a position on a polar plot. Plasticity rules were employed phenomenologically,

as exponential decay of the modulation along a trajectory in a polar plot, with

the target gain and phase (new position in the polar plot) being defined by the

vestibular mismatch. Equations were solved numerically using Matlab. See

Supplementary Data for details.

Statistical analyses. Statistical tests were performed with GraphPad (Prism 3.0,

GraphPad Software) or with SPSS 11 (SPSS). Unless stated otherwise, data were

compared with two-tailed paired or unpaired Student’s t tests, as appropriate.

We also used two-way repeated-measures ANOVA and, where data were non-

normally distributed (Shapiro-Wilk test), the Mann-Whitney U test. The level

of significance was set at P o 0.05.

49. Stahl, J.S., van Alphen, A.M. & De Zeeuw, C.I. A comparison of video and magneticsearch coil recordings of mouse eye movements. J. Neurosci. Methods 99, 101–110(2000).

50. Clements, J.D. & Bekkers, J.M. Detection of spontaneous synaptic events with anoptimally scaled template. Biophys. J. 73, 220–229 (1997).

doi:10.1038/nn.2348 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Disparity- and velocity-based signals for three-dimensional motion perception in human MT+

Bas Rokers, Lawrence K Cormack & Alexander C Huk

How does the primate visual system encode three-dimensional motion? The macaque middle temporal area (MT) and the human

MT complex (MT1) have well-established sensitivity to two-dimensional frontoparallel motion and static disparity. However,

evidence for sensitivity to three-dimensional motion has remained elusive. We found that human MT1 encodes two binocular cues

to three-dimensional motion: changing disparities over time and interocular comparisons of retinal velocities. By varying important

properties of moving dot displays, we distinguished these three-dimensional motion signals from their constituents, instantaneous

binocular disparity and monocular retinal motion. An adaptation experiment confirmed direction selectivity for three-dimensional

motion. Our results indicate that MT1 carries critical binocular signals for three-dimensional motion processing, revealing an

important and previously overlooked role for this well-studied brain area.

Most research on motion processing uses computer-generated patternsthat move across two-dimensional displays. Real objects, however,move through a three-dimensional world. To support the binocularperception of three-dimensional motion, the visual system must extractinformation from different patterns of dynamic stimulation on theretinae. Much is known about how the visual system processes stimulito encode two-dimensional motion in a fixed depth plane and how staticbinocular disparities are processed to represent position in depth1,2.

Much less is known, however, about how the visual system combinestwo-dimensional motion and binocular disparity to compute three-dimensional motion. Most notable is the dearth of evidence fordirect roles of the MT and medial superior temporal (MST) visualareas in three-dimensional motion, which is surprising given theircentral role in two-dimensional motion processing and well-establisheddisparity sensitivity3–9.

The visual system could rely on one or both of two binocularcues to compute three-dimensional motion direction: changes inbinocular disparity over time (changing disparity) and/or the differ-ence in the interocular image velocities at corresponding points onthe two retinae (interocular velocity difference, IOVD). The changingdisparity cue is sufficient for the perception of motion throughdepth10,11, and recent work has shown that the IOVD cue makes arobust, independent contribution12–17.

We carried out a series of four functional magnetic resonanceimaging (fMRI) experiments to identify distinct three-dimensionalmotion signals in the human visual system, to elucidate the proces-sing of the disparity-based (changing disparity) and/or velocity-based(IOVD) cues, and to test for three-dimensional motion directionselectivity. Across all of our experiments, MT+ responses were consis-tently selective for stimuli that specified three-dimensional motion.

This selectivity was distinct from responses to instantaneous dispari-ties and monocular motions. We found evidence of three-dimensionalmotion selectivity in V3A and lateral occipital (LO) complex, but nonein primary visual cortex. The pattern of blood oxygen level–dependent(BOLD) responses suggests that the computation of three-dimensionalmotion occurs subsequent to V1 and that three-dimensional motionsignals are evident at the population level in MT+.

RESULTS

Observers viewed the stimuli through a mirror stereoscope while wemeasured BOLD fMRI responses in a variety of visually responsiveregions of interest (ROIs), including V1, V2, V3, V3A/B, hV4, LO1/LO2and MT+ (combined putative MT/MST). In Experiments 1–3, displaysalternated in a blocked design between conditions every 12 s. Experi-ment 4 used an event-related adaptation protocol. Observers continu-ally performed a challenging detection task to control attention (seeOnline Methods).

Experiment 1: selectivity for three-dimensional motion

When an object moves directly toward or away from an observerfixating a stationary point, regions on the retinae corresponding to theobject’s location are stimulated by opposite directions of horizontalmotion. These dichoptic horizontal opponent motions must be dis-tinguished from similar patterns of retinal stimulation that do notspecify three-dimensional motion, such as locally opposing horizontalmotions on the same retina18,19 or opposing vertical motions, whetheracross the two retinae (dichoptic) or on the same retina (monocular).Horizontal dichoptically opponent motion contains both changes inbinocular disparity over time (the changing disparity cue) and inter-ocular velocity differences (the IOVD cue). To encode either of these

Received 4 February; accepted 27 April; published online 5 July 2009; doi:10.1038/nn.2343

Neurobiology, Psychology, Center for Perceptual Systems, Institute for Neuroscience, and Imaging Research Center, The University of Texas at Austin, Austin, Texas, USA.Correspondence should be addressed to B.R. (rokers@mail.utexas.edu).

1050 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

cues, the visual system must distinguish the orientation, the directionand the eye of origin of visual motion signals.

We measured fMRI responses while observers viewed displays com-prising 32 moving dot pairs (Fig. 1a). Dots in each pair had eitheropposing horizontal or opposing vertical directions of motion (Fig. 1a)and were presented either dichoptically (each dot in a pair to a differenteye) or monocularly paired (both dots in a pair in the same eye)(Fig. 1a). Dots pairs oscillated back and forth, starting at a random(but opposite) phase on their sinusoidal trajectory to avoid full-fieldcoherent monocular motion and to minimize oculomotor drive. Wequantified the amplitude of fMRI response modulation as displaysalternated between moving (12 s) and stationary (12 s).

MT+ responded preferentially to horizontally opponent, dichopticmotion: the only stimulus that simulated three-dimensional motion.We examined the average time course in MT+ as the dots alternatedbetween moving and stationary for the four conditions (Fig. 1b). MT+responses were B50% larger for dichoptic opponent motion thanfor monocular opponent motion when the dot paths were horizontal(t test on amplitudes, t16 ¼ 4.77, P ¼ 2 � 10�4), but were nearlyidentical for vertical motion (t16 ¼ 0.82, P ¼ 0.425).

We examined the response amplitudes for all of the areas that westudied (Fig. 1c). The difference in response to dichoptically separatedas compared with monocularly paired stimuli was substantially largerfor horizontal than for vertical stimuli in MT+ (repeated measuresANOVA, interaction between presentation (dichoptic/monocular) anddirection (horizontal/vertical), F1,35 ¼ 10.22, P ¼ 0.006). This patternof results was unique to MT+ (P 4 0.6 for all other ROIs).

The larger MT+ response to dichoptically separated compared withmonocularly paired horizontal opponent motion suggests that MT+carries signals that are specific to three-dimensional motion. The lackof difference for vertical motions demonstrates orientation specificity

of the responses consistent with the horizontal offset between the twoeyes. This dependence on orientation also confirms that the differencein responses between dichoptically separated and monocularly paireddisplays was not a result of differential patterns of monocular dotdensity between the conditions or of differences in overall patterns ofdichoptic stimulation.

These results suggest that MT+ processes three-dimensional motiondistinctly from similar stimuli that do not specify three-dimensionalmotion. Of course, horizontal dichoptically opponent motion, as withany real moving object, contains changing disparity and IOVD cues (aswell as instantaneous binocular disparities). Although psychophysicalevidence supports the notion that both cues are used12–15, it is unclearhow they are computed and combined. We therefore isolated each cuein two additional experiments: one that tested for a contribution ofthe changing disparity cue (Experiment 2) and one that tested for acontribution of the IOVD cue (Experiment 3). We then confirmedthat three-dimensional motion responses were direction selective inan event-related adaptation experiment (Experiment 4).

Experiment 2: disparity-based cue to three-dimensional motion

The changing disparity cue can be isolated in dynamic random dotstimuli that contain systematic changes in disparity over time, but donot contain any coherent monocular motion20,21. Such displays onlycontain disparity-based motion signals in the cyclopean view (that is,after binocular combination); each eye’s half image looks like incoher-ent random dot noise.

We examined this changing disparity–isolating stimulus (Fig. 2a).On each stimulus frame, dots were repositioned in random imagelocations and then assigned one of two disparities depending on thequadrant in which they fell (Fig. 2a). Disparities changed smoothlyand systematically over time to produce a percept of wedges movingsinusoidally through depth, with adjacent quadrants moving in oppo-site directions (Fig. 2a). Such stimuli do not contain coherent mono-cular motions and thus no IOVD cue, nor do they contain globalcoherent motion that might drive eye movements.

We measured fMRI responses as observers viewed a display thatalternated between this changing disparity stimulus (changing dis-parity motion) and a spatiotemporally scrambled version containingthe same number of dots and the same distribution of dot disparities

0 6 12 18 24–1

0

1

Dic

hopt

icM

onoc

ular

Vertical

L L

L LR

Horizontal

R R

R

a

c

Vertical

Horizontal

V1 V2 V3 V3A hV4 LO MT+0

0.5

1.0

0 6 12 18 24–1

0

1MT+ VerticalHorizontal

DichopticMonocular

fMR

I res

pons

e (∆

% B

OLD

)

MT+

fMR

I res

pons

e (∆

% B

OLD

)

b

Time (s) Time (s)

Dichoptic (MTD)MonocularDichopticMonocular

Figure 1 Stimuli and results for Experiment 1. (a) Stimulus conditions. Left

(L) and right (R) eye half images for each of four conditions are shown. Each

image contained 32 dots (14 are shown). Each dot oscillated sinusoidally

(arrows illustrate motions) and the two dots in a pair moved in anti-phase.

Dots moved along horizontal (left column) or vertical trajectories (right

column). In dichoptic conditions (top row), paired dots were split between

the left and the right images. In monocular conditions (bottom row), paired

dots were presented in the same eye. The horizontal-dichoptic stimuluswas perceived as three-dimensional motion; in other conditions, dots were

perceived as twinkling, but without any three-dimensional motion. Color

codes condition throughout all figures. Left and right half images are fusible

stereo-pairs. (b) MT+ responses were selective for three-dimensional motion.

Time series of fMRI response across moving and stationary epochs (averaged

across observers) in MT+ are shown for the four conditions. Left, MT+

responses to horizontal motion are larger for dichoptic than for monocular

presentation. Right, MT+ responses to vertical motion are similar for

dichoptic and monocular presentation. Error bars represent ± s.e.m. (c) MT+

showed the clearest selectivity for three-dimensional motion. fMRI response

amplitudes are shown across all visual areas, averaged across subjects. MT+

responses showed a significant interaction between orientation (horizontal/

vertical) and presentation (dichoptic/monocular), demonstrating a sensitivity

to three-dimensional motion separate from possible sensitivities to horizontal

or dichoptic motion alone (see text for statistical information). This selectivity

was not evident in other visual areas. Error bars represent ± s.e.m.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1051

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

(see Online Methods). Perceptually, the displays alternated betweencrisp, wedge-shaped surfaces moving through depth and a dynamicthree-dimensional cloud of dots with no coherent structure or motion.These stimuli were indistinguishable when viewed monocularly.

When we plotted the average time course of the MT+ response asthe display alternated between changing disparity motion and spatio-temporally scrambled motion, the response modulation was readilyapparent (Fig. 2b). This contrasts with the V1 response, which is notonly much weaker, but also inverted, probably as a result of theabundance of unstructured disparities continuously present in thespatiotemporally scrambled blocks (Fig. 2b).

To verify that the MT+ modulation was a result of a changing disparitysignal per se (rather than to spatial or temporal disparity structure presentin the changing disparity blocks), we ran three additional conditions:spatially scrambled (scrambling the spatial positions of the dots whileleaving the temporal structure intact), temporally scrambled (scramblingthe order in which the frames were presented while leaving the spatialstructure intact; Fig. 2b) and anti-correlated (pairing each white dotin one eye with a black dot in the other eye and vice versa). Thislast manipulation greatly degrades the disparity signal on which thechanging disparity cue relies13–15,21,22. Each condition alternated againstthe spatiotemporally scrambled stimulus.

In all three conditions, the response modulation was substantiallyreduced relative to the primary changing disparity motion condition,indicating that MT+ responses were largest for spatiotemporal patternsof disparity that specify three-dimensional motion. We plotted theresponse amplitudes for all of the areas studied in the main changingdisparity condition and the three other conditions (Fig. 2c). Comparisonacross areas showed a strong contrast between the earliest areas (for

example, V1, V2 and V3) and later areas (V3A, LO and MT+; Fig. 2c). InMT+, responses varied significantly between the changing disparitycondition and the other three conditions (ANOVA main effect of motiontype, F3,71 ¼ 16.90, P ¼ 0.003; changing disparity versus: spatiallyscrambled, t22 ¼ 2.45, P ¼ 0.023; temporally scrambled, t22 ¼ 8.04,P ¼ 5 � 10�8; anti-correlated, t22 ¼ 3.42, P ¼ 3 � 10�3).

Even though the spatially and temporally scrambled conditionscontained the same overall distribution of disparities, the spatio-temporal structure uniquely present in the changing disparity condi-tion yielded considerably larger MT+ responses. Similar modulationsof response were also observed in LO (F3,71 ¼ 16.80, P ¼ 0.003) andV3A (F3,71 ¼ 18.30, P ¼ 0.002), but not in earlier areas (P 4 0.06).

The relatively weak responses to the temporally scrambled stimulussuggest that MT+ is selective for motion from the changing disparitycue, distinct from disparity-defined edges or other depth-definedstructure generated by the presence of wedges in our displays. Thetemporally scrambled frames had the same number and averagemagnitude of disparity-defined edges, but lacked the temporal struc-ture consistent with smooth three-dimensional motion. We furtherverified this by repeating the main changing disparity condition withdisplays that contained different numbers of wedges (SupplementaryFig. 1). Note that anti-correlated displays containing only the changingdisparity cue yielded a B50% response reduction. In Experiment 3,when IOVDs were present, the identical manipulation had a verydifferent effect on MT+ responses.

Experiment 3: velocity-based cue to three-dimensional motion

Having established a changing disparity response in MT+, we tested forthe presence of an IOVD signal. Although it was possible in Experiment2 to remove the IOVD cue from the stimulus and to isolate thechanging disparity cue, it is not geometrically possible to do theconverse. One can, however, selectively degrade the changing disparitysignal using binocularly anti-correlated stimuli without affecting themonocular velocities on which the IOVD cue relies13,14.

We measured fMRI responses as subjects viewed displays that alter-nated between moving (12 s) and stationary (12 s). We presented either

0 6 12 18 24

0

0.5

a

b

c

V1 V2 V3 V3A hV4 LO MT+

0

0.5

0 6 12 18 24

0

0.5

–0.5 –0.5

MT+ V1

RL

Cyclopean

Motion type

Tim

e

CD motion

CD versus STS

fMR

I res

pons

e (∆

% B

OLD

)fM

RI r

espo

nse

(∆%

BO

LD)

CD versus STSSS versus STSTS versus STS

TS versus STS

CD versus STS – anti-correlated

Time (s)Time (s)

Figure 2 Stimuli and results for Experiment 2. (a) Changing disparity (CD)

motion stimulus. The left pairs of patches show the left and right eye’s half

images. The rightmost patches depict a scheme of the cyclopean percept.

Binocular disparities between left and right images defined a spatiotemporal

pattern of quadrants oscillating sinusoidally in depth. The locations of

individual dots on each stimulus frame were randomized, removing any

coherent monocular motion. The display alternated between this changing

disparity motion stimulus (12 s) and a spatiotemporally scrambled (STS)version (12 s). (b) MT+ responses were selective for changing disparity–

defined three-dimensional motion. Time series of fMRI response in MT+

(left) and V1 (right) are shown for both changing disparity motion (blue)

and temporally-scrambled changing disparity-motion (red) versus STS motion

(gray). The MT+ response was large to changing disparity motion, but small

to STS motion. V1 responses were generally small, but note the inversion of

response to changing disparity and STS motion. (c) MT+ showed the clearest

selectivity for changing disparity–defined three-dimensional motion. Average

fMRI response amplitudes across all visual areas are shown for changing

disparity motion, spatially scrambled changing disparity motion (SS),

temporally scrambled changing disparity motion (TS) and anti-correlated

changing disparity motion (all versus spatiotemporally scrambled changing

disparity motion). MT+ responses were large to changing disparity motion,

smaller to scrambled changing disparity motion, negligible to temporally

scrambled changing disparity motion and greatly reduced when the changing

disparity display was anti-correlated. V3A responses were smaller, but

followed a similar pattern. LO responses were also selective for changing

disparity motion, but the response to temporally scrambled changing

disparity motion was relatively large, despite the fact that changing disparitymotion was not perceived in this condition.

1052 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

motion through depth (MTD) displays or motion within depth(MWD) displays that contained dots moving in frontoparallel planesoccupying the same range of disparities traversed by the MTD stimulus(Fig. 3a). Dot pairs oscillated sinusoidally with random starting phase.The only difference between the MTD and MWD stimuli was therelationship of a dot pair’s motion across the two eyes, which was eitherin phase (MWD) or anti-phase (MTD) (Fig. 3a).

Displays contained fully correlated dots, fully anti-correlated dots(Fig. 3a), or 50/50% or 25/75% correlated/anti-correlated dots (datanot shown). For the correlated displays, the dots either appeared to bemoving toward and away through depth (MTD) or side-to-side atvarious depths in the same three-dimensional volume (MWD). Forthe anti-correlated displays, the dots either appeared to be movingtoward and away through depth (MTD, just as for the correlateddisplays) or side-to-side with little or no relative depth (MWD, asanti-correlation degraded the disparity signal). This loss of percepts ofthree-dimensional position in light of preserved percepts of three-dimensional motion qualitatively replicates our previous psycho-physical isolation of the IOVD cue using anti-correlation14.

MT+ responses to MWD decreased as binocular correlation decreased(Fig. 3b). In contrast, MT+ responses to MTD did not (and mayhave increased slightly; Fig. 3b). These slopes were both different fromzero and from one another (95% confidence interval on the differencebetween the slopes = [0.16, 0.54]; see Fig. 3 for additional confidenceintervals). If MT+ responses were only driven by net monocular motion,we should have observed identical responses across all conditions. Indeed,the same analysis for V1 showed that neither the MTD nor the MWDslope was significantly different from zero and, notably, they were not

different from one another (95% confidence interval on the differencebetween the slopes = [–0.15, 0.25]; Fig. 3b).

The smaller MWD response to anti-correlated dots than to correlateddots is consistent with observed MT+responses in previous fMRI studiesof disparity processing23. This reduction probably reflects the reduceddisparity-based component of the MT+ response. Given that the onlydifference between the MWD and MTD stimuli is whether the dots werein phase across the eyes, we interpret the lack of a similar MTD reductionas revealing a distinct contribution of the (remaining) IOVD signal.

We analyzed the responses across visual areas (omitting the inter-mediate correlations; Fig. 3c). The significant interaction betweencorrelation (correlated/anti-correlated) and direction of motion(MTD/MWD) in MT+ (ANOVA, F1, 35 ¼ 9.36, P ¼ 8 � 10�3) supportsour parametric analysis of slopes (Fig. 3c). V3A showed a similarpattern (F1,35 ¼ 6.92, P ¼ 0.018), but no other areas did (all P 4 0.1).

This pattern indicates that MT+ carries a signal associated withMTD even when strong disparity-based (changing disparity) cues areweak or absent. It is therefore likely that, in addition to responding tothe changing disparity cue (Experiment 2), MT+ can also be driven bythe IOVD cue to three-dimensional motion.

Experiment 4: selective adaptation to three-dimensional motion

If, as the preceding experiments indicate, MT+ is important in three-dimensional motion perception, we should expect to find directionselectivity for three-dimensional motion. We performed an fMRI-adapta-tion experiment to test whether adaptation to unidirectional three-dimensional motion (containing both changing disparity and IOVDcues) differentially affected fMRI responses to a subsequent probe stimulusthat moved in either the same or opposite direction as the adaptor. In eachscanning session an adapting random dot field moved either toward oraway from the observer. After 100 s of adaptation, observers viewed a seriesof 7.5-s trials, each containing a 4-s top-up adaptor followed by a 1-s probe(with same or opposite directions; Fig. 4a).

We observed strong direction-selective adaptation; MT+ responseswere smaller when the probe stimulus moved in the same direction asthe adaptor compared with when the adaptor moved in the oppositedirection (F1,1476 ¼ 85.72, P E 0; Fig. 4b). Pronounced effects wereevident in every observer in every scanning session (SupplementaryFig. 2) and were consistent with results from ongoing psychophysicalstudies (Supplementary Fig. 3). We observed similar, but smaller,effects in V3A (F1,1476 ¼ 14.94, P ¼ 0.0001) and LO (F1,1476 ¼ 13.86,

a

b

c

Cor

rela

ted

Ant

i-cor

rela

ted

MTD MWD

L

L

L

LR

R R

R

MWD

MTDCorrelated

Percentage of correlated dots

Anti-correlatedCorrelatedAnti-correlated

V1 V2 V3 V3A hV4 LO MT+0

0.5

1.0

100 50 25 0

0.8

1.0

1.2

MT+ V1

100 50 25 0

0

0.2

MTD

MWD

Motion type

fMR

I res

pons

e (∆

% B

OLD

)fM

RI r

espo

nse

(∆%

BO

LD)

Figure 3 Stimuli and results for Experiment 3. (a) Stimulus conditions. Each

pair of patches represents a left and right eye half image. In MTD conditions,

dots in each dichoptic pair moved in opposite horizontal directions in the

two eyes. In MWD conditions, dots in each pair moved in the same horizontal

directions (in phase). The range of disparities was the same in MWD and

MTD conditions. In correlated conditions, dot pairs were of the same contrast

polarity. In anti-correlated conditions, dot pairs were of opposite contrast

polarity. In anti-correlated displays, MWD stimuli appeared flat, but MTDstimuli still appeared to contain motion through depth. (b) MT+ responses

to motion containing IOVDs were robust to anti-correlation. Average fMRI

response amplitudes are shown for MTD and MWD as a function of the

percentage of correlated dots (100%, perfect correlation; 0%, complete

anti-correlation). MT+ responses (left) to MWD decreased as the percentage

of correlation decreased (95% confidence interval, on the slope = [–0.33,

–0.03], units of percentage BOLD change per 100% correlation), but MT+

responses to MTD increased with anti-correlation. V1 responses (right)

showed no such dependence. (c) MT+ showed selectivity for IOVD-biased

three-dimensional motion. Average fMRI response amplitudes across all areas

are shown. MT+ responses to MWD decreased as the percentage of correlated

dots decreased, whereas responses to MTD increased. A similar, but weaker,

pattern of response was evident in V3A. Trends in other areas were

non-significant (see text for statistical information).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1053

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

P ¼ 0.0002), smaller, but still significant, effects in V2 and V3(P o 0.05), and no reliable effects in V1 (P ¼ 0.58; Fig. 4b). Thisconfirms the importance of MT+ in the processing of three-dimen-sional motion (as well as the secondary roles of V3A and LO), asidentified in the first three experiments.

We confirmed that these effects were not inherited from the adapta-tion of earlier monocular two-dimensional direction mechanisms bypresenting the left and right eye’s probe stimuli in immediate succession,rather than simultaneously (Fig. 4a). These staggered probes yieldedsmall adaptation effects of constant magnitude across areas (marginal inMT+, F1,1477 ¼ 3.63, P ¼ 0.057). The pattern of response is consistentwith downstream areas inheriting V1 adaptation, but the magnitude ofmonocular two-dimensional adaptation cannot account for the largethree-dimensional motion adaptation effects that we observed in MT+.

To quantify selectivity for three-dimensional motion, we computeda direction-selectivity index reflecting the difference in response tothe opposite- versus same-direction probes normalized by their sum(Online Methods). The direction-selectivity index for three-dimensionalmotion was largest in MT+ (Fig. 4c). Other areas showed lesser degreesof selectivity, which appeared to grow in magnitude across the hierarchy(that is, V1 o V2 o V3 o V3A, LO o MT+). We plotted the selectivityindices for the staggered probes (Fig. 4c). Notably, the three-dimensionaldichoptic index was markedly different from the staggered (monocular)index only in MT+ (P o 10�4 bootstrapped, P 4 0.05 in all other areas).

The distinction between the patterns of three-dimensional dichopticand two-dimensional monocular direction-selective adaptation enhancethe conclusions drawn from the prior blocked design experiments: MT+carries strong three-dimensional direction signals. Although V3A and LOmay also process three-dimensional signals, their selectivity could not bedistinguished from sensitivity to monocular motions. The results in twosubregions in MT+ (putative human MTand MST) were generally similaracross experiments (Supplementary Fig. 4). The greater sensitivity of theadaptation protocol did reveal a somewhat larger three-dimensional

motion direction-selectivity index in MST (0.15 ± 0.02) than in MT(0.09 ± 0.02). The conservative method for identification of MTand MSTleft many MT+ voxels unassigned (Online Methods), so the overall MT+selectivity is not the simple average of MT and MST selectivity.

DISCUSSION

Our results demonstrate that the two binocular signals produced bythree-dimensional motion, changing disparity and IOVD, are pro-cessed in human MT+. Despite extensive investigation of the role ofhuman MT+ (as well as macaque MT and MST) in the perception ofboth frontoparallel (two dimensional) motion and static dispar-ity3,7,24,25, electrophysiological evidence for three-dimensional motionsensitivity has been middling. Some physiological work in monkeysand cats has suggested the existence of cells tuned to trajectories ofthree-dimensional motion26–31, and a study on motion parallax haspointed to a role for motion-selective MT neurons in three-dimensionalperception32, but recent work has not observed responses that areunambiguously selective for three-dimensional motion in macaqueMT5,7. It seems incongruous, however, that a brain area might haveevolved simply to process two-dimensional motion and static depth.

Our results differ from some prior electrophysiological studies, whichoptimized stimuli to suit the two-dimensional frontoparallel selectivity ofindividual neurons5. In contrast, we optimized our stimuli to yield robustthree-dimensional motion percepts to human observers. A previousfMRI study did find responses to structure from two-dimensionalmotion in human MT+33. Another fMRI study emphasized responsesto cyclopean stereomotion anterior to canonical human MT+20. Ourresults are difficult to compare directly with those of that study because ofdifferences in stimuli (we only used smooth changing-disparity motionalternating with a variety of scrambled controls and separately investi-gated the contribution of coherent monocular motions) and differencesin data analysis (we first identified and subdivided MT+ and then madeadditional measurements in these ROIs to quantify the responses tomotion though depth). Although our findings in MT+ do not precludethe involvement of areas outside MT+, they do demonstrate clear three-dimensional motion selectivity in canonical MT+ and motivate futurework to assess the contributions of subregions and adjacent areas20.

Our results also suggest a role for V3A and lateral occipital areas(LO1/LO2), which are less well understood in both human and monkeymodels. Our findings suggest that they deserve further consideration inthe processing of disparity and velocity signals. In contrast, we observedvery little evidence for three-dimensional motion selectivity in V1.

0 6 12 18–0.25

–0.25

0

0.25

0.50

0 6 12 18

0

0.25

0.50

V1 V2 V3 V3A LO MT+

0

0.1

0.2

a

b

c

3D dichoptic 2D monocular

L R

Tim

e L R

Same Opposite Same OppositeA

dapt

erP

robe

MT+

fMR

I res

pons

e (∆

% B

OLD

) V1

Time (s)

3D motion directionSameOpposite

Adaptation type3D dichoptic2D monocular

hV4

Dire

ctio

n-se

lect

ivity

inde

x

Figure 4 Stimuli and results for Experiment 4. (a) Stimulus conditions.

Each pair of circular patches represents the left and right eye half images.

In the three-dimensional (3D) dichoptic condition (left), a top-up three-

dimensional motion adaptor containing random dots appeared to drift either

away (shown) or toward the subject. The subject subsequently viewed a probe

stimulus that moved in either the same or opposite direction. In the two-

dimensional (2D) monocular conditions (right), the two eyes’ probe stimuli

were presented in immediate temporal succession, rather than simul-taneously; the two conditions were otherwise identical. (b) MT+ responses

reflect direction-selective adaptation to MTD. Time series of fMRI response in

MT+ (left) and V1 (right) are shown for probe stimuli during the three-

dimensional motion adaptation condition. In MT+, but not V1, responses to

opposite-direction probes (orange) were larger than to same-direction probes

(yellow). Error bars are smaller than plotting symbols. (c) MT+ showed the

strongest three-dimensional motion direction selectivity. Average direction-

selectivity indices across all areas for three-dimensional motion (orange bars)

and monocular two-dimensional motion (cyan bars) adaptation. Three-

dimensional motion direction selectivity was strongest in MT+, intermediate

in areas V3A and LO, and not evident in V1. Direction selectivity to two-

dimensional monocular motion was weak, but present across visual areas. Error

bars indicate bootstrapped 68% confidence intervals (equivalent to ± 1 s.e.m.).

1054 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Although our measurements could have lacked the sensitivity orresolution to identify three-dimensional motion selectivity in V1, wesuggest that V1 most likely serves to extract the building blocks34–36 forthree-dimensional motion computations that are ultimately performedin extrastriate areas. Subcortical pathways may also contribute to thesecomputations, possibly bypassing V1 altogether.

There are numerous reasons why the pattern of results across allfour experiments cannot be explained by sensitivity to disparity alone5.First, the first three experiments contrasted, in different ways, responsesto stimuli with the same distributions of disparities. Second, weobserved patterns of responses to three-dimensional motion acrossareas that differed from previous fMRI characterizations of sensitivityto static disparities37. Third, the results of our adaptation experimentdemonstrated direction selectivity by comparing responses to stimulithat differed solely in their direction of three-dimensional motion,while containing identical patterns of disparities.

Furthermore, we observed a response reduction to anti-correlatedstimuli in MT+, but only for stimuli that did not contain IOVDs. Thus,the effects of anti-correlation were different for displays that did and didnot contain IOVDs, revealing a neural computation for three-dimensionalmotion that cannot be explained in terms of sensitivity to disparity alone.

In summary, our results reveal an important and previously over-looked role for human MT+ in three-dimensional motion processing.Our results provide evidence that MT+ responds to both changingdisparities and IOVDs that specify movement through three-dimensionalspace. This three-dimensional motion sensitivity derived from bothdisparity- and velocity-based cues motivates reconsideration of thewell-studied sensitivities to frontoparallel motion and static disparityin MT and MST. Likewise, our observation of an IOVD signal in MT+may explain why several computations in MT appear to be mono-cular38,39. Such eye-specific processing may not be an inherited artifactof two-dimensional motion processing, but rather a reflection of acomputational strategy for computing interocular velocity differences.Canonical neurophysiological models of MT motion processing40–42

will need extension to incorporate the comparison of monocularvelocity signals, as well as sensitivity to changing disparities over time.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe thank D. Ress and T. Czuba for assistance with magnetic resonance imagingand comments on the manuscript and P. Neri for commenting on an earlierversion of the manuscript. This work was supported by a National ScienceFoundation CAREER Award (BCS-0748413), a Mind Science FoundationResearch Grant, and a pilot scanning grant from the University of Texas atAustin Imaging Research Center to A.C.H., and a Netherlands Organisationfor Scientific Research grant to B.R. (2006/11353/ALW).

AUTHOR CONTRIBUTIONSThe authors jointly conceived the project, conducted the experiments and wrote themanuscript. B.R. programmed the visual displays and conducted the data analyses.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Born, R.T. & Bradley, D.C. Structure and function of visual area MT. Annu. Rev. Neurosci.28, 157–189 (2005).

2. Cumming, B.G. & DeAngelis, G.C. The physiology of stereopsis. Annu. Rev. Neurosci.24, 203–238 (2001).

3. DeAngelis, G.C. & Newsome, W.T. Organization of disparity-selective neurons inmacaque area MT. J. Neurosci. 19, 1398–1415 (1999).

4. Huk, A.C., Dougherty, R.F. & Heeger, D.J. Retinotopy and functional subdivision ofhuman areas MT and MST. J. Neurosci. 22, 7195–7205 (2002).

5. Maunsell, J.H. & Van Essen, D.C. Functional properties of neurons in middle temporalvisual area of the macaque monkey. II. Binocular interactions and sensitivity to binoculardisparity. J. Neurophysiol. 49, 1148–1167 (1983).

6. Albright, T.D. Direction and orientation selectivity of neurons in visual area MT of themacaque. J. Neurophysiol. 52, 1106–1130 (1984).

7. DeAngelis, G.C. & Newsome, W.T. Perceptual ‘‘read-out’’ of conjoined direction anddisparity maps in extrastriate area MT. PLoS Biol. 2, e77 (2004).

8. Nguyenkim, J.D. & DeAngelis, G.C. Disparity-based coding of three-dimensionalsurface orientation by macaque middle temporal neurons. J. Neurosci. 23, 7117–7128(2003).

9. Smith, A.T. & Wall, M.B. Sensitivity of human visual cortical areas to the stereoscopicdepth of a moving stimulus. J. Vis. 8, 1–12 (2008).

10. Julesz, B. Foundations of Cyclopean Perception (The University of Chicago Press,Chicago, 1971).

11. Cumming, B.G. & Parker, A.J. Binocular mechanisms for detecting motion-in-depth.Vision Res. 34, 483–495 (1994).

12. Brooks, K.R. Interocular velocity difference contributes to stereomotion speed percep-tion. J. Vis. 2, 218–231 (2002).

13. Harris, J.M. & Rushton, S.K. Poor visibility of motion in depth is due to early motionaveraging. Vision Res. 43, 385–392 (2003).

14. Rokers, B., Cormack, L.K. & Huk, A.C. Strong percepts of motion through depth withoutstrong percepts of position in depth. J. Vis. 8, 1–10 (2008).

15. Fernandez, J.M. & Farell, B. Motion in depth from interocular velocity differencesrevealed by differential motion aftereffect. Vision Res. 46, 1307–1317 (2006).

16. Beverley, K.I. & Regan, D. Evidence for the existence of neural mechanisms selectivelysensitive to the direction of movement in space. J. Physiol. (Lond.) 235, 17–29 (1973).

17. Shioiri, S., Saisho, H., & Yaguchi, H. Motion in depth based on inter-ocular velocitydifferences. Vision Res. 40, 2565–2572 (2000).

18. Heeger, D.J., Boynton, G.M., Demb, J.B., Seidemann, E. & Newsome, W.T. Motionopponency in visual cortex. J. Neurosci. 19, 7162–7174 (1999).

19. Qian, N., Andersen, R.A. & Adelson, E.H. Transparent motion perception as detection ofunbalanced motion signals. I. Psychophysics. J. Neurosci. 14, 7357–7366 (1994).

20. Likova, L.T. & Tyler, C.W. Stereomotion processing in the human occipital cortex.Neuroimage 38, 293–305 (2007).

21. Norcia, A.M. & Tyler, C.W. Temporal frequency limits for stereoscopic apparent motionprocesses. Vision Res. 24, 395–401 (1984).

22. Cumming, B.G. & Parker, A.J. Responses of primary visual cortical neurons to binoculardisparity without depth perception. Nature 389, 280–283 (1997).

23. Bridge, H. & Parker, A.J. Topographical representation of binocular depth in the humanvisual cortex using fMRI. J Vis 7, 1–14 (2007).

24. Bradley, D.C., Qian, N. & Andersen, R.A. Integration of motion and stereopsis in middletemporal cortical area of macaques. Nature 373, 609–611 (1995).

25. Dodd, J.V., Krug, K., Cumming, B.G. & Parker, A.J. Perceptually bistable three-dimensional figures evoke high choice probabilities in cortical area MT. J. Neurosci.21, 4809–4821 (2001).

26. Akase, E., Inokawa, H. & Toyama, K. Neuronal responsiveness to three-dimensionalmotion in cat posteromedial lateral suprasylvian cortex. Exp. Brain Res. 122, 214–226(1998).

27. Toyama, K., Komatsu, Y., Kasai, H., Fujii, K. & Umetani, K. Responsiveness of Clare-Bishop neurons to visual cues associated with motion of a visual stimulus in three-dimensional space. Vision Res. 25, 407–414 (1985).

28. Cynader, M. & Regan, D. Neurons in cat visual cortex tuned to the direction of motion indepth: effect of positional disparity. Vision Res. 22, 967–982 (1982).

29. Zeki, S.M. Cells responding to changing image size and disparity in the cortex of therhesus monkey. J. Physiol. (Lond.) 242, 827–841 (1974).

30. Poggio, G.F. & Talbot, W.H. Mechanisms of static and dynamic stereopsis in foveal cortexof the rhesus monkey. J. Physiol. (Lond.) 315, 469–492 (1981).

31. Regan, D. & Cynader, M. Neurons in cat visual cortex tuned to the direction of motionin depth: effect of stimulus speed. Invest. Ophthalmol. Vis. Sci. 22, 535–550 (1982).

32. Nadler, J.W., Angelaki, D.E. & DeAngelis, G.C. A neural representation of depth frommotion parallax in macaque visual cortex. Nature 452, 642–645 (2008).

33. Orban, G.A., Sunaert, S., Todd, J.T., Van Hecke, P. & Marchal, G. Human cortical regionsinvolved in extracting depth from motion. Neuron 24, 929–940 (1999).

34. Ponce, C.R., Lomber, S.G. & Born, R.T. Integrating motion and depth via parallelpathways. Nat. Neurosci. 11, 216–223 (2008).

35. Barlow, H.B., Blakemore, C. & Pettigrew, J.D. The neural mechanism of binocular depthdiscrimination. J. Physiol. (Lond.) 193, 327–342 (1967).

36. Hubel, D.H. & Wiesel, T.N. Receptive fields and functional architecture of monkeystriate cortex. J. Physiol. (Lond.) 195, 215–243 (1968).

37. Tsao, D.Y. et al. Stereopsis activates V3A and caudal intraparietal areas in macaques andhumans. Neuron 39, 555–568 (2003).

38. Tailby, C., Majaj, N. & Movshon, T. Binocular integration of pattern motion signals by MTneurons and by human observers [Abstract]. J. Vis. 7, 95a (2007).

39. Majaj, N.J., Tailby, C. & Movshon, J.A. (2007). Motion opponency in area MT of themacaque is mostly monocular [Abstract]. J. Vis. 7, 96a (2007).

40. Rust, N.C., Mante, V., Simoncelli, E.P. & Movshon, J.A. How MTcells analyze the motionof visual patterns. Nat. Neurosci. 9, 1421–1431 (2006).

41. Perrone, J.A. & Thiele, A. A model of speed tuning in MT neurons. Vision Res. 42,1035–1051 (2002).

42. Simoncelli, E.P. & Heeger, D.J. A model of neuronal responses in visual area MT. VisionRes. 38, 743–761 (1998).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1055

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSSubjects. fMRI data were collected in three subjects (the three authors, males

aged 33–44), all with normal or corrected-to-normal vision. All were experi-

enced psychophysical observers in both motion and depth experiments.

Experiments were undertaken with the written consent of each subject and

all procedures were approved by the University of Texas at Austin Institutional

Review Board and the University of Texas at Austin Imaging Research Center

safety guidelines for magnetic resonance research. Each observer participated in

a total of 19 scanning sessions.

Magnetic resonance imaging. Magnetic resonance imaging was performed at

the University of Texas at Austin Imaging Research Center on a GE Signa HD

3T scanner, using a GE 8-channel phased array head coil.

Anatomical imaging. A whole-brain anatomical volume at 1-mm3 resolution

was acquired for each subject. Brain tissue was segmented into gray matter,

white matter and cerebrospinal fluid by an automated algorithm followed by

manual refinement43. The inplane anatomical volume from each fMRI session

was then co-registered with this whole brain anatomical volume with an

accuracy of B1 mm (ref. 44).

fMRI. For Experiments 1–3, we used a two-shot spiral BOLD fMRI sequence45

((2.2-mm)3 voxels, 3-s volume acquisition duration, repetition time ¼ 1.5 s,

echo time ¼ 30 ms, 73 degree flip angle), with 24 pseudo-coronal slices starting

at the occipital pole and continuing anterior to provide occipital and parietal

coverage. Data from the first 24 s of each fMRI scan were discarded to minimize

transient effects of both the scanner’s magnetic saturation and observer’s

hemodynamics.

For Experiment 4, we used a two-shot spiral sequence with coarser spatial

resolution and finer temporal resolution ((3.2-mm)3 voxels, 1.5-s volume

acquisition duration, repetition time ¼ 750 ms, echo time ¼ 30 ms, 56 degree

flip angle), with 14 quasi-axial slices covering the posterior visual cortices,

oriented roughly parallel with the calcarine sulcus. Data from the first and last

two trials (15 s) of each fMRI scan were discarded to ensure that magnetic and

hemodynamic steady-state had been reached and so that BOLD responses for

each trial were similarly convolved with the responses from preceding and

following trials.

In each session, T1-weighted inplane anatomical images were acquired

for co-registration. Each fMRI time series of each voxel was highpass filtered

(0.015 Hz cutoff frequency) to compensate for the slow signal drift typical in

fMRI signals. Each voxel’s time series was also divided by its mean intensity

to convert the data from arbitrary image intensity units to percent signal

modulation (% BOLD signal change) and to compensate for variations

in mean image intensity across space. The resulting time series were then

averaged across voxels in a given cortical area or ROI (see below) restricted to

the portion of the gray matter that was responsive to the stimulus after

reference scan restriction.

Defining the visual areas. The fMRI data were analyzed in each of the

visual cortical areas separately for each subject. Mapping of visual areas V1,

V2, V3, V3A, hV4, LO, MT+, MT and MST was performed in separate

experimental sessions for each subject using standard techniques (see Supple-

mentary Fig. 5)4,46,47.

Statistical data analysis. For Experiments 1–3, we computed the average cycle

time series responses shown in Figures 1b and 2b by averaging the fMRI time

series in each visual area across stimulus cycles (six per scan), repeats (three per

condition) and subjects (three) at each time point. S.e.m. were computed for

each time point, and were generally only slightly larger than the size of the

plotting symbols. We estimated the fMRI response amplitudes (Figs. 1c, 2c and

3b,c) using standard Fourier-based block-design methods implemented by the

Stanford VISTA group software package, mrVISTA/mrLoadRet18 (see Supple-

mentary Fig. 5).

For the analysis of response as a function of binocular correlation (Fig. 3b),

we fit the response amplitudes to each condition (MTD and MWD) with a line

using a least-squares algorithm. We then tested whether the slopes (or differences

between slopes) were significantly different from zero using bootstrapping; we

resampled the data with replacement 104 times, fit each resampled dataset with a

line and calculated the 95% confidence intervals on the resulting distributions of

bootstrapped slopes. Experiment 4 used an event-related adaptation protocol and

analysis scheme that was identical to that described previously48.

General visual apparatus, stimuli and task. Subjects viewed three-dimensional

motion displays using a custom-built magnetic resonance–compatible mirror

stereoscope. Stimuli were presented on a 36.5- � 27.4-cm rear projection screen

(19.6 � 14.6 degrees of visual angle) illuminated by a 60 Hz LCD projector

(linearized with 109.4-cd m�2 mean luminance). The viewing distance was

109 cm. Dots in the display were anti-aliased to obtain subpixel position

accuracy. We used a pair of adjustable (pan/tilt) mirrors (one above each eye)

to project the appropriate part of the stimulus display onto each eye. A mirror

that is inclined with respect to the visual axis will produce a rotation of the

projected image when slanted about its vertical axis. Observers visually

corrected this rotation by counter-rotating the two halves of the visual display

through the display software, implemented using the Psychophysics Toolbox49.

A pair of adjustable septums (one immediately between the mirrors and one

further along the optical path) blocked each eye’s (undesired) view of the

contralateral half image. This apparatus has been used and discussed in

additional detail in our prior psychophysical work dissociating the changing

disparity and IOVD cues14.

In all of the experiments that involved moving dot pairs (Experiments 1, 2

and 4), the two dots in each pair had the same center of motion, regardless

of whether they were presented in the same monocular half image or across

the two half images. In Experiments 1–3, 32 dots pairs were presented in a

7.6-degree circular gray background aperture. In each monocular half image,

half of the dots were white and half of the dots were black. Each dot in one

half image was paired with a dot in the other half image; there were no

unpaired dots in any displays. Dot size was 0.15 degree and binocular disparity

ranged between ±22 arcmin. The average dot density was 0.83 dots per degree2.

Dots were spaced in the circular apertures to minimize spatial interactions

across nearby dot pairs by requiring that the center of motion of each dot

(corresponding to its 0 arcmin disparity point) was at least 0.88 degrees away

from the same point of any dot belonging to another pair. A fixation point with

nonius lines and four reference points outside of the aperture were present at

all times to aid with proper vergence, fixation and half-image rotation. In

Experiment 4, a number of features were modified to maximize the psycho-

physical adaptation effects. Details are reported below and in Supplementary

Figure 6.

Behavioral task. To control attention across experimental conditions, the

observers performed a challenging visual dot-change task throughout all fMRI

runs50. Observers were required to press a button within 1 s of detecting a

change in contrast polarity (Experiments 1 and 3) or the color (Experiments 2

and 4) of a randomly selected dot pair. Time of dot pair change was drawn

from a 4-s average exponential distribution with a 1-s minimum and 12-s

maximum (Supplementary Fig. 6).

Stimulus details. The stimuli in Experiment 1 (Fig. 1a) served to contrast

signals generated by stimuli that yield three-dimensional motion percepts

(dichoptically separated opposite-direction horizontal motion) with signals

generated by stimuli that do not produce three-dimensional motion percepts

(monocularly paired or vertical opposite-direction motion).

Moving dots were presented in circular patches in the left and right eye half

images in four different experimental conditions. Dots were presented in pairs,

and dots in each pair oscillated in opposite directions on a sinusoidal trajectory

producing a maximum separation of ±22 arcmin. In horizontal conditions, the

dots moved along horizontal trajectories. In vertical conditions, the dots moved

along vertical trajectories. In dichoptic conditions, the oppositely moving dots

in each pair were split between the left and the right half images. In mono-

cular conditions, each oppositely moving dot pair was randomly positioned in

either the left or the right eye. The horizontal-dichoptic stimulus consisted of

changing horizontal binocular disparities and horizontal IOVDs and produced

a percept of three-dimensional motion through depth (MTD). The other

conditions contained similar motion properties (that is, the same orientation,

but with monocular dot pairings, or the same dichoptic dot pairings, but with

different orientation) and were generally perceived as twinkling moving dots

(that is, motion opponent displays18,19).

NATURE NEUROSCIENCE doi:10.1038/nn.2343

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Experiment 2. The stimuli in Experiment 2 (Fig. 2a) served to contrast

disparity signals that produce strong three-dimensional motion percepts

(changing disparity over time) with disparity signals that do not produce

strong three-dimensional motion percepts (spatially scrambled, temporally

scrambled or anti-correlated versions of the stimuli).

Dots were presented in the circular patches in the left and right eye half

images in four different experimental conditions. In each condition, dot pair

disparity varied between ±22 arcmin and the 12-s presentation of the experi-

mental stimulus alternated with a 12-s spatiotemporally scrambled version in a

blocked design. In the main changing disparity motion condition, the binocular

disparities of the dots were systematically manipulated over time to create a

spatiotemporal pattern of quadrants oscillating at 1.6 Hz toward and away from

the observer. This was accomplished by assigning the same disparity to all dots

in a quadrant and sinusoidally modulating that disparity over time. Adjacent

quadrants moved in opposite directions (anti-phase). Critically, the image-plane

locations of individual dots on each stimulus frame (60-Hz refresh rate) were

randomized. This removed any coherent monocular motion from the display,

but preserved the systematic changes in each quadrant’s disparity over time20.

Likewise, the presence of opposite directions of motion through depth across

the quadrants removed full-field cyclopean motion and any corresponding drive

to change vergence. We compared BOLD responses to spatially scrambled,

temporally scrambled and anti-correlated versions of the same stimulus.

Experiment 3. The stimuli in Experiment 3 (Fig. 3a) served to contrast retinal

motion signals (IOVDs) that produce strong three-dimensional motion per-

cepts (MTD stimuli) with retinal motion signals that do not produce strong

three-dimensional motion percepts (MWD stimuli). We used binocular anti-

correlation to investigate the contribution of IOVDs, when the contribution of

changing disparity was strongly reduced.

We presented 32 dots pairs in the circular patches of the left and right eye

half images in eight different experimental conditions. In MTD conditions, the

dots in each dichoptic pair moved horizontally in opposite directions (anti-

phase; note that this stimulus is identical to the dichoptic-horizontal stimulus

described in Experiment 1). In MWD conditions, the dots in each dichoptic

pair moved in the same horizontal direction (same phase), at a randomly

selected disparity. This manipulation is equivalent to changing in-phase dot

pair sinusoidal motion (MWD) to anti-phase motion (MTD). The monocular

motions and range of disparities were identical across MTD and MWD

conditions. In correlated conditions, the dots in each dichoptic pair were of

the same contrast polarity. In anti-correlated conditions, the dots in each

dichoptic pair were of opposite contrast polarity.

Note that anti-correlated MWD displays appeared perceptually flat, but anti-

correlated MTD displays were still perceived as containing three-dimensional

motion. This selective robustness of MTD to anti-correlation reveals the

contribution of the IOVD cue. We have used the same stimulus manipulation

to dissociate the IOVD cue from the changing disparity cue in quantitative

psychophysical experiments14.

Experiment 4. This experiment employed a direction-selective adaptation

protocol to investigate the selectivity of cortical areas to specific directions

of three-dimensional motion. Each three-dimensional motion adaptation

scanning session started with 100 s of initial adaptation to unidirectional

three-dimensional motion (either toward or away). The adaptation direction

was constant throughout each scanning session, which contained 12 adapta-

tion runs. Each scanning run contained 36 7.5-s trials in random order. A

trial contained 4 s of top-up adaptation, a 1.25-s interstimulus interval (mean

gray screen), a 1-s probe stimulus and a final 1.25-s intertrial interval. In 14 of

the 36 trials, the probe stimuli moved in the same direction as the adaptor, in

another 14 it moved in the opposite direction, and in the remaining eight trials

the adaptor was followed by a blank (no probe) stimulus. These adaptor-blank

trials were used to estimate the baseline response to the adaptor (see

Supplementary Fig. 5). Each subject participated in two scanning sessions,

one with three-dimensional motion adaptation toward and one with adapta-

tion away.

The monocular two-dimensional adaptation control experiment presented

the same stimuli with slightly different timing. The 4-s top-up adaptation was

followed by a 0.75-s interstimulus interval and a 2-s probe stimulus, during

which the two monocular images were shown for 1 s each in random order

(followed by a 0.75-s intertrial interval). The three-dimensional adaptation and

two-dimensional monocular (staggered probe) control experiments were thus

nearly identical, except for the temporally staggered (and thus monocular)

presentation of the probe stimulus in this control experiment. Each subject

participated in two scanning sessions, one for each direction of three-dimen-

sional motion adaptation (toward and away).

43. Wandell, B.A., Chial, S. & Backus, B.T. Visualization and measurement of the corticalsurface. J. Cogn. Neurosci. 12, 739–752 (2000).

44. Nestares, O. & Heeger, D.J. Robust multiresolution alignment of MRI brain volumes.Magn. Reson. Med. 43, 705–715 (2000).

45. Glover, G.H. & Lai, S. Self-navigated spiral fMRI: interleaved versus single-shot. Magn.Reson. Med. 39, 361–368 (1998).

46. Engel, S.A. et al. fMRI of human visual cortex. Nature 369, 525 (1994).47. Sereno, M.I. et al. Borders of multiple visual areas in humans revealed by functional

magnetic resonance imaging. Science 268, 889–893 (1995).48. Larsson, J., Landy, M.S. & Heeger, D.J. Orientation-selective adaptation to first- and

second-order patterns in human visual cortex. J. Neurophysiol. 95, 862–881 (2006).49. Brainard, D.H. The Psychophysics Toolbox. Spat. Vis. 10, 433–436 (1997).50. Huk, A.C., Ress, D. & Heeger, D.J. Neuronal basis of the motion aftereffect reconsidered.

Neuron 32, 161–172 (2001).

doi:10.1038/nn.2343 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Sensory transformations and the use of multiplereference frames for reach planning

Leah M M McGuire & Philip N Sabes

The sensory signals that drive movement planning arrive in a variety of ‘reference frames’, and integrating or comparing

them requires sensory transformations. We propose a model in which the statistical properties of sensory signals and their

transformations determine how these signals are used. This model incorporates the patterns of gaze-dependent errors that we

found in our human psychophysics experiment when the sensory signals available for reach planning were varied. These results

challenge the widely held ideas that error patterns directly reflect the reference frame of the underlying neural representation and

that it is preferable to use a single common reference frame for movement planning. We found that gaze-dependent error patterns,

often cited as evidence for retinotopic reach planning, can be explained by a transformation bias and are not exclusively linked to

retinotopic representations. Furthermore, the presence of multiple reference frames allows for optimal use of available sensory

information and explains task-dependent reweighting of sensory signals.

Humans use various sensory signals when interacting with the environ-ment. We can reach to pick up a coin that we see in front of us or transferthe coin from one hand to another without looking. Using multiplesensory modalities for planning similar movements is potentially prob-lematic, as different sensory signals arrive in different reference frames.Specifically, early visual pathways represent stimulus location relativeto current gaze location (a retinotopic representation), whereas pro-prioceptive signals represent hand location relative to the shoulderor trunk (a body-centered representation). To utilize these signals,some of them must be transformed between reference frames. Althoughsensory transformations may appear mathematically simple, we foundthat transformations can incur a cost by adding bias and variability1,2

into the transformed signal. Thus, sensory transformations likelyinfluence the flow of information in motor planning circuits.

It has been argued that transforming sensory signals into a commonrepresentation would simplify reach planning3–7, and many researchershave attempted to characterize this representation. Many psychophysicalstudies have focused on the patterns of reach error using the assumptionthat the reference frame for movement planning directly determines thespatial pattern of errors. Such studies argue for both retinotopic5,8–11 andhand- or body-centered12,13 planning. Studies of primate physiology andhuman functional magnetic resonance imaging have also found evidencefor a range of neural representations for movement planning14–23. Thesedisparate results suggest that a single common representation for reachplanning is unlikely22–26. We argue that the presence of noisy sensorytransformations makes it advantageous to represent movement planssimultaneously in multiple reference frames.

We examined how the representations of movement plan dependon the available sensory inputs, focusing on the effects of gaze location.

A well-studied gaze-dependent error is the retinal eccentricity effect,where subjects overestimate the distance between the gaze locationand a visually peripheral target when pointing to the target27. Becausethese errors are most parsimoniously described as an overshoot in aretinotopic reference frame, they are cited as evidence for retinotopicreach planning8–10. We found that the magnitude and direction ofgaze-dependent errors depended heavily on the available sensoryinformation. These results were interpreted using a model of move-ment planning in which sensory signals are combined in a statisticallyprincipled manner in two separate reference frames. Our modelprovides an explanation for these gaze-dependent reach errors: theyarise when sensory information about target location is transformedbetween representations using an internal estimate of gaze directionthat is biased toward the target. We thus demonstrate that spatialpatterns of reach errors do not necessarily directly reflect the referenceframe of the underlying neural representation.

RESULTS

Measuring gaze-dependent reach errors

We first examined how the pattern of reach errors depends on thesensory signals available during the planning and execution of amovement. Specifically, we manipulated information about targetlocation and initial hand position, the two variables that are neededto compute a movement vector. Target information was varied byhaving subjects reach to either visual targets (VIS), proprioceptivetargets (the left index finger, PROP) or targets consisting of simulta-neous visual and proprioceptive signals (left index finger with visualfeedback, VIS + PROP). Information about initial hand position wasvaried by having subjects reach either with (feedback) or without

Received 23 March; accepted 1 June; published online 13 July 2009; doi:10.1038/nn.2357

W. M. Keck Center for Integrative Neuroscience, Department of Physiology, and the Neuroscience Graduate Program, University of California, San Francisco, California, USA.Correspondence should be addressed to P.N.S. (sabes@phy.ucsf.edu).

1056 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

(NoFB) visual feedback of the right (reaching) hand before movementonset, although feedback was never available during the movement. Foreach of the six resulting trial types, we measured movement errors assubjects reached to nine different target locations with gaze held on oneof two fixation points (Supplementary Fig. 1).

A comparison of reach endpoints for an example subject at themidline target shows how reach errors depended on the availablesensory signals (Fig. 1). The mean error changed markedly as afunction of gaze location and these gaze-dependent effects differedacross trial types. During VIS and VIS + PROP trials (Fig. 1a–d), reachendpoints were biased away from the gaze location (the retinal eccen-tricity effect) and the magnitude of the effect decreased with increasingsensory information (compare Figs. 1a and 1b–d). In PROP trials, asmall bias toward gaze location was observed instead (Fig. 1e,f). Thesepatterns were consistent across targets (Supplementary Section 1 andSupplementary Fig. 2). In addition to these gaze-dependent effects,there was a gaze-independent bias in reaching that could differ acrosstargets and trial types. Although there was a trend toward overshootingthe target, the pattern of this bias, measured across targets and trialtypes in a separate gaze-free trial condition, was idiosyncratic fromsubject to subject (Supplementary Section 1 and Supplementary Fig. 3),making these patterns difficult to interpret. We therefore focused on theconsistent gaze-dependent effects.

To isolate the gaze-dependent effects, we analyzed reach errors inpolar coordinates about the midpoint of the eyes and subtracted thegaze-free errors (Online Methods, Supplementary Section 1 andSupplementary Fig. 4). When we examined radial (depth) reach errors,we found a tendency to overshoot the target; however, these errors didnot differ between the two gaze locations (Supplementary Section 1and Supplementary Fig. 5). When we examined angular reach errors(Fig. 2), we observed both a general leftward bias (see SupplementarySection 1) and significant differences as a function of gaze location(P o 0.001 and P ¼ 0.003). Gaze-dependent effects were thus confinedto the angular reach errors.

We next examined how these gaze-dependent effects varied with trialtype (Fig. 2a–f). The retinal eccentricity effect was observed in VIS/NoFB trials (Fig. 2a); subjects made rightward (positive) reach errorswhen fixating to the left of the target and leftward (negative) reach

errors when fixating to the right of the target. A similar pattern wasobserved in VIS/feedback trials, but with smaller magnitude (Fig. 2b; seealso ref. 8). In contrast, during PROP trials, the reach endpoint wascloser to the fixation point; for example, leftward errors were made whenfixating to the left of the target (Fig. 2e,f). The gaze-dependent errorpatterns in the VIS + PROP trial types appeared to be a combination ofthose observed in VIS and PROP trials (Fig. 2c,d). For all trial types, theerrors for the two gaze-locations aligned qualitatively when angularerror was plotted in a retinotopic reference frame: that is, as a function oftarget relative to gaze (Fig. 2). This alignment might appear to supporta retinotopic representation for reach planning8–10. However, as thesepatterns differ markedly across trial types, they cannot be readilyexplained in terms of a fixed retinotopic bias, suggesting the need fora different explanation of these error patterns.

A planning model: integration across reference frames

We developed a model of reach planning that accounts for thepattern of gaze-dependent errors observed in our data. The modelhas two important features: the presence of multiple representationsfor movement planning and a bias in the transformation betweenthose representations.

The model begins with sensory inputs signaling target location,initial hand position and gaze location. As sensory signals are inher-ently variable, we modeled them as Gaussian likelihoods of truelocation given the sensory input, with likelihood variance reflectingthe reliability of the sensory modality28. Visual signals arrive in aretinotopic representation and proprioceptive signals arrive in a body-centered representation. Each available signal is then transformed into

38

a b

d

f

c

e

NoFB FB

34

38

34

38

34

38

34

38

34

38

34

–5 5

Fixation leftV

ISV

IS +

PR

OP

PR

OP

y (c

m)

x (cm)

Fixation right

0

–5 50

–5 50

–5 50

–5 50

–5 50

cm

Figure 1 Reach errors at the center target for an example subject for all trial

types. (a–f) Lines indicate mean reach error for each gaze position, ellipses

represent s.d., + indicates fixation points and � indicates reach targets. The

origin (not shown) is located directly below the midpoint of the eyes. FB,feedback trial.

Target position (deg)

–10 deg fixation +10 deg fixation

4

a b

d

f

c

e

NoFB FB

Ang

ular

rea

ch e

rror

(de

g)

–4

–20

VIS

P < 0.001 P < 0.001

P < 0.001

P < 0.001

P < 0.001

P = 0.003

VIS

+ P

RO

PP

RO

P

200

–20 200

–20 200

–20 200

–20 200

–20 200

0

4

–4

0

4

–4

0

4

–4

0

4

–4

0

4

–4

0

Figure 2 Average angular reach error across subjects for each trial condition.

(a–f) Negative values indicate reach endpoints to the left of the target, and

positive values indicate reach endpoints to the right of the target. Before

averaging, linearly interpolated gaze-free errors for each trial type were

subtracted (Supplementary Section 1 and Supplementary Fig. 4). Error

bars indicate standard errors. P values were determined with a paired

permutation test50. Insets show angular reach error plotted as a function

of target position relative to gaze (that is, in a retinotopic coordinate frame).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1057

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

the non-native reference frame (Fig. 3a). Because subjects’ headpositions were fixed during the experiment, this complex nonlineartransformation29 can be approximated simply by adding or subtractingthe gaze location: that is, by convolving their distributions (Fig. 3a andOnline Methods)1. However, because the internal estimate of gazelocation is also uncertain, this transformation adds variability tothe signals. When both sensory modalities are available, the nativeand transformed signals are integrated in both reference framerepresentations (VIS + PROP target condition; Fig. 3a). Movementvectors are then computed in each representation by convolving thetarget and initial hand distributions (subtraction). It should be notedthat, because the sensory transformation adds variability, each spatialvariable is more reliably represented in one or the other of theserepresentations depending on the availability and reliability of therelevant visual and proprioceptive signals (Supplementary Section 2).

The sensory transformations in our model also introduce a biasinto the estimates of transformed variables. In particular, we positthat the internal estimate of gaze direction used to transform targetlocation is biased toward the target. The bias was modeled as aBayesian prior on gaze location1 centered on the target. Because thetransformation consists of adding or subtracting gaze location,depending on the direction of the transformation, the prior effec-tively biases the transformed target estimate either away from ortoward the direction of gaze (Fig. 3a). Native (untransformed)target representations remain unbiased. Because the gaze prior iscentered on the estimated target location, the variance of the prior(Fig. 3b,c) was assumed to scale with the uncertainty of the internaltarget estimate leading to the bias patterns shown in Figure 3d.Therefore, the gaze-dependent errors in the model depend on therelative weighting of the various sensory signals, depending on boththe availability and reliability of sensory inputs and on the methodof reading out the final movement vector.

We considered three possible output schemes for reading the plannedmovement vector from the model (Supplementary Fig. 6). Either theretinotopic (RET) or body-centered (BODY) representations can eachbe read out directly. Alternatively, the two movement vector estimatescan themselves be combined to form an integrated (INTEG) readout.The contributions of the two representations to the INTEG readout

depend on their relative reliability. It should benoted that these calculations use the simplify-ing assumption that the signals being combinedare independent (Supplementary Section 2).

Each of the three potential readouts provided a quantitative predic-tion of reach errors as a function of hand, target and gaze locations. Theonly free parameters in the model were the variances of individualsensory inputs and the gaze priors (Online Methods). The values of theproprioceptive variances were based on previously reported values30.Four parameters remained: visual variance, gaze variance, and twoscaling factors relating the variance of the gaze prior to the variancein visual and proprioceptive target signals. We fit these parameters to

Fixation point Native signalTransformed signalIntegrated signal

VIS targeta

b c d10

Var

ianc

e (c

m2 )

–30 30Target eye (deg)

0 –30

–4

4

0

30Target eye (deg)

0 –30 30

Target eye (deg)

RET BODY

5 cm

BODY RET

0

6

2

10

Var

ianc

e (c

m2 )

Gaz

e bi

as (

deg)

6

2

p(TB) = p(TR) ⊗ p(G)p(TR) = p(TB) ⊗ p(–G)

p(TR) = p(TB) ⊗ p(–G)p(TB) = p(TR) ⊗ p(G)

VIS + PROP target PROP target

Ret

inot

opic

repr

esen

tatio

nB

ody-

cent

ered

repr

esen

tatio

nTr

ansf

orm

atio

n

True visual target

True prop target

σ2g σ2

g

σ2prior

σ2prior

σ2post σ2

post

4a

c

e

b

d

f

NoFB FB

Ang

ular

rea

ch e

rror

(de

g)

–4–20

VIS

VIS

+ P

RO

PP

RO

P

200

–20 200

–20 20

RETBODYINTEG

0

–20

Target position (deg)

200

–20 200

–20 200

0

4

–4

0

4

–4

0

4

–4

0

4

–4

0

4

–4

0

Figure 4 Model fits for gaze-dependent reach errors. (a–f) Black and gray

lines show mean (standard error) errors across subjects for each trial

condition, after subtracting the overall mean separately for each of the six

trial types. Colored lines show best-fit model predictions. Solid lines, gaze

right; dashed lines, gaze left.

Figure 3 The bias and variance injected

by sensory transformations in the model.

(a) Examples of the bias and variance that

arise during transformation of target information.

Circles represent the posterior distribution

of target location (95% confidence limits).

Transformed signals have greater uncertainty and

are biased either toward (retinotopic) or away from(body centered) the gaze location. In VIS + PROP

trials, the native and transformed signals are

integrated into a combined estimate of location

(filled circles). (b,c) Posterior variance in

estimated gaze location (transformation variance,

black), gaze likelihood variance (gray) and

variance of gaze prior (dashed) (see Online

Methods) for target transformation from

retinotopic to body-centered space (b) and

target transformation from body-centered to

retinotopic space (c). (d) Bias in gaze estimate

used in target transformations. Values in all

panels were computed from INTEG model fit.

1058 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

the gaze-dependent errors (Fig. 2) after mean correction (see OnlineMethods) using least-squares regression (Supplementary Section 2,Supplementary Table 1 and Supplementary Fig. 7).

Model fits of constant and variable reach errors

We first considered how well the three output models fit the observedpatterns of gaze-dependent errors (Fig. 4a–f). When the model was fitwith a single-representation readout, RET or BODY, it failed to predicterrors for all trial types. This is because only transformed target signalscontain a gaze-dependent error, and thus when the readout was in thenative reference frame of a unimodal target, no error was predicted(Fig. 4a,b,e,f). Although both readouts had errors for bimodal targets(VIS + PROP; Fig. 4c,d), errors in the RET readout were a result ofthe transformed proprioceptive signal and would therefore be in thewrong direction (Fig. 3a). Thus, for the RET readout, the fitting proce-dure nullified the effect of the transformed proprioceptive signal byforcing the model to rely only on vision: that is, by driving the visualvariance parameter toward zero (Supplementary Table 1). In contrast,when the movement vector representations were combined in theINTEG readout, the model performed well across all trial types. Thisreadout captured both changes in the magnitude of gaze-dependenterrors (VIS versus VIS+PROP and feedback versus NoFB) and the signreversal that we observed with PROP targets. Itaccomplished this by differentially weightingthe two reference frames across trial conditions,with both representations making substantivecontributions (Supplementary Section 2 andSupplementary Fig. 8). Indeed, a scheme thatsimply switches between the RET and BODYreadouts depending on task would not capturethe data well. First, both the RET and BODYreadouts failed to capture the differences in themagnitude of gaze-dependent errors that weobserved between feedback and NoFB condi-tions. Second, to predict the observed errors,the switching scheme would need to rely pre-dominantly on the more variable transformedsignals, rather than on native signals, a sub-optimal arrangement.

In addition to fitting the gaze-dependenterror patterns, our model predicted the

differences in movement variability across trial types (Fig. 5a). Becausecomputations in the model were assumed to be noise free, model outputvariability was entirely the result of variability in the sensory inputs,shaped by the model computations, and did not require any additionalparameter fitting. The INTEG model fit provided an accurate predictionfor the changes in output variance across trial types and was better thanthe two single-representation fits. These predictions came from separateparameter fits for each readout. However, the model parameterspresumably reflect actual variances in the neuronal representations ofsensory inputs. Using a single set of variances, for example, the INTEGfit, we looked at how variability in the movement plan depends onreadout. We found that the INTEG readout generally yielded a lowervariance estimate (Fig. 5b), as it made better use of all available sensorysignals (although the extent of this advantage depended on the statisticalproperties of the sensory transformations; Supplementary Section 2and Supplementary Fig. 9). In contrast to the idea that a single coordi-nate frame should dominate movement planning3–10,15,18, this analysisillustrates that using multiple representations of a movement plan yieldsmore reliable performance across tasks.

Model predictions for previously published datasets

We tested the model, fit to our own dataset, on a similar, previouslypublished dataset8 containing visual target trials with an expandedrange of movements (that is, more start and gaze locations). These datashowed the retinal eccentricity effect, and, as described above, the effectmagnitude was smaller when visual feedback of the hand was available(data reproduced in Fig. 6a,b). In addition, there was a component ofthe reach error that correlated with the relative positions of the handand target (data reproduced Fig. 6c,d). Our model captured all of theseimportant features in the dataset (Fig. 6e–h).

Our model explains another very different empirical result, againwithout additional parameter fitting. We have previously shown thatvisual information about initial hand location is weighted more heavilywhen reaching to visual targets (as in VIS/feedback trials) than whenreaching to proprioceptive targets (as in PROP/feedback trials)2. Wepreviously proposed that this sensory reweighting was a result of thecost (for example, variability) incurred by sensory transformations2. Thepresent model makes this cost explicit, with quantitative predictionsof the angular error that should result from artificial shifts in visualfeedback. In both the empirical data and the INTEG readout predictions,visual feedback shifts had a weaker effect when reaching to propriocep-tive targets than when reaching to visual targets (Fig. 7a). This effect was

6RETBODYINTEGData

a bBest fit parameters for each model Same parameters for all models

4

2

Diff

eren

ce in

var

ianc

e fr

omm

ean

(deg

)

Mea

n pl

anni

ng v

aria

nce

(deg

)

0 4

6

8

2–2

–4

VIS

+ P

RO

P/N

oFB

VIS

+ P

RO

P/F

B

VIS

/NoF

B

VIS

/FB

PR

OP

/NoF

B

PR

OP

/FB

VIS

+ P

RO

P/N

oFB

VIS

+ P

RO

P/F

B

VIS

/NoF

B

VIS

/FB

PR

OP

/NoF

B

PR

OP

/FB

Figure 5 Reach variability. (a) Trial-type differences in angular reach

variance. Black lines represent mean (standard error) reach variance across

subjects. Values are derived from the average variance across subjects and

trial conditions in each trial type, with the overall mean subtracted. Colored

lines represent model predictions with each readouts fit parameters. (b) Mean

variability of each model readout scheme, as a function of trial type, using

fixed model parameters (INTEG fit).

–20 20 –20Eye relative to target (deg) Hand relative to target (deg)

20 –20 20 –20 20

–20

4NoFBa

e f g h

b c d30

Targets (deg)

2010

–10–20–30

0

FB NoFB FB

Dat

aIN

TE

G m

odelP

oint

ing

erro

r (d

eg)

0

–4

4

0

–4

4

0

–4

4

0

–4

4

0

–4

4

0

–4

4

0

–4

4

0

–4

20 –20 20 –20 20 –20 20

Figure 6 Model predictions for data from a previous study8 (a,b,e,f). Gaze-dependent pointing error.

(c,d,g,h) Pointing error as a function of initial hand position relative to the target. The average pointing

error from a previous study8 is shown in a–d. Our INTEG model predictions of pointing error with

parameters fit to our data are shown in e–h.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1059

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

quantified in terms of overall weighting of visual versus proprioceptivefeedback (Fig. 7b), which was much greater for VIS targets than forPROP targets. In the model, this re-weighting was a result of the tradeoffbetween the retinotopic and body-centered representations of the move-ment plan (Supplementary Fig. 8), evidenced by the fact that neither theRET nor BODY readout exhibited the effect. This result provides furthersupport for the use of multiple representations in movement planning.

Origins of the gaze bias

We found that a bias in the internal estimate of gaze location can accountfor the complex pattern of gaze-dependent reach errors that we observedacross trial types. This bias might arise as a result of either a ‘covert’saccade plan toward the target or a shift of attention to the target. Wetested these hypotheses by controlling the saccade target or the locusof attention independent of the reach target, but these manipulationsdid not alter the reach error pattern (Supplementary Section 3 andSupplementary Fig. 10). Alternatively, the bias could arise from a priorexpectation that reach targets tend to be foveated, reflecting the fact thateye and hand movements tend to be tightly linked31. Indeed, the modelbias was implemented in this manner (see Online Methods).

Although the source of the gaze bias remains undetermined, wewere able to corroborate its presence using an independent measure.Specifically, we found that a visually peripheral reach target biased asubject’s estimate of straight ahead, and this bias was consistent witha shift in the estimated gaze direction32 toward the target (Supple-mentary Section 3 and Supplementary Fig. 11). This perceptual effect,similar to the errors observed in our reach experiment, was wellmodeled by a Bayesian prior on gaze location centered at the target.

DISCUSSION

Our study sought to test two widely held ideas in the field of sensori-motor control: that spatial patterns of errors for a given movementreflect the underlying neural representation7–13 and that a single refer-ence frame should dominate movement planning3–6,8–10,15,18,33. Weargue that neither of these ideas is correct. First, we found that a single,apparently retinotopic, pattern of reach errors could be explained by amodel in which multiple neural representations are used (for example,a combination of both retinotopic and body-centered reference frames).Second, we found that using more than one representation conferred anadvantage in terms of reduced planning variability.

Spatial patterns of reach errors, especially retinotopic or gaze-centeredpatterns, have been cited as evidence that the neural representation forreach planning occurs in a particular reference frame5,8–13,33. We observedsimilar error patterns, but found that their magnitude and directionalityvaried with the sensory signals that were available for movement plan-ning. This variation was inconsistent with a fixed bias in a single neuralrepresentation. We found that a bias in the transformation of targetinformation between representations could lead to gaze-dependent errorpatterns in both retinotopic and body-centered representations. In fact,the retinal eccentricity effect for visual targets arose in the model onlyfrom the body-centered representation, clearly illustrating that the spatialpattern of reach errors need not be a good indicator of the underlyingneural representation. More generally, we argue that when sensory signalsare used in a statistically optimal manner28,34–38, the same information iscontained in multiple neural representations and there is no requiredrelationship between the behavioral output and any single representationof the movement plan.

We propose that a biased transformation can account for gaze-dependent reaching errors, focusing on azimuthal (left-right) reacherrors and azimuthal gaze shifts. When gaze is shifted vertically(up-down), a similar pattern of vertical gaze-dependent errors isobserved39, which is qualitatively consistent with our model. We didnot observe gaze-dependent effects in the radial (depth) reach errorsof our subjects (Supplementary Fig. 4), probably because gaze wasnot varied in depth. When this manipulation is done, a complexpattern of gaze-dependent depth errors emerges33, presumably reflect-ing the complex binocular, three-dimensional geometry of the eyes29.Modeling these patterns will require substantially more complex modelrepresentations and transformations.

Although we have focused on gaze-dependent errors, a variety ofother error patterns are amenable to similar analysis. For example,several studies exploring the effect of head rotations on reach accuracyhave found errors in reaching toward the direction in which the head isoriented40,41, a pattern that is consistent with a bias of the perceivedmidline of the head toward the direction of gaze41. In addition, theleftward bias in reaching that we observed (Fig. 2) can be modeled as arightward bias in the proprioceptive estimate of the right hand(Supplementary Section 1). The idiosyncratic gaze-independenterror patterns exhibited by individual subjects (SupplementaryFig. 2) may also be explained by subject-specific biases on sensoryvariables42. All of these biases could be the result of prior expectationson sensory variables or their correlations36, although alternate sourcesof errors are plausible, such as impoverish neural representations43.

We argue that the variability of sensory transformations make itadvantageous to use multiple representations for a movement plan.First, we found that using multiple reference frames in a weightedfashion improved movement variability across trial types (Fig. 5b).Next, we found that only multiple reference frames in the model couldaccount for the reweighting of visual feedback of the hand as a resultof target type (Fig. 7)2. This is because the relative variability of the tworepresentations, and hence their contribution to the output, dependson the sensory modality of the target. However, it should be noted thatthe model does not account for all of the observed reweighting(Fig. 7). This may be explained by a reduced weighting of vision as aresult of a lack of belief by the subject that the visual cursor is at theirfinger44. Still, our results suggest that, although a single representationmay simplify the flow of information3–6, it does not make optimal useof that information in movement planning.

Previous studies have reported patterns of movement errors2,12,26 orgeneralization of motor learning45,46 that cannot be explained succinctlyin a single reference frame. Such error patterns often depend on the

30a bPROP/FB data

VIS/FB model

VIS/FB data

PROP/FB model

Right visual shiftLeft visual shift

10

1.0

0.8

0.6

0.4

0.2

0

–10

–30–100 0 100

Target angle (deg)VIS/FB PROP/FB

RETBODYINTEG

Ang

ular

err

or (

deg)

Wei

ghtin

g of

vis

ual i

nfor

mat

ion

Figure 7 Changes in sensory weighting with target modality. (a) Mean angular

error induced by artificial shifts in the visual feedback of the hand before

movement onset. Data are from a previous study2. Error bars represent

standard error. Model predictions use the INTEG readout, with parameters fit

to our data. (b) Relative weighting of visual versus proprioceptive information

about initial hand position in movement planning for reaches to VIS and

PROP targets. Error bars represent s.d. across subjects. Dashed lines show

model predictions for each readout scheme.

1060 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

availability of sensory signals2,12 and are explained in terms of changesin the underlying reference frame44,46 or an intermediate representationof movement planning26,45. Our results provide a different explanation;movements are always represented in multiple reference frames, indepen-dent of the task, and it is the statistical reliability of these representationsthat determines their relative weighting.

This model is consistent with the neurophysiological literature,where a variety of spatial representations have been observed acrossthe reach planning network24,47. Reach-related areas have been foundto have retinotopic coding15,48, head- or body-centered coding19,49, andhand- and shoulder-centered coding16,21–23, as well as mixed represen-tations14,16–18,21–23,49. The two separate representations of movementplan in our model could be found in subsets of these cortical areas.Alternatively, the same computation could be performed using aheterogeneous neural population that contains both retinotopic andbody-centered components, that is, mixed representation. Indeed,these implementations are at two ends of a continuum and physiolo-gical evidence seems to point to a middle ground in which all of thecortical areas for reach planning make use of mixed representations,but the parietal cortex has a more retinotopic character and the frontalcortex is more hand or body centered24,47. As our results indicate,however, the ultimate answer is likely to be found not by finer assays ofneural reference frames, but rather by comparing activity in these areasacross tasks with different sensory information4,14,49.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe would like to thank S.M. Beurze, S. van Pelt, W.P. Medendorp and S.J. Soberfor generously providing their data for model comparison. This work wassupported by the National Eye Institute (R01 EY-015679), the National Instituteof Mental Health (P50 MH77970) and the McKnight Endowment Fund forNeuroscience. L.M.M.M. was supported by a graduate fellowship from theNational Science Foundation.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Schlicht, E.J. & Schrater, P.R. Impact of coordinate transformation uncertainty onhuman sensorimotor control. J. Neurophysiol. 97, 4203–4214 (2007).

2. Sober, S.J. & Sabes, P.N. Flexible strategies for sensory integration during motorplanning. Nat. Neurosci. 8, 490–497 (2005).

3. Buneo, C.A. & Andersen, R.A. The posterior parietal cortex: sensorimotor interface forthe planning and online control of visually guided movements. Neuropsychologia 44,2594–2606 (2006).

4. Cohen, Y.E. & Andersen, R.A. A common reference frame for movement plans in theposterior parietal cortex. Nat. Rev. Neurosci. 3, 553–562 (2002).

5. Engel, K.C., Flanders, M. & Soechting, J.F. Oculocentric frames of reference for limbmovement. Arch. Ital. Biol. 140, 211–219 (2002).

6. Lacquaniti, F. & Caminiti, R. Visuo-motor transformations for arm reaching. Eur. J.Neurosci. 10, 195–203 (1998).

7. Soechting, J.F. & Flanders, M. Sensorimotor representations for pointing to targets inthree-dimensional space. J. Neurophysiol. 62, 582–594 (1989).

8. Beurze, S.M., Van Pelt, S. & Medendorp, W.P. Behavioral reference frames for planninghuman reaching movements. J. Neurophysiol. 96, 352–362 (2006).

9. Henriques, D.Y., Klier, E.M., Smith, M.A., Lowy, D. & Crawford, J.D. Gaze-centeredremapping of remembered visual space in an open-loop pointing task. J. Neurosci. 18,1583–1594 (1998).

10. Pouget, A., Ducom, J.C., Torri, J. & Bavelier, D. Multisensory spatial representations ineye-centered coordinates for reaching. Cognition 83, B1–B11 (2002).

11. McIntyre, J., Stratta, F. & Lacquaniti, F. Viewer-centered frame of reference for pointing tomemorized targets in three-dimensional space. J. Neurophysiol. 78, 1601–1618 (1997).

12. Carrozzo, M., McIntyre, J., Zago, M. & Lacquaniti, F. Viewer-centered and body-centeredframes of reference in direct visuomotor transformations. Exp. Brain Res. 129,201–210 (1999).

13. McIntyre, J., Stratta, F. & Lacquaniti, F. Short-term memory for reaching to visualtargets: psychophysical evidence for body-centered reference frames. J. Neurosci. 18,8423–8435 (1998).

14. Avillac, M., Deneve, S., Olivier, E., Pouget, A. & Duhamel, J.R. Reference frames forrepresenting visual and tactile locations in parietal cortex. Nat. Neurosci. 8, 941–949(2005).

15. Batista, A.P., Buneo, C.A., Snyder, L.H. & Andersen, R.A. Reach plans in eye-centeredcoordinates. Science 285, 257–260 (1999).

16. Batista, A.P. et al. Reference frames for reach planning in macaque dorsal premotorcortex. J. Neurophysiol. 98, 966–983 (2007).

17. Battaglia-Mayer, A. et al. Eye-hand coordination during reaching. II. An analysis of therelationships between visuomanual signals in parietal cortex and parieto-frontal asso-ciation projections. Cereb. Cortex 11, 528–544 (2001).

18. Buneo, C.A., Jarvis, M.R., Batista, A.P. & Andersen, R.A. Direct visuomotor transforma-tions for reaching. Nature 416, 632–636 (2002).

19. Lacquaniti, F., Guigon, E., Bianchi, L., Ferraina, S. & Caminiti, R. Representing spatialinformation for limb movement: role of area 5 in the monkey. Cereb. Cortex 5, 391–409(1995).

20. Medendorp, W.P., Goltz, H.C., Vilis, T. & Crawford, J.D. Gaze-centered updating of visualspace in human parietal cortex. J. Neurosci. 23, 6209–6214 (2003).

21. Pesaran, B., Nelson, M.J. & Andersen, R.A. Dorsal premotor neurons encode the relativeposition of the hand, eye and goal during reach planning. Neuron 51, 125–134 (2006).

22. Wu, W. & Hatsopoulos, N. Evidence against a single coordinate system representation inthe motor cortex. Exp. Brain Res. 175, 197–210 (2006).

23. Wu, W. & Hatsopoulos, N.G. Coordinate system representations of movement directionin the premotor cortex. Exp. Brain Res. 176, 652–657 (2007).

24. Burnod, Y. et al. Parieto-frontal coding of reaching: an integrated framework. Exp. BrainRes. 129, 325–346 (1999).

25. Caminiti, R., Ferraina, S. & Mayer, A.B. Visuomotor transformations: early corticalmechanisms of reaching. Curr. Opin. Neurobiol. 8, 753–761 (1998).

26. Carrozzo, M. & Lacquaniti, F. A hybrid frame of reference for visuo-manual coordination.Neuroreport 5, 453–456 (1994).

27. Bock, O. Contribution of retinal versus extraretinal signals towards visual localization ingoal-directed movements. Exp. Brain Res. 64, 476–482 (1986).

28. Knill, D.C. & Pouget, A. The Bayesian brain: the role of uncertainty in neural coding andcomputation. Trends Neurosci. 27, 712–719 (2004).

29. Blohm, G. & Crawford, J.D. Computations for geometrically accurate visually guidedreaching in 3-D space. J. Vis. 7, 1–22 (2007).

30. van Beers, R.J., Sittig, A.C. & Denier van der Gon, J.J. The precision of proprioceptiveposition sense. Exp. Brain Res. 122, 367–377 (1998).

31. Land, M.F. & Hayhoe, M. In what ways do eye movements contribute to everydayactivities? Vision Res. 41, 3559–3565 (2001).

32. Balslev, D. & Miall, R.C. Eye position representation in human anterior parietal cortex.J. Neurosci. 28, 8968–8972 (2008).

33. Van Pelt, S. & Medendorp, W.P. Updating target distance across eye movements indepth. J. Neurophysiol. 99, 2281–2290 (2008).

34. Ernst, M.O. & Banks, M.S. Humans integrate visual and haptic information in astatistically optimal fashion. Nature 415, 429–433 (2002).

35. Helbig, H.B. & Ernst, M.O. Optimal integration of shape information from vision andtouch. Exp. Brain Res. 179, 595–606 (2007).

36. Kording, K.P. & Wolpert, D.M. Probabilistic mechanisms in sensorimotor control.Novartis Found. Symp. 270, 191–198; discussion 198–202, 232–197 (2006).

37. van Beers, R.J., Sittig, A.C. & Gon, J.J. Integration of proprioceptive and visual position-information: an experimentally supported model. J. Neurophysiol. 81, 1355–1364 (1999).

38. Yuille, A.L. & Bulthoff, H.H. Bayesian decision theory and psychophysics. in Perceptionas Bayesian Inference 123–136 (Cambridge University Press, Cambridge, UK, 1996).

39. Henriques, D.Y. & Crawford, J.D. Direction-dependent distortions of retinocentric spacein the visuomotor transformation for pointing. Exp. Brain Res. 132, 179–194 (2000).

40. Henriques, D.Y. & Crawford, J.D. Role of eye, head and shoulder geometry in theplanning of accurate arm movements. J. Neurophysiol. 87, 1677–1685 (2002).

41. Lewald, J. & Ehrenstein, W.H. Visual and proprioceptive shifts in perceived egocentricdirection induced by eye position. Vision Res. 40, 539–547 (2000).

42. Vindras, P., Desmurget, M., Prablanc, C. & Viviani, P. Pointing errors reflect biases in theperception of the initial hand position. J. Neurophysiol. 79, 3290–3294 (1998).

43. Soechting, J.F. & Flanders, M. Errors in pointing are due to approximations insensorimotor transformations. J. Neurophysiol. 62, 595–608 (1989).

44. Kording, K.P. & Tenenbaum, J.B. Causal inference in multisensory integration. in NIPS737–744 (MIT Press, Cambridge, Massachusetts, 2006).

45. Ahmed, A.A., Wolpert, D.M. & Flanagan, J.R. Flexible representations of dynamics areused in object manipulation. Curr. Biol. 18, 763–768 (2008).

46. Kluzik, J., Diedrichsen, J., Shadmehr, R. & Bastian, A.J. Reach adaptation: whatdetermines whether we learn an internal model of the tool or adapt the model of ourarm? J. Neurophysiol. 100, 1455–1464 (2008).

47. Battaglia-Mayer, A., Caminiti, R., Lacquaniti, F. & Zago, M. Multiple levels of representa-tion of reaching in the parieto-frontal network. Cereb. Cortex 13, 1009–1022 (2003).

48. Cohen, Y.E. & Andersen, R.A. Reaches to sounds encoded in an eye-centered referenceframe. Neuron 27, 647–652 (2000).

49. Mullette-Gillman, O.A., Cohen, Y.E. & Groh, J.M. Eye-centered, head-centered andcomplex coding of visual and auditory targets in the intraparietal sulcus. J. Neurophysiol.94, 2331–2352 (2005).

50. Good, P.I. Permutation Tests: A Practical Guide to Resampling Methods for TestingHypotheses, 2nd edn. (Springer-Verlag, New York, 2000).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1061

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSExperimental setup. Subjects were seated in a virtual reality setup (Supple-

mentary Fig. 1). The right arm rested on top of a thin (6 mm) and rigid

horizontal table. The left arm remained under the table. When used as a reach

target, the left index finger touched the underside the table with the wrist

supine. Thus, the two hands never came in contact. The location of both index

fingers was monitored using an infrared tracker (Optotrak 3020, Northern

Digital). The subjects’ view of their arms was blocked by a mirror through

which they viewed a rear-projection screen (Supplementary Fig. 1). This

provided the illusion that visual objects appeared to lie in the plane of the table.

The rig was enclosed in black felt and the room was darkened to minimize

additional visual cues. Head movements were lightly restrained with a chin rest

and eye movements were monitored with an ISCAN infrared eye tracker.

Task design. Nine potential reach targets were located on the table on a 35-cm

arc centered at the point directly below the midpoint of the two eyes

(Supplementary Fig. 1). The targets (8-mm radius green disks, when visual)

were located at ±201, ±151, ±61, ±21 and 01 relative to midline. Two gaze

fixation points (5-mm radius red disks) were located on this arc at ±101. Visual

feedback was given with an 8-mm radius disk centered on the index finger:

white for the right hand and blue for the left.

All trials consisted of four steps. First, subjects moved their right index finger

to a fixed starting location. On feedback trials, the start location was indicated

with a 10-mm radius green disk and visual feedback of the right hand was

illuminated. On NoFB trials, neither feedback nor target were visible and

subjects were guided to the start location using the arrow field method, which

provides no feedback of absolute hand position2. Second, when the right index

finger came to rest within 10 mm of the start location, one of the fixation

points appeared. Subjects were required to maintain fixation at the fixation

point for the remainder of the trial. Third, after fixation, the reach target was

specified. For VIS and VIS + PROP trials, the target disk appeared. In VIS +

PROP trials, feedback of the left hand appeared and subjects moved the left

index finger to the target, and the target disk was then extinguished leaving the

blue feedback disk. For PROP trials, an arrow field was used to guide the

unseen left hand to the unseen target location. Fourth, after target specification,

there was a 500-ms delay before an audible go tone was played and subjects

reached to the target. On feedback trials, both feedback and the start disk were

extinguished at the go tone and remained off for the rest of the trial. Finally,

subjects were required to hold the final reach position for 500 ms. Subjects

practiced the various trial types before beginning the experiment.

Eight subjects (two females and six males) participated in the experiment.

Subjects were right-handed, had no known neural or motor deficits, had

normal or corrected-to-normal vision, and gave written informed consent

before participation. The experiment was divided into two sessions, which were

performed on different days to minimize fatigue. One session contained

feedback trials and the other contained NoFB trials, and session order was

randomized across subjects. Each session contained six repetitions of each of

the 54 conditions (three target types � nine targets � two fixation points), for a

total of 324 trials (not including error trials, which were repeated). These trials

were followed by trials in which gaze location was unconstrained. This set

consisted of six repetitions of nine trial conditions (three target types � three

targets), bringing the total number of trials for each session to 378. The order of

presentation across conditions was randomized in each repetition.

Data analysis. For each trial, reach endpoint was defined as the position at

which movement speed first fell to 5 mm s�1. Reach targets and endpoints were

converted into polar coordinates about an origin located directly below the

midpoint of the two eyes. Angular reach error is the angular difference between

the endpoint and target, with positive values indicting reach endpoints to the

right of the target and negative values indicating reach endpoints to the left of

the target. For the plots, permutation tests and model fitting, the angular reach

errors were corrected by subtracting the linearly interpolated free-gaze errors

(separately for each subject and trial type) to minimize the effects of idiosyn-

cratic gaze-independent errors while preserving the relationship between error,

gaze and target (see Supplementary Figs. 3–5). The significance of gaze-

dependent effects was tested by a paired permutation test of main effect of

gaze location50.

Model. Our model of reach planning describes how statistical representations

of sensory inputs are used to compute a movement vector plan. Five sensory

signals are potentially available, modeled as independent Gaussian probability

distributions centered on the true locations, X, that is, Gaussian likelihoods,

N(X,s 2), with an isotropic covariance matrix of s2I: vision of the right

fingertip is represented by p fvjFRð ÞBNðFR;s2vÞ, proprioception of the right

fingertip is represented by pðfp j FBÞBNðFB; s2pÞ, vision of the target is

represented by pðtv jTRÞBNðTR; s2vÞ, proprioception of the target is repre-

sented by pðtp jTBÞBNðTB; s2pÞ and the felt gaze position is represented by

pðg jGÞBNðG; s2gÞ (equation (1)). Lower-case variables are sensory signals,

with subscripts denoting sensory modality (v for visual and p for proprioceptive).

Upper-case variables are true locations with subscripts denoting reference frame

(R for the retinotopic location and B for body-centered location).

When a signal is unavailable in a given trial type (for example, fv in

NoFB trials), the likelihood is the uniform distribution. The likelihood

represents variability in a sensory signal x given the true location X. The

computations in the model, however, depend on uncertainty in X given x,

that is, on the posterior distributions pðX j xÞ. Bayes’ rule relates these

two distributions:

pðX j xÞ / pðx jXÞpðXÞ ð2Þ

where p(X) represents prior information about the location. We usually assume

that the prior is flat, so the posterior is proportional to the likelihood (although

a nontrivial prior is described below).

All of the computations in the model make locally optimal use of the signals

required for a computation, assuming that those signals are independent.

Indeed, only two operations are performed by the network: signal integration

and addition (or subtraction). Integration is the process of combining

information about variable X from two signals x1 and x2. By Bayes rule

(equation (2)) and the definition of independence, the integrated posterior is

the product of the two input distributions:

pðX j x1; x2Þ / pðX j x1ÞpðX j x2Þ ð3Þ

The resulting posterior is also Gaussian, with mean and variance:

mXjx1; x2¼

mXjx1

s2Xjx1

+mXjx2

s2Xjx2

!s2

Xjx1 ;x2; s2

Xjx1 ;x2¼ 1

s2Xjx1

+1

s2Xjx2

!�1

ð4Þ

In equation (4) and below, mXjx and s2Xjx represent the mean and variance of

the distribution pðXjxÞ, respectively. If either input distribution is uniform, that

is, if a sensory signal is absent, the integrated posterior is equal to the other

input. Note that the integrated variance is smaller than either of the input

variances. The second operation, addition, is where the network computes the

posterior of a variable Z ¼ X + Y from input signals x and y. In this case, the

output posterior is given by convolving the inputs:

pðZ j y; xÞ ¼ pðX j xÞ � pðY j yÞ ð5Þ

The result is also a Gaussian:

mZjx;y ¼ mXjx + mY jy ; s2Zjx;y ¼ s2

Xjx + s2Y jy ð6Þ

Note that for both integration and addition, the output mean is a weighted

sum of the input means, with either constant (unity) weights or weights that

depend on the input variances.

Given the sensory signals described in equation (1), the model first builds

internal representations of fingertip and target locations in both retinotopic and

body-centered reference frames. These representations integrate all available

sensory signals, requiring the transformation of non-native signals. For example,

in computing the retinotopic representation of target, the proprioceptive signal

tp must be transformed. Because head position is fixed, TR ¼ TB � G, and the

transformation follows equation (5):

pðTR j tp; gÞ ¼ pðTB j tpÞ � pð�G j gÞ ð7Þ

Parallel transformations convert fp into a retinotopic representation and tv

and fv into body-centered representations. This can yield two independent

estimates of the same variable, which are then integrated according to

NATURE NEUROSCIENCE doi:10.1038/nn.2357

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

equation (3). The retinotopic representation of target, for example, has the

posterior distribution:

pðTR j tv; tp; gÞ / pðTR j tvÞ pðTR j tp; gÞ ð8Þ

Of course, in VIS or PROP trials, one of the input distributions is uniform.

The model next computes retinotopic and body-centered representations of

the desired movement vector. Because the true value of the instructed move-

ment vector is MV ¼ T – F, these computations also follow equation (5):

pðMVU j tv; tp; fv; fp; gÞ ¼ pðTU j tv; tp; gÞ � pðFU j fv; fp; gÞ;U ¼ R; B

ð9Þ

Note that both of the inputs to equation (9) depend on gaze. If the same

estimate of gaze is used in all transformations, then the inputs are not fully

independent, and equation (6) is only an approximate solution (see Supple-

mentary Section 2 and Supplementary Fig. 9).

Finally, the model selects a planned movement vector using one of three

readout schemes. The RET and BODY readouts are just the mean values of the

MVR and MVB posteriors, respectively, from equation (9). For the INTEG

readout, the model integrates these two posteriors according to equation (3), as

if they were independent:

pðMVINTEG j tv; tp; fv; fp; gÞ / pðMVRjtv; tp; fv; fp; gÞ p ðMVB j tv; tp; fv; fp; gÞð10Þ

Note that these three readouts are maximum a posteriori estimates given the

respective posteriors. This computation is only correct when the input signals

are independent, which is not always true (see Supplementary Section 2).

The final component of the model is a bias in the transformation of target

position between reference frames. We modeled this bias as a systematic mis-

estimation of gaze location as a result of a Bayesian prior. The prior takes the

form of a Gaussian distribution, p(G)BN(tprior, s2prior). The mean of the prior

distribution tprior is a Gaussian random variable with mean TB and variance

proportional to that of the transformed target variable. The variance of the prior,

s2prior, is a product of the variance of the target distribution being transformed

and a scaling factor that depends on the modality of the transformed signal (see

Supplementary Table 1). This prior effects the gaze estimate used for transform-

ing target locations, but not the transformation of finger locations.

Because the model is composed entirely of integration (equation (3)) and

addition (equation (5)) operations, all expected values, including the move-

ment vector readouts, can be written as weighted sums of the means of the

initial sensory inputs, with coefficients depending only on the variances of

those signals (equations (4) and (6)). The trial-by-trial variances of the readouts

(Fig. 6) can be computed from these coefficients and the input variances in

equation (1).

Model fitting. The only model parameters are the sensory variances in equation

(1) and the variance of the gaze prior, s2prior. The proprioceptive variances

were set a priori on the basis of previously published estimates30. The variance

of visual signals, s2v, is assumed to scale linearly with the distance of the

stimulus from center of gaze and this scale factor is the first free parameter.

The variance of the gaze signal, s2g is the second. The variance of the gaze

prior, s2prior, is assumed to scale with variance of the target variable being

transformed. Two scale factors, one for each target modality, make up the

remaining free parameters. We fit these four free parameters to the average

angular movement errors, after mean correction (Fig. 4). The model generates

Cartesian movement vectors, which are then converted into polar coordi-

nates to obtain angular errors. The fitting procedure minimized the sum

square prediction error across trial conditions using the Matlab optimization

toolbox (function fmincon, Mathworks). Optimization was repeated 100 times

with random initial parameter values, and the final parameters were largely

insensitive to initial values.

Modeling other datasets. The model predictions for previously published

datasets (Figs. 6 and 7) used the model parameters fit to our own dataset with

the INTEG readout. When modeling the reach errors for a previous study8

(Fig. 6), we translated their target array into our workspace and rescaled the

spacing of the targets to maintain the same azimuthal separation. The data for

Figure 7 come from the gaze-fixed trials in another previous study2 (Supple-

mentary Fig. 3).

doi:10.1038/nn.2357 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Prefrontal and striatal dopaminergic genes predictindividual differences in exploration and exploitation

Michael J Frank1–3, Bradley B Doll1–3, Jen Oas-Terpstra4 & Francisco Moreno4

The basal ganglia support learning to exploit decisions that have yielded positive outcomes in the past. In contrast, limited

evidence implicates the prefrontal cortex in the process of making strategic exploratory decisions when the magnitude of potential

outcomes is unknown. Here we examine neurogenetic contributions to individual differences in these distinct aspects of motivated

human behavior, using a temporal decision-making task and computational analysis. We show that two genes controlling striatal

dopamine function, DARPP-32 (also called PPP1R1B) and DRD2, are associated with exploitative learning to adjust response

times incrementally as a function of positive and negative decision outcomes. In contrast, a gene primarily controlling prefrontal

dopamine function (COMT) is associated with a particular type of ‘directed exploration’, in which exploratory decisions are made

in proportion to Bayesian uncertainty about whether other choices might produce outcomes that are better than the status quo.

Quantitative model fits reveal that genetic factors modulate independent parameters of a reinforcement learning system.

Individuals differ in their choices and neural responses whenconfronted with decision uncertainty1,2. Some people are motivatedby having achieved desirable outcomes and are driven to work harderto attain even better ones, whereas others are primarily motivated toavoid negative outcomes3. However, individuals often don’t knowwhich outcomes should be considered positive until they comparethem to those obtained from other decision strategies (for example, doyou choose to return to the same failsafe sushi restaurant or to try anew one because it might be even better?). This classic problem ofwhether to sample other options or maintain the current strategyfor maximizing reward is known as the exploration/exploitationdilemma4–7. Here we examine neurogenetic contributions to exploi-tative and exploratory behavior.

In part, individual differences in personality variables are thought toreflect different parameters within the dopaminergic motivationalsystem8. Dopaminergic genetic components that alter function in thestriatum (and indirectly its interactions with frontal cortex9) differ-entiate between individuals who are more adept at learning frompositive as compared to negative decision outcomes, via modulationof striatum and its interactions with frontal cortex9–11. Specifically, afunctional polymorphism within the DARPP-32 gene—whereby car-riers of two copies of the ‘T’ allele (T/T carriers) show greater DARPP-32 mRNA expression than those with at least one copy of the ‘C’ allele(C carriers)9—is predictive of ‘Go learning’ to reproduce behaviors thatyield positive outcomes10. The DARPP-32 protein is highly concen-trated in the striatum, is phosphorylated by D1 dopamine receptorstimulation, and is required for striatal D1 receptor–mediated synapticplasticity and behavioral reward learning12–14. Although DARPP-32 isalso present in D2-containing neurons, stimulation of D2 receptors

dephosphorylates DARPP-32 and does not mediate its effects onreward learning13. Conversely, polymorphisms within the DRD2 genepredictive of striatal D2 receptor density are associated with ‘NoGolearning’ to avoid behaviors that yield negative outcomes10,11: indivi-duals with two copies of the DRD2 ‘T’ allele (T/T carriers) have greaterstriatal D2 receptor density15. These findings converge with the notionthat dopamine has a key role in reinforcement learning16 and, inparticular, that dopamine acts in the striatum to support learning frompositive and negative outcomes via D1 and D2 receptors in separateneuronal striatonigral and striatopallidal populations17,18. The findingsalso converge with rodent data showing that the transition to exploi-tative behavior is associated with the development of highly stabilizedstriatal firing patterns19.

Although the role of striatal dopamine in reinforcement exploitationis relatively well established, the neurobiological correlates of explora-tion are far less developed. Computational considerations suggest thatan adaptive heuristic is to explore in proportion to one’s uncertaintyabout the consequent outcomes4,6,7,20. Such computations mightdepend on neuromodulation within the prefrontal cortex (PFC)7.Functional neuroimaging evidence implicates anterior and orbitalPFC in computations of uncertainty2,21 and in the making of explora-tory decisions in a reinforcement learning environment6. Further,models and experimental data suggest that orbital PFC representsreward magnitudes, which are required to compute the expected valueof decisions, especially over delays6,22–24. At the genetic level, the geneCOMT, encoding catechol-O-methyltransferase (COMT), substantiallyaffects PFC dopamine levels and, in turn, PFC-dependent cognitivefunction25. COMT is an enzyme that breaks down dopamine; an alleleof COMTencoding valine at chr22:18331271 (known as the ‘val’ allele)

Received 13 March; accepted 28 April; published online 20 July 2009; doi:10.1038/nn.2342

1Departments of Cognitive & Linguistic Sciences, 2Psychology and 3Psychiatry, Brown Institute for Brain Science, Brown University, Providence, Rhode Island, USA.4Department of Psychiatry, University of Arizona, Tucson, Arizona, USA. Correspondence should be addressed to M.J.F. (michael_frank@brown.edu).

1062 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

is associated with greater enzymatic efficacy, and therefore lower PFCdopamine levels, than a methionine-encoding (‘met’) allele. Theenzyme has a comparatively minor role in striatum owing to itsrelatively sparse expression and to the presence of potent dopaminetransporters and autoreceptors25–28.

We assessed these motivational components, including exploitation,exploration, and probability versus magnitude learning, within a single‘‘temporal utility integration task’’29. We hypothesized that geneticmarkers of striatal dopaminergic function (DARPP-32 and DRD2)would be predictive of response time adaptation to maximize rewards.In contrast, we hypothesized that a genetic marker of prefrontaldopaminergic function (COMT) would be predictive of uncertainty-based exploration and enhanced representation of reward magnitudes.

RESULTS

Temporal integration of expected value

Participants observed a clock arm that completed a revolution over5 s, and they could stop the clock with a key press in an attempt towin points. Rewards were delivered with a probability and magni-tude that varied as a function of response time (RT, Fig. 1). Thefunctions were designed such that the expected value (EV; prob-ability � magnitude) increased, decreased or remained constant(IEV, DEV or CEV) with increasing response times (Fig. 1). Thus, inthe DEV condition, faster RTs yielded more points on average, suchthat performance benefited from Go learning to produce furtherspeeded RTs. In contrast, fast RTs in the IEV condition yieldedbelow-average outcomes, such that performance benefited fromNoGo learning to produce adaptively slower responding. The CEVcondition was included for a within-subject baseline RT measure forcomparison with IEV and DEV. Because all RTs are equivalentlyrewarding in the CEV condition, participants’ RT in this conditioncontrolled for individual differences in overall motor responding.Given this baseline, an ability to adaptively integrate expected valuewould be indicated by relatively faster responding in the DEVcondition and slower responding in the IEV condition. Dopami-nergic manipulations in Parkinson’s patients have opposite effectson these measures, likely via modulation of striatal dopamine29.

We also included a fourth condition (constant expected value–reverse, CEVR) in which reward probability increased while magnitudedecreased. This condition serves two purposes. First, because both CEVand CEVR have equal expected values across time, any difference in RTin these two conditions can be attributed to a participants’ potentialbias to learn more about reward probability than about magnitude or

vice versa. Second, CEVR provides another measure of avoidancelearning. That is, despite the constant expected value, a bias to learnfrom negative outcomes will produce slowed responses because of theirhigh probability of occurrence at early response times.

Overall, participants showed robust learning (Fig. 2; also see Fig. 5 inthe Supplementary Data Analysis for RTs for each genotype). Com-pared to the baseline CEV condition, RTs in the IEV condition weresignificantly slower (F1,67 ¼ 28.5, P o 0.0001), whereas those in theDEV condition were significantly faster (F1,67 ¼ 6.7, P ¼ 0.01).

There were no effects of any gene either on baseline RTs in the CEVcondition or on overall response time (all P values 4 0.25). Never-theless, within-subject RT modulations due to reward structure werepredictably altered by striatal genotype (Fig. 3). Individuals with theDARPP-32 T/T genotype showed enhanced Go learning, with faster RTsin the last block of the DEV condition (F1,64 ¼ 4.4, P ¼ 0.039), and,marginally, relative to CEV (DEVdiff; F1,64 ¼ 3.1, P ¼ 0.08, an effect thatwas significant across all trials; P o 0.05). DARPP-32 alleles had noeffect on NoGo learning (IEV RTs, or IEVdiff; P values 4 0.8).Conversely, DRD2 T/T carriers, who have the highest striatal D2receptor density10,15, showed marginally slower RTs in IEV, indicativeof enhanced NoGo learning (F1,66 ¼ 3.3, P ¼ 0.07 for both IEV andIEVdiff), but no effect on Go learning (P values 4 0.3). Modelingresults reported below, together with CEVR performance, morestrongly support the conclusion that DARPP-32 and DRD2 allelesmodulate learning to speed and slow RTs from positive and negativeoutcomes. Finally, there was no effect of COMT on any of thesemeasures (P values 4 0.35). This constellation of genetic effectsconverge with those found previously10 but extend them to acompletely different task context, dependent measure and sample.Moreover, these same RT adaptations due to reward structure aresensitive to dopaminergic manipulation in Parkinson’s disease29.

Further analysis revealed genetic contributions to learning fromprobability relative to magnitude of reinforcement, as assessed bycomparing RTs in the CEVR condition (alone and relative to CEV;P ¼ 0.02, Supplementary Data Analysis). Specifically, individuals

543210543210Time (s)Time (s)

CEVRIEVDEVCEV

05

1015202530354045505560

Exp

ecte

d va

lue

(fre

q ×

mag

) Expected valueReward magnitude

0

50

100

150

200

250

300

350

CEVRIEVDEVCEV

Num

ber

of p

oint

s ga

ined

54321Time (s)

00.00.10.20.30.40.50.60.70.80.91.0

Pro

babi

lity CEVR

IEVDEVCEV

Reward frequencya b

c d

50403020100Trial

50403020100Trial

1,000

1,500

2,000

RT

(m

s)

2,500

3,000 All subjects: model fits

1,000

1,500

2,000

2,500

3,000

RT

(m

s)

CEV

All subjects: dataa b

IEVCEVR

DEVCEV

IEVCEVR

DEV

Figure 2 Response times as a function of trial number, smoothed (with

weighted linear least-squares fit) over a ten-trial window. (a) In all 69

participants. (b) Computational model.

Figure 1 Task conditions: decreasing expected value (DEV), constant

expected value (CEV), increasing expected value (IEV) and constant expected

value–reverse (CEVR). The x axis corresponds to the time after onset of the

clock stimulus at which the response is made. The functions are designed

such that the expected value at the beginning in DEV is equal to that at the

end in IEV so that at optimal performance, subjects should obtain the same

average reward in both IEV and DEV. Faster responses were accompanied by

longer intertrial intervals so that reward rate is roughly equalized acrossconditions. (a) Example clock-face stimulus. Each trial ended when the

subject made a response or otherwise when the 5 s duration elapsed. The

number of points won on the current trial was displayed. (b) Probability

of reward occurring as a function of response time. (c) Reward magnitude

(contingent on probability in b). (d) Expected value across trials for each time

point. Note that CEV and CEVR have the same EV.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1063

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

with enhanced D2 function showed significantly greater sensitivity tofrequent negative outcomes in CEVR, again consistent with enhancedNoGo learning. There was also some evidence that carriers of theCOMT met allele were more sensitive to reward magnitudes (Fig. 1 inSupplementary Data Analysis).

Trial-to-trial RT adaptation: exploration?

Although, on average, participants incrementally changed response timesdependent on reward structure, single-subject data revealed large RTswings from one trial to the next (Fig. 4). These swings did not reflectadaptive changes following rewards or lack thereof 29. Instead, preli-minary analyses indicated that RT swings simply reflected a regressionto the mean, whereby faster-than-average responses were more likely tobe followed by relatively slower responses and vice versa (P o 0.0001;Supplementary Data Analysis). As will be clear, however, these RTswings reflect more than just a statistical necessity and are likely torepresent participants’ tendency to explore the space of responses todetermine the reward structure. We investigated this effect in themathematical reinforcement learning (RL) model developed below.

Computational model

We previously simulated performance in thistask using an a priori neural network modelof the basal ganglia29. The model simulatesinteractive neural dynamics among cortico-striatal circuits and accounts for variouseffects of dopaminergic manipulation on actionselection and reinforcement learning17,30–32.Simulated treatments with medications thatstimulate dopamine receptors induce speededRTs in the DEV condition as a result of D1-dependent Go learning in striatonigral cells.However, the same increased dopaminerelease impedes the ability to slow down inIEV due to excessive D2 receptor stimula-tion on striatopallidal cells and concomitantimpairments in NoGo learning. Simulateddopamine depletion produces the oppositeresult: less speeding in DEV but better slowingin IEV and CEVR, mirroring the performanceof Parkinson’s patients on the task29.

Here we develop an abstract mathematicalmodel designed to quantitatively fit individualparticipants’ response times on a trial-to-trialbasis. The purpose of this modeling is three-fold: (i) to demonstrate the core computational

principles by which the more complex neural model captures theincremental RT changes as a function of reward prediction error;(ii) to augment the model to capture strategic exploratory behavior asa function of reward uncertainty; and (iii) to determine whetherbest-fitting model parameters for both exploitative and exploratorydecisions are predictably modulated as a function of genotype10.

The point of departure for the model is the central assumptioncommon with virtually all reinforcement models, namely that partici-pants develop an expected value V(t) for the reward they expect to gainin a given trial t. This value is updated as a function of each rewardexperience using a simple delta rule:

Vðt + 1Þ ¼ VðtÞ+ adðtÞwhere a is a learning rate that modifies the extent to which values areupdated from one trial to the next and d is the reward prediction errorreported by dopamine neurons16,33, which is simply the reward out-come (Rew) minus the prior expected value:

dðtÞ ¼ RewðtÞ � VðtÞ

IEV_diff (NoGo)

Contrast

DEV_diff (Go)

–150

0

150

300

450

600

750

900

Res

pons

e tim

e di

ff (m

s)

Go/NoGo RT learningCOMT gene

metval/valC/C, C/T

T/T

Go/NoGo RT learningDRD2 gene

IEV_diff (NoGo)DEV_diff (Go)

Contrast

–150

0

150

300

450

600

750

900

Res

pons

e tim

e di

ff (m

s)

IEV_diff (NoGo)DEV_diff (Go)

Contrast

–150

0

150

300

450

600

750

900

Res

pons

e tim

e di

ff (m

s) T/TC/C, C/T

Go/NoGo RT learningDARPP-32 gene

a b c

Figure 3 Relative within-subjects biases to speed RTs in DEV relative to CEV (DEVdiff ¼ CEV � DEV) and to slow RTs in IEV (IEVdiff ¼ IEV � CEV). Values

represent mean (s.e.m.) in the last quarter of trials in each condition. (a) DARPP-32 gene. (b) DRD2 gene. (c) COMT gene.

4030Trial

201004030Trial

201000

1,000

2,000RT

(m

s) 3,000

4,000

5,000

0

1,000

2,000RT

(m

s) 3,000

4,000

5,000

DataGoNoGo

Single-subject IEV

NoGoGoData

Single-subject DEV

4030Trial

201000

1,000

2,000RT

(m

s) 3,000

4,000

5,000

DataGoNoGo

Single-subject CEVR

403020Trial

1000

1,000

2,000

3,000

4,000

5,000

RT

(m

s)

NoGoGoData

Single-subject CEVa b

c d

Figure 4 Trial-to-trial RT adjustments in a single subject. (a–d) Shown are data in CEV (a), CEVR (b),

DEV (c) and IEV (d). Model Go and NoGo terms (magnified by four times) accumulate as a function

of positive and negative prediction errors. Go dominates over NoGo in DEV and the reverse in IEV,

but these incremental changes do not capture trial-by-trial dynamics. For this subject, aG ¼ 0.63

and aL ¼ 0.74 (ms per point).

1064 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

This value integration is posited to be computed by brain areasupstream of dopamine neurons comprising the ‘‘critic,’’ which learnsas a function of prediction errors to faithfully represent expectedvalue5,34,35. Our model further shares the assumption that these sameprediction error signals train the ‘‘actor’’ in the striatum34. This processcan occur in at least two ways. First, we model a simple, likely implicit,process whereby accumulated positive prediction errors translate intoapproach-related speeded responses (Go learning), whereas accumu-lated negative prediction errors produce relative avoidance and slowedresponses (NoGo learning)29,32. These processes are posited to rely onD1 and D2 receptor mechanisms in separate populations of striatoni-gral and striatopallidal cells17,29,32,36. Because of these differentiallearning mechanisms, we use different learning rates, and for each:

Goðs; a; t + 1Þ ¼ Goðs; a; tÞ+ aGd+ðtÞNoGoðs; a; t + 1Þ ¼ NoGoðs; a; tÞ+ aNd�ðtÞ

where aG controls D1-dependent speeding from positive predictionerrors (d+) and aN controls D2-dependent slowing from negativeprediction errors (d�), for action a and clock-face state s. On eachtrial RTs were predicted to speed or slow according to differencesbetween current Go and NoGo values.

In addition to this implicit process capturing putative striatalcontributions to approach/avoidance, we also model a more strategicprocess in which participants separately keep track of reward structurefor different (‘fast’ and ‘slow’) responses (Supplementary Data Ana-lysis). With these action representations, participants need only adaptRTs in proportion to the difference between their expected rewardvalues. This would allow, for example, participants to delay respondingwhen slow RTs yield larger rewards on average (as in IEV) or to speedup when they do not. We model this process using Bayesian integration,assuming subjects represent the prior distributions of reward predic-tion errors separately for fast and slow responses and update them as afunction of experience via Bayes’ rule:

Pðyjd1 . . . dnÞ / Pðd1 . . . dnjyÞPðyÞwhere y reflects the parameters governing the belief distribution aboutthe reward prediction errors for each response, and d1...dn are theprediction errors observed thus far (on trials 1 to n). Simply stated,Bayes’ rule implies that the degree to which each outcome modifiesparticipants’ beliefs about obtainable rewards depends on their priorexperience and, given this prior, the likelihood that the outcome wouldoccur. As experience is gathered, the means of the posterior distributionsaccurately represent reward structure in each condition (see below).

We considered that participants either track the probability of areward prediction error (that is, the probability that a dopamine burstoccurs) using beta distributions beta(Z,b) or track the magnitude ofexpected rewards represented by normal distributions N(m,s2). Wefocus here on the beta distribution implementation, which provided abetter fit to the behavioral data. Nevertheless, all genetic resultspresented below held when using normal distributions and a Kalmanfilter (Supplementary Data Analysis). In either case, RTs were pre-dicted to adapt in proportion to the difference between the bestestimates of reward structure for fast and slow responses; that is, thefollowing term was added to the RT prediction: r[sslow(s,t) – sfast(s,t)],where r is a free parameter.

We also modeled other parameters that contribute to RT in this task,including simple baseline response speed (irrespective of reward),captured by free parameter K; autocorrelation between the currentand previous RT (l) regardless of reward; and a tendency to adapt RTstoward the single largest reward experienced thus far (‘going for gold’,parameter u). Finally, we posited that exploratory strategies would

� fast

�slow

�fast

�slow

� fast

�slow

�slow

� fast � fast �slow

�fast

�slow

504030

Trial

20100504030

Trial

201000

0.2

0.4

0.6

0.8Single-subject IEVSingle-subject DEV

0

0.2

0.4

0.6

0.8

1.00.80.60.40.20

p(�s,a > 0)

1.00.80.60.40.20

p(�s,a > 0)

FastSlow

FastSlow

DEV beta distributions IEV beta distributionsa b

c d

Parameter

��GN

DRD2 (T – C)DARPP-32 (T – C)COMT (met – val)

–0.20

–0.10

0.00

0.10

0.20

Gen

e-pa

ram

eter

diff

eren

ce

RL model: summaryGene-parameter dissociation

*

*

*

Figure 5 Genetic effects on reinforcement model parameters. DARPP-32 T/T

carriers showed relatively greater learning rates from gains than losses

(aGN ¼ aG – aN) compared to C carriers. DRD2 T/T carriers showed the

opposite pattern. The COMT gene did not affect learning rates, but met

carriers had significantly higher uncertainty-based explore parameter (e)values (which are divided by 104 to allow them to be displayed on the

same scale) than did val/val participants. Error bars, s.e.m.

Figure 6 Evolution of action-value distributions.

(a,b) Beta probability density distributions

representing the belief about the likelihood of

reward prediction errors following fast and slow

responses, averaged across all subjects’ data.

The x axis shows the probability of a positive

prediction error and the y axis represents the

belief in each probability, with the mean value mrepresenting the best guess. Dotted lines reflect

distributions after a single trial; dashed lines,

after 25 trials; solid lines, after 50 trials. (See

Supplementary Video 1 for dynamic changes in

these distributions across all trials for a single

subject.). Differences between mfast and mslow were

used to adjust RTs to maximize reward likelihood.

The s.d. s was taken as an index of uncertainty.

Exploration was predicted to modulate RT indirection of greater uncertainty about whether

outcomes might be better than the status quo.

(c,d) Trajectory of means and s.d. for a single

subject in DEV and IEV conditions. Uncertainties

s decrease with experience. Corresponding beta

hyperparameters Z and b are shown in

Supplementary Data Analysis.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1065

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

contribute to participants’ RT adjustments, as participants sampled theoutcomes available to determine which response was most adaptive.This process is modeled as a dynamic Explore process depending onBayesian uncertainty, which is elaborated further below and is hypothe-sized to rely on prefrontal cortex–dependent processes. The completeRT update is thus as follows:

RTðs; tÞ ¼K + lRTðs; t � 1Þ � Goðs; a; tÞ+ NoGoðs; a; tÞ+ r½mslowðs; tÞ � mfastðs; tÞ�+ n½RTbest � RTavg�+ Exploreðs; tÞ

For each subject, a single set of best fitting parameters was derivedacross all conditions. The model captures the qualitative pattern ofresults, with predicted RT changing as a function of reward structure(Fig. 2b; see Fig. 6 in Supplementary Data Analysis for model fits foreach genotype). Positive prediction errors are most prevalent for earlyresponses in DEV, and accordingly model RTs are fastest in thiscondition. Negative prediction errors are most prevalent in IEV andCEVR, leading to slowed model responses.

We hypothesized that these relative learning rate parameters fordetermining exploitative responses would be modulated by striatalgenotype. Indeed, DARPP-32 T/T carriers, who should have increasedstriatal D1-dependent learning10,13,14, had relatively larger aG as com-pared to aN than did C carriers, suggesting relatively greater sensitivityto positive than negative prediction errors (Fig. 5; F1,65 ¼ 4.0,P ¼ 0.05). Conversely, DRD2 T/T carriers, with relatively greater D2receptor density29, showed relatively greater learning from negativeprediction errors (F1,66 ¼ 5.3, P ¼ 0.02). Relative learning rates werenot modulated by COMT genotype (P 4 0.2), and other than theExplore parameter, no other parameters differed as a function of anygenotype (all P values 4 0.2).

Uncertainty-based exploration

The above model provides an account of incremental RT changes as afunction of reward prediction error, and it provides evidence for themechanisms posited to mediate these effects in neural networks29.Nevertheless, inspection of individual subject data reveals morecomplex dynamics than those observed in the averaged data (Fig. 4).These plots show RTs across trials for an arbitrary single participant,along with model Go and NoGo terms. Asymptotically, the participantconverges on a faster RT in DEV, and slower RT in IEV, relative to CEV.

However, at the more fine-grained scale, there are often large RTswings from one trial to the next that are not captured by modellearning mechanisms.

We hypothesized that these RTswings are rational, in that they mightreflect exploratory strategies to gather statistics of reward structure.Several solutions have been proposed to manage the exploration/exploitation tradeoff. If performance is unsatisfactory over extendedperiods, stochastic noise can simply be added to behavioral outputs,promoting random exploratory choices7. Alternatively, exploration canbe strategically directed toward particular choices in proportion to theamount of information that would be gained, regardless of pastperformance4,6,37,38. Our model embodies the assumption thatexploratory decisions occur in proportion to the participant’s relativeuncertainty about whether responses other than those currently beingexploited might yield better outcomes. This assumption builds on priormodeling in which exploration is encouraged by adding an ‘uncertaintybonus’ to the value of decision options having uncertain out-comes4,6,37,38. Here we posit that exploration occurs in proportion touncertainty about the probability that the explored option will yield apositive reward prediction error (or, in alternative models, uncertaintyabout the expected value of such rewards or reward prediction errors;Supplementary Data Analysis). The Bayesian framework for integrat-ing reward statistics provides a natural index of uncertainty: the s.d. ofthe prior distributions39, which decreases after sampling a given action(albeit at a slower rate for more variable outcomes).

Initially, distributions representing belief about reward structure foreach response category are wide, reflecting maximum uncertainty(Fig. 6). As experience with each option is gathered, the distributionsevolve to reflect the underlying reward structures, such that the meanbelief is higher for fast responses in DEV and for slow responses in IEV.Moreover, the s.d., and hence uncertainties, decrease with experience.This process is analogous to estimating the odds of a coin flip resultingin heads or tails, with uncertainty about those odds decreasing withthe number of observations. With these distributions, the relativeuncertainties for fast and slow responses in a given trial can be usedas a rational heuristic to drive exploration. In particular, the Exploreterm of the model is computed as follows:

Explore ðs; tÞ ¼ e½ssjs;a ¼ Slow � ssjs;a ¼ Fast�

where e is a free parameter that scales exploration in proportion torelative uncertainty and sd|s,a ¼ Slow and sd|s,a ¼ Fast are the standard

met/metval/metval/val

met/metval/metval/val

COMT gene-dose effectsRelative exploration due to uncertainty

COMT gene-dose effectsRelative exploration due to uncertainty

–1.25–1.00–0.75–0.50–0.25

0.000.250.500.751.001.25

∆z (�

– �)

∆z (�

– �)

–1.00

–0.75

–0.50

–0.25

0.00

0.25

0.50

0.75

1.00

Single Subject, CEV

5045403530Trial252015105

–4,000

–3,000

–2,000

–1,000

0

1,000

2,000

3,000

4,000

RT

Diff

(m

s)

ExplorationModel Exp termRT diff

met/metval/metval/val

COMT gene-dose effectsUncertainty-exploration parameter

0.000.050.100.150.200.250.300.350.400.450.50

� (×

1e4

)

a b c d

Figure 7 COMT gene predicts directed exploration toward uncertain responses. (a) RT swings (change in RT from the previous trial) in a single met/met subject

in the CEV condition and the corresponding model uncertainty-based Explore term (amplified to be on the same RT scale). See Supplementary Video 2 for this

subject’s evolution of beta distributions in CEV. (b) Effect of COMT gene dose on the uncertainty-based exploration parameter e. (c,d) Gene-dose effects were

also observed when comparing relative contributions of e compared with a reverse-momentum parameter g (c) and a lose-switch parameter k (d). Relative z

scores are plotted here to permit comparison of parameter scaling quantities of different magnitudes. Error bars, s.e.m.

1066 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

deviations quantifying uncertainty about reward prediction error like-lihood given slow and fast responses, respectively. Thus, with suffi-ciently high e, RT swings are predicted to occur in the direction ofgreater uncertainty about the likelihood that outcomes might be betterthan the status quo.

Overall, including this uncertainty-based exploration term provideda better fit to trial-by-trial choice than the base model withoutexploration (and penalizing the fit for the additional parameters; seeSupplementary Data Analysis). Although the model cannot determi-nistically predict RT swings (which reflect the output of multipleinteracting processes, including those sensitive to previous reinforce-ment), there is nevertheless a reliable positive correlation between themodel’s uncertainty-based exploratory predictions and participants’actual RT swings from on trial to the next (r4,214 ¼ 0.31, P o 0.0001;Fig. 7; Fig. 3 in Supplementary Data Analysis).

Moreover, this relationship was particularly evident for carriersof the COMT met allele (Fig. 3 in Supplementary Data Analysis),supporting a role for PFC neuromodulatory control over explorationas a function of decision uncertainty. The e parameter that scalesexploration in proportion to uncertainty was significantly higheramong met allele carriers (Fig. 5; F1,67 ¼ 8.2, P ¼ 0.006). Further,there was a monotonic gene-dose effect, with e values beinglargest in met/met participants, intermediate in val/met andsmallest in val/val carriers (Fig. 7b; F1,67 ¼ 9.5, P ¼ 0.003). Nosuch effects on e were observed for DARPP-32 or DRD2 genotypes(P values 4 0.5).

Notably, the COMT exploration effects appear to be specific touncertainty. First, overall RT variability (in terms of s.d.) did not differas a function of genotype (P 4 0.2). Second, a number of foilmodels attempting to account for RT swings without recourse touncertainty confirmed that only the uncertainty-based explorationparameter can account for COMT effects (Supplementary DataAnalysis). For example, we included a ‘reverse momentum’ parameterg, which predicted RT swings to counter a string of progressivelyspeeded or slowed responses, regardless of uncertainty. Althoughthis model provided a reasonable fit to RT swings overall, the uncer-tainty model was superior only in carriers of the COMT met allele(Supplementary Data Analysis). We also included a ‘lose-switch’parameter k, which predicted RTs to adjust from fast to slow or viceversa following a negative prediction error. Notably, there wereCOMT gene-dose effects not only on raw e values but also on theirrelative weighting compared to either g or k (P values o 0.004;Fig. 7c,d). This result implies that the contribution of COMT to RTswings is specific to uncertainty.

DISCUSSION

Individuals differ substantially in their motivational drives. The presentfindings demonstrate three distinct aspects of value-based decision-making associated with independent genetic factors (see summaryFig. 5). These genes modulate specific aspects of dopaminergic func-tion in brain areas thought to support exploration and exploita-tion6,7,10,19. Behaviorally, exploitative choices were manifest by RTdifferences between conditions in which rewards could on average bemaximized by responding earlier (DEV) or later (IEV) in the trial,compared to baseline (CEV) conditions. Modeling showed that striatalgenetic effects are accounted for by individual differences in learningrates from positive and negative prediction errors and their couplingwith response speeding and slowing. This result is nontrivial: striatalgenes could have affected exploitation by modulating the extent towhich RTs are adjusted as a function of mean reward value estimates(that is, the r parameter). Similarly, whereas trial-to-trial RT swings

were readily viewable in single-subject data (Fig. 4), the specificcomponents due to uncertainty-based exploration, and individualdifferences therein, were only extracted with the computational analysis.

Our observation that DARPP-32 and DRD2 modulate reinforcementlearning in the temporal decision-making domain is consistent withsimilar genetic effects in choice paradigms10 and with data frompatients with Parkinson’s disease, on and off medication, in this sametask29. Recent rodent studies show direct support for the model’s dualD1 and D2 mechanisms of synaptic plasticity17,18.

The present human genetic data provide support for the mechanismsposited in models of striatal dopamine, in which accumulated rewardprediction errors over multiple trials produce speeded responses,whereas negative prediction errors slow responses29,40. Our assumptionthat DARPP-32 genetic effects reflect striatal D1 receptor–mediated Golearning is supported by evidence that the DARPP-32 protein is highlyconcentrated in the striatum12 and is critical for D1- but not D2-dependent synaptic plasticity and behavioral reward learning13,14. Thesedata also converge with effects of pharmacological manipulation ofstriatal D1 receptors on appetitive approach and response speeding toobtain rewards in monkeys and rats36,41.

Similarly, our assumption that DRD2 genetic effects reflect primarilystriatal D2 receptor–mediated learning is supported by evidence thatT/T carriers show enhanced striatal D2 receptor density29,42. Theore-tically, striatal D2 receptors are thought to be necessary for learning instriatopallidal neurons when dopamine levels are low17, as is the caseduring negative prediction errors43–45 or as a result of Parkinson’sdisease29,30. Indeed, synaptic potentiation in striatopallidal neurons iselevated under conditions of dopamine depletion18. Conversely, ratswith reduced striatal D2 receptor density46 are less sensitive to aversiveoutcomes, persisting in taking addictive drugs even when this isfollowed by shocks47.

Perhaps less clear is the precise neurobiological mechanism by whichCOMT modulates uncertainty-based exploration. Indeed, themechanisms of exploration are understudied compared to those ofexploitation. Nevertheless, neuroimaging studies reveal that in non-reinforcement-learning contexts, anterior prefrontal cortical regionsreflect Bayesian uncertainty21, and that this same region is activatedwhen participants make exploratory decisions in a RL environment6.Our findings provide the first evidence for exploratory decisions thatoccur in proportion to uncertainty about whether other responsesmight produce better outcomes than the status quo. This explorationstrategy is strongly motivated by prior theoretical work6,7,38 and seemsto be highly dependent on genetic function in the prefrontal cortex.Furthermore, the COMT effects on trial-to-trial ‘lose-shift’ behavior inchoice paradigms that we originally reported10 might be more parsi-moniously explained by uncertainty-based exploratory mechanisms.Indeed, in that study, met carriers showed greater propensity to shiftonly in the initial trials of the task, when reward structure was mostuncertain. Thus, these exploratory strategies may be viewed as anattempt to minimize uncertainty.

In contrast to the multiple extant neural models of exploitation,there is a dearth of models investigating how neuronal populations canlearn to represent quantities of uncertainty as a function of experience.Nevertheless, the sorts of Bayesian probability distributions requiredfor the uncertainty computations used here are naturally coded inpopulations of spiking neurons48,49. Thus, future research shouldexamine how such representations can be learned and whether pre-frontal dopamine supports the uncertainty computations per se, theactive maintenance of relative uncertainties in working memory acrosstrials, or simply the final decision to override exploitative strategies inorder to explore when uncertainty is sufficiently high.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1067

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSWe thank S. Williamson and E. Carter for help with DNA analysis andadministering cognitive tasks to participants, and N. Daw, P. Dayan, andR. O’Reilly for helpful discussions. This research was supported byUS National Institutes of Mental Health grant R01 MH080066-01.

AUTHOR CONTRIBUTIONSM.J.F., B.B.D. and F.M. designed the study; M.J.F. conducted the modeling andanalyzed the behavioral data; B.B.D. collected data; J.O.-T. and F.M. extracted theDNA and conducted genotyping; M.J.F., B.B.D. and F.M. wrote the manuscript.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://www.nature.com/

reprintsandpermissions/.

1. Scheres, A. & Sanfey, A.G. Individual differences in decision making: Drive and RewardResponsiveness affect strategic bargaining in economic games. Behav. Brain Funct. 2,35 (2006).

2. Hsu, M., Bhatt, M., Adolphs, R., Tranel, D. & Camerer, C.F. Neural systems responding todegrees of uncertainty in human decision-making. Science 310, 1680–1683 (2005).

3. Frank, M.J., Woroch, B.S. & Curran, T. Error-related negativity predicts reinforcementlearning and conflict biases. Neuron 47, 495–501 (2005).

4. Gittins, J.C. & Jones, D. A dynamic allocation index for the sequential designof experiments. in Progress in Statistics (eds. Gani, J., Sarkadi, K. & Vincze, I.),241–266 (North Holland Publishing Company, Amsterdam, 1974).

5. Sutton, R.S. & Barto, A.G. Reinforcement Learning: An Introduction (MIT Press,Cambridge, Massachusetts, USA, 1998).

6. Daw, N.D., O’Doherty, J.P., Dayan, P., Seymour, B. & Dolan, R.J. Cortical substrates forexploratory decisions in humans. Nature 441, 876–879 (2006).

7. Cohen, J.D., McClure, S.M. & Yu, A.J. Should I stay or should I go? How the human brainmanages the trade-off between exploitation and exploration. Phil. Trans. R. Soc. Lond. B362, 933–942 (2007).

8. Depue, R.A. & Collins, P.F. Neurobiology of the structure of personality: dopamine, facili-tation of incentive motivation, and extraversion. Behav. Brain Sci. 22, 491–517 (2001).

9. Meyer-Lindenberg, A. et al. Genetic evidence implicating DARPP-32 in human frontos-triatal structure, function, and cognition. J. Clin. Invest. 117, 672–682 (2007).

10. Frank, M.J., Moustafa, A.A., Haughey, H.M., Curran, T. & Hutchison, K.E. Genetic tripledissociation reveals multiple roles for dopamine in reinforcement learning. Proc. Natl.Acad. Sci. USA 104, 16311–16316 (2007).

11. Klein, T.A. et al. Genetically determined differences in learning from errors. Science318, 1642–1645 (2007).

12. Ouimet, C.C., Miller, P.E., Hemmings, H.C., Walaas, S.I. & Greengard, P. DARPP-32, adopamine- and adenosine 3’:5’-monophosphate-regulated phosphoprotein enriched indopamine-innervated brain regions. III. Immunocytochemical localization. J. Neurosci.4, 111–124 (1984).

13. Stipanovich, A. et al. A phosphatase cascade by which rewarding stimuli controlnucleosomal response. Nature 453, 879–884 (2008).

14. Calabresi, P. et al. Dopamine and cAMP-regulated phosphoprotein 32 kDa controls bothstriatal long-term depression and long-term potentiation, opposing forms of synapticplasticity. J. Neurosci. 20, 8443–8451 (2000).

15. Hirvonen, M. et al. Erratum: C957T polymorphism of the dopamine D2 receptor (DRD2)gene affects striatal DRD2 availability in vivo. Mol. Psychiatry . 10, 889 (2005).

16. Montague, P.R., Dayan, P. & Sejnowski, T.J. A framework for mesencephalic dopaminesystems based on predictive Hebbian learning. J. Neurosci. 16, 1936–1947 (1996).

17. Frank, M.J. Dynamic dopamine modulation in the basal ganglia: a neurocomputationalaccount of cognitive deficits in medicated and nonmedicated Parkinsonism. J. Cogn.Neurosci. 17, 51–72 (2005).

18. Shen, W., Flajolet, M., Greengard, P. & Surmeier, D.J. Dichotomous dopaminergiccontrol of striatal synaptic plasticity. Science 321, 848–851 (2008).

19. Graybiel, A.M. Habits, rituals, and the evaluative brain. Annu. Rev. Neurosci. 31,359–387 (2008).

20. Kakade, S. & Dayan, P. Dopamine: generalization and bonuses. Neural Netw. 15,549–559 (2002).

21. Yoshida, W. & Ishii, S. Resolution of uncertainty in prefrontal cortex. Neuron 50,781–789 (2006).

22. Frank, M.J. & Claus, E.D. Anatomy of a decision: striato-orbitofrontal interactions inreinforcement learning, decision making, and reversal. Psychol. Rev. 113, 300–326(2006).

23. Roesch, M.R. & Olson, C.R. Neuronal activity related to reward value and motivation inprimate frontal cortex. Science 304, 307–310 (2004).

24. Rudebeck, P.H., Walton, M.E., Smyth, A.N., Bannerman, D.M. & Rushworth, M.F.S.Separate neural pathways process different decision costs. Nat. Neurosci. 9,1161–1168 (2006).

25. Meyer-Lindenberg, A. et al. Midbrain dopamine and prefrontal function inhumans: interaction and modulation by COMT genotype. Nat. Neurosci. 8, 594–596(2005).

26. Slifstein, M. et al. COMT genotype predicts cortical-limbic D1 receptor availabilitymeasured with [11C]NNC112 and PET. Mol. Psychiatry 13, 821–827 (2008).

27. Gogos, J.A. et al. Catechol-O-methyltransferase-deficient mice exhibit sexuallydimorphic changes in catecholamine levels and behavior. Proc. Natl. Acad. Sci. USA95, 9991–9996 (1998).

28. Forbes, E.E. et al. Genetic variation in components of dopamine neurotransmissionimpacts ventral striatal reactivity associated with impulsivity. Mol. Psychiatry 14, 60–70(2009).

29. Moustafa, A.A., Cohen, M.X., Sherman, S.J. & Frank, M.J. A role for dopamine intemporal decision making and reward maximization in parkinsonism. J. Neurosci. 28,12294–12304 (2008).

30. Frank, M.J., Seeberger, L.C. & O’Reilly, R.C. By carrot or by stick: cognitive reinforce-ment learning in parkinsonism. Science 306, 1940–1943 (2004).

31. Santesso, D., Evins, A., Frank, M., Cowman, E. & Pizzagalli, D. Single dose of adopamine agonist impairs reinforcement learning in humans: evidence from event-related potentials and computational modeling of striatal-cortical function. Hum. BrainMapp. 30, 1963–1976 (2009).

32. Wiecki, T.V., Riedinger, K., Meyerhofer, A., Schmidt, W.J. & Frank, M.J. A neurocompu-tational account of catalepsy sensitization induced by D2 receptor blockade in rats:context dependency, extinction, and renewal. Psychopharmacology (Berl.) 204,265–277 (2009).

33. Bayer, H.M. & Glimcher, P.W. Midbrain dopamine neurons encode a quantitative rewardprediction error signal. Neuron 47, 129–141 (2005).

34. O’Doherty, J. et al. Dissociable roles of ventral and dorsal striatum in instrumentalconditioning. Science 304, 452–454 (2004).

35. O’Reilly, R.C., Frank, M.J., Hazy, T.E. & Watz, B. PVLV: the primary value and learnedvalue Pavlovian learning algorithm. Behav. Neurosci. 121, 31–49 (2007).

36. Nakamura, K. & Hikosaka, O. Role of dopamine in the primate caudate nucleus in rewardmodulation of saccades. J. Neurosci. 26, 5360–5369 (2006).

37. Sutton, R.S. Integrated architectures for learning, planning and reacting based onapproximating dynamic programming. Proceedings of the Seventh International Con-ference on Machine Learning (Porter, B.W. & Mooney, R.J., eds.) 216–224 (MorganKaufmann, Palo Alto, California, USA, 1990).

38. Dayan, P. & Sejnowski, T.J. Exploration bonuses and dual control. Mach. Learn. 25,5–22 (1996).

39. Daw, N.D., Niv, Y. & Dayan, P. Uncertainty-based competition between prefrontal anddorsolateral striatal systems for behavioral control. Nat. Neurosci. 8, 1704–1711(2005).

40. Niv, Y., Daw, N.D., Joel, D. & Dayan, P. Tonic dopamine: opportunity costs and the controlof response vigor. Psychopharmacology (Berl.) 191, 507–520 (2007).

41. Dalley, J.W. et al. Time-limited modulation of appetitive Pavlovian memory by D1 andNMDA receptors in the nucleus accumbens. Proc. Natl. Acad. Sci. USA 102,6189–6194 (2005).

42. Zhang, Y. et al. Polymorphisms in human dopamine D2 receptor gene affect geneexpression, splicing, and neuronal activity during working memory. Proc. Natl. Acad.Sci. USA 104, 20552–20557 (2007).

43. Hollerman, J.R. & Schultz, W. Dopamine neurons report an error in the temporalprediction of reward during learning. Nat. Neurosci. 1, 304–309 (1998).

44. Satoh, T., Nakai, S., Sato, T. & Kimura, M. Correlated coding of motivation and outcomeof decision by dopamine neurons. J. Neurosci. 23, 9913–9923 (2003).

45. Bayer, H.M., Lau, B. & Glimcher, P.W. Statistics of midbrain dopamine neuron spiketrains in the awake primate. J. Neurophysiol. 98, 1428–1439 (2007).

46. Dalley, J.W. et al. Nucleus accumbens D2/3 receptors predict trait impulsivity andcocaine reinforcement. Science 315, 1267–1270 (2007).

47. Belin, D., Mar, A.C., Dalley, J.W., Robbins, T.W. & Everitt, B.J. High impulsivity predictsthe switch to compulsive cocaine-taking. Science 320, 1352–1355 (2008).

48. Zemel, R.S., Dayan, P. & Pouget, A. Probabilistic interpretation of population codes.Neural Comput. 10, 403–430 (1998).

49. Ma, W.J., Beck, J.M., Latham, P.E. & Pouget, A. Bayesian inference with probabilisticpopulation codes. Nat. Neurosci. 9, 1432–1438 (2006).

1068 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

ART ICLES

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSSample. We tested 73 healthy participants who were recruited from the

University of Arizona undergraduate psychology subject pool and who provided

informed written consent. Two subjects declined genetic sampling and are

excluded from analysis. Failed genetic assays eliminated a further two COMT

samples, two DRD2 samples and three DARPP-32 samples. The remaining 69

subjects (46 female) had a mean age of 19 (s.e.m. ¼ 0.2) and comprised 48 self-

identified as Caucasian, 14 Hispanics, 2 Asians, 1 African-American and 4

subjects who categorized themselves as ‘Other’. The breakdown of COMT

genotypes was 19:43:7 (val/val:val/met:met/met). The breakdown of DRD2

genotypes was 31:38 (C carriers:T/T homozygotes). The breakdown of

DARPP-32 genotypes was 38:29 (T/T:C carriers; note that in our prior report

the T/T genotype was incorrectly referred to as A/A, and C carriers as G carriers,

due to mislabeling of the base-pair complement10. Thus, the T/T subjects here

reflect the same genotype previously associated with enhanced Go learning.)

Genetic effects were independent: there was no association between the dis-

tribution of any one polymorphism and any other (for example, DRD2 genotype

was not predictive of COMT genotype, and so on Fisher’s exact test, P 4 0.3).

All genotypes were in Hardy-Weinberg equilibrium (P values 4 0.1), with the

exception of COMT (w2[1] ¼ 5.6, P o 0.05). This deviation is likely to be due

to heterogeneity in the population; in an analysis of individuals self-identifying

as Caucasian alone, Hardy-Weinberg equilibrium was not violated (P 4 0.1).

Genotyping. Genotyping procedures were carried out in the Molecular

Psychiatry Laboratory at the University of Arizona. DNA samples were

extracted from saliva samples using Oragene DNA Collection Kits (DNAGen-

otek). Genomic DNA was amplified using standard PCR protocols.

Dopamine- and adenosine-3¢,5¢-monophosphate (cAMP)-regulating phos-

phoprotein SNP (encoded by DARPP-32, rs907094). Genomic DNA was

amplified for the DARPP-32 (also called PPP1R1B) SNP using standard PCR

protocols. Amplification of the 404-bp region was carried out using the sense

primers DD-F 5¢-GCATTGCTGAGTCTCACCTGCAGTCT-3¢ and antisense

primers DD-R 5¢-ATTGGGAGAGGGACTGAGCCAAGGATGG-3¢ in a reaction

volume of 25 ml consisting of 2.5 ng of DNA, 0.25 mM dNTPs, 0.25 mM each

sense and antisense primers, 1� Qiagen PCR buffer and 1.5 U Taq DNA

polymerase (Qiagen). Thermocycling conditions consisted of an initial dena-

turation step of 95 1C for 5 min followed by 35 cycles of 94 1C for 30 s, 72 1C

for 60 s, and 72 1C for 60 s, with a final extension step of 72 1C for 10 min. PCR

products were sequenced using the ABI 3730XL DNA Analyzer (Applied

Biosystems) and visualized using Chromas Vs. 2.13 (Technelysium).

COMT rs4680. Genomic DNA was amplified for the Comt4680 polymorphism

using standard PCR protocols. Amplification of the 109-bp region was carried

out using the sense primers Comt-F 5¢-TCTCCACCTGTGCTCACCTC-3¢ and

antisense primers Comt-R 5¢-GATGACCCTGGTGATAGTGG-3¢ in a reaction

volume of 25 ml consisting of 2.5 ng of DNA, 0.25 mM dNTPs, 0.25 mM each

sense and antisense primers, 1� Qiagen PCR buffer and 1 U Taq DNA

polymerase (Qiagen). Thermocycling conditions consisted of an initial dena-

turation step of 95 1C for 5 min followed by 35 cycles of 95 1C for 15 s, 54 1C

for 20 s, and 72 1C for 30 s, with a final extension step of 72 1C for 5 min. The

restriction enzyme NlaIII (5 U, New England Biolabs) was added to a 20-ml

aliquot of the PCR product and digested for 2 h at 37 1C. Five microliters of the

digested PCR product was added to 4 ml of Orange G DNA loading buffer

and loaded onto a 3% agarose gel. Images were captured via the Gel Doc XR

System (Bio-Rad).

DRD2 rs6277. Optimization of tetra-primer ARMS PCR for the detection of

the DRD2 polymorphism was performed empirically using primers designed by

original software developed by the founders of the tetra-primer ARMS PCR

method and available on the website http://cedar.genetics.soton.ac.uk/

public_html/primer1.html, with a Tm optimized to 72 1C and a GC content

of 48.7%.

Genomic DNA was amplified for the DRD2 polymorphism using tetra-

primer ARMS PCR protocol as described50. Amplification of the total 2,950-bp

region was carried out using the outer sense primers DRD2-F 5¢-ACGGCTC

ATGGTCTTGAGGGAGGTCCGG-3¢ and outer antisense primers DRD-R 5¢-CCAGAGCCCTCTGCCTCTGGTGCAGGAG-3¢ as well as inner sense primers

DRD-Fi 5¢-ATTCTTCTCTGGTTTGGCGGGGCTGGCA-3¢ and inner antisense

primers 5¢-CGTCCCACCACGGTCTCCACAGCACTACC-3¢ in a reaction

volume of 25 ml consisting of 2.5 ng of DNA, 0.25 mM dNTPs, 0.025 mM

outer sense and antisense primers, 0.25 mM inner sense and antisense primers,

1� Qiagen PCR buffer and 2 U Taq DNA polymerase (Qiagen). Thermocycling

conditions consisted of an initial denaturation step of 95 1C for 5 min followed

by 35 cycles of 94 1C for 30 s, 72 1C for 60 s, and 72 1C for 60 s, with a final

extension step of 72 1C for 10 min. Five microliters of the PCR product was

added to 4 ml of Orange G DNA loading buffer and loaded onto a 3% agarose

gel and run in 0.5� TAE buffer for 20 min at 72 V. The gels were prestained

with GelStar Nucleic Acid Gel Stain and images were captured with the Gel Doc

XR System (Bio-Rad).

Genotyping for DRD2 was carried in triplicate, and identification of

each individual allele was conducted by three independent observers with

100% agreement.

Ethnicity. Because there was some heterogeneity in the sample (14 subjects

were Hispanic), it was critical to establish whether genetic effects could have

been due occult stratification. To this end, we reanalyzed the data with the

14 Hispanic individuals omitted and found very similar patterns of results for

each genotype. Similar results also were found when omitting all individuals

not self-identifying as Caucasian. We also reanalyzed all the data and included

an additional factor into the general linear model according to whether subjects

were Hispanic or not. In this analysis, all genetic effects remained significant

and there was no effect of ethnicity, nor was there an interaction between

ethnicity and genotype (P values 4 0.25). Again, similar findings were

included if the factor coded whether subjects were self-identifying as Caucasian

or non-Caucasian. Finally, Hardy-Weinberg equilibrium data were also ana-

lyzed when excluding Hispanic and other individuals not self-identifying as

Caucasian, and no genotype frequencies deviated from equilibrium.

Task methods. Task instructions were as follows:

‘‘You will see a clock face. Its arm will make a full turn over the course of

5 seconds. Press the ‘spacebar’ key to win points before the arm makes a full turn.

Try to win as many points as you can!

‘‘Sometimes you will win lots of points and sometimes you will win less. The

time at which you respond affects in some way the number of points that you can

win. If you don’t respond by the end of the clock cycle, you will not win

any points.

‘‘Hint: Try to respond at different times along the clock cycle in order to learn

how to make the most points. Note: The length of the experiment is constant and is

not affected by when you respond.’’ This hint was provided to prevent

participants from responding quickly simply to leave the experiment early

and in an attempt to equate reward rate (that is, rewards per second) across

conditions. In addition, earlier responses were associated with longer intertrial

intervals so that the statement that the length of the experiment was constant

was roughly accurate. However, because subjects might be averse to waiting

through long intertrial intervals, and because we also wished to reduce the

predictability of the onset of the next trial’s clock face stimulus, we set the

intertrial interval to (5,000 – RT)/2. Thus, faster responses were associated with

longer wait times, but the onset of each trial was temporally unpredictable.

The order of condition (CEV, DEV, IEV, CEVR) was counterbalanced across

participants. A rest break was given between each of the conditions (after every

50 trials). Subjects were instructed at the beginning of each condition to

respond at different times in order to try to win the most points but were not

told about the different rules (for example, IEV, DEV). Each condition was also

associated with a different color of clock face to facilitate encoding that the

participant was in a new context, with the assignment of condition to color

counterbalanced. Participants completed 50 trials of one condition before

proceeding to the next, for a total of 200 trials.

To prevent participants from explicitly memorizing a particular value of

reward feedback for a given response time, we also added a small amount of

random uniform noise (±5 points) to the reward magnitudes on each trial.

Analysis. General linear models were used for all statistical analysis. COMT

gene-dose effects were tested by entering the number of met alleles expressed by

each subject as a continuous variable. Behavioral analyses, except where

indicated, examined RTs in the last quarter (12 trials) of each condition, by

doi:10.1038/nn.2342 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

which time participants were likely to have learned the reward structure of the

particular clock face29. (Although it is possible to compute learning from the

first to last quarter of each condition, some participants learned to discriminate

reward structure even in the first quarter, minimizing the difference across

quarters. We therefore focused our analyses on the last quarter, in which

performance was expected to stabilize. Further, the model-based analyses

converge with those derived from these behavioral measures without confining

analysis to any part of the learning curve.) In some analyses, the degrees of

freedom are 1 less than they should be because a computer crash occurred for

one subject who therefore did not complete all conditions.

Model methods. In all models, we used the Simplex method with multiple

starting points to derive best-fitting parameters for each individual participant

that minimized the sum of squared error (SSE) between predicted and actual

RTs across all trials. A single set of parameters was derived for each subject

providing the best fit across all task conditions. Data were smoothed with a

five-trial moving average for fitting of sequential time-series responses,

although similar results were produced without such smoothing, just with

larger overall SSEs for all models. Model fits were evaluated with Akaike’s

Information Criterion (AIC), which penalizes model fits for models with

additional parameters:

AIC ¼ 2k + n½logð2pSSE=nÞ+ 1�

where k is the number of parameters, n is the number of data points to be fit

and SSE is the sum of squared error between the model predictions and actual

response times across all trials for each subject. The model with the lowest AIC

value is determined to be the best fit.

Exploit model. There are several ways in which RTs might be modeled in this

task. Our first aim was to derive a simple model to approximate the

mechanisms embodied within our a priori neural network model of the basal

ganglia, which predicted the double dissociation between RTs in the DEV and

IEV conditions dependent on dopaminergic medication status in Parkinson’s

disease29. Because that model is complex and involves multiple interacting

brain areas, we sought to capture its core computations in abstract form and to

then fit free parameters of this reduced model to individual subject data, which

in turn can be linked to striatal dopaminergic genes. A similar procedure was

used in a choice rather than RT task10.

We modeled the incremental RT changes in the different conditions via

separate Go and NoGo parameters that learn from positive and negative

prediction errors and serve to speed and slow RTs, respectively. These para-

meters correspond to D1- and D2-dependent learning in striatonigral and

striatopallidal neurons. The terms ‘Go’ and ‘NoGo’ are shorthand descriptions

of the functions of the two pathways in the neural model, whereby Go and

NoGo activity separately report the learned probability that a given action in

the current state would produce a positive and negative outcome, respectively.

In choice paradigms, the probability that an action is taken is proportional to

the relative (Go – NoGo) activity for that action, as compared to all other

actions. Here, as far as the striatum is concerned in the model, there is only one

action (‘‘hit the spacebar’’), and the relative (Go – NoGo) activity simply

determines the speed at which that action is executed.

Positive and negative prediction errors are computed relative to current

expected value V, which are then used to update V estimates for subsequent

trials and also to train the Go and NoGo striatal values. This scheme is

reminiscent of ‘‘actor-critic’’ reinforcement learning models5,34, where the

critic is the V system, the prediction errors of which are reflected in phasic

dopaminergic signals, and the actor comprises Go and NoGo striatal

neuronal populations17,29.

The expected value V was initialized to 0 at the beginning of the task. The

final V value at the end of each condition was carried over to the beginning of

the next, on the assumption that any rewards obtained at the beginning of a

condition are compared relative to their best estimate of expected value in the

task at large (for example, 50 points might be interpreted as a positive

prediction error if in the last block they had on average obtained 20 points,

but would be a negative prediction error if their previous average point value

was 100). Go and NoGo values were initialized to 0 and accumulated as a

function of reward prediction errors for each state (clock face). (Although the

Go and NoGo terms accumulate monotonically as a function of experience, in

the neural model, Go synapses are weakened following negative prediction

errors and NoGo synapses are strengthened, preventing these values from

saturating. Here the contributions of Go and NoGo terms were small enough

for this to not be necessary; however, adding a decay term to Go/NoGo values

to prevent increases without bound did not change the basic pattern of results.)

Finally, due to model degeneracy, a was held constant and was set to

0.1 to allow integration of history, allowing other Go/NoGo learning para-

meters to vary freely. This same critic learning rate was used in the neural

network implementation29.

Bayesian integration of expected value. The Go and NoGo learning mechan-

isms capture a relatively automatic process in which the striatum speeds or

slows responses after positive or negative prediction errors, respectively,

independent of the RTs that produced those reinforcements. This mechanism

may result from the architecture of the basal ganglia, which supports approach

and avoidance behavior for positive and negative outcomes. This mechanism is

also adaptive in the current task if participants’ initial responses are faster than

the midpoint (as was typically the case), in which case positive prediction errors

predominate in DEV and negative prediction errors predominate in IEV, leading

to speeding and slowing, respectively. The improved behavioral fit (including

penalty for additional parameters) provided by including these mechanisms

suggests that these tendencies capture some of the variance in this task. However,

note that these mechanisms are not necessarily adaptive in all cases: for example,

slow responses that produce positive prediction errors (for example, in IEV)

would lead to subsequent speeding according to this mechanism.

We posited that in addition to Go/NoGo learning, subjects would attempt

to explicitly keep track of the rewards experienced for different responses and

then produce those responses that had been rewarded most. It is unrealistic

to assume that participants track reward structure for all possible response

times. Instead, we employed a simplifying (and perhaps more plausible)

assumption that participants simply track reward structure for responses

categorized as ‘‘fast’’ or ‘‘slow.’’ Given that the reward functions are monotonic

(and assuming subjects believe this to be the case), one only needs to track

rewards separately for fast and slow responses to determine which has the

highest expected value, and to respond faster or slower in proportion to the

difference in these values.

We thus categorized each response depending on whether it was faster or

slower than the participants local mean RTavg, which was itself tracked with

the delta rule:

RTavgðtÞ ¼ RTavgðt � 1Þ + a½RTðt � 1Þ � RTavgðt � 1Þ�

(This choice for tracking average RT was not critical; all results are similar even

if simply defining fast and slow according to the first and second halves of the

clock. However, using an adaptive local mean RT is more general and may

prove useful if the reward functions are nonmonotonic.)

We represented participants’ beliefs about reward structure for these two

response categories in Bayesian terms, assuming participants represent not only

a single value of each response but rather a distribution of such values and,

crucially, the uncertainty about them39. In particular, we posited that partici-

pants would track the estimated likelihood of obtaining a positive reward

prediction error for each response, or the magnitude of such prediction errors,

as a function of the past set of dopamine bursts reported by midbrain

dopamine neurons. Any probability distribution in the exponential family of

distributions can be represented in a population of spiking neurons48,49, so

a priori it is not clear whether it is more plausible for participants to track

simply the probability of a dopamine burst occurring at all or to instead

represent the magnitude of the typical prediction error. Model fits to data were

clearly superior for probability simulations, which we focus on here; never-

theless, as reported in the Supplementary Data Analysis, all genetic findings

hold when modeling reward magnitudes (or reward prediction error magni-

tudes) with a Kalman filter.

We represented the likelihood of reward prediction errors for each state s

and fast or slow action a as beta distributions beta(Zs,a,bs,a) (see below). The

probability of a reward prediction error can be represented as a binomial

process, and the beta distribution is the conjugate prior to the binomial

distribution. This implies that the application of Bayes’ rule to update the

NATURE NEUROSCIENCE doi:10.1038/nn.2342

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

prior distribution results in a posterior distribution that is itself also a beta

distribution with new parameters. (Strictly speaking, a binomial process

assumes that each observation is independent. This assumption is violated in

the case of reward prediction errors because a given reward value may be

interpreted as a positive or negative prediction error depending on prior

reinforcement context. The beta distribution is nevertheless a simplifying

assumption that provided a substantial improvement to behavioral fit. Further-

more, we also modeled a version in which we tracked the probability of

obtaining a nonzero reward, rather than a reward prediction error. In this

model, we also binarized responses such that ‘‘fast’’ and ‘‘slow’’ responses were

categorized according to those that were in the first and second halves of the

clock. In this case, each observation is indeed independent, and all core results

continued to hold.)

The probability density function of the beta distribution is as follows:

f ðx; Z; bÞ ¼ xZ�1ð1 � xÞb�1

R 10 zZ�1ð1 � zÞb�1dz

where the integral in the denominator is the beta function B(Z,b) and is a

normalization factor that ensures that the area under the density function is

always 1. The defining parameters of the posterior distribution for each state s

are calculated after each outcome using Bayes’ rule:

PðZ;bjd1 . . . dnÞ ¼Pðd1 . . . dnjZ; bÞPðZ; bÞR R

Pðd1 . . . dnjZ; bÞdZdb¼ Pðd1 . . . dnjZ;bÞPðZ;bÞ

Pðd1 . . . dnÞ

Explore model. Because of the conjugate prior relationship between binomial

and beta distributions, this update is trivial without having to directly compute

Bayes’ equation above. The Z and b parameters are updated for each state or

action by simply incrementing the prior Z and b hyperparameters after each

instance of a positive or negative prediction error, respectively (see Fig. 4 in

Supplementary Data Analysis for trajectories of hyperparameters for a

single subject):

Zs;aðt + 1Þ ¼Zs;aðtÞ+ 1 if ds;a;t 40

Zs;aðtÞ otherwise

(

bs;aðt + 1Þ ¼bs;aðtÞ+ 1 if ds;a;to0

bs;aðtÞ otherwise

(

The participant can then compare the means of each posterior distribution

and adjust RTs so as to increase the probability of obtaining a reward prediction

error. The mean of the beta distribution is simply m ¼ Z/(Z + b). Thus, this

component of the exploitation model predicts that subjects adjust RTs accord-

ing to r[mslow(s,t) – mfast(s,t)], where r is a free parameter scaling the degree to

which participants use these mean estimates in adapting their RTs.

In addition to the Go/NoGo learning and Bayesian integration mechanisms,

model fits to data were also substantially improved by a mechanism in which

participants adapted RTs toward that which had produced the single largest

reward thus far (‘going for gold’), regardless of the reward probability. This

tendency was captured by free parameter u and was not associated with any

genotype (nor was it required for the core results of the paper to hold, but it

may be useful for future studies of the neural and genetic mechanisms of this

behavior). We modeled this by keeping track of the RT that yielded rewards

that were at least 1 s.d. greater than all rewards observed thus far in the block

and adapting all subsequent RTs toward this value. Further, participants’

response on one trial may be heavily influenced by that of the previous trial,

independent of value. Accordingly, we introduce a parameter l to capture

individual differences in this influence of previous responses.

Thus, the full RT model is as follows:

RTðs; tÞ ¼ K + lRTðs; t � 1Þ � Goðs; a; tÞ+ NoGoðs; a; tÞ+ r½mslowðs; tÞ

� mfastðs; tÞ�+ n½RTbest � RTavg�+ Exploreðs; tÞ

The computations of the final Explore term is discussed next.

One of the central advantages of the Bayesian framework is that it provides

an estimate not only of the ‘best guess’ (the mean, or expected value m of the

beta distribution) but also the uncertainty about that mean, quantified by the

s.d. s of that distribution. We attempted to predict RT swings from one trial to

the next, hypothesizing that RT swings reflect exploration when participants are

uncertain about whether they might obtain better outcomes. The s.d. of the

beta distributions for each state (clock-face) can be computed analytically in

each trial as a measure of uncertainty:

ss; aðtÞ ¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiZs; aðtÞbs;aðtÞ

ðZs;aðtÞ+ bs;aðtÞÞ2ðZs;aðtÞ+ bs;aðtÞ+ 1Þ

!vuut

The model Explore term was applied on each trial as a function of the relative

differences in uncertainty about the likelihood of reward prediction errors given

fast and slow responses:

Explore ðs; tÞ ¼ e½sdjs;a ¼ Slow � sdjs;a ¼ Fast�

In this way, exploratory-based RT swings are predicted to occur in the direction

of greater uncertainty (thereby acting to reduce this uncertainty). Note that for

trials immediately following an exploratory RT swing, as it stands this

implementation would roughly double-count exploration because the lparameter already reflects autocorrelation between the previous and current

RT (where in this case the previous trial was an exploratory swing). To mitigate

against this double counting, we set the Explore term to 0 in trials immediately

following an exploratory RT swing (defined as a change in RT that was in the

same direction predicted by the uncertainty Explore term). The results were not

sensitive to this particular implementation, however. (For example, similar

findings were found without resetting Explore to 0 but instead including a

parameter into the RT estimate that reflects the effects of previous RT swings

from trial n – 2 to n – 1 (in addition to l, which accounts for the raw RT in trial

n – 1). This additional parameter was negative, such that a large RT swing in

trial n – 1 was predictive of a swing in the opposite direction in trial n. In this

model, without Explore being re-set, all genetic findings remained significant,

including the COMT gene-dose Explore effect; P ¼ 0.01.)

A number of models of RT swings were compared in an effort to determine

whether COMT effects were specific to uncertainty.

Sutton (1990) exploration bonus. In this model, exploration is increasingly

encouraged for options that had not been explored for several trials. Specifi-

cally, exploration is predicted to increase with the square root of the number of

trials since making that choice, scaled by free parameter z:

RT0ðs; tÞ ¼ RTðs; tÞ+ BffiffiffiffiffiffiffiðnÞ

pif RTðs; t � 1Þ . . .RTðs; t � nÞoRTavgðt � iÞ

RTðs; tÞ � BffiffiffiffiffiffiffiðnÞ

potherwise

�

‘‘Lose-switch’’ model. In this model, RT swings are predicted to occur after

negative prediction errors, such that participants switch to a slower response if

the previous response was fast and vice versa. The degree of adaptation was

scaled by free parameter k.

RT0ðs; tÞ ¼RTðs; tÞ + k if ds; a; t � 1o0; RTðs; t � 1Þo RTavgðt � 1ÞRTðs; tÞ � k if ds; a; t � 1o0; RTðs; t � 1Þ � RTavgðt � 1ÞRTðs; tÞ otherwise

8<:

‘‘Regression to the mean’’ model. Here responses are predicted to speed or

slow as a function of whether the previous response was faster or slower than

the local mean, regardless of the outcome. The degree of adaptation was scaled

by free parameter x.

RT0ðs; tÞ ¼ RTðs; tÞ+ x if RTðs; t � 1Þo RTavgðt � 1ÞRTðs; tÞ � x if RTðs; t � 1Þ � RTavgðt � 1Þ

�

where RT¢(s,t) is the new RT prediction including regression to the mean.

‘‘Reverse momentum’’ model. This model attempts to capture periodic

changes in RT whereby subjects reverse the direction of their responses if they

had progressively sped up or slowed down over the last number of trials. The

degree of RT adjustment was predicted to linearly increase with the number of

preceding responses that had been progressively speeded or slowed, and scaled

by a free parameter g. Further, this RT reversal was predicted to occur only if the

number of progressively speeded or slowed responses exceeded a minimum

doi:10.1038/nn.2342 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

threshold y, also a free parameter (this parameter allows for variability in the

period of RT swings and was required for the good fits described below).

RT0ðs; tÞ ¼RTðs; tÞ+ gn if RTðs; t � 1ÞoRTðs; t � 2Þo . . .RTðs; t � nÞ . . . ; n4yRTðs; tÞ � gn if RTðs; t � 1Þ4RTðs; t � 2Þ4 . . .RTðs; t � nÞ . . . ; n4yRTðs; tÞ otherwise

8<:

Model comparison results are presented in the Supplementary Data Analysis.

50. Ye, S., Dhillon, S., Ke, X., Collins, A.R. & Day, I.N. An efficient procedure forgenotyping single nucleotide polymorphisms. Nucleic Acids Res. 29, e88-1–e88–8(2001).

NATURE NEUROSCIENCE doi:10.1038/nn.2342

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Targeted disruption of cocaine-activated nucleusaccumbens neurons prevents context-specificsensitizationEisuke Koya1, Sam A Golden1, Brandon K Harvey1, Danielle H Guez-Barber1, Alexander Berkow1,Danielle E Simmons1, Jennifer M Bossert1, Sunila G Nair1, Jamie L Uejima1, Marcelo T Marin1,Timothy B Mitchell1, David Farquhar2, Sukhen C Ghosh2, Brandi J Mattson1 & Bruce T Hope1

Learned associations between effects of abused drugs and

the drug administration environment are important in drug

addiction. Histochemical and electrophysiological studies

suggest that these associations are encoded in sparsely

distributed nucleus accumbens neurons that are selectively

activated by drugs and drug-associated cues. Although

correlations have been observed between nucleus accumbens

neuronal activity and responsivity to drugs and drug cues, no

technique exists for selectively manipulating these activated

neurons and establishing their causal role in behavioral effects

of drugs and drug cues. Here we describe a new approach,

which we term the ‘Daun02 inactivation method’, that

selectively inactivates a minority of neurons previously activated

by cocaine in an environment repeatedly paired with cocaine

to demonstrate a causal role for these activated neurons in

context-specific cocaine-induced psychomotor sensitization

in rats. This method provides a new tool for studying the causal

roles of selectively activated neurons in behavioral effects of

drugs and drug cues and in other learned behaviors.

Learned associations between the effects of abused drugs and the drugadministration environment play important roles in drug addiction1–9.This form of learning involves the encoding of highly detailed informa-tion about drug effects and complex sets of cues in the drug admin-istration environment, and it thus requires a correspondingly highdegree of specificity in the underlying pattern of neural activity.Identification and manipulation of the activated neurons that mediatethese learned associations are crucial to understanding the neurobio-logy of drug addiction.

Neuronal activity within the nucleus accumbens mediates many drug-related learned behaviors1,3,10,11, including context-specific psychomotorsensitization7,8,12,13. Psychomotor sensitization refers to the progressiveincrease in cocaine-induced locomotor activity and stereotypy afterrepeated drug exposure14,15, whereas ‘context-specific’ psychomotorsensitization refers to psychomotor sensitization that is selectively

expressed in the drug-associated environmental context but not in a non-drug-paired environment7–9,16. Results from histochemical studies usingimmediate-early gene markers of neuronal activation indicate that afterrepeated exposure to cocaine in a specific environmental context, onlya minority of sparsely distributed accumbens neurons are selectivelyactivated by cocaine in the drug-paired environment but not in a non-drug-paired environment17–20. Thus context-specific neuronal activationcorrelates with context-specific expression of psychomotor sensitization18.

The implication of the above findings is that only a small number ofactivated nucleus accumbens neurons mediate context-specific psycho-motor sensitization and other behavioral effects of drugs that are modu-lated by the drug administration environment. Unfortunately, currentexperimental methods such as brain lesioning, intracranial injectionsof pharmacological agents or electrical stimulation target both activatedand non-activated neurons, and thus they cannot be used to selectivelymanipulate the activated neurons and assess a causal role for these neuronsin behavior. This issue is partly addressed by in vivo electrophysiologystudies that have found temporal correlations between nucleus accumbensneuronal activity and exposure to drugs or drug-associated cues21–23.However, results from these electrophysiology studies, like those fromhistochemical studies, are correlational and do not by themselves indicatewhether the activated neurons play causal roles in behavior.

Here, we describe a new method, termed the ‘Daun02 inactivationmethod’, that we used to selectively inactivate only those neurons activatedby cocaine in an environment repeatedly paired with drug injections.Our results indicate that small subsets (2–3%) of nucleus accumbensneurons are selectively activated by cocaine in a specific environmentand mediate context-specific psychomotor sensitization.

RESULTS

We used context-specific sensitization of cocaine-induced locomotoractivity as an animal model of learned associations between the drug andits administration environment7,9. In this model, cocaine is administeredrepeatedly to rats in a novel environment outside their home cage.Subsequent cocaine-induced locomotor responses are robustly sensitized

Received 1 April; accepted 9 June; published online 20 July 2009; doi:10.1038/nn.2364

1Behavioral Neuroscience Branch, Intramural Research Program, National Institute on Drug Abuse, National Institutes of Health, Department of Health and Human Services,Baltimore, Maryland, USA. 2Department of Experimental Therapeutics, University of Texas M.D. Anderson Cancer Center, Houston, Texas, USA. Correspondence should beaddressed to B.T.H. (bhope@intra.nida.nih.gov).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1069

TECHNICAL REPORT

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

in the drug-paired environment, but not in distinct non-drug-pairednovel environments, for at least 6 months after the last cocaineexposure9,13,18. In Experiment 1, we injected one set of rats (Paired)with cocaine or saline once daily for 7 d in square locomotor activitychambers with smooth floors (context A) and another set of rats (Non-paired) in round bowls with wood chip bedding (context B). Repeatedcocaine injections increased cocaine-induced locomotor activity levelsfrom day 1 to day 7 in the activity chambers (37%) and round bowls(39%) (data not shown). Seven days later, on test day, we injected all ratsin the locomotor activity chambers with either cocaine or saline. In thepaired groups (Fig. 1a), repeated cocaine injections enhanced, and thussensitized, cocaine-induced locomotor activity, whereas cocaine-inducedlocomotion was not enhanced in the non-paired groups (Fig. 1b; three-way ANOVA interaction of prior drug exposure (cocaine, saline), testinjection (cocaine, saline) and test environment (paired, non-paired):F1,56 ¼ 8.9, P ¼ 0.004). Thus, cocaine test injections elicited a sensitizedlocomotor response only in the environment previously paired withrepeated cocaine injections, indicating that a learned association betweenthe effects of cocaine and the environment in which it was administeredis critical for the expression of this form of sensitization.

We had previously identified a neural correlate of context-specificsensitization to cocaine in the nucleus accumbens13. The ability ofcocaine to induce the neural activity marker Fos24–27 in nucleusaccumbens is strongly enhanced for at least 6 months after sensiti-zation in a novel environment13,17,20. Here, we found that specificenvironments modulate this enhanced Fos response. Repeated cocaineinjections enhanced cocaine-induced Fos expression in the Pairedgroups (Fig. 1c) but not in the Non-paired groups (Fig. 1d; three-way ANOVA interaction of prior drug exposure (cocaine, saline), testinjection (cocaine, saline) and test environment (paired, non-paired):F1,55 ¼ 7.7, P ¼ 0.008). As with the locomotor response, cocaine testinjections produced an enhanced Fos response only in the environ-ment previously paired with repeated cocaine injections. Consideredalongside our previous findings that the cocaine administration envir-onment modulates not only neuronal activation but also the specificpattern of neurons that are activated during expression of context-specific sensitization18, the present finding indicates that the activity ofa small proportion of nucleus accumbens neurons correlates with, andcould mediate, learned associations underlying this behavior.

However, it remains unknown whether these activated neurons play acausal role in mediating the context-specific sensitized response. Untilnow, no tools existed to selectively manipulate the small number of

sparsely distributed neurons that are activated during a drug-relatedlearned behavior. This difficulty is exemplified by the results of doublelabeling for Fos and the neuronal marker NeuN: only 2–3% of allnucleus accumbens neurons were activated by cocaine during thesensitized locomotor response (Fig. 1e, Supplementary Fig. 1). Howcan one manipulate these neurons without affecting the surroundingneurons? We developed a new technique for inactivating only thosenucleus accumbens neurons activated during context-specific sensitizedlocomotion. We used c-fos–lacZ transgenic rats developed by J. Morganand T. Curran (St. Jude Children’s Hospital, Memphis, Tennessee, USA),which contain a transgene with a c-fos promoter regulating expression ofthe bacterial lacZ gene, encoding the protein b-galactosidase28 (Fig. 2a)that can be detected with X-gal staining (see below and; Supplemen-tary Fig. 2). When injected into cocaine-sensitized c-fos–lacZ rats inthe drug-paired environment, cocaine induced b-galactosidase exclu-sively in Fos-expressing nucleus accumbens neurons, as determined byquantitative analysis of the colocalization of Fos and b-galactosidasein the nucleus accumbens; likewise, all Fos was expressed only inb-galactosidase nuclei (Fig. 2b). This allowed us to use a prodrug calledDaun02, which is converted by b-galactosidase to daunorubicin29

(Fig. 2a)—a compound shown to reduce calcium ion (Ca2+)-dependentaction potentials in neuroblastoma cells30—to inactivate neuronsexpressing both b-galactosidase and Fos. Thus, we could selectivelyinactivate b-galactosidase–expressing neurons activated in thepresence of cocaine and paired stimuli in the administrationenvironment and monitor the degree of inactivation by measuringb-galactosidase expression.

The timeline for all Daun02 inactivation experiments is shown inFigure 3a. In Experiment 2, we sensitized c-fos–lacZ rats to cocaineusing the same procedure described above for wild-type rats. Repeatedcocaine injections increased cocaine-induced locomotor activity levelsby B38% from day 1 to day 7 (F1,58 ¼ 21.3, P o 0.0001), which iscomparable to the increase seen in wild-type rats in Experiment 1.Seven days later, on induction day, we injected the rats with eithercocaine to induce b-galactosidase or saline as a control that does notsignificantly induce b-galactosidase. Ninety minutes later, when b-galactosidase was near maximal levels, we bilaterally infused Daun02 orvehicle into the nucleus accumbens and then returned the rats to theirhome cages for a further 3 d to allow cell-specific inactivation to occur.On test day, we injected all rats with cocaine to allow us to assesscocaine-induced locomotor activity and nucleus accumbens b-galac-tosidase expression. Thus, all values shown in Figure 3b,c indicate

Figure 1 Environment modulates cocaine-induced

locomotor activity and nucleus accumbens

neuronal activity in sensitized rats (Experiment 1).

(a,b) Prior repeated cocaine injections enhanced

cocaine-induced locomotor activity in Paired rats

that received cocaine injections (15 mg per kg

body weight, intraperitoneally (i.p.)) in the same

locomotor activity chambers (context A indicatedschematically by a square; panel a), but not in

Non-paired rats that received the same repeated

cocaine injections in a different environment

(context B indicated schematically by a circle;

panel b). (c,d) In parallel, prior repeated cocaine

injections enhanced cocaine-induced Fos

expression in Paired rats (c) but not in Non-paired

rats (d). IR, immunoreactive. Values are expressed

as the mean ± s.e.m. using n ¼ 7–8 rats per

group. * indicates a significant difference relative

to cocaine challenge following repeated saline administration. (e) Fos and NeuN immunohistochemistry in the nucleus accumbens of sensitized rats after the

cocaine test injection in the Paired environment. Green-labeled nuclei indicate Fos expression; red-labeled nuclei indicate expression of the general neuronal

nuclei marker NeuN; green/yellow-labeled nuclei are double-labeled for both Fos and NeuN expression. White bar, 100 mm.

Anterior commissure

Fos-IR

NeuN-IR

Repeated injectionsSalineCocaine

Dis

tanc

etr

avel

ed (

m)

100

200

300

Paired

*

0

Non-paireda b

Saline Cocaine Saline CocaineTest injection Test injection

Fos

-IR

nuc

lei

per

mm

2

0

20

40

60

80 *c d

e

Repeatedinjections

A A AB

Repeatedinjections

Testinjection

Testinjection

Experiment 1: Context-specific sensitization of cocaine-induced locomotion and Fos

1070 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

T ECHNICAL REPORT

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

levels after cocaine injections on test day, and ‘Saline’ and ‘Cocaine’ onthe x axes refer to injections on induction day, 3 d before test day.

Nucleus accumbens injections of Daun02 attenuated cocaine-induced locomotor activity (Fig. 3b) in rats that previously receivedcocaine, but not saline, on induction day (interaction of Daun02infusions (Daun02, vehicle) with induction-day injection (cocaine,saline): F1,67 ¼ 5.03, P ¼ 0.028). Indeed, Daun02 reduced cocaine-induced locomotor activity to levels similar to the acute cocaineresponse observed on day 1 of the sensitization treatment. Daun02also attenuated cocaine-induced b-galactosidase expression (Fig. 3c) inrats that had previously received cocaine, but not saline, on inductionday (interaction of Daun02 infusions with induction-day injection:F1,53 ¼ 4.52, P ¼ 0.038). For perspective, when the number ofb-galactosidase nuclei induced following cocaine injections (Fig. 3c)is compared to the number induced following saline control injections(Experiment 4; Supplementary Fig. 3), Daun02 treatment resulted in50% fewer b-galactosidase nuclei induced specifically by cocaine. As ananatomical control, infusions of Daun02 into caudate putamen 1 mmdorsal to our test site did not alter sensitized locomotor responses tococaine test injections (P ¼ 0.36). In addition, when wild-type rats were

treated with the same procedure used in Experiment 2, Daun02 did nothave a significant effect on cocaine-induced locomotor activity, whichfurther indicates a lack of nonspecific effects on neurons that do notproduce b-galactosidase (Supplementary Fig. 4). Altogether, thesefindings demonstrate that Daun02 inactivation requires cocaine-induced b-galactosidase found in strongly activated nucleus accumbensneurons. If Daun02 had produced nonspecific inactivation of nucleusaccumbens neurons, then Daun02 infusions would also have attenuatedcocaine-induced locomotor activity and b-galactosidase expression inrats injected with saline on induction day, which was not the case. Wenext examined whether Daun02-mediated attenuation of cocaine-induced locomotion and b-galactosidase was associated with non-specific effects in the nucleus accumbens. To assess the possibility thatDaun02-mediated effects were due to generalized neurotoxicity in theaccumbens, we examined cresyl violet–stained sections of nucleusaccumbens from rats in Experiment 2. These sections did not show anygross alterations or decrease in size of nucleus accumbens architecture orany indication of gliosis at higher-level magnification (SupplementaryFig. 5). To examine the possibility that Daun02 effects were due tononspecific impairments of accumbens function, we assessed the effectsof Daun02 on nonspecific activation of nucleus accumbens neuronsin Experiment 3. We treated rats identically to those in Experiment 2except that we infused a mixture of AMPA and picrotoxin (a GABAA

receptor antagonist) into the nucleus accumbens on test day to producerobust (over 12-fold more b-galactosidase expression than cocaine)nonspecific activation of accumbens neurons. Daun02 did not affect

Figure 3 Timeline for Daun02 inactivationexperiments. (a) Rats were injected once daily

for 7 d with cocaine (15 mg per kg body weight

(mg/kg), i.p) or saline. On induction day, after 7 d

of withdrawal, the rats were injected with either

cocaine (20 mg/kg) or saline. Ninety minutes

later, rats received an intracranial infusion of 2 mg

of Daun02 or vehicle bilaterally into the nucleus

accumbens (Acb). On test day, 3 d after

intracranial infusions, the rats were injected with

cocaine (15 mg/kg) or saline. In Experiment 2,

Daun02 attenuated cocaine-induced locomotion

and b-galactosidase expression in nucleus

accumbens of c-fos–lacZ rats. Rats were

repeatedly injected with cocaine from day 1 to

day 7, and 7 d later, on induction day, rats were

injected with cocaine or saline and then given

nucleus accumbens Daun02 or vehicle infusions.

On test day, 3 d later, all rats were injectedwith cocaine. (b,c) Daun02 attenuated cocaine-

induced locomotor activity (b) and nucleus

accumbens b-galactosidase levels (c) on test day when rats had been previously injected with cocaine, but not saline, on induction day. Black dashed line indicates

maximum levels of cocaine-induced b-galactosidase (as determined from vehicle induction group) and gray dashed line indicates baseline b-galactosidase levels

(as determined from saline test injection group in Experiment 4; Supplementary Fig. 3). Values are expressed as mean ± s.e.m. distance traveled during 1 h

following cocaine test injections (n ¼ 14–21) and mean ± s.e.m. density of b-galactosidase–labeled nuclei in nucleus accumbens (n ¼ 14–16). * indicates

a significant difference relative to that in rats infused with vehicle following cocaine injections on induction day.

Daun02(prodrug)

Daunorubicin(reduces excitability)

a bNeuronal activity

c-fos promoter

c-fos 3′ sequence

lacZ

c-fos 5′ sequence

Fos

Merged

β-gal

β-gal

- 3-Nitro-4-hydroxy-

- CO2

O OH

OH

OH

OH

O O

HO

HO

O

O

OO

O

O

C

OH

CH3

CH3

NHC

CH3O

O OH

OH

O

OO

C

OH

CH3

CH3O

CH2CH2OH

O2N

OH

HO

OO

OCH3

NH

O OH

OH

O

OO

C

OH

CH3

CH3O

HO

OCH3

NH2

CCH2

O2N

benzyl alcohol

a

b c

Saline Cocaine

VehicleDaun02

*

Induction-day injection

Cocaine-induced locomotion

0

100

60

80

40

20

Saline Cocaine

β-ga

l nuc

lei p

er m

m2

Cocaine-induced nucleusaccumbens β-galactosidase

*

Induction-day injection

VehicleDaun02

Dis

tanc

e tr

avel

ed (

m)

100

200

300

400

0

Experiment 2: Daun02 effects on cocaine-sensitized locomotion and β-galactosidase

Experimental outline: Daun02 inactivationRepeated injections

(context A)

Exp. 2: CocaineExp. 3: Cocaine Exp. 4: CocaineExp. 5: SalineExp. 6: Cocaine

Cocaine or SalineCocaineCocaineCocaineCocaine

Injection:

Acb injection:Daun02 or vehicle

Context:

AAAA

A or B

Induction dayTest day

(context A)

7 d 90 min 3 d

CocaineAcb AMPA+picrotoxin

SalineCocaineCocaine

Figure 2 Schematic mechanism for Daun02 inactivation in c-fos–lacZ rats.

(a) The c-fos–lacZ transgene contains a c-fos promoter that drives expression

of lacZ, which encodes the protein b-galactosidase (b-gal) (adapted from

ref. 39). b-galactosidase can catalyze the conversion of the prodrug

Daun02 into daunorubicin (adapted from ref. 29), which reduces cellular

excitability30. (b) Cocaine-induced neuronal activity induces b-galactosidase

expression (red-labeled nuclei) and Fos expression (green-labeled nuclei)

in neurons in the nucleus accumbens of sensitized c-fos–lacZ rats. Nucleidouble-labeled for both b-galactosidase and Fos appear yellow to orange in

the Merged image panel and indicate colocalization of b-galactosidase and

Fos proteins. White bar, 50 mm.

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1071

TECHNICAL REPORT

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

AMPA+picrotoxin-induced locomotion (Fig. 4a, P ¼ 0.94) or thenumber of b-galactosidase–labeled nuclei (Fig. 4b, P ¼ 0.39, imagesin 4c). This lack of effect indicates that when cocaine was injectedon induction day in Experiment 2, subsequent Daun02-dependentdecreases in cocaine-induced locomotion and b-galactosidase were notdue to nonspecific inactivation of accumbens neurons. Taken together,our results indicate that Daun02 inactivation after expression of context-specific sensitization does not appear to produce significant nonspecificeffects on nucleus accumbens structure and function.

In Experiment 4, we assessed the effects of Daun02 inactivation onspontaneous or novelty-induced locomotor activity. Daun02 infusionson induction day did not alter locomotor activity (P ¼ 0.65) orb-galactosidase expression (P ¼ 0.72) after saline test injections(Supplementary Fig. 3). This finding indicates that Daun02inactivation of nucleus accumbens neurons is specific for onlycocaine-induced locomotor activity and b-galactosidase expression.

In Experiment 5, we examined whether Daun02 inactivation waseffective after an acute cocaine injection. We repeatedly injected saline,instead of cocaine, into c-fos–lacZ rats and then on induction day injectedthe rats with cocaine and then provided infusions of Daun02 or vehicle inthe nucleus accumbens. On test day, Daun02 did not alter subsequentcocaine-induced locomotion (P ¼ 0.61; Supplementary Fig. 6), indicat-ing that Daun02 inactivation requires prior sensitization to cocaine.

As shown in Experiment 1, the learned association mediatingcontext-specific locomotor sensitization is expressed only in thecocaine-paired environment. By extension, the specific neurons med-iating this learned association should also be activated only in thecocaine-paired environment. Therefore, cocaine injections in an alter-nate non-paired environment on induction day should not activatethese context-specific activated neurons, and consequently Daun02should not attenuate subsequent sensitized responding in the pairedenvironment on test day. We tested this prediction in Experiment 6, inwhich we administered repeated cocaine injections to c-fos–lacZ rats inlocomotor activity chambers (context A) according to our usual

procedure, but on induction day we injected half of these rats withcocaine in the paired environment (context A) and the other half withcocaine in an alternate non-paired environment (context B) followed,in each case, by nucleus accumbens infusions of Daun02 or vehicle. Ontest day, we injected all rats with cocaine in context A and assessedlocomotor activity. Daun02 attenuated cocaine-induced locomotoractivity (Fig. 5; interaction of Daun02 infusions (Daun02, vehicle)with induction-day context (A, B): F1,55 ¼ 5.94, P ¼ 0.018) in rats thatreceived cocaine on induction day in the paired environment (contextA), but not in rats that received cocaine in the alternate non-pairedenvironment (context B). The lack of effects from Daun02 wheninfused in the alternate environment indicates that the neuronsselectively activated by cocaine in the cocaine-paired environmentremained intact and preserved the learned association.

DISCUSSION

Our Daun02 inactivation procedure allowed us to demonstrate a causalrole between a small subset of nucleus accumbens neurons selec-tively activated by cocaine in the drug-paired environment andcontext-specific locomotor sensitization to the drug. Daun02 disruptedbehavior only when these neurons were activated by cocaine in thepaired environment and did not produce nonspecific impairments innucleus accumbens function or structure. We found that Daun02 waseffective only when neurons were activated by injection of cocaine, butnot saline, on induction day. Daun02 was effective in attenuatingcocaine-induced, but not novelty-induced or AMPA+picrotoxin-induced, locomotion and accumbens activity on test day. Daun02was effective only in reducing sensitized cocaine-induced locomotoractivity but not locomotion following the first acute injection ofcocaine. Most importantly, Daun02 required context-specificcocaine-induced activation of a small subset of nucleus accumbensneurons in the drug-paired environment but had no effect whencocaine was injected in an alternate non-paired environment oninduction day. Overall, the dependence of Daun02 on stimuli-inducedendogenous neuronal activity in behaving animals makes it possible toexamine the role of context-specific activation of a small proportion ofselectively activated neurons in drug-induced behavior.

The learned association between cocaine and the drug administra-tion environment following context-specific locomotor sensitization tococaine appears to be mediated by enhanced activation of a smallnumber of sparsely distributed neurons in the nucleus accumbens.

Experiment 6: Daun02 effects in different contexts

Cocaine-induced locomotion

VehicleDaun02

Induction-day context

100

300

200

0Dis

tanc

e tr

avel

ed (

m)

BA

*

Figure 5 Experiment 6: Although Daun02 attenuated locomotor activity when

cocaine was previously injected in the Paired environment (context A), it did

not alter cocaine-induced locomotor activity on test day in c-fos–lacZ rats

that had been injected on induction day with cocaine in an alternate non-

paired environment (context B). Cocaine-induced locomotion on test day wassignificantly higher than on day 1 for both context A and context B groups

infused with vehicle. Values are expressed as mean ± s.e.m. distance traveled

during 1 h after test injections (n ¼ 10–15 for context A; n ¼ 17 for context

B). * indicates a significant difference relative to that in rats infused with

vehicle following cocaine injections on induction day.

Dis

tanc

e tr

avel

ed (

m)

AMPA+picrotoxin-induced locomotion

β-ga

l nuc

lei p

er m

m2

AMPA+picrotoxin-inducedaccumbens β-galactosidase

0

50

100

150

200

250

0

500

1,000

1,500a

c

b

Cocaine+Daun02Cocaine+vehicle

Induction-day treatment AMPA+picrotoxin-induced β-galactosidase in nucleus accumbens

Experiment 3: Daun02 did not alter nonspecific nucleus accumbens activation

Vehicle Daun02Vehicle Daun02

Figure 4 Experiment 3. (a,b) Daun02 infusion on induction day did not

alter locomotor activity (a) and b-galactosidase expression (b) in nucleus

accumbens of c-fos–lacZ rats following nucleus accumbens AMPA+picrotoxin

infusions on test day. (c) Representative images of X-gal staining for

visualization of b-galactosidase. Black bar, 100 mm. Values are expressed as

mean ± s.e.m. distance traveled during 1 h and mean ± s.e.m. density of

b-galactosidase–labeled nuclei in nucleus accumbens following intra-

accumbens administration of AMPA+picrotoxin (n ¼ 7–8).

1072 VOLUME 12 [ NUMBER 8 [ AUGUST 2009 NATURE NEUROSCIENCE

T ECHNICAL REPORT

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

Results from our previous study indicate that the exact pattern of theseactivated neurons is determined at least in part by cocaine-inducedinteroceptive cues and associated environmental cues present duringdrug injections18. Following a cocaine sensitization treatment in thisprevious study, test injections of cocaine in the formerly cocaine-pairedenvironment activated 87% of the same neurons that were repeatedlyactivated in nucleus accumbens during prior repeated cocaine admin-istration. This overlap in activation pattern was significantly less whentest injections of cocaine were administered in an alternate unpairedenvironment. The results from this previous study are congruent withour present results indicating that Daun02 disrupted sensitized loco-motion only when cocaine was injected in the cocaine-paired environ-ment on induction day but not when it was injected in an alternateunpaired environment. Together these studies lead us to speculate thatduring repeated cocaine injections in a novel environment, the learnedassociation between environment and cocaine effects becomes encodedin specific neuronal ensembles31,32.

Our Daun02 inactivation technique has the broader potential toreveal the causal roles that selectively activated subsets of neurons playin a variety of learned behaviors, such as cue-induced drug seeking33–35,fear conditioning36 and spatial learning37. By revealing behaviorallyrelevant alterations produced in a small number of selected neurons,the present technique in rats and a recently reported technique inmice38, combined with electrophysiological and molecular character-ization of these neurons during a learning task, provide new tools tostudy the neuronal mechanisms of learned behaviors.

METHODS

Methods and any associated references are available in the onlineversion of the paper at http://www.nature.com/natureneuroscience/.

Note: Supplementary information is available on the Nature Neuroscience website.

ACKNOWLEDGMENTSThis research was supported by the Intramural Research Program of the NationalInstitute on Drug Abuse. We thank E. Wentzell for her excellent editorial assistance,K. Knestaut for breeding the cfos-lacZ rats and T. Adams-Deutsch, C. Pickensand K. Wihbey for technical assistance. We are especially thankful to Y. Shahamfor his comments over the years and for his help in writing this manuscript, andto R. Wise for providing the inspiration for this study.

AUTHOR CONTRIBUTIONSE.K. and B.T.H. designed the behavioral experiments. E.K. performed theseexperiments with S.A.G., A.B., D.E.S., J.M.B., S.G.N., J.L.U., M.T.M., T.B.M.,B.J.M. and B.T.H. E.K. performed the histological experiments. D.F. and S.C.G.synthesized the Daun02 compound. B.K.H. and D.H.G.-B. examined the prodrugproperties of Daun02. E.K. and B.T.H. analyzed the data and wrote the paper.

Published online at http://www.nature.com/natureneuroscience/.

Reprints and permissions information is available online at http://npg.nature.com/

reprintsandpermissions/.

1. Everitt, B.J. & Robbins, T.W. Neural systems of reinforcement for drug addiction: fromactions to habits to compulsion. Nat. Neurosci. 8, 1481–1489 (2005).

2. Siegel, S. Morphine analgesic tolerance: its situation specificity supports a Pavlovianconditioning model. Science 193, 323–325 (1976).

3. Wise, R.A. Dopamine, learning and motivation. Nat. Rev. Neurosci. 5, 483–494 (2004).4. Hyman, S.E. & Malenka, R.C. Addiction and the brain: the neurobiology of compulsion

and its persistence. Nat. Rev. Neurosci. 2, 695–703 (2001).5. Crombag, H.S., Bossert, J.M., Koya, E. & Shaham, Y. Review. Context-induced relapse to

drug seeking: a review. Phil. Trans. R. Soc. Lond. B 363, 3233–3243 (2008).6. Stewart, J., de Wit, H. & Eikelboom, R. Role of unconditioned and conditioned drug effects

in the self-administration of opiates and stimulants. Psychol. Rev. 91, 251–268 (1984).7. Badiani, A. & Robinson, T.E. Drug-induced neurobehavioral plasticity: the role of

environmental context. Behav. Pharmacol. 15, 327–339 (2004).8. Vezina, P. & Leyton, M. Conditioned cues and the expression of stimulant sensitization in

animals and humans. Neuropharmacology 56 Suppl 1, 160–168 (2009).

9. Anagnostaras, S.G. & Robinson, T.E. Sensitization to the psychomotor stimulant effectsof amphetamine: modulation by associative learning. Behav. Neurosci. 110,1397–1414 (1996).

10. Koob, G.F. Drugs of abuse: anatomy, pharmacology and function of reward pathways.Trends Pharmacol. Sci. 13, 177–184 (1992).

11. Roberts, D.C., Corcoran, M.E. & Fibiger, H.C. On the role of ascending catecholaminer-gic systems in intravenous self-administration of cocaine. Pharmacol. Biochem. Behav.6, 615–620 (1977).

12. Delfs, J.M., Schreiber, L. & Kelley, A.E. Microinjection of cocaine into the nucleusaccumbens elicits locomotor activation in the rat. J. Neurosci. 10, 303–310 (1990).

13. Hope, B.T., Simmons, D.E., Mitchell, T.B., Kreuter, J.D. & Mattson, B.J. Cocaine-induced locomotor activity and Fos expression in nucleus accumbens are sensitized for6 months after repeated cocaine administration outside the home cage. Eur. J. Neurosci.24, 867–875 (2006).

14. Robinson, T.E. & Becker, J.B. Enduring changes in brain and behavior produced bychronic amphetamine administration: a review and evaluation of animal models ofamphetamine psychosis. Brain Res. 396, 157–198 (1986).

15. Vanderschuren, L.J. & Kalivas, P.W. Alterations in dopaminergic and glutamatergictransmission in the induction and expression of behavioral sensitization: a critical reviewof preclinical studies. Psychopharmacology (Berl.) 151, 99–120 (2000).

16. Vezina, P., Giovino, A.A., Wise, R.A. & Stewart, J. Environment-specific cross-sensitiza-tion between the locomotor activating effects of morphine and amphetamine. Pharma-col. Biochem. Behav. 32, 581–584 (1989).

17. Crombag, H.S., Jedynak, J.P., Redmond, K., Robinson, T.E. & Hope, B.T. Locomotorsensitization to cocaine is associated with increased Fos expression in the accumbens,but not in the caudate. Behav. Brain Res. 136, 455–462 (2002).

18. Mattson, B.J. et al. Context-specific sensitization of cocaine-induced locomotor activityand associated neuronal ensembles in rat nucleus accumbens. Eur. J. Neurosci. 27,202–212 (2008).

19. Rademacher, D.J., Napier, T.C. & Meredith, G.E. Context modulates the expression ofconditioned motor sensitization, cellular activation and synaptophysin immunoreactiv-ity. Eur. J. Neurosci. 26, 2661–2668 (2007).

20. Todtenkopf, M.S., Mihalakopoulos, A. & Stellar, J.R. Withdrawal duration differentiallyaffects c-fos expression in the medial prefrontal cortex and discrete subregions of thenucleus accumbens in cocaine-sensitized rats. Neuroscience 114, 1061–1069 (2002).

21. Hollander, J.A. & Carelli, R.M. Cocaine-associated stimuli increase cocaine seeking andactivate accumbens core neurons after abstinence. J. Neurosci. 27, 3535–3539 (2007).

22. Carelli, R.M. Activation of accumbens cell firing by stimuli associated with cocainedelivery during self-administration. Synapse 35, 238–242 (2000).

23. Carelli, R.M. & Wightman, R.M. Functional microcircuitry in the accumbens underlyingdrug addiction: insights from real-time signaling during behavior. Curr. Opin. Neurobiol.14, 763–768 (2004).

24. Sgambato, V., Abo, V., Rogard, M., Besson, M.J. & Deniau, J.M. Effect of electricalstimulation of the cerebral cortex on the expression of the Fos protein in the basalganglia. Neuroscience 81, 93–112 (1997).

25. Berretta, S., Robertson, H.A. & Graybiel, A.M. Neurochemically specialized projectionneurons of the striatum respond differentially to psychomotor stimulants. Prog. BrainRes. 99, 201–205 (1993).

26. Morgan, J.I. & Curran, T. Stimulus-transcription coupling in the nervous system:involvement of the inducible proto-oncogenes fos and jun. Annu. Rev. Neurosci. 14,421–451 (1991).

27. Canales, J.J. & Graybiel, A.M. Patterns of gene expression and behavior induced bychronic dopamine treatments. Ann. Neurol. 47, S53–S59 (2000).

28. Kasof, G.M. et al. Kainic acid-induced neuronal death is associated with DNA damageand a unique immediate-early gene response in c-fos-lacZ transgenic rats. J. Neurosci.15, 4238–4249 (1995).

29. Farquhar, D. et al. Suicide gene therapy using E. coli beta-galactosidase. CancerChemother. Pharmacol. 50, 65–70 (2002).

30. Santone, K.S., Oakes, S.G., Taylor, S.R. & Powis, G. Anthracycline-induced inhibition ofa calcium action potential in differentiated murine neuroblastoma cells. Cancer Res. 46,2659–2664 (1986).

31. Pennartz, C.M., Groenewegen, H.J. & Lopes da Silva, F.H. The nucleus accumbensas a complex of functionally distinct neuronal ensembles: an integration of behavioural,electrophysiological and anatomical data. Prog. Neurobiol. 42, 719–761 (1994).

32. Graybiel, A.M. Habits, rituals, and the evaluative brain. Annu. Rev. Neurosci. 31,359–387 (2008).

33. Bossert, J.M., Poles, G.C., Wihbey, K.A., Koya, E. & Shaham, Y. Differential effects ofblockade of dopamine D1-family receptors in nucleus accumbens core or shell onreinstatement of heroin seeking induced by contextual and discrete cues. J. Neurosci.27, 12655–12663 (2007).

34. Crombag, H.S. & Shaham, Y. Renewal of drug seeking by contextual cues after prolongedextinction in rats. Behav. Neurosci. 116, 169–173 (2002).

35. Lu, L., Grimm, J.W., Hope, B.T. & Shaham, Y. Incubation of cocaine craving after with-drawal: a review of preclinical data. Neuropharmacology 47 Suppl 1, 214–226 (2004).

36. Phelps, E.A. & LeDoux, J.E. Contributions of the amygdala to emotion processing: fromanimal models to human behavior. Neuron 48, 175–187 (2005).

37. Moser, E.I. & Paulsen, O. New excitement in cognitive space: between place cells andspatial memory. Curr. Opin. Neurobiol. 11, 745–751 (2001).

38. Han, J.H. et al. Selective erasure of a fear memory. Science 323, 1492–1496 (2009).39. Schilling, K., Luk, D., Morgan, J.I. & Curran, T. Regulation of a fos-lacZ fusion gene: a

paradigm for quantitative analysis of stimulus-transcription coupling. Proc. Natl. Acad.Sci. USA 88, 5665–5669 (1991).

NATURE NEUROSCIENCE VOLUME 12 [ NUMBER 8 [ AUGUST 2009 1073

TECHNICAL REPORT

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

ONLINE METHODSSubjects. Male c-fos–lacZ transgenic rats28 and Sprague-Dawley rats (Charles

River) were housed individually in standard plastic cages in a temperature-

and humidity-controlled room. They were maintained on a 12 h/12 h

reverse light/dark cycle (lights on at 21:00 h) and allowed free access to food

and water. They were acclimatized to these housing conditions for a mini-

mum of 7 d before drug treatments. Experimental procedures were

approved by the National Institute on Drug Abuse Animal Care and

Use Committee.

Experiment 1: cocaine-induced locomotor sensitization in wild-type rats.

Male Sprague-Dawley rats were divided into eight groups of 7–8 rats each. We

administered seven once-daily injections of cocaine (15 mg per kg body weight

(mg/kg), i.p.) or saline to four groups of ‘Paired’ rats in locomotor activity

chambers (context A: square chamber, dark, silent, smooth floor) and to four

groups of ‘Non-paired’ rats in a distinct novel environment (context B: round

chamber, brighter lighting, music by Marilyn Manson played continuously,

wood chip bedding on floor). Seven days later, we placed rats in the locomotor

activity chambers (context A) for 30 min and then injected all rats with cocaine

(20 mg/kg, i.p.) or saline in the locomotor activity chambers and assessed

locomotor activity for 60 min. Two hours after the test injections, all rats were

deeply anesthetized with isoflurane and perfused transcardially with 4%

paraformaldehyde.

Fos immunohistochemistry. Brain sections were processed for Fos

immunohistochemistry and quantified as described previously13,17. We used

antibody to Fos (1:10,000 dilution of Ab-5, Calbiochem) and developed

the sections using an Elite ABC kit from VectorLabs and diaminobenzi-

dine (DAB).

Fos and NeuN double labeling. Eight more rats were treated identically to

the Paired groups above, and all rats were perfused with paraformaldehyde 2 h

after test injections of cocaine (20 mg/kg i.p.) or saline (n ¼ 4). The general

procedure for fluorescence immunohistochemistry and quantification was

similar to that described previously18. Sections were incubated with antibody

to Fos (1:10,000 dilution of Ab-5, Calbiochem) and antibody to NeuN (1:2,000

dilution of MAB377, Chemicon) and labeled with secondary antibodies

conjugated to Alexa 488 and 568 fluorophores, respectively.

b-galactosidase and Fos immunohistochemistry with c-fos–lacZ rats. We

used transgenic c-fos–lacZ rats that had been bred for 35–40 generations on a

Sprague-Dawley background28. We first examined colocalization of Fos and

b-galactosidase expression in these rats. All c-fos–lacZ rats received repeated

cocaine injections as described above for Paired groups. Rats were perfused

with paraformaldehyde 2 h after the test injections of cocaine (20 mg/kg i.p.).

Sections were incubated with antibody to Fos (1:10,000 dilution of Ab-5,

Calbiochem) and polyclonal goat antibody to b-galactosidase (1:1,000 dilution

of 4600-1409, Biogenesis) and labeled with secondary antibodies conjugated to

Alexa 488 and 568 fluorophores, respectively.

Experiments 2–6: implanting cannulae into c-fos–lacZ rats. We implanted

guide cannulae into the nucleus accumbens for Daun02 inactivation experi-

ments using previously described surgical procedures13. Transgenic rats (250–

350 g) were anesthetized with equithesin (including 60 mg/kg sodium

pentobarbital plus 225 mg/kg chloral hydrate; i.p.). For nucleus accumbens

infusions, permanent guide cannulae (23 G; Plastics One) were implanted

bilaterally at a 101 angle with the tip 1 mm dorsal to the nucleus accumbens;

coordinates were anteroposterior +1.6 mm, mediolateral ± 3.0 mm and

dorsoventral �6.0 mm relative to bregma40. Guide cannula coordinates for

control infusions into the caudate putamen were anteroposterior +1.6 mm,

mediolateral ± 3.0 mm and dorsoventral �5.0 mm relative to bregma. After

cannula implantation, rats were allowed to recover for at least 5 d before

experiments were begun.

Experimental procedure for Daun02 inactivation. The detailed timeline for

Daun02 experiments is shown in Figure 3a. In Experiment 2, c-fos–lacZ rats

received seven once-daily injections of cocaine (15 mg/kg i.p.), using the same

procedure described above for Paired rats. On ‘induction day’, 7 d later,

we injected all rats with cocaine (20 mg/kg, i.p.) or saline in the locomotor

activity chambers. Ninety minutes later, 2 mg of Daun02 (ref. 29) or vehicle

(50% artificial cerebrospinal fluid (ACSF; composition in mM: NaCl, 148;

KCl, 2.7; CaCl2, 1.2; MgCl2, 0.8; pH 7.4), 50% DMSO) in a 0.5-ml volume was

bilaterally infused over 2 min into the nucleus accumbens or the dorsal

control site (the caudate putamen). Injectors extended 1 mm beyond the end

of the guide cannulae. Accumbens cannula placement was verified in

cresyl violet–stained sections (Supplementary Fig. 7). Rats were returned to

their home cages for 3 d after infusions—the 3-d time point being chosen

based on unpublished pilot studies on the time course of the effectiveness of

Daun02. On test day, rats were placed in locomotor activity chambers for

30 min and then injected with cocaine (15 mg/kg i.p.), and locomotor activity

was monitored for 60 min. Rats were perfused with paraformaldehyde 2 h after

test injections. Brain sections were processed for X-gal histochemistry.

For Experiments 3–6 and wild-type controls, the same experimental timeline

(for example, once-daily repeated injections, induction day and test day) and

cocaine doses were used, but with the following modifications. For Experiment

3, the same procedure was used except that on test day rats were infused with

AMPA (1.0 mM, 00-5, Ascent Scientific) plus picrotoxin (2.7 mM, P176,

Sigma) bilaterally in the nucleus accumbens. For wild-type controls, two

groups of wild-type rats were repeatedly injected with cocaine once daily for

7 d, and on induction day all rats were injected with cocaine and then

given Daun02 or vehicle infusions. On test day, all rats were then injected

with cocaine. In Experiment 4, rats were repeatedly injected with cocaine

once daily for 7 d, and on induction day all rats were injected with cocaine

and then given Daun02 or vehicle infusions. On test day, all rats were then

injected with saline. In Experiment 5, rats were repeatedly injected with

saline once daily for 7 d, and on induction day all rats were injected with

cocaine and then given Daun02 or vehicle infusions. In Experiment 6, rats

were repeatedly injected with cocaine once daily for 7 d in locomotor activity

boxes (context A). On induction day, half of these rats were injected

with cocaine in the round chambers (context B; described above for Experi-

ment 1) while the other half were injected with cocaine in the activity

boxes (context A). Each half was further subdivided into two groups that

received either Daun02 or vehicle infusions into the nucleus accumbens. On

test day, all rats were injected with cocaine in the locomotor activity boxes

(context A).

X-gal histochemistry. Coronal sections (30 mm) were cut in a cryostat at

�20 1C and collected in 10 mM PBS. Free-floating sections were washed three

times for 10 min each in PBS and then incubated in reaction buffer (0.1 M

X-gal, 100 mM sodium phosphate, 100 mM sodium chloride, 5 mM EGTA,

2 mM MgCl2, 0.2% Triton X-100, 0.05 M K3FeCN6, 0.05 M K4FeCN6) for 3 h

at 37 1C with gentle shaking. Sections were washed with three times for 10 min

each in PBS and mounted onto chrom-alum–coated slides. Slides were dried,

dehydrated through a graded series of alcohol (70%, 95%, 95%, 100% and

100% ethanol) and cleared with Citrasolv (Fisher Scientific) before cover-

slipping with Permount (Fisher Scientific).

Quantification of histochemical results. Bright-field images of nucleus

accumbens were captured under a 5� objective lens and digitized using a

CCD camera (Coolsnap Photometrics, Roper Scientific Inc.) attached to a Zeiss

Axioskop 2 light microscope. b-galactosidase–expressing nuclei characterized

by blue nuclear staining were counted from these images using IPLab software

for Macintosh (Scanalytics) (Supplementary Fig. 2). We chose a threshold level

that detected moderate to darkly stained nuclei but not lightly stained nuclei.

We counted nuclei from three rectangular sampling areas of 0.15 mm2 each

around the nucleus accumbens injection site from 2–3 coronal sections per rat.

Image capture and quantification were conducted by an observer blind to the

experimental conditions.

Data analysis. Locomotor activity and histochemical data were analyzed using

factorial two- or three-way analyses of variance (ANOVA). Fisher’s PLSD

post hoc tests were used to make specific group comparisons when appropriate.

Effects were considered significant when P o 0.05. In Experiments 2–6, 12 out

of the 199 rats that ran o70 m for 2 d or more during repeated cocaine

injections were eliminated from the study due to abnormally low cocaine

NATURE NEUROSCIENCE doi:10.1038/nn.2364

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.

responses during sensitization. Of the remaining 187 rats, 8 rats with test-day

locomotion values 2 s.d. beyond the mean were excluded. Rats were also

excluded from analysis when injection sites were outside the intended bound-

aries, for example when they were too rostral (B2.5 mm anterior to bregma) or

too caudal (B1.2 mm anterior to bregma). The rostral nucleus accumbens

was chosen because nearly all cocaine-induced Fos is expressed within the

rostral but not caudal portions of the nucleus accumbens17. Finally, we

excluded rats with obvious signs of necrosis and rats that we later determined

to be c-fos–lacZ negative.

40. Paxinos, G. & Watson, C. The Rat Brain in Stereotaxic Coordinates (Academic Press,New York, 1998).

doi:10.1038/nn.2364 NATURE NEUROSCIENCE

©20

09 N

atu

re A

mer

ica,

Inc.

All

rig

hts

res

erve

d.