nature neuroscience july 2002

editorialConstitutional protection for animals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611

news and viewsMagnetic stimulation reveals the distribution of language in a normal population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613Glyn W. Humphreys and Peter PraamstraSEE ARTICLE, PAGE 695

NMDA receptors lose their inhibitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614David M. LovingerSEE ARTICLE, PAGE 641

Short circuiting the circadian clock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 616Anthony N. van den Pol and Karl Obrietan

Be caught napping: you’re doing more than resting your eyes . . . . . . . . . . . . . 618Pierre Maquet, Philippe Peigneux, Steve Laureys and Carlyle SmithSEE ARTICLE, PAGE 677

book reviewTo slice or not to slice? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621The Intact and Sliced Brainby Mircea SteriadeREVIEWED BY DAVID A. MCCORMICK

brief communicationsRegulation of dendritic development by the ARF exchange factor ARNO . . . . . 623Delia J. Hernández-Deviez, James E. Casanova and Jean M. Wilson

Neurobiological evidence for hedonic allostasis associated with escalating cocaine use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625Serge H. Ahmed, Paul J. Kenny, George F. Koob and Athina Markou

contents

http://neurosci.nature.com

volume 5 no 7 july 2002

Sleep is known to be importantfor learning, but how it exerts thiseffect remains mysterious.Mednick and colleagues nowreport that performance on avisual shape discrimination taskdeteriorated across multiple train-ing sessions within a given day,but this deterioration could beprevented or reversed by a briefmidday nap. This benefit was spe-cific to the trained task and to theregion of the visual field that wasengaged by it. Photo courtesy ofPhoto Researchers. See pages 618and 677.

nature neuroscience • volume 5 no 7 • july 2002 i

Suppressing dendriticbranching with ARNO.

Page 623.

Nature Neuroscience (ISSN 1097-6256) is published monthly by Nature America Inc., 345 Park Avenue South, New York, NY 10010-1707. Editorial Office: 345 Park Avenue South, New York, NY10010-1707. Tel: (212) 726 9200, Fax: (212) 696 9635. Annual subscription rates: USA/Canada: US$199/US$213 (personal), US$99/US$106 (student), Canada add 7% for GST: 140911595RT001;U.K./Europe: £185 (personal), £105 (student); Rest of world (excluding China, Japan, Korea): £235 (personal), £110 (student); Japan: Contact Nature Japan K.K., MG Ichigaya Building 5F, 19-1 Haraikatamachi, Shinjuku-ku, Tokyo 162-0841. Tel: 81 (03) 3267 8751, Fax: 81 (03) 3267 8746. Authorization to photocopy for internal or personal use, or internal or personal use of specif-ic clients, is granted by Nature Neuroscience to libraries and others registered with the Copyright Clearance Center (CCC) Transactional Routing Service, provided the base fee of $9.00 an article (or$1.00 a page) is paid direct to CCC, 27 Congress Street, Salem, MA 01970, USA. Back issues: US$45, Canada add 7% for GST; Periodicals postage paid at New York, NY and additional mailingoffices. CPC PUB AGREEMENT #40032744. POSTMASTER: Send address changes to Nature Neuroscience Subscription Department, P.O. Box 5054, Brentwood, TN 37024-5054. Printed by Publish-ers Press, Inc., Lebanon Junction, KY, USA. Copyright © 2002 Nature America Inc. Printed in USA.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

contents

nature neuroscience • volume 5 no 7 • july 2002 ii

Parallel processingof natural scenes.

Page 629.

NMDA receptors lose theirinhibitions to alcohol.

Pages 614 and 641.

Imaging multivesicular releaseat single synapses.

Page 657.

Dyskinesias following neural transplantation in Parkinson’s disease . . . . . . . . . . 627Peter Hagell, Paola Piccini, Anders Björklund, Patrik Brundin, Stig Rehncrona, Håkan Widner, Lesley Crabb, Nicola Pavese, Wolfgang H. Oertel, Niall Quinn, David J. Brooks and Olle Lindvall

Parallel processing in high-level categorization of natural images . . . . . . . . . . . 629Guillaume A. Rousselet, Michèle Fabre-Thorpe and Simon J. Thorpe

Global effects of feature-based attention in human visual cortex . . . . . . . . . . . . 631Melissa Saenz, Giedrius T. Buracas and Geoffrey M. Boynton

reviewGenetically engineered mouse models of neurodegenerative diseases . . . . . . . 633Philip C. Wong, Huaibin Cai, David R. Borchelt and Donald L. Price

articlesDARPP-32 and regulation of the ethanol sensitivity of NMDA receptors in the nucleus accumbens . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641R. E. Maldve, T. A. Zhang, K. Ferrani-Kile, S. S. Schreiber, M. J. Lippmann, G. L. Snyder, A. A. Fienberg, S. W. Leslie, R. A. Gonzales and R. A. MorrisettSEE NEWS AND VIEWS, PAGE 614

Synaptotagmin function in dense core vesicle exocytosis studied in cracked PC12 cells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649Ok-Ho Shin, Josep Rizo and Thomas C. Südhof

Facilitation at single synapses probed with optical quantal analysis . . . . . . . . . . 657Thomas G. Oertner, Bernardo L. Sabatini, Esther A. Nimchinsky and Karel Svoboda

Three-dimensional orientation tuning in macaque area V4 . . . . . . . . . . . . . . . . 665David A. Hinkle and Charles E. Connor

Filtering of neural signals by focused attention in the monkey prefrontal cortex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 671Stefan Everling, Chris J. Tinsley, David Gaffan and John Duncan

The restorative effect of naps on perceptual deterioration . . . . . . . . . . . . . . . . . 677Sara C. Mednick, Ken Nakayama, Jose L. Cantero, Mercedes Atienza, Alicia A. Levin, Neha Pathak and Robert StickgoldSEE NEWS AND VIEWS, PAGE 618

Visual features of intermediate complexity and their use in classification . . . . . . 682Shimon Ullman, Michel Vidal-Naquet and Erez Sali

Morphology of Heschl’s gyrus reflects enhanced activation in the auditory cortex of musicians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 688Peter Schneider, Michael Scherg, H. Günter Dosch, Hans J. Specht, Alexander Gutschalk and André Rupp

Degree of language lateralization determines susceptibility to unilateral brain lesions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 695S. Knecht, A. Flöel, B. Dräger, C. Breitenstein, J. Sommer, H. Henningsen, E. B. Ringelstein and A. Pascual-LeoneSEE NEWS AND VIEWS, PAGE 613

Segmenting nonsense: an event-related potential index of perceived onsets in continuous speech . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 700Lisa D. Sanders, Elissa L. Newport and Helen J. Neville

errata/corrigenda . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704

classified advertising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . see back pages

Auditory cortex morphologyin professional musicians.

Page 688.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

nature neuroscience • volume 5 no 7 • july 2002 611

It appears certain that Germany’s constitution will soon requirethe state to protect the welfare of animals. The Bundestag, the lowerhouse of parliament, voted 543 to 19 in favor of the amendmentin May, and the upper house (Bundesrat) is expected to approvethe measure very soon. With over 80% of German citizens favoringthe amendment, the animal rights movement has achieved a majorvictory. The failure of scientists to make a convincing public case fortheir work is particularly worrisome because another critical leg-islative battle lies ahead. If this fight is also lost, the likely result willbe great damage to German biomedical research.

The amendment will add “and animals” to Article 20a of theGerman Basic Law, so that it will read, “The state takes respon-sibility for protecting the natural foundations of life and animalsin the interest of future generations.” Previously the courts haveinterpreted “life” to mean human life, but the new wording islikely to suggest that animals have a legal right to be protectedfrom avoidable pain. Freedom of academic research was alreadyprotected in the German Basic Law. Thus, in the past, when animal welfare conflicted with academic freedom, the courtsweighed the rights of researchers more heavily. After the amend-ment, however, animal protection and academic freedom willhave equal legal status, and it is not clear how this new balancewill affect future court decisions. For now, animal welfare lawswill remain the same, but local authorities and courts may inter-pret them differently. In addition, the legislature may come underincreasing pressure to pass additional laws on a variety of ani-mal-protection issues. “Even the politicians do not know andcannot predict what the consequences of this change to the con-stitution will be,” says Kuno Kirschfeld, who heads the commit-tee on animal legislation at the Max Planck Society.

German animal activists have worked for the last decade toachieve constitutional protection for animal rights, an aim sup-ported by the ruling Social Democratic party and the Green party.Two years ago, a similar attempt was narrowly defeated, fallingjust short of the two-thirds necessary to amend the constitution. Inthat vote, the amendment was blocked by the Christian Demo-crats, who have now reversed their position following a contro-versial decision by the high court (Bundesverfassungsgericht) inJanuary. The court ruled that Muslims may practice halal slaugh-tering, in which animals are not stunned before being killed. Thedecision outraged the public and increased the pressure on par-liament to grant constitutional protection to animals. With anelection approaching, the Christian Democrats announced inMarch that they would support the constitutional amendment.

Allowing the opposition to control the terms of the debate hasbeen a critical mistake for the research community. Scientists havenot effectively communicated to the public that animal experimentsare necessary for medical progress. “People are not informed about

animal research and the problems the constitutional change maycause,” says Ivar Aune, a spokesman for the Society for Health andResearch, Germany’s main lobby group in defense of biomedicalresearch. Aune notes that 85% of people surveyed answer yes to thequestion, “Are you against animal research?”, yet only 15% say yeswhen asked, “If animal research is the only way to find a cure forcancer or AIDS, are you still against it?” This discrepancy suggeststhat opposition to animal research, though broad, is not very solidand could be changed by further information. Many scientists feelthat Germany’s main funding agencies, the Max Planck Instituteand the Deutsche Forschungsgemeinschaft (DFG), have not taken astrong enough public stand in defense of animal experiments.

German animal protection laws are already very strict. Everythree years, researchers must reapply to their local government for alicense to do experiments on animals. According to a previous courtdecision, the local authorities are prohibited from applying theirown ethical standards to these applications; instead they must deter-mine whether the proposed research is permissible under Germanlaw. On the rare occasions when the local authority refuses an appli-cation, the researcher can go to court and ask a judge to approve it.This process is likely to provide the legal test cases that will ultimatelydetermine the practical effects of the constitutional amendment.

Under current law, researchers can choose whether or not togo to court to defend their work, but animal activists are fightingto overcome this major disadvantage (from their perspective) inthe legal system. Activists are lobbying vigorously for a law allow-ing animal protection groups to sue researchers on behalf of theiranimal subjects. There is a precedent for such permission in Ger-man law; certain interest groups can file lawsuits against viola-tions of environmental protection laws that do not harm thempersonally. A similar law pertaining to experimental animalswould give activists broad power to harass scientists. Regardlessof the final outcome, such lawsuits could be very damaging toresearchers’ careers, as cases can take years to reach the consti-tutional court and the law provides for laboratories to be shutdown until such court proceedings are resolved.

To keep this weapon out of the hands of animal activists,researchers will need to succeed in convincing politicians and thepublic that animal experiments are valuable to society. This con-stitutional amendment should serve as a wake-up call to scientists,not only in Germany, but in other countries where animal rightsactivists are exploring legal challenges to animal research1. If sci-entists do not argue more effectively in favor of animal research,they will lose other legal and legislative battles. As Stefan Treue of theGerman Primate Center in Göttingen concludes, “My view on whatwent wrong is that we were too quiet for too long.”

1. Legal challenges to animal experimentation. Nat. Neurosci. 3, 523 (2000).

editorial

Constitutional protection for animals

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


Since the pioneering work of Paul Broca(1861), damage to left cerebral cortexhas been known to disturb languageprocessing. Broca described a patientwith impaired speech production after alesion to the left inferior frontal cortex1.Subsequently, Wernicke (1874) doc-umented a patient with poor speechcomprehension following damage tothe left posterior superior temporallobe2. These early studies providedthe foundation for the now-standardview that language functions are typ-ically lateralized in the left hemi-sphere. The role in languagefunctions of the right hemisphere,however, has remained more difficultto verify. Now, using transcranialmagnetic stimulation (TMS) to inter-fere with neuronal function, Knechtand colleagues3 have been able tothrow new light on the causal role ofright hemisphere involvement in lan-guage processing.

Traditional evidence on theinvolvement of the right hemispherein language comes from two sources.One piece of evidence concernspatients who present with some lan-guage disturbance after a lesion of theright hemisphere. This relativelyuncommon pattern is often attributedto individuals having crossed lateral-ization of function, associated with dif-ferences in handedness4. It can beargued, at least in these cases, that theright hemisphere is necessary for nor-mal language function. (Not all lefthanders have right-hemisphere lan-guage dominance, but people whoseright hemisphere is language dominant

whether a patient was truly right hemi-sphere dominant for language and sosuffered language problems after aright hemisphere lesion, or whether thelesion somehow prevents normal lan-guage functions in the left hemisphere

from operating (perhaps by someinhibitory function).

Similarly, we do not know if, inother cases, any recovery of languageis due to the right hemisphere takingover some functions or due to directrecovery within the left hemisphere.This circularity problem is not solvedby recent advances in functionalbrain imaging. Functional brainimaging can provide informationconcerning which brain areas areactive in language tasks after a patienthas suffered a brain lesion, and stud-ies have shown that right-hemisphereregions are more active in recoveredaphasics than in normal controls,

considered as a group5. However, thereare also considerable individual dif-ferences in lateralization of languagefunction in the normal population,and when compared with the spread

of results in normals, then recoveredaphasics do not differ6. From suchresults, it is difficult to judge whether(i) the right hemisphere activations arenecessary for language functions inthese individuals (patients and con-trols), (ii) the patients are showingrecovery because of new recruitment ofright-hemisphere areas, or (iii) thepatients show recovery because they hadsome language functions in their righthemisphere to begin with.

In the past ten years, investigatorshave developed TMS as a new means ofinvestigating brain function throughdirect intervention. In this technique, abrief magnetic pulse (or a train of puls-es) can be applied to the scalp. The puls-es induce a local electric field in the

Magnetic stimulation reveals thedistribution of language in a normalpopulationGlyn W. Humphreys and Peter Praamstra

Language is classically considered to be a function of the left side of the brain. Now aninterference technique, transcranial magnetic stimulation, in healthy subjects shows that theright-side language activity detected in some people is indeed functionally relevant.

do tend to be left handed.) The secondpiece of evidence comes from recoveryof function after a brain lesion. Here itcan be argued that the recovery of lan-guage function, following damage tothe left hemisphere, is either due to the

right hemisphere taking over some ofthe impaired functions or due to dor-mant right-hemisphere language capa-bilities being released from inhibitionfrom the dominant left hemisphere4.However, there is a difficulty with thesetraditional arguments, which essen-tially concern individual differences inlanguage lateralization: the argumentsare typically circular. Measures of lan-guage lateralization on the individualsconcerned, before their lesion, are usu-ally not available. Hence it cannot beverified whether the individuals hadunusual lateralization pre-morbidly—with right side dominance or with lan-guage equally distributed across leftand right hemispheres. Without thisknowledge, we simply do not know

news and views

The authors are at the Behavioural BrainSciences Centre, School of Psychology,University of Birmingham, Edgbaston,Birmingham B15 2TT, UK.e-mail: [email protected]

Fig. 1. TMS induces a local electric field in the surfaceof the brain beneath the stimulator, and thereby dis-rupts local brain activity released from inhibition by thedominant. Image courtesy of Alvaro Pascual-Leone.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

614 nature neuroscience • volume 5 no 7 • july 2002

surface of the brain underneath thestimulator, and consequent changes inbrain activity (Fig. 1). This enables theinvestigator to intervene in the neuralprocesses underlying cognitive function,because a stimulated brain area mayincrease or decrease in activity accord-ing to the stimulation parameters. Con-sequent behavioral changes can then beassessed. There are various means ofgenerating neuronal change throughTMS, including event-related, single-pulse or repetitive TMS. Knecht and col-leagues3 used a different ‘blocked’procedure, in which participantsreceived a sustained block of TMS puls-es for 600 seconds before doing languagetasks. This procedure leads to more pro-longed neuronal and behavioral changes,as measured in the subsequent task(here, matching the appropriate word toa picture). The pulses were applied inone of three sites: either over the approx-imate location of Wernicke’s area or itshomologue in the right hemisphere, orover a midline region of occipital cortex(presumably without specialized lan-guage functions). Participants were sep-arated according to whether they wereleft-hemisphere language dominant orright-hemisphere language dominant orhad bilateral language representation,according to differences in brain activity

is relatively complex and may be accom-plished using several possible represen-tations (for example, the names of theobjects or their semantics). Further stud-ies, in which experimenters vary the rela-tionships between the picture and thewords, will be informative here. It is alsounclear why responses were speeded bystimulation of the non-dominant hemi-sphere (for example, after TMS to theright hemisphere in left-dominant indi-viduals). Knecht and colleagues3 arguethat this effect was due to a narrowing ofa lexical search when identifying the pic-ture, but this needs to be verified by vary-ing the size of the lexical search directly.Nevertheless, this important study pro-vides perhaps the clearest evidence todate that language functions reside in theright hemisphere of some individuals.

1. Broca, P. Bull. Soc. Anat. 6, 330–350 (1861).

2. Wernicke, C. Der Aphasiche Symptomen-komplex (Cohen & Weigert, Breslau, Poland,1874).

3. Knecht, S. et al. Nat. Neurosci. 5, 695–699(2002).

4. Code, C. Language, Aphasia and the RightHemisphere (Wiley, London, 1987).

5. Weiller, C. et al. Ann. Neurol. 37, 723–732(1995).

6. Warburton, E., Price, C. J., Swinburn, K. &Wise, R. J. S. J. Neurol. Neurosurg. Psychiatry66, 155–161 (1999).

in a word generation task (measuredthrough functional transcranial Dopplersonography).

Striking differences emerged betweengroups in the TMS effects. Stimulationof the left hemisphere above Wernicke’sarea tended to slow picture–wordmatching by left-dominant participants,while, in contrast, speeding responses ofright-dominant participants. Stimula-tion of the homologue area in the righthemisphere slowed matching by right-dominant patients while speedingresponses by left-dominant participants.There were few effects from occipitalstimulation. This study, then, providesevidence that altering activation in theright hemisphere (via TMS) affects alanguage-processing task in individualswith right-hemisphere lateralization—importantly, the lateralization was deter-mined before testing. This clearlysuggests that, for these individuals, theright hemisphere is necessary for lan-guage. For instance, it cannot be thatright-hemisphere stimulation led toinhibition of left-hemisphere language,as exactly the opposite results occurredwith individuals with left-dominant lan-guage functions.

Several questions remain, includingwhich exact language functions wereaffected by TMS. Picture–word matching

news and views

NMDA receptors lose theirinhibitionsDavid M. Lovinger

A new study suggests that interactions between dopamineand glutamate neurotransmitter pathways are important inregulating the inhibitory effects of alcohol on brain function.

The transition from controlled, ‘social’drinking to uncontrolled alcohol abuseis a key step in the development of alco-holism. During this process, the braindevelops changes in synaptic transmis-sion that produce tolerance to the pres-ence of alcohol and also promote adesire or craving for alcohol. The neur-

al changes that promote alcohol abuseare thought to involve synaptic plastici-ty within the brain’s intrinsic reward cir-cuitry, which includes the nucleusaccumbens (NAc) and the ventraltegmental area (VTA). Two neurotrans-mitters within this reward system,dopamine and glutamate, are importantin intoxication and alcohol addiction,but how interactions between these sig-nals contribute to alcohol-related neur-al changes is unclear. One paradoxicalfinding is that alcohol inhibits theNMDA type of glutamate receptor,which contributes to intoxication1. How-

ever, NMDA receptors also promoteneural plasticity, such as lasting increas-es in synaptic communication in theNAc2,3. Inhibition of these receptors byalcohol therefore should prevent themfrom participating in the synaptic plas-ticity believed to be critical for the devel-opment of uncontrolled drinking.

In this issue, Maldve and coworkersprovide a possible solution to this prob-lem by demonstrating that activation ofdopamine receptors relieves alcoholinhibition of NMDA receptors in theNAc4. This molecular ‘disinhibition’could promote tolerance to alcohol, butmore importantly in the context of alco-hol abuse and alcoholism, would freeNMDA receptors to participate in thegeneration of synaptic changes neces-sary for alcohol addiction. The newfindings increase our understanding ofkey molecular events related to thedevelopment of alcoholism, and couldlead to new pharmacotherapies for usein treatment of the disorder.

Maldve et al.4 found that ethanolinhibition of NMDA receptors was

The author is Chief of the Laboratory forIntegrative Neuroscience at the NationalInstitute on Alcohol Abuse and Alcoholism, Rm. 158H Park 5 Bldg, 12420 Parklawn Drive,Rockville, Maryland 20852, USAe-mail: [email protected]

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

greatly reduced following activation of D1 dopamine receptors in cultured neurons and brain slices. Thisdopamine-stimulated disinhibitiondepended on a series of intracellular sig-naling events involving activation ofadenylate cyclase, subsequent activationof the cyclic AMP-dependent proteinkinase (PKA), and phosphorylation ofthe protein phosphatase inhibitorDARPP-32 (Fig. 1). DARPP-32 seems tobe a key molecular switch in this path-way, as this disin-hibition was absent inmice lacking DARPP-32. Activation ofD1 receptors also increased phosphory-lation of the NR1 subunit of the NMDAreceptor, presumably because phospho-rylated DARPP-32 inhibits the proteinphosphatase PP-1 that normally removesthe phosphate group at serine 897 with-in the NR1 protein.

Several steps in this process remainto be explored in more detail. For exam-ple, the mechanism suggested in thispaper likely works in conjunction witha D1 receptor–mediated activation ofPKA that phosphorylates NR1 at residue897 (ref. 5; Fig. 1). The relative impor-tance of kinase activation and phos-phatase inhibition in the D1 receptorstimulation of NR1 phosphorylationneeds to be worked out in the contextof alcohol’s inhibition of the NMDAreceptor. Furthermore, it is not yet clearif phosphorylation of the NMDA recep-tor directly produces decreased alcoholsensitivity or if there is another, as yetunidentified, phosphoprotein with arole in this scenario. If phosphorylationof NR1 is the key to decreased inhibi-tion, then the intramolecular changesthat render the NMDA receptor less sen-sitive to alcohol will require elucidation.These unanswered questions shouldprovide fertile ground to extend theexciting findings presented in this paper,which may lead to development of phar-macological or genetic manipulations atseveral steps in the intracellular molec-ular cascade that could alter the patternof alcohol effects. Interestingly, in vivoapplication of a D1 agonist also pro-

receptors in dependence-related in-creases in alcohol drinking should helpto evaluate this model.

The observations made by Maldve et al.4 also have the potential to explainsome aspects of rapid tolerance to alco-hol, in which individuals become lesssensitive to alcohol’s intoxicating effecteven within a single drinking session.Alcohol was found to increase phos-phorylation of DARPP-32 in the pre-sent report, although not to the extentseen following activation of D1 recep-tors. This increased phosphorylation,in conjunction with alcohol enhance-ment of NAc dopaminergic transmis-sion, should decrease ethanol inhibi-tion of NMDA receptors and removeimpairment of neuronal activity evenwithin a single intoxication experience.In support of this, another study implicated NMDA receptor phospho-rylation in the development of acutealcohol tolerance12.

The sensitivity of NMDA receptorsto alcohol varies under different condi-tions, for example, as a function ofdevelopment and brain region13,14. Pro-tein phosphorylation likely provides oneexplana-tion for these differences inNMDA receptor alcohol sensitivity. Therich interactions between glutamate andother neurotransmitter systems andintracellular signaling pathways could

duces increased phospho-rylation of DARPP-32 andNR1, and alcohol interactswith these biochemicalpathways6, indicating thatthe D1R/DARPP-32/NR1interaction is involved inalcohol’s effects in theintact animal.

The interactions amongalcohol,dopamine and glu-tamate may be an impor-tant step in the develop-ment of uncontrolled or‘binge’ drinking, with D1receptors and DARPP-32likely candidates as media-tors of this transition. Asthe authors point out, mice lacking eitherthe D1 receptor or DARPP-32 do notshow the increases in alcohol intake nor-mally observed in humans and other ani-mals after repeated exposure to increasingconcentrations of alcohol7,8, although itmust be pointed out that drinking in themouse models only leads to blood alco-hol levels that would be considered mod-erate in humans, and never approachesthe high levels observed in many alco-holics. Thus, it is not yet clear how thesefindings relate to the development ofhuman alcoholism.

The location of the dopamine–glutamate interaction proposed in thecurrent study within the NAc is impor-tant because this brain region is impli-cated in reward and alcohol addiction.As the authors speculate, this suggestsan elegant hypothesis wherebyD1/DARPP-32 disinhibition of NMDAreceptors could free these receptors topromote the neural changes that bringabout alcohol abuse. Findings thatalcohol exposure increases the activityof dopaminergic neurons in the VTAthat project to the NAc9,10 and thatNAc dopamine levels increase duringdrinking11 provide additional details ofhow the dopaminergic system maybecome ‘revved up’ as a result of alco-hol consumption. Continued explo-ration of the role of NAc NMDA

Fig. 1. Intracellular molecular events linking dopamine receptor activation toincreased NMDA receptor phosphorylation and decreased inhibition by alcohol.Without D1 receptor activation, alcohol inhibits NMDA receptor function.Activation of D1 receptors stimulates adenylate cyclase (AC), which catalyzes cAMPformation, increasing PKA activity, and subsequent phosphorylation of DARPP-32 andthe NMDA receptor subunit NR1. Alcohol also stimulates DARPP-32 phosphoryla-tion. Phosphorylated DARPP-32 inhibits protein phosphatase PP2A, and this reducesNMDA receptor dephosphorylation. The net increase in NMDA receptor phospho-rylation relieves the inhibition of the receptor by alcohol.


news and views

?

D1-type receptor

Dopamine

AC

cAMP

PKA

DARPP-32-PO3–

DARPP-32

PP-1

NR1-PO3– NR1

Alcohol

AlcoholIvelisse Robles

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


tailor the cell’s response to continuingalcohol exposure, yielding different pat-terns of initial effect and tolerancedepending on the brain region affectedand the behavior studied. In the presentcase, the dopamine–glutamate interac-tions observed in the NAc are likely tobe involved in the transition to excessivedrinking, whereas differences in alcoholsensitivity among NMDA receptors dueto fyn kinase activation in the hip-pocampus12 might be involved in cog-nitive function. As the authors of thecurrent study point out, the proteinkinases and NMDA receptor subunitproteins implicated in tolerance differacross brain regions, and the dopamine-mediated disinhibition they havedescribed is unlikely to occur indopamine-poor brain regions that con-tain alcohol-sensitive NMDA receptors,such as the cerebellum.

Examination of the relationshipbetween receptor subtypes and post-translational modifications such as

H. & Lopes da Silva, F. H. Eur. J. Neurosci. 5,107–117 (1993).

3. Kombian, S. B. & Malenka, R. C. Nature 368,242–246 (1994).

4. Maldve, R. E. et al. Nat. Neurosci. 5, 641–648(2002).

5. Snyder, G. L. et al. J. Neurosci. 18,10297–10303 (1998).

6. Edwards, S. et al. Alcohol. Clin. Exp. Res. 26,173–180 (2002).

7. El-Ghundi, M. et al. Eur. J. Pharmacol. 353,149–158 (1998).

8. Risinger, F. O. et al. J. Neurosci. 21, 340–348(2001).

9. Brodie, M. S., Shefner, S. A. & Dunwiddie, T. V. Brain Res. 508, 65–69 (1990).

10. Gessa, G. L. et al. Brain Res. 348, 201–203(1985).

11. Gonzales, R. A. & Weiss, F. J. Neurosci. 18,10633–10671 (1998).

12. Miyakawa, T. et al. Science 278, 698–701(1997).

13. Lovinger, D. M. J. Pharmacol. Exp. Ther. 274,164–172 (1995).

14. Yang, X. et al. J. Pharmacol. Exp. Ther. 278,114–124 (1996).

phosphorylation with respect to alcoholsensitivity will no doubt provide impor-tant clues to alcohol effects on theNMDA receptor. It may be premature toextrapolate from studies of dissociatedcells and brain slices to the behavior ofanimals, especially with respect to themultifaceted pharmacological effects ofalcohol. However, the use of animalmodels such as genetically altered micewill help us to determine if altering pre-sumed molecular targets of alcoholaction in the brain alters alcohol-relat-ed behaviors. Viewed in this context, itis likely that a better understanding ofthe molecular effects of alcohol, in turn,should help us to explain several aspectsof alcohol intoxication, tolerance anddependence. Ultimately, such studiescould lead to better pharmacotherapiesfor alcohol abuse and alcoholism.

1. Diamond, I. & Gordon, A. S. Physiol. Rev.77, 1–20 (1997).

2. Pennartz, C.M., Ameerun, R., Groenewegen,

news and views

Short circuiting the circadianclockAnthony N. van den Pol and Karl Obrietan

A recent report in Cell shows that the transcription andtranslation cycles that drive the molecular circadian clockmay also be regulated by electrical activity.

Circadian rhythms pervade all aspectsof life. Although the cellular clocks thatdrive these near 24-hour cycles areentrained to environmental cues, theycontinue to keep time and to drivebehavioral and physiological rhythmseven in the absence of environmentalsignals. Now, a recent report in Cell addsa new dimension to our understandingof the mechanisms underlying time-keeping in these cellular clocks.Nitabach et al.1 used a clever potassiumchannel gene-targeting approach toselectively silence the electrical activityof critical cells in the Drosophila central

clock, the ventral subset of lateral pace-maker neurons (LNv). As might beexpected, this electrical silencingblocked the ability of the pacemakerneurons to send the circadian signal toother brain regions, thereby stoppingcircadian rhythms of behavior. But moreinterestingly, silencing the electricalactivity of these neurons also eliminat-ed the cycles of the molecular clock. Thisis an unexpected finding, as currentmodels of molecular clock function donot include ion channel feedback.

Drosophila have enjoyed great popu-larity among molecular chronobiologistssince the identification of a single-genemutation that is capable of eliminatingcircadian rhythms2. The molecular basisfor circadian timing—postulated to arisefrom a limited number of gene productsthat interact to form a core clock feed-back loop—has been worked out ingreat detail3–5 (Fig. 1). In Drosophila,this loop involves the proteins Period,

Timeless, Clock and Cycle. The pendu-lum of this clock swings into motion bythe formation of a transcription factor, adimer of Clock and Cycle. This basichelix-loop-helix transcription factorfacilitates the expression of period andtimeless genes. Period and Timelessdimerize in the cytoplasm, allowingthem to move into the nucleus, andinhibit Clock/Cycle-dependent tran-scription. Degradation of the Period/Timeless complex relieves transcrip-tional inhibition, thus permitting a newround of period and timeless expression.Doubletime, a clock kinase, affects clocktiming by reducing the stability of Peri-od. The cycle of this complete loop takesabout 24 hours, and is postulated tofunction as the timing mechanism of thecircadian clock. The output of the mol-ecular clock can be viewed as transcrip-tional modulation of other downstreamgenes by Clock and Cycle. Although alarge number of genes, including somethat code for ion channels, show circa-dian variations in expression5, whichones are directly modulated by coreclock transcriptional regulators remainsto be determined.

Despite our extensive knowledgeabout the molecular clock, little isknown about how electrical activityinteracts with the system. This is animportant question, as the complexfeedback loops among the core clockgenes are fascinating, but relevant only

Anthony N. van den Pol is in the Department ofNeurosurgery, Yale University Medical School,333 Cedar Street, New Haven, Connecticut06520, USA. Karl Obrietan is in theDepartment of Neuroscience, Ohio StateUniversity, 333 West 10th Avenue, Columbus,Ohio 43210, USA.e-mail: [email protected]

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

because they alter the physiology of thehost cell. For example, the circadianrhythm of neuronal activity leads to arelease of chemical signals, such as pig-ment dispersing factor, at circadianintervals. The expression of the clock isseen by the host organism only as theoutput signal sent from the clock cells toother parts of the brain, based in mam-mals on action potential–dependentchronotransmitter release6.

Nitabach et al.1 addressed the cou-pling of the molecular clock with elec-trical activity by manipulating ionchannels in Drosophila clock cells. Theauthors generated transgenic flies thatselectively expressed one of two potassi-um channels in the LNv, ORK or Kir2.1.These potassium channel variants werechosen because they have a high openprobability at resting membrane poten-tial, and they were targeted to the LNvusing a promoter for a protein selective-ly expressed in that region, pigment dis-persing factor7. Due to the expectedpotassium efflux from open channelspredicted by the Nernst equation, theclock cells were putatively made perma-nently silent, unable to receive or sendsynaptic signals. Under constant condi-tions, these flies showed no behavioralrhythms. Surprisingly, the flies also dis-played no circadian variation in levels ofPeriod and Timeless, and showed areduction of Timeless transport into thenucleus. This suggests that electricalsilence stopped the molecular clock.Control experiments showed that globalexpression of these two potassium chan-nels throughout the nervous system was


cryptochrome protein8 to light. Thefinding that light can temporarily rescuethe diurnal cycling of clock-related pro-teins also suggests that open potassiumchannels do not simply damage the LNv.

How could events regulated by plas-ma membrane ion channels block themolecular clock? Does the block reveala new integral component of clock mod-ulation, or does it instead indicateunusual secondary mechanisms thatinfluence clock function under anabnormal hyperpolarized state? Previ-ous studies of mammalian or Aplysiaclock neurons showed that after actionpotentials were blocked with tetrodotox-in (TTX), the clock kept time, but didnot send the circadian message to otherbrain regions; when TTX applicationwas stopped, the clock timing was at pre-cisely the point predicted for a clock thathad continued to keep time in theabsence of action potentials6,9. It wouldbe useful to know whether TTX wouldhave the same effect in Drosophila.Nonetheless, the mechanism underlyingthe substantive difference between elim-inating neuronal activity by blockingvoltage-activated sodium channels (withTTX, clock keeps time) and by openingpotassium channels (targeted expressionof ORK or Kir2.1, clock stops) remainsto be determined. Neither treatment istoxic to the clock cells. Each would elim-inate the response to synaptic input andspike output.

Although little information is avail-able on the interface between cellularactivity and molecular clock mecha-nisms, based on what we know, we canspeculate on several possibilities of howopen potassium channels may block theclock. Long-lasting, activity-dependentchanges in second messengers, linked tocalcium or phosphorylation, could inter-fere with clock dynamics at several sites;for instance, reducing Period phospho-rylation would prevent its degradationand oscillation, or reducing active trans-port of the Period–Timeless dimer intothe nucleus may have the same effect(Fig. 1). Doubletime, which enhancesPeriod degradation in a phosphoryla-tion-dependent manner10, may also beinvolved. The mammalian ortholog ofDoubletime is regulated by a calciumbinding phosphatase11. Thus, ifDrosophila Doubletime were regulatedin a similar manner, membrane hyper-polarization might lead to reduced cyto-plasmic calcium, which in turn couldaffect Doubletime-mediated regulationof Period stability. Another possibility is

lethal, and that non-conducting molec-ular variants of the potassium channelgenes did not influence the clock.

To really understand what aspect ofaltered electrical activity stopped themolecular clock, it would be useful toknow to what degree ORK or Kir2.1hyperpolarized the membrane potential,decreased input resistance and shuntedelectrical activity in the fly LNv.Although Kir2.1 was more effective thanORK in blocking behavioral rhythms, itis not clear whether this was due to func-tional differences between the two potas-sium channels, or to differences inexpression levels. If the detailed physio-logical effects of potassium channelexpression in the LNv were known, thendifferential levels of attenuation ofbehavioral rhythms reported could becorrelated with the specific physiologi-cal behavior of the potassium channel-expressing clock cells. AlthoughDrosophila have been the animal ofchoice for molecular clock research, verylittle cellular physiology has been donein this preparation to address criticalquestions of electrical activity and cou-pling to the molecular clock. This short-coming may be due to difficulties inrecording from these small cells, andmay also reflect the molecular/geneticbias of the labs working in this area.

Although targeted expression ofpotassium channels blocked the endoge-nous rhythm under constant dark con-ditions, Period and Timeless still cycledif flies were maintained under normaldiurnal light cycles. This might reflect adirect response of LNv photoreceptive

news and views

K+

K+

K+

K+ efflux

Timeless

PO4

Period

Clock Cycle

PeriodX

-

++

CREB

Clock Cycle

Hyperpolarization, increased conductance, changes in 2nd messengers.

Timeless

DoubletimeCalcium/phosphatase

Kinase

NAD

Nuclear translocation

Period

Fig. 1. Possible ways in which electrical inactivity might interact with the molecular circadian clock.Amy Center

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

Ever felt burned out after several hoursof intense intellectual work? Ever tried ashort nap after lunch to improve yourproductivity? In this issue, Mednick etal. suggest that you should. Their datashow that a nap restores performanceafter repetitive training on the same dayon a perceptual learning task1.

What happens in your brain duringsleep? According to one hypothesis,sleep is involved in restoring brain func-tion after the wear and tear of the day’sactivities. Such use-dependent process-es are intimately linked to the neuronalworkload of the preceding waking peri-od (Fig. 1b, top). Another hypothesis is


that an inducible modulator of tran-scription, perhaps the CREB/CRE tran-scriptional pathway, could couplemembrane ion channel activity to thecore clock feedback loop. CRE-mediat-ed gene expression is regulated in a cir-cadian-dependent manner in Drosophila,and disruption of CREB signaling dis-rupts Period oscillations12. Because theCREB/CRE transcriptional pathway isresponsive to neuronal activity, hyper-polarizing the membrane potentialwould likely attenuate CRE-dependenttranscription, which in turn couldadversely affect the cycling of clockgenes. Circadian variations in cellularmetabolic state relating to nicotimamideadenine dinucleotide cofactors may alsoinfluence clock timing properties13.Given that metabolic state can be influ-enced by neuronal electrical activity,blocking electrical activity in pacemakerneurons may blunt a potentially impor-tant mechanism by which rhythm gen-

1. Nitabach, M. N., Blau, J. & Holmes, T. C.Cell 109, 485–495 (2002).

2. Konopka, R. J. & Benzer, S. Proc. Natl. Acad.Sci. USA 68, 2112–2116 (1971).

3. Williams, J. A. & Sehgal, A. Annu. Rev.Physiol. 63, 729–755 (2001).

4. Reppert, S. M. & Weaver, D. R. J. Biol.Rhythms 15, 357–364 (2000).

5. McDonald, M. J. & Rosbash, M. Cell 107,567–578 (2001).

6. Schwartz, W. J., Gross, R. A. & Morton, M.T. Proc. Natl. Acad. Sci. USA 84, 1694–1698(1987).

7. Renn, S. C., Park, J. H., Roshbash, M., Hall, J. C. & Taghert, P. H. Cell 99, 791–802(1999).

8. Emery, P. et al. Cell 95, 669–679 (1998).

9. Eskin, A. J. Neurobiol. 8, 273–299 (1977).

10. Price, J. L. et al. Cell 94, 83–95 (1998).

11. Cegielska, A., Gietzen, K. F., Rivers, A. &Virshup, D. M. J. Biol. Chem. 273,1357–1364 (1998).

12. Belvin, M. P., Zhou, H. & Yin, J. C. Neuron22, 777–787 (1999).

13. Rutter, J., Reick, M., Wu, L. C. & McKnight, S. L. Science 293, 510–514 (2001).

eration is modulated. Any of the path-ways influenced here by open potassiumchannels may normally function as anintegral part of the molecular clock orto entrain it to environmental cycles.

Is work on Drosophila circadianclocks relevant to other organisms? Themammalian circadian timing system,although somewhat more complicated,and with a greater recognized cast ofclock-related genes, relies on transcrip-tional/translational feedback loops sim-ilar to those found in Drosophila.Orthologs of fly clock genes are foundin the hypothalamic suprachiasmaticnucleus, the central circadian pacemak-er in the mammalian brain. Many of thegenes have functional roles generallyparallel to those in Drosophila, and mayalso show similar responses to hyperpo-larizing clock neurons. As with anyprovocative paper, Nitabach et al.1

answers some interesting questions, andraises even more.

news and views

Be caught napping: you’redoing more than restingyour eyesPierre Maquet, Philippe Peigneux, Steven Laureys and Carlyle Smith

Sleep is suggested to repair fatigue or to enhance memoryconsolidation. A new paper shows that the beneficial effect ofsleep is specific to the task and the brain regions engaged by it.

that sleep is at least favorable, if notindispensable, for memory consolida-tion in the brain. These experience-dependent processes entail expansionof the behavioral repertoire after expo-sure to a new environment. In contrastto restoration, consolidation might goon long after the last training sessionhas ended (Fig. 1b, bottom)2.

The texture discrimination taskdescribed by Karni et al.3 has beenexceptionally useful for the behavioralstudy of these two sleep functions inhumans. In this task, a target consistingof a horizontal or vertical array of threediagonal bars is displayed briefly againsta background of horizontal bars, fol-lowed by a blank screen, then by a mask.The interval between the target and themask (interstimulus interval, or ISI) isvariable, and the ISI needed to achieve80% correct responses is taken as a mea-sure of perceptual ability. This proce-

dure has become a classic task to inves-tigate the role of sleep in memory con-solidation. After a single trainingsession, performance on this taskimproves only after subjects have sleptduring the first night after training2 (ISIdecreases by 30 ms, Fig. 1a). This slowlearning process is sensitive to REMsleep deprivation4 and is optimal onlywhen subjects have had both non-REMsleep (early in the night) and REM sleep(late in the night)5.

Now Mednick et al.1 show thatsubjects’ performance deteriorates ifthey are trained on the task four timesat regular intervals on the same day. Themore they are trained, the worse theirperformance becomes (about 40 msincrease in ISI after the fourth session;Fig. 1 in ref. 1). This deterioration canbe avoided if the subjects are allowed tonap for 30 to 60 minutes at the begin-ning of the afternoon, as expected fromprevious studies (for instance, ref. 6).Importantly, the authors went on toshow that performance does not deteri-orate if stimuli in the last training ses-sion are presented in the visualhemifield contralateral to the one usedin the initial training sessions, that is,when the probed visual cortex is con-tralateral to the initially trained cortex.This observation leads to the importantconclusion that the deterioration inbrain function is task dependent andregionally specific. In other words, it isnot a global effect of fatigue on the

Carlyle Smith is in the Department ofPsychology, Trent University, Peterborough,Canada. The other authors are at the CyclotronResearch Centre, University of Liège, Liège,Belgium.e-mail: [email protected]

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


brain, but a specific effect on thebrain regions involved in the task. A60-minute nap is more efficientthan a 30-minute nap to recover theinitial level of performance, themain difference being more timespent in non-REM sleep during longnaps. These data can be viewed asevidence for a restorative effect ofsleep, especially non-REM sleep, onregional brain function. They echoother examples of regionally specif-ic, use-dependent sleep processes inwhich increased stimulation duringwakefulness is followed by aenhanced intensity of sleep only inthe corresponding sensory cortex7,8.For instance, in humans, after pro-longed vibratory stimulation of thehand, slow-wave activity (an EEGmarker of the intensity of slow-wavesleep) is specifically increased overthe controlateral somatosensorycortex during subsequent sleep7.These previous studies, however, didnot demonstrate a regionally specif-ic functional benefit of sleep.

Although the authors discusstheir findings in terms of thepreservation of information in thecerebral networks as steps in mem-ory formation and consolidation, thebehavioral data only suggest therestoration of an optimal brain func-tion during sleep. At the cellular andmolecular levels, the cerebral correlatesof the effects reported by Mednick et al.1 are still a matter of speculation.The restorative mechanisms are stillpoorly understood and might involve,for instance, the elimination of toxiccompounds (for example, the responseto oxidative stress by glutathione9) andthe reconstitution of energy stores(such as brain glycogen10). Likewise, themechanisms of consolidation of mem-ory traces during sleep are stillunknown, although gene transcriptionand protein synthesis are probablyinvolved11. What is known, however, isthat neural activity enhances the localrelease of many compounds like NO,adenosine or cytokines (IL1β, TNFα ,growth factors, interferons α and γ, andmany others)12. The level of some ofthese molecules increases with sleepneed, suggesting that they might signalthat neurons have been working for an

1. Mednick, S. et al. Nat. Neurosci. 5, 677–681(2002).

2. Stickgold, R., James, L. & Hobson, J. A. Nat.Neurosci. 3, 1237–1238 (2000).

3. Karni, A. & Sagi, D. Proc. Natl. Acad. Sci.USA 88, 4966–4970 (1991).

4. Karni, A., Tanne, D., Rubenstein, B. S.,Askenasy, J. J. & Sagi, D. Science 265,679–682 (1994).

5. Gais, S., Plihal, W., Wagner, U. & Born, J.Nat. Neurosci. 3, 1335–1339 (2000).

6. Bonnet, M. H. & Arand, D. L. J. Sleep Res. 4,71–77 (1995).

7. Kattler, H., Dijk, D. J. & Borbely, A. A. J.Sleep Res. 3, 159–164 (1994).

8. Vyazovskiy, V., Borbely, A. A. & Tobler, I. J. Sleep Res. 9, 367–371 (2000).

9. Inoue, S., Honda, K. & Komoda, Y. Behav.Brain Res. 69, 91–96 (1995).

10. Benington, J. H. & Heller, H. C. Prog.Neurobiol. 45, 347–360 (1995).

11. Maquet, P. Science 294, 1048–1052 (2001).

12. Krueger, J. M., Obal, F. J., Fang, J., Kubota, T.& Taishi, P. Ann. NY Acad. Sci. 933, 211–221(2001).

13. Strecker, R. E. et al. Behav. Brain Res. 115,183–204 (2000).

14. Brandt, J. A. et al. Brain Res. 898, 105–112(2001).

extended period of time12,13. Thesemolecules are not only released in ause-dependent manner but participatein a cascade of molecular events thateventually promote sleep, and especial-ly non-REM sleep (Fig. 1 in ref. 12).

Finally, some authors suggest thatseveral of these molecular compounds,especially growth factors, not onlyinduce sleep but are also involved insynaptic plasticity12,14 . However, theirrole remains to be established in theconsolidation of memory traces duringsleep. The view thus emerges that atask-dependent, regionally specific setof neural networks challenged by pre-vious waking activity could be identi-fied by the production and release of anumber of molecules. These com-pounds would promote sleep and leadto a cascade of processes involved incerebral restoration, brain plasticity orboth. To gain a comprehensive under-standing of these mechanisms, futureresearch will have to characterize theseprocesses more fully.

news and views

nightday nightday nightday day

nap

Restoration

Consolidation? ?

Det

erio

ratio

nIm

prov

emen

t

Per

form

ance

Fig. 1. The role of sleep in memory. (a) The texture detection task is thought to rely on plastic changesin the primary visual cortex. With four sessions on the same day, performance worsens monotonically(red dots), unless the subjects can nap (purple square and dots). After a single training session, the per-formance improves after the first night’s sleep. This improvement is significant if the subjects can sleepduring the first part of the night (mainly non REM sleep, white dot) but is optimal only if they sleep thewhole night (both non REM sleep and REM sleep, first blue dot). Performance still improves after thesecond and third nights (last two blue dots). (b) The hypothetical time course of restorative processes(top) and consolidation of memory traces (bottom). Restoration and memory processing are usuallyconsidered as two different sleep functions. At the cellular level, it is not known whether restorationand consolidation processes are distinct or have mechanisms in common.

Amy Center

a

b

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

Regulation of dendriticdevelopment by the ARFexchange factor ARNODelia J. Hernández-Deviez1, James E. Casanova2 and Jean M. Wilson1

1 Department of Cell Biology and Anatomy, College of Medicine, University of Arizona, Box 245044, Tucson, Arizona 85718, USA

2 Department of Cell Biology, University of Virginia Health System, 1300 Jefferson Park Avenue, Charlottesville, Virginia 22908, USA

Correspondence should be addressed to J.M.W. ([email protected])

Published online: 28 May 2002, doi:10.1038/nn865

Here we analyzed the role of ARF6, a member of the ADP-ribosylation factor (ARF) family of small GTPases, in dendriticarbor development in rat hippocampal neurons in culture. Over-expression of the inactive form of the GTP exchange factor ARNO(ARF nucleotide binding site opener) or inactive ARF6 enhanceddendritic branching, whereas coexpression of either Rac1 (a mem-ber of the Rho family of small GTPases known to control den-dritic dynamics and growth1–4) or active ARF6 with inactiveARNO eliminated the enhanced branching effect. These resultsindicate that the ARF family of small GTPases contributes to theregulation of dendritic branching, and that ARF6 activation turnson two independent pathways that suppress dendritic branchingin vivo: one through Rac1 and the other through ARF6.

Dendritogenesis is a dynamic process that continues through-out the life of a neuron. The correct development and stabilizationof dendritic arbors is essential for neurons to be able to receive,integrate and process information5. During development, neuronsthat migrate into the hippocampus actively extend processes toestablish precise connections6. We found that ARNO is presentduring these early events. Immunolabeling showed that on embry-onic day 17 (E17) there was strong ARNO labeling in the ventric-ular and marginal zones (Fig. 1a) and weaker labeling of cells withinthe hippocampus. Both MAP2 kinase–positive and –negative cellswere labeled (data not shown), consistent with ARNO’s role as aubiquitous ARF regulator. On E18, concentrated labeling of cellsat the periphery was absent, but more cells within the substanceof the hippocampus were strongly labeled (Fig. 1b), perhaps reflect-ing cell migration. Labeling was found throughout the cell bodiesand in processes (Fig. 1c). Notably, the subcellular distribution ofARNO showed that it was localized to sites of active processextension: the lamellar extensions and ruffling areas at early stages(Fig. 1d) and the tips of processes in later stages (Fig. 1e). ARNOwas localized to the perinuclear region at all stages. Immunoblotanalysis and previous results7 showed that ARNO was present inboth embryonic and adult hippocampus (Fig. 1f), and that it waspredominantly limited to the membrane fraction on E18. AsARNO functions at membrane surfaces, this partitioning intoembryonic membranes indicates that ARNO is highly activatedduring early dendritogenesis. Further immunoblot analysis showedthat ARF1/3 and ARF6 were present in embryonic and adult hip-pocampus (Fig. 1f), indicating a role for these molecules in bothdeveloping and mature brain. These morphological and bio-chemical localizations are consistent with a direct role for ARNOand its ARF substrates in dendritic initiation and remodeling.

To investigate the role of ARNO in dendritogenesis, we over-expressed wild-type and inactive forms of ARNO8 in culturedneurons. ARNO is recruited to membrane phosphoinositidesvia its pleckstrin homology domain, and when the inactive formis overexpressed, it may displace endogenous ARNO and inhib-it nucleotide exchange on ARF molecules. Untransfected cellscontained a few thick dendrites (Fig. 2a), and neurons overex-pressing wild-type ARNO showed no quantifiable effects on den-dritic complexity (Fig. 2b and h). Overexpression of inactiveARNO (ARNO-E156K), by contrast, resulted in a pronouncedchange in the dendritic arbor (Fig. 2c and h). Cells expressingARNO-E156K had many dendritic branches, which formed ameshwork around the cell body (Fig. 2c), causing a 4.5-foldincrease in the number of dendritic tips (Fig. 2h). These resultsindicate that ARNO could be a powerful regulator of dendriticdevelopment and branching.

ARNO acts as a GTP exchange factor for both ARF1 andARF6 in vitro9,10. We found that dominant-negative ARF6(ARF6-T27N) resulted in a dendritic morphology similar to thatseen with inactive ARNO. Cells expressing ARF6-T27N contained

brief communications


Fig. 1. ARNO is expressed in developing hippocampus. Red, anti-ARNOantibodies9. Blue, DAPI nuclear staining. (a) On embryonic day 17 (E17),cells in the ventricular and marginal zones showed high ARNO expression(arrows). There was weaker labeling in the substance of the hippocampus(arrowheads). (b) On E18, strongly labeled cells were seen throughouthippocampus (arrows) and some with weaker labeling (arrowheads). Scalebars, 50 µm (a, b). (c) Neuron in situ: ARNO was present in the cell bodyand in an extension (arrows). N, nucleus. (d, e) Neuronal culturesimmunolabeled with anti-ARNO (green) and the nuclear stain propidiumiodide (red). (d) At 1d in vitro, ARNO at lamellar extensions (arrows) andruffles (arrowhead). (e) Later at 3 d in vitro, ARNO in the tips of processes(arrows). Arrowhead, axon; N, nucleus. Scale bars, 5 µm (c–e). (f) Supernatant (sup) and pellet from embryonic and adult hippocampiimmunoblotted using anti-ARNO, ARF6 and ARF1/3 antibodies. Allexperiments were approved by the Institutional Animal Care Use andCommittee of the University of Arizona.

a b

c

ed

gf

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om



many short dendrites with a complex dendritic arbor (Fig. 2dand h). In contrast, overexpression of ARF6 caused no quantifi-able effects on complexity (Fig. 2h). Overexpression of the con-stitutively active form of ARF6 (ARF6-Q67L), however, resultedin a small decrease in the number of dendritic tips (Fig. 2h).

If the effects of ARNO-E156K were mediated through ARF6,coexpression of active ARF6 and ARNO-E156K should reversethe enhanced dendritic complexity. Indeed, coexpression of thesemolecules resulted in a complete loss of dendritic processes (Fig. 2f and h). Although we cannot rule out the possibility thatother ARF6-specific GTP exchange factors were involved, ourresults are consistent with the hypothesis that ARNO affects den-drite dynamics through an ARF6 pathway. Although ARF6-Q67Lcompletely reversed the effect of ARNO-E156K on dendriticbranching, it is possible that ARF1 is also involved in theseevents. Attempts to test this hypothesis by overexpressing wild-type and mutant forms of ARF1 resulted in cell death.

During cell migration, Rac1 is activated downstream ofARF6 (ref. 11). To determine if Rac1 is a downstream effector ofARF6 during dendritogenesis, we first analyzed the effects ofexpressing Rac1 or dominant-negative Rac1 (Rac1-N17) ondendritic morphology. Notably, overexpression of Rac1-N17resulted in a phenotype (Fig. 2e) very similar to that of cells

overexpressing ARNO-E156K, including a 3-fold increase inthe number of dendritic tips (Fig. 2h). In contrast, overexpres-sion of wild-type Rac1 did not affect dendritic complexity (Fig. 2h). In mature hippocampal neurons, Rac1-N17 did notsignificantly affect dendritic branching3, suggesting that regu-lation of dendritogenesis differs between mature neurons andthose in early stages of arbor development.

We predicted that dendritic branching would be negativelyregulated by ARNO and ARF6 signaling. ARF6 activation turnson two independent pathways, one through Rac1 and the otherthrough phosphoinositide kinases12, both of which contributeto the suppression of dendrite branching. Thus, if the effects ofRac-N17 expression on dendrite complexity were mediatedthrough the ARNO and ARF6 pathway, then coexpression ofRac1 with ARNO-E156K should block this increase in dendrit-ic complexity. We found that neurons cotransfected withARNO-E156K and Rac1 had less dendritic complexity than didcells expressing only ARNO-E156K (Fig. 2g and h), indicatingthat the enhanced dendritic complexity was in part due toARNO-mediated effects on the Rac1 pathway. In a parallel sup-pression experiment, however, neurons coexpressing activeARF6 and Rac1-N17 showed a loss of dendritic extensions (Fig. 2h). These results indicate that ARF6-stimulating path-ways can act through ARF6 effectors directly or through Rac1 tomodulate dendritic dynamics.

The control of dendritic arborization has important impli-cations for nervous system function5,13. ARF6 and ARNO maybe responsible for the precise developmental control of both ini-tiation and branching early in dendritogenesis, as well as for therapid changes induced by synaptic activity and neurotrophicfactors later on. This regulation could influence the extent ofsynaptic inputs and the integration of information that is cru-cial for normal nervous system functioning. The next criticalstep is to determine the factors that regulate ARNO activity indeveloping neurons.

AcknowledgmentsWe thank J. Settleman for the Rac plasmids and P. Oyarbide and M. Duran for

assistance with statistical analysis. We also thank M. Ramaswami, R. Parton,

V. Faundez and N. McMullen for critical reading of the manuscript. This work

was supported by National Institutes of Health grants DK43329 (to J.M.W.) and

AI32991 (to J.E.C.). D.H.D. was supported by Consejo Nacional de

Investigaciones Científicas y Tecnológicas, Venezuela and University of Los

Andes, Mérida-Venezuela.

Competing interests statementThe authors declare that they have no competing financial interests.

RECEIVED 21 MARCH; ACCEPTED 18 APRIL 2002

1. Threadgill, R., Bobb, K. & Ghosh, A. Neuron 19, 625–634 (1997).2. Li, Z., Van Aelst, L. & Cline, H. T. Nat. Neurosci. 3, 217–225 (2000).3. Nakayama, A. Y., Harms, M. B. & Luo, L. J. Neurosci. 20, 5329–5338

(2000).4. Luo, L. Nat. Rev. Neurosci. 1, 173–180 (2000).5. Scott, E. K. & Luo, L. Nat. Neurosci. 4, 359–365 (2001).6. Altman, J. & Bayer, S. A. J. Comp. Neurol. 301, 325–342 (1990).7. Suzuki, I. et al. Brain Res. Mol. Brain Res. 98, 41–50 (2002).8. Frank, S. R., Hatfield, J. C. & Casanova, J. E. Mol. Biol. Cell 9, 3133–3146

(1998).9. Frank, S., Upender, S., Hansen, S. H. & Casanova, J. E. J. Biol. Chem. 273,

23–27 (1998).10. Chardin, P. et al. Nature 384, 481–484 (1996).11. Santy, L. C. & Casanova, J. E. J. Cell Biol. 154, 599–610 (2001).12. Honda, A. et al. Cell 99, 521–532 (1999).13. Cline, H. T. Curr. Opin. Neurobiol. 11, 118–126 (2001).

Fig. 2. ARNO mediates dendrite initiation and branching through ARF6and Rac1. (a–c) Neurons transfected after 1 d in vitro and immunola-beled after 9 d with anti-MAP2 (green) and anti-myc (ARNO, red), anti-HA (ARF6, red), or anti-FLAG (Rac1, red). (a) Mock transfectants and(b) cells expressing wild-type ARNO or (c) ARNO-E156K. (d–g) Cellsexpressing (d) ARF6-T27N (e) Rac1-N17 (f) ARNO-E156K and ARF6-Q67L or (g) ARNO-E156K and Rac1. Arrowheads, axons. Arrows, den-dritic branches. Scale bars, 10 µm. (h) Dendrite complexity wasquantified by counting dendritic tips. Changes were compared usingone-way analysis of variance (ANOVA) and expressed as the mean num-ber of dendritic tips per cell. Data are means ± s.d. (n = 10). P < 0.05 (*)and P < 0.01 (**) compared to control, P < 0.01 compared to ARNO-E156K (**, red), P < 0.01 compared to Rac1-N17 (**, blue). Cells shownare near the mean. Cell survival was between 65% (ARF6 expressionplasmids) and 85% (other plasmids).

a

c

b

d

e g

f

h

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

tion of ICSS thresholds, rats were given differential access to acontinuous schedule of intravenous cocaine self-administration(0.25 mg/injection): 0 hour (naive rats, n = 6), 1 hour (short-access or ShA rats, n = 9) or 6 hours per day (long-access or LgArats, n = 11)11. Repeated prolonged access to cocaine or heroin(6 hours or more) produces an escalation in drug intake notobserved with limited access (1 hour) to the drug11,12. ICSSthresholds were measured daily 3 hours and 17–22 hours aftereach self-administration session, the second time point occur-ring 1 hour before each next session. This design allowed us toprobe the functional state of brain reward systems during thewithdrawal periods separating repeated self-administration ses-sions (Supplementary Methods online). All experimental pro-cedures were approved by the Animal Care and Use Committeeof The Scripps Research Institute.

ICSS thresholds remained unchanged and stable for the dura-tion of the experiment in both drug-naive controls and ShA rats(Fig. 1a). In contrast, ICSS thresholds in LgA rats were progres-sively elevated between sessions to 30% above baseline (Fig. 1a).ICSS thresholds deviated more and more from baseline in LgA ratsbecause elevated ICSS thresholds failed to return to baseline beforeeach next self-administration session. This effect was not due to adecreased ability to respond, as no difference between groups wasobserved in response latencies for ICSS (Fig. 1c). The gradual ele-vation in thresholds was associated with a dramatic escalation ofboth first-hour (Fig. 1b) and total cocaine consumption (from75.5 ± 13.9 to 125 ± 4.3 injections; data not shown). The slope ofelevation in reward thresholds measured 1 hour before access tococaine was highly correlated (r = 0.78, P < 0.01) with the slopeof escalation in total cocaine intake (Fig. 1d). Also, in all LgA rats,daily levels of total cocaine intake were positively correlated withdaily ICSS thresholds (percent change from baseline) measured 1 hour before (range, 0.27–0.91) or 3 hours after (0.23–0.82) dailyaccess to cocaine. This positive correlation was significant in 8 of 11 LgA rats (P < 0.05; mean r = 0.73; range, 0.57–0.91). These find-ings may suggest an insidious process of escalation wherebydecreased reward function contributes to increased cocaine intake,which in turn further decreases brain reward function.

Neurobiological evidencefor hedonic allostasisassociated withescalating cocaine useSerge H. Ahmed1,2,3, Paul J. Kenny2,3, George F. Koob2

and Athina Markou2

1 Laboratoire de Neuropsychobiologie des Désadaptations, Université de Bordeaux 2, CNRS-UMR 5541, 146 rue Léo-Saignat, Bordeaux 33076, France

2 Department of Neuropharmacology, The Scripps Research Institute, 10550 North Torrey Pines Road, La Jolla, California 92037, USA

3 The first two authors contributed equally to this work.

Correspondence should be addressed to S. H. A. ([email protected])

Published online: 10 June 2002, doi:10.1038/nn872

A paradoxical aspect of the transition to drug addiction is thatdrug users spend progressively more time and effort to obtaindrug hedonic effects that continually decrease with repeated expe-rience1,2. According to the hedonic allostasis hypothesis3,increased craving for and tolerance to the hedonic effects of drugsresult from the same chronic alteration in the regulation of brainreward function (allostasis). Here we show in rats that repeatedwithdrawals from prolonged cocaine self-administration pro-duces a persistent decrease in brain reward function that is high-ly correlated with escalation of cocaine intake and that reducesthe hedonic impact of cocaine.

During acute withdrawal from prolonged exposure to vari-ous drugs of abuse, intracranial self-stimulation (ICSS) thresh-olds—an operational measure of brain reward function—increaseabove basal levels but return to baseline hours afterwards4–7. Thiselevation in thresholds associated with acute with-drawal reflects a transient decrease in brain rewardsensitivity that counterbalances the threshold-low-ering effects of drugs8,9. Unknown, however, iswhether this transient decrease in reward worsensand becomes chronic with repeated withdrawalsand whether it is associated with the developmentof compulsive drug use. Male Wistar rats (300–350 g) with bipolar electrodes in the posterior lat-eral hypothalamus were trained in a discrete-trial,current-intensity ICSS protocol10. After stabiliza-

Fig. 1. Relationship between elevation in ICSSreward thresholds and cocaine intake escalation. (a) Percent change from baseline ICSS thresholds; (b) percent change from baseline response latencies(3 hours and 17–22 hours after each self-administra-tion session; first data point 1 hour before the firstsession). (c) Number of cocaine injections earnedduring the first hour of each session. (d) Correlationbetween the slope of escalation in total cocaineintake and the slope of elevation in ICSS thresholds inLgA rats. Slope coefficients were computed by fittingthe self-administration and ICSS data with a linearfunction. *P < 0.05 compared to drug-naive and/orShA rats, tests of simple main effects.



a b

c d

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

One day after escalation testing, all rats received an intravenousinjection of vehicle and 3 hours later an injection of the unit doseof cocaine available during self-administration. ICSS testing began1 min after each injection. In all groups, cocaine significantly low-ered ICSS thresholds relative to thresholds measured after vehicleadministration, reflecting the reward-facilitating effect of cocaine(Fig. 2). The net threshold-lowering effect of cocaine (thresholdsafter vehicle minus thresholds after cocaine administration) didnot vary significantly between the groups (Fig. 2, inset). Never-theless, in LgA rats, the chronic elevation in basal ICSS rewardthresholds shifted the net threshold-lowering effect of cocaineupward, thereby preventing the threshold from reaching the sameabsolute level as in controls after the same challenge (Fig. 2).

Two days after escalation testing, daily access of LgA rats tococaine was reduced from 6 hours to 1 hour for 8 consecutivedays; this duration of access does not induce drug intake escala-tion (refs. 11,12 and present data). During the post-escalationphase, ICSS thresholds were measured only 1 hour before eachself-administration session. Reward thresholds (Fig. 3a) and first-hour cocaine intake (Fig. 3b) of LgA rats remained significantlyelevated above control levels for at least 8 days after cessation ofprolonged access to cocaine.

This study supports the hedonic allostasis hypothesis of drugaddiction3 by showing that the transient counteradaptive reactionafter acute withdrawal from prolonged cocaine exposure5 showsa residual hysteresis with repeated withdrawals that is highly cor-related with cocaine intake escalation and that leads to the estab-lishment of a persistent deficit (at least eight days) in regulation ofbrain reward function. These findings suggest that hedonic respon-siveness to the environment decreases progressively during repeat-ed withdrawals from prolonged access to cocaine, increasing theanimal’s motivation to seek the threshold-lowering effects ofcocaine to reverse this hedonic deficit. The persistence of this hedo-nic deficit may be part of the neurobiological basis for continuedcraving and increased vulnerability to relapse associated with drugaddiction3. The present findings also suggest that tol-erance to the hedonic effects of cocaine does notresult from a decreased effect of cocaine on basalreward thresholds, consistent with findings of no

Fig. 3. Persistent elevation in ICSS reward thresholdsafter cessation of prolonged access to cocaine self-administration. (a) Percent change from baseline ICSSthresholds. ICSS measurements were made daily onehour before access to cocaine. (b) Number of cocaineinjections earned per one-hour session. In both graphs,LgA rats were significantly different from other groupsthroughout post-escalation testing (P < 0.05, Newman-Keuls tests).



change in cocaine pharmokinetics or pharmacodynamics aftercocaine intake escalation (S.H.A. et al., unpublished data). Rather,tolerance results from the establishment of a new basal rewardthreshold that shifts the unchanged threshold-lowering effect ofcocaine upward and therefore prevents thresholds from reachingthe same absolute level as before repeated, prolonged exposure tococaine. Thus, more doses are progressively needed to maintainthe same hedonic effect, further aggravating the dysregulation ofbrain reward function. Thus, hedonic allostasis provides a parsi-monious explanation to one of the enduring paradoxes of drugaddiction by showing how a subject becomes both hedonicallydependent on a drug and tolerant to its hedonic effects. Treatmentsthat block elevation in brain reward thresholds produced by chron-ic cocaine thus would be predicted to block escalation of cocaineintake and could be tested as new therapies for addiction13.

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

AcknowledgmentsSupported by the Centre National de la Recherche Scientifique (S.H.A.), the

Peter McManus Charitable Trust (P.J.K.) and grants from the National Institute

on Drug Abuse (DA04398 to G.F.K. and DA11946 to A.M.). This is manuscript

number 14258-NP from The Scripps Research Institute.


RECEIVED 26 FEBRUARY; ACCEPTED 23 APRIL 2002

1. Koob, G. F. Neuron 16, 893–896 (1996).2. Robinson, T. E. & Berridge, K. C. Brain Res. Rev. 18, 247–291 (1993).3. Koob, G. F. & Le Moal, M. Neuropsychopharmacology 24, 97–129 (2001).4. Leith, N. J. & Barrett, R. J. Psychopharmacologia 46, 19–25 (1976).5. Markou, A. & Koob, G. F. Neuropsychopharmacology 4, 17–26 (1991).6. Schulteis, G., Markou, A., Cole, M. & Koob, G. F. Proc. Natl. Acad. Sci. USA

92, 5880–5884 (1995).7. Epping-Jordan, M. P., Watkins, S. S., Koob, G. F. & Markou, A. Nature 393,

76–79 (1998).8. Kornetsky, C. & Esposito, R. U. Fed. Proc. 38, 2473–2476 (1979).9. Wise, R. A. Annu. Rev. Neurosci. 19, 319–340 (1996).10. Markou, A. & Koob, G. F. Physiol. Behav. 51, 111–119 (1992).11. Ahmed, S. H., Walker, J. R. & Koob, G. F. Neuropsychopharmacology 22,

413–421 (2000).12. Ahmed, S. H. & Koob, G. F. Science 282, 298–300 (1998).13. Markou, A., Kosten, T. R. & Koob, G. F. Neuropsychopharmacology 18,

135–174 (1998).

Fig. 2. Acute effect of cocaine on ICSS reward thresholds. Percentchange from baseline ICSS thresholds following intravenous saline orcocaine administration (0.25 mg/injection). Bars in inset represent per-cent difference between thresholds after vehicle and cocaine adminis-tration. *P < 0.05 compared to vehicle, tests of simple main effects.

a b

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

Fig. 1. Relationship between elevation in ICSS rewardthresholds and cocaine intake escalation. (a) Percentchange from baseline ICSS thresholds; (b) percentchange from baseline response latencies (3 hours and17–22 hours after each self-administration session;first data point 1 hour before the first session). (c) Number of cocaine injections earned during thefirst hour of each session. (d) Correlation betweenthe slope of escalation in total cocaine intake and theslope of elevation in ICSS thresholds in LgA rats.Slope coefficients were computed by fitting the self-administration and ICSS data with a linear function. *P < 0.05 compared to drug-naive and/or ShA rats,tests of simple main effects.


Neurobiological evidence for hedonic allostasis associated with escalating cocaine useSerge H. Ahmed, Paul J. Kenny, George F. Koob and Athina MarkouNat. Neurosci. 5, 625–626 (2002)

A mistake was introduced during the preparation of this paper. In the AOP version, panel labels b and c in Fig. 1 were mistakenlyswitched. The lower left panel should be labeled b, and the upper right panel should be labeled c. This mistake has been corrected inthe HTML version and will appear correctly in print. The PDF version available online has been appended.

erratum

a c

b d

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om



Dyskinesias followingneural transplantation inParkinson’s diseasePeter Hagell1,4, Paola Piccini6, Anders Björklund2, Patrik Brundin3, Stig Rehncrona5, Håkan Widner3,4,Lesley Crabb7, Nicola Pavese6, Wolfgang H. Oertel9, Niall Quinn8, David J. Brooks6 and Olle Lindvall1,4

1 Section of Restorative Neurology, 2Division of Neurobiology and 3Section for Neuronal Survival, Wallenberg Neuroscience Center, SE-221 84 Lund, Sweden

4 Division of Neurology and 5Division of Neurosurgery, Department of Clinical Neuroscience, University Hospital, SE-221 85 Lund, Sweden

6 MRC Clinical Sciences Centre and Division of Neuroscience, Faculty of Medicine, Imperial College, Hammersmith Hospital, Du Cane Road, London W12 0NN, UK

7 National Hospital for Neurology and Neurosurgery and 8Sobell Department of Motor Neuroscience and Movement Disorders, Queen Square, London, WC1N 3BG, UK

9 Department of Neurology, University of Marburg, DE-35033 Marburg, Germany

Correspondence should be addressed to O.L. ([email protected])


Severe dyskinesias during the ‘off ’ phases (periods of increasedParkinson’s disease (PD) disability) have been observed followingintrastriatal transplantation of human embryonic mesencephal-ic tissue1. Here we retrospectively analyzed 14 patients who werefollowed for up to 11 years after grafting, and found that dyski-nesias (abnormal involuntary movements and postures) increasedduring postoperative off phases, but were generally of mild tomoderate severity. Dyskinesia severity was not related to the mag-nitude of graft-derived dopaminer-gic re-innervation, as judged by18F-labeled 6-L-fluorodopa (FD)positron emission tomography(PET), indicating that off-phasedyskinesias probably did not resultfrom excessive growth of grafteddopaminergic neurons.

All 14 patients showed peak ‘on-phase’ (periods of little or no PD-related motor disability) dyski-nesias before grafting (Table 1). Of 11 cases for which we had access topreoperative video recordings duringoff phases, six showed some degree ofoff-phase dyskinesias before trans-plantation. Mild to moderate foot orneck dystonia (cramp-like abnormalposturing) was observed in threecases, and a fourth had marked footdystonia. A fifth patient manifestedvery mild repetitive movements of theright leg and generalized dystonicpostures. Finally, one patient showedmild, action-induced choreiformneck movements and hand dystonia.

Hyperkinesias (predominantly choreiform movements) anddystonias increased during off phases after transplantation (Table 2).There were no significant (P > 0.05) changes in the severity of peakon-phase dyskinesias (Table 2) or in the percentage of time spentin on-phase with dyskinesias (Table 1). In addition, when the over-all degree of dyskinesias in each patient was expressed as a globalclinical dyskinesia rating scale (CDRS)2 score (Table 2, see foot-note), there was a significant (P = 0.001) increase during off phas-es but not during peak on phases. Maximum off-phase global CDRSscores after grafting correlated with peak on-phase scores at thesame time point (Spearman’s Rho, rs = 0.634, P = 0.015; for detailsof all correlations, see Supplementary Table 1 online), but not withpreoperative on- or off-phase scores. Off-phase hyperkinesias anddystonias typically appeared concurrently, either in the same or dif-ferent body part(s), as choreiform movements intermingled withbrief dystonic postures. Repetitive, stereotypic or ballistic move-ments were also seen. In eight patients, dyskinesias were mild andcaused no distress or disability; maximum postoperative off-phaseglobal CDRS scores ranged between 0.5 and 4.5 (median 2.5,interquartile range 1.1–3.8). In the remaining six patients, maxi-mum postoperative off-phase global CDRS scores ranged between9 and 18 (median 12, interquartile range 10.1–15.4). Only in onecase did this constitute a clinical therapeutic problem. Of the sixpatients with preoperative off-phase dyskinesias, dystonia hadincreased in three and decreased in three patients at the last post-operative assessment; hyperkinesias had increased in all six patients.

Differential development of off- and on-phase dyskinesiaswas observed after grafting: two patients with virtually no pre-operative off-phase dyskinesias developed mild dyskinesias in offphases after grafting. Concomitantly, their pronounced preop-erative on-phase dyskinesias were reduced by >50% postopera-tively. Another patient with no preoperative, but pronouncedpostoperative, off-phase dyskinesias showed virtually no changein peak on-phase dyskinesias. In one of the patients with the mostpronounced postoperative off-phase dyskinesias, L-dopa anddopamine agonists were withdrawn for up to nine weeks, withno apparent effect on the dyskinesias.

Table 1. Characteristics of patient group (n = 14) and transplantation procedure.

At first At maximum postoperative P-valuetransplantation off-phase dyskinesiasa

Age/duration of PD (years) 52.0 ± 7.0/11.9 ± 2.2 –– ––

Hoehn & Yahr stageb 3.25 (3–4.25)c –– ––Daily dose of L-dopa equivalentsd 932.5 ± 477.9 565.4 ± 474.9 0.0004UPDRS motor scoreb,e 42.5 (40.25–55)c 27 (19.25–36)c 0.002Time spent in ‘off ’ (%)f 36.6 ± 21.1 24.7 ± 20.2 0.095Time spent in ‘on’ withdyskinesias (%)f 20.6 ± 16.7 14.4 ± 16.6 0.227VMs implanted in the putamen/ 6.3 ± 2.8/1.1 ± 1.0 –– ––caudate nucleusg

Of 18 patients who were transplanted, four with non-idiopathic PD10 were excluded. Twelve patients were graftedstereotaxically with fresh dissociated VM tissue from 5–9 week-old (postconception) human embryos10,11. In twopatients, the tissue was stored at 4°C in a hibernation medium that contained tirilazad mesylate and GDNF for1–8 d before implantation. Data are mean ± s.d. except where indicated. Comparisons were done with pairedStudent’s t-tests and—for the UPDRS—Wilcoxon signed-ranks test; α = 0.05 (2-tailed). Study procedures wereapproved by research ethical committees in Lund, London and Munich. aRecorded at the time of the highest post-operative off-phase dyskinesia scores for each patient. bAs assessed in practically defined ‘off’ (in the morning≥12 h after the last dose of anti-parkinsonian medication)12. cMedian (interquartile range). dOne-hundred L-dopaequivalents = 100 mg of standard L-dopa = 133 mg of controlled-release L-dopa = 10 mg of bromocriptine = 1 mgof pergolide = 5 mg of ropinirole = 2 mg of apomorphine. eOverall parkinsonian symptomatology assessed withthe UPDRS motor examination score13. fMean daily time, as recorded by the patients during the preceding month.gPer patient. Grafts were placed unilaterally in the putamen (n = 2) and putamen + caudate nucleus (n = 2), bilater-ally in the putamen + unilaterally in the caudate nucleus (n = 1), bilaterally in the putamen (n = 4) and bilaterally inthe putamen + caudate nucleus (n = 5).

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om



The severity of postoperative off-phase dyskinesias tendedto correlate negatively with preoperative putaminal FD uptake(rs = –0.549, P = 0.064; for scatter-plots, see Supplementary Fig. 1online). This finding indicates that the manifestation of off-phasedyskinesias after grafting, similar to that of L-dopa–induced on-phase dyskinesias3–5, can be related to the baseline severity of stri-atal dopaminergic denervation. Our observations argue againstthe notion that L-dopa–induced on-phase and graft-evoked off-phase dyskinesias have identical underlying mechanisms. Themixed type of dyskinesias seen in off-phases after transplantationis different from that typical of off-phase and on-phase in PD6,7.There was a significant correlation (rs = 0.634; P = 0.015) betweenthe severity of peak on and off dyskinesias postoperatively, andthe peak on-phase dyskinesia scores were higher in patients withmore pronounced off-phase dyskinesias. Thus, in several cases,these adverse effects seemed to be additive after transplantation.However, gradual reduction or no change of peak on-phase dys-kinesias after grafting, despite continued development of off-phase dyskinesias, was also seen. This raises the possibility thatgrafts can not only induce or worsen off-phase dyskinesias butalso may ameliorate peak on-phase dyskinesias5,8,9.

Off-phase dyskinesias were not associated with the most markedsymptomatic relief. No correlation was found between the improve-ment of off-phase unified PD rating scale (UPDRS) motor scoresand the maximum global postoperative off-phase CDRS scores (rs= 0.003, P = 0.991). The degree of motor improvement and reduc-tion in medication did not differ between patients with mild andthose with more pronounced postoperative dyskinesias. The evo-lution of off-phase dyskinesias typically followed a time course thatdiffered from that of symptomatic relief. In the three bilaterallygrafted patients who showed the highest postoperative off-phasedyskinesia scores, maximum improvement of UPDRS motor scoresoccurred by 12 months after transplantation, whereas off-phasedyskinesias reached their maximum at 24–48 months. These dataindicate that the mechanisms of graft-derived clinical improvement,believed to be restoration of striatal dopaminergic neurotransmis-sion, most likely differ from those of off-phase dyskinesias.

Our data do not support the idea that off-phase dyskinesias arecaused by overgrowth of grafted dopamine neurons. Neither theFD uptake in the scan done closest in time to maximum postop-


erative off-phase dyskinesias (rs = –0.132,P = 0.652) nor the increase compared topreoperative values (rs = –0.267, P = 0.401) correlated with off-phase glob-al CDRS scores or differed betweenpatients with mild and more pronounceddyskinesias. Nevertheless, such dyskine-sias could still be related to dopaminergicmechanisms in the striatum. Small graftscould give rise to dopamine spill-over thatreaches supersensitive receptors outsiderestricted islands of reinnervated striatalareas. Alternatively, off-phase dyskinesiasmight depend on transplantation-evokedchanges in the host striatum or on non-dopaminergic components in the grafts.Thus, global postoperative off-phaseCDRS scores correlated with the numberof ventral mesencephalon (VM) implant-ed in the putamen (rs = 0.562, P = 0.037).Conspicuously, the two patients whoreceived tissue that had been stored for1–8 days developed more pronounced off-

phase dyskinesias (global CDRS score: median 15, interquartilerange 12–18) than patients implanted with fresh tissue (median3.5, interquartile range 1.6–10.1). Similarly, the most dramaticpost-grafting off-phase dyskinesias reported from other centerswere observed after implantation of tissue that had been culturedfor up to four weeks1.

Our findings are contrary to the notion that off-phase dyski-nesias are characteristic for dopamine cell replacement per se andprovide no evidence that this side effect should stop the furtherdevelopment of a cell therapy for PD. However, the underlyingmechanisms must be better understood so that off-phase dyski-nesias following neural transplantation can be avoided.


AcknowledgmentsThis study was supported by the British and Swedish Medical Research Council,

the United Kingdom Parkinson’s Disease Society, the Gemeinnützige Hertie

Stiftung, the Skane County Council Research and Development Foundation and

the Kock, Wiberg, Söderberg and King Gustav V and Queen Victoria

Foundations.


RECEIVED 13 NOVEMBER 2001; ACCEPTED 14 MARCH 2002

1. Freed, C. R. et al. N. Engl. J. Med. 344, 710–719 (2001).2. Hagell, P. & Widner, H. Mov. Disord. 14, 448–455 (1999).3. Boyce, S., Rupniak, N. M., Steventon, M. J. & Iversen, S. D. Clin.

Neuropharmacol. 13, 448–458 (1990).4. Nutt, J. G. Ann. Neurol. 47 (Suppl. 1), 160–166 (2000).5. Lee, C. S., Cenci, M. A., Schulzer, M. & Björklund, A. Brain 123, 1365–1379

(2000).6. Cubo, E. et al. Arch. Neurol. 58, 1379–1382 (2001).7. Luquin, M. R. et al. Mov. Disord. 7, 117–124 (1992).8. Widner, H. et al. N. Engl. J. Med. 327, 1556–1563 (1992).9. Hauser, R. A. et al. Arch. Neurol. 56, 179–197 (1999).10. Lindvall, O. & Hagell, P. Prog. Brain Res. 127, 299–320 (2000).11. Lindvall, O. et al. Arch. Neurol. 46, 615–631 (1989).12. Langston, J. W. et al. Mov. Disord. 7, 2–13 (1992).13. Fahn, S. et al. in Recent Developments in Parkinson’s Disease Vol. 2 (eds. Fahn,

S., Marsden, C. D., Calne, D. B. & Goldstein, M.) 153–163 (MacMillanHealthcare Information, Florham Park, New Jersey, 1987).

Table 2. Occurrence of dyskinesias after neural grafting.

Preoperativea Maximum off- Latest P-valued

phase dyskinesiasb assessmentc

Practically defined ‘off ’ phasee

Hyperkinesias 0 (0–0) 4 (0.9–7.5) 3.2 (0.9–7.5) 0.00007Dystonia 0.5 (0–3) 2.8 (1.4–10.1) 2.8 (0–6.8) 0.042

Peak L-dopa induced ‘on’ phasef

Hyperkinesias 10 (5.6–13.8) 6 (3.8–11.5) 7 (3.4–14.9) 0.926Dystonia 5 (1–7.4) 2 (1.9–4.2) 2 (1.8–9.2) 0.924

Dyskinesias were retrospectively scored in random order from videos with the CDRS (maximum score =28)2, by one rater (P. H.) who was blind to the recording dates. Intrarater reliability (intraclass correlation,hyperkinesias = 0.98, dystonia = 0.88) was established a priori with separate video sequences2. The latestavailable preoperative video and videos from approximately 12 and 24 months after grafting and (for patientsfollowed beyond that) the latest available postoperative recording were examined. Preoperative videos duringoff and on periods were unavailable for three and two patients, respectively. Data are median (interquartilerange). Unpaired and paired comparisons of CDRS scores, as described in the main text, were done withtwo-tailed Mann-Whitney U-tests and Wilcoxon signed-ranks tests, respectively (α = 0.05). In the main text,overall dyskinesias in each patient are also expressed with a global CDRS score, derived as the sum of thehighest hyperkinesia or dystonia ratings from each body part (maximum score, 28; intrarater intraclass coeffi-cient, 0.98).aOff, n = 11; peak on, n = 12. bAt a mean of 39.8 months (range 11–132 months) postoperatively;n = 14. cAt a mean of 44.6 months (range 15–132 months) postoperatively; n = 14. dA Friedman test wasused. eAssessed in the morning ≥12 h after the last dose of anti-parkinsonian medication12. fPeak ‘on’ follow-ing intake of an individually standardized L-dopa dose, which was the same at each assessment.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

Parallel processing inhigh-level categorizationof natural imagesGuillaume A. Rousselet, Michèle Fabre-Thorpe andSimon J. Thorpe

Centre de Recherche Cerveau and Cognition (UMR 5549, CNRS-UPS), Facultéde Médecine de Rangueil, 133 route de Narbonne, 31062 Toulouse, France

Correspondence should be addressed to G.A.R.([email protected])


Models of visual processing often include an initial parallel stagethat is restricted to relatively low-level features, whereas activa-tion of higher-level object descriptions is generally assumed torequire attention1–4. Here we report that even high-level objectrepresentations can be accessed in parallel: in a rapid animal ver-sus non-animal categorization task, both behavioral and elec-trophysiological data show that human subjects were as fast atresponding to two simultaneously presented natural images asthey were to a single one. The implication is that even complexnatural images can be processed in parallel without the need forsequential focal attention.

High-order representations, up to the semantic level, can beaccessed very rapidly from brief picture presentations5,6. Event-related potential (ERP) experiments show that complex processingof natural scenes is achieved 150 ms after stimulus onset7. Thus,when humans are asked to decide whether a briefly presented pho-tograph contains an animal, the ERPs in response to targets anddistractors diverge sharply from 150 ms. There is evidence that thesedifferences reflect a real visual decision rather than physical differ-ences between stimulus categories8. The scenes used in such exper-iments typically contain several objects, suggesting that there is atleast some degree of parallelism in the underlying processing. Toexplore this issue, we analyzed whether processing speed is affectedwhen subjects are asked to process two pictures simultaneously.

Twenty subjects (mean age, 32.5 ± 10.9) performed a modifiedversion of the animal versus non-animal go/no-go task used inprevious studies7,8 (see Supplementary Fig. 1 and Supplemen-tary Methods online). In 20 blocks of 96 trials, single brief pre-

sentations (20 ms) of one image appearing 3.6° to the left or rightof a central fixation point were randomly mixed with the samenumber of dual presentations in which two images were flashedsimultaneously at the same eccentricities. In both conditions, ananimal target was presented on half of the trials. Target location(left versus right hemifield) was equiprobable.

Notably, subjects were able to process dual and single pre-sentations at the same speed (Fig. 1a). This is shown by both themedian reaction times (RTs, 390 versus 391 ms, respectively) andby the latencies of the earliest responses which were equal orshorter with two images than with one image (means of 255 ver-sus 260 ms, respectively; see Supplementary Table 1 online).

Subjects tended to be more accurate in the one-image condi-tion (90.4%) than with dual images (86.7%). This accuracy decreasewas predicted by a simple parallel model of processing (Fig. 1b) inwhich each of two simultaneously presented images is processedby a separate and independent mechanism, and both mechanismseventually converge on a single output system (see SupplementaryMethods). Further support for a parallel processing model comesfrom the tight fit between the experimental and the predictedcumulative performance accuracy (d’) curves (Fig. 1b).

The similarity in processing speed between the two condi-tions was confirmed by electrophysiological data (Fig. 2). Asso-ciated ERPs were averaged off-line for each condition anddifference waves were obtained by subtracting the ERP for cor-rect distractor trials from the ERP for correct target trials. Dif-ferential activation, probably generated within high-orderextrastriate visual areas9, was clearly seen at both occipito-temporal and frontal sites (see Supplementary Fig. 2 online).There was no effect of image condition on the onset of the dif-ferential occipital activity. Target and distractor signals divergedsharply around 140–150 ms after stimulus onset with anenhanced occipital negativity on target trials. This differentialoccipital activity became significant at similar latencies in both



Fig. 1. Behavioral results. (a) Reaction time distributions. Number ofresponses are expressed over time, with time bins of 5 ms. Correctresponses or ‘hits’ (thick top curves) are shown for the one target alone(gray) or for the target flanked by distractor (black). False alarms (thinbottom curves) are shown for the one distractor alone (gray) or for thetwo distractors (black). (b) Performance time course functions and pre-dictions of a parallel model of processing. Average performance accu-racy (in d’ units) is plotted as a function of processing time (in ms) forone image (gray curve) and for two images (black curve). The dynamic d’was calculated from the cumulative number of hits and false alarms ateach successive 10 ms time step. The predicted curve from the modelwas calculated using the probabilities of hits and false alarms calculatedfrom the experimental data in the one-image condition. A global fall inaccuracy from 90.4% in the one-image condition to 87.7% in the two-image condition was predicted by our model (see SupplementaryMethods). The experimental procedures were authorized by the localethical committee (CCPPRB No. 9614003) and all subjects gaveinformed consent to participate.

a

b

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

These physiological results directly supportpsychological models in which the competitivebottleneck is situated at a high level of integra-tion13,14. Further support comes from behavioralfindings that show that this animal versus non-animal categorization task can be done simulta-neously with another attentionally demandingtask15. We have no evidence, however, to suggestthat the ability to process more than one image atthe same time extends beyond the specific caseof two images presented simultaneously in thetwo hemifields. Further experiments will berequired to explicitly test whether the system canprocess other stimulus arrangements in parallel,

such as two images presented within the same hemifield or fourimages presented simultaneously.

Taken together, our data show that high-level object catego-rization of natural scenes can be done in parallel very rapidly andwithout the need for sequential focal attention. Whereas classicmodels of allocation of attentional resources consider ‘early’vision as being early in complexity and restrict low-level visionto the lower part of the cortical hierarchy (namely V1 and V2),early vision might more appropriately be considered as process-ing that is early in time.


AcknowledgmentsThis work was supported by the Cognitique program (COG35 and 35b).

Financial support was provided to G.A.R. by a PhD grant from the French

Government.


RECEIVED 24 JANUARY; ACCEPTED 29 APRIL 2002

1. Treisman, A. Philos. Trans. R. Soc. Lond. B Biol. Sci. 353, 1295–1306 (1998).2. Wolfe, J. M. Vis. Res. 34, 1187–1195 (1994).3. Kinchla, R. A. Annu. Rev. Psychol. 43, 711–742 (1992).4. McElree, B. & Carrasco, M. J. Exp. Psychol. Hum. Percept. Perform. 25,

1517–1539 (1999).5. Potter, M. C. J. Exp. Psychol. [Hum. Learn.] 2, 509–522 (1976).6. Biederman, I. Science 177, 77–80 (1972).7. Thorpe, S., Fize, D. & Marlot, C. Nature 381, 520–522 (1996).8. VanRullen, R. & Thorpe, S. J. J. Cogn. Neurosci. 13, 454–461 (2001).9. Fize, D. et al. Neuroimage 11, 634–643 (2000).10. Schall, J. D. Nat. Rev. Neurosci. 2, 33–42 (2001).11. Freedman, D. J., Riesenhuber, M., Poggio, T. & Miller, E. K. Science 291,

312–316 (2001).12. Sasaki, K., Gemba, H., Nambu, A. & Matsuzaki, R. Neurosci. Res. 18, 249–252

(1993).13. Duncan, J. Psychol. Rev. 87, 272–300 (1980).14. Chun, M. M. & Potter, M. C. J. Exp. Psychol. Hum. Percept. Perform. 21,

109–127 (1995).15. Li, F. F., VanRullen, R., Koch, C. & Perona, P. Proc. Natl. Acad. Sci. USA (in

press).

Fig. 2. Grand average ERPs and associated differen-tial activities. Grand average ERPs are plotted forcorrect target trials (thick line) and for correct dis-tractor trials (thin line). Results are shown for con-tralateral occipital electrodes (a) and all frontalelectrodes (b), for the one-image (top panel) andthe two-image (middle panel) conditions. Bottom,differential activity between the one- (gray) and two-image (black) conditions.


conditions (152 ms with one image versus 150 ms with twoimages, P < 0.0005) and then developed at the same rate, withthe same slope and amplitude in both conditions.

Differential activity was also seen at frontal sites (Fig. 2b),starting at around 160–170 ms in both conditions and becom-ing significant (P < 0.0005) at about the same latency: 173 ms(one image) and 175 ms (two images). At 190 ms after stimulusonset, the differential activity recorded in the one-image conditionbegan to diverge from that of the two-image condition, develop-ing with a steeper slope and finally reaching a higher amplitude.

These behavioral and electrophysiological results provide strongevidence that processing speed is unchanged between the one- andtwo-image conditions. Furthermore, the slight accuracy impair-ment (<4%) with two images can be explained using a very simplemodel in which the two images are processed by separate mecha-nisms that pool their outputs. The brief image presentations andinitial lateralization of visual inputs to the contralateral striate visu-al cortex indicate that each hemisphere could work in parallel on adifferent visual scene. This interpretation is strengthened by thehigh lateralization of the differential occipital activity.

The RT distributions (Fig. 1a) show that the number of ‘go’responses in the two-image condition, although initially similarto that seen in the one-image condition, was considerably loweraround the mean RTs. This effect might be explained by someform of competitive process occurring in the two-image condi-tion. Given the strong similarity between the occipito-temporaldifferential activity in the two conditions (Fig. 2a), it seemsunlikely that this competition affects the initial visual process-ing. Competition is more likely to occur later on at the point of‘sensorimotor decision’10. Evidence for a late competitive processat frontal sites comes from the late divergence seen between theone- and two-image conditions after 190 ms. High-level repre-sentations in occipito-temporal visual areas would be activatedindependently in each hemisphere. At frontal sites, by contrast,when integration of the outputs of the two cerebral hemispheresis needed for decision-making, competition could result fromfrontal processes related either to category-specific decision-making11 or to response inhibition on no-go trials12.


a b

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

Global effects of feature-based attention in humanvisual cortexMelissa Saenz1,2, Giedrius T. Buracas1 and Geoffrey M. Boynton1

1 The Salk Institute for Biological Studies, 10010 N. Torrey Pines Road, La Jolla, California 92037,USA

2 Department of Neurosciences, University of California San Diego, La Jolla, California 92093, USA

Correspondence should be addressed to G.M.B. ([email protected])


The content of visual experience depends on how selectiveattention is distributed in the visual field. We used functionalmagnetic resonance imaging (fMRI) in humans to test whetherfeature-based attention can globally influence visual corticalresponses to stimuli outside the attended location. Attention toa stimulus feature (color or direction of motion) increased theresponse of cortical visual areas to a spatially distant, ignoredstimulus that shared the same feature.

Visual attention influences local neuronal responses and psy-chophysical performance for stimuli at an attended location1,2.Another, feature-based mechanism of attention may globallyinfluence responses to stimuli outside the attended location thatshare features with the attended stimulus3,4.

Here we asked subjects (without shifting gaze) to attend toone direction of motion (the ‘target field’) within two overlap-ping fields of upward and downward moving dots on one sideof a central fixation point (Fig. 1a). Subjects were asked toignore a single field of dots moving up or down on the otherside. Dots had limited lifetimes (200 ms) to prevent subjectsfrom tracking individual dots.

Subjects did a speed discrimination task at threshold (79%correct, measured by staircase procedure before scanning).

Each of the three fields of dots moved at a baseline speed dur-ing one interval and slightly faster during the other interval.Subjects indicated the faster interval in the target field with akey press. A cue (0.5° line at fixation) signaled subjects to shiftattention between upward and downward fields every 20 s dur-ing the four-minute fMRI scan (Fig. 1a). Dots on the ignoredside did not change direction. Thus, conditions alternatedbetween ‘same’ (target field direction matches ignored stimu-lus) and ‘different’ (target field in the opposite direction).Before scanning, subjects trained for several hours until per-formance was stable. During scans, feedback was given duringthe intertrial interval.

Echo-planar imaging (EPI) was done with a Siemens(Munich, Germany) Vision 1.5-Tesla scanner (4 × 4 × 4 mmvoxels, 16 slices, TR = 2 s). We analyzed the blood oxygenationlevel–dependent (BOLD) response to the ignored stimulus inV1, V2, V3, V3A and MT+, the probable human homologue ofthe visual motion–responsive macaque areas MT and MST5.Visual areas were identified with standard fMRI cortical map-ping and flattening techniques6,7. We restricted analysis to pre-selected voxels within each visual area that responded to areference stimulus presented at the experimental stimulus loca-tion. As stimuli were presented in the periphery, responses tothe left and right stimuli were separated into different brainhemispheres (with one possible exception for area MT+, below).

All visual areas responded more strongly to the ignored stim-ulus when it moved in the same direction as the target field (Fig. 1b). Response amplitudes to the ignored stimulus were cal-culated for each visual area (Fig. 1c) as the amplitude of the best-fitting sine wave, phase adjusted for a typical hemodynamicdelay8. Stimulus, eye position (Supplementary Methods online),task difficulty and locus of spatial attention did not change, andso this modulation must be due to feature-based attention.

If subjects inadvertently shifted spatial attention to theignored stimulus in the ‘same’ condition, that should haveimpaired task performance1,9, but it was similar in both con-ditions (same, 87.6% correct; different, 87.0%, P > 0.05). Inseparate psychophysical trials, performance was impaired



Fig. 1. Feature-based attention to motion. (a) Stimuli (not to scale)were circular apertures (radius 5°) of coherently moving random dots inthe lower visual field (2.5° below and centered 11° to left and right offixation, baseline speed 10°/s). Arrow, field of 50 dots moving upward or downward. Dots were white (560 cd/m2) on a gray background (230 cd/m2). Dashed circle (not present in the experiment), spatial focusof attention. (b) fMRI time series of BOLD response (same versus dif-ferent) to ignored stimulus for MT+, averaged across three subjects and24 repetitions per subject. (c) Response amplitudes to ignored stimulus.(d) Response amplitudes to attended stimulus. (e) Attentional responseamplitudes as a percentage of stimulus-evoked response. Data in (c–e)are mean ± s.e.m. During each trial, stimuli were presented for twosequential 1-s intervals separated by a 100-ms interval in which only thefixation point was present. Trials started every 3.3 s. The order ofspeeds was independently randomized for each field of dots on everytrial, and the baseline speed was randomly and independently jitteredacross trials in all three fields of dots. Scans were counterbalanced forthe attended side (left/right), the starting attended direction (up/down)and the direction of motion on the ignored side (up/down). Three sub-jects with normal visual acuity participated, and all gave writteninformed consent. These experiments were approved by the SalkInstitutional Review Board.

a b

c d

e

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om



significantly when subjects were instructed to divide attentionacross left and right stimuli (Supplementary Methods online).Such distribution of attention would decrease fMRI responses tothe attended stimulus10 during the ‘same’ condition, resulting innegative response amplitudes (different > same), which we didnot observe (Fig. 1d). Area MT+ showed small but significant (P< 0.05) modulation of the response to the attended stimulus bycondition (Fig. 1d), perhaps because its large receptive fields11

may have included parts of the visual field responding to theignored stimulus.

To normalize the effect of attention to the baseline activityof each region in this task, we had our subjects perform a variantof the task, in which the ignored stimulus appeared when sub-jects attended its matching direction (‘same’) and disappearedwhen subjects attended its opposing direction (‘off ’). Weexpressed the attentional response as a percentage of the stim-ulus-evoked response (100 × (same – different)/(same – off);Fig. 1e). Thus 0% would indicate no attentional modulation,and 100% would indicate that attention modulated the responseto the ignored stimulus as much as removing it.

To determine the generality of this effect, we used color asthe attended feature in a second experiment with three subjectswith normal visual acuity and color perception. The attendedstimulus comprised overlapping fields of stationary red andgreen dots, and the ignored stimulus was a single field of redor green dots (Fig. 2a). We placed stimuli in the upper visualfield to include the suspected human homologue of the color-responsive macaque area V4, for which only an upper visual

field representation has been identified12. Subjects performeda threshold-increment luminance detection task. Other aspectsof the experimental design were unchanged (SupplementaryMethods online).

The fMRI response to the ignored stimulus was modulatedby feature-based attention to color in areas V1, V2, V3, V3A, V4(Fig. 2b) and MT+. Response amplitudes to the ignored stim-ulus were stronger during the ‘same’ condition, when its colormatched the attended color (Fig. 2b and c). As in the first exper-iment, there was no significant difference in task performancebetween ‘same’ (89.5%) and ‘different’ (87.2%, P > 0.05) con-ditions. Additionally, the BOLD response to the attended stim-ulus did not vary with condition in any area (Fig. 2d). As above,we normalized the attentional response amplitudes (Fig. 2c) tothe response elicited by cycling the ignored stimulus on and offduring the task (Fig. 2e). In both experiments, attentional mod-ulation in area MT+ was large relative to the stimulus-evokedresponse, consistent with larger effects of attention at later stagesof cortical processing13.

Our results demonstrate spatially global neuronal modulationdue to feature-based attention across multiple early stages of corticalvisual processing. Activity modulation by spatial attention occursin each of these cortical visual areas14,15. A feature-based mecha-nism of attention may thus work in parallel with a spatial mechanism to influence the earliest stages of cortical visual processing. Furthermore, these results are consistent with theproposed feature-similarity gain model3, whereby feature-basedattention modulates the gain of cortical neurons tuned to the attend-ed feature, anywhere in the visual field. The global effect of feature-based attention could be centrally involved in selecting the locationof behaviorally relevant stimuli. A feature-based increase in signalstrength would be useful for identifying and highlighting relevantperipheral stimuli during visual search, or for identifying parts ofthe same object by grouping stimuli with common features.

Note: Supplementary information is available on the Nature Neuroscience

website.

AcknowledgmentsSupported by the National Institutes of Health (EY12925) and the National

Science Foundation. We thank R.O. Duncan, G. Stoner, J. Reynolds and I. Fine

for comments on the manuscript.


RECEIVED 21 FEBRUARY; ACCEPTED 26 APRIL 2002

1. Posner, M. I., Snyder, C. R. & Davidson, B. J. J. Exp. Psychol. 109, 160–174(1980).

2. Desimone, R. & Duncan, J. Annu. Rev. Neurosci. 18, 193–222 (1995).3. Treue, S. & Martinez Trujillo, J. C. Nature 399, 575–579 (1999).4. McAdams, C. J. & Maunsell, J. H. J. Neurophysiol. 83, 1751–1755 (2000).5. Watson, J. D. et al. Cereb. Cortex 3, 79–94 (1993).6. Engel, S. A., Glover, G. H. & Wandell, B. A. Cereb. Cortex 7, 181–192 (1997).7. Sereno, M. I., McDonald, C. T. & Allman, J. M. Cereb. Cortex 4, 601–620

(1994).8. Heeger, D. J., Boynton, G. M., Demb, J. B., Seidemann, E. & Newsome, W. T.

J. Neurosci. 19, 7162–7174 (1999).9. Lee, D. K., Koch, C. & Braun, J. Percept. Psychophys. 61, 1241–1255 (1999).10. Vandenberghe, R. et al. J. Neurosci. 17, 3739–3750 (1997).11. Duffy, C. J. & Wurtz, R. H. J. Neurophysiol. 65, 1329–1345 (1991).12. Tootell, R. B. & Hadjikhani, N. Cereb. Cortex 11, 298–311 (2001).13. McAdams, C. J. & Maunsell, J. H. R. J. Neurosci. 19, 431–441 (1999).14. Gandhi, S. P., Heeger, D. J. & Boynton, G. M. Proc. Natl. Acad. Sci. USA 96,

3314–3319 (1999).15. Martinez, A. et al. Nat. Neurosci. 2, 364–369 (1999).

Fig. 2. Feature-based attention to color. (a) Stimuli (not to scale) werecircular apertures of stationary red and green random dots in the uppervisual field (2.5° above, centered 11° to left and right of fixation). R or G,field of 50 red or green dots on gray background. (b) fMRI time series inresponse to ignored stimulus for V4, averaged across three subjects and24 repetitions per subject. (c) Response amplitudes to ignored stimulus.(d) Response amplitudes to attended stimulus. (e) Attentional responseamplitudes as a percentage of stimulus-evoked response. Data in (c–e)are mean ± s.e.m.

a b

c d

e

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

review


The neurodegenerative diseases represent a challenge for scienceand medicine because of their prevalence, cost, lack of mechanism-based treatments, and impact on individuals and caregivers1–6.Genetic risk factors influence these age-associated, chronic illnesses,such as Alzheimer’s disease, motor neuron diseases, Parkinson’sdisease (PD), trinucleotide repeat diseases and prion disorders.They are characterized by dysfunction and death of specific pop-ulations of neurons and by the presence, in many instances, ofintracellular or extracellular protein aggregates. Although symp-tomatic treatments are available, there are no mechanism-basedtreatments. Recent research, particularly in animal models, hasbegun to provide new insights into the mechanisms of these dis-orders and has identified new targets for therapy.

The identification of mutations in specific genes causing eachof these neurodegenerative diseases has provided new opportu-nities to investigate the molecular participants in disease process-es and to explore pathogenic mechanisms using transgenicapproaches. In autosomal dominant genetic disorders, the mutantproteins often do not exhibit reductions in their normal func-tions, but instead acquire toxic properties that directly or indi-rectly affect the functions and viability of neurons. Introducingmutant genes into mice can reproduce some features of these dis-eases1,6–11. Autosomal recessive diseases, which usually lack thefunctional protein encoded by the mutant gene, can often bemodeled by gene knockout strategies. In both groups of disor-ders, gene knockout or overexpression of the genes that influ-ence pathogenic pathways have provided insight into diseasemechanisms and potential therapeutic targets. These model sys-tems can also be used to evaluate novel treatments and expeditethe path to clinical trials.

Here we focus on the familial forms of AD (FAD) and twoforms of motor neuron disease, familial amyotrophic lateralsclerosis (FALS) and spinal muscular atrophy (SMA). In thecase of FAD and FALS, we anticipate that understanding the

Genetically engineered mousemodels of neurodegenerativediseases

Philip C. Wong1,2,4, Huaibin Cai1,2,4, David R. Borchelt1,2,4 and Donald L. Price1,2,3,4

Departments of Pathology1, Neuroscience2 and Neurology3, and the Division of Neuropathology4, The Johns Hopkins University School of Medicine, 558 Ross Research Building, 720 Rutland Avenue, Baltimore, Maryland 21205-2196, USA

Correspondence should be addressed to P. C. W. ([email protected])

Recent research has significantly advanced our understanding of the molecular mechanisms ofneurodegenerative diseases, including Alzheimer’s disease (AD) and motor neuron disease. Here weemphasize the use of genetically engineered mouse models that are instrumental for understandingwhy AD is a neuronal disease, and for validating attractive therapeutic targets. In motor neuron dis-eases, Cu/Zn superoxide dismutase and survival motor neuron mouse models are useful in testingdisease mechanisms and therapeutic strategies for amyotrophic lateral sclerosis (ALS) and spinalmotor atrophy, respectively, but the mechanisms that account for selective motor neuron lossremain uncertain. We anticipate that, in the future, therapies based on understanding disease mech-anisms will be identified and tested in mouse model systems.

inherited illnesses will shed light on the more common spo-radic forms of AD and ALS. In the case of SMA, all cases areattributable to the same genetic mechanism.

Alzheimer’s diseaseProblems with memory and cognition appear during the seventhdecade in most individuals with AD, but may appear earlier, par-ticularly in familial cases12. Mental functions and activities ofdaily living become progressively impaired. The clinical signs ofAD result from selective degeneration of neurons in brain regionscritical for memory, cognitive performance and personality13,14.Dysfunction and death of these neurons lead to reduced num-bers of synaptic markers in their target fields14,15; the disruptionof synaptic communication is manifested by mental impairmentsand, finally, severe dementia14.

Two types of intracellular and extracellular protein aggregatesfound in the brain are a pathological hallmark of AD (Fig. 1).Neurofibrillary tangles are inclusions located within cell bodiesand proximal dendrites, and within filamentous swellings in dis-tal axons and synaptic terminals. Hyperphosphorylated isoformsof the microtubule-associated protein tau, which assemble intopoorly soluble paired helical filaments, are central feature of theseneurofibrillary tangles10. The extracellular aggregates in the brainof individuals with AD result from elevated levels of Aβ, a 4 kDamyloid peptide derived by cleavage of the amyloid precursorprotein (APP). Aβ monomers form oligomers and multimers,which assemble into protofilaments and then fibrils7,14,16,17. Even-tually, Aβ fibrils are deposited as the amyloid cores of neuritic orsenile plaques (amyloidosis), which are complex structures alsocontaining dystrophic neurites, astrocytes and microglia. Bothneurofibrillary lesions and plaques are preferentially localized tothe cortex, hippocampus and amygdala.

In some individuals with early-onset AD, the illness may beinherited as an autosomal dominant (that is, only a single copy of

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


the mutant gene is necessary to cause the disease). Such muta-tions are identified in at least three different genes: APP, PS1 andPS2 (refs. 1,6,18,19). APP is a type I transmembrane proteinexpressed in many different cell types, but particularly abundantin neurons (Fig. 2a). Pathogenic Aβ peptides are generated viacleavage of APP by BACE1 (β-site APP cleaving enzyme 1)20–24

and γ-secretase1,7 (Fig. 2b). The levels and distributions of APPand these pro-amyloidogenic cleavage enzymes in neurons, andin particular BACE1, are hypothesized to be the principal deter-minants of high levels of Aβ in the brain25. Formation of Aβ isprecluded by cleavage of APP within the Aβ domain by α-secretase or BACE2 (ref. 26; Fig. 2c).

A variety of APP mutations reported in cases of FAD are nearcleavage sites involved in formation of Aβ (Fig. 2d). The APP 717mutation is located near the C-terminus of Aβ and facilitates γ-secretase activity, leading to increased secretion of the longer andmore toxic Aβ peptide, Aβ42 (ref. 7; Fig. 2b). This longer Aβ42peptide is thought to promote the formation of Aβ aggregates andamyloid plaques. The APPswe mutation, a double mutation at theN-terminus of Aβ, enhances BACE1 cleavage and is associated withelevated levels of Aβ peptides, including Aβ42. In contrast, APPmutations within the Aβ peptide domain (for example, APP-E693Q, A692G or E693G) do not elevate the level of Aβ but maycause amyloidosis by increasing Aβ oligomer or protofibril for-mation27. Thus, a common feature of all FAD-linked APP muta-tions is the enhancement or facilitation of Aβ protofibril formation.

Another family of genes identified in FAD is the prese-nilins28–30. PS1 and PS2 encode highly homologous 43- to 50-kDmultipass transmembrane proteins that are processed to stableN-terminal and C-terminal fragments31, and are widely expressedat low abundance in the central nervous system. PS1 influencesAPP processing, but it is not clear whether PS1 itself acts as theprotease (γ-secretase), functions as a cofactor critical for the activ-ity of γ-secretase, or exerts its influence via trafficking of APP tothe proper compartment for γ -secretase cleavage32–36. What-ever the case, PS1 is now recognized as one of the critical elementof the γ-secretase complex. Nicastrain, a type I transmembraneglycoprotein, is another important component of this complex37.The PS1 gene has been reported to harbor more than 80 differentFAD mutations (AD mutation database, http://molgen-www.uia.ac.be), whereas only a small number of mutations havebeen found in PS2-linked families. The vast majority of abnor-

malities in PS genes are missense mutations thatresult in single amino acid substitutions, which ingeneral seem to influence γ-secretase activity andincrease the generation of the Aβ42 peptide.

Mouse models relevant to ADTo generate animal models of Aβ amyloidosis, manygroups have produced transgenic mice that expresswild-type APP, APP fragments, Aβ and FAD-linkedmutant APP and PS1. Some of the mutant APP mice,although they do not reproduce the full phenotypeof AD, represent excellent models of Aβ amyloidosis.

Expression of APP minigenes that encodeFAD-linked APP mutants (swe and 717), using sev-eral different promoters, leads to elevated levels of Aβ,diffuse Aβ deposits and plaques in the hippocampusand cortex of these mice38. The severity is influencedby the level of transgene expression and the specificmutation. In addition to Aβ deposits, these plaquescontain neurites (some showing hyperphosphorylat-ed tau immunoreactivity), astrocytes and microglia;

however, neurofibrillary tangles are not present. A variety of defectsare found in different mutant lines, including mild loss of neu-rons, learning deficits, problems in object recognition memory,and problems with alternation-spatial reference and working mem-ory39. Interestingly, synaptic abnormalities in hippocampal cir-cuits seem to precede the deposition of Aβ into plaques40,41.

Appropriate mouse models that display both amyloid plaquesand neurofibrillary tangles have not been entirely successful. In anattempt to obtain mice with both plaques and tangles, mutant APPtransgenic mice were mated to mice expressing the P301L taumutant42, a mutation linked to familial frontotemporal dementiawith parkinsonism (FTDP)2. Although these lines do have moretangles, mice bearing both mutant tau and APP are problematicas a faithful model of FAD because the FTDP mutation alone isassociated with increased tangles. Similarly, tau pathology can beinduced by introducing Aβ fibril into P301L tau mutant mice43.More appropriate models of AD might require co-expression ofmutant APP and all six isoforms of wild type human tau.

Mice expressing both mutant PS1 and mutant APP developaccelerated Aβ amyloidosis in the central nervous system. Coex-pression of the human A246E mutant PS1 and APPswe elevateslevels of Aβ in brain, and these mice develop numerous amyloiddeposits, dystrophic neuritis, and glial responses in the hip-pocampus and cortex44. Mutations known to be more malignantin the human disorder also produce accelerated Aβ depositionin mouse models. These mice also demonstrate that the key par-ticipants in Aβ amyloidosis (APP, PS1 and BACE1) are colocalizedin neurites immediately proximal to sites of Aβ formation inbrain, supporting the concept of a neuronal origin for Aβ.

In an effort to understand the functions of the Alzheimer’srelated genes, researchers have ablated most of them. Thisapproach is somewhat problematic with regard to APP becauseof the homologous amyloid precursor-like proteins, APLP1 andAPLP2. Homozygous APP–/– mice are viable and fertile, but seemto have subtle decreases in locomotor activity and forelimb gripstrength45. The absence of substantial phenotypes in APP–/– micemay be related to functional redundancy of APLP1 and APLP2.Consistent with this idea, APLP2–/– mice appear normal, but micewith either both APP and APLP2 targeted alleles or both APLP1and APLP2 null alleles show significant postnatal lethality46.

Similar approaches have focused on the proteins implicatedin β- and γ-secretase activities. Gene targeting of the presenilins

review

BOB CRIMI

Amyloid

Neurite

NeuriteTangles

Fig. 1. Pathological features of Alzheimer’s disease. Schematic diagram of a neuronshowing an extracellular amyloid plaque in the target field with a central core of Aβfibrils surrounded by dystrophic neurites, astrocytes and microglia. Within the cellbody, proximal dendrites and distal axons are intracellular protein aggregates, calledneurofibrillary tangles, which are paired helical filaments assembled from hyper-phosphorylated tau protein.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

is complicated because PS1 interacts functionally with Notch, areceptor protein involved in critical cell-fate decisions duringdevelopment47. PS1–/– mice do not survive beyond the early post-natal period and show severe perturbations in the axial skeleton,ribs and spinal ganglia—all defects in somitogenesis that resem-ble a partial Notch1 null phenotype48. However, it is clear thatPS1 is involved in γ-secretase activity; in cell culture, deletion ofPS1, or substitution of particular aspartate residues, leads toreduced levels of γ-secretase cleavage products and levels of Aβ36.PS2 null mice are viable and fertile, though they develop age-associated mild pulmonary fibrosis and hemorrhage. Mice het-erozygous for PS1 and lacking PS2 survive in relatively goodhealth, but mice lacking PS1 and heterozygous for PS2 die mid-way through gestation with full Notch1 null-like phenotype49.

To study the role of PS1 in vivo in adult mice, two groups gen-erated conditional PS1-targeted mice lacking PS1 expression inthe forebrain after embryonic development50,51. As expected, theabsence of PS1 resulted in decreased Aβ generation, further estab-lishing that PS1 is critical for γ-secretase activity in the brain. Thefinding that these forebrain-specific PS1 knockout mice do nothave significant morphological, physiological or overt behavioralabnormalities suggest that γ-secretase inhibitors may be usefulas therapeutic agents for Aβ amyloidosis.

BACE1 null mice are viable and healthy, have no obvious phe-notype or pathology, and can mate successfully52–54. Important-ly, in cortical neurons from BACE1 null embryos, there is nocleavage at the +1 and + 11 sites of Aβ52, and the secretion of Aβpeptides is abolished even in the presence of elevated level ofexogenous APP. Moreover, Aβ peptides are not produced inbrains of BACE1 null mice. These results establish that BACE1 isthe β-secretase required to cleave APP to generate the N-terminiof Aβ. Although behavioral studies on BACE1 null mice are nec-essary to determine the effect of the absence of BACE1, resultsthus far indicate that BACE1 is an excellent therapeutic target fordrug development for AD.

These mouse models have revealed that neurons are the majorsource for Aβ production in the brain. Because BACE1 is theprincipal β-secretase in neurons, and BACE2 may serve to limitthe secretion of Aβ peptides (Fig. 2), we hypothesized that therelative levels of BACE1 and BACE2 activities are major deter-minants of Aβ amyloidosis52. In this model, the secretion of Aβpeptides would be expected to be the highest in neurons andbrain as compared to other cell types or tissues because neurons

express high levels of BACE1 coupled with low expression ofBACE2. Seemingly inconsistent with this hypothesis is a studyshowing high levels of BACE1 mRNA expression in the pan-creas24. Given that APP is expressed in pancreas, why do AD anddiabetes mellitus not occur together? It now appears that someof the pancreatic BACE1 mRNAs are alternatively spliced to gen-erate a BACE1 isoform that is incapable of cleaving APP55. Takentogether with the observations that pancreas possesses low lev-els of BACE1 protein and activity20, these results are consistentwith the view that a high ratio of BACE1 to BACE2 activity leadsto selective vulnerability of neurons to Aβ amyloidosis, whereaspancreatic cells are spared.

Potential therapeutics for ADModel systems have great value for evaluating experimental treat-ments. Although they do not model the full phenotype of AD,these mutant transgenic mice represent excellent models of Aβamyloidois and are highly suitable for identification of thera-peutic targets. Although both β-and γ-secretase activities repre-sent therapeutic targets for the development of novel proteaseinhibitors for AD, the demonstration that BACE1 is the principalβ-secretase in cultured neurons52 and in brain53 provides anexcellent rationale for focusing on the design of novel therapeu-tics to inhibit BACE1 activity in brain. Importantly, in contrast toPS1–/– mice, the BACE1–/– mice seem to be normal52–54. Fur-thermore, BACE1-deficient neurons fail to secrete Aβ even whenco-expressing the APPswe and mutant PS1 genes, and these micedo not exhibit Aβ plaques in the brain (H.C., D.L.P. and P.C.W.,unpublished data).

Given the role of PS1 in γ-secretase activity56, development ofPS1 inhibitors is an important avenue of investigation for potentialtherapeutics. However, because of presenilins’ role in Notch pro-

review


Non-amyloidogenic cleavages

...EVKM DAEFRHDSGYEVHHQKLVFFAEDVGSNKGAIIGLMVGGVVIAT VIVITLVMLKKKQ...

α-secretase

α-secretase

α-CTF C83 19,20-CTF C80, C79

BACE2

BACE1

BACE1

BACE2

α-secre

tase

BACE2

γ-secretase

γ-secretase

p3 fragments

671

FAD-linked APP mutations


γ-secretase

Amyloidogenic cleavages


β-CTF C99

11-CTF C89

BACE1 BACE1

BACE1

γ-secretase

γ-secretase

Aβ1-40, 1-42Aβ11-40, 11-42

671

APP

K670N+M671L A692G D694N

E693KE693Q E693G

I716V

V717GV717F V717I

NH2

1 671 770

Aβ COOHcytoplasmic

Fig. 2. Amyloid precursor protein (APP) and secretase cleavages. (a) Schematic of APP, a type I transmembrane protein. The Aβ region isindicated in red. Transmembrane segment is demarcated by the gray lines.(b) Amyloidogenic cleavages. Aβ region indicated in (a) is expanded toshow the amino acid sequences of Aβ and flanking regions. The sequentialactions of β-site APP-cleaving enzyme (BACE1) and γ-secretase generateAβ1-40, 1-42 and Aβ 11-40, 11-42. Arrows indicate cleavage sites forBACE1 and γ-secretase. (c) Non-amyloidogenic cleavages. The sequentialactions of α-secretase or BACE2 and γ-secretase generate p3 fragments.Arrows indicate α-secretase, BACE2, and γ-secretase cleavage sites. Notethat α-secretase or BACE2 cleaves APP within the Aβ domain, precludingthe formation of amyloidogenic Aβ peptides. In non-neuronal cells, APP isprimarily processed by α-secretase or BACE2. (d) FAD-linked mutationsin APP. The Swedish mutation (K670N+M671L) increased the cleavageefficiency of BACE1, whereas mutations within the transmembranedomain promoted the cleavage by γ-secretase to increase Aβ42.Mutations within the Aβ domain seemed to enhance Aβ oligomer orprotofibril formation. Large arrow indicates cleavage by γ-secretase–likeprotease to generate APP intracellular domain.

a

b

c

d

Amy Center

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


cessing57, it may be valuable to try to design therapeutics that inhib-it the γ-secretase activity of PS1 selectively, without affecting theactivity involved in Notch1 processing58. This is important becauseseveral cell populations, hematopoetic stem cells in particular, useNotch signaling for cell-fate decisions even in the adult59.

Both Aβ immunization (with Freund’s adjuvant) and passivetransfer of Aβ antibodies reduce levels of Aβ and plaque burden inmutant APP transgenic mice in both prevention and treatmenttrials60,61. Efficacy seems to be related to antibody titer. The mech-anisms of enhanced clearance are not certain, but one possibilityis that a small amount of Aβ antibody reaches the brain, binds toAβ peptides, promotes the disassembly of fibrils, and via the Fcantibody domain, attracts activated microglia to remove Aβ60.Another possibility, not mutually exclusive, is that serum anti-bodies serve as a sink to draw the amyloid peptide from the braininto the circulation, thus changing the equilibrium of Aβ in dif-ferent compartments and promoting removal from the brain62.Immunization seems to attenuate learning and behavioral deficitsin at least two cohorts of mutant APP mice63,64. It is clear fromthe studies in mice that achieving adequate levels of antibody titeris critical because these levels predict amyloid clearance in mutanttransgenic mice. Unfortunately, recently initiated Phase 2 trialshave been suspended because of severe “inflammation” in theCNS of a subset of patients (n > 15). This is particularly disap-pointing, given that Phase 1 trials with Aβ and adjuvant were notassociated with any adverse events, and considering the successof immunotherapy in transgenic mice.

Amyotrophic lateral sclerosisALS manifests as muscle weakness and atrophy, along with spas-tic paralysis, which result from selective degeneration of spinaland corticospinal motor neurons, respectively. The disease affectsthe size, appearance and metabolism of these cells and evolves instages. First, neurofilamentous swellings occur in proximal axon-al segments accompanied by ubiquitin-positive aggregates65.Next, motor axons retract and become disconnected from thedenervated muscles. At this stage, trophic support is compro-mised, and cell bodies shrink and dendrites are attenuated. Neu-rons die in the final stages and exhibit several characteristics ofapoptosis66. Ultimately, the numbers of motor neurons in brain-stem nuclei and spinal cord are reduced, and large pyramidalneurons in motor cortex are lost. The clinical signs are most close-ly linked to the disconnection of the synaptic terminals of theseneuronal populations from their targets.

Approximately 10% of cases of ALS are familial, and, inmost of these cases, the disease is inherited in an autosomaldominant pattern67. Approximately 15–20% of patients withautosomal dominant FALS (∼ 2% of all ALS cases) have muta-tions in the gene that encodes cytosolic Cu/Zn superoxide dis-mutase (SOD1), an antioxidant enzyme that catalyzes theconversion of ·O–

2 to O2 and H2O2. To date, ∼ 90 different mis-sense mutations have been identified in the SOD1 gene67. Thesemutations are scattered throughout the protein and are notpreferentially localized near the active site or the dimer inter-face. Although some FALS SOD1 mutants show reduced enzy-matic activities, many retain full activity. The mutant enzymecauses selective neuronal degeneration through a gain of toxicproperty, consistent with autosomal dominant inheritance(reviewed in ref. 68). The presence of ubiquitin aggregates con-taining mutant SOD1 protein in affected neurons raises thepossibility that the defect is due to essential molecules beingsequestered away by the aggregates. Alternatively, the misfold-ed mutant SOD1 could catalyze aberrant reactions.

Other chromosomal loci linked to FALS show autosomaldominant, autosomal recessive or X-linked inheritance patterns67.A gene termed ALS2, on chromosome 2, is linked to juvenile ALSin several families. ALS2 encodes a protein termed alsin, whichshares homology to GTPase regulatory proteins (guanine-nucleotide exchange factors) that participate in critical cellularfunctions including signal transduction, regulation of thecytoskeleton and intracellular trafficking69,70. Mutations in ALS2are inherited as autosomal recessive and lead to a premature trun-cation, suggesting that the disease is associated with a loss of ALS2function. Thus, genetic approaches to ablate the gene encodingALS2 may provide useful mouse models of this rare form of ALS.

SOD1 mutant miceMice expressing a variety of mutant SOD1s (found in FALS)develop progressive weakness and muscle atrophy, and show theprototypical cellular stages of ALS71–73. The G37R SOD1 trans-genic mice provide an excellent illustration of these models. G37RSOD1, which retains full SOD activity, accumulates to 3–12 timesthe endogenous levels in the spinal cord, and the levels of themutant protein influence the age of onset. Toxic SOD1 is trans-ported anterograde in axons, and early on, it accumulates inaxons, where it is associated with structural pathology. Approx-imately 2–3 months before the appearance of clinical signs, SOD1accumulates in irregular, swollen, intraparenchymal portions ofmotor axons, and the axonal cytoskeleton and axonal transportare abnormal. Vacuoles, thought to represent degenerating mito-chondria, are present in enlarged axons and in dendritic swellings.(The latter is reminiscent of changes seen in excitotoxicity, whichis suggested to be involved in ALS74,75.) The cell bodies of someneurons show SOD1, ubiquitin and phosphorylated NF-Himmunoreactive inclusions76. Once Wallerian degeneration (char-acterized by degenerating myelin) of large axons is obvious, themice are usually weak. Eventually, the number of motor neuronsis reduced. These mouse models recapitulate the major clinicaland pathological hallmarks of ALS.

Potential therapies for ALSMutant SOD1 mice have been used to test pharmacological andgene-based therapies68,77–80. Several potential therapeutics havebeen tested with unencouraging results, including vitamin E andselenium, riluzole and gabapentin, and the copper chelator d-penicillamine. At present, treatment with creatine seems to havethe most robust pharmacological influence on the disease80; oraladministration of creatine to G93A SOD1 mice resulted in a dose-dependent improvement in motor tasks and extended survival.

The molecular mechanisms whereby mutant SOD1 causesselective motor neuron death have not yet been defined. One pro-posal is that the toxic property of mutant SOD1 involves mutation-induced conformational changes in SOD1 that result inaberrant oxidative activities81. In this scenario, cell dysfunctionand death could be initiated by aberrant oxidative chemistriescatalyzed by the copper atom bound in the active site of mutantSOD1 (ref. 81). To test this hypothesis, multiple lines of mutantSOD1 mice were crossed with mice lacking the specific copperchaperone protein (CCS) required for Cu loading of SOD1. Inac-tivation of the CCS gene in mice demonstrates that CCS isrequired for efficient copper incorporation into SOD1 in mam-mals, and the phenotypes of the CCS null mice resemble thoseof the SOD1 null mice82. Metabolic [64]Cu labeling studies inmutant SOD1 mice lacking CCS show that copper incorporationinto wild-type and mutant SOD1 is significantly diminished with-out the CCS83. Motor neurons in mice lacking the CCS have an

review©

2002

Nat

ure

Pu

blis

hin

g G

rou

p

htt

p:/

/neu

rosc

i.nat

ure

.co

m

increased rate of death after facial nerve axotomy, a response pre-viously shown for mice deficient in SOD1. Thus, CCS is neces-sary for efficient copper incorporation into SOD1 in motorneurons. However, although the absence of the CCS results in asignificant reduction in the level of copper-loaded mutant SOD1,it has no effect on the onset, progression or pathology of motorneuron disease in mutant SOD1 mice83. These results demon-strate that aberrant, Cu-dependent activity of mutant SOD1 isunlikely to be involved in the pathogenesis of FALS.

Although the molecular mechanisms underlying mutantSOD1-linked familial ALS remain unclear, several pathogenicmechanisms other than the copper hypothesis have been proposed.Mutant SOD1-containing aggregates participate in the pathogen-esis of SOD1-linked FALS84–86. The identification of mutant SOD1-containing aggregates early in the pathogenesis of mutant SOD1mice supports their importance in disease progression.

A pathological hallmark of ALS is the accumulation of neu-rofilaments in proximal axons and cell bodies of motor neurons.To test the role of neurofilaments in mutant SOD1–inducedmotor neuron disease, SOD1 mutant mice were crossbred to sev-eral lines of mice with altered distributions of neurofilaments.The progeny of SOD1 mutant mice crossed to mice expressingan NF-H-β-galactosidase fusion protein (NF-H-lacZ), whichcrosslinks neurofilaments and prevents their export to axons, hasno effect on disease progression87. In contrast, the lifespan ofSOD1 mutant mice was moderately increased in the absence ofneurofilaments when NF-L (neurofilament light chain) was ablat-ed88. However, crosses with mice over-expressing wild-type NF-L or NF-H (neurofilament heavy chain) resulted in sparing ofmotor neurons, attenuation of disease progression and increasedlife span79,89. Given the conflicting results, it remains unclear howthe distribution of neurofilaments influences motor neuron dis-ease induced by mutant SOD1 (ref. 68).

Spinal muscular atrophySMA is an autosomal recessive disease characterized by muscleweakness and atrophy in infants and children90. SMA is classi-fied as Type I, II or III based on the of age of onset and the degreeof functional disability. Infants with Type I SMA become weakbefore six months of age and die before two years. The incidenceof Type I SMA (also known as Werdnig–Hoffmann disease) isestimated at ∼ 1:10,000 live births, with a carrier rate frequencybetween 1:50 and 1:80. At the cellular level, motor neurons showchromatolysis and accumulations of phosphorylated neurofila-ments in cell bodies, and motor roots degenerate, leading to den-ervation of skeletal muscle. The final result is apoptosis and lossof large motor neurons90.

SMA Type I, II and III are linked to a single highly complexgenetic locus on chromosome 5. Investigators initially identi-fied two SMA candidate genes: survival motor neuron (SMN)91

and neuronal apoptosis inhibitory protein (NAIP)92. The regionof chromosome 5 containing these two genes is duplicated. Thecopy of the SMN gene closer to the telomere is termed SMN1,and the homologous copy closer to the centromere is termedSMN2. SMN is expressed from both SMN genes, however, andwhereas SMN1 produces full-length transcripts, SMN2 tran-scripts are alternatively spliced, resulting in mainly truncatedtranscripts lacking exon 7. SMN1 is now recognized as theSMA-determining gene, with NAIP possibly involved in modi-fying the severity of the disease. Mutations in SMN1 are pre-sent in nearly 100% of affected individuals90, and partial orcomplete deletions are detected in over 95% of cases. Homozy-gous deletion of SMN1 has not been found in unaffected indi-

viduals, and SMN2 has not been shown to be deleted in anyaffected individual. The level of full-length SMN protein inspinal motor neurons governs whether the individual has SMAType I, II or III (Type I having the least).

SMN is expressed in large motor neurons, but also in manyother cells, with the highest levels seen in the hippocampus andcerebellum. SMN is part of a multiprotein complex involved inbiogenesis of small nuclear ribonucleoprotein (snRNP), and isthought to be involved in the processing of small nuclear RNAs.In the cytoplasm, the SMN complex is associated with snRNPSm core proteins that are involved in the assembly of spliceosomalsnRNP complexes93. The complex represents a functional unitof the splicesomal machinery.

SMA-linked SMN mutants are defective in binding to Smproteins because mutant SMN cannot form the large oligomersthat are essential for high-affinity binding. Although abnor-malities of spliceosomal snRNP biogenesis and metabolism arethought to be involved in the pathogenesis of SMA, it is notclear how SMN deletions cause the abnormalities in motorneurons associated with SMA.

SMN knockout miceGiven the causative role of SMN loss in over 90% of SMA cases,deletion of SMN is an appropriate approach toward the devel-opment of a mouse model for the disease. SMN is highly con-served between mice (single copy) and humans (two copies,SMN1 and SMN2), with 82% identity94. However, SMN nullembryos do not survive past the peri-implantation stage, corre-sponding to the initiation of embryonic RNA transcription. Theseresults are consistent with the view that SMN function is essential,most likely due to its role in the biogenesis of spliceosomalsnRNPs and pre-mRNA splicing. Smn+/– heterozygous mice95

showed approximately 50% reduction of SMN protein in thespinal cord, which resulted in a progressive loss of motor neu-rons between birth and 6 months of age. The phenotype of thesemice resembles SMA type III.

To test whether human SMN2 can complement the embry-onic lethality of Smn–/– embryos, and to generate a mouse modelfor SMA, several groups generated transgenic mice expressinghuman SMN2 and crossbred them to Smn+/– mice to produceSMN2 transgenic mice lacking the endogenous mouse SMN96,97.These Smn–/–; SMN2 mice have abnormalities in the spinal cordand skeletal muscles similar to those seen in cases of SMA. Thus,SMN2 can partially compensate for the endogenous mouse SMN,and the variable phenotypes observed in Smn–/–; SMN2 micerecapitulate those seen in SMA Type I, II or III. The level of full-length SMN protein in these Smn–/–; SMN2 mice correlates withthe severity of the disease. Smn–/–; SMN2 mice exhibit severalclinical characteristics: Type I mice do not develop fur and dieby postnatal day 10; Type II mice are inactive and die between 2to 4 weeks of age; and Type III mice survive and breed normallybut have defects in tail size. These studies strongly support theidea that the level of intact SMN protein determines the severityof phenotypes in SMA.

These mouse models of SMA are useful in understanding dis-ease mechanisms and for testing therapeutic strategies. It is clearfrom the human and mouse pathology that the level of full-lengthSMN protein correlates with the severity of the disease. Sodiumbutyrate is effective in elevating the level of SMN protein (fromthe SMN2 gene) in lymphoid cell lines derived from SMApatients98. This compound was thus used to test for therapeuticeffects in mouse models of SMA98. Treatment of SMA-like micewith sodium butyrate led to an increase in the level of SMN pro-

review


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


tein in spinal cord and a concomitant amelioration of diseasephenotype98. These studies indicate that sodium butyrate maybe an effective therapeutic for SMA.

CONCLUSIONSThe identification of genes mutated or deleted in the inheritedforms of neurodegenerative diseases has allowed investigators tocreate in-vivo and in-vitro model systems relevant to a wide vari-ety of human neurological disorders. Genetically engineered micethat recapitulate some of the features of human diseases can pro-vide important new information about the neurobiology of thesediseases. These new models allow investigators to examine themolecular mechanisms by which mutant proteins cause selectivedysfunction and death of neurons. Moreover, they can be used totest pathogenic pathways by crossing these mice with either mutat-ed or deleted alleles of other molecular players in the pathogenicprocess. The results of these approaches provide us with a betterunderstanding of the pathogenic mechanisms of the diseases, andshould lead to the design of new therapeutic strategies.

In this review, we emphasized the value of transgenic and genetargeted models and the lessons they provided for understand-ing these and other neurodegenerative diseases. Specifically,BACE1–/– mice and APP/PS1 transgenic models have providedextraordinary new insights into the mechanisms of amyloidoge-nesis and the reasons why AD is a brain amyloidosis. For ALS,SOD1 transgenic models lacking CCS demonstrate that aberrantactivities dependent on copper-loaded SOD1 are unlikely to bethe pathogenic mechanism.

The mouse models have not only provided substantialprogress in understanding the molecular mechanisms of neu-rodegeneration, they are also instrumental in identifying targetsfor mechanism-based therapeutics, such as BACE1. These genet-ically engineered models are valuable for testing a variety of ther-apeutic approaches, including Aβ immunotherapy in AD,creatine in ALS, and sodium butyrate in SMA. In summary,investigation of the pathogeneses of neurodegenerative diseasesusing transgenic models and other approaches has made spec-tacular progress over the past few years, and we anticipate thatmore promising therapies based on our present understandingof the disease mechanisms will continue to be identified. Trans-genic models should provide a highly useful tool for quicklyassessing which therapies should be pursued.

AcknowledgmentsThe authors thank colleagues from JHMI, particularly S. Sisodia, M. Lee,

G. Thinakaren, E. Koo, J. Subramaniam, L. Martin, V. Koliatsos, A. Bergin,

L. Brujin, C. Pardo, B. Rabin, T. Crawford, M. Becher, P. Hoffman, J. Griffin,

J. Rothstein, J. Troncoso, T. Li, V. Culotta and D. Cleveland as well as those at

other institutions (J. Gitlin) for contributions to the original work cited in this

review and for discussions. Supported by grants from the U. S. Public Health

Service (AG05146, AG07914, AG10480, AG10491, AG14248, NS07435,

NS20471, NS37145, NS10580, NS37771, NS40014, NS38377, NS38065) as well

as the Metropolitan Life Foundation, Adler Foundation, and Bristol-Myers

Squibb Foundation.

RECEIVED 14 JANUARY; ACCEPTED 30 MAY 2002

1. Price, D. L., Tanzi, R. E., Borchelt, D. R. & Sisodia, S. S. Alzheimer’s disease:genetic studies and transgenic models. Annu. Rev. Genet. 32, 461–493(1998).

2. Lee, V. M. Y. & Trojanowski, J. Q. Neurodegenerative taupathies: humandisease and transgenic mouse models. Neuron 24, 507–510 (1999).

3. Lin, X., Cummings, C. J. & Zoghbi, H. Y. Expanding our understanding ofpolyglutamine diseases through mouse models. Neuron 24, 499–502(1999).

4. Dunnett, S. B. & Bjorklund, A. Prospects for new restorative andneuroprotective treatments in Parkinson’s Disease. Nature 399, A32–A39(1999).

5. Yamamoto, A., Lucas, J. J. & Hen, R. Reversal of neuropathology and motordysfunction in a conditional model of Huntington’s disease. Cell 101, 57–66(2000).

6. Hardy, J. & Gwinn-Hardy, K. Genetic classification of primaryneurodegenerative disease. Science 282, 1075–1079 (1998).

7. Selkoe, D. J. Alzheimer’s disease: genes, proteins and therapy. Physiol. Rev.81, 741–766 (2001).

8. Price, D. L., Sisodia, S. S. & Borchelt, D. R. Genetic neurodegenerativediseases: the human illness and transgenic models. Science 282, 1079–1083(1998).

9. Schilling, G. et al. Intranuclear inclusions and neuritic pathology intransgenic mice expressing a mutant N-terminal fragment of huntingtin.Hum. Mol. Genet. 8, 397–407 (1999).

10. Goedert, M., Spillantini, M. G. & Davies, S. W. Filamentous nerve cellinclusions in neurodegenerative diseases. Curr. Opin. Neurobiol. 8, 619–632(1998).

11. Zoghbi, H. Y. & Orr, H. T. Polyglutamine diseases: protein cleavage andaggregation. Curr. Opin. Neurobiol. 9, 566–570 (1999).

12. Albert, M. S. & Drachman, D. A. Alzheimer’s disease. What is it, how manypeople have it, and why do we need to know? Neurology 55, 166–168 (2000).

13. Mesulam, M. M. Neuroplasticity failure in Alzheimer’s disease: bridging thegape between plaques and tangles. Neuron 24, 521–529 (1999).

14. Price, D. L. & Sisodia, S. S. Mutant genes in familial Alzheimer’s disease andtransgenic models. Annu. Rev. Neurosci. 21, 479–505 (1998).

15. Beach, T. G. et al. The cholinergic deficit coincides with Aβ deposition atthe earliest histopathologic stages of Alzheimer disease. J. Neuropathol. Exp.Neurol. 59, 308–313 (2000).

16. Serpell, L. C., Blake, C. C. F. & Fraser, P. E. Molecular structure of a fibrillarAlzheimer’s Aβ fragment. Biochemistry 39, 13269–13275 (2000).

17. Lansbury, P. T. Jr. Evolution of amyloid: what normal protein folding maytell us about fibrillogenesis and disease. Proc. Natl. Acad. Sci. USA 96,3342–3344 (1999).

18. Tanzi, R. E. New frontiers in Alzheimer’s disease genetics. Neuron 32,181–184 (2001).

19. Selkoe, D. J. Clearing the brain’s amyloid cobwebs. Neuron 32, 177–180(2001).

20. Sinha, S. et al. Purification and cloning of amyloid precursor protein β-secretase from human brain. Nature 402, 537–540 (1999).

21. Hussain, I. et al. Identification of a novel aspartic protease (Asp 2) as β-secretase. Mol. Cell. Neurosci. 14, 419–427 (1999).

22. Lin, X. et al. Human aspartic protease memapsin 2 cleaves the β-secretasesite of β-amyloid precursor protein. Proc. Natl. Acad. Sci. USA 97,1456–1460 (2000).

23. Yan, R. et al. Membrane-anchored aspartyl protease with Alzheimer’sdisease β-secretase activity. Nature 402, 533–537 (1999).

24. Vassar, R. et al. β-secretase cleavage of Alzheimer’s amyloid precusorprotein by the transmembrane aspartic protease BACE. Science 286,735–741 (1999).

25. Wong, P. C., Price, D. L. & Cai, H. The brain’s susceptibility to amyloidplaques. Science 293, 1434–1435 (2001).

26. Farzan, M., Schnitzler, C. E., Vasilieva, N., Leung, D. & Choe, H. BACE2, aβ-secretase homolog, cleaves at the β site and within the amyloid-β regionof the amyloid-β precursor protein. Proc. Natl. Acad. Sci. USA 97,9712–9717 (2000).

27. Nilsberth, C. et al. The ‘Arctic’ APP mutation (E693G) causes Alzheimer’sdisease by enhanced Aβ protofibril formation. Nat. Neurosci. 4, 887–893(2001).

28. Sherrington, R. et al. Cloning of a gene bearing missense mutations inearly-onset familial Alzheimer’s disease. Nature 375, 754–760 (1995).

29. Levy-Lahad, E. et al. Candidate gene for the chromosome 1 familialAlzheimer’s disease locus. Science 269, 973–977 (1995).

30. Rogaev, E. I. et al. Familial Alzheimer’s disease in kindreds with missensemutations in a gene on chromosome 1 related to the Alzheimer’s diseasetype 3 gene. Nature 376, 775–778 (1995).

31. Thinakaran, G. et al. Endoproteolysis of presenilin 1 and accumulation ofprocessed derivatives in vivo. Neuron 17, 181–190 (1996).

32. Sisodia, S. S. & George-Hyslop, P. H. gamma-secretase, Notch, Aβ andAlzheimer’s disease: where do the presenilins fit in? Nat. Rev. Neurosci. 3,281–290 (2002).

33. Esler, W. P. & Wolfe, M. S. A portrait of Alzheimer secretases—new featuresand familiar faces. Science 293, 1449–1454 (2001).

34. Vassar, R. & Citron, M. Aβ-generating enzymes: recent advances in β- andγ-secretase research. Neuron 27, 419–422 (2000).

35. De Strooper, B. et al. Deficiency of presenilin-1 inhibits the normal cleavageof amyloid precursor protein. Nature 391, 387–390 (1998).

36. Wolfe, M. S. et al. Two transmembrane aspartates in presenilin-1 requiredfor presenilin endoproteolysis and secretase activity. Nature 398, 513–517(1999).

37. Kopan, R. & Goate, A. Aph-2/Nicastrin: an essential component of gamma-secretase and regulator of notch signaling and presenilin localization.Neuron 33, 321–324 (2002).

38. Calhoun, M. E. et al. Neuronal overexpression of mutant amyloid precursor

review©

2002

Nat

ure

Pu

blis

hin

g G

rou

p

htt

p:/

/neu

rosc

i.nat

ure

.co

m

protein results in prominent deposition of cerebrovascular amyloid. Proc.Natl. Acad. Sci. USA 96, 14088–14093 (1999).

39. Chen, G. et al. A learning deficit related to age and β-amyloid plaques in amouse model of Alzheimer’s disease. Nature 408, 975–979 (2000).

40. Mucke, L. et al. High-level neuronal expression of Aβ1-42 in wild-typehuman amyloid protein precursor transgenic mice: synaptotoxicity withoutplaque formation. J. Neurosci. 20, 4050–4058 (2000).

41. Hsia, A. Y. et al. Plaque-independent disruption of neural circuits inAlzheimer’s disease mouse models. Proc. Natl. Acad. Sci. USA 96,3228–3233 (1999).

42. Lewis, J. et al. Enhanced neurofibrillary degeneration in transgenic miceexpressing mutant tau and APP. Science 293, 1487–1491 (2001).

43. Götz, J., Chen, F., Barmettler, R. & Nitsch, R. M. Tau filament formation intransgenic mice expressing P301L tau. J. Biol. Chem. 276, 529–534 (2001).

44. Borchelt, D. R. et al. Accelerated amyloid deposition in the brains oftransgenic mice co-expressing mutant presenilin 1 and amyloid precursorproteins. Neuron 19, 939–945 (1997).

45. Zheng, H. et al. β-amyloid precursor protein-deficient mice show reactivegliosis and decreased locomotor activity. Cell 81, 525–531 (1995).

46. Heber, S. et al. Mice with combined gene knock-outs reveal essential andpartially redundant functions of amyloid precursor protein familymembers. J. Neurosci. 20, 7951–7963 (2000).

47. Struhl, G. & Greenwald, I. Presenilin is required for activity and nuclearaccess of notch in drosophila. Nature 398, 522–525 (1999).

48. Wong, P. C. et al. Presenilin 1 is required for Notch1 and Dll1 expression inthe paraxial mesoderm. Nature 387, 288–292 (1997).

49. Donoviel, D. et al. Mice lacking both presenilin genes exhibit earlyembryonic patterning defects. Genes Dev. 13, 2801–2810 (1999).

50. Yu, H. et al. APP processing and synaptic plasticity in presenilin-1conditional knockout mice. Neuron 31, 713–726 (2001).

51. Feng, R. et al. Deficient neurogenesis in forebrain-specific presenilin-1knockout mice is associated with reduced clearance of hippocampalmemory traces. Neuron 32, 911–926 (2001).

52. Cai, H. et al. BACE1 is the major β-secretase for generation of Aβ peptidesby neurons. Nat. Neurosci. 4, 233–234 (2001).

53. Luo, Y. et al. Mice deficient in BACE1, the Alzheimer’s β-secretase, havenormal phenotype and abolished β-amyloid generation. Nature 4, 231–232(2001).

54. Roberds, S. L. et al. BACE knockout mice are healthy despite lacking theprimary β-secretase activity in brain: implication for Alzheimer’s diseasetherapeutics. Hum. Mol. Genet. 10, 1317–1324 (2001).

55. Bodendorf, U., Fischer, F., Bodian, D., Multhaup, G. & Paganetti, P. A splicevariant of β-secretase deficient in the amyloidogenic processing of theamyloid precursor protein. J. Biol. Chem. 276, 12019–12023 (2001).

56. Li, Y. M. et al. Photoactivated γ-secretase inhibitors directed to the activesite convalently label presenilin 1. Nature 405, 689–693 (2000).

57. Huppert, S. S. et al. Embryonic lethality in mice homozygous for aprocessing-deficient allele of Notch 1. Nature 405, 966–970 (2000).

58. Petit, A. et al. New protease inhibitors prevent γ-secretase-mediatedproduction of Aβ40/42 without affecting Notch clevage. Nat. Cell Biol. 3,507–511 (2001).

59. Hadland, B. K. et al. γ -secretase inhibitors repress thymocyte development.Proc. Natl. Acad. Sci. USA 98, 7487–7491 (2001).

60. Bard, F. et al. Peripherally administered antibodies against amyloid β-peptide enter the central nervous system and reduce pathology in a mousemodel of Alzheimer’s disease. Nat. Med. 6, 916–919 (2000).

61. Schenk, D. et al. Immunization with amyloid attenuates Alzheimer-disease-like pathology in the PDAPP mouse. Nature 400, 173–177 (1999).

62. DeMattos, R. B. et al. Peripheral anti-Aβ antibody alters CNS and plasmaAβ clearance and decreases brain Aβ burden in a mouse model ofAlzheimer’s disease. Proc. Natl. Acad. Sci. USA 98, 8850–8855 (2001).

63. Janus, C. et al. Aβ peptide immunization reduces behavioural impairmentand plaques in a model of Alzheimer’s disease. Nature 408, 979–982 (2000).

64. Morgan, D. et al. A β peptide vaccination prevents memory loss in ananimal model of Alzheimer’s disease. Nature 408, 982–985 (2000).

65. Ince, P. G. in Amyotrophic Lateral Sclerosis (eds. Brown, R. H. Jr., Meininger,V. & Swash, M.) 83–112 (Martin Dunitz, London, 2000).

66. Martin, L. J. Neuronal death in amyotrophic lateral sclerosis is apoptosis:possible contribution of a programmed cell death mechanism. J. Neuropathol. Exp. Neurol. 58, 459–471 (1999).

67. Andersen, P. M., Morita, M. & Brown, R. H. Jr. in Amyotrophic LateralSclerosis (eds. Brown, R. H. Jr., Meininger, V. & Swash, M.) 223–250 (MartinDunitz, London, 2000).

68. Cleveland, D. W. & Rothstein, J. D. From Charcot to Lou Gehrig:deciphering selective motor neurons death in ALS. Nature 2, 806–819(2001).

69. Hadano, S. et al. A gene encoding a putative GTPase regulator is mutated infamilial amyotrophic lateral sclerosis 2. Nat. Genet. 29, 166–173 (2001).

70. Yang, Y. et al. The gene encoding alsin, a protein with three guanine-nucleotide exchange factor domains, is mutated in a form of recessiveamyotrophic lateral sclerosis. Nat. Genet. 29, 160–165 (2001).

71. Gurney, M. E. et al. Motor neuron degeneration in mice that express ahuman Cu, Zn superoxide dismutase mutation. Science 264, 1772–1775(1994).

72. Wong, P. C. et al. An adverse property of a familial ALS-linked SOD1mutation causes motor neuron disease characterized by vacuolardegeneration of mitochondria. Neuron 14, 1105–1116 (1995).

73. Bruijn, L. I. et al. ALS-linked SOD1 mutant G85R mediates damage toastrocytes and promotes rapidly progressive disease with SOD1-containinginclusions. Neuron 18, 327–338 (1997).

74. Jackson, M. & Rothstein, J. D. in Amyotrophic Lateral Sclerosis (eds. Brown,R. H. Jr., Meininger, V. & Swash, M.) 263–278 (Martin Dunitz, London,2000).

75. Shaw, P. J. in Amyotrophic Lateral Sclerosis (eds. Brown, R. H. Jr., Meininger,V. & Swash, M.) 113–144 (Martin Dunitz, London, 2000).

76. Williamson, T. L. & Cleveland, D. W. Slowing of axonal transport is a veryearly event in the toxicity of ALS-linked SOD1 mutants to motor neurons.Nat. Neurosci. 2, 50–56 (1999).

77. Julien, J.-P. Amyotrophic lateral sclerosis: unfolding the toxicity of themisfolded. Cell 104, 581–591 (2001).

78. Pasinelli, P., Borchelt, D. R., Houseweart, M. K., Cleveland, D. W. & Brown,R. H. Caspase-1 is activated in neural cells and tissue with amyotrophiclateral sclerosis-associated mutations in copper-zinc superoxide dismutase.Neurobiology 95, 15763–15768 (1998).

79. Couillard-Després, S. et al. Protective effect of neurofilament NF-Hoverexpression in motor neuron disease induced by mutant superoxidedismutase. Proc. Natl. Acad. Sci. USA 95, 9626–9630 (1998).

80. Klivenyi, P. et al. Neuroprotective effects of creatine in a transgenic animalmodel of amyotrophic lateral sclerosis. Nat. Med. 5, 347–350 (1999).

81. Estévez, A. G. et al. Induction of nitric oxide-dependent apoptosis in motorneurons by zinc-deficient superoxide dismutase. Science 286, 2498–2500(1999).

82. Wong, P. C. et al. Copper chaperone for superoxide dismutase is essential toactivate mammalian Cu/Zn superoxide dismutase. Proc. Natl. Acad. Sci.USA 97, 2886–2891 (2000).

83. Subramaniam, J. R. et al. Mutant SOD1 causes motor neuron diseaseindependent of copper chaperone-mediated copper loading. Nat. Neurosci.5, 301–307 (2002).

84. Johnston, J. A., Dalton, M. J., Gurney, M. E. & Kopito, R. R. Formation ofhigh molecular weight complexes of mutant Cu, Zn-superoxide dismutasein a mouse model for familial amyotrophic lateral sclerosis. Proc. Natl.Acad. Sci. USA 97, 12571–12576 (2000).

85. Bruijn, L. I. et al. Aggregation and motor neuron toxicity of an ALS-linkedSOD1 mutant independent from wild-type SOD1. Science 281, 1851–1854(1998).

86. Beaulieu, J. M., Jacomy, H. & Julien, J. P. Formation of intermediatefilament protein aggregates with disparate effects in two transgenic mousemodels lacking the neurofilament light subunit. J. Neurosci. 20, 5321–5328(2000).

87. Eyer, J., Cleveland, D. W., Wong, P.C. & Peterson, A. C. Pathogenesis of twoaxonopathies does not require axonal neurofilaments. Nature 391, 584–587(1998).

88. Williamson, T. L. et al. Absence of neurofilaments reduces the selectivevulnerability of motor neurons and slows disease caused by a familialamyotrophic lateral sclerosis-linked superoxide dismutase 1 mutant. Proc.Natl. Acad. Sci. USA 95, 9631–9636 (1998).

89. Kong, J. & Xu, Z. Overexpression of neurofilament subunit NF-L and NF-Hextends survival of a mouse model for amyotrophic lateral sclerosis.Neurosci. Lett. 281, 72–74 (2000).

90. Crawford, T. & Pardo, C. A. The neurobiology of childhood spinal muscularatrophy. Neurobiol. Dis. 3, 97–110 (1996).

91. Lefebvre, S. et al. Identification and characterization of a spinal muscularatrophy-determining gene. Cell 80, 155–165 (1995).

92. Roy, N. et al. The gene for neuronal apoptosis inhibitory protein is partiallydeleted in individuals with spinal muscular atrophy. Cell 80, 167–178(1995).

93. Fischer, U., Liu, Q. & Dreyfuss, G. The SMN-SIP1 complex has an essentialrole in spliceosomal snRNP biogenesis. Cell 90, 1023–1029 (1997).

94. Pagliardini, S. et al. Subcellular localization and axonal transport of thesurvival motor neuron (SMN) protein in the developing rat spinal cord.Hum. Mol. Genet. 9, 47–56 (2000).

95. Jablonka, S., Schrank, B., Kralewski, M., Rossoll, W. & Sendtner, M.Reduced survival motor neuron (Smn) gene dose in mice leads to motorneuron degeneration: an animal model for spinal muscular atrophy typeIII. Hum. Mol. Genet. 9, 341–346 (2000).

96. Hsieh-Li, H. M. et al. A mouse model for spinal muscular atrophy. Nat.Genet. 24, 66–70 (2000).

97. Monani, U. R. et al. The human centromeric survival motor neuron gene(SMN2) rescues embryonic lethality in Smn–/– mice and results in a mousewith spinal muscular atrophy. Hum. Mol. Genet. 9, 333–339 (2000).

98. Chang, J.-G. et al. Treatment of spinal muscular atrphy by sodium butyrate.Proc. Natl. Acad. Sci. USA 98, 9808–9813 (2001).

review


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

articles


A major neural system in drug reinforcement involves mesocor-ticolimbic pathways, whereby dopaminergic ventral tegmentalarea (VTA) inputs regulate corticoaccumbens glutamatergic (Glu)systems1. The NMDA receptor, a subtype of glutamate receptor,has been implicated as a critical component of the neuroadap-tive processes underlying addiction to abused drugs. Indeed,processes similar to NMDA receptor–dependent synaptic plas-ticity have been observed to be functionally linked with cocainereinforcement. Trafficking of (±)-α-amino-3-hydroxy-5-methylisoxazole-4-propionic acid (AMPA) receptors to the post-synaptic density, a physiological process that underlies synapticplasticity in hippocampal pyramidal neurons, occurs on VTAneurons following a single dose of cocaine2. Additionally, exci-tatory drive involving NMDA receptor activation from limbicstructures that target the VTA reinitiates cocaine responding fol-lowing extinction3. These findings suggest that NMDA receptoractivation with concomitant glutamatergic synaptic potentiationmay be necessary for induction of drug reinforcement4.

Accordingly, the NMDA receptor seems to be a mediator ofmany neuroadaptive processes underlying ethanol addiction, espe-cially in relation to excessive ethanol intake or ‘binge drinking’5. Atconcentrations corresponding to blood alcohol levels commonlyassociated with a significant level of intoxication in humans (20–50 mM or about 100–250 mg %, whereas legal intoxication inmost states is 80 mg %), ethanol (EtOH) inhibits the activationof native NMDA receptors in many brain structures and neuronal

DARPP-32 and regulation of theethanol sensitivity of NMDAreceptors in the nucleus accumbens

R. E. Maldve1,2, T. A. Zhang1,2, K. Ferrani-Kile2, S. S. Schreiber1,2, M. J. Lippmann1,2, G. L. Snyder3, A. A. Fienberg3, S. W. Leslie1,2, R. A. Gonzales1,2 and R. A. Morrisett1,2

1 The Waggoner Center for Alcohol and Addiction Research, 2500 Speedway, The University of Texas at Austin, Austin, Texas 78712-1074, USA2 The College of Pharmacy, The University of Texas at Austin, Austin, Texas 78712-1074, USA3 The Laboratory of Molecular and Cellular Neuroscience, The Rockefeller University, New York, New York 10021, USA

Correspondence should be addressed to R.M. ([email protected])


The medium spiny neurons of the nucleus accumbens receive both an excitatory glutamatergicinput from forebrain and a dopaminergic input from the ventral tegmental area. This integrationpoint may constitute a locus whereby the N-methyl-D-aspartate (NMDA)-subtype of glutamatereceptors promotes drug reinforcement. Here we investigate how dopaminergic inputs alter theethanol sensitivity of NMDA receptors in rats and mice and report that previous dopamine recep-tor-1 (D1) activation, culminating in dopamine and cAMP-regulated phosphoprotein-32 kD(DARPP-32) and NMDA receptor subunit-1 (NR1)-NMDA receptor phosphorylation, stronglydecreases ethanol inhibition of NMDA responses. The regulation of ethanol sensitivity of NMDAreceptors by D1 receptors was absent in DARPP-32 knockout mice. We propose that DARPP-32mediated blunting of the response to ethanol subsequent to activation of ventral tegmental areadopaminergic neurons initiates molecular alterations that influence synaptic plasticity in this cir-cuit, thereby promoting the development of ethanol reinforcement.

preparations6,7. NMDA receptor currents and long-term poten-tiation (LTP) induction are both sensitive to ethanol concentra-tions achieved during binge drinking8. Long-term exposure tosuch concentrations of ethanol enhances NMDA receptor func-tion and induces epileptiform activity and neurotoxicity followingwithdrawal from chronic exposure9,10. Drug discrimination stud-ies indicate that strong NMDA receptor inhibition supplants forethanol responding11. Thus, evidence indicates the NMDA recep-tor is a target of ethanol action, especially in relation to neuroad-aptive responses seen in binge-level intake and underlies a varietyof outcomes in chronic abusers including ethanol dependence,amnesia (blackouts), withdrawal seizures and neurotoxicity.

However, a paradox exists concerning the involvement ofNMDA receptors in ethanol reinforcement. If these receptors pro-mote reinforcement, as for other drugs of abuse, then ethanol inhi-bition of NMDA receptors should inhibit addiction to this drug.Instead, as described above, there is sufficient evidence to suggestthat NMDA receptors mediate, at least in part, some componentsof ethanol reinforcement. Thus, we reasoned that some mecha-nism must exist that attenuates ethanol sensitivity of NMDA recep-tors in circuits involved in ethanol reinforcement. Coincidently,the phosphoprotein DARPP-32 has been identified as an intracel-lular regulator of ethanol reinforcement12. DARPP-32 is a com-ponent of a postsynaptic signaling system in the medium spinytarget neurons of the nucleus accumbens (NAcc) and communi-cates signals in response to dopamine D1 receptor activation13.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


articles

and picrotoxin (25 µM). NMDA EPSCs were also slightlyenhanced by reduction of extracellular Mg2+ to 0.5 mM8,10.NMDA EPSCs from control slices of NAcc, elicited by lowamplitude stimulation of local afferents, were inhibited bybath application of ethanol (Fig. 1). The peak amplitude ofthe NMDA EPSCs was reduced by over 50% in 75 mMethanol (Fig. 1a) and responses returned to baseline levelsupon washout. However, medium spiny neurons in NAcc

slices pretreated for 15 minutes with the D1 receptor agonist, SKF38393 (25 µM), before an identical exposure to ethanol (75 mM),displayed a substantial reduction in the sensitivity to ethanol incomparison with that observed in control accumbal slices (Fig.1b). The time course of the responses to ethanol in this pair ofneurons depicted was identical (Fig. 1c) and returned to near-baseline NMDA EPSC levels after ethanol washout. Finally, thecumulative data from the population of neurons tested in thisexperiment indicate that under conditions of D1 receptor acti-vation, the ethanol sensitivity of pharmacologically isolatedNMDA EPSCs is reduced by about half that seen in native,untreated preparations (Fig. 1d).

To more definitively determine the synaptic site of D1 recep-tor modulation of the ethanol inhibition of the NMDA EPSC,we measured whole-cell currents elicited in medium spiny neu-rons after 3–7 days in culture in perforated patch mode to pre-vent intracellular perfusion from the recording electrode;whole-cell currents were elicited by direct NMDA application(INMDA) via pressure ejection onto the soma of medium spinyneurons at a holding potential of –60 mV (Fig. 1e). DirectNMDA agonist application also revealed a suppression ofethanol sensitivity of NMDA responses in medium spiny neu-rons pretreated with SKF 38393 (25 µM for 15 min, Fig. 1e andf). These data were significant (P < 0.015) , in agreement withobservations with NMDA EPSCs and strongly suggest thatpostsynaptic D1 receptor–mediated processes regulate theethanol sensitivity of NMDA receptors.

Fig. 1. Patch clamp analysis reveals D1 receptor modulation of theEtOH sensitivity of accumbal-medium spiny NMDA synaptic cur-rents. NMDA synaptic currents (average of 5 traces) elicited by localstimulation under control conditions (a) and 20 min following bathapplication of the dopamine D1 receptor agonist, SKF 38393 (25 µM, b). In each case, the current elicited in baseline conditionand in the presence of ethanol (75 mM) are indicated. Stimulus arti-facts are truncated for clarity. (c) The time course of the inhibitionof the NMDA synaptic current by ethanol of the cells shown in (a)and (b), comparing ethanol inhibition in the absence and presence ofthe D1 agonist; washout of ethanol is also depicted in this timecourse plot. (d) The cumulative data of ethanol inhibition at 15–20 min after application in control cells versus those pretreated withSKF 38393 (25 µM) for all cells tested. The asterisk indicates thatthese groups differed significantly (P < 0.025, Student’s t-test com-parison of 5 averaged responses in each condition from 4 cells ineither group). (e) D1 receptor modulation of ethanol sensitivity ofNMDA whole-cell currents, elicited by direct agonist application tocultured medium spiny neurons (3–7 days in vitro). The top pair oftraces show NMDA-activated currents in the absence and presenceof ethanol in a typical control cell, whereas typical ethanol inhibitionin cells pretreated with the D1 agonist (SKF 38393, 25 µM) is shownin the bottom pair of traces. (f) The cumulative data of ethanol inhi-bition at 15–20 min post-application in control cells versus thosepretreated with SKF 38393 (25 µM) for all cells tested. The asteriskindicates that these groups differed significantly (P < 0.015, Student’st-test comparison of 5 averaged responses in each condition from 4 control and 5 SKF 38393-pretreated neurons).

D1 receptors stimulate the phosphorylation of DARPP-32 on Thr-34 via adenylyl cyclase and protein kinase A, thereby convertingDARPP-32 to an inhibitor of protein phosphatase-1 (PP-1). Thisdecreases endogenous phosphatase activity and enhances phos-phorylation of one of the major subunits of the NMDA receptor,NR1 (refs. 14 and 15). Mice lacking DARPP-32 display no dose-dependence for ethanol responding, as they do not increase theirintake when ethanol concentrations are increased, especially athigher levels (> 10% v/v)12. In contrast, wild-type mice increasetheir intake linearly as increasing concentrations are made avail-able. These data suggest that the reinforcement process that occursduring long-term access to increasing concentrations of ethanol isregulated by the activity of DARPP-32.

Here we report that dopamine D1-receptor activation dimin-ishes the ethanol sensitivity of NMDA receptors via the proteinkinase A/DARPP-32 cascade in principal neurons of the NAcc,and we propose a molecular mechanism that contributes toethanol reinforcement.

RESULTSSKF 38393 reduces EtOH inhibition of NMDA receptorsWe investigated regulation by the D1 receptor of the ethanol sen-sitivity (see Methods) of NMDA receptors in medium spiny neu-rons of the NAcc with whole-cell slice patch-clamp recordings.We pharmacologically isolated NMDA excitatory postsynapticcurrents (EPSCs) by blocking non-NMDA and GABA receptorswith 6,7-dinitroquinoxaline-2,3(1H,4H)-dione (DNQX; 10 µM)

SKF 38393 pre-treatment

20pA

20 ms

NMDA EPSC (DNQX/PTX)Control

Baseline

EtOH (75 mM)

Baseline

EtOH (75 mM)

0

10

20

30

40

50

60

Control SKF 38393

*

Control

50pA

80 ms

Baseline

EtOH (75 mM)SKF 38393 pre-treatment

Baseline

EtOH (75 mM)INMDA

0 10 20 30 400

20

40

60

80

100

120 +EtOH

I EP

SC

(% b

asel

ine)

min

SKF38393 Control

EtO

H in

hib

itio

n o

f I

(

%)

NM

DA

0

10

20

30

40

50

60

Control SKF 38393

*

EtO

H in

hib

itio

n o

f N

MD

A E

PS

C (

%)

a b

c d

e f

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

articles


Fig. 2. D1 receptors activate a protein kinase A–dependentprocess that regulates NMDA receptor sensitivity to ethanol in thenucleus accumbens. A field potential recording protocol usingdirect activation of postsynaptic NMDA receptors at 2-min inter-vals was used. (a) The amplitude of individual field potentialresponses to NMDA application (200 µM) from control slices (top)and following D1 receptor activation (bottom, 25 µM). (b) NMDAfield potentials elicited as described in (a). Ethanol inhibition ofNMDA responses was analyzed under control conditions (top), inthe presence of the D1 agonist (middle, SKF 38393, 25 µM) and inthe presence of the D1 agonist and the D1 antagonist (SCH 23390,bottom). (c) Cumulative data for all slices studied. Ethanol inhibi-tion was tested at two concentrations (35 mM, raw traces notshown, and 75 mM) in the absence and presence of the D1 agonist,SKF 38393 (25 µM). The single asterisk indicates that ethanol inhi-bition in the SKF 38393 group is significantly different from the con-trol group (35 mM, P < 0.005, n = 4 control slices, n = 7 slicesfollowing SKF 38393 pre-treatment; 75 mM, P < 0.0001, n = 8 slicesfrom different rats in each group). The double asterisk indicatesthat ethanol inhibition (75 mM) between the SKF 38393 and theSKF 38393/SCH 23390 groups was significantly different (P < 0.015,n = 4 for SCH/SKF group) and no difference between the controland SCH 23390 groups existed (P = 0.1). (d) The effect of the pro-tein kinase A antagonist, H89, on the D1 receptor modulation ofthe ethanol sensitivity of NMDA receptors. Slices were testedunder conditions of strong D1 receptor activation (SKF 38393, 100 µM, alone top traces) and following pretreatment with the pro-tein kinase A antagonist, H89 (1 µM). (e) The cumulative ethanolinhibition data for all slices tested for experiments from (d). Thesingle asterisk indicates that H89 was significantly different from theSKF alone group (P < 0.002, n = 5 control, SKF 38393 group, n = 4H89/SKF group). The double asterisk indicates the cumulative datafrom an experiment testing the ability forskolin (Fsk, raw traces notshown) to mimic the action of the D1 agonist on NMDA receptorethanol sensitivity (P < 0.001 versus ethanol with no addition, n = 4 slices in forskolin group; P < 0.04 versus SKF 38393 (100 µM)).A lower concentration of forskolin (10 µM) also elicited a significantreduction (P < 0.02, n = 5) in ethanol sensitivity, which was notapparent when slices were exposed to an identical concentration ofthe inactive analogue, dideoxy-forskolin (dd-Fsk).

conditions (top traces), with SKF 38393 alone (25 µM, middletraces) and with SKF 38393 pre-treated with the selective D1receptor antagonist, SCH 23390 (15 µM, bottom traces). Ethanolinhibited the NMDA response by over 50% in the control con-dition and by about 15% in the presence of the D1 agonist. Pre-vious treatment of the preparation with the D1 receptorantagonist prevented the reduction in ethanol sensitivity ofNMDA responses by the receptor agonist (40% ethanol inhibi-tion). Cumulative data for all slices tested under these conditionsagain indicate approximately a 50% reduction in ethanol sensi-tivity induced by D1 receptor activation at the agonist concen-tration tested (25 µM, Fig. 2c). When slices were pretreated withthe D1 antagonist before the D1 agonist, the sensitivity of theneuronal population was restored to a level not significantly dif-ferent from that observed under control conditions (Fig. 2c). Wealso tested inhibition at a lower concentration of ethanol (35 mM) and observed a similar, approximately 50% reduction inethanol sensitivity in the presence of the D1 agonist (Fig. 2c).

A well-established pathway has been described in mediumspiny neurons whereby D1 receptor activation initiates a proteinkinase A–dependent phosphorylation cascade. Accordingly, we

0

10

20

30

40

50NMDA

0

10

20

30

40

50

SKF 38393

*

**

35 mM 75 mM EtOH

*

0.75mV

300 ms

SKF 38393 EtOH

H89/SKF 38393 EtOH

0.5mV

300 ms

0 20 40 60 800

20

40

60

80

100

120

0 20 40 60 800

20

40

60

80

100

120

min

ACSF +CPPeneWash+EtOH Wash

Wash+EtOH/

SKF 38393SKF

38393 Wash+CPPene

min

+H89dd-Fsk (10

M)

Fsk(50

M)µ µ

µ

Fsk (10 M)

+H89

NM

DA

res

po

nse

(%

co

ntr

ol)

Control +/–

+/–

EtOH

SKF 38393 EtOH

SKF 38393/ SCH 23390 EtOH

**

*

*

Control

SKF 38393/SCH 23390

SKF 38393

a

b c

de

+/–

+/–

+/–

EtO

H in

hib

itio

n (

%)

EtO

H in

hib

itio

n (

%)

D1/protein kinase A mechanisms and ethanol sensitivityTo pharmacologically identify the receptor and biochemical path-ways involved in the regulation of ethanol sensitivity in the nucle-us accumbens, we used a field potential recording technique thatdirectly activates postsynaptic NMDA receptors16. This techniqueallowed prolonged and stable recordings (>90 min) and avoid-ed potential pitfalls that occur with prolonged whole-cell record-ing. As observed using patch clamp recordings of individualneurons (Fig. 1), previous activation of D1 dopamine receptorsmarkedly decreased the ethanol sensitivity of field potentialNMDA responses (Fig. 2a). After washout of ethanol, applica-tion of the selective NMDA receptor antagonist, D-CPPene (1 µM), resulted in virtually a complete block of the NMDA fieldpotential (Fig. 2a). Thus, this technical approach is appropriatefor investigation of the molecular mechanisms underlying D1receptor suppression of NMDA receptor sensitivity to ethanol.These results also verify our previous findings using patch clamprecordings with both synaptic and direct NMDA receptor stim-ulation and indicate that D1 regulation of the ethanol sensitivityof NMDA responses can be detected at the population level.

We next addressed the pharmacological involvement of D1receptors in SKF 38393 modulation of the ethanol sensitivity ofNMDA receptors in NAcc (Fig. 2b). Ethanol (75 mM) inhibitionof NMDA responses was elicited in slices treated under control

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


articles

determined whether inhibition of this enzyme could interruptthis pathway and thereby occlude D1 receptor–dependent regu-lation of ethanol sensitivity. We also sought to determine whetherD1 receptor reduction of ethanol sensitivity could be enhancedunder conditions of near-maximal D1 receptor activation alone(SKF 38393, 100 µM, Fig. 2d) as well as after pretreatment withH89 (1 µM), an antagonist selective for protein kinase A (Fig. 2d).Slices pretreated with SKF 38393 (100 µM) also had a markedlydecreased sensitivity to ethanol that was greater than that observedat the lower concentration of the D1 agonist (Fig. 2e). However,when slices were pretreated with the protein kinase A antagonist,ethanol sensitivity of NMDA responses was restored to near con-trol levels (Fig. 2e). These data indicate that protein kinase A mustbe critical in this cascade, which culminates in the regulation ofethanol sensitivity of NMDA receptors in nucleus accumbens.

To confirm further the involvement of the protein kinase Apathway in D1 receptor regulation of NMDA receptor sensitivity toethanol, we also measured ethanol sensitivity in the presence ofthe adenylyl cyclase activator, forskolin (50 µM, Fig. 2e, middlebar graphs). Direct activation of PKA with forskolin resulted in adecrement in ethanol inhibition (75 mM) of the NMDA responseto a level of about 11%, significantly less than that observed undercontrol conditions; this decrease in ethanol sensitivity induced byforskolin was also reversed by the protein kinase A antagonist,

H-89 (Fig. 2e, middle graphs). We alsotested the effect of a lower concentrationof forskolin (10 µM), which also induceda significant decrease in ethanol sensitiv-ity relative to control inhibition (Fig. 2e,right bar graphs). In comparison, theethanol sensitivity determined in the

presence of the inactive analogue, dideoxy-forskolin, was signifi-cantly less than that observed in forskolin and not different fromcontrol values (Fig. 2e, right graphs).

DARPP-32 and NR1 phosphorylation in NAccThe D1 receptor activated protein kinase A–DARPP-32 cascadehas been suggested to induce the phosphorylation of numeroustarget proteins, including NMDA receptors, in medium spinyneurons. We hypothesized that this cascade may encompass amechanism through which D1 receptor activation decreases theethanol sensitivity of NMDA receptors. Therefore, we measuredthe time course of phosphorylation of DARPP-32, as well asNMDA-NR1 subunits, using immunoblot techniques in slices

Fig. 3. Phosphorylation of DARPP-32 andNMDA receptors coincides with the timecourse of diminished ethanol sensitivity. (a, b) The time course of DARPP-32 andNR1 phosphorylation, respectively, mea-sured using immunoblot techniques innucleus accumbens slices treated as in elec-trophysiological recordings in Figs. 1 and 2.Lanes are indicated at the bottom of (b) andare identical for all treatment groups forboth proteins and include basal, SKF 38393alone (25 µM, 1–20 min, left blots) and 5min following pretreatment with SCH23390 (20 min, 15 µM, right blots). Treatedversus untreated groups were significantlydifferent across all times (P < 0.0015 forpNR1, P < 0.0005 for phospho-DARPP-32groups, n = 4–5 separate experiments foreither phosphoprotein by ANOVA). (c) Confocal microscopic-double labelingimmunofluorescent detection in mediumspiny neurons of the nucleus accumbens ofthe localization of NR1 (fluorescein isothio-cyanate (FITC), top) and DARPP-32 (rho-damine B isothianate (RITC), middle) withsuperimposed images at the bottom.

pN

R1

(% c

ontr

ol)

1–

SKF38393:SCH23390:

400

300

200

100

0

400

300

200

100

0

500

10–

0–

5–

20 min –

–20

525

pDARPP

tDARPP

NR1

pNR1

pD

AR

PP

/to

tal D

AR

PP (

% c

on

tro

l)

10 µ

pD

AR

PP

/to

tal D

AR

PP

(%

co

ntr

ol)

0

25

50

75

100

125

150

Control 1 5 10 20 min

pDARPP

Fig. 4. Ethanol directly increases DARPP-32 phosphorylation in nucleusaccumbens. (a) The time course of DARPP-32 phosphorylation inducedby ethanol (25 mM) in a typical experiment. (b) The cumulative datashowing DARPP-32 phosphorylation following ethanol treatment(ANOVA for 1–20-min time points, P < 0.01, n = 5 experiments).

a

b

a

b

c

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

articles


treated as in the previous electrophysiological experiments.DARPP-32 phosphorylation was stimulated within 1 minute ofagonist application and reached a maximum level after 10 min-utes and remained stable over the ensuing 10 minutes (Fig. 3a).These data indicate that DARPP-32 was maximally phosphory-lated under conditions in which the ethanol sensitivity wasassessed electrophysiologically. DARPP-32 phosphorylation wasnot observed in slices pre-treated with the D1 antagonist SCH23390 (15 µM) before or 5 minutes following subsequent addi-tion of the D1 agonist (Fig. 3a). We analyzed the level of totalDARPP-32 in all samples following subsequent stripping and re-probing with an antibody specific for total DARPP-32. No changein total DARPP-32 protein was observed following D1 receptoractivation (Fig. 3a). The cumulative data from all slices tested inthis manner for DARPP-32 phosphorylation normalized to totalDARPP-32 in each sample indicates that the maximal increasein DARPP-32 phosphorylation was almost 300% over basal lev-els with a significant increase in DARPP-32 phosphorylationwithin 1 min of agonist application (Fig. 3a).

The phosphorylation of NR1 at Ser-897 via a protein kinaseA/DARPP-32–dependent process has been suggested to resultfrom D1 receptor activation. Therefore, if the regulation ofethanol sensitivity of these neurons is mediated via this pathway,this may be a critical step through which the ethanol sensitivity ofNMDA receptors is modified. Therefore, we also assessed thestate of phosphorylation of NR1 in the same samples generatedfor the measurements of DARPP-32 phosphorylation (Fig. 3b).We observed an increase in NR1 phosphorylation following D1receptor activation of similar magnitude and time course to thatof DARPP-32 phosphorylation. The stimulation of NR1 phos-phorylation was completely abolished by previous treatment withthe D1 antagonist, and these treatments did not modify theendogenous levels of NR1. The cumulative data for these sam-ples indicate that the peak level of NR1 phosphorylation was over300% greater than control levels, and the peak phospho-NR1level was reached before that observed with phospho-DARPP-32, suggesting a high degree of regulation of NR1 phosphoryla-tion in nucleus accumbens (see Discussion).

Our patch clamp and immunoblot studies are suggestive ofcolocalization of DARPP-32 and NMDA receptors within theprincipal neurons of the nucleus accumbens, the medium spinyneurons. However, there are no published reports that provideany direct immunocytochemical evidence for DARPP-32 andNMDA receptor colocalization in any type of neuron. To direct-ly verify that these proteins are indeed colocalized in these cells,

Fig. 5. Mice lacking DARPP-32 are insensitive to D1 receptor modu-lation of ethanol sensitivity of NMDA receptors. (a) Field potentialrecordings of NMDA responses in slices of nucleus accumbens frommice lacking (DARPP-32–/–, left traces) and from mice containingDARPP-32 (DARPP-32+/+, right traces). Traces are averagedresponses from 5 responses elicited in the conditions indicated; SKF38393 and ethanol applications were performed for 15–20 min ineach condition; bottom traces are superimposed for comparison.Artifacts due to pressure ejection of NMDA are truncated for clarity.(b) The cumulative data from all slices tested from separate mice. Thebasal level of ethanol inhibition between DARPP-32–/– and DARPP-32+/+ mice was not significantly different (raw traces not shown, P =0.61, n = 7 DARPP-32+/+ mice and n = 8 DARPP-32–/– mice).However, the ethanol sensitivity in slices from DARPP-32–/– mice wasnot altered 20 min after application of the D1 agonist while ethanolinhibition in the DARPP-32+/+ slices was reduced substantially (stu-dent’s t-test, P < 0.001, n = 7–8 mice).

we used double-labeling immunofluorescence techniques andconfocal imaging of NAcc (Fig. 3c). Freshly frozen cryostat sections prepared from identically aged rats, as for the previouselectrophysiological recordings, were used for these immunocy-tochemical studies, and samples were prepared under permeabi-lizing conditions to identify all cell-surface and internalizedproteins. NR1 (Fig. 3c, top) was labeled with a secondary anti-body derived from fluorescein; DARPP-32 was labeled with arhodamine-derived secondary in the same samples (Fig. 3c, mid-dle) and under these conditions, the colocalization of DARPP-32 and NR1 was readily apparent (Fig. 3c, bottom). Consistentwith the density of these principal neurons of the accumbens(95% are medium spiny neurons), these proteins colocalized tosome degree in virtually all the neurons we observed.

Ethanol induction of DARPP-32 phosphorylationAn additional question of interest concerns the direct effect ofethanol exposure on DARPP-32 phosphorylation. Slices weretreated as in the previous experiments and exposed to ethanol ata moderate concentration (25 mM). Ethanol did induce a sig-nificant, albeit moderate, level of DARPP-32 phosphorylation(Fig. 4a). The cumulative data for the time course of ethanoleffects on DARPP-32 phosphorylation indicate that ethanolinduced a peak of about 39% over control phospho-DARPP-32levels within 10 minutes that did not decline substantially with-in the next 10 minutes (Fig. 4b).

Ethanol sensitivity in mice lacking DARPP-32The time course and pharmacological results of the immunoblotstudies provide strong evidence indicative of the involvement ofDARPP-32 in the phosphorylation of the NMDA-NR1 subunit;however, this evidence is indirect and circumstantial. To directlydemonstrate the involvement of the DARPP-32 cascade, weassessed the ethanol sensitivity in DARPP-32–/– mice using theNMDA pressure ejection electrophysiological analysis described inFig. 2 as used in rats. Such an expectation is indicated because D1receptor–dependent phosphorylation of NR1 is absent in DARPP-32–/– mice15. Recordings from NAcc slices from DARPP-32–/– andDARPP-32+/+ mice were pretreated with the D1 agonist for 15–20minutes (SKF 38393, 25 µM) and subsequent bath application ofethanol (75 mM) resulted in inhibition of the NMDA response,which differed substantially between groups (Fig. 5). As previ-ously observed in rats, DARPP-32+/+ mice displayed a markedreduction in ethanol sensitivity when exposed in the presence ofthe D1 agonist. However, ethanol inhibition of NMDA responses

0.3mV

300 ms

SKF38393

EtOH+SKF 38393

EtOH EtOH/SKF 38393

*

Super-imposed

NMDA NMDA

0

10

20

30

40

50

60DARPP-32

DARPP-32

DARPP-32 DARPP-32 –/–

–/–

+/+

+/+

EtO

H in

hib

itio

n (

%)

a b

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


articles

in DARPP-32–/– mice displayed a much greater sensitivity toethanol, independent of the presence of the D1 agonist. Cumu-lative data from all slices tested from individual mice indicate thatactive DARPP-32 is necessary for D1 receptor regulation of theethanol sensitivity of NMDA receptors (Fig. 5b). There was aslight, albeit insignificant, trend that the DARPP-32–/– mice dis-played a greater sensitivity to ethanol under basal conditions(without SKF 38393). This indicates that there is little if any basaltone of D1 receptor function influencing NMDA sensitivity toethanol, and suggests that a wide range of regulatory control isavailable via this pathway. The previous immunoblot analysis (Fig. 3b) is supportive of this conclusion in that only under con-ditions of substantial D1 receptor activation did we observe a sub-stantial level of NMDA-NR1 phosphorylation.

DISCUSSIONIn animal models and in humans, ethanol consumption increasesupon repeated exposure, a phenomenon that may be associatedwith use-dependent adaptation in reward circuits. However, micelacking DARPP-32 do not display such an increase in ethanol intakeeven when the available ethanol concentration is increased to a rel-atively high range12. This result suggests a lack of neuroadaptiveprocesses that might promote intake with repeated exposure toincreasing concentrations. We propose a mechanism whereby theD1-receptor/DARPP-32 pathway modulates ethanol reinforce-ment. Whereas self-administration studies indicate that DARPP-32is an important component of this neuroadaptive process, addi-tional findings critical to this mechanism have been reported. First,mice lacking D1 receptors fail to exhibit increases in ethanol self-administration in tandem with the availability of increasing con-centrations of ethanol in a pattern surprisingly similar to that seenin DARPP-32 knockout mice12,17. Second, enhancement of accum-bal DA release, directly measured using in vivo microdialysis, occursin temporal synchrony with operant ethanol self-administration18.Finally, ethanol directly increases VTA firing rates, possibly due toinhibition of a delayed rectifier potassium channel19. These findingssupport the contention that endogenous activation of D1 recep-tors and DARPP-32 phosphorylation may encompass steps thatpromote ethanol reinforcement.

Previous evidence suggests similarities between glutamatergicprocesses underlying synaptic potentiation and processes underly-ing drug reinforcement. NMDA receptor–dependent potentiationof excitatory synapses on VTA neurons occurs following a singledose of cocaine and glutamatergic excitation from limbic structuresre-initiates cocaine responding2,3. Accordingly, we propose thatalterations in glutamatergic transmission that are triggered byNMDA receptor activation are involved in the induction of ethanolreinforcement. Indeed, a direct involvement for NMDA receptors inethanol reward has been reported, as NMDA antagonists can sup-plant for ethanol in conditioned place preference studies20.

We propose that if NMDA receptor activation is critical forestablishing ethanol reinforcement, then some mechanism mustexist that occludes inhibitory actions of ethanol on accumbalNMDA receptors. The present findings of D1/DARPP-32 regu-lation of ethanol sensitivity of NMDA receptors, combined withprevious findings in the literature, suggest a mechanism thatwould contribute to a positive-feedback system promoting therecurrent intake of ethanol. Upon exposure to ethanol, depolar-ization of VTA neurons induces an increase in the release of DAin the NAcc, thereby activating D1 receptors and initiating theprotein kinase A/DARPP-32 cascade. This cascade results in thephosphorylation of DARPP-32, thereby decreasing the activityof protein phosphatases and increasing the phosphorylation of

Ser-897 on the NR-1 subunit. Phosphorylation at this site resultsin a decreased sensitivity of the NMDA receptor to ethanol, whichmaintains a near normal level of NMDA receptor function andpromotes long-term alterations in glutamatergic transmission.Subsequent activation of distal components of the medial fore-brain bundle and the propagation of ethanol reinforcement sig-nals to forebrain and limbic sites would be promoted by theincreased tone of glutamatergic transmission in accumbens. D1receptor–mediated modulation of the ethanol sensitivity ofNMDA receptors is indeed demonstrable via both synaptic stim-ulation as well as direct agonist activation of postsynaptic NMDAreceptors, indicating that the site of modulation of ethanol sen-sitivity is postsynaptic. A low level of DARPP-32 phosphoryla-tion was induced by ethanol treatment, suggesting that thissignaling pathway is activated during intoxication.

It is important to qualify this hypothesis with caveats. The out-come of conditioning stimulation of NMDA receptors in medi-um spiny neurons is complex and can result in eitherenhancement or depression of either NMDA or non-NMDA com-ponents of glutamatergic synaptic transmission21–23. Additional-ly, the output of the medium spiny neuron is GABAergic andinhibitory, and the nature of NMDA receptor–dependent synap-tic plasticity underlying behavioral neuroadaptive responses inethanol addiction remains to be addressed. In this regard, we havefocused our efforts toward direct measures of ethanol effects onNMDA receptor activity and, accordingly, have sought to identi-fy the intracellular processes regulating ethanol sensitivity.

Phosphorylation-dependent regulation of the ethanol sensitiv-ity of NMDA receptors has previously been demonstrated for atyrosine kinase, fyn, which seems to modulate ethanol tolerance,especially of NR2B receptors24. Concerning NR1, truncation of C0elements has been associated with reduced ethanol sensitivity at asite associated with calcium-dependent desensitization of the recep-tor25. The present findings are the first to demonstrate dopamine-dependent regulation of NMDA receptor sensitivity to ethanol ina mesocorticolimbic structure. Our data suggest that phosphory-lation of Ser-897 of the NMDA receptor decreases the ability ofethanol to inhibit NMDA receptors. Discriminating between directsteric hindrance mechanisms involving an ethanol binding pock-et at or near this residue or more indirect and complex mechanismswill require further analysis. Ser-897 is critically involved in theexpression of NMDA receptor effects on CREB phosphorylation,which are triggered by D1 receptor activation26. These findingsimplicate the PKA/DARPP-32 pathway in the mediation of NMDAreceptor function and its regulation by D1 receptor activation.

The relative time courses of DARPP-32 and NR-1 phospho-rylation in our samples seemed to be virtually synchronous; also,phospho-DARPP-32 generation did not reach levels seen withNR-1 phosphorylation. Complex regulation of the DARPP-32–dependent system that modulates ethanol sensitivity of thesetarget neurons exists and may explain the differences in the timecourse for generation of phospho-DARPP-32 and phospho-NR1.For instance, a negative feedback regulation of DARPP-32 phos-phorylation via NMDA receptor activation of calcineurin hasbeen reported27. Indeed, such an inhibitory effect on phospho-DARPP-32 levels would be more likely at the low concentrationsof the D1 agonist we used.

Inhibition of NMDA receptor–dependent function (for exam-ple, in working memory and spatial processing) by ethanol is wellknown28. Interestingly, cortical and hippocampal sites involved insuch tasks have limited dopaminergic innervation when com-pared with other telencephalic structures that are implicated tohave a much greater involvement in drug reinforcement mech-

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

articles


anisms (for example, the NAcc). We propose that the DA inner-vation onto the NAcc appears to substantially reduce the ability ofethanol to inhibit the function of NMDA receptors, a majorsynaptic mediator of long-term neuroadaptive responses tococaine. By this mechanism, dopamine release triggered byethanol intake would maintain NMDA receptor function prim-ing long-term synaptic alterations that promote use-dependentethanol intake. Therefore, we suggest that this cascade, andDARPP-32 in particular, represent targets for modulation of theprocesses contributing to ethanol addiction.

METHODSMaterials. Forskolin and H89 were obtained from Calbiochem (SanDiego, California). SKF 38393 and SCH 23390 were purchased fromResearch Biochemicals (Natick, Massachusetts). Pronase, protein phos-phatase and protease inhibitors were from Sigma Chemical (St. Louis,Missouri). Anti-phospho-NR1 (Ser-897) was from Upstate Biotechnol-ogy (Waltham, Massachusetts). Cell culture media were from Gibco-Invitrogen (Carlsbad, California).

Preparation and incubation of coronal sections. Male Sprague–Dawleyrat pups three to four weeks old, obtained from the Animal ResourceCenter at The University of Texas at Austin, were decapitated accordingto an IACUC-approved protocol. The brains were dissected in ice-cold,oxygenated artificial cerebrospinal buffer (ACSF; 120 mM NaCl, 25 mMNaHCO3, 3.3 mM KCl, 1.23 mM NaH2PO4 and 10 mM dextrose, 95% O2/5% CO2). Coronal slices (500 µm) were prepared using a vibrat-ing slicer (VibroSlice, Campden Instruments, Sileby, UK) and beginningfrom 2 mm anterior to bregma. The slices were then maintained at 32°Cin ACSF buffer for 45–60 min before experimental analysis. In somecases, more than one slice was used per animal except for the knockoutstudies, where one slice per animal was used.

Culture techniques. The region surrounding the anterior commissuremedial to the ventral striatum from rats three days old was dissected andthen minced on ice and incubated with pronase, washed with DMEM,triturated through fire-polished glass pipets and cultured for 3–7 daysin Neural basal media with B-27 supplement (Gibco). Previous immuno-cytochemical analysis of DARPP-32 in these cultures indicated that vir-tually all cells displaying a neural cytoarchitecture in these culturescontain DARPP-32 and NR1 (data not shown).

Slice electrophysiology and general patch clamp recording methods. Coronal hemisections containing NAcc were transferred to a recordingchamber and continuously perfused with ACSF buffer at 32°C and a flowrate of 1.2 ml/min. The ACSF buffer was identical to that used for tissuepreparation; however, the concentrations of CaCl2 and MgSO4 were adjust-ed to 2 mM and 1 mM, respectively, unless otherwise noted. All electro-physiological recordings were made between a 300–500 µm arc of tissueoutside of the anterior commissure, encompassing the shell of the NAcc.Medium spiny neurons, which encompass more than 95% of the neuronsin this brain region, were electrophysiologically identified by their restingmembrane potentials greater than –70 mV. Recording electrodes were madefrom thin-walled borosilicate glass (TW150F-4, WPI, Sarasota, Florida,1.2–2.2 MΩ) and filled with 135 mM KMeSO3, 8 mM NaCl, 0.5 mMEGTA, 10 mM HEPES, 2 mM MgCl2, 2 mM Tris-ATP, 0.3 mM Tris-GTP,260-270 mOsm, pH 7.2 with KOH. Access resistance was partially com-pensated and monitored throughout all experiments. Recordings whereaccess resistance changed more than 10% during the course of the exper-iment were not included in data analyses and were made using an AxonInstruments Model 200B amplifier filtered at 1 kHz and digitized at 10–20 kHz with a Digidata interface (Axon Instruments, Foster City, Califor-nia). Data acquisition and amplifier control were implemented usingpClamp v7.0 or v8.0 software (Axon Instruments) All data presented areaveraged from epochs of at least five individual trials recorded during the15–20 min following ethanol exposure. Responses were continually mon-itored and recorded before and after that time for stability of ethanol onsetand offset. In all recordings where recordings were maintained followingethanol application, reversal was at least partially observed.

NMDA EPSCs. NMDA-receptor mediated EPSCs elicited by electricalstimulation of local afferents were isolated pharmacologically usingDNQX (10 µM) and picrotoxin (25 µM) and evoked by local stimula-tion via monopolar tungsten electrodes within 200 µm of the recordingsite (100 µs duration, 50–200 µA amplitude, evoked at 0.02 Hz).

NMDA whole-cell currents. Whole-cell currents were recorded using theidentical patch electrode setup as for the synaptic recordings except theywere done in the perforated patch mode with amphotericin B (60 mg/mlstock solutions in DMSO were periodically diluted (every 3 h) into freshstandard intracellular solution at a final concentration of 0.2–0.3% v/v).All recordings were made in the presence of tetrodotoxin (0.5 µM), MgCl2(0.5 mM) and CaCl2 (2 mM). Access resistance was continually moni-tored during cell-attached mode to verify opening of the patch to whole-cell mode. This process normally required 15 min following sealing andrecordings were initiated when the access resistance reached less than 50 MΩ. Currents were elicited via pressure ejection (10–20 p.s.i., 3–5 mspulse at 1-min intervals, Picospritzer, General Valve, Fairfield, New Jer-sey) of a solution of NMDA and glycine (1 mM and 0.1 mM, respective-ly) from a patch pipet placed within 20–30 µM of the cell.

Extracellular NMDA field potentials. Field responses to local applicationof NMDA were recorded using a single micropipette (3-4 MΩ) for simul-taneous pressure microejection of NMDA and potential recording16. Sin-gle pressure pulses of 10–30 ms duration and 10–30 p.s.i. were delivered at2-min intervals in the shell region of the NAcc. Stability of all recordingswas verified and recordings were terminated if responses varied by morethan 20%. NMDA responses were measured at peak amplitude and a min-imum of three responses was averaged for each condition, normally thelast three of a drug administration epoch. All drugs, except for NMDA,were applied by bath perfusion. The NMDA stock solution (10 mM) wasdissolved in ACSF buffer at a final concentration of 200 µM used for pres-sure ejection. The pressure ejection artifacts were omitted for clarity fromthe traces shown; however, the artifact was continually monitored forresponse stability and to detect pipet clogging. Ethanol inhibition waschecked for reversal by washing back to normal recording solution.

Immunoblotting. Coronal sections were incubated for the times specifiedin each experiment. After exposure to the drugs, the NAcc was quicklydissected on ice, transferred to microfuge tubes containing 100 µl of 1%SDS, snap frozen in liquid N2 and stored for analysis within 1–2 days.Samples were sonicated for 10 s and boiled for 10 min.

For analysis of phospho-DARPP-32 protein expression, 25 µg of proteinand low-range molecular weight standards were electrophoresed on a 12%acrylamide SDS-PAGE gel and immunoblotted onto polyvinylidenediflouride (PVDF) membranes (Immobilon-P; Millipore, Bedford, Massa-chusetts). The membranes were blocked for 1 h at room temperature (7.5%nonfat dry milk in 137 mM NaC, 25 mM Tris, 3 mM KCl, 25 mM Tris-HCl, 0.2% Tween 20 and phosphatase inhibitor cocktail(1:100)) and immunoblotted overnight with phospho-DARPP-32 antibody(1:750)29, a monoclonal antibody raised against the phospho-Thr-34 sitephosphorylated or DARPP-32 antibody (1:10000)30, a monoclonal anti-body that was used to measure the total amount of DARPP-32 in the sam-ples. Antibody binding was detected using a goat anti-mouse horseradishperoxidase-linked IgG (1:10000; Bio Rad, Hercules, California) and ECLimmunoblotting detection system (Amersham Pharmacia, Piscataway, NewJersey). For phospho-NR1 (pNR1) analysis, 25 µg of protein and high-rangemolecular weight standards were electrophoresed on a 7.5% acrylamideSDS-PAGE gel and immunoblotted onto PVDF membranes. Detection ofphospho-NR1 was determined using the anti-phospho-NR1 antibody(Upstate Biotechnology), specific for the phospho-NR1 subunit phospho-rylated by PKA on serine residue 897. Antibody binding was detected usinga goat anti-rabbit horseradish peroxidase-linked IgG (1:10000) and ECLimmunoblotting detection system. For western analysis, immunoblots werequantified by digital imaging with a high-resolution laser scanner and ScionImage1.62 densitometry software (NIH, Bethesda, Maryland).

Confocal fluorescence immunohistochemistry. Rats (20–30 days old)were transcardially perfused with 0.9% ice-cold saline followed by 4%paraformaldehyde under equithesin anesthesia. Frozen cryostat sections

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


receptors and ethanol: inhibition of calcium flux and cyclic GMPproduction. J. Neurochem. 52, 1937–1940 (1989).

8. Morrisett, R. A. & Swartzwelder, H. S. Attenuation of hippocampal long-term potentiation by ethanol: a patch clamp analysis of glutamatergic andGABAergic mechanisms. J. Neurosci. 13, 2264–2272 (1993).

9. Thomas, M. P., Monaghan, D. T. & Morrisett, R. A. Evidence for a causativerole of NMDA receptors in an in vitro model of alcohol withdrawalhyperexcitability. J. Pharmacol. Exp. Therap. 287, 87–97 (1998).

10. Thomas, M. P. & Morrisett, R. A. Dynamics of NMDAR-mediatedneurotoxicity during chronic ethanol exposure and withdrawal.Neuropharmacology 39, 218–226 (2000).

11. Grant, K. A., Knisely, J. S., Tabakoff, B., Barrett, J. E. & Balster, R. L. Ethanol-like discriminative stimulus effects on non-competitive N-methyl-D-aspartate antagonists. Behav. Pharmacol. 2, 87–95 (1991).

12. Risinger, F. O., Freeman, P. A., Greengard, P. & Fienberg, A. A. Motivationaleffects of ethanol in DARPP-32 knock-out mice. J. Neurosci. 21, 340–348(2001).

13. Fienberg, A. A. et al. DARPP-32: regulator of the efficacy of dopaminergicneurotransmission. Science 281, 838–842 (1998).

14. Hemmings, J. H. C., Greengard, P., Tung, H. Y. L. & Cohen, P. DARPP-32, adopamine-regulated neuronal phosphoprotein, is a potent inhibitor ofprotein phosphatase-1. Nature 310, 503–505 (1984).

15. Snyder, G. L., Fienberg, A. A., Huganier, R. L. & Greengard, P. Adopamine/D1 receptor/protein kinase A/dopamine- and cAMP-regulatedphosphoprotein (Mr 32 kDa)/protein phosphatase-1 pathway regulateddephosphorylation of the NMDA receptor. J. Neurosci. 18, 10297–10303(1998).

16. Fedorov, N. B. & Reymann, K. G. Simultaneous local pressure microejectionof excitatory amino acids and field potential recording with singlemicropipette in the hippocampal slice. J. Neurosci. Meth. 50, 83–90 (1993).

17. El-Ghundi, M. et al. Disruption of dopamine D1 receptor gene expressionattenuates alcohol-seeking behavior. Eur. J. Pharmacol. 353, 149–158(1998).

18. Gonzales, R. A. & Weiss, F. Suppression of ethanol-reinforced behavior bynaltrexone is associated with attenuation of the ethanol-induced increase indialysate dopamine levels in the nucleus accumbens. J. Neurosci. 18,10663–10671 (1998).

19. Brodie, M. S., Pesold, C. & Appel, S. B. Ethanol directly excites dopaminergicventral tegmental reward neurons. Alcohol. Clin. Exp. Res. 23, 1848–1852(1999).

20. Biala, G. & Kotlinska, J. Blockade of the acquisition of ethanol-inducedconditioned place preference by N-methyl-D-asparate receptor antagonists.Alcohol Alcohol 34, 175–182 (1999).

21. Pennartz, C., Aneerun, R., Groenewegen, H. & Lopes da Silva, F. Synapticplasticity in an in vitro slice preparation of the rat nucleus accumbens. Eur. J. Neurosci. 5, 107–117 (1993).

22. Kombian, S. B. & Malenka, R. C. Simultaneous LTP of non-NMDA- and LTDof NMDA-receptor-mediated responses in the nucleus accumbens. Nature368, 242–246 (1994).

23. Thomas, M. J., Malenka, R. C. & Bonci, A. Modulation of long-termdepression by dopamine in the mesolimbic system. J. Neurosci. 20,5581–5586 (2000).

24. Kalluri, H. & Ticku, M. Effect of ethanol on phosphorylation of theNMDAR2B subunit in mouse cortical neurons. Brain Res. Mol. Brain Res. 68,159–168 (1999).

25. Anders, D., Blevins, T., Smothers, C. & Woodward, J. J. Reduced ethanolinhibition of N-methyl-D-aspartate receptors by deletion of the NR1 C0domain or overexpression of α-actinin-2 proteins. J Biol. Chem. 275,15019–15024 (2000).

26. Rajadhyalsha, A. et al. L-type Ca2+ channels are essential for glutamate-mediated CREB phosphorylation and c-fos gene expression in striatalneurons. J. Neurosci. 19, 6348–6359 (1999).

27. Halpain, S., Girault, J.-A. & Greengard, P. Activation of NMDA receptorsinduces dephosphorylation of DARPP-32 in rat striatal slices. Nature 343,369–372 (1990).

28. Givens, B., Williams, J. & Gill, T. Septohippocampal pathway as a site forthe memory-impairing effects of ethanol. Hippocampus 10, 111–121(2000).

29. Snyder, G. L. et al. Phosphorylation of DARPP-32 and protein phosphataseinhibitor-1 in rat choroid plexus: regulation by factors other than dopamine.J. Neurosci. 12, 3071–3083 (1992).

30. Hemmings, J. H. C. & Greengard, P. DARPP-32, a dopamine-and adenosine3′:5′-monophosphate-regulated phosphoprotein: regional, tissue, andphylogenetic distribution. J. Neurosci. 6, 1469–1481 (1986).

articles

(20 µm thick) were washed 3 with saponin (0.1%) and blocked withgoat serum (10% in 0.1% saponin). Slides were then incubated overnightat 4°C with antibody (1°C) for a total NR1 as used for immunoblotting,then washed in saponin/serum and incubated with Alexa-Fluor-488 goatanti-rabbit 2° antibody (Molecular Probes, Eugene, Oregon) overnightat 4°C and again washed in saponin/serum. The process was then repeat-ed with DARPP-32 antibody (1°C) and Alexa-Fluor-532 goat anti-mouseantibody (2°C). Slides were mounted with ProLong antifade (MolecularProbes)and imaged with an Olympus Fluoview FV-300 confocal micro-scope mounted on an Olympus IX-70 inverted microscope and using a40 oil UPLANAPO obj. (Leeds Inst., Irving, Texas). Excitation was viathe 488 (NR1) or 514 (DARPP-32) lines of a multi-line argon laser(Model 163/263, Spectra-Physics, Mountain View, California) using notchfilters to eliminate the extra lines (Kaiser Optical, Ann Arbor, Michigan).Non-specific staining was verified by incubation of companion slides asabove except that the 1°C antibody was omitted from the protocol. In allcases, such treatment completely eliminated all specific staining.

DARPP-32–/– and DARPP-32+/+ mice. Homozygous C57Bl/6 congenicmice (back-crossed 10 times) lacking DARPP-32 (and their homozygouswild-type littermates) were obtained from The Rockefeller Institute.

Data analysis and statistics. Ethanol sensitivity was calculated as follows.

Simple comparisons were made using student’s t-test and multiple com-parisons used ANOVA comparing all time points to the baseline response.

AcknowledgmentsThis work was supported primarily by NIH R01AA11845 (R.A.M.) and the

Integrative Neuroscience Initiative on Alcoholism (INIA, NIH U01AA13CR

(R.E.M.)), the Texas Commission on Alcoholism and Drug Abuse (TCADA)

and the Waggoner Center on Alcoholism and Addiction Research (WCAAR)

(R.A.M.). Additional support was from NIH R01AA11852 (R.A.G.) and

R01AA11836 (S.W.L.) and T32AA07471 (K.F-K.). The authors thank

R.A. Harris and A. Hendricson for suggestions and discussion, P. Greengard

and P. Ingrassia for DARPP-32 antibodies and DARPP-32 modified mice, and

A. Miao and M. Jia for technical support.


RECEIVED 13 MAY; ACCEPTED 29 MAY 2002

1. Wise, R. A. Drug-activation of brain reward pathways. Drug Alcohol Depend.51, 13–22 (1998).

2. Ungless, M. A., Whistler, J. L., Malenka, R. C. & Bonci, A. Single cocaineexposure in vivo induces long-term potentiation in dopamine neurons.Nature 411, 587–587 (2001).

3. Vorel, S. R., Liu, X., Hayes, R. J., Spector, J. A. & Gardner, E. L. Relapse tococaine-seeking after hippocampal theta burst stimulation. Science 292,1175–1178 (2001).

4. Nestler, E. J. Neurobiology. Total recall-the memory of addiction. Science 292,2266–2267 (2001).

5. Woodward, J. J. Ethanol and NMDA receptor signalling. Crit. Rev. Neurobiol.14, 68–89 (2000).

6. Lovinger, D. M., White, G. & Weight, F. F. Ethanol inhibits NMDA-activatedion currents in hippocampal neurons. Science 243, 1721–1724 (1989).

7. Hoffman, P. L., Rabe, C. S., Moses, F. & Tabakoff, B. N-methyl-D-aspartate

(peak amplitude of NMDA response in ethanol) (peak amplitude of control NMDA response) 1–

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

Synaptotagmins are Ca2+-binding proteins that contain an N-terminal transmembrane region and two C-terminal C2 domains.Synaptotagmins 1 and 2, the most abundant isoforms, are local-ized to synaptic vesicles and endocrine secretory granules1–3, andare essential for fast but not slow Ca2+-dependent exocytosis4–6.Both C2 domains of synaptotagmins 1 and 2 (referred to as C2Aand C2B domains) bind multiple Ca2+ ions, and form Ca2+-depen-dent complexes with negatively charged phospholipids7–11. In addi-tion, the C2 domains participate in other Ca2+-dependentinteractions12, including Ca2+-dependent binding to SNARE com-plexes mediated by the C2A domains, and Ca2+-dependent self-association of synaptotagmins mediated by the C2B domains9,13–16.

The affinity of C2 domains to Ca2+ is low; for example, com-plete Ca2+-binding to synaptotagmin 1 requires more than 10 mM Ca2+ (refs. 5 and 10). The Ca2+ affinity of the C2 domainsis dramatically increased when the domains are bound to Ca2+-dependent ligands such as phospholipids, probably because theseligands provide additional coordination sites for Ca2+ ions17.Consistent with this notion, the apparent Ca2+ affinity of C2domains differs characteristically for each ligand9, and in the caseof phospholipids, the apparent Ca2+ affinity depends on the elec-trostatic charge of the phospholipids5. Furthermore, double C2-domain fragments have higher apparent Ca2+ affinities than C2domains, presumably because the linkage of two autonomousCa2+-binding domains results in cooperativity. For example, inphospholipid complexes, the double C2A/B-domain fragment ofsynaptotagmin 1 exhibits an apparent Ca2+ affinity of 1–6 µMCa2+ (depending on the phospholipids), whereas the individualC2A or C2B domains exhibit affinities of 7–40 µM Ca2+ each5,7,10.

Synaptotagmin function in densecore vesicle exocytosis studied incracked PC12 cells

Ok-Ho Shin1, Josep Rizo2 and Thomas C. Südhof1

1 Center for Basic Neuroscience, Department of Molecular Genetics and Howard Hughes Medical Institute and 2 Departments of Biochemistry and Pharmacology, University of Texas Southwestern Medical Center, 6000 Harry Hines Blvd., Dallas, Texas 75390, USA.

Correspondence should be addressed to T.S. ([email protected])


Ca2+-triggered dense-core vesicle exocytosis in PC12 cells does not require vesicular synaptotagmins1 and 2, but may use plasma membrane synaptotagmins 3 and 7 as Ca2+ sensors. In support of thishypothesis, C2 domains from the plasma membrane but not vesicular synaptotagmins inhibit PC12cell exocytosis. Ca2+ induces binding of both plasma membrane and vesicular synaptotagmins tophospholipids and SNAREs (soluble N-ethylmaleimide-sensitive attachment protein receptors),although with distinct apparent Ca2+ affinities. Here we used gain-of-function C2-domain mutants ofsynaptotagmin 1 and loss-of-function C2-domain mutants of synaptotagmin 7 to examine howsynaptotagmins function in dense-core vesicle exocytosis. Our data indicate that phospholipid- butnot SNARE-binding by plasma membrane synaptotagmins is the primary determinant of Ca2+-triggered dense-core vesicle exocytosis. These results support a general lipid-based mechanism ofaction of synaptotagmins in exocytosis, with the specificity of various synaptotagmins for differenttypes of fusion governed by their differential localizations and Ca2+ affinities.

Moreover, these C2 domain–phospholipid complexes have high-er apparent Ca2+ affinities than single or double C2-domain frag-ments that complex with SNAREs (EC50 ≈ 20–1,000 µM Ca2+)5,9.

The similarity between the apparent Ca2+ affinities of synaptotagmin–phospholipid complexes and the exocytotic Ca2+

sensor in central synapses18,19 suggests that synaptotagmin 1 nor-mally functions in synaptic vesicle exocytosis by forming aCa2+–phospholipid complex. Studies of mutant mice containing apoint mutation in the endogenous synaptotagmin 1 gene, theR233Q mutation, support this hypothesis5. This mutation causesa twofold decrease in the apparent Ca2+ affinity of synaptotagmin 1during phospholipid binding, but has no effect on the apparentCa2+ affinity of the synaptotagmin 1–SNARE complex. The samemutation decreases the Ca2+ affinity of neurotransmitter releaseapproximately twofold5. This result suggests that Ca2+-dependentphospholipid complexes are instrumental for the function of synap-totagmin 1 in synaptic vesicle exocytosis.

In contrast to synaptic vesicle exocytosis, dense-core vesicle exo-cytosis in neuroendocrine cells is largely independent of synapto-tagmins 1 and 2 (refs. 6 and 20). Deletion of vesicularsynaptotagmins impairs the minor fast component of dense-corevesicle exocytosis but leaves the major slower component intact6,and overexpression of synaptotagmin 1 alters fusion pore kineticsduring dense-core vesicle exocytosis21. Besides synaptotagmin 1, atleast two other synaptotagmins (synaptotagmins 3 and 7) are abun-dantly expressed in neuroendocrine cells, and could potentially bemajor Ca2+ sensors for exocytosis9,11,22,23. Synaptotagmins 3 and7 have the same domain structure as classical synaptotagmins, butare localized to plasma membranes23,24.In cracked PC12 cells

articles


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


(see Methods), a widely used system to study Ca2+-triggered exo-cytosis21,25–28, C2 domains from synaptotagmins 3 and 7 severelyinhibit exocytosis, whereas synaptotagmin 1 C2 domains have noeffect11,23. (However, ref. 29 reports a more extensive inhibition bysynaptotagmin 1 C2-domain fragments.) Only two other proteinsinhibit exocytosis in cracked PC12 cells as strongly as C2 domains ofplasma membrane synaptotagmins: the light chains of clostridialneurotoxins27,30, and the SNARE motif of syntaxin 1 (ref. 31). Theselective effects of C2 domains from plasma membrane synapto-tagmins support the notion that these synaptotagmins are Ca2+

sensors for dense-core vesicle exocytosis, but provide no clue totheir mechanism of action.

To explore how synaptotagmins function in dense-core vesi-cle exocytosis, we searched for mutations that convert (activate)the synaptotagmin 1 C2A domain into an inhibitor of PC12 cellexocytosis, or that abolish (inactivate) inhibition of exocytosisin the synaptotagmin 7 C2A domain. These experiments wereguided by the notion that synaptotagmin C2 domains inhibit exo-cytosis at the point when Ca2+ triggers exocytosis, by competingwith endogenous exocytotic Ca2+ sensors, likely endogenoussynaptotagmins. To test if this reaction requires binding by phos-pholipids or by SNAREs, we compared the effects of these muta-tions on PC12 cell exocytosis and on phospholipid- and

SNARE-binding. Our data showed that inhibition of PC12 cellexocytosis by the C2A domain of plasma membrane synaptotag-min 7, and lack of such inhibition by the corresponding C2Adomain from synaptotagmin 1, depends on binding by phos-pholipids but not by SNAREs. However, phospholipid bindingalone cannot explain the inhibitory reaction, suggesting that theC2 domain–phospholipid complex is a specific agent in as yetunidentified tertiary interactions.

RESULTSModulating Ca2+-dependent activities of C2A domains Previous studies showed that the top loops of synaptotagmin C2domains determine their Ca2+-dependent properties32,33. Thesynaptotagmin 1 and 7 C2A domains exhibit an approximatelytenfold difference in apparent Ca2+ affinity and exert distincteffects on PC12 cell exocytosis11,23, but their presumptive Ca2+

binding sites are very similar (Fig. 1a). We wanted to analyze howthese C2 domains differ and by what mechanism the synapto-tagmin 7 C2A domain inhibits PC12 cell exocytosis. To do this,we searched for mutations that selectively change the Ca2+-depen-dent interactions of these C2 domains with phospholipids andSNARE proteins, two interactions that are widely thought tomediate the function of synaptotagmins in exocytosis12.

We first transplanted the top loops of the synaptotagmin 7C2A domain onto the synaptotagmin 1 C2A domain to generategain-of-function mutants. We made seven mutants that incor-porate various combinations of synaptotagmin 7 Ca2+-bindingloops grafted onto the corresponding sequences of the synapto-tagmin 1 C2A domain (Fig. 1b). Using a liposome centrifugation

articles

0 1 2 3 10 30 100

1,000

Total

Ca2+ (µM)

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

0 1 2 3 10 30 100

1,000

Total

Ca2+ (µM)

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

a

b

c

C

N

N

C

D

GI

LEAP

I/P

K/R

H/N

D

S

F RD

D

S

T

Y/L G

G/S

M/F

L/K

P

D

O H

O

O

OO

O

O

O

OO

O

OO

O

Ca3

Ca2

F

Ca1

Y/F

/Y

Loop 1

Loop 3

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

Syt 1-C A2

Syt 1-C A-Loop 12

Syt 1-C A-Loop 22

Syt 1-C A-Loop 32

Syt 1-C A-Loop 1,22

Syt 1-C A-Loop 1,32

Syt 1-C A-Loop 2,32

Syt 1-C A-Loop 1,2,32

Syt 7-C A2

Syt 1-C A2

Syt 1-C A-M173W2

Syt 1-C A-F231W/F234W2

Syt 1-C A-3W2

Syt 1-C A/B2

Syt 1-C A/B-6W2

Syt 7-C A2

Syt 7-C A-F169A2

Syt 7-C A-F229A2

Syt 7-C A-F169A/F229A2

Fig. 1. Sequence determinants of the Ca2+ affinities of C2 domains. (a) Ca2+-binding sites of the C2A domains from synaptotagmins 1 and 7.Only loops 1 and 3 are shown (synaptotagmin 1, residues 140–267;synaptotagmin 7, residues 134–262) because loop 2 does not directlycontact Ca2+ ions. At positions where synaptotagmins 1 and 7 differ, thesynaptotagmin 1 residue is shown first followed by the synaptotagmin 7residue. Amino acids that were mutated are circled. (b) Apparent Ca2+

affinities of mutant synaptotagmin 1 C2A domains containing Ca2+-bind-ing loops from the synaptotagmin 7 C2A domain. The apparent Ca2+

affinities of wild-type and mutant C2 domains (tested as purified GST-fusion proteins) were studied with the liposome centrifugation assay.Liposomes composed of 25% PS/75% PC were incubated with the indi-cated C2 domains (present as purified GST-fusion proteins) at the Ca2+

concentrations shown on top (clamped with Ca2+/EGTA buffers) andcentrifuged, and bound proteins were analyzed by SDS-PAGE andCoomassie blue staining. Data shown are from a single representativeexperiment repeated multiple times. The C2 domains analyzed are thewild-type C2A domains of synaptotagmin 1 (top, Syt 1-C2A) and 7 (bot-tom, Syt 7-C2A), and mutant C2A domains in which the Ca2+-bindingloops 1, 2 and 3 from synaptotagmin 7 were pasted into the corre-sponding sequences of synaptotagmin 1 individually or in combination,as described to the right of the Coomassie-stained gels (Syt 1C2A-loop 1, -loop 2 and so on). Data shown in this panel and in (c) areCoomassie-stained gels from a single representative experimentrepeated multiple times. (c) Apparent Ca2+ affinities of synaptotagmin 1and 7 C2A domains with amino acid substitutions in hydrophobicresidues. The apparent Ca2+ affinities of wild-type and mutant C2domains were tested as purified GST-fusion proteins with the liposomecentrifugation assay. The point mutants analyzed were named accordingto the parent C2 domain and the mutated residues. (For example, Syt 1C2A-M173W indicates that methionine173 in the synaptotagmin 1 C2Adomain was changed into a tryptophan.) The mutants in which threeresidues in the C2A domain and six residues in the double C2A/B-domain fragments of synaptotagmin 1 were substituted for tryptophansare referred to as -3W and -6W, respectively.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

assay10, we then measured the Ca2+ concentration–dependence ofphospholipid binding in these mutants (Fig. 1b). Grafting onesynaptotagmin 7 Ca2+-binding loop onto the synaptotagmin 1C2A domain had small effects on its apparent Ca2+ affinity. Trans-plantation of loops 1 and 3 together (these loops directly ligatethe Ca2+ ions; Fig. 1a) increased the apparent Ca2+ affinity sig-nificantly (from ∼ 10 µM to ∼ 3 µM Ca2+, P < 0.05; Fig. 1b). Inser-tion of all three synaptotagmin 7 Ca2+-binding loops into thesynaptotagmin 1 C2A domain further augmented the apparentCa2+ affinity (to ∼ 2–3 µM Ca2+), although even the triple-loopmutant did not achieve the Ca2+ affinity of the normal synapto-tagmin 7 C2A domain (<1 µM; Fig. 1b). We saw no change inthe phospholipid binding specificity of the triple-loop swapsynaptotagmin 1 C2A domains (Fig. 2). Together, these data sug-gest that all three Ca2+-binding loops of the synaptotagmin 7 C2Adomain contribute to its higher Ca2+ affinity, but that Ca2+ affin-ity also depends on additional differences in the overall struc-tures of the C2A domains.

We next searched for point mutations that alter the Ca2+-dependent interactions of the synaptotagmin 1 and 7 C2Adomains without changing the number of Ca2+ ions bound.The apparent affinity of synaptotagmin C2 domains for Ca2+

during phospholipid binding depends directly on the Ca2+-binding residues in the top loops and indirectly on positivelycharged and hydrophobic amino acids that surround theseresidues5,34. These positively charged and hydrophobic residuesincrease the C2 domains’ apparent Ca2+ affinity, probablybecause they stabilize the Ca2+–phospholipid complexes of theC2 domains by contacting the phospholipid bilayers. Guidedby this model, we introduced mutations into C2A domains thatincrease the hydrophobicity of synaptotagmin 1 or decrease thehydrophobicity of synaptotagmin 7. We chose hydrophobicresidues instead of positively charged amino acids becausehydrophobic residues were more likely to selectively alter phos-pholipid binding but not SNARE binding. Only subtle changeswere introduced that change the hydrophobicity of key residueswithout affecting the surface charge.

Methionine173 of the synaptotagmin 1 C2A domain is essentialfor Ca2+-dependent phospholipid binding, presumably because itinserts into the docked phospholipid bilayer34,35. Enhancing thehydrophobicity of this residue by replacing it with tryptophaninduced a minor increase in phospholipid binding and apparentCa2+ affinity (Fig. 1c). We obtained a similarly small increase byreplacing two other hydrophobic residues that are located on thetop loops of the C2A domain (phenylalanine231 and phenylala-nine234) with tryptophans. However, converting all threehydrophobic residues into tryptophans (referred to as 3W mutant)induced a more dramatic increase in the apparent Ca2+ affinity ofthe synaptotagmin 1 C2A-domain–phospholipid complex (from∼ 10 µM to ∼ 2.5 µM). Conversely, mutating phenylalanine169 inthe synaptotagmin 7 C2A domain (which corresponds to M173 insynaptotagmin 1; numbering is based on the shortest splice vari-ant23) into alanine significantly reduced the apparent Ca2+ affini-ty of the domain. The single mutation in phenylalanineF229 (whichcorresponds to F234 in synaptotagmin 1) and the double pheny-

articles


0 1 2 3 10 30 100

1,00

0

Tota

l

Ca2+ (µM)

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

15% PIP/ 85% PC

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

50% PI/ 50% PC

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

40% PS/ 60% PC

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

50% PE/ 50% PC

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

Syt 7-C A2


Syt 1-C A2

Syt 1-C A-3W2


Syt 7-C A2


Syt 1-C A2

Syt 1-C A-3W2


Syt 7-C A2


Syt 1-C A2

Syt 1-C A-3W2


Syt 7-C A2


Syt 1-C A2

Syt 1-C A-3W2


Fig. 2. Phospholipid-specificity during Ca2+-dependent phospholipidbinding of various synaptotagmin 1 and 7 C2A-domain mutants.Phospholipid binding was assayed as described in Fig. 1 with liposomescontaining different phospholipid compositions. Phospholipid composi-tions are indicated on the left.

2+Ca concentration (µM)

20

60

80

40

100

0

20

60

80

40

100

0

20

60

80

40

100

0

20

60

80

40

100

0

Syt 7-C A2

3 [

H]-P

C b

indi

ngN

orm

aliz

ed3

[ H

]-PC

bin

ding

Nor

mal

ized

0 0.3 0.6 1 3 10 30100 1,000 0 0.3 0.6 1 3 10 30100 1,000

0 0.3 0.6 1 3 10 30100 1,000 0 0.3 0.6 1 3 10 30100 1,000

Syt 1-C A2 Syt 1-C A2

Syt 1-C A2 Syt 1-C A2

Syt 1-C A- Loop 1,2,3

2 Syt 1-C A-3W (3W)

2

Syt 7-C A- F169A/F229A (AA)

2


Fig. 3. Apparent Ca2+ affinities of wild-type and mutant synaptotagmin1 and 7 C2A domains measured with a resin-pulldown assay. GST-fusionproteins of the indicated C2 domains were immobilized on glutathionebeads and incubated with radiolabeled liposomes (25% PS/75% PC) atthe free Ca2+ concentrations shown. In all graphs, binding of a given C2domain is compared to that of the synaptotagmin 1 C2A domain. TheAA mutation in the synaptotagmin 7 C2A domain converts the submi-cromolar apparent Ca2+ affinity of this domain into that of the synapto-tagmin 1 C2A domain.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


lalanine mutation (referred to as AA mutant) had more dramaticeffects, lowering the apparent Ca2+ affinity from less than 1 µM toapproximately 10 µM (Fig. 1c). None of these mutations signifi-cantly altered the phospholipid binding specificity of the respec-tive C2A domains (Fig. 2). We used the GST-pulldown assay as anindependent measure of apparent Ca2+ affinities7 to allow a bet-ter definition of Ca2+-dependent phospholipid binding at low Ca2+

concentrations. The GST-pulldown assay revealed that the appar-ent Ca2+ affinity of the synaptotagmin 7 C2A domain is submi-cromolar, but is converted into an approximately 10-µM apparentCa2+ affinity identical to that of the synaptotagmin 1 C2A domainby the AA mutation (Fig. 3).

Finally, we examined the effects of some of these mutationson the double C2 domain fragment of synaptotagmin 1 (referredto as C2A/B-domain fragment). This fragment had an approxi-mately 5–10-fold higher apparent Ca2+ affinity than either theC2A- or C2B domain alone (Fig. 1c). Although the wild-typeC2A/B-domain fragment of synaptotagmin 1 could be producedeasily, W-mutant C2A/B-domain fragments were very difficultto obtain (data not shown). The only mutant that we could puri-fy in sufficient amounts was a one in which three hydrophobicresidues in both C2 domains were converted into tryptophans(referred as the 6W mutant). Ca2+-dependent phospholipid bind-ing studies with this mutant revealed that it bound to phospho-lipids even in the nominal absence of Ca2+, indicating that C2domains can be converted from a Ca2+-regulated to a constitutivemodule by changing key residues that interact with the phos-pholipid membranes (Fig. 1c).

SNARE binding by C2A-domain mutantsThe C2A domain of synaptotagmin 1 binds to SNARE proteins,although relatively high Ca2+ concentrations are required (>0.1 mM Ca2+)9,14,36. To test if the synaptotagmin 7 C2Adomain also binds to SNAREs, we performed GST-pulldownexperiments with rat brain proteins (Fig. 4). The synaptotag-min 7 C2A domain was 100-fold more potent in capturing

brain SNARE proteins than the synaptotagmin 1 C2Adomain, and was active at lower Ca2+-concentrations (∼ 10 µM versus ∼ 1–2 mM, respectively). All three synapticSNARE proteins were bound (Fig. 4), suggesting that thesynaptotagmin 7 C2A domain interacts with the assembledcore complex. Thus, the synaptotagmin 7 C2A domain, witha higher Ca2+ affinity, performs both major Ca2+-dependentactivities of the synaptotagmin 1 C2A domain (phospholipidand SNARE binding), suggesting that either activity couldpotentially explain the differential effect of synaptotagmin1 and 7 C2 domains in PC12 cell exocytosis23.

We next tested if the mutants of synaptotagmin C2Adomains that alter phospholipid binding have a similar effecton SNARE binding. The synaptotagmin 1 mutants thatincreased the apparent Ca2+ affinity of the phospholipid com-plex did not induce a major enhancement of Ca2+-dependentSNARE interactions (Fig. 4). For example, the 3W mutant ofthe synaptotagmin 1 C2A domain exhibited a phospholipid-dependent apparent Ca2+ affinity of approximately 2 µM Ca2+

(approaching that of the synaptotagmin 7 C2A domain) butstill required about 100 µM Ca2+ for SNARE binding. The dif-ferential effect of the mutations was most pronounced for theAA mutant of the synaptotagmin 7 C2A domain, which had asevere effect on Ca2+-dependent phospholipid binding (Figs. 1–3) but only a marginal effect on SNARE binding (Fig. 4). As a result, Ca2+-dependent phospholipid-bindingproperties of the synaptotagmin 7 AA mutant resembled the

synaptotagmin 1 C2A domain, but its SNARE-binding proper-ties were similar to the synaptotagmin 7 C2A domain.

Effects of C2-domain mutants on PC12 cell exocytosisThe differential effects of the synaptotagmin 1 and 7 C2A-domainmutations on phospholipid and SNARE binding allow us todetermine which of these binding activities are important forCa2+-triggered PC12 cell exocytosis. For this purpose, we pre-loaded PC12 cells with 3H-norepinephrine, permeabilized thecells by freeze-thawing, and preincubated the cracked cells with 6 µM of the various purified C2 domains. Exocytosis was thentriggered by addition of 10 µM Ca2+, a relatively high concen-tration to favor C2 domains with a low apparent Ca2+ affinity.

The results showed that the synaptotagmin 1 C2A domain didnot inhibit exocytosis, but hybrid synaptotagmin 1 C2A domainswith more synaptotagmin 7 Ca2+-binding loops had corre-spondingly more inhibition of exocytosis (Fig. 5a). Similarly, thegain-of-function point mutants of the synaptotagmin 1 C2Adomain that enhanced Ca2+-dependent phospholipid bindingalso enhanced inhibition of exocytosis, whereas the loss-of-function mutants of the synaptotagmin 7 C2A domain abolishedinhibition of exocytosis by this C2 domain (Fig. 5a). We alsoexamined the effect of the wild-type and mutant double C2domain fragment from synaptotagmin 1 because it was suggest-ed that this fragment may be a more effective agent in PC12 cellsthan the single C2A domain29. However, the wild-type doubleC2A/B domain fragment of synaptotagmin 1 was only moder-ately more inhibitory than the single C2A domain of synapto-tagmin 1, whereas the 6W mutation that activatedCa2+-dependent phospholipid binding (Fig. 1c) also activatedinhibition of Ca2+-triggered exocytosis (Fig. 5a). A plot of theamount of secretion observed in the presence of a wild type ormutant C2 domain versus the apparent Ca2+ affinity of that par-ticular C2 domain reveals a precise correlation between inhibi-tion of exocytosis and apparent Ca2+ affinity during phospholipidbinding (Fig. 5b). Together, these data demonstrate that in dense-

articles

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

1 2 3 4 5 6

Synaptobrevin 2 Syntaxin 1 SNAP-25

0 3 10 100

1,0

00

Lysa

te

0 3 10 100

1,0

00

Lysa

te

Ca2+ (µM)

0 3 10

100

1,0

00

Lysa

te

Ca2+ (µM) Ca2+ (µM)

Syt 1-C A2

Syt 1-C A-3W2

Syt 1-C A- Loop 1,2,3

2

Syt 7-C A2

Syt 7-C A- F169A/F229A

2

Fig. 4. Ca2+-dependent binding of wild-type and mutant synaptotagmin 1and 7 C2A domains to the SNARE proteins synaptobrevin/VAMP 2, syn-taxin 1 and SNAP-25. GST-fusion proteins containing the C2A-domainmutants with the largest changes in apparent Ca2+ affinity (Fig. 2) wereimmobilized on glutathione beads and incubated with solubilized rat brainextracts at different concentrations of free Ca2+ clamped with Ca2+/EGTAbuffers. Bound proteins were analyzed by SDS-PAGE and immunoblotting.Two gain-of-function mutants of synaptotagmin 1 and a single loss-of-function mutant of synaptotagmin 7 with the strongest effects were ana-lyzed in comparison to the wild-type C2A domains.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

core vesicle exocytosis from PC12 cells, the phospholipid-dependent apparent Ca2+ affinity of the C2 domains is adetermining factor.

To ensure that the various inhibitory effects of the loss-and gain-of-function mutants were not specific for the sin-gle Ca2+ concentration used in Fig. 5, we measured the effectsof three mutations (the AA mutation of the synaptotagmin 7C2A domain, and the complete loop-swap and the 3W muta-tion for the synaptotagmin 1 C2A domain) on exocytosis overthe entire range of Ca2+-concentrations (Fig. 6). The resultsconfirmed that the synaptotagmin 7 C2A domain severelyinhibited Ca2+-triggered exocytosis at all Ca2+ concentrations,whereas the single and double C2-domain fragment fromsynaptotagmin 1 exerted only small effects. The submicro-molar Ca2+ dependence of exocytosis in cracked PC12 cellsresembled the apparent Ca2+ affinity of the synaptotagmin 7C2A domain in the presence of phospholipids (Fig. 3), and is con-sistent with the Ca2+ dependence of dense-core vesicle exocyto-sis in chromaffin cells37. The AA mutation abolished theinhibitory activity of the synaptoagmin 7 C2A domain at all Ca2+

concentrations. In contrast, the gain-of-function mutations ofthe synaptotagmin 1 C2A domain conferred inhibitory activityonto this domain particularly at high Ca2+ concentrations (Fig. 6). This inhibitory activity was significantly less than thatof the synaptotagmin 7 C2A domain, but more than that observedfor the wild-type double C2A/B-domain fragment of synapto-tagmin 1. Viewed together, these data suggest that although theapparent Ca2+ affinity is a determining factor in inhibition (Fig. 5b), simply increasing the Ca2+ concentration does not con-vert non-inhibitory domains into inhibitory domains.

Specificity of C2 domains in PC12 cell exocytosisThe importance of the Ca2+-dependent C2 domain–phospholipidcomplex in Ca2+-triggered PC12 cell exocytosis indicates thatinteractions between opposing phospholipid membranes, possi-bly mediated by synaptotagmins, are important in PC12 cell mem-brane fusion triggered by Ca2+. One concern is that instead ofacting specifically, the C2 domains inhibit exocytosis by forming anonspecific membrane coat. To address this concern, we com-pared the synaptotagmin C2A domains with the C2 domain ofcytoplasmic phospholipase A2 (cPLA2), which has a high apparentCa2+ affinity, similar to the synaptotagmin 7 C2A domain, but hasa distinct phospholipid-binding specificity33,38–40.

Fig. 5. Comparative analysis the inhibition of Ca2+-triggered exocy-tosis in cracked PC12 cells by wild type and mutant C2A domains.(a) PC12 cells were loaded with 3H-labeled norepinephrine, cracked,and preincubated with the purified C2-domain GST-fusion proteinsshown on top (concentration, 6 µM; see Fig. 1 for a description ofthe mutants). Exocytosis is measured as the amount of norepineph-rine released during a 30-min incubation period after addition ofbuffer containing 0 or 10 µM Ca2+. Data are from a representativeexperiment independently repeated multiple times; responses arenormalized to the control that was set at 100%. Data are means ±s.e.m. (n = 6 for 1–16, n = 3 for 17 and 18; each experiment wasdone in duplicate). (b) Amount of norepinephrine secretion inducedby 10 µM Ca2+ in cracked PC12 cells in presence of various C2domains versus the apparent Ca2+ affinity of these C2 domains asphospholipid complexes. The line shown was calculated by regres-sion analysis (r2 = 0.91) under exclusion of the two points obtainedwith the double C2-domain fragment from synaptotagmin 1 (Syt 1-C2A/B and Syt 1C2A/B-6W). Data shown are means ± s.e.m.; whereerror bars are not visible, they were too small to be shown.

We wanted to ensure that in our buffer systems, with the par-ticular phospholipids used, the apparent Ca2+ affinities of thecPLA2 C2 domain and the synaptotagmin 7 C2A domain werecomparable. Thus, we first examined the Ca2+-dependent phos-pholipid binding properties of the synaptotagmin 1 and 7 C2Adomains and the cPLA2 C2 domain in the same experiments. Ca2+

titrations with a range of phospholipids confirmed previous con-clusions33,38,40 that the C2 domain of cPLA2 preferentially bindsto neutral phospholipids, whereas the synaptotagmin C2 domainspreferentially binds to negatively charged phospholipids (Fig. 7).The apparent Ca2+ affinity of the cPLA2 C2 domain was highestwith 100% neutral phospholipids (<<1 µM Ca2+), and decreasedwith increasing concentrations of negatively charged phospho-lipids. The Ca2+ affinity of the cPLA2 C2 domain was similar tothat of the synaptotagmin 7 C2A domain in the presence of 15%phosphatidylinositol-phosphate or of 25% phosphatidylserine (∼ 1 µM Ca2+; Fig. 7), concentrations of negatively charged phos-pholipids that may correspond to those of plasma membranes(reviewed in ref. 39). At higher concentrations of negatively chargedphospholipids that are probably not physiological, the apparentCa2+ affinity of the cPLA2 C2 domain decreased further (Fig. 7).

We then tested the effect of the cPLA2 C2 domain on Ca2+-triggered exocytosis in cracked PC12 cells by titrating exocytosisin the presence of identical amounts of C2 domains with increas-ing concentrations of Ca2+. The cPLA2 C2 domain caused mod-erate inhibition of exocytosis, especially at higher Ca2+

concentrations. However, inhibition by the cPLA2 C2 domain

articles


a

b100

75

50

25

0

0 2 4 6 8 10

17

Loop swapping Loss of function

14 15 164 5 6 7 8 9 10 11 12 131 2 3

4 5 6 7 8 9 10 11 12 13

Gain of function14 15 161 2 3

41

NE re

lease

(% o

f tot

al)

17 18

17 18

Double C2 domain

Syt

1-C

A

GS

T

Syt

7-C

AS

yt 1

-C A

-Loo

p 1

Syt

1-C

A-L

oop

2

Syt

1-C

A-L

oop

3

Syt

1-C

A-L

oop

1,2

Syt

1-C

A-L

oop

1,3

Syt

1-C

A-L

oop

2,3

Syt

1-C

A-L

oop

1,2,

3

Syt

7-C

A-F

169A

Syt

7-C

A-F

229A

Syt

7-C

A-F

169A

/F22

9A

Syt

1-C

A-M

173W

Syt

1-C

A-F

231W

/F23

4W

Syt

1-C

A-3

WS

yt 1

-C A

/BS

yt 1

-C A

/B-6

W


Syt 7-C A-F229A2

Syt 1-C A2

Syt 1-C A-Loop 22

Syt 1-C A-Loop 1,22

Syt 1-C A-Loop 12

Syt 1-C A-F231W/F234W2

Syt 1-C A-Loop 32

Syt 1-C A-Loop 1,32


Syt 1-C A/B2

Syt 1-C A-M173W2

Syt 1-C A-Loop 2,32

Syt 1-C A-3W2

Syt 1-C A/B-6W2

Syt 7-C A2

Syt 7-C A-F169A2

2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2

Ca -

induc

ed N

E se

cret

ion (%

of c

ontro

l)2+

Apparent Ca -affinity of C -domain/phospholipid complex (µM)2+2

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


was much smaller than the virtually complete inhibition observedwith the synaptotagmin 7 C2A domain (Fig. 8a). To ensure thatthe inhibition by the cPLA2 C2 domain was not limited by proteinavailability, we tested the effects of increasing concentrations of C2domains (Fig. 8b). The inhibition by the cPLA2 C2 domain wasclearly dose-dependent, but flattened at a lower level than thatobserved by the synaptotagmin 7 C2A domain, suggesting thatsimply coating membranes with a C2 domain achieves only par-tial inhibition of PC12 cell exocytosis.

DISCUSSIONCa2+-triggered exocytosis in cracked PC12 cells is strongly inhib-ited by C2 domains from synaptotagmin 7, but is largely unaf-fected by closely related C2 domains from synaptotagmin 1 (refs. 11 and 23). The cracked PC12 cell assay measures Ca2+-triggered exocytosis from secretory vesicles that are alreadydocked and primed; thus, synaptotagmin 7 C2 domains inhibita late step of exocytosis25–28. Ca2+ induces dense-core vesicle exo-cytosis within seconds37. Although this is slow compared to themillisecond time scale of synaptic vesicle exocytosis18,19, inhibi-tion of PC12 cell exocytosis by the C2 domains must happen inthe same time period because it depends on Ca2+-binding to theC2 domains. Here we used the differential effects of C2 domainsfrom synaptotagmin 1 and 7 on PC12 cell exocytosis to explore

how synaptotagmins function in dense-core vesicle exocytosis.We first showed that the synaptotagmin 7 C2A domain exhibiteda higher apparent Ca2+ affinity for SNAREs and for phospho-lipids than the synaptotagmin 1 C2A-domain or C2A/B-domainfragment; thus, both SNARE and phospholipid binding couldpotentially explain the preferential action of the synaptotagmin 7C2A domain in inhibiting exocytosis (Figs. 1–4). We then definedmutations that selectively altered the hydrophobicity of keyresidues in the synaptotagmin 1 and 7 C2A domains, and differ-entially changed their Ca2+-dependent binding to phospholipidsand to SNAREs. When we tested these mutations in the crackedPC12 cell assay, we found that the apparent Ca2+ affinity of a C2-domain/phospholipid complex precisely correlated with itsinhibitory activity in PC12 cell exocytosis, whereas SNARE bind-

articles

0 0.1 0.3 0.6 1 10 1001,000

0 0.1 0.3 0.6 1 10 1001,000

Norm

alize

d NE

rele

ase

20

40

60

80

0

100

0 0.1 0.3 0.6 1 10 1001,000

20

40

60

80

0

100

0 0.1 0.3 0.6 1 10 1001,000 0 0.1 0.3 0.6 1 10 1001,000

20

40

60

80

0

100

20

40

60

80

0

100

0 0.1 0.3 0.6 1 10 1001,000

ControlControl

Control

Control

Control

Control

20

40

60

80

0

100

Norm

alize

d NE

rele

ase

20

40

60

80

0

100

Norm

alize

d NE

rele

ase


Syt 7-C A2


Syt 1-C A2 Syt 1-C A/B2

Syt 1-C A-Loop 1,2,32Syt 1-C A-3W2


Fig. 6. Ca2+-dependence of exocytosis in cracked PC12 cells in the pres-ence of wild-type and mutant C2A domains from synaptotagmin 1 and 7.Cracked PC12 cells, prelabeled with 3H-norepinephrine, were preincu-bated with purified GST-fusion proteins of the indicated C2A domains orthe same concentration of GST alone (‘control’ run separately for eachindividual C2A domain), and norepinephrine release was triggered by theaddition of Ca2+ at the indicated concentrations. Recombinant proteins(6 µM) were added to cracked PC12 cells, and exocytosis was triggeredby Ca2+ at the indicated concentrations. Ca2+ concentrations wereclamped with Ca2+/EGTA buffers. In each experiment, a separate controlincubation containing GST alone was assayed in parallel to exclude inter-experimental variability, resulting in small differences between the con-trol curves for each protein. Only the mutants with the largest effects onCa2+-dependent phospholipid- and SNARE-binding were analyzed.

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

25% PS/ 75% PC

40% PS/ 60% PC

0 1 2 3 10 30 100

1,00

0

Tota

l

Ca2+ (µM)

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

1 2 3 4 5 6 7 8 9

50% PE/ 50% PC

50% PI/ 50% PC

15% PIP/ 85% PC

Syt 1-C A2

Syt 7-C A2

cPLA - C22

Syt 1-C A2

Syt 7-C A2

cPLA - C22

Syt 1-C A2

Syt 7-C A2

cPLA - C22

Syt 1-C A2

Syt 7-C A2

cPLA - C22

Syt 1-C A2

Syt 7-C A2

cPLA - C22

Fig. 7. Comparison of Ca2+-dependent phospholipid binding by theC2A domains of synaptotagmins 1 (Syt 1-C2A) and 7 (Syt 7-C2A) andthe C2 domain of cytoplasmic phospholipase A2 (cPLA2-C2): effects ofphospholipid composition. Equivalent amounts of GST-fusion proteinscontaining the C2 domains were incubated with liposomes composed ofthe indicated ratios of phosphatidylserine (PS), phosphatidylcholine(PC), phosphatidylethanolamine (PE), phosphatidylinositol (PI) andphosphatidylinositolphosphate (PIP) in the presence of the concentra-tions of free Ca2+ shown. Liposomes were pelleted by centrifugation,and bound C2 domains were analyzed by SDS-PAGE and Coomassieblue staining. Data are from a single representative experiment that wasindependently repeated multiple times.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

ing seemed to be irrelevant (Fig. 5b). The most explicit examplewas provided by the AA mutation of the synaptotagmin 7 C2Adomain, which did not significantly alter SNARE binding butgreatly decreased phospholipid binding and abolished the abili-ty of the C2A domain to inhibit PC12 cell exocytosis (Figs. 1–6).

How does the synaptotagmin 7 C2A domain inhibit exocyto-sis? The most straightforward explanation is that the exogenousC2 domains coat target membranes, and thereby interfere withendogenous synaptotagmins whose Ca2+-dependent binding tothese very target membranes is essential for exocytosis. However,inhibition is probably not only due to a simple mechanical inter-ference of the Ca2+-dependent binding of endogenous synapto-tagmins to target membranes. The C2A domain of synaptotagmin 7 severely inhibited exocytosis at all Ca2+ con-centrations, whereas the single or double C2-domain fragmentsof synaptotagmin 1 did not cause severe inhibition at any Ca2+

concentrations (Figs. 5 and 6). At the same time, the apparentCa2+ affinities of the synaptotagmin 1 C2 domains, althoughlower than those of the synaptotagmin 7 C2 domains, were with-in the Ca2+ concentration range tested in the PC12 cell assays. Ifpure membrane coating were responsible for inhibition of exo-cytosis, one would expect the C2 domains with lower apparentCa2+ affinities to become inhibitory at higher Ca2+ concentra-tion, which was not the case. In addition, the C2 domain of cPLA2caused only limited inhibition (Fig. 8), although its apparentCa2+ affinity is comparable to that of the synaptotagmin 7 C2Adomain (Fig. 7). The most parsimonious model explaining thesedata is that Ca2+-dependent synaptotagmin 7 C2A-domain–phos-pholipid complexes participate in additional interactions that arenot formed by the cPLA2 C2 domain. However, alternative expla-nations, such as an unusual effect of the local phospholipid com-position, cannot be ruled out.

We cannot tell if our observations in neuroendocrine PC12 cellexocytosis are directly applicable to synaptic vesicle exocytosis. Thereare many differences between synaptic and dense-core vesicle exo-cytosis, not the least of which seems to be the relative unimpor-tance of vesicular synaptotagmins for Ca2+-triggering20 and theinvolvement of Ca2+ activator protein for secretion (CAPS) in Ca2+-dependent exocytosis41. Thus, neuroendocrine and synaptic vesi-cle exocytosis may exhibit distinct features. Independent of suchdifferences, however, the C2 domains of vesicular synaptotagminsprobably also function as Ca2+-triggered phospholipid complexes insynaptic vesicle exocytosis because the R233Q mutation thatchanges release probability selectively alters phospholipid bindingbut not SNARE binding5. Thus, our data suggest a convergence ofthe mechanism of action of synaptotagmins in different types ofexocytosis as Ca2+-dependent phospholipid-binding machines.

METHODSExpression and purification of recombinant proteins. The synaptotagminC2-domain and PLA2 C2-domain expression vectors in pGEX-KG42 weredescribed previously23,33. The loop swapping mutants were constructedby introducing the following point mutations into pGEX-Syt 1C2A: loop 1, A166Q/L171K/M173F/G174S; loop 2, H198K/T201N; loop 3,A227Q/Y229L/F231Y/K236R/H237N/I239P. For gain or loss of functionmutations of the synaptotagmin 1 and 7 C2A domains, we introducedthe following mutations: pGEX-Syt 1C2A-M173W/F231W/F234W (3W),pGEX-Syt 1C2A/B-M173W/F231W/F234W/V304W/Y264W/I367W(6W) and pGEX-Syt 7C2A-F169A/F229A using a commercial mutagen-esis kit (Stratagene, La Jolla, California); all plasmids were verified byDNA sequencing. Purification of GST-fusion proteins was done as pre-viously described23,42. To remove bacterial contaminants associated withC2B domains43, bacterial extracts containing C2B-domain proteins weretreated with 1,500 units/l benzonase (Novagen, Madison, Wisconsin) for1 h at room temperature, bound to glutathione agarose (0.3 ml per liter

culture) and washed in batch 2× with 15 ml 1 M NaCl in 10 mM Tris pH8.0, 1 mM EDTA; 5× with 20 mM CaCl2 in 10 mM Tris; 5× with 1 MNaCl, 10 mM Tris, 1 mM EDTA; and 3× with 10 mM Tris, 1 mM EDTA,150 mM NaCl before elution.

Phospholipid binding assays. Phospholipid binding assays were doneessentially as described using a pulldown assay with immobilized GST-fusion proteins7,9 or a centrifugation assay with soluble GST-fusion pro-teins10. All binding assays were done in HEPES buffer (50 mMHEPES-NaOH pH 6.8, 100 mM NaCl, 4 mM Na2EGTA) with Ca2+/EGTAbuffers whose composition was calculated with commercial software,EqCal for Windows (Biosoft, Ferguson, Missouri). All buffers were pre-pared in plastic containers with high-resistance MilliQ water and 1 MCa2+ standards (Fluka Chemical, Rankonkoma, New York).

SNARE protein pulldowns. Two unstripped rat brains (∼ 1.5 g/brain;Pel-Freez Biologicals, Rogers, Arkansas) were homogenized with a tissuehomogenizer (Thomas Scientific, Philadelphia, Pennsylvania) in 20 mlHEPES containing a protease inhibitor cocktail (Roche, Indianapolis,Indiana) and 1 mM DTT. One percent Triton X-100 was added, pro-teins were extracted for 1 h at 4°C with rocking and insoluble proteinswere removed by centrifugation (100,000g for 1 h). The supernatant wasused for SNARE protein pulldowns containing 250 µl brain lysate, 30µg GST-Syt C2 domains attached to 30 µl of glutathione agarose beadsand 250 µl of 2× Ca2+/EGTA buffers. The binding reactions were incu-bated at 4°C for 1 h with rocking, beads were washed 6× with 1 ml ofthe corresponding Ca2+/EGTA buffers containing 0.5% Triton X-100,resuspended in 50 µl SDS-sample buffer, and aliquots (30 µl) were ana-lyzed by SDS-PAGE and immunoblotting.

Cracked PC12 cell secretion assays. Cracked PC12 cell secretion assayswere done with freeze-thaw permeabilization of PC12 cells23,44. PC12 cellsin 100-mm plates at approximately 70% confluence (2 days after platingat 40% confluence) were loaded with 4 µl 3H-norepinephrine (NEN, 1 mCi/ml stock solution) and 0.5 mM ascorbic acid for 24 h, washed withphysiological saline solution (145 mM NaCl, 5.6 mM KCl, 2.2 mM CaCl2,0.5 mM MgCl2, 5.6 mM glucose, 15 mM HEPES-NaOH pH 7.4), harvestedby pipetting a stream of Ca2+-free ice-cold PC12 secretion buffer (120 mMK-glutamate, 20 mM K-acetate, 2 mM EGTA, 20 mM HEPES-NaOH

articles


2+Ca concentration (µM) Protein concentration (µM)

Control

Norm

aliz

ed N

E r

ele

ase

00 0.1 0.3 0.6 1 10 100 1,000

20

40

60

80

100

a b

NE

rele

ase (

% o

f to

tal)

50

40

30

20

0 1 2 3 4 5 106 7 8 9

Syt 1-C A2

Syt 7-C A2

Syt 7-C A2

cPLA -C22 cPLA -C22

Fig. 8. Effects of the cPLA2 C2 domain on Ca2+-triggered exocytosis incracked PC12 cells. (a) Ca2+ concentration dependence. Cracked, 3H-norepinephrine-loaded PC12 cells were preincubated with 6 µM of thecPLA2 C2 domain, and exocytosis was triggered by addition of Ca2+ atthe indicated concentrations. Data are from a single experiment in trip-licate that was repeated independently multiple times. (b) Protein con-centration dependence. Cracked 3H-norepinephrine loaded PC12 cellswere preincubated with different concentrations of purified GST-fusionproteins containing the C2 domain of cytoplasmic phospholipase A2(cPLA2) or the C2A domains of synaptotagmins 1 or 7 as indicated.Exocytosis was then triggered by addition of 10 µM Ca2+.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


pH 7.2), washed twice with 6 ml buffer and resuspended in 6-ml buffer ina 15-ml cone tube. For permeabilization, cell tubes were frozen overnightat –80°C and thawed at room temperature, 10 mM EGTA was added andthawed cells were left on ice for 2 h to allow efficient extraction of solubleproteins. Resulting cell ghosts were washed 3× with 6 ml of PC12 secre-tion buffer containing 1% BSA, with centrifugations at 1,000g for 5 min.Secretion reactions (∼ 20 reactions/100-mm plate) were done with stan-dard reaction mixes in 1.5-ml microtubes (total volume, 0.1 ml): washedcell ghosts, 2 mM ATP, 2 mM MgCl2, 10 µl rat brain cytosol (10 g/l) inPC12 secretion buffer with various concentrations of Ca2+ and recombi-nant proteins. Reactions were incubated for 30 min at 30°C and terminat-ed by chilling to 0°C, and samples were centrifuged at 4°C for 3 min at20,800g. Supernatants and the pellets solubilized in 1% Triton X-100 wereanalyzed by liquid scintillation counting.

SDS-PAGE and immunoblotting were done using standard proce-dures45,46. Immunoblots were developed by enhanced chemilumines-cence (Amersham, Piscataway, New Jersey).

AcknowledgmentsWe thank I. Leznicki, A. Roth and E. Borowicz for technical assistance, and S.

Gerber for the phospholipase A2 C2-domain expression plasmid. This study was

supported by grants from the NIH to J.R. (NS37200 and NS40944).


RECEIVED 16 APRIL; ACCEPTED 20 MAY 2002

1. Matthew, W. D., Tsavaler, L. & Reichardt, L. F. Identification of a synapticvesicle-specific membrane protein with a wide distribution in neuronal andneurosecretory tissue. J. Cell Biol. 91, 257–269 (1981).

2. Perin, M. S., Brose, N., Jahn, R. & Südhof, T. C. Domain structure ofsynaptotagmin (p65). J. Biol. Chem. 266, 623–629 (1991).

3. Geppert, M., Archer, B. T. III & Südhof, T. C. Synaptotagmin II: a noveldifferentially distributed form of synaptotagmin. J. Biol. Chem. 266,13548–13552 (1991).

4. Geppert, M., Goda, Y., Hammer, R. E., Li, C., Rosahl, T. W., Stevens, C. F. &Südhof, T. C. Synaptotagmin I: a major Ca2+ sensor for transmitter release at acentral synapse. Cell 79, 717–727 (1994).

5. Fernández-Chacón, R. et al. Synaptotagmin I functions as a Ca2+-regulator ofrelease probability. Nature 410, 41–49 (2001).

6. Voets, T. et al. Intracellular calcium dependence of large dense-core vesicleexocytosis in the absence of synaptotagmin 1. Proc. Natl. Acad. Sci. USA 98,11680–11685 (2001).

7. Davletov, B. & Südhof, T. C. A single C2-domain from synaptotagmin I issufficient for high affinity Ca2+/phospholipid-binding. J. Biol. Chem. 268,26386–26390 (1993).

8. Chapman, E. R. & Jahn, R. Calcium-dependent interaction of the cytoplasmicregion of synaptotagmin with membranes. Autonomous function of a singleC2-homologous domain. J. Biol. Chem. 269, 5735–5741 (1994).

9. Li, C. et al. Ca2+-dependent and Ca2+-independent activities of neural andnonneural synaptotagmins. Nature 375, 594–599 (1995).

10. Fernandez, I. et al. Three-dimensional structure of the synaptotagmin 1 C2B-domain: synaptotagmin 1 as a phospholipid-binding machine. Neuron 32,1057–1069 (2001).

11. Sugita, S., Shin, O-H., Han, W., Lao, Y. & Südhof, T. C. Synaptotagmins form ahierarchy of exocytotic Ca2+-sensors with distinct Ca2+-affinities. EMBO J. 21,270–280 (2002).

12. Südhof, T.C. Synaptotagmins: why so many? J. Biol. Chem. 277, 7629–7632(2002).

13. Bennett, M. K., Calakos, N. & Scheller, R. H. Syntaxin: a synaptic proteinimplicated in docking of synaptic vesicles at presynaptic active zones. Science257, 255–259 (1992).

14. Kee, Y. & Scheller, R. H. Localization of synaptotagmin-binding domains onsyntaxin. J. Neurosci. 16, 1975–1981 (1996).

15. Sugita, S., Hata,Y. & Südhof, T. C. Distinct Ca2+-dependent properties of thefirst and second C2-domains of synaptotagmin I. J. Biol. Chem. 271,1262–1265 (1996).

16. Chapman, E. R., An, S., Edwardson, J. M. & Jahn, R. A novel function for thesecond C2 domain of synaptotagmin. Ca2+-triggered dimerization. J. Biol.Chem. 271, 5844–5849 (1996).

17. Zhang, X., Rizo, R. & Südhof, T. C. Mechanism of phospholipid binding by theC2A-domain of synaptotagmin. Biochemistry 37, 12395–12403 (1998).

18. Bollmann, J. H., Sakmann, B. & Borst, J. G. Calcium sensitivity of glutamaterelease in a calyx-type terminal. Science 289, 953–957 (2000).

19. Schneggenburger, R. & Neher, E. Intracellular calcium dependence oftransmitter release rates at a fast central synapse. Nature 406, 889–893 (2000).

20. Shoji-Kasai, Y. et al. Neurotransmitter release from synaptotagmin-deficientclonal variants of PC12 cells. Science 256, 1821–1823 (1992).

21. Wang, C.T. et al. Synaptotagmin modulation of fusion pore kinetics inregulated exocytosis of dense-core vesicles. Science 294, 1111–1115 (2001).

22. Mizuta, M. et al. Synaptotagmin III is a novel isoform of rat synaptotagminexpressed in endocrine and neuronal cells. J. Biol. Chem. 269, 11675–11678(1994).

23. Sugita, S. et al. Synaptotagmin VII as a plasma membrane Ca2+-sensor inexocytosis. Neuron 30, 459–473 (2001).

24. Butz, S., Fernandez-Chacon, R., Schmitz, F., Jahn, R. & Südhof, T.C. Thesubcellular localizations of atypical synaptotagmins: synaptotagmin III isenriched in synapses and synaptic plasma membranes but not in synapticvesicles. J. Biol. Chem. 274, 18290–18296 (1999).

25. Ahnert-Hilger, G. M., Brautigam, M. & Gratzl, M. Ca2+-stimulatedcatecholamine release from alpha-toxin-cracked PC12 cells: biochemicalevidence for exocytosis and its modulation by protein kinase C and G proteins.Biochemistry 26, 7842–7848 (1987).

26. Chamberlain, L. H., Roth, D., Morgan, A. & Burgoyne, R. D. Distinct effects ofα-SNAP, 14-3-3 proteins, and calmodulin on priming and triggering ofregulated exocytosis. J. Cell Biol. 130, 1063–1070 (1995).

27. Chen, Y. A., Scales, S. J., Patel, S. M., Doung, Y-C. & Scheller, R. H. SNAREcomplex formation is triggered by Ca2+ and drives membrane fusion. Cell 97,165–174 (1999).

28. Avery, J. et al. A cell-free system for regulated exocytosis in PC12 cells. J. CellBiol. 148, 317–324 (2000).

29. Earles, C. A., Bai, J., Wang, P. & Chapman, E. R. The tandem C2-domains ofsynaptotagmin contain redundant Ca2+-binding sites that cooperate to engaget-SNAREs and trigger exocytosis. J. Cell Biol. 154, 1117–1123 (2001).

30. Banerjee, A., Kowalchyk, J. A., Dasgupta, B. R. & Martin, T. F. SNAP-25 isrequired for a late postdocking step in Ca2+-dependent exocytosis. J. Biol.Chem. 271, 20227–20230 (1996).

31. Zhong, P., Chen, Y. A., Tam, D., Chung, D., Scheller, R. H. & Miljanich, G. P.An α-helical minimal binding domain within the H3 domain of syntaxin isrequired for SNAP-25 binding. Biochemistry 36, 4317–4326 (1997).

32. Shao, X., Davletov, B. A., Sutton, R. B., Südhof, T. C. & Rizo, J. Bipartite Ca2+-binding motif in C2-domains of synaptotagmin and protein kinase C. Science273, 248–251 (2001).

33. Gerber, S. H., Rizo, J. & Südhof, T. C. The top loops of the C2-domains fromsynaptotagmin and phospholipase A2 control function specificity. J. Biol.Chem. 276, 32288–32292 (2001).

34. Gerber, S. H., Rizo, J., & Südhof, T. C. Role of electrostatic and hydrophobicinteractions in Ca2+-dependent phospholipid binding by the C2A-domain ofsynaptotagmin 1. Diabetes 51 (suppl. 1), S12–S18 (2002).

35. Chapman, E. R. & Davis, A. F. Direct interaction of a Ca2+-binding loop ofsynaptotagmin with lipid bilayers. J. Biol. Chem. 273, 13995–14001 (1998).

36. Gerona, R. R., Larsen, E. C., Kowalchyk, J. A., & Martin, T. F. The C terminusof SNAP25 is essential for Ca2+-dependent binding of synaptotagmin toSNARE complexes. J. Biol. Chem. 275, 6328–6336 (2000).

37. Voets, T. Dissection of three Ca2+-dependent steps leading to secretion inchromaffin cells from mouse adrenal slices. Neuron 28, 537–545 (2000).

38. Nalefski, E. A. et al. Independent folding and ligand specificity of the C2calcium-dependent lipid binding domain of cytosolic phospholipase A2. J. Biol. Chem. 273, 1365–1372 (1998).

39. Davletov, B., Perisic, O. & Williams, R. L. Calcium-dependent membranepenetration is a hallmark of the C2 domain of cytosolic phospholipase A2whereas the C2A domain of synaptotagmin binds membranes electrostatically.J. Biol. Chem. 273, 19093–19096 (1998).

40. Gennis, R. B. Biomembranes: Molecular Structure and Function (Springer, NewYork, 1989).

41. Walent, J. H., Porter, B. W. & Martin, T. F. A novel 145 kd brain cytosolicprotein reconstitutes Ca2+-regulated secretion in permeable neuroendocrinecells. Cell 70, 765–775 (1992).

42. Guan, K. L. & Dixon, J. E. Eukaryotic proteins expressed in Escherichia coli: animproved thrombin cleavage and purification procedure of fusion proteinswith glutathione S-transferase. Anal. Biochem. 192, 262–267 (1991).

43. Ubach, J. et al. The C2B-domain of synaptotagmin 1 is a Ca2+-bindingmodule. Biochemistry 40, 5854–5860 (2001).

44. Klenchin, V. A., Kowalshyk, J. A. & Martin, T. F. J. Large dense-core vesicleexocytosis in PC12 cells. Methods 18, 204–208 (1998).

45. Laemmli, U. K. Cleavage of structural proteins during the assembly of thehead of bacteriophage T4. Nature 227, 680–685 (1970).

46. Johnston, P. A., Jahn, R. & Südhof, T. C. Transmembrane topography andevolutionary conservation of synaptophysin. J. Biol. Chem. 264, 1268–1273(1989).

articles©

2002

Nat

ure

Pu

blis

hin

g G

rou

p

htt

p:/

/neu

rosc

i.nat

ure

.co

m

Classic studies of the neuromuscular junction identify the vesicleas the quantum of synaptic transmission1. The neuromuscularjunction contains many active zones, defined ultrastructurally aspatches of membrane-associated material studded with dockedsynaptic vesicles. Each active zone releases vesicles independent ofother active zones, and the amplitude distribution of postsynap-tic potentials can be described quantitatively by a binomialmodel1. Most glutamatergic synapses of the CNS, such as thosebetween hippocampal CA3 and CA1 pyramidal neurons(CA3–CA1 synapses), contain one active zone with several (2–20)docked vesicles, apposed to a single postsynaptic density2,3. Inresponse to an action potential (AP), each synapse releases neu-rotransmitter with a characteristic probability, and considerableheterogeneity exists between synapses in their release proper-ties4,5. If each docked vesicle were ready to fuse independently inresponse to an AP, then the simultaneous release of multiplequanta would occur on occasion. For example, if two dockedvesicles each released with a probability of 0.5, then the simulta-neous release of two vesicles would occur with a probability of0.25. Although some studies have found evidence for multi-quantal release6–9, most experiments, using a variety of approach-es, indicate that at most a single vesicle can be released in responseto an AP5,10–14. This ‘univesicular release rule’ stipulates that onereleased vesicle would rapidly (within microseconds) inhibit therelease of other docked vesicles within the same active zone12.Thus, determining if and under which conditions multivesicu-lar release can occur has important implications for under-standing the mechanisms of neurotransmitter release andsynaptic plasticity.

Testing the univesicular release rule requires the technical-ly difficult measurement of transmission at single synapses.One vesicle of neurotransmitter is thought to activate only afew postsynaptic receptors, producing small unitary cur-

Facilitation at single synapsesprobed with optical quantal analysis

Thomas G. Oertner, Bernardo L. Sabatini, Esther A. Nimchinsky and Karel Svoboda

Howard Hughes Medical Institute, Cold Spring Harbor Laboratory, 1 Bungtown Road, Cold Spring Harbor, New York 11724, USA

Correspondence should be addressed to K.S. ([email protected])


Many synapses can change their strength rapidly in a use-dependent manner, but the mechanismsof such short-term plasticity remain unknown. To understand these mechanisms, measurements ofneurotransmitter release at single synapses are required. We probed transmitter release byimaging transient increases in [Ca2+] mediated by synaptic N-methyl-D-aspartate receptors(NMDARs) in individual dendritic spines of CA1 pyramidal neurons in rat brain slices, enablingquantal analysis at single synapses. We found that changes in release probability, produced bypaired-pulse facilitation (PPF) or by manipulation of presynaptic adenosine receptors, wereassociated with changes in glutamate concentration in the synaptic cleft, indicating that singlesynapses can release a variable amount of glutamate per action potential. The relationshipbetween release probability and response size is consistent with a binomial model of vesicle releasewith several (>5) independent release sites per active zone, suggesting that multivesicular releasecontributes to facilitation at these synapses.

rents8,15,16. Most synapses are located within an extensive den-dritic tree and currents measured at the soma are highly fil-tered and attenuated8,17–19. In addition, it is difficult to ensurethat one is recording from a single synapse using electrophys-iological techniques alone8,17. To probe transmission at singlesynapses, we imaged NMDAR-mediated [Ca2+] accumulationsin individual spines20 as a measure of glutamate release. Pre-vious studies using serial section electron microscopy haveshown that most spines participate in only one excitatorysynapse3,21; therefore, NMDAR-mediated synaptic [Ca2+] tran-sients in single spines report release from single active zones.We found that the amount of glutamate released per actionpotential could be modulated with manipulations that changerelease probability. Our findings are quantitatively consistentwith multivesicular release from multiple independent releasesites at single active zones.

RESULTSTo probe transmission at single synapses, we imaged [Ca2+]accumulations in single spines produced by NMDAR currents20

using two-photon laser scanning microscopy22,23. SynapticNMDARs are far from saturated by glutamate release producedby a single AP20,24,25 and NMDAR activation is a quantitativemeasure of glutamate in the synaptic cleft. To monitor gluta-mate release optically, it was necessary to isolate NMDAR-mediated Ca2+ currents, as Ca2+ influx into spines can occurthrough multiple pathways26. Recorded neurons were voltage-clamped slightly above NMDAR reversal potential (+10 mV),inactivating voltage-sensitive Ca2+ channels20. At these depo-larized potentials, the Mg2+ block of NMDARs is relieved. AMPA(α-amino-3-hydroxy-5-methyl-4-isoxazole propionic acid)-typeglutamate receptors were blocked by the AMPA receptor(AMPAR) antagonist NBQX (10 µM). The remaining synaptic

articles


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

[Ca2+] transients were completely blocked by antagonists ofNMDARs (10 µM APV; 97 ± 1%, n = 5)20. We performed ourquantitative imaging in line-scan mode20,22,27, allowing detec-tion of quantal fluorescence transients with a sufficient signal-to-noise ratio to clearly distinguish trials in which neurotransmitterwas (successes) or was not (failures) released (Fig. 1). Synapticresponses could be monitored for 80–200 trials (Fig. 1d and f)and thus allowed the measurement of the probability of failure torelease neurotransmitter at a single synapse. We define the prob-ability of release (of one or more quanta) at a synapse in responseto a single stimulus as P′ = (1-failure probability). P′ varied great-ly across synapses (range, 0.18–0.62; mean, 0.33). The coeffi-cient of variation (c.v.) of the response amplitudes (c.v.= s.d./mean, excluding failures; Fig. 1f) was 0.37 (range,0.22–0.48), similar to the variability of miniature excitatory post-synaptic currents (EPSCs) in slices (c.v. = 0.42)28.

Many types of synapses show a type of short-term plasticitycalled paired-pulse facilitation (PPF)12,29. When two APs invadepresynaptic terminals in close succession (10–300 ms), theamount of transmitter released in response to the second AP isgreater, on average, than to the first. PPF is usually measured forpopulations of synapses and can be quantified asPPF = EPSC(tISI)/EPSC(0), where tISI is the interstimulus inter-val and EPSC(t) is the excitatory postsynaptic current measuredat the soma following a stimulus at time t. Under our experi-mental conditions, PPF peaked somewhere between 10–20 msand decreased with a decay time of ∼ 200 ms. Similar time cours-es were measured for currents dominated by AMPARs (Fig. 2a)and NMDARs (Fig. 2b and c).

To analyze the mechanisms of facili-tation, we imaged PPF of postsynaptic[Ca2+] transients at single synapses(PPFCa; Fig. 3). We distinguishedbetween the average response amplitude(R), where the average was computedover all trials including failures (the usualmeasure of synaptic strength), and theaverage response potency10 (r), where theaverage was computed over successesonly (Fig. 3c and d). PPFCa = R ′ ′ /R ′ ,where R′ and R′′ are the average responseamplitudes to the first and second pulse,respectively (Fig. 3a). The univesicularrelease rule implies that potency shouldnot change during plasticity, and there-fore PPFCa should be equal to the poten-tiation of release probabilities: PPFCa= R′′ /R′ = P′′ /P′, where P′ and P′′ are thesynaptic release probabilities to the firstand second pulse (see Methods). Alter-natively, facilitation could involve anincrease in potency, consistent with mul-tivesicular release. In this case, the ratioof synaptic release probabilities by itselfwould not account for the observed plas-ticity, and PPFCa > P′′ /P′ .

In our measurements, it was possible tosort responses to a pair of stimuli into fourcategories corresponding to the four pos-sible permutations of failures and success-es for each stimulus (Fig. 3c). Eachoutcome has an associated probability (pij),where the subscript i indicates failure (i =

0) or success (i = 1) in response to the first stimulus and the sub-script j indicates failure or success in response to the second stim-ulus (p00 + p01 + p10 + p11 = 1). We can compute synaptic releaseprobabilities as P′′ = (p01 + p11) and P′ = (p10 + p11). For allsynapses probed at tISI = 100 ms, PPFCa > P′′ /P′ (P < 0.01, n =8), suggesting that PPFCa was in part due to increases in potency.

A direct measure of potency changes during PPFCa is givenby comparing the potency of the first response (or the potencyin response to a single stimulus, r) to the potency of the secondresponse, given there was a failure on the first stimulus (r01; Fig. 3c). In this comparison, we used facilitated responses thatproduced release only on the second stimulus so that we couldexclude postsynaptic mechanisms of facilitation. The univesicu-lar release rule predicts the potency ratio (r01/r)* = 1 (* denotesa model prediction). In the example of Fig. 3, r01/r = 1.77 (Fig. 3d and f), which is inconsistent with the univesicular releaserule. For every synapse measured with tISI = 40 ms, we found thatr01/r was significantly (P < 0.05) larger than 1 (R′′ /R′ = 2.30± 0.27; r01/r = 1.53 ± 0.09; n = 6). Similar results held fortISI = 100 ms (R′′ /R′ = 1.95 ± 0.22; r01/r = 1.33 ± 0.07; n = 8) andtISI = 250 ms (R′′ /R′ = 1.64 ± 0.19; r01/r = 1.17 ± 0.06; n = 6; Fig. 4d). These data show that synaptic potency is plastic andhence that the amount of glutamate released by a single actionpotential can be modulated.

Can increased potency during facilitation be explained bythe existence of multiple independent release sites at singleactive zones? If multivesicular release from several sites at thesame active zone occurs, potency and release probability willchange together in a predictable manner. Each active zone con-

10

0

Time (ms)

Flu

ores

cenc

e (a

u)

c

Time (s)

Rat

io (

G/R

)

6004002000

2.0

1.5

1.0

0.5

0.0

e

Ratio (G/R)

2.01.51.00.50.0

Num

ber

f

d

50 ms

G/R0.5

g

a

b

1 mµ

20

10

00 200100

50 ms10 pA

+ 40 mV+ 10 mV

114

12

8

4

0

∆

Fig. 1. Measurement of NMDAR-mediated [Ca2+] transients in single spines. (a) Left, dendritewith several spines (red fluorescence) and right, [Ca2+] transient after synaptic stimulation (greenfluorescence, ∆G). White line indicates position of the line scan. (b) Line scans across spine head(total duration, 450 ms). White triangles indicate time of synaptic stimulation. Red fluorescencedid not change (left), whereas green fluorescence increased rapidly in the spine after synapticstimulation (right). A weak and delayed increase in [Ca2+] due to Ca2+ diffusion is apparent in thedendrite. (c) Time course of fluorescence intensity in the spine head in the [Ca2+]-insensitive(red) and [Ca2+]-sensitive (green) fluorescence channels (single trial, same data as in b). (d) Multiple responses to synaptic stimulation with single pulses (130 trials). Failures of neuro-transmitter release can be clearly distinguished from successes. (e) Response amplitudes overtime. Response amplitudes, failure rates, and resting fluorescence (corresponding to resting[Ca2+]i) were stable (same data as in d); response amplitudes were averaged in a 40 ms windowstarting 50 ms after stimulation (horizontal bar at bottom of d). (f) Histogram of response ampli-tudes. (g) EPSC measured in the soma at nominal holding potentials of +10 mV (black) and +40 mV (gray). The initial fast transient is the stimulus artifact.


articles©

2002

Nat

ure

Pu

blis

hin

g G

rou

p

htt

p:/

/neu

rosc

i.nat

ure

.co

m

(Fig. 4c and d). PPFCa was larger than PPF, as probed withNMDAR-mediated currents (2.30 ± 0.27 versus 1.73 ± 0.17,tISI = 40 ms, n = 6). A quantitative comparison of facilitationamplitudes is complicated by differences in measurement tech-niques: different populations of synapses contribute to theresponses evoked by the first and second stimulus in electro-physiological measurements, whereas the same single synapse isprobed on both trials in imaging measurements.

If single active zones do contain multiple independent releasesites, then pharmacological manipulation of release probabili-ty should also change synaptic potency. At CA1 hippocampal

Fig. 3. Paired-pulse facilitation at a single synapse.(a) Average [Ca2+] transient in response to a singlestimulus and pair of stimuli at tISI = 40 ms. To quan-tify PPFCa, peak amplitudes (R′ and R′′ ) were mea-sured as indicated by the arrows. (b) Responses to45 paired-pulse stimuli. Failures of synaptic trans-mission can be clearly distinguished from successes.(c) The four possible outcomes resulting frompaired-pulse stimulation with their associated prob-abilities (p11, p10, p01, p00). Also indicated is thepotency in response to the first stimulus (r) and thepotency in response to the second stimulus giventhat there was a failure on the first (r01). (d) Timecourse of the success amplitude to a single stimulus(yellow) and to the second stimulus in a pair wherethe first produced a failure (green). Also shown isthe failure to both stimuli (black). (e) Responseamplitudes (yellow = r, green = r01, black = failures,same data as in d). (f) Paired-pulse facilitation of theaverage response (gray) was in part due to facilita-tion in potency (green).

tains several docked vesicles that seem poised for release (∼ 10;equivalent to the readily releasable pool)4. Replenishing thereadily releasable pool takes several seconds4, and the numberof docked vesicles (D) is not likely to change over the durationof a cycle in a PPF experiment unless release occurs. If eachdocked vesicle acts as a release site, fusing independently withprobability p after an AP, then, according to the binomialmodel1, the synaptic release probability in response to the firstpulse is P′ = 1 – (1 – p)D. Owing to facilitation, after the firstpulse the release probability per vesicle is enhanced by a fac-tor α to αp (α > 1). The release probability for the synapse onthe second pulse, given that there was a failure of release to thefirst pulse, is then p01 = 1 – (1 – αp)D. We measured both P′and p01 directly (Fig. 3b and c) and used them to calculate anexpected potency facilitation (r01/r)* = (P ′ – P ′ (1–p01))1/D)/(p01 – p01 (1 – P′))1/D)h (see equation (2), Methods).In the equation above, h is the Hill coefficient for synapses,relating NMDAR activation and the number of vesicles released(range of h, 1–1.4; see Methods). As r and r01 could also bemeasured, we compared the expected potency facilitation(r01/r)* with the measured potency facilitation, r01/r (Fig. 4b).This analysis showed that our measurements were compatiblewith the existence of several independent release sites per activezone and clearly inconsistent with the univesicular release rule(Fig. 4b). Postulating five release sites gives a satisfactory fit tothe data, consistent with the number of docked vesicles report-ed to exist in CA1 active zones3.

How do our single-synapse measurements of PPFCa compareto PPF averaged over multiple synapses and measured at thesoma? The time courses of PPFCa and PPF were indistinguish-able, suggesting that the mechanisms of plasticity are shared

articles


a

b

50 pA

100 ms

c

tISI (ms)

3.0

2.5

2.0

1.5

1.0

3002001000

PP

F

Fig. 2. Paired-pulse facilitation of excitatory postsynaptic currents. (a) EPSCs recorded at –70 mV for interstimulus intervals (tISI)of 20–300 ms (average of six trials). (b) EPSCs recorded in the same cellat +40 mV. (c) Paired-pulse facilitation (PPF) of EPSCs dominated byAMPAR (gray) or NMDAR (black) mediated currents. Inset, AMPAR-mediated current amplitudes were estimated as the peaks of EPSCs at–70 mV, whereas NMDAR current amplitudes were estimated 40 msafter stimulation at +40 mV. For all measurements, the response to asingle stimulus was subtracted from the response to the second pulse.

RR

"'

rr01

3

2

1

c

r

r01

p11

p10

p01

p00

e f

a

50 ms

G/R 0.05

R''

R'

b

50 ms

G/R 0.2

d

50 ms

G/R 0.1

15

10

5

01.00.80.60.40.2

Num

ber

Rat

io

Ratio (G/R)

∆∆

∆

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

synapses, activation of presynaptic A1 adenosine receptors leadsto a reduction of transmitter release30,31. Using [Ca2+] imagingin individual spines, we measured the synaptic release proba-bility, P′, and potency, r. After a baseline period, we applied 2-chloroadenosine (1–5 µM), an agonist of A1 receptors. In eightof nine synapses studied, both P′ and r decreased significantly,and in one synapse, 2-chloroadenosine did not induce a changein either P′ or r (Fig. 5c, red arrows). In some experiments,washout of 2-chloroadenosine partially reversed the reductionof P′, and was associated with a concomitant increase in r (Fig.5c, blue arrows, n = 2). At a concentration of 20 µM, the specificA1 antagonist DPCPX caused corresponding changes in theopposite direction (Fig. 5c, black arrows, n = 3). The findingthat P′ and r consistently changed together is consistent withthe existence of multiple release sites in an active zone.

Although our experiments support the idea of multiplerelease sites per active zone, some other mechanisms also couldexplain our data. As glutamate receptors are not saturated20,24,25,any mechanism that can change the amplitude or time course ofglutamate concentration in the cleft would change the occu-pancy of receptors and the amplitudes of postsynaptic respons-es. In particular, it has been suggested that glutamate releasefrom small vesicles can occur in two modes: the classic all-or-none exocytosis and a graded mode in which glutamate diffus-es through a transient fusion pore32. A use-dependent switchfrom graded mode to all-or-none mode would cause potencypotentiation and could, in principle, account for our data.

Fig. 4. Dissecting PPF in individual synapses. (a) Normalizedpotency plotted against a measure of release probability(tISI = 40 ms, n = 6; tISI = 100 ms, n = 8; tISI = 250 ms, n = 6).Arrows connect points corresponding to the control response(P′ , 1) with points corresponding to the facilitated response(p01/(p01+p00), r01/r). The x-value p01/(p01+p00) is the releaseprobability for the second stimulus given that there was a failureof release on the first. In all but one synapse, increasing releaseprobabilities were correlated with increasing potency (arrowswith positive slope). (b) Measured potency ratio compared topredicted potency ratio for the univesicular release rule (horizon-tal line) and a binomial model with five independent release sites(dotted line with slope = 1, see Methods). Black symbols showthe prediction assuming a Hill coefficient h = 1, red symbols for h= 1.4. Shapes indicate different interstimulus intervals: tISI = 40 ms(circles), 100 ms (squares) and 250 ms (triangles). Filled symbolsmark spines in which the measured r01/r was significantly largerthan 1 (Wilcoxon two-sample test, P < 0.05). (c) The time coursewas similar for PPFCa and PPFi. Values are normalized to PPF attISI = 40 ms. (d) Potency ratio (white bars) and total PPFCa (blackbars) as a function of tISI.


articles

d

1

1.5

2

2.5

40 100 250

RR

"'

rr01

Release probability

Pot

ency

rat

io

/r

r01

a

1.0

1.2

1.4

1.6

1.8

0.2 0.4 0.6 0.8 10

c

tISI (ms) tISI (ms)

PP

F(

)/P

PF

(40

ms)

t ISI R

atio

1.5

1.0

0.5

0.03002001000

AMPA EPSCs NMDA EPSCs [Ca ]2+

Predicted potency ratio

b

Pot

ency

rat

io

/r

r01 1.8

1.6

1.4

1.2

1.0

2.22.01.81.61.41.21.0

2.2

2.0

Fig. 5. Varying release probability pharmacologically changes potency.(a) Calcium transients from a single spine. After 7 min, 10 µM 2-Chloroadenosine was added to the bath. The size of the average successamplitude (black curve) decreased markedly (circles, successes; dia-monds, failures). (b) Average release probability of the experimentabove, calculated in an 11-point sliding window. Pr changed from 0.59(control) to 0.28 (adenosine), indicated by the black horizontal lines. (c) Summary data of 13 spines in different slices treated with 2-Chloroadenosine (2–10 µM, red arrows), with 2-Chloroadenosine andsubsequent washout (blue arrows) or with the A1-antagonist DPCPX(20 µM, black arrows; 5 µM, gray arrows). Filled arrowheads indicate asignificant change in potency (Wilcoxon two-sample test, P < 0.05). Theexample shown in (a) and (b) is marked with an asterisk.

A testable prediction of this model is that the glutamate tran-sient after graded exocytosis will be smaller and longer-lastingthan it is during all-or-none exocytosis32. This implies that therise time of the NMDAR current would be slower33, resulting ina slower rise of the spine [Ca2+] transient, which is proportionalto the integral of the Ca2+ current under our measurement con-ditions34,35. We found, however, that the rise times of [Ca2+] tran-sients in response to single stimuli were indistinguishable fromthose in response to the second of a pair of stimuli (Fig. 6). Thesefindings indicate that under different conditions of release, char-

0.6

0.4

0.2

0.0

2-Chloroadenosine

0 2 4 6 8 10 12Time (min)

Res

pons

e am

p. (

G/R

)a

b 1.0

0.5

0.00 2 4 6 8 10 12

Time (min)

Rel

ease

pro

b.

c

0.6

0.8

1

1.2

1.4

Rel

ativ

e po

tenc

y

0 0.2 0.4 0.6 0.8 1

Release probability

∆

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

acterized by different synaptic release probabilities, cleft gluta-mate transients have the same time course but different ampli-tudes. We conclude that fusion pore modulation is unlikely toaccount for our data.

Using spine Ca2+ accumulation as a reporter of released glu-tamate requires that the Ca2+-sensitive dye is not saturated afterthe release of a single vesicle. Three lines of evidence suggestthat under our experimental conditions, the fluorescence sig-nal was approximately proportional to NMDAR activation.First, because of the large added buffer capacity for Ca2+ pro-vided by the indicator, [Ca2+] accumulations were reduced ∼ 30-fold compared to native conditions35. Based on quantitativemeasurements of NMDAR-mediated [Ca2+] accumulations, theamplitudes of these buffered transients are expected to be onthe order of ∆[Ca2+]syn ≈ 200–300 nM35, which is in the lin-ear regime of our indicator (∆[Ca2+]syn < Kd = 785 nM). Sec-ond, the changes in fluorescence after depolarizing the neuronsfrom –70 to +10 mV were ∼ 1.7 times larger than those inresponse to synaptic stimuli, again indicating that synaptic[Ca2+] transients were far from indicator saturation. Third, inseveral experiments (n = 8), we explicitly tested for linearity(Fig. 7). If the Ca2+ indicator were close to saturation, larger[Ca2+] accumulations would have resulted in relatively com-pressed fluorescence responses. Consequently, our measure-ments of PPFCa and potency ratios (r01/r) would have beenunderestimates of the true values. To address this issue, mea-surements were interleaved between holding potentials close toreversal for the NMDAR (+10 mV) and potentials closer toreversal for Ca2+ (+40 mV). Consistent with the changes in dri-ving force for Ca2+, responses were smaller at the higher hold-ing potentials, r(+40 mV)/r(+10 mV) = 0.56, and were thereforeexpected to be more linear in the case of saturation. However,r01/r was identical at both holding potentials: r01/r (+10 mV) =1.33 ± 0.07, r01/r (+40 mV) = 1.26 ± 0.07 (Fig. 7b). We con-clude that spine [Ca2+] is proportional to NMDAR activation.

DISCUSSIONWe imaged NMDAR-mediated [Ca2+] transients in single spinesevoked by synaptic stimulation. As virtually all spines in the CA1region of hippocampus form only one synapse3,21, [Ca2+] imag-ing allowed us to measure synaptic release probability and poten-cy (the average amplitude of the postsynaptic response to a

successful synaptic transmission) at individual synapses, gener-ating data for quantal analysis at single CA3–CA1 synapses. Usingactivity-dependent and pharmacological modulation of releaseprobability, we found that potency increased with release proba-bility. These data imply that single active zones are capable ofreleasing a variable amount of glutamate per action potential.

Several technical issues, however, complicate the interpre-tation of our experiments. Under some conditions, glutamatereleased from a neighboring, non-imaged synapse can diffuseto postsynaptic receptors at the imaged synapse(‘spillover’)36–38 and contribute to potency facilitation (r01/r >1). We consider this mechanism unlikely because we used stim-uli that activate only a tiny subpopulation of synapses (<5%),implying that the probability of activating multiple varicosi-ties within diffusion distance was small36. Studies specificallydesigned to detect correlations in the response amplitudes ofneighboring spines do not show evidence of spillover20 (E. A.N. & K. S., Soc. Neurosci. Abstr. 27, 155.2, 2001). Consistentwith this finding, published evidence for spillover was collect-ed primarily at low temperatures38 or in the presence of block-ers of astrocytic glutamate transporters36,37, not under ourexperimental conditions. Furthermore, in our experiments,double-failure trials (r00) were indistinguishable from baseline(Fig. 3d), inconsistent with spillover from other active synaps-es that were not imaged.

Even spillover too weak to activate NMDARs, however, couldleave its mark by producing receptors singly bound by glutamate.This glutamate would sum with direct release of glutamate topotentiate NMDAR activation. For two reasons, this spillovermodel cannot account for our data. First, the potency facilita-tion produced in this situation is, at most, 1 + 2∆Pdα, where ∆Pdis the change in the release probability of the non-imaged (‘dark’)synapse, and α is the fraction of glutamate escaping to a neigh-boring synapse (see Methods for derivation). Even assuming alarge α = 0.1 (refs. 39 and 40), one would expect potency facil-itation in the range of 0–15%, which is much smaller than whatwe measured (0–80%, Fig. 4b). Second, in the spillover model,potency facilitation is expected to be independent of release prob-ability and PPF at the imaged synapse (equation (3), Methods).This is inconsistent with the correlation between measured poten-cy ratio and predicted potency ratio (a function of release prob-abilities at the imaged synapse, Fig. 4b). Thus, spillover and

articles


0

0.2

0.4

0.6

0.8

1

1.2

40 100 250 ms

Slo

pe (

% o

f con

trol

)

b

50 ms

Scaled control

Potentiated

aFig. 6. Rise times of [Ca2+] transients under control and potentiatedconditions. (a) Overlay of peak-normalized control (black) and potenti-ated (red) [Ca2+] transients from a single spine. The initial slope of the[Ca2+] transients (dashed) was used as a measure of rise time. Timecourse of the [Ca2+] transients was not changed in the potentiatedresponses (tISI = 40 ms, 30% potency potentiation). (b) Summary datafor all synapses. The slopes of the potentiated responses were not dif-ferent from those of controls for all interstimulus intervals (Wilcoxonmatched-pair signed-rank test, n = 20).

b

0.0

0.5

1.0

1.5

Pot

ency

rat

io

/r

r01

+10 +40 mV

a100 ms

G/R 0.2

r

r

r01

r01

+40 mV

+10 mV

∆

Fig. 7. Change in potency is constant as a function of holding poten-tial. (a) Average success to the first stimulus (r), average success tothe second stimulus given a failure to the first (r01) and average failureat +10 mV (black curves) and at +40 mV (red curves) for a singlesynapse. Trials at different holding potentials were interleaved, tISI =100 ms. (b) Summary data for eight synapses showing that thepotency ratio was the same at both holding potentials and significantly(P < 0.01) larger than 1.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


pooling of glutamate can only account for a small fraction of thepotency facilitation we observed.

Release of Ca2+ from intracellular stores could also con-tribute to postsynaptic [Ca2+] signals under certain condi-tions41 and confound our interpretation of spine [Ca2+] as ameasure of NMDAR activation. However, three lines of evi-dence argue against this possibility. First, release from storeshas not been observed under experimental conditions that aresimilar to ours20,42. Second, most spines do not contain endo-plasmic reticulum (ER)43, so release from stores would mostlikely be initiated in the dendritic shaft, but in our experimentsspines were activated without concomitant dendritic [Ca2+]signals (Fig. 1a and b). Third, [Ca2+] accumulations variedwith holding potential as expected for Ca2+ influx throughNMDARs (Fig. 7). These observations led us to conclude thatintracellular stores did not contribute to the [Ca2+] signal.

At CA3–CA1 synapses, successful glutamate release inresponse to low-frequency stimulation opens only a fewNMDAR receptors on average (2–5 receptors, unpub. observ.).Therefore, failures of postsynaptic [Ca2+] transients could, inrare instances, be due to receptor failures after transmitterrelease. On arrival of the second pulse, some NMDA receptorscould still be singly bound by glutamate without opening33,44.An upper bound on the potency facilitation that could becaused by this mechanism can be estimated as 3.3% for a 40-msISI, 2.5% for a 100-ms ISI, and 1% for a 250-ms ISI—not suf-ficient to account for our data. Furthermore, experiments usingpharmacological manipulations of release probability (Fig. 5)were not affected by the possibility of receptor failures.

Our results indicate that, at higher release probabilities,more glutamate is released per action potential. As neuro-transmitter is released in quanta corresponding to individualvesicles, and as single active zones contain multiple dockedvesicles, the most parsimonious explanation of our data is thatmultiple vesicles can be released in response to a single AP andthe probability of multivesicular release increases with synap-tic release probability. The monotonic relationship betweenrelease probability and potency that we found (Figs. 4b and5c) is in quantitative agreement with a binomial model thatincludes several independent release sites. We suggest that theserelease sites correspond to the docked vesicles that are seenunder the electron microscope2.

Some studies show, using minimal stimulation, that potencyremains unchanged under conditions that modulate release, con-sistent with the univesicular release rule5,10. Differences betweenthese studies and our imaging experiments may be due to dif-ferences in preparation (neonatal versus juvenile slice) or toexperimental conditions (room temperature versus 34°C). Anoth-er possibility is that different methods (electrophysiology versusimaging) select for different types of synapses. For example, ourmeasurements probably excluded the smallest spines, presum-ably corresponding to the smallest synapses. Owing to the strongdendritic filtering of currents arising from distant synapses8,18,failure analysis using somatic patch-clamp recordings probablyselects for large synapses close to the soma. Resolving the dis-crepancy between these electrophysiological measurements andour optical measurements will require further studies that per-haps apply both types of analysis to the same synapse.

According to the binomial model, if the release probability(Pr) of a small synapse is low, most successful transmissions willbe due to the release of a single vesicle (for example, 92% ofreleases are single-vesicle releases for a synapse with five releasesites and Pr = 0.2). As Pr rises, the fraction of multivesicular

events increases: for the same synapse at P = 0.6, 67% of releas-es are univesicular and 33% are multivesicular. The finding thatthe spine Ca2+ transients reflect the predicted changes in cleftglutamate concentration confirms that NMDA receptors are farfrom saturation after the release of a single vesicle20. At eachsynapse, the coupling between changes in synaptic release prob-ability and potency will result in larger unitary synaptic cur-rents during high-frequency stimulation (bursts) and smallerunitary currents in response to isolated APs. The potentiatedCa2+ accumulations observed during facilitation (Fig. 3d) mayselectively trigger some kinds of use-dependent postsynapticplasticity45. The possibility of multivesicular release implies thateven synapses with reliable responses to a single AP (high synap-tic release probability) could still increase the number ofreleased vesicles per AP, thus expanding the effective dynamicrange for facilitation. Finally, it is possible that univesicularrelease at some synapses predominantly activates NMDARs andmultivesicular release at other synapses activates lower-affini-ty AMPARs. This is another possible presynaptic explana-tion32,46 for a subset of silent synapses47,48.

METHODSPreparation and electrophysiology. Horizontal hippocampal slices (350 µm thick) were prepared from Wistar rats 16–19 days old in accor-dance with the animal care and use guidelines of Cold Spring HarborLaboratory, using a chilled cutting solution containing 110 mM cholinechloride, 25 mM NaHCO3, 25 mM D-glucose, 11.6 mM sodium ascor-bate, 7 mM MgSO4, 3.1 mM sodium pyruvate, 2.5 mM KCl, 1.25 mMNaH2PO4 and 0.5 mM CaCl2. Slices were incubated in gassed (95% O2and 5% CO2) physiological saline (127 mM NaCl, 25 mM NaHCO3,25 mM D-glucose, 2.5 mM KCl, 1.5 mM MgCl2, 1.5 mM CaCl2 and1.25 mM NaH2PO4) at 34°C for 30–45 min and then at room temper-ature until used. Experiments were done at 34°C in physiological salinecontaining 0.01 mM NBQX, 0.01 mM bicuculline and 0.01 mM serine(Sigma, St. Louis, Missouri). Whole-cell patch electrodes (3–6 MΩ)contained 135 mM CsMeSO3, 10 mM HEPES, 10 mM sodium phos-phocreatine, 5 mM glutathione, 4 mM MgCl2, 4 mM Na2-ATP, 0.4 mMNa-GTP, 0.6 mM Fluo5F and 0.04 mM Alexa Fluor 594 (MolecularProbes, Eugene, Oregon). Cells were depolarized to +10 mV to relievethe Mg2+ block of NMDARs and to inactivate voltage-sensitive Ca2+

channels. Synaptic transmission was evoked by short current pulsesdelivered with a glass pipette (2–3 µm tip)20. Paired pulses and singlepulses were alternated every five seconds.

Two-photon imaging. We used a custom built two-photon laser scan-ning microscope49 consisting of a Ti:sapphire laser (Mira, Coherent,Santa Clara, California) tuned to λ ∼ 810 nm, a 63× 0.9NA Objective(Olympus, Melville, New York) and a Zeiss scan lens (Zeiss, Thorn-wood, New York). Fluorescence was detected in epifluorescence andtransfluorescence (through an oil-immersion condenser, Zeiss, NA = 1.4) modes using photomultiplier tubes (R3896, Hamamatsu,Hamamatsu City, Japan). Image acquisition was controlled by customsoftware written in Matlab (MathWorks, Natick, Massachusetts). Inthe transfluorescence pathway, a 565-nm dichroic mirror was used toseparate green and red fluorescence. BG22-colored glass filters and607/45 barrier filters were placed respectively in the ‘green’ (shorterwavelength) and ‘red’ (longer wavelength) pathways to eliminate trans-mitted or reflected excitation light. (All filters and dichroic mirrorswere from Chroma, Battleboro, Vermont.) Neurons were filled throughthe patch electrode for more than 15 min before imaging. To measurea fluorescence signal proportional to [Ca2+], we used large concen-trations (600 µM) of a medium-affinity (Kd = 785 nM under phys-iological conditions, data not shown) Ca2+ indicator, Fluo5F(Molecular Probes), detected as green fluorescence. Fluo5F is too dimat rest to reliably image spines, so we added a Ca2+-insensitive fluo-rophore (Alexa Fluor 594) to the pipette solution, detected in the redchannel (Fig. 1 a–c). Stimulated synapses on higher-order apical den-drites (70–380 µm from the soma) were identified in frame scans with

articles©

2002

Nat

ure

Pu

blis

hin

g G

rou

p

htt

p:/

/neu

rosc

i.nat

ure

.co

m

an on-line analysis program written in IGOR (Wavemetrics, LakeOswego, Oregon) (Fig. 1a). In the paired-pulse experiments, the acti-vated spine was centered in a 15 × 15 µm window and probed usingline scans with a temporal resolution of 2 ms (Fig. 1b). In the phar-macological experiments, where high temporal resolution was not arequirement, four frame scans (64 × 64 pixel) were acquired in eachtrial (two before and two after the stimulus). Only synapses in whichstable responses from well-isolated Ca2+ sources could be verified wereanalyzed. Spine volumes were in the range from 0.032 to 0.183 µm3,mean volume 0.081 µm3 (n = 21, collected in a separate set of experi-ments using identical selection criteria).

Data analysis. As the resting fluorescence of Fluo5F is very low, calcu-lating ∆F/F introduces large errors due to shot noise. Thus we used theratio of green:red fluorescence intensity, ∆G/R; this measure is inde-pendent of absolute dye concentration and robust to movement arti-facts. To avoid problems with background subtraction, care was takento avoid ejecting dye from the pipette into the slice35. Cells in whichthe surrounding slice showed detectable background fluorescence afterdye loading were not used. The only background correction that wasdone was subtraction of the photomultiplier tube dark current. Thethreshold for detection of successes was set individually for each spinein a plot of all individual trials from that spine (Fig. 3b). Only spinesthat showed a clear separation between failures and successes wereincluded in the analysis. The sorting threshold was set in the middle ofthe gap between failure and success trials. First, successes to the firststimulus (r1, which may have failed or succeeded to the second stimu-lus) were detected, removed from the plot, and averaged separately.Second, successes to the second stimulus (r01) were separated fromcomplete failures of transmission (r00) and averaged separately (Fig. 3d). Tests of significance used the Wilcoxon two-sample test unlessotherwise noted. All measurements are given as mean ± s.e.m.

Quantal analysis of short-term synaptic plasticity. The number of suc-cesses divided by the number of trials is the probability of release for asynapse, P′, in response to a single pulse (or the first pulse in a pair). Letγ be the fluorescence signal produced by the release of a single vesicle.The success amplitude to a single stimulus is the potency r = γnh, wheren is the average number of vesicles released on success trials only and h isthe effective Hill coefficient, describing the response of synaptic recep-tors as a function of the number of vesicles released. If glutamate releasedfrom different vesicles interacts with distinct subsets of receptors, thenh = 1. If glutamate from different vesicles interact with the same popu-lation of receptors, then h = 1.4, the Hill coefficient for the concentra-tion–response curves for NMDARs50.

In a paired-pulse experiment, responses were sorted into (i) completefailures of transmission (with probability p00), (ii) responses to the sec-ond stimulus only (p01) and (iii) responses to the first stimulus and eitherfailure (p10) or success (p11) to the second stimulus. Then the averagefluorescence response to the first pulse, including failures, is R′ =(p11 + p10)r. In general, in response to a pair of stimuli, the potency ofthe second pulse may vary depending on whether there was release onthe first pulse. We denote as r11 and r01 the potencies of the second pulsegiven that release did (r11) or did not (r01) occur on the first pulse. Theaverage response to the second pulse in a pair is then R′′ = p11r11 + p01r01.The PPF is the average response to the second pulse relative to the aver-age response to the first pulse, PPF = R′′ /R′ = (p11r11 + p01 r01)/(p11 +p10)r. In the special case where at most one vesicle can be released perAP (univesicular release rule), r = r01 = r11, and PPF is simply equal tothe relative change in release probability on the second and first pulse:PPF = P′′ /P′ = (p01 – p10)/(p11 + p10).

Models of short-term synaptic plasticity at single synapses. We comparedthe number of vesicles released by successes on pulse one (n) and pulse two,under the condition that there was no release on the first pulse (n01). Theseare related to potencies as r = γnh and r01 = γn01

h. The potencies can be mea-sured (Fig. 3) together with their associated release probabilities P′ and p01,respectively. In a simple model of release, an active zone contains D dockedvesicles where each vesicle can release independently with a release proba-bility p. Thus, each docked vesicle serves as an independent release site.Conditions of PPF increase the vesicle release probability to αp, where

α > 1. The number of vesicles released by successes is then as follows:

This yields the following ratio:

(1)

According to the binomial model, the release probabilities are P′ = 1 –(1 – p)D and p01 = 1 – (1 – αp)D . We can compute the number of releasedvesicles by solving these expressions for α in terms of measurable quan-tities and inserting into equation (1):

(2)

To relate the potencies to the ratio of vesicles released, we use the fol-lowing relationship:

Potency facilitation due to spillover. Could spillover explain thepotency facilitation we observed (Fig. 4b)? We considered the situationwhen spillover occurs between two nearby synapses: the ‘imaged’synapse (i) and the non-imaged ‘dark’ synapse (d). Spillover is char-acterized by the parameter α , the fraction of glutamate released atone synapse that reaches a neighboring synapse. In previous mea-surements, we saw clear and complete transmission failures, and noresponse correlations between neighboring spines20 (E. A. N. & K. S.,Soc. Neurosci. Abstr., 27, 155.2, 2001), so we assumed that spilloverof glutamate alone is not sufficient to produce NMDAR activation,but rather produces a population of singly bound receptors. Spilloverof glutamate combines, however, with directly released glutamate toproduce potency facilitation. Seeking to derive a worst-case scenario,we used a Hill coefficient of 2. Then the potency at the imagedsynapse, to first order in α , is the following:

r ≈ γc2(1 + 2Pdα)

Here, c is an effective concentration of glutamate in the cleft, and γ is aconstant. Note that r is independent of the release probability at theimaged synapse, Pi. The potency facilitation follows:

(3)

AcknowledgmentsWe thank Z. Mainen, R. Malinow, M. Maravall and R. Yasuda for comments on themanuscript, and T. Pologruto for software development. This work was support-ed by grants from the Swartz Initiative for Computational Neuroscience (to T.G.O.),the Helen Hay Whitney Foundation (to B.L.S.), the Pew and Mathers Founda-tions and the National Institutes of Health (NIH).


RECEIVED 18 MARCH; ACCEPTED 8 MAY 2002

1. Katz, B. The Release of Neural Transmitter Substances (Thomas, Springfield,Illinois, 1969).

2. Harris, K. M. & Sultan, P. Variation in the number, location and size ofsynaptic vesicles provides an anatomical basis for the nonuniform probability

≈ (1 + 2(Pd" – Pd')α)r01

r

=r01

r

n01

n

h

=n01

n

P' – P'(1 – p01)1/D

p01 – p01(1 – P')1/D

= αn01

n

P'

p01

n = and n01 =pD

P'

αpD

p01

articles


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


of release at hippocampal CA1 synapses. Neuropharmacology 34, 1387–1395(1995).

3. Schikorski, T. & Stevens, C. F. Quantitative ultrastructural analysis ofhippocampal excitatory synapses. J. Neurosci. 17, 5858–5867 (1997).

4. Dobrunz, L. E. & Stevens, C. F. Heterogeneity of release probability,facilitation, and depletion at central synapses. Neuron 18, 995–1008(1997).

5. Hanse, E. & Gustafsson, B. Quantal variability at glutamatergic synapsesin area CA1 of the rat neonatal hippocampus. J. Physiol. 531, 467–480(2001).

6. Auger, C., Kondo, S. & Marty, A. Multivesicular release at single functionalsynaptic sites in cerebellar stellate and basket cells. J. Neurosci. 18, 4532–4547(1998).

7. Tong, G. & Jahr, C. E. Multivesicular release from excitatory synapses ofcultured hippocampal neurons. Neuron 12, 51–59 (1994).

8. Isaac, J. T. et al. An investigation of the expression mechanism of LTP ofAMPA receptor–mediated synaptic transmission at hippocampal CA1synapses using failures analysis and dendritic recordings. Neuropharmacology37, 1399–1410 (1998).

9. Wadiche, J. I. & Jahr, C. E. Multivesicular release at climbing fiber-purkinjecell synapses. Neuron 32, 301–313 (2001).

10. Stevens, C. F. & Wang, Y. Facilitation and depression at single centralsynapses. Neuron 14, 795–802 (1995).

11. Redman, S. Quantal analysis of synaptic potentials in neurons of the centralnervous system. Physiol. Rev. 70, 165–198 (1990).

12. Regehr, W. G. & Stevens, C. F. in Synapses (eds. Cowan, W. M., Sudhof, T. C.& Stevens, C. F.) 135–175 (Johns Hopkins Univ. Press, Baltimore, 2001).

13. Auger, C. & Marty, A. Quantal currents at single-site central synapses. J. Physiol. 526, 3–11 (2000).

14. Korn, H., Triller, A., Mallet, A. & Faber, D. S. Fluctuating responses at acentral synapse: n of binomial fit predicts number of stained presynapticboutons. Science 213, 898–901 (1981).

15. Spruston, N., Jonas, P. & Sakmann, B. Dendritic glutamate receptor channelsin rat hippocampal CA3 and CA1 pyramidal neurons. J. Physiol. 482,325–352 (1995).

16. Faber, D. S., Young, W. S., Legendre, P. & Korn, H. Intrinsic quantalvariability due to stochastic properties of receptor–transmitter interactions.Science 258, 1494–1498 (1992).

17. Magee, J. C. & Cook, E. P. Somatic EPSP amplitude is independent of synapselocation in hippocampal pyramidal neurons. Nat. Neurosci. 3, 895–903(2000).

18. Stuart, G. & Spruston, N. Determinants of voltage attenuation in neocorticalpyramidal neuron dendrites. J. Neurosci. 18, 3501–3510 (1998).

19. Williams, S. R. & Stuart, G. J. Dependence of EPSP efficacy on synapselocation in neocortical pyramidal neurons. Science 295, 1907–1910 (2002).

20. Mainen, Z. F., Malinow, R. & Svoboda, K. Synaptic calcium transients insingle spines indicate that NMDA receptors are not saturated. Nature 399,151–155 (1999).

21. Harris, K. M. & Stevens, J. K. Dendritic spines of CA1 pyramidal cells in therat hippocampus: serial electron microscopy with reference to theirbiophysical characteristics. J. Neurosci. 9, 2982–2997 (1989).

22. Yuste, R. & Denk, W. Dendritic spines as basic functional units of neuronalintegration. Nature 375, 682–684 (1995).

23. Denk, W. & Svoboda, K. Photon upmanship: why multiphoton imaging ismore than a gimmick. Neuron 18, 351–357 (1997).

24. McAllister, A. K. & Stevens, C. F. Nonsaturation of AMPA and NMDAreceptors at hippocampal synapses. Proc. Natl. Acad. Sci. USA 97, 6173–6178(2000).

25. Umemiya, M., Senda, M. & Murphy, T. H. Behavior of NMDA and AMPAreceptor-mediated miniature EPSCs at rat cortical neuron synapses identifiedby calcium imaging. J. Physiol. 521, 113–122 (1999).

26. Sabatini, B. L., Maravall, M. & Svoboda, K. Ca2+ signaling in dendritic spines.Curr. Opin. Neurobiol. 11, 349–356 (2001).

27. Mainen, Z. F. et al. Two-photon imaging in living brain slices. Methods 18,231–239 (1999).

28. Bekkers, J. M., Richerson, G. B. & Stevens, C. F. Origins of variability inquantal size in cultured hippocampal neurons and hippocampal slices. Proc.Natl. Acad. Sci. USA 87, 5359–5362 (1990).

29. Zucker, R. S. & Regehr, W. G. Short-term synaptic plasticity. Annu. Rev.Neurosci. 12, 13–31 (1989).

30. Wu, L. G. & Saggau, P. Adenosine inhibits evoked synaptic transmissionprimarily by reducing presynaptic calcium influx in area CA1 ofhippocampus. Neuron 12, 1139–1148 (1994).

31. Wu, L. G. & Saggau, P. Presynaptic inhibition of elicited neurotransmitterrelease. Trends Neurosci. 20, 204–212 (1997).

32. Choi, S., Klingauf, J. & Tsien, R. W. Postfusional regulation of cleft glutamateconcentration during LTP at ‘silent synapses.’ Nat. Neurosci. 3, 330–336(2000).

33. Clements, J. D. & Westbrook, G. L. Activation kinetics reveal the number ofglutamate and glycine binding sites on the N-methyl-D-aspartate receptor.Neuron 7, 605–613 (1991).

34. Neher, E. The use of Fura-2 for estimating Ca buffers and Ca fluxes.Neuropharmacology 34, 1423–1442 (1995).

35. Sabatini, B. S., Oertner, T. G. & Svoboda, K. The life-cycle of Ca2+ ions inspines. Neuron 33, 439–452 (2002).

36. Asztely, F., Erdemli, G. & Kullmann, D. M. Extrasynaptic glutamate spilloverin the hippocampus: dependence on temperature and the role of activeglutamate uptake. Neuron 18, 281–293 (1997).

37. Arnth-Jensen, N., Jabaudon, D. & Scanziani, M. Cooperation betweenindependent hippocampal synapses is controlled by glutamate uptake. Nat.Neurosci. 5, 325–331 (2002).

38. Diamond, J. S. Neuronal glutamate transporters limit activation of NMDAreceptors by neurotransmitter spillover on CA1 pyramidal cells. J. Neurosci.21, 8328–8338 (2001).

39. Barbour, B. An evaluation of synapse independence. J. Neurosci. 21,7969–7984 (2001).

40. Rusakov, D. A. & Kullmann, D. M. Extrasynaptic glutamate diffusion in thehippocampus: ultrastructural constraints, uptake, and receptor activation. J. Neurosci. 18, 3158–3170 (1998).

41. Emptage, N., Bliss, T. V. P. & Fine, A. Single synaptic events evoke NMDAreceptor-mediated release of calcium from internal stores in hippocampaldendritic spines. Neuron 22, 115–124 (1999).

42. Kovalchuk, Y., Eilers, J., Lisman, J. & Konnerth, A. NMDA receptor–mediatedsubthreshold Ca2+ signals in spines of hippocampal neurons. J. Neurosci. 20,1791–1799 (2000).

43. Spacek, J. & Harris, K. M. Three-dimensional organization of smoothendoplasmic reticulum in hippocampal CA1 dendrites and dendritic spinesof the immature and mature rat. J. Neurosci. 17, 190–203 (1997).

44. Patneau, D. K. & Mayer, M. L. Structure–activity relationships for amino acidtransmitter candidates acting at N-methyl-D-aspartate and quisqualatereceptors. J. Neurosci. 10, 2385–2399 (1990).

45. Bliss, T. V. P. & Collingridge, G. L. A synaptic model of memory: long-termpotentiation in the hippocampus. Nature 361, 31–39 (1993).

46. Renger, J. J., Egles, C. & Liu, G. A developmental switch in neurotransmitterflux enhances synaptic efficacy by affecting AMPA receptor activation.Neuron 29, 469–484 (2001).

47. Liao, D., Hessler, N. A. & Malinow, R. Activation of postsynaptically silentsynapses during pairing-induced LTP in CA1 region of hippocampal slice.Nature 375, 400–404 (1995).

48. Isaac, J. T., Nicoll, R. A. & Malenka, R. C. Evidence for silent synapses:implications for the expression of LTP. Neuron 15, 427–434 (1995).

49. Sabatini, B. L. & Svoboda, K. Analsyis of calcium channels in single spinesusing optical fluctuation analysis. Nature 408, 589–593 (2000).

50. Sather, W., Dieudonne, S., MacDonald, J. F. & Ascher, P. Activation anddesensitization of N-methyl-D-aspartate receptors in nucleated outside-outpatches from mouse neurons. J. Physiol. 450, 643–672 (1992).

articles©

2002

Nat

ure

Pu

blis

hin

g G

rou

p

htt

p:/

/neu

rosc

i.nat

ure

.co

m

Tuning for 2D orientation in the image plane is prominent inmany areas of visual cortex and in multiple species1–5. Becauseorientation tuning is so prevalent, especially in early visualareas, most models of object recognition incorporate orienta-tion analysis as the major initial stage in image processing6–9.An important issue for such models is whether orientation sig-nals are purely 2D or at least partially 3D. Orientation in thereal world is 3D. Elongated image elements (such as lines andedges) are usually slanted in depth with respect to the viewer.3D orientation is perceptible based on binocular disparity andother depth cues, and could be useful for appreciating shape-in-depth. Here we observed explicit tuning for 3D orientationin area V4. This implies that orientation processing in the ven-tral pathway is 3D, and that initial stages of shape analysisinvolve depth information.

RESULTSWe studied 286 V4 neurons in two awake, fixating rhesus mon-keys. We deliberately selected neurons tuned for 2D bar ori-entation (that is, orientation in the image or frontoparallelplane). For each neuron, optimum 2D bar orientation, color,width and binocular position disparity were determined in pre-liminary tests. These optimum values were then held constantwhile stereoscopic slant (rotation about an axis in the imageplane positioned at the stimulus center and orthogonal to barorientation) was varied at 11.25° intervals.

For the example neuron in Fig. 1a, the optimum 2D orien-tation was 146.25° (counterclockwise from horizontal). Pre-liminary tests showed that this neuron was not tuned forposition in depth (binocular position disparity) of standardfrontoparallel bars. Stimuli in the slant test were therefore cen-tered in depth at the fixation plane (0° disparity). Thus, eachslanted bar appeared to be half in front of and half behind thefixation plane. Bar length was set at twice the estimated clas-sical receptive field (CRF) diameter, and bars were drifted back

Three-dimensional orientationtuning in macaque area V4

David A. Hinkle and Charles E. Connor

Department of Neuroscience, Johns Hopkins University School of Medicine and Zanvyl Krieger Mind/Brain Institute, 338 Krieger Hall, 3400 North Charles Street, Johns Hopkins University, Baltimore, Maryland 21218, USA

Correspondence should be addressed to C.E.C. ([email protected])


Tuning for the orientation of elongated, linear image elements (edges, bars, gratings), firstdiscovered by Hubel and Wiesel, is considered a key feature of visual processing in the brain. It hasbeen studied extensively in two dimensions (2D) using frontoparallel stimuli, but in real life mostlines, edges and contours are slanted with respect to the viewer. Here we report that neurons inmacaque area V4, an intermediate stage in the ventral (object-related) pathway of visual cortex,were tuned for 3D orientation—that is, for specific slants as well as for 2D orientation. The tuningfor 3D orientation was consistent across depth position (binocular disparity) and position within the2D classical receptive field. The existence of 3D orientation signals in the ventral pathway suggeststhat the brain may use such information to interpret 3D shape.

and forth at 4°/s across the CRF in the direction orthogonal to2D bar orientation, over a total path length of twice the CRFdiameter. Thus, the entire extent of the CRF was stimulated.Stereoscopic slant was conveyed by the binocular disparity gra-dient along the bar edges (which can also be thought of as ori-entation disparity) and the disparity gradient of small textureelements on the bar surface. The 2D appearance of the stim-uli (length, width, color, texture element size) remained con-stant except for disparity-related changes (Fig. 1a). This neuronshowed graded tuning for stereoscopic slant, with a peakresponse (within the tested range) of 66.5 ± 6.7 spikes/s (mean± s.e.m.) to stimuli slanted 78.75° with respect to the imageplane (left side near/right side far). Analysis of responses inseparate subregions of the CRF showed that tuning for stereo-scopic slant was consistent across 2D position (Methods).Responses to left and right eye images presented alone werewell below maximum and displayed no pattern that mightexplain the binocular responses.

To be useful for shape perception, 3D orientation tuningshould remain consistent across different stimulus positions,not only in the image plane but also in depth. To test this, wepresented an effective bar slant (67.5°, left near/right far; Fig. 1b) and an ineffective slant (67.5°, right near/left far) atstereoscopic depth positions ranging from –1.0° to 1.0° dis-parity. As in the first test, the entire CRF was sampled by usinglong bars and a long drift path. The response differencebetween the effective and ineffective slants remained consis-tent across the tested disparity range as well as across 2D posi-tion. We also tested the full range of slant values at threedisparities (–0.5°, 0.0°, 0.5°), and the resulting slant-tuningfunctions were similar at all three depths (Fig. 1c). These resultsdemonstrate that slant tuning was position invariant across thecell’s 3D receptive field. They also confirm that slant tuningwas not an artifact produced by tuning for depth (binocularposition disparity) at a particular 2D position.

articles


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


Consistent results were obtained for the full range of 3D ori-entations. In Fig. 2, orientation of the bar-shaped icons repre-sents 2D stimulus orientation (tested at 22.5° intervals), andradial position indicates stereoscopic slant, ranging from 0° (flator in the image plane) in the center ring and increasing at 22.5°intervals such that the bar end farthest from the center is alwaysnearest to the viewer (Fig. 2a). The Fig. 1 example neuronshowed graded tuning in this 3D orientation domain (Fig. 2b),with a peak consistent with the previous tests. Many neuronsexhibited similar tuning for non-zero slants (Fig. 2c and d), whilesome were tuned for orientations in the image plane (Fig. 2e).Other cells exhibited 2D orientation tuning but were insensitiveto slant (Fig. 2f).

In low-amplitude regions of these tuning surfaces, responserates are ambiguous and could reflect either non-optimal 2Dorientation or non-optimal slant. Complex temporal codesmight disambiguate these alternatives, but there were no cleartemporal response pattern differences between non-optimalstimuli (Fig. 3). This type of ambiguity is inevitable for any non-monotonic or multi-dimensional rate code. Individual neuronsfiring at their maximum rates may provide unambiguous sig-nals for optimum stimuli, but lower-amplitude responses canonly be resolved at the population level.

Approximately half our sample of orientation-tuned V4 neu-rons (157/286; 55%) displayed significant (P < 0.05, Mantel’s

test10,11) tuning for stereoscopic slant, as in Fig. 1a. For 104 ofthese neurons, slant tuning remained consistent across stereo-scopic depth, as in Fig. 1b and c (Methods). Thus, approximate-ly one-third (104/286; 36%) of orientation-tuned V4 neuronsappeared to encode 3D orientation in a depth-invariant manner.Of these, 50 were tested with the full range of 3D orientations (asin Fig. 2), and all tuning functions were single-peaked andsmoothly graded in this domain.

The tuning functions we observed depended on 3D linear ori-entation, not surface slant. We tested 14 cells tuned for 3D ori-entation, using both bars with surface texture and bars with nosurface texture. Orientation tuning was equivalent under thesetwo conditions for all cells. Thus, surface texture was not a nec-essary cue for 3D orientation tuning; the disparity gradient ofthe bar edges (orientation disparity) was sufficient. More impor-tantly, we tested 22 cells tuned for 3D orientation using circulartextured surface stimuli (covering twice the CRF diameter) at thesame slant angles and with the same texture elements as the barstimuli. Only one of these 22 cells was significantly tuned (P <0.05, Mantel’s test) for slant of the surface stimuli the numberexpected by chance at the 5% criterion level. Thus, surface ori-entation alone was not sufficient to produce the tuning functionswe observed; 3D linear orientation was the necessary cue. Thisshould not be taken as evidence that V4 contains no representa-tion of surface slant. We deliberately sampled from cells sensi-tive to linear orientation and may have excluded cells moreresponsive to texture12.

The 3D orientation tuning that we observed did not dependon stimulus length, even though apparent length varied with slantin most tests (Fig. 1a; apparent length was varied so that 2Dlength would remain constant). We tested 21 cells tuned for 3Dorientation at multiple lengths (typically twice and half the CRFdiameter) and in every case obtained comparable tuning curves.More importantly, we observed no instances of double tuningpeaks for opposing slants, which would be predicted in every caseif tuning depended on apparent length.

We quantified slant-tuning strength by fitting a cubic splinefunction to the data points (Fig. 1a) and finding the maximum

articles

Fig. 1. 3D orientation tuning of a representative V4 neuron. (a) 3Dslant tuning at the optimum 2D orientation. The stimuli were driftingbars of optimum width (0.41°) and color (cyan). Bar length and driftpath length were both 13.28° (twice the estimated CRF diameter). Thecolored bar icons in the top row (‘top view’) indicate apparent 3D slantsof the bar stimuli as they would appear from above if the disparity-defined depth structure were real. 3D slant was conveyed only by dis-parity cues; 2D appearance was constant (see ‘front view’) except fordisparity-related changes. Bar color represents mean response rateacross three repetitions, as indicated by the scale bar on the left of thex–y plot. The x–y plot shows mean response rate ± standard error (ver-tical bars) for binocular stimuli and for left and right eye images pre-sented alone. The curve connecting binocular stimulus responses is acubic spline fit performed in Matlab. This neuron is tuned for barsslanted such that the left end is near and the right end is far. (b) Theresponse difference between left near and right near slants is maintainedacross binocular position disparities (stereoscopic depth positions)ranging from –1.0° to 1.0°. The x-axis indicates binocular position dis-parity of the slanted bar stimulus center point. Responses to the opti-mum (left near) stimulus were lower in this test because a higherpercentage of the stimuli were effective in driving the cell. Most cells inthis study showed such a difference between the two tests. We havefound that repetitive stimulation with highly effective stimuli typicallydepresses response rates in V4 (ref. 49). (c) Slant tuning is consistentacross three position disparity values (–0.5°, 0.0°, 0.5°).

Frontview

Topview

67.5 45 22.5 0 22.5 45 67.5

0

10

20

30

40

50

60

70 BinocularLeft eyeRight eye

Stimulus slant with respect to image plane (deg)

Sp

ikes

/s

−1.0 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1.0

0

10

20

30

40

Slant:67.5º left near67.5º right near

Binocular disparity (deg)

Sp

ikes

/s

67.5 45 22.5 0 22.5 45 67.5

0

20

40

60

80

100Disparity:

–0.5º0º0.5º

Stimulus slant with respect to image plane (deg)

Sp

ikes

/s

a

b

c

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

(max) and minimum (min) function values. Tuning strengthwas defined as (max – min)/(max). The distribution of tuningstrength values across all neurons collected from monkeys trainedonly to fixate (Fig. 4a) and the distribution across all neuronscollected during a delayed match-to-sample task (Fig. 4b; Meth-ods) were statistically indistinguishable (randomization Kol-mogorov-Smirnov test; P = 0.15). In the subsample of 104 cellsmeeting our criteria for depth-invariant 3D orientation tuning(Fig. 4c), mean tuning strength was 0.80 (that is, a 5:1 ratiobetween max and min). Tuning peak positions for these cellswere distributed across the entire domain of 3D bar orientations(Fig. 4d). The distribution included slant tuning at 2D orienta-tions near horizontal, which would not be expected if the mech-anism depended solely on disparity gradients along the longedges of the bar stimuli (orientation disparity). In these cases,slant tuning was presumably based on the disparity gradient ofsurface texture elements or the disparities of the bar ends. Thereare some tuning peaks near 0° slant, which might be expectedjust based on narrow 2D orientation tuning as slant increases,orientations of the left and right eye images diverge from theoptimum (especially when optimum 2D orientation is near ver-tical). However, most slant-tuning peaks (84/104) were signifi-cantly (P < 0.01) different from 0°. The direction of slant tuningwas not correlated with receptive field position (P = 0.39; ran-domization test based on the circular correlation13 between angleof receptive field position relative to fixation and optimum slantdirection projected onto the image plane). However, most recep-tive fields in our sample were located in the lower right visualquadrant. (Neurons were sampled from the left hemisphere inboth monkeys.) The resulting limited angular range (∼ 90°) forreceptive field position reduces the power of any test for corre-lation with slant direction.

DISCUSSIONOrientation of elongated image elements is considered to be oneof the primary dimensions for neural representation of shape. Ourresults show that the ventral visual pathway carries robust signalsfor 3D orientation. Approximately one-third of orientation-sensitive V4 neurons show 3D slant tuning that remains consis-tent not only across position in the image plane but also acrossdepth (binocular position disparity). Such 3D position invari-ance would be useful for encoding 3D shape in a way that couldnot be disrupted by small changes in object location. 3D posi-tion invariance is also computationally difficult to achieve14. Thatthe visual system has solved this complex problem argues for thefunctional significance of 3D orientation tuning. A separate groupof 53 neurons in our sample were sensitive to slant but changedresponses across depth; these cells could conceivably carry bothslant and position-in-depth information. More complex 3D shapeinformation is represented at later processing stages15–18. These

and other findings19–22 indicate a rich representation of depthinformation in the ventral pathway.

Many current object recognition theories incorporate 2D ori-entation extraction as the first stage in shape processing; 3D struc-ture is inferred at later stages if at all. Our results suggest thatearly stages in shape-processing models should include 3D ori-entation processing as well. This suggestion is compatible withboth ‘viewpoint-invariant’6–8 and ‘viewpoint-dependent’23–25

models, both of which would benefit from explicit signals for 3Dlinear orientation such as we observed in area V4. Viewpoint-invariant mechanisms could take advantage of 3D orientationsignals to help construct an internal 3D object model. Viewpoint-dependent mechanisms could rely on 3D orientation signals tohelp distinguish similar views of similar objects that differ only indepth structure. Our results are incompatible with a strictly 2Dversion of viewpoint-dependent recognition.

Previous work in cats and monkeys has not revealed a clearrepresentation of 3D orientation in areas V1 and V2, thoughmany neurons in these areas are tuned for binocular position dis-parity26–28. Some neurons in anesthetized cat primary visual cor-tex29,30 and area 21a (ref. 31) have different orientationpreferences in the two eyes. Such differences could represent selec-tivity for orientation disparity (slanted lines and edges produceslightly different orientations in the two eye images) and might beuseful for perception of 3D orientation29. However, careful analy-sis suggests that the orientation preference differences in cat V1 donot produce useful signals for orientation disparity30. A smallpercentage of monkey V2 neurons seems to be sensitive to ori-entation disparity, even without horizontal disparity cues32. Inawake monkey V1, some neurons demonstrate consistent selec-tivity for orientation disparity, but this is probably a consequenceof tuning for position disparity33. Thus, although some V1/V2cells are sensitive to orientation disparity, current evidence does

articles


Fig. 2. 3D orientation-tuning examples. (a) The stimulus domain com-prised the full range of 2D orientations sampled at 22.5° intervals. Ateach 2D orientation, 3D slant was varied at 22.5° intervals for a total ofseven slants. Slant values (0°, 22.5°, 45°, 67.5°) indicate stereoscopicrotation of the bar stimulus with respect to the image plane and aboutan axis positioned at its center and orthogonal to its long axis. (b–f) Colored bar icons depict 2D stimulus orientation. Distance fromthe center of the plot corresponds to 3D slant (see a). The conventionfor slant direction is that the bar end farthest from the center is nearestto the viewer. The eight bar icons at 0° slant are replicated around thecenter circle for symmetry. Bar color represents mean response rate, asindicated by the scale bar to the right of each plot.

Slant

0

22.545

67.5

0

20

40

60

Sp

ikes

/s

0

10

20

30

40

0

5

10

Sp

ikes

/s

0

10

20

0

10

20

30

Sp

ikes

/s

a b

c d

e f

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


not indicate that this represents explicit 3D orientation process-ing. Of course, populations of orientation-tuned V1/V2 neuronswith small receptive fields and a range of depth tuning peakswould implicitly encode 3D orientation. Selective connectivitywith such cells could produce the explicit 3D orientation tuningwe observed in V4. Also, the presence of 3D orientation tuning inV4 implies that similar tuning properties exist in subsequentstages of the ventral pathway (in IT cortex).

Tuning for 3D orientation of extended surfaces seems to beprominent in the dorsal visual pathway. Neurons in the caudalintraparietal area (CIP) show tuning for 3D surface orientationbased on binocular disparity gradients in contours, surface textureor both34,35. Neurons in area MT show robust tuning for 3D ori-entation of moving dot texture surfaces (J. D. Nguyenkim & G. C.DeAngelis, Soc. Neurosci. Abstr. 27, 165.8, 2001). Areas MT andMST are also sensitive to stereoscopic position in depth36,37 andinteractions between depth and motion37–39. Although depthposition tuning has been described in the ventral pathway20–22, itis only in the dorsal pathway that a clear causal connectionbetween neural responses and depth perception has been estab-lished using combined behavioral and microstimulation meth-ods40. The functional role of depth sensitivity in the ventral

pathway is less certain. Presumably, tuning for surface orienta-tion and depth position in the dorsal pathway relates to 3Daspects of large-scale space and motion perception. Tuning for3D orientation of linear shape elements and other depth-relat-ed properties in the ventral pathway more likely relate to 3Dobject perception.

The only perceptual cues for slant in our experiment werebinocular disparity along bar contours and binocular disparityof texture elements on the bar surface. The texture cues wereadded to enhance the 3D percept, but contour cues alone seemedto be sufficient, as 3D orientation tuning functions were repli-cated using stimuli with no texture. Psychophysical experimentsshow that a variety of disparity-based and non-disparity-based(monocular) cues are used to infer slant41–45. The cells that westudied here may be equally sensitive to monocular slant cuessuch as perspective and texture size gradient. Alternatively, theymay show only standard 2D orientation tuning in response to2D images, albeit at a less than maximum response level. Furtherexperiments will be needed to ascertain the complete set of slantcues to which individual neurons are sensitive, but our datademonstrate that disparity cues by themselves are sufficient tosupport robust 3D orientation tuning.

articles

Fig. 3. Raster plot of individual-trial responses for the Fig. 2c example cell. Each trial is represented by a horizontal row of action potentials (tickmarks). The position of each action potential corresponds to the position of the bar stimulus during the sweep across the CRF. Successive sweepswere in alternate directions. Arrows to the left of each row indicate sweep direction. (Bar orientation was orthogonal to sweep direction.) (a) Densesample of slant values (11.25° increments) at the optimum 2D orientation (vertical). (b) Full 3D orientation tuning test. Raster plots are arranged asin Fig. 2c. (The number of repetitions is smaller in this test because of the greater number of stimuli.)

78.7

567

.50

56.2

545

.00

33.7

522

.50

11.2

500

.00

11.2

522

.50

33.7

545

.00

56.2

567

.50

78.7

5T

op

nea

r / b

ott

om

far

Bo

tto

m n

ear

/ to

p f

arS

lan

t

10°

a b

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

METHODSStimuli and data collection. Stereoscopic visual stimuli were generatedon an Octane workstation (Silicon Graphics, Mountain View, Califor-nia) using a NuVision liquid crystal shutter (MacNaughton, Beaverton,Oregon) attached to the display screen. We compensated for cross-talkbetween the two eye channels by adding to each eye’s image a low-contrast, negative version of the opposite eye image. This procedure pro-duced stimuli that appeared to be free of any interference between eyechannels. Eye position was monitored with a scleral coil system (River-bend Instruments, Birmingham, Alabama)46 and a video-based system(ISCAN, Burlington, Massachusetts). In one monkey, the positions ofboth eyes were monitored during experiments using two scleral coils. In the other monkey, the position of only one eye was monitored dur-ing experiments, but the positions of both eyes were monitored during experiment-like trials, using the video system. Analyses of eye positiondata for both monkeys showed that there was no significant correlationbetween eye vergence angle and stimulus slant or any other relevant stim-ulus parameter. Thus, stereoscopic stimulus effects reported here werenot due to vergence eye movements. Microelectrodes were used to iso-late single units in the V4 lower field parafoveal representation on theprelunate gyrus and adjoining banks of the lunate and superior tempo-ral sulci. Detailed recording procedures have been described20,47.

The behavioral task during data collection for both monkeys was tomaintain fixation within a 0.5° radius window centered on a 0.1° whitespot straight ahead, for a period of up to 5.5 s per trial, while stimuli(rectangular bars of optimum color and width) were flashed or drift-ed in the CRF. To further ensure stereoscopic perception during theexperiment, one animal was trained (partway through the study) toperform a delayed match-to-sample (DMS) slant discrimination taskon the bar stimuli in the CRF. Results were similar during DMS andfixation behavior, and were combined for most of the analyses pre-sented here. The combined data set comprised 105 neurons studiedduring the fixation task and 181 neurons studied during the DMS task.All procedures conformed to National Institutes of Health and USDAguidelines and were approved by the Johns Hopkins University AnimalCare and Use Committee.

Most neurons were tested with drifting bars, but some (22 of the 157cells with significant slant tuning) were tested additionally (15) or alter-natively (7) with flashed bars. For drifting bar tests, only one stimuluswas presented per trial. Drift rate was 4°/s, and drift path length wastwice the CRF diameter. The number of sweeps depended on CRF size,and varied from one to five sweeps per trial, for total presentation timesof 2.5–5.5 s. For flashed bar tests, five different stimuli were flashed dur-ing each trial in random order (presentation time, 0.75 s; interstimulusinterval, 0.25 s). CRF diameter was estimated from eccentricity basedon the V4 diameter/eccentricity relationship reported in ref. 48. Thosedata suggest that the average V4 CRF diameter is approximately 1° +0.625 × eccentricity, and this accorded well with our estimates based onhand plotting. For the neuron in Fig. 1, with an eccentricity of 9.02°,the estimated CRF diameter was 6.64°. The texture elements were linesof width 0.04° and length 0.24°, positioned at random orientations andrandom locations with an average density of seven elements per squareddegree. Texture element size and spacing were constant across all parts ofall stimuli; thus there were no monocular texture cues for slant or depth.

Data analysis. Response rates for drifting and flashed bar stimuli werecalculated by summing action potentials across the entire presentationperiod and dividing by presentation time. Response rates were averagedacross three to five presentations. Background rate, measured duringcomparable trial periods in which no stimulus was presented, was sub-tracted from all stimulus response rates.

For the 100 depth-invariant slant-tuned neurons studied with driftingbars, response differences across 2D position were evaluated using analy-sis windows corresponding to different quarters of the CRF. Most neu-rons were responsive across all four quarters (81/100); some respondedonly in two or three quarters (19/100), indicating that the eccentricity-based formula sometimes overestimated CRF diameter. Slant-tuningfunctions were consistent for all neurons across all CRF quarters in whichthey were responsive.

All neurons with significant slant tuning were tested with one effectiveslant and one ineffective slant at disparities ranging from –1.0° to 1.0° at0.2° intervals, as in Fig. 1b. Our criterion for consistent tuning acrossdepth was that the effective slant response was greater than the ineffective slant response for all disparities at which either stimulus produced a response greater than half maximum. In addition, 39 neu-rons that passed this test were also studied with the full range of slantvalues at one or more additional disparities, as in Fig. 1c. In all 39 cases, slant-tuning functions were similar at the different disparities. Thirty-four slant-tuned neurons were tested with right and left eye images

articles


Fig. 4. Population results. (a) Distribution of 3D slant-tuning strengths for neu-rons studied during the fixation task. Index values greater than 1.0 (collected hereinto one bin) represent cases in which the min response (Results) was below 0(below background response rate). (b) Distribution of tuning strengths for neu-rons studied during the delayed match-to-sample task. (c) Distribution of tuningstrengths for 104 neurons with significant tuning that remained consistent acrossdepth position (Results). (d) Distribution of tuning peaks across the 3D orienta-tion domain (Fig. 2a) for the same 104 neurons. Tuning peaks were derived fromcubic spline fits to slant-tuning data at the optimum 2D orientation (Fig. 1a). Redcircles represent neurons with slant-tuning peaks significantly different from 0°.Blue circles represent tuning peaks not significantly different from 0°.

0.0 0.2 0.4 0.6 0.8 1.0 ∞

Nu

mb

er o

f ce

lls

Slant-tuning strength

0

5

10

15

20

0.0 0.2 0.4 0.6 0.8 1.0 ∞

Nu

mb

er o

f ce

lls


0

10

20

30

0.0 0.2 0.4 0.6 0.8 1.0 ∞

Nu

mb

er o

f ce

lls


0

5

10

15

20

a

b

c

d©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


presented alone. There were no cases in which responses to these monoc-ular stimuli seemed to explain slant-tuning functions.

Among neurons tested with the full range of 3D bar orientations,the 2D orientation tuning peaks (circumferential dimension in Fig. 2)were the same for slanted and flat (0° slant) bars (for example, Fig. 2d).The mean 2D orientation peak difference between slanted (optimumslant) and flat (0° slant) bars was just 12.2° (standard deviation, 16.8°)for the 41/50 neurons with significant orientation tuning for both slant-ed and flat bars (P < 0.05; ANOVA). Thus, tuning in the circumferen-tial dimension in Fig. 2 was due to the 2D component of barorientation, not surface slant.

Significant difference from tuning for 0° slant was determined by treat-ing each spike as an observation associated with the slant value of thebar stimulus that evoked that spike. A two-tailed t-test was used to eval-uate whether the mean of the resulting response strength distributionacross the slant domain was significantly different from 0°. The assump-tions underlying a standard t-test are presumably violated in this situa-tion, so we adopted a stricter significance criterion (P < 0.01) than wedid in other tests.

AcknowledgmentsWe thank S.L. Brincat, G.F. Poggio and R. von der Heydt for comments on the

manuscript. Some analyses were suggested by B.G. Cumming. Technical support

was provided by W. Nash, W. Quinlan and B. Sorenson. This work was

supported by the National Institute of Neurological Disorders and Stroke and by

the Pew Scholars Program in the Biomedical Sciences.


RECEIVED 27 FEBRUARY; ACCEPTED 29 MAY 2002

1. Desimone, R. & Schein, S. J. Visual properties of neurons in area V4 of themacaque: sensitivity to stimulus form. J. Neurophysiol. 57, 835–868 (1987).

2. Gross, C. G., Rocha-Miranda, C. E. & Bender, D. B. Visual properties ofneurons in inferotemporal cortex of the macaque. J. Neurophysiol. 35, 96–111(1972).

3. Hubel, D. H. & Wiesel, T. N. Receptive fields of single neurones in the cat’sstriate cortex. J. Physiol. (Lond.) 148, 574–591 (1959).

4. Hubel, D. H. & Wiesel, T. N. Receptive fields and functional architecture intwo nonstriate visual areas (18 and 19) of the cat. J. Neurophysiol. 28, 229–289(1965).

5. Hubel, D. H. & Wiesel, T. N. Receptive fields and functional architecture ofmonkey striate cortex. J. Physiol. (Lond.) 195, 215–243 (1968).

6. Biederman, I. Recognition-by-components: a theory of human imageunderstanding. Psychol. Rev. 94, 115–147 (1987).

7. Dickinson, S. J., Pentland, A. P. & Rosenfeld, A. From volumes to views: anapproach to 3D object recognition. CVGIP: Image Understanding 55, 130–154(1992).

8. Marr, D. & Nishihara, H. K. Representation and recognition of the spatialorganization of three-dimensional shapes. Proc. R. Soc. Lond. B Biol. Sci. 200,269–294 (1978).

9. Riesenhuber, M. & Poggio, T. Hierarchical models of object recognition incortex. Nat. Neurosci. 2, 1019–1025 (1999).

10. Mantel, N. The detection of disease clustering and a generalized regressionapproach. Cancer Res. 27, 209–220 (1967).

11. Manly, B. F. J. Randomization and Monte Carlo Methods in Biology 114–118(Chapman & Hall, London, 1991).

12. Hanazawa, A. & Komatsu, H. Influence of the direction of elementalluminance gradients on the responses of V4 cells to textured surfaces. J. Neurosci. 21, 4490–4497 (2001).

13. Mardia, K. V. & Jupp, P. E. Directional Statistics 248–249 (Wiley, Chichester,2000).

14. Anzai, A., Ohzawa, I. & Freeman, R. D. Neural mechanisms for processingbinocular information: complex cells. J. Neurophysiol. 82, 909–924 (1999).

15. Janssen, P., Vogels, R. & Orban, G. A. Macaque inferior temporal neurons areselective for disparity-defined three-dimensional shapes. Proc. Natl. Acad. Sci.USA 96, 8217–8222 (1999).

16. Janssen, P., Vogels, R. & Orban, G. A. Three-dimensional shape coding ininferior temporal cortex. Neuron 27, 385–397 (2000).

17. Janssen, P., Vogels, R. & Orban, G. A. Selectivity for 3D shape that reveals

distinct areas within macaque inferior temporal cortex. Science 288,2054–2056 (2000).

18. Janssen, P., Vogels, R., Liu, Y. & Orban, G. A. Macaque inferior temporalneurons are selective for three-dimensional boundaries and surfaces. J. Neurosci. 21, 9419–9429 (2001).

19. Dobbins, A. C., Jeo, R. M., Fiser, J. & Allman, J. M. Distance modulation ofneural activity in the visual cortex. Science 281, 552–555 (1998).

20. Hinkle, D. A. & Connor, C. E. Disparity tuning in macaque area V4.Neuroreport 12, 365–369 (2001).

21. Watanabe, M., Tanaka, H., Uka, T. & Fujita, I. Disparity-selective neurons inarea V4 of macaque monkeys. J. Neurophysiol. 87, 1960–1973 (2002).

22. Uka, T., Tanaka, H., Yoshiyama, K., Kato, M. & Fujita, I. Disparity selectivityof neurons in monkey inferior temporal cortex. J. Neurophysiol. 84, 120–132(2000).

23. Poggio, T. & Edelman, S. A network that learns to recognize three-dimensional objects. Nature 343, 263–266 (1990).

24. Tarr, M. J. & Pinker, S. Mental rotation and orientation-dependence in shaperecognition. Cognit. Psychol. 21, 233–282 (1989).

25. Vetter, T., Hurlbert, A. & Poggio, T. View-based models of 3D objectrecognition: invariance to imaging transformations. Cereb. Cortex 5, 261–269(1995).

26. Barlow, H. B., Blakemore, C. & Pettigrew, J. D. The neural mechanism ofbinocular depth discrimination. J. Physiol. (Lond.) 193, 327–342 (1967).

27. Hubel, D. H. & Wiesel, T. N. Stereoscopic vision in macaque monkey: cellssensitive to binocular depth in area 18 of the macaque monkey cortex. Nature225, 41–42 (1970).

28. Poggio, G. F. & Fischer, B. Binocular interaction and depth sensitivity instriate and prestriate cortex of behaving rhesus monkey. J. Neurophysiol. 40,1392–1405 (1977).

29. Blakemore, C., Fiorentini, A. & Maffei, L. A second neural mechanism ofbinocular depth discrimination. J. Physiol. 226, 725–749 (1972).

30. Nelson, J. I., Kato, H. & Bishop, P. O. Discrimination of orientation andposition disparities by binocularly activated neurons in cat striate cortex. J. Neurophysiol. 40, 260–283 (1977).

31. Wieniawa-Narkiewicz, E., Wimborne, B. M., Michalski, A. & Henry, G. H.Area 21a in the cat and the detection of binocular orientation disparity.Ophthalmic Physiol. Opt. 12, 269–272 (1992).

32. von der Heydt R., Hanny, P., Dursteler, M. R. & Poggio, G. F. Neuronalresponses to stereoscopic tilt in the visual cortex of the behaving monkey.Invest. Ophthalmol. Vis. Sci. 22 (Suppl.), 12–14 (1982).

33. Bridge, H. & Cumming, B. G. Responses of macaque V1 neurons to binocularorientation differences. J. Neurosci. 21, 7293–7302 (2001).

34. Shikata, E., Tanaka, Y., Nakamura, H., Taira, M. & Sakata, H. Selectivity of theparietal visual neurones in 3D orientation of surface of stereoscopic stimuli.Neuroreport 7, 2389–2394 (1996).

35. Taira, M., Tsutsui, K. I., Jiang, M., Yara, K. & Sakata, H. Parietal neuronsrepresent surface orientation from the gradient of binocular disparity. J. Neurophysiol. 83, 3140–3146 (2000).

36. Maunsell, J. H. & Van Essen, D. C. Functional properties of neurons inmiddle temporal visual area of the macaque monkey. II. Binocularinteractions and sensitivity to binocular disparity. J. Neurophysiol. 49,1148–1167 (1983).

37. Roy, J. P., Komatsu, H. & Wurtz, R. H. Disparity sensitivity of neurons inmonkey extrastriate area MST. J. Neurosci. 12, 2478–2492 (1992).

38. Bradley, D. C. & Andersen, R. A. Center-surround antagonism based ondisparity in primate area MT. J. Neurosci. 18, 7552–7565 (1998).

39. Bradley, D. C., Chang, G. C. & Andersen, R. A. Encoding of three-dimensional structure-from-motion by primate area MT neurons. Nature392, 714–717 (1998).

40. DeAngelis, G. C., Cumming, B. G. & Newsome, W. T. Cortical area MT andthe perception of stereoscopic depth. Nature 394, 677–680 (1998).

41. Cagenello, R. & Rogers, B. J. Anisotropies in the perception of stereoscopicsurfaces: the role of orientation disparity. Vision Res. 33, 2189–2201 (1993).

42. Ernst, M. O., Banks, M. S. & Bulthoff, H. H. Touch can change visual slantperception. Nat. Neurosci. 3, 69–73 (2000).

43. Gillam, B. & Ryan, C. Perspective, orientation disparity and anisotropy instereoscopic slant perception. Perception 21, 427–439 (1992).

44. Mitchison, G. J. & McKee, S. P. Mechanisms underlying the anisotropy ofstereoscopic tilt perception. Vision Res. 30, 1781–1791 (1990).

45. Ninio, J. Orientational versus horizontal disparity in the stereoscopicappreciation of slant. Perception 14, 305–314 (1985).

46. Robinson, D. A. A method of measuring eye movement using a scleral searchcoil in a magnetic field. IEEE Trans. Bio-Med. Electron. 10, 137–145 (1963).

47. Pasupathy, A. & Connor, C. E. Shape representation in area V4: position-specific tuning for boundary conformation. J. Neurophysiol. 86, 2505–2519(2001).

48. Gattass, R., Sousa, A. P. & Gross, C. G. Visuotopic organization and extent ofV3 and V4 of the macaque. J. Neurosci. 8, 1831–1845 (1988).

49. Pasupathy, A. & Connor, C. E. Responses to contour features in macaque areaV4. J. Neurophysiol. 82, 2490–2502 (1999).

articles©

2002

Nat

ure

Pu

blis

hin

g G

rou

p

htt

p:/

/neu

rosc

i.nat

ure

.co

m

In human vision, blindness to ignored inputs is a conspicuousfeature of selective attention1. When attention is focused on onevisual event, there is often little awareness or processing of others.Here we consider the role of prefrontal cortex in filtering out2

unwanted signals.Several current accounts of attentional function propose an

important role for prefrontal cortex (PFC)3,4. A central aspect ofprefrontal function is thought to be flexibility, with cells adapt-ing to code information of specific relevance to current behav-ior4,5. In the monkey, studies of working memory6 and visualsearch7,8 have shown selective neural responses to relevant or tar-get inputs in the lateral PFC and the frontal eye field. Human neuroimaging studies also suggest that the PFC is important inattention and awareness9, with substantially stronger prefrontalresponses to consciously detected inputs10,11.

In both behavioral and physiological studies, spatial cueinghas been a useful way to analyze attentional functions12–14. Foruse in the monkey, we adapted a spatial cueing task used inhuman event-related potential and other studies13 (Fig. 1 andMethods). We presented the animals with a stream of visualobjects pictured on a computer screen. Throughout the experi-ment, only three objects were involved: a single target (fish) andtwo non-targets (bear, hamburger). In the ‘unilateral’ condition,objects were presented sequentially at a single location in the leftor right visual field. The task was to maintain central gaze untilthe appearance of the target, and then to fixate it. In the ‘bilat-eral’ condition, objects appeared simultaneously in left and rightvisual fields. One location was cued at the start of the trial; thetask was to attend just to this side, again waiting for the target onthis side and fixating it when it appeared. During these tasks, werecorded properties of neurons in the lateral PFC.

Filtering of neural signals byfocused attention in the monkeyprefrontal cortex

Stefan Everling1,2,3, Chris J. Tinsley1,2, David Gaffan2 and John Duncan1

1 MRC Cognition and Brain Sciences Unit, 15 Chaucer Road, Cambridge CB2 2EF, UK2 Department of Experimental Psychology, Oxford University, South Parks Road, Oxford OX1 3UD, UK3 Present address: Departments of Physiology & Psychology, University of Western Ontario, London, Ontario N6A 5C2, Canada

Correspondence should be addressed to J.D. ([email protected])


Prefrontal cortex is thought to be important in attention and awareness. Here we recorded theactivity of prefrontal neurons in monkeys carrying out a focused attention task. Having directedattention to one location, monkeys monitored a stream of visual objects, awaiting a predefinedtarget. Although neurons rarely discriminated between one non-target and another, they commonlydiscriminated between targets and non-targets. From the onset of the visual response, thistarget/non-target discrimination was effectively eliminated when the same objects appeared at anunattended location in the opposite visual hemifield. The results show that, in prefrontal cortex,filtering of ignored locations is strong, early and spatially global. Such filtering may be important inblindness to unattended signals—a conspicuous aspect of human selective attention.

Neural activity in the unilateral condition allowed us to askwhat kind of visual information is represented in PFC for thistask. The results showed prominent representation of the tar-get/non-target distinction. Data from the bilateral condition test-ed how this information is filtered from an unattended visuallocation. The results showed effectively complete filtering, begin-ning with the onset of the visual response.

RESULTSBehavioral dataNeural data were obtained from two monkeys over a total of 67experimental sessions (see Methods). In the unilateral task, mon-keys on average correctly completed 86% of the trials, lookedtoward a non-target stimulus on 13% and missed the target stim-ulus on 1%. In the bilateral task, they performed correctly in 80%of the trials, looked toward a non-target stimulus on 5%, lookedtoward a target stimulus at the wrong location on 14% andmissed the target on 1%. Breaks in fixation were not counted inthese trial percentages.

Unilateral conditionActivity in both conditions was recorded for a total of 161 PFCneurons. Many neurons responded to the presentation of theobjects. For the unilateral condition, our analyses focused on objectselectivity, that is, differential response based on object identity.

A first analysis concerned differential response to the two dif-ferent non-targets. For each cell, we performed two-way analy-sis of variance (ANOVA) with factors cued location (ipsilateralor contralateral to recording site) and non-target object (bear,hamburger). In this analysis, 18 neurons showed a main effect oflocation, but only 1 showed a main effect of object and 0 showed

articles


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

responses in 10-ms bins, beginning at stimulus onset (see Meth-ods). During trials in which the stimuli appeared at the preferredlocation, the activity started to differ between target and non-target objects in the period 110–120 ms after stimulus onset (t-test across cells, P < 0.05). This discrimination occurred wellbefore the onset of saccades toward target objects (average sac-cade latency, 190 ms in monkey A, 181 ms in monkey B), and last-ed until 300 ms or more after stimulus presentation. For stimuli atthe non-preferred location, these small analysis bins showed nosignificant differences during the first 200 ms after stimulus onset.

In Figs. 2a and b, rasters show spikes for each single presen-tation of a target (blue) or non-target (red), aligned on stimulusonset. For targets, rasters are ordered by saccadic latency, circlesindicating saccade onset. Especially for the neuron in Fig. 2b, thedata suggest the neural activity is more closely time-locked to tar-get onset than to saccade initiation; onset latency of the neuralresponse is approximately the same for the slowest and fastest sac-cades. To analyze this for the whole population, for each cell wesorted trials into those with fast, medium or slow saccadic laten-cies (division of saccadic latency distribution into thirds, or ter-tiles). For each tertile, we took onset of the neural response to beindicated by the first two consecutive 10-ms bins in which activ-ity exceeded two standard deviations of baseline activity. Onsetlatency was defined as the start of the first of these bins. Responseonsets could be reliably defined in 25 of 33 cells. For these 25 cells,mean saccadic latencies in fast, medium and slow tertiles wererespectively 166, 184 and 218 ms. Neural onset latencies for thethree tertiles (respectively 126, 158 and 141 ms) were not signifi-cantly different (ANOVA, P > 0.20). Timed from saccadic ratherthan stimulus onset, neural onset latencies for the three tertileswere respectively –40, –26 and –78 ms, the difference this timebeing significant (ANOVA, P < 0.05). These data suggest neuralresponses time-locked to stimulus onset rather than saccade.

The conclusion that responses in these cells reflect target iden-tification rather than saccade initiation is confirmed by two furtherpieces of evidence. In Fig. 3a, mean neural activity is shown forthree kinds of stimulus: targets with correct saccades, non-targetswith saccades correctly withheld and non-targets leading to anincorrect saccade. These data come just from trials with stimuliin (and saccades to) each cell’s preferred location; data have beencombined for 19 cells with reasonable (>4) numbers of error trials. For non-target stimuli, there was little or no neural responsewhether or not a saccade was made. Across these 19 cells, responses to targets and non-targets with saccades were significantly different (P < 0.002), whereas responses to non-targets with and without saccades were not (P > 0.40; t-tests onactivity 100–200 ms after stimulus onset; see Methods). For a sub-set of target-selective cells, data were also gathered in a gap sac-cade task (Fig. 3b; see Methods). In this task, the monkey simplymoved his eyes to one of the two peripheral locations when thefixation point turned off and a peripheral stimulus turned on.

Fig. 1. The focused attention task and recording locations. (a) Examplestimulus sequences in unilateral and bilateral conditions. Each trialbegan when the monkey fixated (curved arrow) a small dot in the cen-ter of the screen. A cue (a white square) appeared to left or right, fol-lowed by a stream of stimuli in just the cued location (unilateralcondition) or in both locations (bilateral condition). Central fixation(dotted circle) was to be maintained until a target (fish) appeared at thecued location, at which point an immediate saccade (dotted arrow) tothis target was required. (b) Location of recording sites, with numbersof cells showing significant object (target versus non-target) selectivity(excitatory only, n = 33, see text).


an interaction (P < 0.01; see Methods). In contrast, a secondanalysis concerned differential response to non-targets versustargets. For these ANOVAs, responses to the two non-targets werepooled, so that the ‘object’ factor was simply target versus non-target. This time, there were significant main effects of location in39 neurons, of object in 34 neurons, and there was a significantinteraction in 21 neurons. In total, 44 of the 161 neurons (27%)showed some form of target/non-target selectivity (main effector interaction in this second ANOVA).

For the remainder of this report, we focus on those neuronswith target/non-target selectivity that gave excitatory respons-es to the target (33 of 44 or 75%; see Fig. 2 for exampleresponses). We defined each neuron’s preferred location as thelocation that yielded the maximal response for the target object in the unilateral condition. Consistent with previousreports15–17, we found a mild preference for the contralateralhemifield (18 of 33 or 55% preferred contralateral stimuli and15 of 33 or 45% preferred ipsilateral stimuli).

For the neuron in Fig. 2a, there was a strong excitatoryresponse to targets in the preferred location, and a somewhatweaker response to targets in the non-preferred location. Noresponse was seen to non-targets. For the neuron in Fig. 2b, bothtargets and non-targets gave some response in the preferred loca-tion, though the target response was stronger. For the popula-tion as a whole (Fig. 2c), there was a strong response to targets,but little, if any, response to non-targets.

The population data show that differentiation of targets andnon-targets began early, around 100 ms after stimulus onset. Toestablish this formally, we compared mean target and non-target

articles

Cue

Stimulus 1

Stimulus 2

Stimulus 3

300 ms

500 ms

700 ms

100 ms

500 ms

300 ms

300 ms

BilateralUnilaterala

b

Arcuate sulcus

Monkey A

Principalsulcus

Monkey B

Recording location1 object-selective neuron2–5 object-selective neurons6 object-selective neurons

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

Again, data in Fig. 3b come just from stimuli in (and saccades to)the cell’s preferred location. As compared to responses of the sameneurons in the main task, the gap task produced both lower base-line activity and little or no response surrounding the peripheralstimulus and associated saccade. Across cells, responses in the twotasks were significantly different (t-test on period 50 ms before to50 ms after saccade onset, P < 0.005).

These results from the unilateral condition are in line with pre-vious reports showing that prefrontal neurons classify stimuli intobehaviorally relevant categories, defined either by immediateresponses18 or future significance16,19,20. In our case, such codingof behavioral relevance took the form of selective visual responseto ‘go’ stimuli16, with no differentiation between different objects(different non-targets) having the same behavioral significance.

Bilateral conditionWe then examined the effect of attention on target responses inthe bilateral condition. Both for the whole population of target-selective neurons (Fig. 4a), and for the individual neurons (Fig. 4b, shaded region in Fig. 4a), we measured responses toeach possible bilateral stimulus array, either with attention direct-ed to the preferred location (blue lines) or to the non-preferredlocation (red lines).

In each case, strong responses were seen for an attended tar-get, especially on the preferred side, but these were eliminatedwhen targets were unattended. In the interval 100–200 ms after stimulus onset, we found significant differences between

attention to preferred and non-preferred sides for each arraywith one target and one non-target. When the target appearedat the preferred location with a non-target at the non-preferredlocation, 76% (25 of 33) of the neurons gave a larger responsewhen the monkey attended to the target object (t-test acrosscells, P < 0.005). Significant differences (t-test, P < 0.05) wereobtained for 42% (14 of 33) of the individual neurons. Thesedifferences started very early, in the period 100–110 ms afterstimulus onset (t-test in 10-ms bins, P < 0.05). Similar butsmaller differences were obtained for arrays in which the tar-get appeared at the non-preferred location and the non-target atthe preferred location. In this condition, 67% (22 of 33) of theneurons had a stronger response when the monkey attended to the target object (t-test, P < 0.05). Significant differences (t-test, P < 0.05) were found for 27% (9 of 33) of the neurons.We also found that many PFC neurons exhibited a slightdecrease later in their activity when two non-target objectsappeared and attention was directed to the preferred location. Inthe window 200–300 ms after stimulus onset, 76% (25 of 33)of the neurons showed lower activity for attend-preferred thanfor attend-non-preferred (t-test across cells, P < 0.005).

articles


Non-preferred location

Target Non-target

Preferred location

Target Non-target

Act

ivity

(sp

ikes

/s)

10

15

20

25

30

5

n = 33

Time from stimulus onset (ms)

0 100 200 300 400


0 100 200 300 400

33

0

0

20

40

60

80

20

40

60

80

Act

ivity

(sp

ikes

/s)

Act

ivity

(sp

ikes

/s)

a

b

c

Fig. 2. Unilateral condition; activity of prefrontal cortex (PFC) neuronsthat showed object (target versus non-target) selectivity. (a, b) Activityof single PFC neurons. (c) Activity of the population (all target/non-target selective neurons with excitatory target responses). Shaded areas indicate presentation of visual stimuli. Rasters show spikesfor each single presentation of non-target (red, top) or target (blue,bottom), aligned on stimulus onset. For targets, circles show onset of the saccade. In this and subsequent figures, display icons show preferred location on the right, though actual side varied between cells. Red and blue icon outlines indicate color key, and were not present in actual display.

0 100 200 300 400


Act

ivity

(sp

ikes

/s)

b

0

10

20

30

40

0

10

20

30

40

Act

ivity

(sp

ikes

/s)

n = 19

n = 20

a

Correct target

Correct non-target

Error non-target

Unilateral target

Gap task

Fig. 3. Neural activity for selected stimulus-response combinations. (a) Unilateral condition; mean activity of 19 target-selective neurons fortargets with correct saccades, non-targets with saccades correctly with-held and non-targets with erroneous saccades. (b) Unilateral condition(targets) compared with gap task; mean activity of 20 target-selectiveneurons. All data concern stimuli in preferred location only.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

non-preferred location (Fig. 5b); here, 27 of 33 (81%) ofneurons had a higher preferred-location selectivity if thatlocation was attended (t-test, P < 0.002).

Inhibitory neuronsThough here we have focused on excitatory neurons, resultswere similar for those 11 neurons with target/non-targetselectivity and inhibitory responses. Typically, inhibition wasstronger for targets, and again was filtered out at the unat-tended location.

DISCUSSIONIn our task, the primary form of object selectivity seen inPFC was distinction between targets and non-targets. In itself,this result suggests the ability of PFC to focus on stimulusdistinctions of relevance to the trained task5,16,20. A numberof recent neuroimaging studies have shown strong frontalresponses to targets in a stimulus sequence, with less responseto non-targets21,22. Such results are directly analogous to ourfinding of selective prefrontal responses to target objects.

Our major results, however, concern effects of spatial cue-ing in the bilateral condition. In our task, we observed strongfiltering of the PFC response to unattended targets. This spa-tial filtering of object identity in the ignored hemifield waseffectively complete, and began with the onset of the visualresponse. Such results are closely reminiscent of the blind-ness to unattended inputs seen in behavioral studies1.

Filtering may be especially strong and early in spatial cue-ing tasks. Inevitably, tasks will vary in how rapidly the relevantand irrelevant objects can be distinguished and irrelevant pro-cessing suppressed. In a previous study of PFC, for example,monkeys entered the position of a specified target object intoworking memory, ignoring other objects in the visual display6.In that task, evidence for selective coding of the target object’sposition began a little later than in the present case, around 140 ms from stimulus onset and following an initial on-discharge.In the frontal eye field, cells show stronger responses to targetthan to non-target stimuli in visual search. This attentional mod-ulation occurs relatively late if target identity varies from trial totrial7, but has been reported from the onset of visual activity ifmonkeys have long experience searching for the same target23.

Attentional modulation—a relative enhancement of responseto an attended stimulus, or a relative suppression of response to an ignored stimulus—has been reported in many parts of the visual system, including striate and prestriate cortex24–28,parietal cortex29,30 and inferotemporal cortex31,32. Though dif-ferences in task, training regime and such prevent direct comparisons with the present data, our results do suggest an

Fig. 4. Bilateral condition; activity in the selected cell sample forthe different stimulus combinations of target and non-targetobjects. (a) Mean spike density for attention to the preferred loca-tion and attention to the non-preferred location. Stimulus com-binations (see display icons) from top to bottom: targets in bothlocations; target in preferred location and non-target in non-preferred location; target in non-preferred location and non-targetin preferred location; non-targets in both locations. Red and blueicon outlines indicate color key, and were not present in actual dis-play. (b) For each stimulus combination (top to bottom, same orderas left panel), activities of individual PFC neurons are plotted forattention to the preferred location (abscissa) versus attention tothe non-preferred location (ordinate). Plotted values are spike den-sities from shaded regions of left panel, after baseline subtraction.


Similar to the discharge behavior in the unilateral condition,the neurons discriminated between target and non-targetobjects as long as these were attended. Though true for bothlocations, this was most apparent for the distinction betweentargets and non-targets in the preferred location (Fig. 4a, uppertwo panels versus lower two panels). On trials in which themonkey attended to the preferred location (blue lines), neuralactivity started to differ between target and non-target objectsin the period 110–120 ms after stimulus onset (t-test in 10-msbins, P < 0.05). This difference between targets and non-targets at the preferred location vanished if the monkey attend-ed to the opposite hemifield (red lines), although the visualstimuli were identical in both conditions.

To evaluate this effect in individual neurons, we computedselectivity indexes for the distinction between targets and non-targets at the preferred location (see Methods). For the subset ofdisplays with a non-target at the non-preferred location (Fig. 5a),22 of 33 neurons (67%) had a higher preferred-location selec-tivity if the monkey attended to that location (t-test, P < 0.02).We obtained a similar result for displays with a target at the

articles


Atte

nd n

on-p

refe

rred

loca

tion

activ

ity (

spik

es/s

)

0 800100 200 300 400 500 600 700

Preferredlocation

tend preferred location

Attend non-preferred locationp

t

referre

Act

ivity

(sp

ikes

/s)

5

10

15

20

25n = 33

Preferredlocation

tend preferred location

Attend non-preferredlocation

Act

ivity

(sp

ikes

/s)

5

10

15

20

25n = 33

PreferredlocationAttend preferred location

Attend non-preferred location

Act

ivity

(sp

ikes

/s)

5

10

15

20

25n = 33

Preferredlocation

Attend preferred location

Attend non-preferred locationAttend

Attend

Att

Act

ivity

(sp

ikes

/s)

5

10

15

20

25

n = 33

a

0

–20 –10 0 10 20 30 40 50–20

–10

10

20

30

40

50P = 0.28n = 33

P < 0.005n = 33

P < 0.05n = 33

Attend preferred location activity (spikes/s)

P < 0.005n = 33

0

–20 –10 0 10 20 30 40 50–20

–10

10

20

30

40

50

0

–20 –10 0 10 20 30 40 50–20

–10

10

20

30

40

50

0

–20 –10 0 10 20 30 40 50–20

–10

10

20

30

40

50

b

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

especially strong and global filtering of unattended inputs in PFC.Of particular significance is that, in many visual areas, thestrongest attentional filtering is spatially local. In a previous exper-iment similar to ours, monkeys were presented with streams ofstimuli in two locations while monitoring one of the locationsfor a potential target24. In both V2 and V4, strong filtering of theunattended location occurred only if stimuli were close together,within the same cell’s receptive field. With stimuli in oppositevisual fields, as in our study, filtering was weak or absent. Localeffects like this have been reported in a number of other studies,including studies of V2, V4 and the MT/MST complex25,26,28.(For an exception, see ref. 33.) This implies that much task-irrelevant information remains present at these early visual levels.In inferior temporal cortex, filtering of unattended inputs maybe hemisphere-specific, occurring only or most strongly if attend-ed and unattended stimuli lie in the same visual hemifield31,32.In human behavior, in contrast, the effects of selective attentionare more global. Attending to one input is associated withreduced processing of others throughout the visual field12,34,35,or sometimes even a different sensory modality36. In our data,similarly, suppression of the unattended input occurred eventhough attended and unattended locations lay in opposite hemi-fields. Our results suggest that, in the PFC, filtering of ignoredinputs may reach a level commensurate with the strong, globaleffects of selective attention in human behavior.

METHODSTask. Each trial started with the presentation of a fixation point (FP, whitedot, 0.2°) together with the outlines of two white boxes (2.5° × 2.5°) 6°left and right of the FP. As the monkey fixated the FP, a solid white square(2° × 2°) was flashed for 100 ms either inside the left or right box, indi-cating the cued side for the trial. After a delay period of 700 ms, asequence of stimulus presentations started. On the cued side, 1–4 stim-uli (each picture about 2° × 2°) were presented in turn, each remaining for300 ms, with a stimulus onset asynchrony (SOA) of 800 ms from onestimulus to the next. In the bilateral condition only, stimuli were alsosimultaneously presented on the uncued side. On the cued side, thesequence consisted of 0–3 non-targets followed by a single target (prob-ability of target was 0.3 for each stimulus until the fourth, for which prob-ability of target was 1.0). The monkey was required to maintain centralfixation (window, 2° × 2°) until the target appeared, then immediatelyto fixate it (response window 400 ms from target onset). On the uncuedside, the sequence consisted of a mixture of targets and non-targets. Anincorrect saccade or no response to the target stimulus on the cued sideresulted in a 2-s ‘time out’ followed by termination of the trial with noreward. Otherwise, a juice reward immediately followed the successfulsaccade to the cued target.

Recording methods. Using standard surgical techniques, two male rhe-sus monkeys (Macaca mulatta) were prepared for chronic, head-fixedsingle neuron recording and eye movement monitoring. All procedureswere approved by the Home Office of the United Kingdom and were incompliance with the guidelines of the European Community (EUVD86/609/EEC) for the care and use of laboratory animals. Horizontal andvertical eye movements were sampled at 1000 Hz using a magnetic searchcoil system (David Northmore Inst., Newark, Delaware). Neuron activ-ity was recorded extracellularly from the lateral prefrontal cortex with

commercially available dura-puncturing tungsten microelectrodes. Arraysof 2–6 electrodes were driven within the recording chamber by custom-designed screw mini-microdrives. The microdrives were mounted on aDelrin grid (Crist Inst., Hagerstown, Maryland) with 1-mm spacingbetween adjacent locations inside the recording chamber. Neural activi-ty was amplified, filtered, and stored for off-line cluster separation withthe Plexon MAP system (Plexon, Dallas, Texas). To ensure a relativelyunbiased sampling of PFC neural activity, we did not pre-screen neu-rons for task-related responses. Instead, we advanced the electrodes untilthe activity of one or more neurons was well isolated, and then data col-lection commenced. Within a session, the average number of stimuluspresentations receiving a correct response was 77 for the unilateral con-dition and 190 for the bilateral condition.

After the conclusion of the chronic experiments, recording locations weredetermined stereotaxically. The precise position of the recording chamberand the orientation of the Delrin grid within the chamber were measured.The eye coil and the dental acrylic implant were removed. The trephina-tion was enlarged by about 5 mm posteriorly and 5 mm medially. The duramater was cut to expose the arcuate and the posterior part of the principalsulcus. Several readings were taken to obtain the location and shape of bothsulci. The dura mater was then sewn, and the wound was closed.

Data analysis. Except for the specific error analysis (Fig. 3), stimuli asso-ciated with errors (broken or incorrect fixation, failure to fixate a cued-side target) were excluded from analyses of neural activity. Also excludedwere data from the fourth stimulus presentation on a trial, because inthis case, the attended object was always predictable. Data were neuralresponse rates after subtraction of baseline activity calculated over a 200-ms interval ending at stimulus onset.

Object selectivity in the unilateral condition was assessed using two-way ANOVAs on neural activity in the interval 100–300 ms after stimu-lus onset, evaluated at P < 0.01. Except as noted in the text, t-tests forother contrasts concerned neural activity in a window 100–200 ms fromstimulus onset. To prevent any influence of new visual input after an eyemovement, these t-tests responses to target stimuli were excluded if sac-cadic latency was below 150 ms (7% of the saccades in monkey A and6% in monkey B).

To determine the start of differences in neural activity between twoconditions, spike trains were convolved with a postsynaptic activationfunction with a binwidth of 1 ms37. This asymmetric activation wave-form is designed to mimic an excitatory postsynaptic potential. The meanactivity of a single neuron for a certain condition was calculated by aver-aging the single trial activities. Comparisons between two conditions

articles


0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Preferredlocation

Non-preferredlocation

Preferredlocation

Non-preferredlocation

Selectivity, attend preferred location

Sel

ectiv

ity, a

ttend

no

n-pr

efer

red

loca

tion

Selectivity, attend preferred location

Sel

ectiv

ity, a

ttend

no

n-pr

efer

red

loca

tion

P < 0.05n = 33

a

bP < 0.005n = 33

Fig. 5. Influence of attention on object selectivity at the preferred loca-tion. The object selectivity index of individual PFC neurons is plotted forattention to the preferred location (abscissa) against attention to the non-preferred location (ordinate). Higher values of selectivity indicate greaterdifferentiation between displays with target versus non-target at the pre-ferred location. Preferred-location selectivity is separately calculated fordisplays with non-target (a) or target (b) at non-preferred location.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


were then made with t-tests on mean activity in 10-ms bins. Convolu-tion with the same activation function was also used in construction ofpost-stimulus histograms.

As a measure of stimulus selectivity (target versus non-target) in thepreferred location (see text and Fig. 5), the selectivity index was definedas the following6.

S = 2 – (ra + rb)/rmax

Here, ra = activity with target in preferred location; rb = activity withnon-target in preferred location; rmax = maximum activity.

Gap saccade task. For a subset of neurons, data were also gathered ina standard gap saccade task. Each trial started with the presentation ofa fixation point (FP; white filled circle, 0.2° diameter). The monkeymaintained steady fixation on this point for 700–900 ms, after whichthe FP was extinguished, and there was a period of 200 ms of no visu-al stimuli (gap period) before a peripheral target stimulus (white cir-cle, 0.2° diameter) was presented. The target was presented at one oftwo possible horizontal locations, either 6° to the left or 6° to the rightof the FP. Across trials, locations were pseudorandomly interleaved withequal probability. The monkey received a juice reward if it started fix-ation, maintained steady fixation during the visual fixation and the gapperiods and made a saccade to the target within 400 ms after its appear-ance. Otherwise, the trial was terminated and a 2-s time-out periodwas imposed.

AcknowledgmentsThe authors thank E.K. Miller and T. Norden-Krichmar for assistance in setting

up Cortex, W. Asaad for supplying the SpikeToolbox, and S. Mygdal and

M. Brown for surgical assistance.



1. Rock, I., Linnett, C. M., Grant, P. & Mack, A. Perception without attention:results of a new method. Cogn. Psychol. 24, 502–534 (1992).

2. Broadbent, D. E. Perception and Communication (Pergamon, London, 1958).3. Desimone, R. & Duncan, J. Neural mechanisms of selective visual attention.

Annu. Rev. Neurosci. 18, 193–222 (1995).4. Miller, E. K. & Cohen, J. D. An integrative theory of prefrontal cortex

function. Annu. Rev. Neurosci. 24, 167–202 (2001).5. Duncan, J. An adaptive coding model of neural function in prefrontal cortex.

Nat. Rev. Neurosci. 2, 820–829 (2001).6. Rainer, G., Asaad, W. F. & Miller, E. K. Selective representation of relevant

information by neurons in the primate prefrontal cortex. Nature 393,577–579 (1998).

7. Schall, J. D., Hanes, D. P., Thompson, K. G. & King, D. J. Saccade targetselection in frontal eye field of macaque. I. Visual and premovementactivation. J. Neurosci. 15, 6905–6918 (1995).

8. Hasegawa, R. P., Matsumoto, M. & Mikami, A. Search target selection inmonkey prefrontal cortex. J. Neurophysiol. 84, 1692–1696 (2000).

9. Lumer, E. D., Friston, K. J. & Rees, G. Neural correlates of perceptual rivalryin the human brain. Science 280, 1930–1934 (1998).

10. Beck, D. M., Rees, G., Frith, C. D. & Lavie, N. Neural correlates of changedetection and change blindness. Nat. Neurosci. 4, 645–650 (2001).

11. Dehaene, S. et al. Cerebral mechanisms of word masking and unconsciousrepetition priming. Nat. Neurosci. 4, 752–758 (2001).

12. Posner, M. I., Snyder, C. R. R. & Davidson, B. J. Attention and the detection ofsignals. J. Exp. Psychol. 109, 160–174 (1980).

13. Van Voorhis, S. & Hillyard, S. A. Visual evoked potentials and selectiveattention to points in space. Percept. Psychophys. 22, 54–62 (1977).

14. Corbetta, M., Miezin, F. M., Shulman, G. L. & Petersen, S. E. A PET study ofvisuospatial attention. J. Neurosci. 13, 1202–1226 (1993).

15. Rainer, G., Asaad, W. F. & Miller, E. K. Memory fields of neurons in theprimate prefrontal cortex. Proc. Natl. Acad. Sci. USA 95, 15008–15013(1998).

16. Sakagami, M. & Niki, H. Encoding of behavioral significance of visual stimuliby primate prefrontal neurons: relation to relevant task conditions. Exp.Brain Res. 97, 423–436 (1994).

17. Funahashi, S., Bruce, C. J. & Goldman-Rakic, P. S. Mnemonic coding ofvisual space in the monkey’s dorsolateral prefrontal cortex. J. Neurophysiol.61, 331–349 (1989).

18. Watanabe, M. Prefrontal unit activity during delayed conditional Go/No-Godiscrimination in the monkey. II. Relation to Go and No-Go responses. BrainRes. 382, 15–27 (1986).

19. Watanabe, M. Prefrontal unit activity during delayed conditional Go/No-Godiscrimination in the monkey. I. Relation to the stimulus. Brain Res. 382,1–14 (1986).

20. Freedman, D. J., Riesenhuber, M., Poggio, T. & Miller, E. K. Categoricalrepresentation of visual stimuli in the primate prefrontal cortex. Science 291,312–316 (2001).

21. Jiang, Y., Haxby, J. V., Martin, A., Ungerleider, L. G. & Parasuraman, R.Complementary neural mechanisms for tracking items in human workingmemory. Science 287, 643–646 (2000).

22. Shulman, G. L., Ollinger, J. M., Linenweber, M., Petersen, S. E. & Corbetta,M. Multiple neural correlates of detection in the human brain. Proc. Natl.Acad. Sci. USA 98, 313–318 (2001).

23. Bichot, N. P., Schall, J. D. & Thompson, K. G. Visual feature selectivity infrontal eye fields induced by experience in mature macaques. Nature 381,697–699 (1996).

24. Luck, S. J., Chelazzi, L., Hillyard, S. A. & Desimone, R. Neural mechanisms ofspatial selective attention in areas V1, V2, and V4 of macaque visual cortex. J. Neurophysiol. 77, 24–42 (1997).

25. Moran, J. & Desimone, R. Selective attention gates visual processing in theextrastriate cortex. Science 229, 782–784 (1985).

26. Reynolds, J. H., Chelazzi, L. & Desimone, R. Competitive mechanismssubserve attention in macaque areas V2 and V4. J. Neurosci. 19, 1736–1753(1999).

27. Roelfsema, P. R., Lamme, V. A. F. & Spekreijse, H. Object-based attention inthe primary visual cortex of the macaque monkey. Nature 395, 376–381(1998).

28. Treue, S. & Maunsell, J. H. Attentional modulation of visual motionprocessing in cortical areas MT and MST. Nature 382, 539–541 (1996).

29. Bushnell, M. C., Goldberg, M. E. & Robinson, D. L. Behavioral enhancementof visual responses in monkey cerebral cortex. I. Modulation in posteriorparietal cortex related to selective visual attention. J. Neurophysiol. 46,755–772 (1981).

30. Gottlieb, J. P., Kusunoki, M. & Goldberg, M. E. The representation of visualsalience in monkey parietal cortex. Nature 391, 481–484 (1998).

31. Chelazzi, L., Duncan, J., Miller, E. K. & Desimone, R. Responses of neurons in inferior temporal cortex during memory-guided visual search. J. Neurophysiol. 80, 2918–2940 (1998).

32. Sato, T. Effects of attention and stimulus interaction on visual responses ofinferior temporal neurons in macaque. J. Neurophysiol. 60, 344–364 (1988).

33. Seidemann, E. & Newsome, W. T. Effect of spatial attention on the responsesof area MT neurons. J. Neurophysiol. 81, 1783–1794 (1999).

34. Baynes, K., Holtzman, J. D. & Volpe, B. T. Components of visual attention.Alterations in response pattern to visual stimuli following parietal lobeinfarction. Brain 109, 99–114 (1986).

35. Duncan, J. et al. Systematic analysis of deficits in visual attention. J. Exp.Psychol. Gen. 128, 450–478 (1999).

36. Dell’Acqua, R., Turatto, M. & Jolicoeur, P. Cross-modal attentional deficits inprocessing tactile stimulation. Percept. Psychophys. 63, 777–789 (2001).

37. Thompson, K. G., Hanes, D. P., Bichot, N. P. & Schall, J. D. Perceptual andmotor processing stages identified in the activity of macaque frontal eye fieldneurons during visual search. J. Neurophysiol. 76, 4040–4055 (1996).

articles©

2002

Nat

ure

Pu

blis

hin

g G

rou

p

htt

p:/

/neu

rosc

i.nat

ure

.co

m

As life in our culture becomes more demanding, average nightlysleep is decreasing in all segments of the population (Stein, M.,National Sleep Foundation Poll, 2001, www.sleepfoundation.org/PressArchives/lessfun_lesssleep.html). ‘Power naps’—brief periodsof daytime sleep lasting an hour or less—improve alertness, pro-ductivity and mood1,2, especially under sleep-deprived conditions3,4,during nightshift work5 and during prolonged periods of driving6.Although naps have been shown to enhance psychomotor speed aswell as short-term memory acquisition7,8, the effect of daytime napson previously learned information is not known. The finding thatpower naps are common among people reporting daily informa-tion overload indicates that napping supports a previously unknownmechanism of off-line information processing, perhaps related tothat which normally occurs during nocturnal sleep9–16.

We investigated the phenomenon of information overload atthe perceptual level. Typically in visual perception tasks, fast learn-ing happens in the first minutes to hours of training17,18. Previousstudies using a visual texture discrimination task (TDT)19 showthat a slower phase of perceptual learning also exists, whichdepends on nocturnal sleep after training9,14,19–22. The slow phaseof improvement becomes evident only after at least six hours ofnocturnal sleep14, and sleep deprivation the night after trainingeliminates the normal post-sleep improvement, even when mea-sured after two full nights of recovery sleep15. The improvementseen in subjects who sleep for eight hours during the night aftertraining correlates with the proportion of deep, slow wave sleep(SWS) in the first quarter of the night and with the proportion ofrapid eye movement sleep (REM) in the last quarter14. These resultsindicate that a full night of sleep is important for maintenance andconsolidation of experience-dependent learning, and that withoutat least six hours of sleep, this potential consolidation is lost. Thesestudies do not, however, address the question of how power napsof an hour or less could aid in such information processing.

Here we show that perceptual performance declined on theTDT with repeated, within-day training. In the context of this

The restorative effect of naps onperceptual deterioration

Sara C. Mednick1, Ken Nakayama1, Jose L. Cantero2, Mercedes Atienza2, Alicia A. Levin2, Neha Pathak2 and Robert Stickgold2

1 Department of Psychology, Harvard University, 33 Kirkland Street, Cambridge, Massachusetts 02138, USA2 Laboratory of Neurophysiology and Department of Psychiatry, Massachusetts Mental Health Center, Harvard Medical School,

74 Fenwood Road, Boston, Massachusetts 02115, USA

Correspondence should be addressed to S.C.M. ([email protected])


Human performance on visual texture discrimination tasks improves slowly (over days) in theabsence of additional training. This ‘slow learning’ requires nocturnal sleep after training and is lim-ited to the region of visual space in which training occurred. Here, we tested human subjects fourtimes in one day and found that with repeated, within-day testing, perceptual thresholds actuallyincreased progressively across the four test sessions. This performance deterioration was preventedeither by shifting the target stimuli to an untrained region of visual space or by having the subjectstake a mid-day nap between the second and third sessions.

deterioration, we found that (i) a daytime nap, but not an equiv-alent period of rest without visual input, reversed the deteriora-tion, (ii) the deterioration was retinotopically specific and (iii)neither an increase in subject motivation nor a decrease in taskdifficulty improved performance.

RESULTSCan too much practice be detrimental?To investigate whether repeated within-day testing on a percep-tual learning task can impair performance, subjects were tested onthe TDT four times in a single day (at 9 a.m., 12 p.m., 4 p.m. and7 p.m.). Each session lasted approximately 60 minutes. For eachsession, the speed of perceptual processing was calculated as thethreshold target-to-mask interstimulus interval (ISI) needed toachieve 80% accuracy. Thirty subjects were randomly assignedto one of three groups: control, long nap or short nap. Controlsubjects (n = 10) showed a 52% slowing in perceptual process-ing across the four test sessions (Fig. 1, filled circles; P = 0.0003,repeated measures analysis of variance (ANOVA) and post hoctests). Thus, with each successive session, subjects needed increas-ingly longer exposures to the stimuli to reliably identify targets.Performance deteriorated despite all testing being done within12 hours of morning awakening, a time when one would not nor-mally expect to see cognitive impairment, and without prior sleepdeprivation. Subjects averaged 6.92 ± 0.77 (s.d.) hours of sleepon the night before testing.

Can daytime sleep reverse perceptual deterioration?As nocturnal sleep is known to enhance alertness and to consol-idate TDT learning9,14,19,22, we asked whether a daytime napmight stop or even reverse the process of deterioration seen withrepeated within-day testing. The remaining 20 subjects were ran-domly assigned to a long (60-minute) or short (30-minute) napcondition. All subjects, including no-nap controls, performedthe task four times during the day; experimental subjects took a

articles


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


nap at 2 p.m.—midway between the second and third test ses-sions. As predicted, napping significantly affected subsequentperformance (P = 0.001, group × session interaction, mixed-model ANOVA): short naps prevented the normal deteriorationthat was seen in test sessions 3 and 4 (Fig. 1, open circles) andlong naps reversed the deterioration seen in the second test ses-sion (Fig. 1, open triangles). Thus, whereas controls showed a14.1-ms increase in ISI threshold between the second and thirdsessions, the short nap group showed no change (<1-ms increase)and the long nap group showed a 20.9-ms decrease (P = 0.03,paired t-test). The short nap group showed no significant changein ISI thresholds across the last three sessions (P = 0.94), but didshow significantly lower thresholds in the fourth session than didcontrols (P = 0.01). The long nap group showed significantly bet-ter performance than controls on both the third and fourth ses-sions (P = 0.03). We compared the distributions of sleep stagesin long naps versus short naps (Table 1) and found no signifi-cant differences in the amount of time spent in stage 1 or stage 2sleep. There was, however, a difference for SWS and REM sleep:the long naps contained 4.1–4.4 times more SWS and REM sleepthan the short naps did. This matches the SWS and REM depen-dency reported for overnight improvement on the task14, sug-gesting that processes occurring during SWS and/or REMunderlie sleep-induced perceptual recovery.

We compared naps taken on the day of TDT testing with base-line naps taken on a different day (see Methods) and found thatthe long nap group spent significantly more time in SWS on thetest day (27.4 versus 20.9 minutes, P < 0.05, one-tailed paired t-test) at the expense of time in stage 2 sleep (17.9 versus 25.0 minutes, P = 0.06) (Fig. 2). This large increase in SWS in test-day naps (31% over baseline) suggests that SWS is crucial for post-nap performance, perhaps through stabilizing and consolidatingplastic neuronal changes from earlier in the day. Such a functionhas previously been proposed for nocturnal SWS (refs. 23–25).

The role of REM, however, is less clear.The increase in time spent in REM duringtest-day naps compared with baseline naps(60% increase) was larger than that seen forSWS, but was not statistically significant(Fig. 2). Similar post-training REM sleepincreases have been reported in animal stud-ies of sleep-dependent learning26. In theshort nap condition, there was no signifi-cant change in SWS (9.2 minutes on testday versus 6.7 minutes baseline) or REM(0.8 minutes on test day versus 2.0 minutes

baseline) duration. Nap duration was controlled in both groupsand neither group showed a significant difference in nap durationbetween baseline and test-day naps (P > 0.27).

Sleep versus restTo test whether the benefits of naps resulted from the absence ofvisual input rather than from sleep itself, subjects (n = 9) repeat-ed the long nap protocol; but instead of sleeping, they rested qui-etly while blindfolded. Wake–sleep state was continuouslymonitored physiologically to ensure maintained wakefulness.The hour of rest without visual input did not have a restorativeeffect: subjects still showed performance decrements at 4 p.m.and 7 p.m. ISI thresholds increased by an average of 29.0 msbetween the second and fourth session (P < 0.05), a performancedecrement nearly identical (P = 0.31) to the 32-ms increase seenin controls (Fig. 1).

Motivation and task difficultyTo test whether a decrease in motivation contributed to the per-formance deterioration, we informed subjects (n = 10) after theirsecond session that their performance had worsened, and that theywould receive a cash bonus if they subsequently returned to theirbaseline performance. Even with this incentive, none of the sub-jects regained baseline performance during the third or fourth ses-sions, and mean ISI thresholds were 32.2 ms longer on the fourthsession compared to the first session (P = 0.001).

To test whether the performance decrement resulted specif-ically from exposure to more difficult trials, we tested ten sub-jects throughout the day in four sessions, with all blocks in thefirst three sessions run at the longest ISI (400 ms). Subjects werethen tested in the fourth session with the standard 25 blocks ofdecreasing ISIs, and their performance in the fourth session didnot differ from that of controls (P = 0.50). Thus, despite hav-ing an easier task in the first three sessions, subsequent perfor-mance still declined.

What aspect of processing is impaired?Several mechanisms might underlie this deterioration. A gener-alized fatigue effect, mediated by a decrease in alertness or atten-tional resources, is one possibility that is consistent with our data.Alternatively, we propose that specific neural networks in pri-mary visual cortex gradually become saturated with informationthrough repeated testing, preventing further perceptual process-ing. This would cause a training-specific deterioration in per-

articles

Fig. 1. Beneficial effect of napping. Performance of control and bothshort and long nap groups across same-day sessions. ISI thresholds fromthe first session defined baseline performance, and were subtractedfrom thresholds for subsequent sessions to determine relative change inperformance.

Table 1. Characteristics of short and long naps.

S1 (min) S2 (min) SWS (min) REM (min) ∆Threshold (ms)Short naps 6.2 ± 1.1 14.3 ± 1.5 6.6 ± 2.1 2.0 ± 1.4 4.5 ± 10.7Long naps 5.5 ± 0.7 17.9 ± 3.0 27.4 ± 3.8 8.9 ± 3.4 20.8 ± 7.6P value 0.85 0.22 0.0001 0.07

The mean times spent in each sleep stage during short (30-min) and long (60-min) naps are presented asmin ± s.e.m. S1, stage 1; S2, stage 2; SWS, slow wave sleep (stages 3 and 4); REM, rapid eye movementsleep. P values from unpaired t-test comparing times in each stage for long versus short naps.∆Threshold, the difference between ISI thresholds on the second (pre-nap) and third (post-nap) tests,expressed as ms ± s.e.m.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

ceptual processing. Whereas the generalized fatigue hypothesispredicts that decrements in performance would be widespreadand largely task-independent, our hypothesis of a training-specificdeterioration predicts that the performance decrements wouldbe restricted to behaviors mediated by the specific neural net-works previously involved in processing the target stimuli. Wereasoned that because TDT learning does not transfer tountrained portions of the visual field20, there should be notraining-specific deterioration if stimuli are presented to anuntrained region of the visual cortex.

To test this hypothesis, we trained 24 subjects as before, butfor half of them we switched the target stimuli to the con-tralateral visual field for the fourth and final same-day test ses-sion. Performance of the switch group did not differsignificantly from that of the control group across the first threesessions, but, unlike the control group, the switch groupshowed significant recovery in the fourth session (Fig. 3, P =0.002, ANOVA group × session interaction and post hoc test).Performance during the switch condition was not significant-ly worse than it was during the first session, indicating that thebehavioral deterioration observed in the trained visual quad-rant did not transfer to the untrained contralateral quadrant.These results strongly support the training-specific deteriora-tion hypothesis, and are contrary to the predictions of the gen-eralized fatigue hypothesis.

Further evidence against the generalized fatigue hypothesiscomes from the dissociation between improved performance andsubjective levels of sleepiness. If the steady decrease in perfor-mance throughout the day in the control group resulted from ageneral fatigue effect, then one should see a parallel increase inreported sleepiness. But no such increase was seen, and meanlevels of subjective sleepiness on the first and last tests were iden-tical (P = 0.45, repeated measures ANOVA). Similarly, the switchgroup showed no significant change in sleepiness across sessions(P = 0.49). In contrast, sleepiness decreased from the first to thelast session in the nap groups (P < 0.03, ANOVA and post hoctests). Thus, the switch group showed the same improvement inperformance as the nap groups did, but without a similar decreasein subjective sleepiness; compared to controls, the switch group

showed the same maintenance in degree of sleepiness, but nodeterioration in performance.

Data from the switch group also eliminated another possibleexplanation, that of a strictly circadian effect. Although the con-trol data could have been explained as a circadian rather thanrepetition effect, the fact that shifting the stimulus to the con-tralateral visual field for the last session reversed this decreaseeliminates this possibility. Thus, when subjects were tested at 7 p.m. with stimuli in an untrained region of visual space, theyperformed as well as they had at 9 a.m. the same morning.

DISCUSSIONTwo learning components occur with TDT testing: a fast, within-session component and a slow, sleep-dependent component16.Our study identified a third consequence of TDT training: withrepeated, same-day training, people require progressively longerISIs for texture discrimination. Such perceptual deteriorationhas not previously been reported for repeated same-day testingon other visual tasks. Task procedure may contribute to this dif-ference, such as whether the task measures vernier acuity27, res-olution acuity27 or texture discrimination19, whether stimuliare presented foveally17,18, parafoveally27 or at more peripher-al eccentricities19, and whether stimulus presentations are long(100–150 ms)17,18,27 or short (17 ms)19. Furthermore, a numberof perceptual learning protocols train subjects over severaldays17,18,27–29 rather than within-day, as in the current study,making it unclear whether fast or slow learning is occurring.With these caveats in mind, the present study shows that someforms of neural plasticity that require sleep for subsequent con-solidation and improvement of perception may actually hinderperformance before sleep.

Our three main findings—that there was a normal decline inTDT performance across the day with repeated exposure to thetask, that this decline was specific to previously trained regionsof visual space and that performance was restored by daytimenapping—have important implications. First, as circadian influ-ences have been ruled out, the performance decline must resultfrom specific neuronal changes induced by the initial testing peri-od. Second, as brain regions involved in higher levels of visual

articles


Fig. 2. Comparison of test-day and baseline long naps. The number ofminutes spent in stage 1 (S1), stage 2 (S2), rapid eye movement (REM)and slow wave sleep (SWS; stage 3 and 4) during baseline (gray bars) andtest-day (solid bars) naps. * P < 0.05; (*) P = 0.06.

Fig. 3. Beneficial effect of shifting stimulus location. Solid bars, deterio-ration in performance of control subjects during the third (T3) andfourth (T4) sessions of the day. Gray bars, deterioration in performanceof experimental ‘switch group’ subjects during T3 followed by recoveryduring T4 when stimulus location was shifted.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


processing lack retinotopic specificity, the critically affected neu-rons are most likely located in early visual processing areas. Final-ly, these initially affected neural networks must be further alteredduring napping to reverse the performance decrement.

We propose that the performance decrement seen here was adirect consequence of a mechanism for preserving informationthat has been processed but has not yet been consolidated intomemory by sleep. As this hypothetical limited-capacity mecha-nism becomes saturated with task-specific information, the localneural network’s ability to process on-line information during taskperformance worsens, resulting in the performance decrement.Several findings support this relationship between informationprocessing and the performance decrement. First, the retinotopicspecificity of the performance decrement is consistent with previ-ous studies showing that TDT learning is similarly retinotopic19

and dependent on early stages of visual processing19,30. Second,post-training sleep is known to be critical for stabilization and con-solidation of TDT learning9,14,15 and we now have shown that a60-minute nap reverses the performance decrement.

Although we cannot exclude a function for REM in thisprocess, it seems most likely that SWS has the central role.Although both SWS and REM have previously been implicatedin the nocturnal, sleep-dependent consolidation of this task9,14,the REM-dependent period appears four hours after the SWS-dependent period has ended14—well beyond the timeframe ofthese naps. We posit that during SWS, mechanisms of corticalplasticity lead to secondary changes in the TDT-trained neuralnetworks. Thus SWS serves as the initial processing stage of expe-rience-dependent, long-term learning and as the critical stage forrestoring perceptual performance. Roles for SWS in memory con-solidation have been proposed by others9,20,23–25.

This example of a training-induced deterioration in perfor-mance has several additional implications. First, it indicates thatthe cognitive benefits of sleep can be studied over a very shorttime period and do not require sleep deprivation or overnightsessions of sleep. This provides a more favorable set of conditionsto study the involvement of sleep in information processing andperformance. Second, it suggests that the psychological sensa-tion of ‘burnout,’ described anecdotally as increased irritationand frustration along with decreased effectiveness after prolongedcognitive effort, may not reflect a general mental fatigue, butrather the specific need of an overused local neural network toenjoy the restorative benefits of sleep.

METHODSA total of 129 undergraduates gave informed consent to participate in thestudy, which was approved by the Harvard University Department of Psy-chology internal review board. All subjects had normal or corrected-to-normal vision and no history of neurological, mental or physical illness.Each subject was tested on the TDT four times in one day: at 9 a.m.,12 p.m., 4 p.m. and 7 p.m. Tests normally lasted 60–75 min and includ-ed 1,250 trials. Participants were asked to discriminate the shape of atarget in one of the lower quadrants of the display at 2.5–5.0° eccen-tricity from the center. Either a ‘T’ or an ‘L’ appeared in the center,where fixation was maintained. (Except where noted, target arraysalways appeared in the lower left quadrant.) The target consisted of ahorizontally or vertically oriented array of three diagonal bars againsta background of horizontal bars. For each trial, the following sequencewas shown: target screen for 16 ms, blank screen for a variable periodof time, and then a mask for 16 ms. After each trial, subjects report-ed both the letter (T or L) at the central fixation point and the orien-tation of the diagonal bar array (horizontal or vertical). For eachsession, the speed of perceptual processing was calculated as thethreshold target-to-mask ISI needed to achieve 80% accuracy. The ISIthreshold provides a measure of the minimal effective stimulation

needed for perception of the target. This task was carried out as pre-viously described14,19. To examine the amount of change from base-line, difference scores were calculated by subtracting the threshold forthe first session from those of the second, third and fourth. Subjectsin all groups kept sleep logs for the week before testing, and no sig-nificant between-group differences were found in sleep patterns.

Thirty subjects were randomly assigned to one of the nap conditionsor to the previously described control condition, permitting statisticalcomparisons between all three groups. Naps began at 2 p.m. and wererecorded polysomnographically, with standard electroencephalographic(EEG), electro-oculographic (EOG) and electromyographic (EMG) mea-sures. Subjects were allowed to sleep until they had completed either afull half-hour or a full hour of polysomnographically identified31 sleepand then were woken by the experimenter. Sleep stages were subsequentlyrescored off-line. On re-scoring, one subject in the short nap group wasfound to have slept for only 9.5 min, and was excluded from all analy-ses. Subjects were recorded during naps on two separate days: the TDTtest day and a control day (‘baseline nap’) either one week before or oneweek after the test day, with the order balanced across subjects. Subjectsin the short nap group averaged 29.1 ± 2.9 (mean ± s.d.) min of sleep onthe test day; subjects in the long nap group averaged 59.6 ± 5.5 min.

For the quadrant switch study (24 switches; 22 controls), peripheraltargets were presented in the lower left or lower right quadrant (balancedacross subjects) for the first three test sessions. Then, for the fourth session,the targets were switched to the opposite lower quadrant for the switchgroup. For the quiet rest condition, nine subjects followed the long napprotocol but were instructed not to sleep during the nap hour. Subjectswere blindfolded to prevent visual stimulation and listened to an audiotape of short stories. Each subject’s wake state was monitored with theNightcap32 (HealthDyne Technologies, Marietta, Georgia, USA) and sub-jects were alerted at the first indication of impending sleep. At the startof the third session, subjects (n = 10) in the motivation protocol wereoffered a bonus of $25 if they could regain their initial performance levelduring the next two sessions. Subjects (n = 12) in the fixed ISI protocolperformed the standard protocol but with the ISI set to 400 ms for all 25blocks during test sessions 1–3. For the fourth session, subjects followedthe standard protocol with decreasing ISIs across the 25 blocks. All subjectsrated their sleepiness using the Stanford sleepiness scale33, which ratessubjective sleepiness on a seven-point scale.

AcknowledgmentsThis research was supported by grants from the National Institutes of Health

(MH 48,832 and NS 26,985) and AFOSR (83-0320) and by fellowships to J.L.C.

and M.A. from the Spanish Ministry of Education and the NATO Scientific

Program, respectively.


RECEIVED 10 DECEMBER 2001; ACCEPTED 2 MAY 2002

1. Takahashi, M. & Arito, H. Maintenance of alertness and performance by abrief nap after lunch under prior sleep deficit. Sleep 23, 813–819 (2000).

2. Dinges, D. F. & Broughton, R. J. (eds.) Sleep and Alertness: Chronobiological,Behavioral and Medical Aspects of Napping (Raven, New York, 1989).

3. Bonnet, M. H. & Arand, D. L. Impact of naps and caffeine on extendednocturnal performance. Physiol. Behav. 56, 103–109 (1994).

4. Bonnet, M. H. The effect of varying prophylactic naps on performance,alertness and mood throughout a 52-hour continuous operation. Sleep 14,307–315 (1991).

5. Rosa, R. R. Napping at home and alertness on the job in rotating shiftworkers. Sleep 16, 727–735 (1993).

6. Horne, J. A. & Reyner, L. A. Counteracting driver sleepiness: effects ofnapping, caffeine, and placebo. Psychophysiology 33, 306–309 (1996).

7. Harma, M., Knauth, P. & Ilmarinen, J. Daytime napping and its effects onalertness and short-term memory performance in shiftworkers. Int. Arch.Occupational Environ. Health 61, 341–345 (1989).

8. Taub, J. M. Effects of habitual variations in napping on psychomotorperformance, memory and subjective states. Int. J. Neurosci. 9, 97–112(1979).

9. Gais, S., Plihal, W., Wagner, U. & Born, J. Early sleep triggers memory for

articles©

2002

Nat

ure

Pu

blis

hin

g G

rou

p

htt

p:/

/neu

rosc

i.nat

ure

.co

m

early visual discrimination skills. Nat. Neurosci. 3, 1335–1339 (2000).10. Plihal, W. & Born, J. Effects of early and late nocturnal sleep on priming and

spatial memory. Psychophysiology 36, 571–582 (1999).11. Smith, C. Sleep states and memory processes. Behav. Brain Res. 69, 137–145

(1995).12. Smith, C. T., Conway, J. M. & Rose, G. M. Brief paradoxical sleep deprivation

impairs reference, but not working, memory in the radial arm maze task.Neurobiol. Learn. Mem. 69, 211–217 (1998).

13. Stickgold, R. Sleep: off-line memory reprocessing. Trends Cogn. Sci. 2,484–492 (1998).

14. Stickgold, R., Whidbee, D., Schirmer, B., Patel, V. & Hobson, J. A. Visualdiscrimination task improvement: a multi-step process occurring duringsleep. J. Cogn. Neurosci. 12, 246–254 (2000).

15. Stickgold, R., James, L. & Hobson, J. A. Visual discrimination learningrequires post-training sleep. Nat. Neurosci. 2, 1237–1238 (2000).

16. Karni, A. & Sagi, D. The time course of learning a visual skill. Nature 365,250–252 (1993).

17. Fahle, M. Human pattern recognition: parallel processing and perceptuallearning. Perception 23, 411–427 (1994).

18. Fahle, M. & Edelman, S. Long-term learning in vernier acuity: effects ofstimulus orientation, range and of feedback. Vision Res. 33, 397–412 (1993).

19. Karni, A. & Sagi, D. Where practice makes perfect in texture discrimination:evidence for primary visual cortex plasticity. Proc. Natl. Acad. Sci. USA 88,4966–4970 (1991).

20. Plihal, W. & Born, J. Effects of early and late nocturnal sleep on declarativeand procedural memory. J. Cogn. Neurosci. 9, 534–547 (1997).

21. Stickgold, R., Schirmer, B., Patel, V., Whidbee, D. & Hobson, J. A. Visualdiscrimination learning: both NREM and REM are required. Sleep 21(Suppl.), 256 (1998).

22. Karni, A., Tanne, D., Rubenstein, B. S., Askenasy, J. J. M. & Sagi, D.Dependence on REM sleep of overnight improvement of a perceptual skill.Science 265, 679–682 (1994).

23. Contreras, D., Destexhe, A. & Steriade, M. Intracellular and computationalcharacterization of the intracortical inhibitory control of synchronizedthalamic inputs in vivo. J. Neurophysiol. 78, 335–350 (1997).

24. Destexhe, A., Contreras, D. & Steriade, M. Spatiotemporal analysis of localfield potentials and unit discharges in cat cerebral cortex during natural wakeand sleep states. J. Neurosci. 19, 4595–4608 (1999).

25. Sejnowski, T. J. & Destexhe, A. Why do we sleep? Brain Res. 886 (Suppl.)208–223 (2000).

26. Smith, C. & Rose, G. M. Evidence for a paradoxical sleep window for placelearning in the Morris water maze. Physiol. Behav. 59, 93–97 (1996).

27. Bettina, B. L., Levi, D. M. & Reich, L. N. Vision Res. 34, 1679–1690 (1995).28. Ball, K. & Sekuler, R. Direction-specific improvement in motion

discrimination. Vision Res. 27, 953–963 (1987).29. Watanabe, T., Nanez, J. E. & Saksaki, Y. Perceptual learning without

perception. Nature 413, 844–848 (2001).30. Karni, A., Weisberg, J., Lalonde, F. & Ungerleider, L. G. Slow changes in

primary and secondary visual cortex associated with perceptual skill learning:an fMRI study. Neuroimage 3 (Suppl.), S543 (1995).

31. Rechtschaffen, A. & Kales, A. A Manual of Standardized Terminology,Techniques and Scoring System for Sleep Stages of Human Subjects (BrainInformation Service, University of California, Los Angeles, 1968).

32. Cantero, J. L., Atienza, M., Hobson, J. A. & Stickgold, R. Nightcap: a reliablehome monitor for detecting normal sleep onset. Sleep 25, 238–245 (2002).

33. Hoddes, E., Zarcone, V., Smythe, H., Phillips, R. & Dement, W. C.Quantification of sleepiness: a new approach. Psychophysiology 10, 431–436(1973).

articles


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


articles

A fundamental question in the study of visual processing is theproblem of ‘feature selection’: which features of an image areextracted and represented by the visual cortex? Several brain areasare involved in visual object processing, and different features arerepresented at different stages. In the earliest processing stages,which involve the retina, lateral geniculate nucleus (LGN) and pri-mary visual cortex (V1), the image is represented by simple localfeatures such as center–surround receptive fields and oriented linesand edges. This encoding can arise from the computational prin-ciples of decorrelation and redundancy reduction1–3 or from faith-ful reconstruction of the input using sparse encoding4–6. After thisearly processing, moderately complex features are represented inareas V4 and the adjacent region TEO, and finally, partial or com-plete object views are represented in anterior regions of infer-otemporal (IT) cortex7–11.

Here we show, by computational analysis and simulations,that features of intermediate complexity and partial object viewsare optimal for visual object classification. These features wereautomatically selected when the system was set to maximize theinformation delivered with respect to a class of images, and thusserve as basic building blocks for representing the class. The sim-ulations show that IC features are more informative than verysimple or very complex ones, and that during visual classifica-tion, the extracted features have the capacity to generalize broad-ly to new exemplars within the class.

It has previously been proposed that complex objects are rep-resented in the visual cortex in terms of simpler elements suchas wavelets or Gabor basis functions4,5, both of which have beenused in object recognition models12. These ‘building blocks’ areuniversal in the sense that they are equally applicable to all naturalimages. Alternatively, we propose that the visual system encodesfeatures of intermediate complexity that are class-specific, that

Visual features of intermediatecomplexity and their use inclassification

Shimon Ullman, Michel Vidal-Naquet and Erez Sali

Department of Computer Science and Applied Mathematics, The Weizmann Institute of Science, PO Box 26, Rehovot 76100, Israel

Correspondence should be addressed to S.U. ([email protected])


The human visual system analyzes shapes and objects in a series of stages in which stimulus featuresof increasing complexity are extracted and analyzed. The first stages use simple local features, andthe image is subsequently represented in terms of larger and more complex features. These includefeatures of intermediate complexity and partial object views. The nature and use of these higher-order representations remains an open question in the study of visual processing by the primate cor-tex. Here we show that intermediate complexity (IC) features are optimal for the basic visual task ofclassification. Moderately complex features are more informative for classification than very simpleor very complex ones, and so they emerge naturally by the simple coding principle of informationmaximization with respect to a class of images. Our findings suggest a specific role for IC features invisual processing and a principle for their extraction.

is, selected for encoding images within a class of related images.These features are used after the encoding of simple features inV1 but before the encoding of complex object views in anterior ITcortex, and they are specifically selected to support visual classi-fication—one of the basic tasks of visual perception. Here wepresent examples of such features, the coding principle used toextract them (maximizing information for classification), theiradvantages and their biological implications.

RESULTSFeatures extracted by maximizing informationFrom a training set of 138 roughly-frontal face images and 40 side-view images of cars (examples in Fig. 1b), we extracted sets of ICfeatures or ‘fragments’ that are optimal building blocks for encod-ing those images (Fig. 1a, c and d). The fragments were extractedon the basis of maximizing the information delivered about theset of faces (or cars) using a search procedure. The search stored alarge number (>10,000) of sub-images (see Methods), measuredthe information delivered with respect to the training set for eachsub-image and extracted the most informative fragments.

The amount of information delivered by a candidate frag-ment about the class of images was calculated using the mutualinformation equation:

I(C, F) = H(C) – H(C F) (1)

In this equation, I(C, F) is the mutual information betweenthe fragment F and the class C of images, and H denotes entropy.This is a natural measure of information conveyed by F about C,as it measures how the uncertainty about the presence of the classC in the image is reduced by the possible presence of the frag-ment F in the image. In simulations, we found that simplified

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

articles


approximations to the mutual information measure, which maybe easier to implement biologically, were also effective.

The most informative face fragments extracted by this pro-cedure were selected successively (Fig. 1a). After finding the frag-ment with the highest mutual information score, the searchidentified the fragment that delivered the maximal amount ofadditional information, and so on. The i+1 fragment was select-ed to increase the mutual information of the fragment set bymaximizing the minimal addition in mutual information withrespect to each of the first i fragments. (A more comprehensivecriterion would be to select the i+1 fragment that maximizes theadditional mutual information with respect to the joint distrib-ution of all i previously selected fragments together, but this is amore demanding computation that requires substantially moretraining data.) After extracting the first eight fragments, the searchextracted additional fragments at the same locations, arrangedby location and decreasing mutual information (Fig. 1d).

Superiority of IC featuresThe most informative fragments were typically fragments ofintermediate size (Fig. 2). The amount of mutual information,defined as the fragment’s ‘merit’ (equation 1), was calculated forfragments of different size centered at the same image location.The merit typically peaked at an intermediate size (median 11%,s.d. 16% of object size). The superiority of intermediate-size frag-ments can be explained as the interplay of two factors: specifici-ty and relative frequency. A large face fragment can providereliable indication of the presence of a face in an image, although

the likelihood of encountering such a fragment in a novel faceimage is low. Consequently, the information carried by such afragment with respect to the class is limited. A smaller fragmenthas a higher likelihood of appearing in different face images, butthe likelihood of its presence in non-face images is also higher.The optimal fragments we found are considerably more complexthan V1-like receptive fields, but still correspond to local imagestructures rather than to the global shape templates that are usedin some current visual recognition models13,14.

We also found a superiority effect with respect to changes inresolution. A common approach in computer vision is to processimages at multiple resolutions15, and it has been suggested thatthe mammalian visual system also performs multi-resolutionprocessing using multiple receptive field sizes16. A plot of themerit of a face fragment as a function of image resolution (Fig. 2d, see Methods) shows that the mutual information peakedat intermediate resolution. This could explain why intermediateresolution face templates are computationally useful for facedetection17. For car images, high-merit fragments were typicallyof low resolution. On the whole, the highly informative featureswere of intermediate complexity: intermediate size at high reso-lution and larger size at intermediate resolution.

The features were selected by this procedure to support gener-alization and classification rather than economic reconstructionof the image5,18. An important difference between classificationand reconstruction schemes is that in classification, fragments aredetermined by images within as well as outside the class, resultingin features that are more informative than those selected for effi-

a b c

1st

2nd

3rd

4th

MeritWeight

0.206.5 5.5 6.45 5.45

0.18 0.18 0.17 0.16 0.11 0.10 0.093.52 2.9 2.9 2.86

d e

Fig. 1. Intermediate complexity visual features were chosen by maximizing delivered information with respect to a class of objects. (a) The best eightface features found, arranged around a face from the learning set (in decreasing order of mutual information from the top, moving counter-clockwise).(b) Examples of faces and cars in the training set. (c) Selected car fragments. (d) Additional fragments organized by type; first row same as (a) withmerit (mutual information in bits) and weight (log2 of the likelihood ratio, equation 2) shown. Each column corresponds to a single type, rowsarranged in decreasing levels of merit. (e) One fragment represented in terms of simpler sub-fragments.©

2002

Nat

ure

Pu

blis

hin

g G

rou

p

htt

p:/

/neu

rosc

i.nat

ure

.co

m


articles

Fig. 2. Superiority of intermediate fragments.Mutual information (merit, ) and weight () as afunction of fragment size (a–c) and resolution (d).(a,b) Examples of size effect on mutual informationand weight for two fragments. Horizontal axis, rela-tive size in terms of fragment area (the size of maxi-mum mutual information is defined as 1). Verticalaxis, merit × 100 (equation 1) and weight (equation2). (c) Average size effect on mutual information (n= 15). Horizontal axis as in (a and b). Vertical axis,relative mutual information ± s.d. (the maximalmutual information is defined as 1). (d) Effect ofimage resolution on mutual information and weight.The decrease in information was significant for a15% resolution change (P < 0.05). Horizontal axis,relative resolution in pixels across the fragment (theresolution of maximum mutual information isdefined as 1). Vertical axis as in (a).

cient reconstruction. In addition, representa-tions resulting from classification schemes con-tain overlapping fragments at multiple scalesthat may be redundant for reconstructing the training set.

Informative features are useful for classificationWe tested how useful the intermediate features were for classifica-tion, compared with more local and more global features. The maindifficulty in classification arises from the variability in shape with-in a natural class of objects19. We compared the matching of a novelimage using composition of fragments to that using full-face shapes(Fig. 3). The full-face condition searched the database for the mostsimilar stored face. The fragments condition matched the novel facewith fragments extracted from the same image set as the full-facecondition. The fragment-based approximation was markedly bet-ter (as judged by eight observers), showing that a modest number ofappropriate building blocks can be used in different combinationsto deal with shape variation within a class of face images.

We tested the generalization capacity of informative features byusing them to classify novel images that were substantially differ-ent from the training set. For the purpose of classification, the fea-tures were organized by type: different fragments covering the sameregion of the face, such as the hairline region, were grouped togeth-er into the same fragment-type group (Fig. 1d). We used this orga-nization because, in classifying an image, the best-matching fragments from each type are selected first, and then com-bined to produce the final decision. The features of a common typecan be represented in terms of simpler fragments (Fig. 1e), leadingto a hierarchical representation in which an intermediate fragmentis defined by the conjunction of lower level features. In this hierar-chical representation, an intermediate feature cannot be simplycharacterized by a preferred sub-image. It is defined instead by thepresence and arrangement of its own constituents, similar to the‘critical features’10 of intermediatecomplexity IT units.

Once a fragment F was detected in the image, the strength ofthe evidence it supplied for the presence of an object of class Cwas measured by the likelihood ratio (R) of fragment F beingfound inside and outside of class C:

(2)

This ratio is commonly used in signal detection, and it is anoptimal detection criterion for the presence of an object from C

R(F) =P(F C)

P(F C)

given the fragment F. We used w = log2(R(F)) as the ‘weight’ ofthe fragments. Note that the merit and weight of a feature aretwo different criteria, as a fragment can have a high weight butstill have a low merit. It is efficient for the visual cortex to usefeatures with high merit and to use them according to theirweights, which can be implemented by synaptic strengths. Toclassify an image as a face or non-face, the following formula wasused (see Methods for more details):

(3)

where Fikstands for the ith fragment of type k. This means that

for each type (such as ‘hairline region’), the maximal response of allfragments of this type is selected. The maximum operation is takenalso over different retinal positions within a region 25% of objectsize. We used a scheme similar to a biological model incorporat-ing the maximum operation20. The fragments detected from eachtype were then combined by summing their weights and compar-ing the sum to a threshold. The threshold was set to the lowestvalue that gave no false classification on a collection of test imagesthat contained no faces. Once the threshold was fixed, 200 novelface and 200 novel non-face images were tested yielding 97%detec-tion and 2.1% false detection, showing that a biologically plausiblecombination of informative features is competitive with currentclassification systems21 (see Fig. 4 for examples). Of particularinterest is the ability to generalize to novel exemplars, includingface paintings that are markedly different from the training set. Toshow the location of the detected object, each fragment was used toestimate the location of the object center (Fig. 4, white boxes). Todetect objects at different scales, the input image was re-sampled and analyzed at multiple (2–4) scales15.

We compared the classification results obtained using optimalfragments with those obtained using fragments that were selectedin a similar manner but with a fixed size. These fixed-size frag-ments were either smaller (one-third the size, 4% of average facearea) or larger (33% of face area) than the optimal fragments. Theywere spaced over a regular 5 × 6 grid, covering the same total areaas the optimal fragments, and applied to a new image set. The ICfragments were significantly better (95.6% detection, 0% falsealarms) than both the small (97% detection, 30.4% false alarms)

Σwik max(Fik

) > θk ik

15

10

5

00 1 2 3 0

01 2 3 4

123456

0 1 2 30

0.20.4

0.60.8

1.2

1

00.20.4

0.60.8

1.2

1

0 0.5 1 1.5 2

Mer

it (

×100

), w

eig

ht

Mer

it (

×100

), w

eig

ht

Mer

it (

×100

), w

eig

ht

Rel

ativ

e m

utu

al in

fo

Relative object size

a b

cd

Relative object size

Relative object size Relative resolution

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

articles


and large (39% detection, 0% false alarms) fragments. Thus, inter-mediate fragments selected for informativeness achieved marked-ly better classification than larger or smaller ones. The selectedfeatures were also more informative than were wavelet features ofdifferent orientations and scales. We further examined the use ofinformative fragments within a back-propagation neural network,with the fragments replacing the first-layer features selected by thenetwork itself. Classification results improved markedly, indicat-ing that incorporating informative fragments in standard networkmodels can enhance their classification performance.

Image rearrangement and spatial relationsA representation using multiple-scale, overlapping fragments wasuseful for enforcing the correct overall arrange-ment of the features. As there was no explicit rep-resentation of the exact location of the fragmentsor of their spatial relationships, the scheme mighthave confused a given shape with a shape con-structed from the same fragments arranged in adifferent configuration. However, owing to the useof overlapping fragments at multiple scales, cor-rect configurations were preferred by our model.The model’s response to faces at different levels ofrearrangement (Fig. 5) shows that with increaseddegree of rearrangement, fragments were increas-ingly lost and the overall response (equation 3)decreased. In contrast, uniform displacement ofthe entire figure by a similar amount had a small-er effect on the response. Similar effects have pre-viously been shown20,22,23 for the identification(rather than classification) of simple shapes.

Physiological studies show a gradual decreasein response as a function of rearrangement inmacaque IT neurons24, as well as in human visu-

Fig. 4. Face and car detection examples showing broadgeneralization. Fragments in Fig. 1 and equation (3)were used. (a) Detected faces and cars marked by out-line squares. (Cars were detected in reduced-resolution images.) (b) Individual face-like features canappear occasionally in non-face images, but the con-junction of a sufficient number to exceed threshold ishighly unlikely. (c) Images were tested at several scalesby re-sampling the input image.

al cortex (mapped by functional magnetic resonance imaging)25.Information regarding the relative arrangement of fragments iscaptured because the fragments themselves have a ‘jigsaw puzzle’property22: their shapes determine their possible interactions,and their assembly is often unique23. This is a useful propertyof IC features that is not included in schemes that use universalV1-type features directly for recognition and classification12,21.

DISCUSSIONOur results have two main implications for the problem of featureselection in visual processing. First, they show that visual featuresof intermediate complexity emerge naturally from a coding prin-ciple of maximizing the delivered information with respect to aclass of objects. Regardless of the particular mechanism used toextract the features, our results explain the relative advantage ofIC features for visual classification. Second, they show that visu-al features based on combinations of object fragments provide arich set of potential features from which informative ones can beeffectively selected. This stands in contrast to back-propagationand other network models for feature selection. Such models typ-ically start from randomly selected features and then seek toimprove them locally by small changes, and therefore perform a

Fragments Novel Full face Fig. 3. Approximating novel faces by fragments. Novel faces (middlecolumn) were approximated by full-face images from a data set (right)and by fragments extracted from the same set (left). For the full-facecondition, a novel face image was matched against the existing images inthe data set, and the best matching face was selected. In the fragmentscondition, the novel face was matched using the best matching fragmentof each type (borders between fragments were blurred to create asmoother image). The fragments improved the compensation for intra-class shape variability.

a

b c

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


articles

local search in the very large space of all possible shapes. Incontrast, the fragment-based scheme performs a moreglobal search in a restricted space of features, composedof combinations of common object parts. This results inmore informative features, probably because the unconstrainedsearch converges to a local optimum that is lower in informationthan the features obtained by fragment selection.

Unlike in many previously described schemes3–6, the featuresthat emerged here are not universal, but shaped by visual expe-rience with particular classes of objects7,10,11. For the task of visu-al object classification, these features are more informative thansimple generic features used by some recognition models, andalso more informative than global complex features selected byother schemes13,14. As a result, the IC fragment–based featuresprovide an efficient basis for classification and generalization. InIT cortex, many units encode partial rather than complete objectviews. It would be of interest to compare empirically the featurespreferred by IT with fragments selected by the model, and alsoto test whether such features can be shaped, as predicted, by clas-sification experience. With respect to resolution, the model sug-gests that if cortical cells are tested, the response of some unitswith large receptive fields will saturate at intermediate rather thanfull resolution.

In view of these findings and the fundamental role of classi-fication and generalization in vision, the visual system is likelyto use a similar coding principle that favors informative features,selected for their use in classification. Although specific neuralimplementations of the fragment selection process are beyondthe scope of this discussion, they may be based on neural con-nections that are facilitated by the co-occurrence of a feature anda class, and depressed by the occurrence of one without the other.Finally, in the overall hierarchy of features used by the visual sys-tem, this work suggests a distinct role for visual features of inter-mediate complexity. Local V1-like features provide an efficientencoding of natural images in general, global object views areuseful for the identification of specific objects under familiarviewing conditions26, and intermediate complexity features areoptimally suited to support generalization and classification.

METHODSImage sets. Training images for fragment extraction were 138 face imagesand 40 car images. Faces were roughly rescaled to 40 columns horizon-tally. Non-class images were a random collection including landscapes,fruits and toys with a similar gray-level range. Examples of the imagesas well as informative fragments extracted from them can be viewed athttp://www.wisdom.weizmann.ac.il/∼ michel/fbc/fbc.html. The site alsocontains additional details on the fragment extraction computation andcomparisons with alternative features.

Procedure. The search for informative fragments examined candidatefragments at multiple locations and sizes, where a fragment is a sub-image of some size p × q taken from one of the images. These sub-imageswere searched for in every database image and their mutual informationwas computed. The fragment-to-image comparison used a weighted sum

Fig. 5. Detection response (equation 3) decreases with degreeof image scrambling. The original image was cut into 2 × 2, 4 × 4or 8 × 8 sub-images, then scrambled. Left, average response (n = 10) for different degrees of scrambling. Example of onescrambled image shown above the curve. Vertical axis, theresponse for the original image is defined as 1. Horizontal axis,level of scrambling (8 denotes 8 × 8 sub-images). For compari-son, response is shown for three face (middle) and non-faceimages (right).

of gray-level gradient and orientation differences. We also tested nor-malized cross-correlation and the ordinal measure (in ref. 27) andobtained similar classification results.

To compute the mutual information I(C,F) in equation (1), we mea-sured the frequency of detecting a given fragment F in the database ofimages that contain or do not contain objects in the class, and assumedP(C) = 0.05. The entropy H(x) of a random variable x is given by–ΣP(x)log(P(x)). Here C and F are binary variables: F = 1 if the fragmentis found in the image and 0 otherwise; C = 1 if the image belongs to theclass in question and 0 otherwise. I(C,F) is then given by:

The measured frequencies depend on the detection threshold used;for each fragment, the threshold that maximized the mutual infor-mation was selected.

To make the search more efficient, the algorithm was divided intotwo stages. The first stage identified the approximate location and sizeof candidate fragments, by searching over a restricted range of sizes andlocations (steps of three pixels). The second stage made additional com-parisons centered around the locations and sizes of the fragments iden-tified in the first stage. We used a range of sizes from one-half the areaof Fi to twice the area of Fi, in integral number of pixels. The processwas repeated for 20 image resolutions spaced linearly over a total factorof ten. Image resolution was reduced by convolving the image with asmoothing kernel (cubic spline in Matlab, Mathworks, Natick, Massa-chusetts) that reduced the high-frequency cutoff of the image, and thenby re-sampling the image with a smaller number of pixels. For car frag-ments, low-resolution fragments were typically superior and all wereselected in low (20 × 40) resolution images. The most informative frag-ments were selected as described in the text. After the first eight facefragments, additional fragments were searched by types. For each type(for example, hairline region) the general location in the images wasmarked manually and the search proceeded in the marked regions. Intotal, we used 48 face and 28 car fragments (we also tested twice andthree-times as many fragments, which resulted in only a small improve-ment in performance). For each fragment, merit was computed bymutual information (equation 1). The weight of fragment Fi wasdefined by the log likelihood ratio (equation 2). For Fi = 1 (fragmentdetected), wi(1) = log2 [P(Fi = 1/C)/ P(Fi = 1/ C)], similarly for wi(0).We used equivalently the weight wi(1) – wi(0) when Fi = 1, and 0 oth-erwise. We used P(C) = 0.05, but found that fragment selection is insen-sitive to this value.

Classification experiments were done on a new set of 600 images, 200for each of three classes (face, car and non-class). To detect a face, for exam-ple, all fragments were searched at each image location. Detected frag-ments were combined within a 40-pixel search window using equation (3).This combination rule assumes conditional independence between frag-ments given the class variable, and often gives good results28. We also

I(C, F) = –P(C)Log(P(C)) –P(C)Log(P(C) + P(F)((P(C F)

Log(P(C F)) + P(C F)Log P(C F)))+ P(F)((P(C F)

Log(P(C F)) + P(C F)Log(P(C F))) (4)

1.4

1.2

0.8

0.6

0.4

0.2

1

1 2 4 80

Scrambling

Rel

ativ

e re

spo

nse

Unscrambled

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

articles


applied a more complex combination rule that takes into account pairwisecorrelations between fragments to a test set of car images, which resulted ina small increase in overall performance, and combination by a back-prop-agation network which resulted in an additional performance increment.

AcknowledgmentsWe thank J. Golberger, M. Bar and N. Rubin for helpful discussions. Supported

by Grant 99-28 CN-QUA.05 from the James S. McDonnell Foundation and by

the Moross Laboratory at the Weizmann Institute. Face images for testing were

in part from the to Carnegie Mellon University (CMU) face images database

http://www.cs.cmu.edu/∼ har/faces.html#upright.



1. Barlow, H. B. & Foldiak, P. in The Computing Neuron (eds. Durbin, R., Miall,C. and Mitchison, G.) 54–72 (Addison-Wesley, Reading, Massachusetts,1989).

2. Atick, J. J. & Redlich, N. A. What does the retina know about natural scenes?Neural Comput. 4, 196–210 (1992).

3. Bell, A. J. & Sejnowski, T. J. The ‘independent components’ of natural scenesare edge filters. Vision Res. 37, 3327–3338 (1997).

4. Field, D. J. What is the goal of sensory coding? Neural Comput. 6, 559–601(1994).

5. Olshausen, B. & Field, D. J. Emergence of simple-cell receptive fieldproperties by learning a sparse code for natural images. Nature 381, 607–609(1996).

6. Vinje, W. E. & Gallant, J. L. Sparse coding and decorrelation in primary visualcortex during natural vision. Science 287, 1273–1276 (2000).

7. Rolls, E. T. Neural organization of higher visual functions. Curr. Opin.Neurobiol. 1, 275–278 (1991).

8. Fujita, I., Tanaka, K., Ito, M. & Cheng, K. Columns for visual features ofobjects in monkey inferotemporal cortex. Nature 360, 343–346 (1992).

9. Gallant, J. L., Braun, J. & Van Essen, D. C. Selectivity for polar, hyperbolic,and cartesian gratings in macaque visual cortex. Science 259, 100–103 (1993).

10. Tanaka, K. Neuronal mechanisms of object recognition. Science 262, 685–688(1993).

11. Logothetis, N. K., Pauls, J., Bulthoff, H. H. & Poggio, T. Shaperepresentation in the inferior temporal cortex of monkeys. Curr. Biol. 5,552–563 (1995).

12. Wiskott, L., Fellous, J. M., Krüger, N., & von der Malsburg, C. Facerecognition by elastic bunch graph matching. IEEE Trans. Pattern Anal. Mach.Intell. 19, 775–779 (1997).

13. Turk, M. & Pentland, A. Eigenfaces for recognition. J. Cogn. Neurosci. 3,71–86 (1990).

14. Belhumeur, P. N., Hespanha, J. P. & Kriegman, D. J. Eigenfaces versusfisherfaces: recognition using class specific linear projection. IEEE Trans.Pattern Anal. Mach. Intell. 19, 711–720 (1997).

15. Ballard, D. H. & Brown, C. M. Computer Vision (Prentice-Hall, InglewoodCliffs, New Jersey, 1982).

16. DeValois, R. L., Albrecht, D. G. & Thorell, L. G. Spatial frequency selectivityof cells in the macaque visual cortex. Vision Res. 22, 545–559 (1982).

17. Brunelli, R. & Poggio, T. Face recognition: features versus templates. IEEETrans. Pattern Anal. Mach. Intell. 15, 1042–1052 (1993).

18. Lee, D. D. & Seung, H. S. Learning the parts of objects by non-negative matrixfactorization. Nature 401, 788–791 (1999).

19. Rosch, E., Mervis, C. B., Gray, W. D., Johnson, D. M. & Boyes-Braem, P. Basicobjects in natural categories. Cognit. Psychol. 8, 382–439 (1976).

20. Riesenhuber, M. & Poggio, T. Hierarchical models of object recognition incortex. Nat. Neurosci. 2, 1019–1025 (1999).

21. Yang, M. H., Kriegman, D. J. & Ahuja, N. Detecting faces in images: a survey.IEEE Trans. Pattern Anal. Mach. Intell. 24, 34–58 (2002).

22. Mel, B. W. SEEMORE: combining color, shape, and texture histogrammingin a neurally inspired approach to visual object recognition. Neural Comput.9, 777–804 (1997).

23. Ullman, S. & Soloviev, S. Computation of pattern invariance in brain-likestructures. Neural Net. 12, 1021–1036 (1999).

24. Vogels, R. Effects of image scrambling on inferior temporal corticalresponses. Neuroreport 10, 1811–1816 (1999).

25. Grill-Spector, K., Kushnier, T., Hendler, T., Edelman, S., Itzchak, Y. & Malach,R. A sequence of object-processing stages revealed by fMRI in humanoccipital lobe. Hum. Brain Mapp. 6, 316–328 (1998).

26. Logothetis, N. K. & Sheinberg, D. L. Visual object recognition. Annu. Rev.Neurosci. 19, 577–621 (1996).

27. Bhat, D. & Nayar, K. S., Ordinal measures for image correspondence. IEEETrans. Pattern Anal. Mach. Intell. 20, 415–423 (1998).

28. Friedman, N., Geiger, D. & Goldszmidt, M. Bayesian network classifiers.Mach. Learn. 29, 131–163 (1997).©

2002

Nat

ure

Pu

blis

hin

g G

rou

p

htt

p:/

/neu

rosc

i.nat

ure

.co

m


articles

Most current knowledge on cerebral processing of music comesfrom studies of normal listeners1–7. Positron emission tomogra-phy (PET)8,9, magnetic resonance tomography (MRI)10–12, elec-troencephalography (EEG)13–16 and magnetoencephalography(MEG)17–19 have all been used to identify differences betweenmusicians and non-musicians. EEG and MEG studies typicallyreport the late auditory evoked response (the evoked neuronalactivity that peaks about 100 ms after the onset of a sound).Intracranial EEG20, scalp EEG21 and MEG22,23 studies haveshown that the earliest evoked activity of the auditory cortexoccurs around 15–30 ms after stimulus onset with a negative–positive complex (N19-P30) in the auditory evoked potential.This complex reflects the postsynaptic neuronal activities of theinitial thalamocortical input to the auditory cortex, correspondsto the N19m-P30m component of the auditory evoked magnet-ic field and originates from the medial half of the first transversetemporal gyrus of Heschl20–23.

Cytoarchitectonic24–26, myeloarchitectonic26,27 and histo-chemical27–29 studies show that the granular core field inhumans, often referred to as the primary auditory cortex (PAC),is largely confined to the medial two-thirds of Heschl’s gyrus(HG), namely the anteromedial portion of Heschl’s gyrus(amHG)27–29. Considerable individual differences have beenreported, however, in the size and location of PAC along HG andrelative to the location of sulcal boundaries27–31. VolumetricMRI studies, therefore, can only provide gross anatomical land-marks for HG using the first transverse sulcus (FTS) as the ante-rior boundary and Heschl’s sulcus (HS) or sulcus intermedius(SI) as posterior boundaries32–34.

Plasticity in the frequency representation of the primary corefield indicates that it is involved in the fine discrimination of pitchand tonal pattern35,36. We therefore reasoned that the greater

Morphology of Heschl’s gyrusreflects enhanced activation in theauditory cortex of musicians

Peter Schneider1,2, Michael Scherg2, H. Günter Dosch1, Hans J. Specht1, Alexander Gutschalk2

and André Rupp2

1Department of Physics, University of Heidelberg, Philosophenweg 12, D-69120 Heidelberg, Germany2Department of Neurology, University Hospital Heidelberg, INF 400, D-69120 Heidelberg, Germany

Correspondence should be addressed to P.S. ([email protected])


Using magnetoencephalography (MEG), we compared the processing of sinusoidal tones in theauditory cortex of 12 non-musicians, 12 professional musicians and 13 amateur musicians. We foundneurophysiological and anatomical differences between groups. In professional musicians ascompared to non-musicians, the activity evoked in primary auditory cortex 19–30 ms after stimulusonset was 102% larger, and the gray matter volume of the anteromedial portion of Heschl’s gyruswas 130% larger. Both quantities were highly correlated with musical aptitude, as measured by psy-chometric evaluation. These results indicate that both the morphology and neurophysiology ofHeschl’s gyrus have an essential impact on musical aptitude.

tonal musical aptitude37 of musicians might have an anatomicalcorrelate in the auditory cortex. We found a substantial differ-ence in the magnetic field component (N19m-P30m) of the PACevoked by sinusoidal tones in musicians as compared to non-musicians. These neurophysiological differences correspondedwith anatomical differences in HG morphology and with differ-ences in musical aptitude.

RESULTSAuditory evoked responsesAfter grouping our 37 subjects a priori into non-musicians, amateur musicians and professional musicians according to musical ability, we recorded each subject continuously over bothhemispheres with a whole-head MEG system while sinusoidaltones with carrier frequencies of 100, 220, 500, 1,100, 2,500 and5,600 Hz were presented. Tones were 100% amplitude-modulatedto record the fast, steady-state response of the auditory cortex23.From the averaged steady-state responses, the primary N19m-P30m responses were computed off-line by deconvolution23. Thesource activity of the auditory cortex was calculated using equiv-alent dipoles21,38,39 in the medial portion of the left and right HG(Fig. 1 and Methods).

The primary N19m-P30m source activity of the auditory cor-tex evoked by tones with a carrier frequency of 500 Hz wasmarkedly different between subject groups (Fig. 2). Signal ampli-tudes were about twice as large in professional musicians as innon-musicians. Similar response differences were seen at theother five frequencies, but the responses were largest at 500 Hzin all groups. The primary N19m-P30m source activity averagedover all frequencies and both hemispheres was 102 ± 16% largerin professional musicians than it was in non-musicians (F1,22 =58.7, P < 0.0001). The increase was highly significant at each

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

articles


frequency and ranged from 78% (1,100 Hz, P < 0.0001) to 144%(5,600 Hz, P < 0.0001).

We next averaged the peak-to-peak N19m-P30m dipoleamplitudes for each group (Fig. 3a). Amateur musicians showed an intermediate average increase of 37 ± 11% over non-musicians (F1,23 = 7.8, P < 0.05). The differencebetween amateur musicians and non-musicianswas significant only in the low frequency range(<1,000 Hz, P < 0.05). There was a frequency ×group interaction (F10,165 = 2.9, P < 0.01) thatranged from 77% (100 Hz, P < 0.01) to 14%(2,500 Hz, nonsignificant (n.s.)).

In professional musicians, dipole amplitudeswere significantly larger in the right than in the lefthemisphere at all frequencies. On average, theN19m-P30m signal was 21 ± 9% larger in the righthemisphere (F1,11 = 47.3, P < 0.0001). Comparedto non-musicians, professional musicians had an

Fig. 2. Auditory evoked N19m-P30m signals and 3Dgray matter surface reconstructions of HG for all sub-jects aligned in the same order. Both the neurophysio-logical and the anatomical data show a large increase inprofessional musicians and a smaller increase in ama-teur musicians. Left, dipole strength of the primarycortical response at 500 Hz. Source activities of theright (thick lines) and left (thin lines) hemispheres aresuperimposed. Right, highlighted areas show theamHG for each subject, aligned in the same order asthe primary evoked responses.

average signal that was 115 ± 18% larger over all frequencies inthe right hemisphere (F1,22 = 73.4, P < 0.0001) and 87 ± 17%larger in the left hemisphere (F1,22 = 43.6, P < 0.0001). Non-musi-cians, by contrast, did not have significantly larger dipole ampli-tudes in the right than in the left hemisphere (5 ± 9%, F1,11 = 3.6,n.s). In amateur musicians, dipole amplitudes were 19 ± 14%larger in the right than they were in the left hemisphere (F1,12 =17.8, P < 0.01). This was significant at three frequencies (100 Hz,500 Hz and 1,100 Hz, P < 0.05).

There was a pronounced difference between the early and lateauditory cortical responses. Whereas the early N19m-P30m com-plex was much larger in musicians, the late N100m component,which was evoked by the onset of the sinusoidal tones, showedsimilar amplitudes in all three groups over all frequencies (Fig. 3b).For all groups, the N100m was largest around 1,000 Hz.

Morphology of Heschl’s gyrusThe large neurophysiological difference between musicians and non-musicians at the level of the PAC coincided with a large

Fig. 1. The auditory stimulus, evoked magnetic fields and corticalanatomy. (a) Stimulus waveform. A modulation frequency of 26–37 Hzwas superimposed on sinusoidal tones with carrier frequencies of100–5,600 Hz to measure the responses to tone onset and to each mod-ulation cycle. (b) Typical averaged response at an MEG sensor over theright auditory cortex shows middle latency onset components P30m andP50m, long latency components N100m and the sustained field (SF). Theresponses to the modulation cycles appear superimposed on the SF. (c) Typical early N19m-P30m response of the PAC after deconvolutionof the modulated signals. (d, e) Source model with one equivalent dipolein each hemisphere depicted in sagittal and transversal T1-weighted MRIimages. The transversal section is parallel to the supratemporal plane.The source activity is modeled with dipoles drawn in the left and righthemispheres. (f) Three-dimensional (3D) gray matter surface recon-struction of the right HG. The FTS defines the anterior boundary andthe most posterior HS defines the posterior boundary. The first trans-verse HG is sometimes divided by the SI, a shallow sulcus which does notextend over its full length.

a

b

c

d

e

f

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


articles

morphological difference. From the three-dimensional (3D) gray matter surface recon-structions of the right and left HG (Fig. 2,right), we estimated the volumes of the gray and white matter ofHG and its macroanatomical subdivisions (Fig. 1f). The wholeHG was defined as extending to the lateral border of the tempo-ral plane, with the FTS as the anterior boundary and the mostposterior HS as the posterior boundary. The anterior portion ofHeschl’s gyrus (aHG) was defined by the SI, if present, or other-wise by the most anterior HS as posterior boundary33, and theamHG was defined as the medial two-thirds of aHG.

The MRI-based volumetry of these structures showed muchlarger gray matter volumes in musicians than in non-musicians,particularly in amHG (Table 1). Averaged over both hemispheres,the amHG gray matter volume was 130 ± 23% larger in profes-sional musicians (F1,22 =71.2, P < 0.0001). The mean differencesby hemisphere were 122 ± 15% in the right and 136 ± 19% in theleft (F1,22 = 2.0, n.s.). The difference was considerably less, butstill significant, when the gray matter volumes of the whole HGor aHG were considered (aHG, 67 ± 19%, F1,22 = 18.1, P < 0.001;HG, 37 ± 15%, F1,22 = 8.3, P < 0.01). White matter volumes werenot significantly different between professional musicians andnon-musicians, apart from a slightly larger amHG volume inprofessionals (30 ± 16%, F1,22 = 4.5, P < 0.05). Gray matter had aconsiderably larger influence than did white matter on this

Fig. 3. Frequency dependence of the primaryN19m-P30m and the late N100m dipolemoments. Solid lines depict the data for theright hemisphere with standard error bars.Dashed lines depict the corresponding meangroup data for the left hemisphere with slightlysmaller early dipole moments. (a) The profes-sional musicians (circles) showed much largerearly response signals over all frequencies thannon-musicians (squares); the amateur musicians(triangles) had larger signals only at frequenciesbelow 1,000 Hz. (b) N100m dipole momentswere similar in all groups.

difference (tissue × group interaction, F1,22 = 77.1, P < 0.0001).Compared with non-musicians, amateur musicians showed sig-nificantly more gray matter volume in only the anterior portionsof HG (amHG, 60 ± 21%, F1,23 = 11.4, P < 0.01; aHG, 61 ± 22%,F1,23 = 10.7, P < 0.01). White matter volumes were not signifi-cantly different between amateurs and non-musicians.

For all three groups, there was a strong correlation (r = 0.87, P < 0.0001) between the neurophysiological (individual peak-to-peak N19m-P30m dipole amplitudes averaged over all frequenciesand both hemispheres) and the anatomical (mean gray mattervolume of amHG) parameters (Fig. 4a). This functional–anatomical correlation was also significant within each group:non-musicians (r = 0.68, P < 0.01), amateurs (r = 0.67, P < 0.01)and professionals (r = 0.58, P < 0.05). Combining the professionaland amateur musicians into one group supported the strong cor-respondence between the primary auditory source activity andthe gray matter volume of amHG (r = 0.78, P < 0.0001). The correlation over all groups was much weaker (r = 0.32 instead of 0.87, P < 0.05) when the whole HG was used as the anatomical reference.

The gray matter volume of amHG separated professional musi-cians (range 536–983 mm3) from non-musicians (172–450 mm3),

Table 1. Volumetry of Heschl’s gyrus.

Gray matter volume (mm3) White matter volume (mm3)

Area Side Non- Professional Amateur Non- Professional Amateurmusicians musicians musicians musicians musicians musicians

HG L 2,083 ± 212 2,607 ± 215 2,694 ± 411 929 ± 72 861 ± 128 883 ± 116R 1,868 ± 110 2,814 ± 197** 2,256 ± 256 946 ± 90 1,172 ± 165 1,018 ± 188

(L+R)/2 1,976 ± 153 2,710 ± 201** 2,475 ± 308 937 ± 52 1,018 ± 132 951 ± 144δHG –0.11 ± 0.06 0.08 ± 0.04 –0.18 ± 0.10 0.02 ± 0.09 0.31 ± 0.12* 0.22 ± 0.10

aHG L 925 ± 107 1,513 ± 115** 1,435 ± 181* 434 ± 44 516 ± 56 504 ± 82R 824 ± 57 1,406 ± 142** 1,379 ± 170** 353 ± 46 484 ± 65 453 ± 50

(L+R)/2 873 ± 77 1,461 ± 112*** 1,407 ± 138** 394 ± 39 505 ± 48 502 ± 59δaHG –0.11 ± 0.07 –0.07 ± 0.09 –0.04 ± 0.11 –0.21 ± 0.13 –0.06 ± 0.15 –0.11 ± 0.18

amHG L 328 ± 58 776 ± 68*** 523 ± 57* 176 ± 18 220 ± 27 246 ± 35R 296 ± 31 659 ± 39*** 448 ± 47* 160 ± 22 227 ± 27 208 ± 26

(L+R)/2 311 ± 27 716 ± 39*** 494 ± 46** 172 ± 15 223 ± 18* 227 ± 28δamHG –0.10 ± 0.12 –0.16 ± 0.10 –0.15 ± 0.11 –0.09 ± 0.13 0.03 ± 0.18 –0.16 ± 0.15

*P < 0.05, **P < 0.01, ***P < 0.001 (ANOVA, professionals/amateurs versus non-musicians). Values given as mean ± s.e.m. δ, hemispheric asymmetry (Methods).

a b

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

articles


and amateur musicians showed an intermediate gray matter volume (189–798 mm3). The total volume of HG, including white and gray matter, showed a larger variance and could not separate the groups (non-musicians, 1,955–4,694 mm3; profes-sionals, 2,629–6,297 mm3; amateurs, 2,151–7,603 mm3). Theasymmetry measures (Methods) showed only one significanteffect: the total volume of HG was 14% larger in the right hemi-sphere of professional musicians (right, 3,986 ± 305 mm3; left,3,468 ± 263 mm3; δHG = 0.14 ± 0.04; F1,11 = 11.7, P < 0.01).

Correlation with musical aptitudeFor all three groups, there was a high correlation between theN19m-P30m signal amplitude and musical aptitude as mea-sured by the AMMA tonal test (Fig. 4b). Both the primarysource activity and the tonal score of musical aptitude com-pletely separated the professional musicians from the non-musicians. The amateur musicians showed an intermediaterange of musical aptitude and dipole amplitudes that over-lapped with the two other groups. Within groups, the correla-tion was significant for non-musicians (r = 0.55, P < 0.05), butnot for amateurs (r = 0.19, n.s.) or professionals (r = 0.05, n.s.).In general, professional musicians had high AMMA scores, highsignal amplitudes and large gray matter volumes of amHG.When analysis was restricted to amateur and professional musi-cians, the correlation was significant (r = 0.52, P < 0.01).

Similarly, the gray matter volume of amHG was highly cor-related with musical aptitude (Fig. 4c). Within groups, the cor-relation was significant for non-musicians (r = 0.71, P < 0.001)and amateurs (r = 0.56, P < 0.05), but not for professionals (r = 0.40, n.s.). When all amateurs and professionals were com-bined, however, the correlation was highly significant (r = 0.70,P < 0.0001). This correlation was smaller when considering thegray matter volume of aHG in its full lateral extent (r = 0.44, P< 0.01) and was nonsignificant when the whole gray mattervolume of HG was calculated (r = 0.26, n.s.). No correlationwas found between musical aptitude and white matter volumesof HG.

Under the assumption that anatomical size determined thesignal strength, a partial correlation was calculated to eliminatethe influence of amHG gray matter volume of on the correla-tion between N19m-P30m amplitude and AMMA score. Thispartial correlation was only r = –0.04 (n.s.), indicating thatanatomical size was the key parameter.

Influence of external variablesWe found no influence of the covariates sex, age or head sizeon either the early dipole amplitudes or on the gray matter vol-ume of amHG. To exclude influences of attention during MEGrecording and of the frequency modulation in the stimulus, wecarried out two additional control sessions in a subgroup of 24stochastically selected subjects. While watching a video, sub-jects detected deviant tones of a different frequency (1,100 Hzinstead of the standard 500 Hz) and indicated them by buttonpress in the attention experiment. We found no significant effectof attention on the primary N19m-P30m component. Withinnoise limits, the N19m-P30m signals for the onset of pure sinu-soidal tones agreed with the signals deconvoluted from themodulated tones23,40.

DISCUSSIONHere we found a large difference in the early neurophysiologicalactivity of the auditory cortex in musicians versus non-musicians,using simple tonal stimuli. In addition, we found strong correla-tions of this activity with the gray matter volume of amHG andwith musical aptitude. Using partial correlations, we showed thatthe gray matter volume of amHG was the key parameter influ-encing the early evoked response of the auditory cortex. The larger gray matter volume in professional musicians was mostpronounced for amHG (130% greater than in non-musicians)and dropped to 37% more volume than in non-musicians whenthe whole HG was used for anatomical reference. Together withevidence from previous EEG20 and MEG21–23 studies that local-ized the origin of the primary auditory-evoked N19-P30 sourceactivity to amHG, our findings provide evidence for the aug-mentation of PAC gray matter in musicians.

This functional–anatomical interpretation is consistent withthe microanatomical24–29 finding that amHG comprises mostof the primary granular core field. However, the macroanatom-ically defined amHG is only an approximate measure of thelocation and extent of PAC, because there is considerable indi-vidual variability27–31. Non-primary cortical fields are mostlikely to be found near the lateral and posterior edges ofamHG2,27–29. Thus, the larger volume of gray matter in musi-cians may comprise PAC as well as surrounding belt areas. Thestrong functional–anatomical correspondence at the level ofamHG is probably related to the stimulation with sinusoidaltones. Whereas functional MRI41,42 and PET43 studies have shown

Fig. 4. Correlations between early neurophysiological source activity, amHG gray matter volume and musical aptitude. (a) The N19m-P30mdipole moment was strongly correlated with the mean gray matter volume of amHG. Values were averaged over the right and left hemispheres.(b, c) The tonal raw score of musical aptitude (AMMA test) was highly correlated with both the N19m-P30m dipole moment (b) and the graymatter volume of amHG (c).

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


articles

that simple tonal stimuli primarily activate the auditory coreregions, the fine time resolution of MEG and EEG allowed us todistinguish the initial primary source activity evoked by pure tones,and to compare the data with human intracranial recordings20.

Our anatomical data extend earlier findings of enlarged Hes-chl’s gyri in musicians. Morphometric post-mortem case stud-ies44 have reported an abnormally large anterior ‘Heschlconvolution’ in two eminent musicians. Furthermore, a histo-logical study45 has reported larger volumes of cytoplasm innerve cells in layers 3–6 of Heschl’s first transverse convolutionin two professional musicians as compared to non-musicians.This is in accordance with our finding that thicker amHG graymatter, not white matter, was the main factor contributing tothe morphological volume difference between musicians andnon-musicians.

At the neurophysiological level, we found a striking differ-ence between the early and late evoked responses of the audi-tory cortex. The finding that the late N100m component wasnot enhanced in musicians is in agreement with a previousMEG study that reports an increase of 25% in the amplitude ofN100m with piano tones, but not with sinusoidal tones17. It hasbeen suggested that structural differences in the white matterof the planum temporale of musicians10 (which has been iden-tified as the predominant generating substrate of N100m20)underlies this effect. In combination, these and our findingsindicate that the early and late auditory evoked responses reflectdifferent stages and areas of functional processing in the humanauditory cortex. Thus, the increase in the activity and structureof particular regions of the auditory cortex in musicians seemsto be related to the processing of specific stimulus propertiesand may therefore reflect multiple structural and neurophysio-logical specializations in the auditory cortex of musicians.

The increase in the early dipole source activity was larger inthe right hemisphere of professional musicians, whereas thegray matter volumes of amHG did not show significant hemi-spheric differences. This result should be interpreted with cau-tion because MEG is largely insensitive to radial current flow.Hence, MEG cannot measure the full size of the net dipolesource vector, and the hemispheric differences could be due, inpart, to different effective orientations in the groups of profes-sionals and non-musicians. Despite this uncertainty about thefull dipole moment, the correlation between the primary sourceactivity and the gray matter volume of amHG was highly sig-nificant for all anatomical measures, irrespective of whether theright hemisphere alone, the left hemisphere alone, or bothtogether were considered.

What are the causes for the striking increases in the graymatter volume of the anteromedial portion of HG and the earlyneurophysiological activities of the auditory cortex in musi-cians? No influence of attention was found. The role of musi-cal practice40,49, however, remains unclear. In the group ofmusicians, there was a ceiling effect in the correlation betweenAMMA score and dipole amplitude (Fig. 4b). Therefore, onlythe amateur musicians are an appropriate sample for studyingthe influence of musical practice on the dipole amplitude. Thenumber of amateur musicians in this study was too small, how-ever, to obtain significant results regarding the effects of start-ing age and intensity of musical practice in childhood.

Functional long-term and short-term plasticity might influ-ence the amplitude of the late auditory evoked N100mresponse17,46, as well as the frequency representation and tem-poral information processing in the PAC35,36 of non-humanprimates. On the other hand, developmental stability around

age seven has been demonstrated for the human HG andplanum temporale in morphometric34,47 and myeologenetic48

studies. This maturation age is consistent with that observedfor the psychometric variable of our study, musical aptitude.The level of musical aptitude reached by the age of nine remainsthe same throughout life37. In conclusion, our results indicatethat the morphology and neurophysiology of HG have an essen-tial impact on musical aptitude. The question remains, however, whether early exposure to music49 or a genetic pre-disposition50 leads to the functional and anatomical differencesbetween musicians and non-musicians.

METHODSSubjects. Thirty-seven right-handed adults with normal hearing weredivided into three groups: 12 non-musicians (age range 26–43 years; 6 men, 6 women), 12 professional musicians (age 29–55; 6 men, 6 women) and 13 amateur musicians (age 24–62; 7 men, 6 women). Asthere is no standard definition of ‘musician’8–19, we classified the threesubject groups as follows: non-musicians had never played an instru-ment beyond standard school education, professional musicians hadundergone a professional music education ending with a diploma andwere actively performing at the time of examination, and amateur musi-cians had received special instruction in one or more musical instru-ments. This classification was validated by the advanced measure of musicaudiation (AMMA) tonal test37 (see below).

A stochastically chosen subgroup of 24 subjects, divided into 8 non-musicians (age 26–43, 4 men, 4 women), 7 professional musicians (age29–51, 3 men, 4 women) and 9 amateur musicians (age 24–62, 5 men, 4 women) participated in additional experiments. Averaged over thegroups, there were no significant differences in age, sex or head size. Allparticipants gave their informed consent to the study. Experimental pro-cedures were approved by the local ethics committee.

AMMA test. This test, which has been standardized with more than 5,000students, presents 30 pairs of short melodies. The repeated melody hasa small change in pitch (10 pairs) or rhythm (10 pairs) or is unchanged(10 pairs). Subjects detected the modification in a three-way forced choicetask. As our study used tonal stimuli, only tonal raw test scores were cal-culated. The tonal raw score was calculated as 20 plus the number of cor-rect responses, minus the number of false alarms. Non-musicians scoredless than 25 on a scale of 0–40 (range 17–24), professional musicians hada raw score of at least 26 (range 26–39) and amateur musicians scoredin an intermediate range (18–33).

Stimuli. 100% amplitude-modulated pure tones with a duration of 1 sincluding approximately 30 modulation cycles (Fig. 1a) were presentedbinaurally at a level of 50 dB SL (sensation level). The tones were deliveredin blocks with one fixed carrier frequency through shielded transducersconnected to the subject through 90 cm plastic tubes and foam earpieces.In one session, six blocks with carrier frequencies of 100, 220, 500, 1,100,2,500 and 5,600 Hz were presented. These carrier frequencies were cho-sen to be equidistant on a logarithmic frequency scale corresponding tothe musical interval of a major ninth. Within each block, 430 tones werepresented at seven different modulation frequencies in the range of 26–37 Hz in pseudorandom order, with an interstimulus interval rangingfrom 1.0 to 1.2 s. To minimize stimulus artifacts, tone polarity wasreversed from one tone to the next. The modulated tones sounded sim-ilar to pure tones with an additional small roughness.

In the additional tonal experiment, unmodulated tones were present-ed with the same frequencies ranging from 100 to 5,600 Hz. To obtainabout 2,000 averages per frequency, the duration of the tones was reducedto 150 ms including a 20 ms rise and fall time. Interstimulus interval wasreduced to the pseudo-randomized 400–600 ms range.

In the main and additional tonal experiments, subjects listened pas-sively to the sounds while watching a silent video of their own choice. Inthe additional attention experiment, modulated tones of 500 Hz (stan-dards) and 1,100 Hz (deviants) were presented in an oddball task andsubjects were asked to indicate the deviants by button press.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

Procedures. MEG was recorded continuously over both hemisphereswith a whole-head Neuromag-122 system (Neuromag, Helsinki, Finland)in a magnetically shielded room using a band-pass filter of 0.01–250 Hzand a sampling rate of 1,000 Hz. For coregistration with MRI, the loca-tions of four indicator coils and 35 surface points, including the nasionand two pre-auricular points, were digitized. The head position in theDewar was determined at the beginning of each recording block. Foreach carrier frequency, about 400 artifact-free responses were averagedoff-line by excluding trials with eye-blinks or MEG gradients greater than600 fT/cm. One of the resulting 122 averaged MEG waveforms over theright auditory cortex is depicted in Fig. 1b. About 2–5 noisy channelswere excluded from further analysis. The amplitude of the N100m com-ponent was measured against the prestimulus baseline by subtracting theaverage signal over a 100 ms interval before tone onset. The fast prima-ry N19m-P30m component was analyzed after applying a zero phaseshift band-pass filter from 20 Hz (12 dB/octave) to 120 Hz (24 dB/octave).

Deconvolution technique. To obtain the primary N19m-P30m compo-nents from the steady-state responses, we applied a linear deconvolutiontechnique23 to the period modulation cycles. Responses were averagedover all cycles at each modulation frequency. The averaged cyclic respons-es (total ∼ 12,000) were then combined to deconvolute the time courseof the transient N19m-P30m response elicited by each modulation cycle(Fig. 1c). Intrinsically, this deconvolution technique23 uses the varyingphase lag and overlap of the underlying components at each modulationfrequency for the reconstruction. The deconvoluted responses had a verylarge signal-to-noise ratio and were highly similar to the N19m-P30mresponse recorded with short transient stimuli. The advantage of thedeconvolution technique was that it separated the primary N19m-P30mresponse (Fig. 1c) from later overlapping responses such as P50m-N100m(Fig. 1b), which originate outside the PAC. Decomposed and transientN19m-P30m responses had similar source localizations in the medialportion of the anterior HG23.

Source analysis. The BESA software (MEGIS Software GmbH, Graefelf-ing, Germany) was used to model the source activity of the auditory cor-tex with one equivalent dipole in each hemisphere38,39. When the dipoleswere fitted to the measured magnetic evoked fields in the 19–30 ms range,locations near the anteromedial portion of the HG were found with amean deviation of 3 mm in the posterior direction and 5 mm in the supe-rior direction. As dipole depth is the weakest parameter in MEG dipolefitting and has a strong inverse correlation with dipole amplitude, a fixeddepth value was defined using the mean dipole depth over all frequen-cies and subjects40. This depth point was 10 mm from the medial bound-ary of the anterior HG for N19m-P30m, and 22 mm from the boundaryfor N100m. Thus, the dipole sources were seeded systematically40 alongthe individual anterior HG at these distances, laterally from its medialboundary. Location parameters in the anterior–posterior and inferior–superior directions were retained from the fit procedure, and orienta-tions were fitted individually to N19m-P30m and the peak of N100m.Using this two-dipole model, robust source waveforms were calculatedfor both components. These provided an image of brain function in termsof the magnitude and timing of the compound source currents in theregion of HG.

Morphometry. To obtain morphological measures of the right and leftHG, the surfaces of the gray and gray–white matter boundaries were ren-dered from the individual T1-weighted 3D-MRI images (Philips EdgeSystem, Eindhoven, Netherlands; 1.5 T, 1 mm slices, Fig. 1d–f) using thesegmentation tools of the BrainVoyager program (Brain Innovation B.V.,Maastricht, Netherlands). The inclusion range of image gray values wascalculated from the individual intensity histograms by identifying thepeaks corresponding to gray and white matter and their half-amplitudeside lobes. For gray matter surface segmentation, both the white and graymatter peaks were included with their side-lobes. For white matter seg-mentation, only the white matter peak was included up to the midpointbetween the peaks of white and gray matter. These rendered surfaces werethen used together with macroanatomically defined anterior, posterior,medial and inferior boundaries to calculate the white and gray mattervolumes of HG, aHG and amHG. FTS defined the anterior boundary in

all cases (Fig. 1f). SI defined the posterior boundary of aHG in 35 of 74hemispheres. The first, most anterior HS was used in 39 hemispheres.The medial boundary was drawn from the medial end of FTS to themedial end of the most posterior HS. The inferior boundary was derivedby an intersecting surface running from the depth of FTS to the depthof HS. Hemispheric asymmetry was determined by the coefficient δ =(VR – VL)/(0.5 × (VR + VL)).

AcknowledgmentsWe thank P. Berg for helpful comments, K. Sartor and C. Stippich for providing

the 3D MRI scans and R. Goebel for his support with the BrainVoyager program.



1. Zatorre, R. J. & Peretz, I. The biological foundations of music. Ann. NY Acad.Sci. 930, (2001).

2. Zatorre, R. J. & Binder, J. R. in Brain Mapping the Systems (eds. Toga, A. W. &Maziotta, J. G.) 365–402 (Academic, San Diego, California, 2000).

3. Liégeois-Chauvel, C., Peretz, I., Bahaï, M., Laguitton, V. & Chauvel, P.Contribution of different cortical areas in the temporal lobes to musicprocessing. Brain 121, 1853–1867 (1998).

4. Maess, B., Koelsch, S., Gunter, T. C. & Friederici, A. D. Musical syntax isprocessed in Broca’s area: an MEG study. Nat. Neurosci. 4, 540–545 (2001).

5. Blood, A. J., Zatorre, R. J., Bermudez, P. & Evans, A. C. Emotional responsesto pleasant and unpleasant music correlate with activity in paralimbic brainregions. Nat. Neurosci. 2, 382–387 (1999).

6. Schiavetto, A., Cortese, F. & Alain, C. Global and local processing of musicalsequences: an event-related brain potential study. Neuroreport 10, 2467–2472(1999).

7. Patel, A. D. & Balaban, E. Human pitch perception is reflected in the timing ofstimulus-related cortical activity. Nat. Neurosci. 4, 839–844 (2001).

8. Zatorre, R. J., Evans, A. C. & Meyer, E. Neural mechanisms underlyingmelodic perception and memory for pitch. J. Neurosci. 14, 1908–1919 (1994).

9. Zatorre, R. J., Perry, D. W., Beckett, C. A., Westbury, C. F. & Evans, A. C.Functional anatomy of musical processing in listeners with absolute pitchand relative pitch. Proc. Natl. Acad. Sci. USA 95, 3172–3177 (1998).

10. Schlaug, G., Jäncke, L., Huang, Y. & Steinmetz, H. In vivo evidence ofstructural brain asymmetry in musicians. Science 267, 699–701 (1995).

11. Schlaug, G., Jäncke, L., Huang, Y., Staiger, J. F. & Steinmetz, H. Increasedcorpus callosum size in musicians. Neuropsychologia 33, 1047–1055 (1995).

12. Münte, T., Kohlmetz, C. & Altenmüller, E. Superior auditory spatial tuning inconductors. Nature 409, 580 (2001).

13. Altenmüller, E. Electrophysiological correlates of music perception in thehuman brain. Eur. Arch. Psychiatry Neurol. Sci. 235, 342–354 (1986).

14. Wayman, J. W., Frisina, R. D. & Walton, J. P. Effects of musical training andabsolute pitch ability on event-related activity in response to sine tones. J. Acoust. Soc. Am. 91, 3527–3531 (1992).

15. Crummer, G. C., Walton, J. P., Wayman, J. W., Hantz, E. C. & Frisina, R. D.Neural processing of musical timbre by musicians, nonmusicians, andmusicians possessing absolute pitch. J. Acoust. Soc. Am. 95, 2720–2727(1994).

16. Koelsch, S., Schröger, E. & Tervaniemi, M. Superior pre-attentive auditoryprocessing in musicians. Neuroreport 10, 1309–1313 (1999).

17. Pantev, C. et al. Increased auditory cortical representation in musicians.Nature 392, 811–813 (1998).

18. Elbert, T., Pantev, C., Wienbruch, C., Rockstroh, B. & Taub, E. Increasedcortical representation of the left hand in string players. Science 270, 305–307(1995).

19. Hirata, Y., Kuriki, S. & Pantev, C. Musicians with absolute pitch show distinctneural activities in the auditory cortex. Neuroreport 10, 999–1002 (1999).

20. Liégeois-Chauvel, C., Musolino, A., Badier, J. M., Marquis, P. & Chauvel, P.Evoked potentials recorded from the auditory cortex in man: evaluation andtopography of the middle latency components. Electroencephalogr. Clin.Neurophysiol. 92, 204–214 (1994).

21. Scherg, M. & von Cramon, D. Evoked dipole source potentials of the humanauditory cortex. Electroencephalogr. Clin. Neurophysiol. 65, 344–360 (1986).

22. Rupp, A. et al. Fast temporal interactions in human auditory cortex.Neuroreport 11, 3731–3736 (2000).

23. Gutschalk, A. et al. Deconvolution of 40 Hz steady-state fields reveals twooverlapping source activities of the human auditory cortex. Clin.Neurophysiol. 110, 856–868 (1999).

24. Braak, H. The pigment architecture of the human temporal lobe. Anat.Embryol. 154, 214–240 (1978).

25. Galaburda, A. & Sanides, F. Cytoarchitectonic organization of the humanauditory cortex. J. Comp. Neurol. 190, 597–610 (1980).

26. Rademacher, J., Caviness, V. S., Steinmetz, H. & Galaburda, A. M.

articles


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


articles

Topographical variation of the human primary cortices: implications forneuroimaging, brain mapping, and neurobiology. Cereb. Cortex 3, 313–329(1993).

27. Hackett, T. A., Preuss, T. M. & Kaas, J. H. Architectonic identification of thecore region in auditory cortex of macaques, chimpanzees and humans. J. Comp. Neurol. 441, 197–222 (2001).

28. Rivier, F. & Clarke, S. Cytochrome oxidase, acetylcholinesterase, andNADPH-diaphorase staining in human supratemporal and insular cortex:evidence for multiple auditory areas. Neuroimage 6, 288–304 (1997).

29. Wallace, M. N., Johnston, P. W. & Palmer, A. R. Histochemical identificationof cortical areas in the auditory region of the human brain. Exp. Brain Res.143, 499–508 (2002).

30. Rademacher, J. et al. Probabilistic mapping and volume measurement ofhuman primary auditory cortex. Neuroimage 13, 669–683 (2001).

31. Morosan, P. et al. Human primary auditory cortex: cytoarchitectonicsubdivisions and mapping into a spatial reference system. Neuroimage 13,684–701 (2001).

32. Steinmetz, H. et al. Cerebral asymmetry: MR planimetry of the humanplanum temporale. J. Comput. Assist. Tomogr. 13, 996–1005 (1989).

33. Penhune, V. B., Zatorre, R. J., MacDonald, J. D. & Evans, A. C.Interhemispheric anatomical differences in human primary auditory cortex:probabilistic mapping and volume measurement from magnetic resonancescans. Cereb. Cortex 6, 661–672 (1996).

34. Leonard, C. M., Puranik, C., Kuldau, J. M. & Lombardino, L. J. Normalvariation in the frequency and location of human auditory cortex. Heschl’sgyrus: where is it? Cereb. Cortex 8, 397–406 (1998).

35. Recanzone, G. H., Schreiner, C. E. & Merzenich, M. M. Plasticity in thefrequency representation of primary auditory cortex followingdiscrimination training in adult owl monkeys. J. Neurosci. 13, 87–103 (1993).

36. Kilgard, P. M. & Merzenich, M. M. Plasticity of temporal informationprocessing in the primary auditory cortex. Nat. Neurosci. 1, 727–731 (1998).

37. Gordon, E. E. Learning Sequences in Music (GIA, Chicago, Illinois, 1997).38. Scherg, M. in Auditory Evoked Magnetic Fields and Electric Potentials Vol. 6

(eds. Grandori, F., Hoke, M. & Romani, G. L.) 165–193 (Karger, Basel,Switzerland, 1990).

39. Scherg, M. & von Cramon, D. Two bilateral sources of the late AEP asidentified by a spatio-temporal dipole model. Electroencephalogr. Clin.Neurophysiol. 62, 32–44 (1985).

40. Schneider, P. Source Activity and Tonotopic Organization of the AuditoryCortex in Musicians and Non-musicians. Thesis, Univ. Heidelberg (2000).

41. Talavage, T. M. et al. Frequency-dependent responses exhibited by multipleregions in human auditory cortex. Hear. Res. 150, 225–244 (2000).

42. Wessinger, C. M. et al. Hierarchical organization of the human auditorycortex revealed by functional magnetic resonance imaging. J. Cogn. Neurosci.13, 1–7 (2001).

43. Zatorre, R. & Belin, P. Spectral and temporal processing in human auditorycortex. Cereb. Cortex 11, 946–953 (2001).

44. Meyer, A. in Music and the Brain (eds. Critchley, M. & Henson, R. A.)255–281 (Heinemann, London, 1977).

45. Somogyi, J. Über das morphologische Korrolat der musikalischenFähigkeiten. Mschr. Psychat. Neurol. 75, 113–169 (1930).

46. Menning, H., Roberts, L. E. & Pantev, C. Plastic changes in the auditorycortex induced by intensive frequency discrimination training. Neuroreport11, 817–822 (2000).

47. Preis, S., Jäncke, L., Schmitz-Hillebrecht, J. & Steinmetz, H. Child age andplanum temporale asymmetry. Brain Cogn. 40, 441–452 (1999).

48. Yakovlev, P. I. & Lecours, A. R. in Regional Development of the Brain in EarlyLife (ed. Minkowski, A.) 3–70 (Blackwell, Oxford, 1967).

49. Monaghan, P., Metcalfe, N. B. & Ruxton, G. D. Does practice shape the brain?Nature 394, 434 (1998).

50. Thompson, P. M. et al. Genetic influences on brain structure. Nat. Neurosci.4, 1253–1258 (2001).

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

Language impairments are most frequently seen after lesions tothe left side of the brain, and occasionally after lesions to the rightside. The mechanisms involved in the recovery of language aftera one-sided brain lesion are uncertain. The undamaged hemi-sphere or undamaged parts of the injured hemisphere may takeover language representation. If so, it is unclear why certainpatients overcome initial deficits, whereas many others remainimpaired. One possibility is that individual differences in theorganization of language render some people more able to recov-er function after unilateral brain damage.

Functional imaging has provided a wealth of information onbrain activity during language processing. Studies with large num-bers of subjects often show individuals with weak lateralization, orbihemispheric activation, during language tasks1–3. Bihemispher-ic increases in blood flow can be artifacts of methodology in somecases, rather than reflections of a truly bilateral language organiza-tion4,5. Data from epilepsy patients who underwent temporary inac-tivation of each hemisphere by injection of anesthetics indicate thatboth hemispheres can be involved in the processing of language6.However, such patients have abnormal brains because of their med-ication-resistant epilepsy, and their bilaterality of language repre-sentation may reflect disorganization resulting from longstandinglesions. It remains unclear whether true bihemispheric representa-tion of language occurs in healthy subjects, accounts for bihemi-spheric activation in functional imaging data and offers resistanceto language deficits after unilateral brain lesions.

To assess the functional relevance of lateralization of lan-guage-related brain activation in healthy subjects, we used tran-sient focal virtual lesions7 induced by transcranial magneticstimulation (TMS).

Degree of language lateralizationdetermines susceptibility tounilateral brain lesions

S. Knecht1, A. Flöel1, B. Dräger1, C. Breitenstein1, J. Sommer1, H. Henningsen1, E. B. Ringelstein1

and A. Pascual-Leone2

1 Department of Neurology, University of Münster, Albert-Schweitzer-Straβe 33, D-48129 Münster, Germany2 Laboratory for Magnetic Brain Stimulation, Harvard Medical School, 330 Brookline Avenue, Kirstein Building KS 454, Boston, Massachusetts 02215, USA

Correspondence should be addressed to S.K. ([email protected])


Language is considered a function of either the left or, in exceptional cases, the right side of thebrain. Functional imaging studies show, however, that in the general population a gradedcontinuum from left hemispheric to right hemispheric language lateralization exists. To determinethe functional relevance of lateralization differences, we suppressed language regions using trans-cranial magnetic stimulation (TMS) in healthy human subjects who differed in lateralization oflanguage-related brain activation. Language disruption correlated with both the degree and side oflateralization. Subjects with weak lateralization (more bilaterality) were less affected by either left-or right-side TMS than were subjects with strong lateralization to one hemisphere. Thus in somepeople, language processing seems to be distributed evenly between the hemispheres, allowing forready compensation after a unilateral lesion.

Subjects were selected from a cohort of 324 to representthe continuum of language lateralization from left- to right-hemisphere dominance3 (Table 1). We assessed language lat-eralization by measuring hemispheric perfusion differencesduring a word generation task using functional transcranialDoppler sonography (fTCD) and functional magnetic reso-nance imaging (fMRI)8. Then, using TMS, we suppressedbrain regions9 to test their causal relation to performance on apicture–word verification task (Methods). To test for nonspe-cific effects, TMS was also applied over a midline occipitalcontrol site (Oz) and during a control task involving matchingof geometric objects.

We found that both the side and the degree of lateralizationof language-related brain activation correlated with a person’ssusceptibility to language disruption by neural deactivation inone hemisphere. In addition, greater bilaterality of the languagesystem, as indicated by low lateralization, allowed for autonomousverbal processing within each hemisphere.

RESULTSWe analyzed our data by dividing subjects into two groups:right language dominant (laterality index (LI) ≤ 0, Methods)and left language dominant (LI > 0). Each subject was stimu-lated with TMS, in different trials, in the left hemispheric Wer-nicke’s language region (CP5) and its right hemispherichomologue (CP6). A repeated-measures analysis of variance(ANOVA) showed a significant interaction for stimulation side× group (F1,18 = 7.74, P < 0.01). Simple main effects of group,analyzed for each stimulation side separately, showed that sub-jects with left language dominance were significantly slower

articles


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


than were subjects with right language dominance after TMSover CP5; the reverse pattern (right dominant > left dominant)was seen for stimulation over CP6 (unpaired t-tests, both t(18)> |2.19|, P < 0.05). Furthermore, simple main effects of stim-ulation side, analyzed for each group separately, showed thatresponse times (RTs) after CP6 stimulation were slower thanthey were after CP5 stimulation for subjects with right lan-guage dominance (paired t-test, t(8) = –2.16, P = 0.06). A ten-dency for the reverse pattern (TMS over CP5 > TMS over CP6)was seen in subjects with left language dominance (paired t-test, P = 0.10). Thus, we found a double dissociation betweenthe side of language lateralization and the side of TMS-induceddisruption in verbal processing (Fig. 1).

Changes in verbal processing speed after TMS of CP5 andCP6 correlated with the degree of language lateralization asassessed by fTCD (Fig. 2). Subjects with strong lateralizationof language-related brain activation showed strong effects frominterference by left or right hemispheric TMS; those with lowlateralization showed only minor effects. The graded effect ofTMS was similar when only disruption of the dominant (leftor right) hemisphere was considered (Fig. 3).

Pooled data from all subjects examined with TMS of thecontrol site Oz (n = 12, Methods) showed no effect of this con-trol TMS on mean RT (628 ms before TMS, 629 ms after TMS;

P = 0.9, paired t-test, Fig. 1), confirming the topographicspecificity of our results. The accuracy of responses dur-ing the picture–word task was not significantly affectedby TMS at either CP5 or CP6 (paired t-tests), and thebefore–after change scores did not correlate with lan-guage lateralization for any of the stimulation sites. Dif-ferences in RT between hands were small (meandifference, 1.5 ms). TMS had no significant effect on thisparameter (3 ms after inhibitory TMS at CP5; 1 ms afterinhibitory TMS at CP6).

There was no statistically significant interactionbetween side of language dominance (determined byfTCD) and effect of TMS on processing speed duringthe non-linguistic control task. There was a nonsignifi-cant correlation between results on the object match-ing task and language lateralization, which showed theinverse pattern to that of the linguistic task (Fig. 2).After right-hemispheric TMS, there was a tendency forslowed geometric object matching with increasing left-side language lateralization (r = 0.24, P = 0.3), where-as the inverse pattern was seen after TMS of the lefthemisphere (r = –0.38, P = 0.1).

DISCUSSIONIn subjects with left-side language dominance, verbal pro-cessing was slowed during transient disruption of the leftbut not the right hemisphere. The opposite pattern wasseen in subjects with right-sided language dominance. Thisdouble dissociation attests to the cross-method validity ofperfusion-sensitive functional imaging and TMS, althoughdifferent language tasks had to be used (word generation

for fTCD and picture–word verification for TMS). The word gen-eration task used for the initial recruitment of subjects by fTCDis predominantly expressive in nature, whereas the picture–wordverification task used for TMS is mostly receptive (Methods).Interhemispheric dissociations of expressive and receptive lan-guage functions have so far been shown only in patients with pre-existing brain lesions10,11. Activation studies usually show aconcordant lateralization of brain regions involved in expressiveand receptive language12,13.

Inhibition of verbal processing after TMS over the left hemi-sphere has been shown in several studies14–18. So far, no datahave been reported on TMS in healthy subjects with right-hemispheric language representation. Verbal slowing afterright-hemisphere TMS in subjects with right-hemisphere lan-guage dominance, as shown here, accords with clinical obser-vations of ‘crossed aphasias’, that is, aphasias after righthemisphere lesions19.

After disruption of the non-dominant hemisphere in sub-jects with either left or right hemisphere language dominance,

articles

Table 1. Subject characteristics.

Subject Gender Age Handedness* LI based on fTCD**1 m 25 – 40 5.442 f 26 90 5.073 f 42 – 60 5.034 m 35 – 10 4.675 f (f) 25 (22) –13 (60) 3.68 (4.20)6 m 32 – 60 2.337 f 25 – 100 1.198 m 23 –60 1.009 f (f) 30 (33) 100 (90) 0.67 (1.25)10 m (m) 27 (26) –20 (100) 0.66 (1.23)11 f 29 100 0.6212 f (f) 22 (23) 73 (100) –0.50 (–0.61)13 f (f) 22 (26) – 80 (–70 –1.99 (–5.01)14 f (f) 22 (24) – 60 (–90) –2.30 (–2.03)15 f (f) 24 (28) – 90 (–100) –2.40 (–2.20)16 m 33 – 100 –2.8917 f (f) 27 (22) 90 (80) –3.58 (–3.52)18 f 29 – 23 –3.6619 f 32 – 88 –3.9220 m 35 70 –4.82

f, female; m, male (values for exchange subjects who took part in the control task).* Handedness was assessed by the Edinburgh Handedness Inventory31 spanning arange from –100 for strong left-handedness to +100 for strong right-handedness.** Positive values indicate lateralization to the left; negative values indicatelateralization to the right hemisphere.

–30

–20

–10

0

10

20

30

CP5 (left) Oz CP6 (right)

Cha

nge

in R

T p

ost

TM

S (

ms)

acce

lera

tion

← →

slo

win

g

Left dominant

TMS site

Right dominant

Fig. 1. Mean changes (± s.e.m.) in verbal processing speed as assessedby reaction time. Interaction between left (CP5) and right (CP6) hemi-spheric stimulation and left dominant (n = 11) and right dominant (n =9) hemispheric language dominance. Also shown are changes in reactiontime (RT) after TMS pulses to the midline occipital control site (Oz).Note that only 6 left and 6 right hemisphere dominant subjects werestimulated at the control site.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

we found an acceleration of verbal processing—anoth-er example of paradoxical functional improvement bya focal brain disruption7,20. At first glance, it could betaken to reflect interhemispheric disinhibition, withthe non-dominant hemisphere inhibiting the domi-nant one during normal, pre-TMS verbal processing.Faster verbal processing could also result from a nar-rowing of the lexical search. The picture–word verifi-cation task in this study consisted of accessing the literalmeaning of concrete nouns. It did not require morecomplex semantic analyses, a function attributed to thenon-dominant hemisphere5,21,22. Inhibition of the non-dominant hemisphere by TMS may have preventedsuch a complex and time-consuming contextual analy-sis. Four subjects (two with left and two with righthemisphere dominance) who became faster after TMSof the non-dominant hemisphere felt that their verbalperformance had markedly deteriorated. These sub-jects may have noticed difficulties in accessing associa-tive semantic information after disruption of theirnon-dominant hemisphere.

Matching of geometric objects during the controltask was affected differently by TMS than was picture–word matching. There was a nonsignificant ten-dency for slowed processing after TMS of the nondominanthemisphere. The discrepancy in results for the linguistic andnon-linguistic tasks implies that modulation of language pro-cessing by TMS was language-specific and not related to a mem-ory component common to both tasks. A slowing of objectmatching after suppression of the rostral part of the superiortemporal cortex in the nondominant hemisphere would also becompatible with the prominent role of this region in extrapo-lation of object-related and space-related information23.

Fig. 2. Correlations between the degree of language lateraliza-tion and degree of language disruption by TMS. Top, threeexample fMRI images from subjects with left, bilateral and rightlanguage dominance (red, maximal activation). Middle and bot-tom, lateralization was estimated by functional transcranialDoppler sonography (fTCD, x-axis), and disruption of languageprocessing is shown on the y-axis (black circles). TMS wasdone over the left hemisphere (middle) and over the righthemisphere (bottom). Orange triangles represent the effect ofTMS on the non-linguistic control task involving matching ofgeometric objects. Positive values on the y-axis indicate anincrease, and negative values a decrease, in response time (RT).

The focus of our study was on language performance in indi-viduals lacking marked hemispheric lateralization of languageprocessing. After TMS of either hemisphere, such subjectsshowed almost no slowing of verbal processing. To the extentthat repetitive TMS is a valid model of focal lesions and henceof cerebrovascular stroke, this finding suggests that individualswith a more bilateral language representation will remain rela-tively unaffected in verbal functioning after stroke to either theright or left hemisphere. We propose that such hemisphericautonomy does not result from simple duplication of the brain’slanguage processing hardware. Rather, in humans with weak lan-guage lateralization, the partially redundant neural network sup-porting language may be more evenly distributed between bothhemispheres than it is in subjects with strong lateralization.

METHODSThe study was approved by the Ethics Committee of the Medical Facul-ty of the University of Münster, Germany. Informed consent was obtainedfrom all subjects.

Functional transcranial Doppler sonography (fTCD). Twenty subjects(Table 1) were selected from a cohort of 324 healthy volunteers previ-

articles


60

30

–30

0

2 4 60

Pearson r = 0.51P = 0.02

Absolute degree of language lateralization (%)

Cha

nge

in R

T (

ms)

pos

t-T

MS

of

dom

inan

t he

mis

pher

eac

cele

ratio

n ←

→ s

low

ing

Fig. 3. Effect of functional lesion by TMS of only the dominant (left orright) hemisphere relative to the absolute degree of language lateraliza-tion by fTCD.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

ously assessed for language dominance by fTCD (Münster FunctionalImaging Study on the Variability of Hemispheric Specialization in Healthand Disease) (Fig. 4)3,8,24,25. Language lateralization had been determinedwith fTCD by measuring relative hemispheric perfusion increases dur-ing word generation3,8,24,25. Subjects were presented a letter on a com-puter screen 5 s after a cueing tone. They were asked to silently find asmany words as possible starting with the displayed letter. Task compli-ance was controlled by instructing subjects to report the words after asecond auditory signal that was delivered 15 s after presentation of theletter. All words had to be reported in a 5-s time window. The next letterwas presented in the same way after a rest period of 60 s. Letters werepresented in random order. Q, X and Y were excluded because very fewwords begin with these letters. No letter was displayed more than once.

Changes in the cerebral blood flow velocity (CBFV) in the basal arteries,an indicator of the downstream increase of the regional metabolic activityduring the language task, were measured by dual transcranial Doppler ul-trasonography of the middle cerebral arteries (MCAs). Ultrasonographywas done with two 2-MHz transducer probes attached to a headband andplaced at the temporal skull windows bilaterally. After automated artifactrejection, data were integrated over the corresponding cardiac cycles, seg-mented into epochs that related to a cueing tone before the language taskand were then averaged. The epochs were set to begin 15 s before, and toend 35 s after, the cueing tone. The mean velocity in the 15-s interval beforecueing (Vpre.mean) was taken as the baseline value. The relative CBFVchanges (dV) during cerebral activation were calculated using the formula:dV = (V(t) – Vpre.mean) * 100/Vpre.mean where V(t) is the CBFV over time.Relative CBFV changes from repeated presentations of letters (on the aver-age 20 runs) were averaged time-locked to the cueing tone.

The functional TCD laterality index (LI) was calculated using theformula:

where ∆V(t) = dV(t)left – dV(t)right is the differencebetween the relative velocity changes of the left andright MCAs. tmax represents the latency of theabsolute maximum of ∆V(t) during an interval of10–18 s after cueing (during verbal processing). Forintegration, a time period of tint = 2 s was chosen.The test-to-retest reproducibility of this procedurebased on the Pearson product moment correlation

LI = ∆V (t) dt1

tint ∫tmax +0.5tint

tmax –0.5tint

Fig. 5. Experimental design. A set of 60 pictures (30different drawings with correct subtitles and 30 iden-tical drawings with incorrect subtitles) was pre-sented in randomized sequence pre- and post-TMS.Response times and accuracy of response were com-pared pre- and post-TMS (paired t-test). In a controltask (lower left), subjects matched geometric objectsin an upper row with those in a lower row (left, cor-rect pairs; right, incorrect pairs).


coefficient was r = 0.95, P < 0.0001 (ref. 25). fTCD has been validatedby direct comparison with the intracarotid amobarbital injection pro-tocol and functional magnetic resonance imaging8,24. Left-hemisphericlanguage dominance was assumed in all cases with a positive lateral-ization index. Right-hemispheric language dominance was definedaccordingly.

Subjects were divided into five categories of lateralization. Within eachcategory, those subjects who had been examined last were contacted andinvited to participate in the present study. Thus, subjects with bilateraland right-hemispheric language representation were overrepresented rel-ative to the general population (Fig. 4). Although moderate to strongleft-hemispheric language lateralization is the epidemiological rule, thefocus here was on the continuum of lateralization and susceptibility thatcould explain why patients with similar brain lesions can differ marked-ly with respect to language impairments26.

Behavioral tasks. Different language tasks had to be used for fTCD andTMS. To produce consistent regional blood flow increases amenable todetection by fMRI or fTCD, neural activations of several seconds in dura-tion are required. To quantify effects of TMS on neural processing speed,linguistic tasks with homogenous material and of short duration arerequired. Verbal processing speed was assessed by measuring RT in apicture–word verification task (Fig. 5). Thirty black-and-white drawingsof concrete objects, selected for name agreement and complexity, werepresented in a randomized order on a computer screen with correct orincorrect subtitles. Incorrect subtitles were randomly chosen and werenot matched with the correct subtitles for semantic or phonematic fea-tures or word length. Pictures and words in this no-match condition weretaken from the same pool as the matching picture–word pairs. Incorrectpairs were correctly rejected in over 95% of trials. Subjects indicated rightor wrong pairings by bimanual key presses. Each subject received eightpractice blocks to overtrain performance and achieve a plateau of RTs.For training, the same material was used as during the TMS interven-tion. Verbal RTs before TMS were subtracted from those after TMS.

A post-hoc control experiment was done to examine whether effectswere linguistic. Twelve of the original subjects and eight new subjectsmatched for language lateralization (Table 1, parentheses) participated.The task involved matching two geometric objects with two identical ordifferent objects of the same size in a lower row (Fig. 5, lower left). A setof 30 correct and 30 incorrect pairings were used for training and test-

articles

Fig. 4. Distribution of lateralization indices (LIs). LIs were based onfTCD for all 324 healthy subjects, 111 subjects with negative (or zero)and 213 subjects with positive handedness values in the Edinburghinventory (based on previously published data3). The percentage perfu-sion difference (x-axis) between the left and the right hemisphere, or,more specifically, the territory of the middle cerebral artery, was mea-sured during word generation.

Training ofpicture–wordverification Mouse Reaction time and accuracy

pre-TMS

Reaction time and accuracypost-TMS

Random order of pictures

Random order of pictures

Lion

Banana

Time

TimeMouse

TMS coilover CP5

CP6

Samples from objectmatching control task:

Midline occipitalcontrol site (Oz)

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

ing in a manner identical to the linguistic task. TMS protocol and RTassessment were as in the original task.

Transcranial magnetic stimulation (TMS). TMS was applied using aMagstim Rapid Stimulator (Magstim, Whitland, UK) with a focal figure-of-eight coil positioned over CP5 or CP6 according to the international10–20 electrode system. These sites are considered to reflect the approx-imate locations of Wernicke’s area and its contralateral homologue,respectively15. They were targeted because of their established role inreceptive language tasks and thus their likely involvement in picture–wordverification27. A midline occipital stimulation site (Oz) served as the con-trol site. TMS was applied at 1 Hz for 600 s at 110% intensity of the motorthreshold28. This TMS protocol has been shown to cause a disruption ofthe function of the targeted brain region and behavioral effects that lastfor several minutes after the stimulation20,29. Each TMS administrationwas followed by a 30-min rest before the next administration to avoidcarryover effects28,30. For each picture, RTs and accuracy of responseswere assessed (Fig. 5).

The relationship between the extent of RT changes after TMS (overCP5 and CP6) and the extent of language lateralization as determinedby fTCD was assessed by Pearson product correlation. ANOVA andpost-hoc analysis were used to test for a double dissociation betweenthe side of TMS and the side of language lateralization as measured byfTCD. For a subset of subjects (2, 3, 5, 6, 7, 9, 12, 13, 14, 15, 17, 18),RTs before and after TMS at Oz were also compared using paired t-tests. To assess possible interference of TMS with motor performance,we measured RTs for each hand separately. We analyzed with pairedt-tests whether differences in RT between hands changed from beforeto after TMS.

AcknowledgmentsThis work was supported by Nachwuchsgruppen-Förderung, Innovative

Medizinische Forschung, and Deutsche Forschungsgemeinschaft.



1. Binder, J. R. et al. Determination of language dominance using functionalMRI: a comparison with the Wada test. Neurology 46, 978–984 (1996).

2. Pujol, J., Deus, J., Losilla, J. M. & Capdevila, A. Cerebral lateralization oflanguage in normal left-handed people studied by functional MRI. Neurology52, 1038–1043 (1999).

3. Knecht, S. et al. Handedness and hemispheric language dominance in healthyhumans. Brain 123, 2512–2518 (2000).

4. Just, M. A., Carpenter, P. A., Keller, T. A., Eddy, W. F. & Thulborn, K. R. Brainactivation modulated by sentence comprehension. Science 274, 114–116(1996).

5. George, M. S. et al. Understanding emotional prosody activates righthemisphere regions. Arch. Neurol. 53, 665–670 (1996).

6. Risse, G. L., Gates, J. R. & Fangman, M. C. A reconsideration of bilateral

language representation based on the intracarotid amobarbital procedure.Brain Cogn. 33, 118–132 (1997).

7. Rafal, R. Virtual neurology. Nat. Neurosci. 4, 862–864 (2001).8. Deppe, M. et al. Assessment of hemispheric language lateralization: a

comparison between fMRI and fTCD. J. Cereb. Blood. Flow. Metab. 20,263–268 (2000).

9. Walsh, V. & Cowey, A. Transcranial magnetic stimulation and cognitiveneuroscience. Nat. Rev. Neurosci. 1, 73–79 (2000).

10. Baynes, K., Eliassen, J. C., Lutsep, H. L. & Gazzaniga, M. S. Modularorganization of cognitive systems masked by interhemispheric integration.Science 280, 902–905 (1998).

11. Kurthen, M. et al. Interhemispheric dissociation of expressive and receptivelanguage functions in patients with complex-partial seizures: an amobarbitalstudy. Brain Lang. 43, 694–712 (1992).

12. Vigliocco, G. The anatomy of meaning and syntax. Curr. Biol. 10, 78–80 (2000).13. Embick, D., Marantz, A., Miyashita, Y., O’Neil, W. & Sakai, K. L. A syntactic

specialization for Broca’s area. Proc. Natl. Acad. Sci. USA 97, 6150–6154(2000).

14. Pascual-Leone, A., Gates, J. R. & Dhuna, A. Induction of speech arrest andcounting errors with rapid-rate transcranial magnetic stimulation. Neurology41, 697–702 (1991).

15. Jennum, P., Friberg, L., Fuglsang-Frederiksen, A. & Dam, M. Speechlocalization using repetitive transcranial magnetic stimulation. Neurology 44,269–273 (1994).

16. Epstein, C. M. Transcranial magnetic stimulation: language function. J. Clin.Neurophysiol. 15, 325–332 (1998).

17. Flitman, S. S. et al. Linguistic processing during repetitive transcranialmagnetic stimulation. Neurology 50, 175–181 (1998).

18. Wassermann, E. M. et al. Repetitive transcranial magnetic stimulation of thedominant hemisphere can disrupt visual naming in temporal lobe epilepsypatients. Neuropsychologia 37, 537–544 (1999).

19. Bakar, M., Kirshner, H. S. & Wertz, R. T. Crossed aphasia. Functional brainimaging with PET or SPECT. Arch. Neurol. 53, 1026–1032 (1996).

20. Hilgetag, C. C., Theoret, H. & Pascual-Leone, A. Enhanced visual spatialattention ipsilateral to rTMS-induced ‘virtual lesions’ of human parietalcortex. Nat. Neurosci. 4, 953–957 (2001).

21. Bottini, G. et al. The role of the right hemisphere in the interpretation offigurative aspects of language. A positron emission tomography activationstudy. Brain 117, 1241–1253 (1994).

22. Gazzaniga, M. S. et al. Collaboration between the hemispheres of acallosotomy patient. Emerging right hemisphere speech and the lefthemisphere interpreter. Brain 119, 1255–1262 (1996).

23. Karnath, H. O., Ferber, S. & Himmelbach, M. Spatial awareness is a functionof the temporal not the posterior parietal lobe. Nature 411, 950–953 (2001).

24. Knecht, S. et al. Non-invasive determination of hemispheric languagedominance using functional transcranial Doppler sonography: a comparisonwith the Wada test. Stroke 29, 82–86 (1998).

25. Knecht, S. et al. Reproducibility of functional transcranial Dopplersonography in determining hemispheric language lateralization. Stroke 29,1155–1159 (1998).

26. Willmes, K. & Poeck, K. To what extent can aphasic syndromes be localized?Brain 116, 1527–1540 (1993).

27. Lichtheim, L. On aphasia. Brain 7, 443–484 (1884).28. Chen, R. et al. Depression of motor cortex excitability by low-frequency

transcranial magnetic stimulation. Neurology 48, 1398–1403 (1997).29. Kosslyn, S. M. et al. The role of area 17 in visual imagery: convergent evidence

from PET and rTMS [published erratum appears in Science 284, 197 (1999)].Science 284, 167–170 (1999).

30. Muellbacher, W., Ziemann, U., Boroojerdi, B. & Hallett, M. Effects of low-frequency transcranial magnetic stimulation on motor excitability and basicmotor behavior. Clin. Neurophysiol. 111, 1002–1007 (2000).

31. Oldfield, R. C. The assessment and analysis of handedness: the Edinburghinventory. Neuropsychologia 9, 97–113 (1971).

articles


©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


articles

To process natural speech, a listener must first break the contin-uous stream of sound into recognizable units. However, there aretypically no reliable pauses between spoken words to indicatewhere one word ends and the next begins. Behavioral studies pro-vide evidence that a wide range of segmentation cues contributeto adults’ ability to segment continuous speech1–3. However, thesebehavioral studies are limited by their inability to establish thetime course of speech segmentation and to distinguish betweenfast, online segmentation and slower linguistic processing thatmay influence performance on specific tasks. Further, behavioralstudies cannot provide direct evidence about the brain systemsinvolved in online segmentation. In addition, it is often difficultto use the same behavioral task with different groups of subjects.For example, evidence of speech segmentation is found in younginfants4,5, bilingual speakers6,7, and monolingual adults usingdifferent tasks. Determining whether these groups are segment-ing speech along the same time course and employing the samemechanisms requires designing tasks that can be accomplishedby, and are equally engaging for, all groups.

The recording of ERPs provides an online measurement ofspeech segmentation suitable for listeners of all ages and back-grounds that also reflects the cortical organization of speech seg-mentation systems. We recently observed that initial syllableselicit a larger negativity around 100 ms (N100) than medial syl-lables presented in continuous speech8. This word-onset effectwas found for initial and medial syllables matched on loudness,length and other acoustic characteristics. However, it remainedpossible that the larger N100s evoked by word onsets index small,uncontrolled physical differences in the different syllable typesor in the syllables preceding them. To be certain that N100 word-

Segmenting nonsense: an event-related potential index of perceivedonsets in continuous speech

Lisa D. Sanders1,3, Elissa L. Newport2 and Helen J. Neville1

1 Department of Psychology, University of Oregon, 1227 University of Oregon, Eugene, Oregon 97403-1227, USA2 Department of Brain and Cognitive Sciences, University of Rochester, Meliora 414, Rochester, New York 14627-0268, USA3 Present address: Department of Linguistics, University of Maryland, 3416 Marie Mount Hall, College Park, Maryland 20742-7505, USA

Correspondence should be addressed to L.D.S. ([email protected])


Speech segmentation, determining where one word ends and the next begins in continuousspeech, is necessary for auditory language processing. However, because there are few directindices of this fast, automatic process, it has been difficult to study. We recorded event-relatedbrain potentials (ERPs) while adult humans listened to six pronounceable nonwords presented ascontinuous speech and compared the responses to nonword onsets before and after participantslearned the nonsense words. In subjects showing the greatest behavioral evidence of wordlearning, word onsets elicited a larger N100 after than before training. Thus N100 amplitudeindexes speech segmentation even for recently learned words without any acoustic segmentationcues. The timing and distribution of these results suggest specific processes that may be central tospeech segmentation.

onset effects index speech segmentation rather than acoustic char-acteristics that correlate with word boundaries, we measuredERPs while presenting the same physical stimuli before and afterthey were perceived as word onsets.

We recorded ERPs in response to six nonsense words pre-sented as continuous speech before and after listeners learnedthe words as lexical items through training. We reasoned that byteaching listeners the lexical items of a nonsense language, theymight begin segmenting continuous streams of those nonsensewords. Thus, we could compare ERPs to the same stimuli whenthey were not segmented as lexical items (before training) andwhen they were segmented as lexical items (after training). Usingthis protocol, we showed that N100 amplitude indexes speechsegmentation in the absence of acoustic segmentation cues.

RESULTSBehaviorThe continuous streams of nonsense words in the present studyhave been used previously to show that mere exposure to distri-butional regularities is sufficient for listeners to learn to distin-guish between nonsense words and part-word items onbehavioral tests5. However, for the subjects in this experiment,performance on the tests given before (Mean percent correct (M)= 53.7%) and after 14 minutes of exposure (M = 53.5%) did notdiffer from each other or from chance. In order to induce seg-mentation more quickly (a previous study5 used 21 minutes ofexposure), we used a training protocol to teach the six nonsensewords to the participants.

Performance on the behavioral test given immediately aftertraining was well above chance (M = 79.5%), indicating that at

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

articles


least some of the words had been learned. There were no differ-ences in accuracy between this test and another test given after asecond 14-minute exposure (M = 79.2%), indicating that listen-ers neither learned new words nor forgot the ones they knewimmediately after training. These scores were combined into asingle post-training accuracy score (M = 79.3%).

Event-related potentialsAcross all subjects, training had no effect on N100 amplitude.However, the difference in N100 amplitude before and after train-ing was highly correlated with individual performance on thepost-training behavioral tests (r = 0.80, P < 0.001). Subjects wholearned more of the words as measured by the behavioral testalso showed larger N100 word-onset effects (Fig. 1).

To determine if the word-onset effect was sig-nificant in the group of subjects that showed thelargest behavioral training effect, participants weredivided into two groups based on a median split ofpost-training accuracy scores. The 9 subjects whoshowed the largest effect of training (before train-ing M = 55.1%, after training M = 90.7%; t8 = 6.95,P < 0.001) improved to a greater extent than the 9subjects who showed the smallest effect of training(before training M = 52.2%, after training M =67.9%; t8 = 4.55, P < 0.01; group × training inter-action, F1,17 = 11.20, P < 0.01).

Only high learners showed a significant effectof training on N100 amplitude (group × training× anterior/posterior interaction, F5,80 = 2.91, P <0.05). For this group, the training × laterality ×anterior/posterior interaction was significant (F5,40= 5.22, P < 0.01). Over lateral electrodes, there wereno main effects or interactions including training.However, over medial and midline electrodes, therewas a main effect of training (F1,8 = 6.29, P < 0.05).

Thus for high learners, word onsets elicited a larger N100 overmedial and midline electrode sites after training (Fig. 2). For lowlearners, there was no effect of training on N100 amplitude.

To compare these results to our previous work8, we did an addi-tional comparison of the ERPs elicited by initial syllables and bymedial and final syllables. Although this comparison lacks theadvantage of contrasting the responses to physically identical stim-uli, it was important to determine if training was specifically influ-encing the processing of word boundaries. Even before training,initial syllables elicited larger N100s than medial and final sylla-bles across anterior electrode sites (position × anterior/posterior,F5,80 = 7.94, P < 0.001; at anterior sites only, position, F1,16 = 13.74,P < 0.001). This difference was likely due to the use of differentsyllables in these positions. However, it is also possible that thepre-training position effect indexed speech segmentation (based onlearning from distributional regularities) that did not influencebehavioral performance.

Importantly, high learners also showed a significant effect oftraining on N100 amplitudes elicted by syllables in different posi-tions (group × position × training interaction, F1,16 = 7.92, P <0.01). For the group of high learners (position × training inter-action, F1,18 = 15.64, P < 0.001), the difference between N100amplitude elicited by initial syllables and medial and final sylla-bles was larger after than before training at anterior electrodesites (high learners after training, position × anterior/posteriorinteraction: F5,40 = 8.548, P < 0.001; at anterior sites only, posi-tion: F1,8 = 21.40, P < 0.001). No such interaction was found forthe group of low learners.

Fig. 1. Performance on behavioral tests after training (percent correct)plotted against difference in N100 amplitude before and after training(before training minus after training). Subjects’ word learning as mea-sured by the behavioral tests correlated with their N100 word-onseteffects.

Fig. 2. ERPs averaged to word onsets before and aftertraining for the subjects showing the largest behaviorallearning effects (high learners). After training, wordonsets elicited a larger N100 at midline and medialelectrode sites. Words also elicited a larger N400 aftertraining.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


articles

For all subjects, the mean amplitude between 200 and 500 ms(N400) also showed effects of training (training × anterior/pos-terior interaction, F5,85 = 4.90, P < 0.001). The four most poste-rior rows of electrodes (reflecting the typical distribution of theN400) showed a main effect of training (F1,17 = 6.57, P < 0.05),indicating that the nonsense words elicited a greater negativityafter training. There were no significant interactions with group,indicating that both high and low learners showed the N400learning effect (Fig. 3).

The presence of N400 effects for both groups might influencethe amplitude of earlier (N100) responses to medial and final syl-lables. However, the interaction of the word-level N400 and thesyllable-level N100 would result in medial and final syllables elic-iting larger N100s, whereas the opposite pattern was found. Fur-thermore, these two components have distinct distributions; theN100 was largest over anterior electrodes, and the N400 waslargest over posterior electrodes.

DISCUSSIONThe N100 word-onset effect for nonsense words in the presentstudy is remarkably similar to our observations in a study ofprocessing real English8. For both real and nonsense words,word onsets elicit larger N100s across midline and medial elec-trode sites. The similarities in these findings are particularlystriking considering the differences between the stimuli in thetwo studies. We observed the N100 word-onset effect in sub-jects listening to their native language, complete with seman-tic, lexical, syntactic, phonological and acoustic information.The N100 word-onset effect was also observed in subjects lis-tening to nonsense sentences that contained only acoustic andphonological segmentation cues. In contrast, the present studyused just 6 nonsense words learned during a 20-minute trainingsession with no associated meanings and no acoustic segmen-tation cues, and again the N100 word-onset effect was observed.

Behavioral studies show that a wide variety of cuescan be used to segment speech; the N100 ERPresponse seems to index the perception of wordonsets regardless of the type or number of seg-mentation cues available.

It is not clear whether the early ERP word-onset effect reflects differences in the way initialand medial sounds are processed or the processof segmentation itself. It is possible that linguis-tic onsets in continuous speech are processed likeacoustic onsets and therefore elicit the same ERPcomponents observed for acoustic onsets. How-ever, mitigating evidence against this interpreta-tion is that the distribution of word-onset effects(medial and midline) was distinct from the morelateral distribution of N100s in general (F1,18 =9.351, P < 0.001). Alternatively, it is also possiblethat listeners direct greater attention to initial thanto medial sounds. The effects of auditory atten-tion to location and pitch are to increase N100amplitude9,10; similar auditory attention effectsmay be elicited by word onsets in continuousspeech.

One study of artificial language learning reports that a late negativity similar to the N400 found in the present study is sensi-tive to learning nonsense words11. After 50 hours of training on a68-word written language, newly learned words elicit a larger neg-ativity between 280 and 360 ms. During the same epoch, real English words elicit a larger negativity than pronounceable non-words or consonant strings. These results are similar to findingsfrom other studies in which words and orthographically legal non-words elicit a larger N400 than consonant strings12,13, and are con-sistent with an interpretation of the N400 as an index of lexicalsearch. In the present study, listeners may not have conducted lex-ical searches at all during pre-training, before they were aware ofrepeating nonsense words in the continuous stream of syllables.

It is important to note that N400 effects, like earlier N100 dif-ferences, indicate that speech has been segmented. That is, beforewe can find any components that are time-locked to onsets in con-tinuous speech, the speech must be processed as if it contains onsets.In the present study, if all syllables were processed in the same man-ner or if the syllables that were processed as onsets were distributedirregularly, N400s would not be time-locked to word onsets.

Interestingly, the group of subjects who showed the smallestbehavioral word-learning effects and no early ERP word-onseteffects (low learners) also had larger N400s after than beforetraining. We observed a similar pattern of results in a study oflate bilinguals listening to their non-native language14. In thatstudy, native Japanese late learners of English did not show N100word-onset effects when listening to English sentences; howev-er, they did show larger N400s in response to words as comparedto nonwords presented in continuous speech. There are severalpossible explanations for these findings. First, the high learners inthe present study and the native speakers in the earlier study seemto have been segmenting speech differently, and in particularfaster, than the respective groups of low learners and non-nativespeakers. That is, N100 amplitude may be indexing fast, online

Fig 3. ERPs averaged to word onsets before and aftertraining for the subjects showing the smallest behaviorallearning effects (low learners). After training, wordselicited a larger N400, similar to that found for the highlearners.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

articles


speech segmentation by the high learners and native speakers,whereas the later N400 effect may reflect slower or more variablesegmentation. A related explanation is that non-native speakersand low learners were segmenting speech, but not using process-es such as allocating greater attention to word onsets; if the N100word onset effect reflects primarily these latter processes, the lackof N100 and presence of N400 in these groups would be consis-tent with this interpretation.

The results of the present study indicate that the N100 word-onset effect in continuous speech cannot be explained solely onthe basis of acoustic differences in initial and medial sounds.Instead, the N100 effect indexes differences in the initial stagesof processing of these two types of sounds. Differential process-ing of initial and medial sounds within a word indicates thatspeech has been segmented; therefore, the ERP word-onset effectcan be used as an online measure of speech segmentation suit-able for a wide variety of stimuli and for listeners of all ages andbackgrounds. In addition, the timing and nature of this effectraise testable hypotheses concerning the specific mechanismsimportant in speech segmentation. Listeners who are more suc-cessful at segmenting speech (as measured by behavioral tests)show earlier segmentation effects. Word-onset effects occur veryearly, suggesting they involve either predictive or very fast, auto-matic processes. Furthermore, word-onset effects are similar evenwhen the available segmentation cues are very different.

METHODSThe procedure was approved by the University of Oregon Office ofHuman Subjects Compliance. Informed consent was obtained from allparticipants. Right-handed monolingual English speakers (n = 18) werefirst given 36 pairs of three-syllable nonsense words presented auditorally.On this pre-test, participants were asked to indicate which of the twoitems seemed more familiar. Each pair consisted of one of the six non-sense words that would later be presented in a continuous stream, andone part-word item constructed from the last syllable of one of the non-sense words followed by the first two syllables of another word.

Following the pre-test, participants were asked to listen carefully toa stream of sounds composed of the six trisyllabic nonsense words(babupu, bupada, dutaba, patubi, pidabu, and tutibu) as described2.The continuous stream was created from a pseudorandom list (the sameitem never occurred consecutively) of the 6 nonsense words repeated200 times each. The list was then modified such that all spaces betweenthe words were removed. From this list, a sound file was synthesizedusing a text-to-speech application. The resulting 14-minute speechstream contained no pauses or other acoustic indications of word onset(e.g., babupudutabatutibubabupubupadapidabu...).

ERPs were recorded from a 29-channel cap containing tin electrodes(Electro-Cap International, Eaton Ohio) during this portion of the exper-iment. Electro-oculogram was recorded from electrodes above and belowand at the outer canthi of the eyes. Impedences at all scalp electrode siteswere maintained below 3 kOhms. The EEG was amplified by Grass ampli-fiers with a bandpass of 0.01 to 100 Hz and sampled every 4 ms duringthe presentations of continuous speech. All electrodes were referencedto a single mastoid (right) online and later re-referenced to the averagemastoid (left and right). Participants were asked to look at a fixationpoint presented on the computer monitor for the duration of the soundstream. They were also asked to remain relaxed, not move, and blinknormally during this part of the experiment.

After 14 minutes of exposure to the stream of continuous nonsensewords, a second behavioral test was given to determine if participantshad learned some of the words by listening to the continuous stream.Because performance on this test did not differ from chance, a train-ing procedure was implemented.

For the training portion of the experiment, participants were specifi-cally instructed to learn the six trisyllabic nonsense words. During the first 10 minutes of training, the nonsense words were presented with a 500 ms ISI. As each word was heard from the speaker, the printed

version of the word was presented on the computer monitor. During thesecond 10 minutes of training, the nonsense words were presented witha 100 ms ISI.

Immediately following training, subjects were given a third test. Par-ticipants were instructed to circle the number corresponding to one ofthe six words they had just learned. ERPs were then recorded for anoth-er 14-minute period while subjects listened to a continuous string of thenonsense words. Following this second ERP recording session, a fourthand final behavioral test was given.

Artifact rejection algorithms were used to reject trials during whichblinks or eye movements occurred before ERPs were averaged. A com-parison of the number of trials rejected before and after training revealedno significant differences. ERPs recorded before and after training wereaveraged to the onsets of each syllable (initial, medial and final). To testthe hypothesis that the same ERP word onset effects described in previ-ous studies would be found for recently learned nonsense words, we mea-sured the peak amplitude between 70 and 130 ms (N100). We alsohypothesized that learned items might elicit an N400, an ERP compo-nent typically elicited by lexical items, so we measured the mean ampli-tude between 200 and 500 ms to test this hypothesis.

These dependent variables were analyzed using a four-factor, repeated-measures ANOVA: training (before, after) × electrode hemisphere (left,right) × electrode laterality (lateral, medial) × electrode anterior/posteriorposition (six levels). Additional ANOVAs were conducted for specific elec-trode sites as was motivated by training and electrode site interactions, aswell as with group (high learners, low learners) as a between-subjects factor.

AcknowledgmentsWe thank R. N. Aslin for help with the stimuli, Y. Yamada for help with data

acquisition, P. Compton and R. Vukcevich for technical support, and D. Coch,

S. Guion, D. Poeppel, M. Posner, M. Spezio and D. Tucker for comments on

previous drafts. Supported by the National Institute on Deafness and Other

Communication Disorders (NIH grants DC00128, DC00481 and DC00167)

and by the National Institute of General Medical Sciences (Institutional NRSA

5-T32-GM07257).


RECEIVED 17 APRIL; ACCEPTED 29 MAY 2002

1. Cutler, A. & Butterfield, S. Rhythmic cues to speech segmentation: evidencefrom juncture misperception. J. Speech Lang. Hear. Res. 31, 218–236 (1992).

2. Saffran, J. R., Newport, E. L. & Aslin, R. N. Word segmentation: the role ofdistributional cues. J. Mem. Lang. 35, 606–621 (1996).

3. Sanders, L. D. & Neville, H. J. Lexical, syntactic, and stress-pattern cues forspeech segmentation. J. Speech Lang. Hear. Res. 43, 1301–1321 (2000).

4. Jusczyk, P. W. Infants’ detection of the sound patterns of words in fluentspeech. Cogn. Psychol. 29, 1–23 (1995).

5. Saffran, J. R., Aslin, R. N. & Newport, E. L. Statistical learning by 8-month-old infants. Science 274, 1926–1928 (1996).

6. Cutler, A., Mehler, J., Norris, D. & Segui, J. Limits on bilingualism. Nature340, 229–230 (1989).

7. Sanders, L. D., Neville, H. J. & Woldorff, M. G. Speech segmentation by nativeand non-native speakers: the use of lexical, syntactic, and stress-pattern cues.J. Speech Lang. Hear. Res. (in press).

8. Sanders, L.D. & Neville, H.J. An ERP study of continuous speech processing:segmentation, semantics, and syntax in native speakers. Cogn. Brain Res. (inpress).

9. Hansen, J. C. & Hillyard, S. A. Endogenous brain potentials associated withselective auditory attention. Electroencephalogr. Clin. Neurophysiol., 49,277–290 (1980).

10. Hink, R. F., Hillyard, S. A. & Benson, P. J. Event-related brain potentials andselective attention to acoustic and phonetic cues. Biolog. Psychol. 6, 1–16 (1978).

11. McCandliss, B. D., Posner, M. I. & Givón, T. Brain plasticity in learning visualwords. Cogn. Psychol. 33, 88–110 (1997).

12. Chwilla, D. J., Brown, C. M. & Hagoort, P. The N400 as a function of the levelof processing. Psychophysiology 32, 274–285 (1995).

13. Holcomb, P. J. & Neville, H. J. Auditory and visual semantic priming in lexicaldecision: a comparison using event-related brain potentials. Lang. Cogn.Processes 5, 281–312 (1990).

14. Sanders, L.D. & Neville, H.J. An ERP study of continuous speech processing:segmentation, semantics, and syntax in non-native speakers. Cogn. Brain Res.(in press).

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om


errata

Receptive-field construction in cortical inhibitory interneuronsHarvey A. Swadlow & Alexander G. GusevNat. Neurosci. 5, 403–404 (2002)

We introduced an error in preparing this article for press. The fifth sentence of the fourth paragraph, “Notably, each of the TC neu-rons showed strong directional selectivity, but these preferred directions differed over a range of 13°,” should have read “over a rangeof 135°.”

Thalamcortical optimization of tactile processing according to behavioral stateMiguel A.L. Nicolelis & Erika E. FanselowNat. Neurosci. 5, 517–523 (2002)

The title of this article contained a typographical error. It should have read:

Thalamocortical optimization of tactile processing according to behavioral state

corrigendumProgressive induction of caudal neural character by graded Wnt signalingUlrika Nordström, Thomas M. Jessell and Thomas EdlundNat. Neurosci. 5, 525-532 (2002)

The authors wish to correct the phrase “rostral-to-caudal shift” on page 528, which should read “rostrocaudal shift”. The erroroccurs twice on this page.

©20

02 N

atu

re P

ub

lish

ing

Gro

up

h

ttp

://n

euro

sci.n

atu

re.c

om

nature neuroscience july 2002

Documents