Nature Neuroscience September 2000 PDF

2000 Nature America Inc. http://neurosci.nature.
com
contents
volume 3 no 9 september 2000
http://neurosci.nature.com
The cover shows the dendritic

complexity of a hippocampal
pyramidal neuron filled with fluo-
editorial
rescently tagged biocytin. Magee
2000 Nature America Inc. http://neurosci.nature.com
and Cook show that dendritic A debate over fMRI data sharing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845
EPSP amplitude increases with
distance from the soma, counter-
balancing the effects of dendritic letters to the editor
filtering. Thus, at the soma, distal
and proximal synapses have equal
efficacy. Photo courtesy of S. Does a stretch-inactivated cation channel integrate osmotic and
Watanabe, C. Bernard and D. peptidergic signals? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 847
Johnston. See pages 849 and 895.
news and views
Distant synapses raise their voices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 849
Nelson Spruston
SEE ARTICLE, PAGE 895
CNTF II, I presume?. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 851

Steven S. Lesser and Donald C. Lo
Probing the olfactory code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853

A new ligand for the C.Giovanni Galizia and Randolf Menzel
CNTF receptor complex. SEE ARTICLE, PAGE 927
Pages 851 and 867.
Coding for visual categories in the human brain . . . . . . . . . . . . . . . . . . . . . . . . . 855
Charles G. Gross
The mechanics behind spines on the move . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 856

10m
John Spiro
control +AMPA recovery
book review
control +AMPA recover
Glutamate receptors and actin Thinking about feeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 857

dynamics in dendritic spines. The Feeling of What Happens: Body and Emotion in the Making of Consciousness
Pages 856 and 887. by Antonio R. Damasio
REVIEWED BY ZACHARY F. MAINEN
Nature Neuroscience (ISSN 1097-6256) is published monthly by Nature America Inc., headquartered at 345 Park Avenue South, New York, NY 10010-1707. Editorial Office: 345 Park
Avenue South, New York, NY 10010. Telephone 212 726 9200, Fax (212) 696 9635. North American Advertising: Nature Neuroscience, 345 Park Avenue South, New York, NY 10010-
1707. Telephone (212) 726-9200. Fax (212) 696-9006. European Advertising: Nature Neuroscience, Porters South, Crinan Street, London N1 9SQ. Telephone (0171) 833 4000. Fax
(0171) 843 4596. New subscriptions, renewals, changes of address, back issues, and all customer service questions in North America should be addressed to Nature Neuro-
science Subscription Department, PO Box 5054, Brentwood, TN 37024-5054. Telephone (800) 524-0384, Direct Dial (615) 377 3322, Fax (615) 377 0525. Outside North America:
Nature Neuroscience, Macmillan Magazines Ltd., Houndsmill, Brunel Road, Basingstoke, RG21 6XS, U.K.. Tel: +44-(0)1256-329242. Fax: +44-(0)1256 812358. Email:
subscriptions@nature.com. Annual subscription rates: U.S./Canada: U.S. $650, Canada add 7% for GST (institutional/corporate), U.S. $199, Canada add 7% for GST (individual
making personal payment BN: 14091 1595 RT); U.K./Europe:435 (institutional/corporate), 185 (individual making personal payment), 99 (student); Rest of world (excluding
Japan): 480 (institutional/corporate), 195 (individual making personal payment), 110 (student); Japan: Contact Japan Publications Trading Co. Ltd., 2-1 Sarugaku-cho 1 chome,
Chiyoda-ku, Tokyo 101, Japan, phone (03) 292-3755. Back issues: U.S./Canada, $45, Canada add 7% for GST; Rest of world: surface U.S. $43, air mail U.S. $45. Reprints: Nature
Neuroscience Reprints Department, 345 Park Avenue South, New York, NY 10010-1707. Subscription information is available at the Nature Neuroscience homepage at http://neu-
rosci.nature.com. POSTMASTER: Send address changes to Nature Neuroscience Subscription Department, P.O. Box 5054, Brentwood, TN 37024-5054. Application to mail periodicals
postage rate is paid at New York, NY. Executive Officers of Nature America Inc: Nicholas Byam Shaw, Chairman of the Board; Jan Velterop, President; Edward Valis, Secretary-Trea-
surer. Printed by Publishers Press, Shepherdsville, KY, USA. Copyright 2000 Nature America Inc.
nature neuroscience volume 3 no 9 september 2000 i

contents
brief communications
Fast synaptic fatigue in shibire mutants reveals a rapid requirement for
dynamin in synaptic vesicle membrane trafficking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 859
F Kawasaki, M Hazen and RW Ordway
Motor timing learned without motor training . . . . . . . . . . . . . . . . . . . . . . . . . . . 860
Molecular memory by DV Meegan, RN Aslin and RA Jacobs
translocation of CaMKII.
Page 881.
commentary
Should the neuroscience community make a paradigm shift to
sharing primary data? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 863
S H Koslow
articles
CLF associates with CLC to form a functional heteromeric ligand for
the CNTF receptor complex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 867
GCA Elson, E Lelivre, C Guillet, S Chevalier, H Plun-Favreau, J Froger, I Suard,
A Benoit de Coignac, Y Delneste, J-Y Bonnefoy, J-F Gauchat and H Gascan

SEE NEWS AND VIEWS, PAGE 851
Control of Mller glial cell proliferation and activation following retinal injury . . . . . . 873
MA Dyer and CL Cepko
Molecular memory by reversible translocation of calcium/calmodulin-
Ensemble responses to odor
dependent protein kinase II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 881
contex in the moth. K Shen, MN Teruel, JH Connor, S Shenolikar and T Meyer
Pages 853 and 927.
Glutamate receptors regulate actin-based plasticity in dendritic spines . . . . . . . . . 887
M Fischer, S Kaech, U Wagner, H Brinkhaus and A Matus
Somatic EPSP amplitude is independent of synapse location in

hippocampal pyramidal neurons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 895
JC Magee and EP Cook
A network of electrically coupled interneurons drives synchronized

inhibition in neocortex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 904
M Beierlein, JR Gibson and BW Connors
The role of -CaMKII autophosphorylation in neocortical
experience-dependent plasticity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 911
S Glazewski, KP Giese, A Silva and K Fox
Competitive Hebbian learning through spike-timing-dependent
Category-specific neuronal synaptic plasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 919
responses in humans. S Song, KD Miller and LF Abbott
Pages 855 and 946.
Multi-unit recordings reveal context-dependent modulation of
synchrony in odor-specific neural ensembles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 927
TA Christensen, VM Pawlowski, H Lei and JG Hildebrand
Fixation neurons in the superior colliculus encode distance between

current and desired gaze positions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 932
A Bergeron and D Guitton
Activity in primary visual cortex predicts performance in a visual detection task . . . . . . . . 940
D Ress, BT Backus and DJ Heeger
Category-specific visual responses of single neurons in the human
medial temporal lobe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 946
G Kreiman, C Koch and I Fried
Motion distorts visual space: shifting the perceived position of

remote stationary objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 954
Motion distorts visual space. D Whitney and P Cavanagh
Page 954.
classified advertising
see back pages
nature neuroscience volume 3 no 9 september 2000 ii

editorial
A debate over fMRI data sharing

Michael Gazzaniga, a prominent cognitive neuroscientist, has second intervals for many minutes, yielding a time-series for each
caused a stir in the neuroimaging community. Gazzaniga is the of many thousands of voxels. Once these four-dimensional maps
director of the newly established National fMRI Data Center, a are generated and corrected for noise and motion artifacts, they
public archive based at Dartmouth College that seeks to provide can be analyzed in many ways. The simplest is to compare the
a repository for the enormous datasets generated by fMRI exper- average activity between different cognitive tasks, in order to iden-
iments. On the face of things, this would seem to be a welcome tify groups of voxels that are significantly activated (or deacti-
move, but Gazzanigas announcement has alarmed many of his vated) by a given task relative to a control condition. More
colleagues, primarily because of his suggestion that several jour- complex methods include parametric designs, in which one looks
nalsincluding Journal of Cognitive Neuroscience, of which he is for graded activation that is correlated with some cognitive vari-
the editorwere planning to require the release of primary data able; event-related analysis, to identify patterns of activity that
as a condition of publication. are temporally correlated with the stimulus or the behavioral out-
In response to Gazzanigas announcement, a circular letter, put; and higher-order analyses in which one looks for correla-
coordinated by Isabel Gauthier of Vanderbilt University and tions between different brain regions in order to make inferences
signed by more than forty researchers, was recently sent to the about their functional interactions.
editors of various leading journals (including both Nature and Finally, the activations must be superimposed onto some stan-
Nature Neuroscience), expressing concerns about the proposed dard anatomical frame of reference. The most common of these
policy and asking them to clarify their positions. Journal of Cog- is Talairach coordinate space, but because of the variability
nitive Neuroscience intends to stand by its stated policy, but most between individual brains, Talairach coordinates do not corre-
other journals are taking a more cautious approach, and are dis- spond precisely to anatomical structures, and nor do the anatom-
tancing themselves from any suggestion that they might compel ical features of the cortex correspond precisely to its functional
their authors to release data. Richard Frackowiak of University subdivisions. There is thus a chain of inferences involved in iden-
College London, who is editor of Neuroimage, put it bluntly in tifying a given activation with a specific brain region.
his public response to Gauthier: I would like to take a strate- The datasets generated by a single study can run to gigabytes
gy of minimal response to this rather unfortunately launched or even terabytes, many orders of magnitude more than can be
proposal. depicted in printed form. Typically, what finally ends up in a pub-
Although Gazzaniga may have misjudged the level of support lished report is a list of locations (often in Talairach coordinates)
he could expect from the community, he has clearly forced the that showed significant activation in the condition of interest.
pace in an important debate. Sharing of published data is stan- These coordinates represent only the center of each activation,
dard practice in many other fields, and in a commentary on page and say nothing about its absolute magnitude or spatial distrib-
863 of this issue, Stephen Koslow of the US National Institute of ution. Although it is common to discuss some of the activations
Mental Health (NIMH) argues that neuroscience risks falling in greater detail, and to illustrate them with images and/or graphs,
behind if it does not adopt similar practices. Gazzanigas initiative the final report nevertheless represents a very reduced subset of
represents a bold attempt to push the fMRI field in this direc- the original data, filtered through a series of transformations and
tion, and as such, the debate deserves wide attention. analyses that are often idiosyncratic. There is no consensus about
To understand the issues, and the difficulties inherent in the right way do these analyses; each has its strengths and weak-
archiving brain imaging data, it is important to consider the steps nesses, and new methods are constantly being developed as the
that go into a typical experiment. Functional MRI detects changes field evolves.
in blood oxygenation and blood flow that occur when brain tis- This would seem to add up to a strong case for sharing the
sue becomes more active. Subjects are therefore scanned as they raw data, along with the associated analytical tools. Supporters
perform various mental tasks, and by comparing the task of inter- of this view identify a number of advantages. Perhaps most
est against some baseline, one can identify brain regions activat- importantly, it would allow meta-analysis, in particular to iden-
ed by the task. Within this general framework, however, there tify common features of tasks that activate specific brain regions.
exist an enormous variety of protocols, both for performing the Marcus Raichle of Washington University in St. Louis describes
scans and for analyzing the resulting datasets. this as an incredibly powerful approach, particularly for brain
The raw output of the scanner must first be converted into structures whose functions are poorly understood. However, it
three-dimensional space, yielding a map of signal amplitude that is still extremely difficult to do this type of analysis; the lists of
is divided into voxels, volume elements that are typically a few activations reported in the published literature are not rich
millimeters across. Each subject may be scanned at two- or three- enough to support detailed comparisons, and attempts to com-
nature neuroscience volume 3 no 9 september 2000 845

editorial
pare raw data across multiple studies are often hampered by the fields such as structural biology, and provided that the data-
lack of common data formats. base can support timed access, individual journals can set time
The ability to examine primary data would also allow limits as they see fit.
researchers to investigate the robustness of published conclusions The National fMRI Data Center is not itself in a position to
to different analytical methods and statistical significance thresh- enforce a policy of data sharing. In general, the strongest enforce-
olds. Any choice of threshold involves a subjective decision about ment for such policies comes from funding bodies, which can
the appropriate trade-off between false positives and false nega- insist on disclosure as a precondition for grant support. NIH, for
tives, and activations that fail to reach a stringent significance example, requires this in fields such as genomic research and
criterion are often not reported. Moreover, as John Gabrieli of structural biology, and grants are designed to cover the costs of
Stanford University notes, there may be a systematic bias to the archiving data and physical samples. Steven Hyman, the direc-
published literature in this regard; it is difficult to write papers tor of NIMH, favors the idea of archiving fMRI data, at least in
describing complex patterns of activation, and so there is often principle, commenting that right now, based on published
a temptation to set a conservative threshold, in order to reduce images, investigators really cannot tell whether their findings are
the number of activations and present a simpler story. more than crudely in agreement. Hyman emphasizes, however,
Not everyone is convinced, however, that these benefits will that in all cases where sharing is mandated, it has been after
actually materialize as a result of the Dartmouth initiative. There extensive consultation with the community. Before any policy is
are two major obstacles that must be overcome if the archive is made regarding fMRI data, he says, it is important that the field
to have any chance of success. First, it must be able to store be engaged in a discussion of what is to be shared, with what tim-
detailed descriptions not only of the imaging data but also of the ing, and how.
psychological context of each study. The field has moved beyond That discussion will need to be broad and inclusive, both sci-
brain mapping (although the name has stuck), and most fMRI entifically and geographically, and its scope will undoubtedly
studies are now designed to test specific hypotheses. Thus, if the extend beyond the Dartmouth initiative (which calls itself a
data are to be interpreted, let alone replicated, they must be national center and has an advisory board that is purely US-
accompanied by detailed descriptions of stimuli, behavioral based). One important forum will be the Organization for
responses and, in some cases, psychological or clinical data about Human Brain Mapping, which has established a task force on
individual subjects. Codifying such diverse datasets represents a neuroinformatics, chaired by Jonathan Cohen of Princeton Uni-
formidable computing problem, and it remains to be seen how versity and comprising researchers from both North America
effectively the Dartmouth team will be able to solve it. and Europe. Cohen emphasizes that his group has no mandate
The second prerequisite for success is that the archive must to impose standards, and might not even make specific recom-
be easy to use. The sheer size of the datasets means that trans- mendations; at this stage, its mission is to reach a consensus on
mitting them is not trivial, and in addition to size, the file for- the issues that need to be considered by the field as a whole if
mats will inevitably be very complex, much more so than for data sharing is to become a reality.
(say) genomic data. The Dartmouth team warns that some In the meantime, what role should journals take? In princi-
growing pains may be expected, and it is not yet clear how ple, every journal is free to impose whatever policy it chooses. In
painful they will be. An early indication will come from a forth- reality, however, editors must balance the interests of their read-
coming special issue of Journal of Cognitive Neuroscience, edited ers with those of their contributors. A journal that fails to pro-
by Mark DEsposito (U.C. Berkeley) and devoted entirely to fMRI vide readers with enough information to evaluate its papers will
studies. The authors of the research papers have all agreed to lack scientific credibility, whereas one that makes excessive
deposit their data into the Dartmouth archive, effectively serv- demands of authors will find itself losing submissions. Journals
ing as its beta-testers. At the time of writing, however, no data operate in a free market, and those whose policies fail to reflect
have been deposited, and DEsposito anticipates that it may be the consensus within their communities will suffer scientifically
many months before the process is complete. and ultimately commercially.
In addition to these technical challenges, several other con- How to balance these factors depends on the field. In genetics,
cerns have been raised. Perhaps the most serious is that many for instance, most journals insist that DNA sequence data be pub-
datasets require months of analysis and lead to multiple pub- licly available at the time of publication. In structural biology,
lications. If the data are released at the time of the first publi- some journals insist on making crystal coordinates available at
cation, the authors could be scooped to other discoveries the time of publication, whereas others allow an embargo period
arising from their work. The concern seems reasonable in (typically 6 to 12 months). In general, any data-sharing policy
principle, and few would question the importance of main- can only be effective if two criteria are met. First, there must be an
taining incentives for authors to do experiments. It is unclear efficient technological infrastructure so that data can be con-
how often this would happen in practiceto publish a scien- tributed, retrieved and reanalyzed without excessive effort. Sec-
tifically credible analysis would require a depth of under- ond, there must be a political consensus within the field that the
standing that could probably only come from discussion with benefits of sharing data outweigh the costs. In the case of neu-
the authorsbut in any case, it can easily be addressed by stor- roimaging data, neither criterion has yet been met, in our view,
ing the data under an embargo for some period after publica- and so although we have no objection to our authors depositing
tion (perhaps a few months) to allow the authors a period of their fMRI data in the National Data Center (or any other simi-
exclusive access. Such arrangements are common in other lar archive), we shall not insist that they do so.
846 nature neuroscience volume 3 no 9 september 2000

letters to the editor
Does a stretch-inactivated cation channel

integrate osmotic and peptidergic signals?
TO THE EDITOR In a recent paper in steady-state Popen is approximately 2-fold email: sachs@buffalo.edu, cmorris@lri.ca or
Nature Neuroscience, Chafke and higher at 0 cm H 2O than at any other ohamill@utmb.edu
Bourque1 assert that mechanisms under- pressure in a range of 175 cm H2O. The
lying osmoreception [in osmosensory two-fold dynamic range for SIC activity REPLY Sachs and colleagues dispute our
neurons] are understood, and more is unimpressive. We are assured that in conclusion that stretch-inactivated cation
specifically, that a stretch-inactivated 12 patches... Popen of peptide stimulated channels serve as a point of molecular
cation channel (SIC) is a point of mole- channels could be modified by changing convergence for osmotic and peptidergic
cular convergence for osmotransduction hydrostatic pressure. This is not data modulation in magnocellular neurose-
and peptide-induced excitation. An entic- nor is the allusion to three [other] patch- cretory cells (MNCs)1. It is argued that
ing idea, but the data do not stand up to es in which channel P open varied as a data reported in this paper cannot be
scrutiny. bell-shaped function of pipette pressure. compared with previous results from our
Gd3+, with its many side effects on cal- Statistics are needed, as well as more cur- laboratory8,9 due to differences in channel
cium2, potassium3, sodium3, non-selec- rent traces. properties and because Gd3+ lacks selec-
tive4 and cation-selective5 channels, is a How reliable is the bell-shaped tivity as a pharmacological probe.
problematic diagnostic tool6. Neverthe- response? Previously8,9 the authors renor- They assert that traces in Fig. 6a of our
less, the evidence offered for molecular malized applied pressures so that 0 cm report1 suggest that Gd3+ increases chan-
convergence is that 10 M Gd 3+ and H2O did not signify atmospheric pres- nel current amplitude, whereas we had
100 M Gd 3+ shorten the mean open sure, but rather the pressure at which previously reported a small (10%)
times (MOT) of peptide-stimulated and NPopen was maximum. Therefore, from reduction in current amplitude9. Unlike
control SICs by similar percentages. Fig. 5b one cannot infer that membrane in the previous study 9 , however, the
This MOT comparison is made against tension increased on either side of 0 cm recordings shown in Fig. 6a were obtained
previously published and highly variable H2O (ref. 11). Furthermore, channels in from different cells, not from a single
findings from which representative cell-attached patches can experience patch. The excerpts shown were simply
excerpts8,9 were 10100 ms bursts, unlike uncontrolled variations in kinetics, so that selected to illustrate the effects of Gd3+ on
the present events (Fig. 3a). normalizing data records to the highest open time. Because it is unlikely that the
Figure 6, illustrating peptide-stimu- activity inevitably leads to bell-shaped trans-patch voltage was precisely the same
lated events modulated by 100 M Gd3+, curves, particularly when the dynamic in these two particular recordings, it is
is disconcerting for several reasons. The range is only two. inappropriate to compare current ampli-
100 M Gd 3+ produces a fast-flickery In previous papers claiming that SICs tudes. We maintain that, at equivalent
block of cation channels (including SICs underlie osmotransduction8,9, the report- voltages, Gd3+ causes a reduction in cur-
in mammalian muscle7), yet the single- ed pressures were about 50-fold smaller rent amplitude.
channel events illustrating Gd3+ block are (2 cm versus 100 cm H 2O; Fig 6b). It is stated that channel openings
distinctly larger than control currents. Several years ago, we alerted Dr. Bourque shown in earlier studies8,9 consisted of
Earlier work 9 showed reduced single- that the published pressures seemed bursts of 10100 ms, unlike the events
channel amplitude with Gd3+, and these extremely lowbut this discrepancy goes shown in the recent paper 1 . We have
old data constitute the controls for Fig. 6. unmentioned and certainly invalidates any acknowledged the complexity of SIC
The finding that the MOT for channel comparison between the present data and kinetics in MNCs9. Although we did not
activity occurring in long bursts8,9 is sim- the older data (Fig. 6b). overtly emphasize bursting in our recent
ilar to the MOT of openings that seldom We are not convinced that mecha- paper1, bursts were nonetheless present
occur in bursts (the new data, Fig 3a) is nisms underlying osmoreception are and can be seen in many of the traces
hardly a flag that identical proteins pro- understood and we disagree that the shown, including in Fig. 6a. More
duced both responses. Also, Gd3+ interacts new data support a molecular conver- importantly, the average mean open
strongly with many anions, including pro- gence. time reported in our earlier study
teins10, so whole-cell responses to 200 M (1.65 0.08 ms; ref. 9) is indeed consis-
Gd3+ (Fig. 6d), which may reflect action Frederick Sachs1, Catherine E. Morris2 tent with the data shown in our recent
on the peptides and/or multiple channel and Owen Hamill3 paper (MOT 1.5 ms; Fig. 3a)1.
types, are not grounds for concluding that 1Physiology and Biophysical Sciences, 320 Cary It is also implied that interactions
identical channels carry peptide-stimu- Hall, SUNY, Buffalo, New York 14214, USA between Gd3+ and exogenously applied
lated and osmotransducing currents. 2Departments of Medicine and Biology, peptides, or other ion channels, might
Ongoing channel activity can be easily University of Ottawa, Ottawa Hospital, 725 have caused the loss of peptide-evoked
characterized as stretch-inactivating by Parkdale Avenue, Ottawa, Ontario K1Y 4K9, and/or resting currents in our experi-
applying brief pulses of pipette suction11, Canada ments. We do not dispute that Gd3+ can
but Chafke and Bourque never use this 3University of Texas Medical Branch, have a variety of effects on other types of
direct approach to show peptide-activated Department of Physiology and Biophysics, ion channels9, and it is indeed conceivable
channels are SICs. Instead, they present 301 University Blvd., Galveston, Texas that Gd3+ might have interfered with the
one doseresponse curve in which the 77555, USA ability of the peptides to activate receptors.

letters to the editor
However, whereas the effects of peptides elsewhere in the paper. It is thus clear to Gd3+ of the peptide-stimulated chan-
routinely outlasted washout by 12 min- from the data that maximal Popen occurs nels1 and of the SICs studied previously9
utes, application of Gd3+ to cells continu- near zero applied pressure. We feel that represents part of the evidence that they
ously exposed to a peptide initiated an our approach is entirely suited to the are identical, this conclusion is further
almost immediate reversal of the response analysis of mechanosensitive gating and supported by the existence of a similar
(Fig. 6c and d), suggesting that this effect that the bell-shaped relationship conductance and ionic selectivity (Fig. 2),
occurred downstream of the receptor. obtained is not an artifact5, but a bona and by the evident stretch-inactivating
Moreover, Gd3+ has no obvious effect on fide indication that the channels present properties of the peptide-stimulated
resting currents in MNCs bathed in hypo- in MNCs are of the SIC variety. channels (Fig. 5). We thus stand by our
tonic solution 9,13 (when channels are Finally, Sachs and colleagues indicate conclusion that SICs provide a point of
inhibited by cell swelling). The inhibitory that the mechanosensitivity of the chan- molecular convergence for osmotic and
effects of Gd3+ on excitatory responses to nels described in our recent paper1 appears peptidergic regulation in MNCs1.
the three peptides, therefore, were not due to be much lower than in our original stud-
to a simple blockade of receptor activation ies8,9. This is due to a technical fault that Charles W. Bourque1 and
or to a nonspecific effect on one or more resulted in a scaling error in the earlier Yassar Chakfe2
additional conductances. studies, and we welcome the opportunity 1Centre for Research in Neuroscience,
Sachs and colleagues question the to clarify this point. In our recent study1, Montreal General Hospital and McGill
appropriateness of our analysis of pipette pressure (adjusted with a syringe) University, 1650 Cedar Avenue, Montreal
mechanosensitive gating. Part of the was monitored via a connection to the QC H3G 1A4, Canada
problem may stem from an impression proximal end of a U-shaped water 2Montreal Neurological Institute, 3801
(see review by Sachs and Morris12) that manometer in which the distal end was University Street, Montreal QC H3A 2B4,
our experimental protocol exclusively open to the atmosphere. Values obtained Canada
involves the use of negative pressures and in this way report absolute steady-state email (C.W.B): mdbq@musica.mcgill.ca
a systematic normalization of the data. pressure, relative to the atmosphere. Unfor-
In fact, our strategy consists of monitor- tunately in the previous studies 8,9, the 1. Chakfe, Y. & Bourque, C. W. Nat. Neurosci. 3,
ing changes in P open that result from apparatus consisted of a sealed tube, such 572579 (2000).
application of a broad range of negative that the pressures applied were severely 2. Mlinar, B. & Enyeart, J. J. J. Physiol. (Lond.)
469, 639652 (1993).
and positive pressures to the recording underestimated. A post hoc calibration of
3. Elinder, F. & Arhem, P. Biophys. J. 67, 7183 (1994).
pipette. We believe that upon exposing a the apparatus (following a query from Dr.
membrane patch to the effects of a range Morris) revealed that values reported in 4. Hase, C. C., Le Dain, A. C. & Martinac, B.
J. Biol. Chem. 270, 1832918334 (1995).
of applied pipette pressures, there will be those two studies8,9 (although linearly relat-
5. Yang, X. C. & Sachs, F. Science 243, 10681071
a point at which tangential membrane ed over the range concerned) reflected only (1989).
tension will be minimal (near zero pres- 22.5% of the pressures actually applied
6. Hamill, O. P. & McBride, D. W. Jr. Pharmacol.
sure) and from which imposed positive during those experiments. Although the Rev. 48, 231252 (1996).
or negative pressures will both provoke error is unfortunate, it in no way invali- 7. Franco, A. Jr., Winegar, B. D. & Lansman, J. B.
proportional increases in tangential dates any of the conclusions reached in Biophys. J. 59, 11641170 (1991).
stretch 12 . We thus surmise that SICs those two papers. When considering the 8. Oliet, S. H. & Bourque, C. W. Nature 364,
would be expected to display a bell- difference in scale, the mechanosensitivity 341343 (1993).
shaped relationship between P open and of the SICs reported in these previous 9. Oliet, S. H. & Bourque, C. W. Neuron 16,
pipette pressure, with maximal P open papers8,9 is indeed comparable to the val- 175181 (1996).
lying near zero pressure. In Fig. 5a of our ues reported in our recent study1. 10. Caldwell, R. A., Clemo, H. F. & Baumgarten,
recent paper 1, the traces show channel We maintain that the quantitative C. M. Am. J. Physiol. 275, C619C621 (1998).
activity recorded at absolute applied similarity between the effects of Gd3+ on 11. Morris, C. E. & Sigurdson, W. J. Science 243,
807809 (1989).
pressures of 100, 0 and +70 cm H 2O, single stretch-inactivating channel MOT
and the graph in Fig. 5b reports how and macroscopic osmoreceptor currents9 12. Sachs, F. & Morris, C. E. in Rev. Physiol.
Biochem. Pharmacol. (eds. Blaustein, M. P. et
P open varied as a function of absolute provides strong evidence for the involve- al.) 178 (Springer, Berlin, 1998).
pipette pressure for the same patch. Pres- ment of these channels in osmotrans- 13. Oliet, S. H. R. Osmoreception in Rat Supraoptic
sure values were not normalized here or duction8. Although the similar sensitivity Neurons. Thesis, McGill Univ. (1994).

news and views
Distant synapses raise their voices

Nelson Spruston
By showing that synaptic strength increases as a function of distance from the soma, Magee
and Cook have solved the long-standing puzzle of how synapses on distal dendrites can
influence action potential initiation.
Elaborately branching dendrites allow part due to the discovery that most neu- the dendritic electrode, increased with
neurons to integrate several thousand rons contain a variety of voltage-gated distance from the soma.
synaptic inputs and permit complex channels in their dendrites (for review, The mechanisms underlying dis-
spatial interactions, which may enhance see ref. 4). The second possibility is that tance-dependent scaling of synaptic
the computational power of individual distal synapses are simply stronger strength are unclear, but will undoubt-
neurons. Yet because synaptic potentials (synaptic scaling). Although some ear- edly be a prominent target for future
attenuate as they travel along dendrites, lier studies supported the idea of dis- investigations. Among the possibilities
the large size of many dendritic trees tance-dependent synaptic scaling (for to be considered are presynaptic differ-
presents a problem. How are synapses discussion, see ref. 1), the present study ences in the size or number of released
on distal dendrites heard in the axon, removes any doubt by showing directly glutamate-containing vesicles and post-
where neuronal output, in the form of that distal synapses on CA1 neurons synaptic differences in the number, con-
an action potential, is generated? In this produce larger excitatory synaptic ductance or open probability of
issue, Jeff Magee and Erik Cook 1 potentials and currents (EPSPs and glutamate receptor channels. In support
demonstrate that neurons solve this EPSCs) than more proximal synapses on of the latter mechanism, distal dendrites
potential problem by increasing the the same neurons. Magee and Cook pre- show increased sensitivity to optically
strength of distal synapses. In effect, dis- sent convincing evidence that the far- uncaged glutamate in CA1 neurons 5 .
tal synapses speak up so they can be ther a synapse is from the soma of a Alterations in the clearance of glutamate
heard at a distance. CA1 neuron, the louder it speaks from the synaptic cleft is another possi-
When I telephone from the U.S. to (Fig. 1). As a result, synaptic potentials ble mechanism that must be considered.
Germany, my voice is transmitted along can attenuate considerably as they prop- One factor that cannot explain Magee
cables over long distances, with no more agate along dendrites, yet their ampli- and Cooks findings is distance-depen-
attenuation than if I were phoning a col- tude at the soma remains independent dent scaling of vesicle release probabili-
league down the hall. To accomplish this of the site of origin.
feat, scientists have developed clever To measure attenuation of
strategies to deal with the inherent ten- EPSPs, Magee and Cook1 used
dency for signals to attenuate as they simultaneous patch-clamp
propagate along electrical or optical recordings from the soma and
telephone cables2. Neurons must solve a dendrite of the same cell.
a similar problem, because synaptic They elicited EPSPs near the
potentials generated in distal dendrites dendritic recording electrode
must travel long distances to reach the using several clever tricks. The
axon. Like undersea telephone cables, most compelling data came
dendrites are long, narrow transmission from local application of a
lines bathed in an electrically conduc- high-osmolarity solution to
tive medium 3 . Current flowing along the dendrite, near the record-
these cables can leak out into the sur- ing electrode. High-osmolari-
rounding medium, resulting in consid- ty solutions are known to
erable signal attenuation. Current leak trigger spontaneous release of
from dendrites is substantial, because neurotransmitter from synap- Fig. 1. Conductance scaling equalizes the strength of EPSPs
arising on distal dendrites. A reconstructed CA1 pyramidal
dendrites lack insulation like the myelin tic terminals. By applying the neuron is shown with a proximal (red, 76 m from the soma)
sheath surrounding axons. solution locally, the release and a distal (blue, 632 m from the soma) synaptic input.
Two possible solutions to this prob- sites should be restricted to Computer simulations of a small synaptic conductance at
lem have been considered. One is that locations close to the dendrit- each site produces large depolarizations at the synapse (red
distal synaptic potentials may be ampli- ic recording electrode. Having and blue), which attenuate to somatic EPSPs (green) of about
fied by voltage-gated channels (synap- done this for a range of posi- 0.2 mV at the soma for the proximal synapse, but a much
tic amplification). This hypothesis has tions on the apical dendrite, smaller EPSP (about 0.03 mV) for the distal synapse. To pro-
received considerable attention, in large Magee and Cook found that duce a somatic EPSP of 0.2 mV in response to activation of
the mean EPSP amplitude the distal synapse, the synaptic conductance 1was scaled up by
a factor of about seven. Magee and Cook show that the
Nelson Spruston is at the Department of Neurobiology measured at the soma was
somatic EPSP amplitude remains constant (with a mean value
and Physiology and Institute for Neuroscience, remarkably independent of of 0.2 mV), regardless of the dendritic position of the acti-
Northwestern University, 2153 N Campus Drive, synapse location, whereas the vated synapse and that this is achieved by distance-depen-
Evanston, Illinois 60208, USA. mean amplitude of the same dent scaling of synaptic conductance. Reconstructed CA1
e-mail: spruston@northwestern.edu EPSPs, when recorded from neuron provided by Tim Mickus.

news and views
common to all neu- Na+ and Ca2+ channels, essentially can-

ronal types. In some celing each other out and resulting in
cell types, passive nor- no net amplification of small EPSPs in
malization may obviate CA1 neurons 8,9 . But if voltage-gated
the need for synaptic channels in dendrites do not amplify
scaling. Other neurons synaptic potentials, what are they there
may use amplification for? One answer is that they mediate
by voltage-gated chan- active propagation of action potentials
nels. Answers to these from the axon back into dendrites 10 ,
questions are likely to which are crucial signals involved in
follow as the methods the induction of synaptic plasticity 11.
used by Magee and Another answer is that they can drive
Cook are applied to regenerative Na+ and Ca 2+ spikes initi-
other neurons. ated in dendrites, which can contribute
Magee and Cooks to the generation of one or more action
study is also likely to potentials in the soma 12 . Dendritic
spawn studies of how and spikes propagate poorly to the soma,
Fig. 2. Distance-dependent variability of temporal summation is when the gradient of however, suggesting that they may have
limited by a high density of hyperpolarization-activated channels in synaptic strength is estab- other, more local functions, perhaps
dendrites. Five EPSPs were simulated at 30 Hz for a proximal (left) lished during development also related to synaptic plasticity.
and distal synapse (right), and the resulting somatic EPSPs are and whether the same Because dendritic spikes require strong
shown. Because the somatic EPSP is filtered more for the distal mechanisms are tapped by synaptic stimulation, they would seem
synapse than for the proximal synapse, activation of the distal activity-induced changes to be involved primarily in modifying
synapse produces more temporal summation. Inclusion of a hyper-
in synaptic strength such a neurons response to concurrent acti-
polarization-activated cation conductance in the model, with an
increasing density as a function of distance from the soma, results in as long-term potentiation vation of many synapses, perhaps with
selective reduction of temporal summation of the somatic EPSP in (LTP). Indeed, one obser- particular spatial patterns.
response to activation of the distal synapse. This normalization of vation of the present study The most compelling example of
temporal summation was shown in a previous paper by Jeff Magee13. is pertinent to the debate subthreshold regulation of synaptic
swirling around the mech- potentials by voltage-gated channels in
anisms of LTP. Even dendrites was reported recently in
though Magee and Cook another Nature Neuroscience paper by
ty. Because the authors averaged only recorded EPSPs evoked from restricted Jeff Magee 13 , who found that a high
EPSPs that were detectable in the den- locations on the dendritic tree, they still density of hyperpolarization-activated
drite, release failures did not affect the observed considerable variability in EPSP cation channels in CA1 dendrites
amplitude of the EPSPs they analyzed. and EPSC amplitude at each site. Thus, strongly influences temporal summa-
One possible explanation for synap- although there is an average tendency for tion of EPSPs. Ralls passive cable the-
tic scaling that the authors needed to EPSPs at distal synapses to be larger, there ory predicts that synaptic potentials
exclude is something called passive nor- is considerable variability superimposed should be filtered by dendritic capaci-
malization. In this mechanism, synap- on this general trend, leaving room for the tance, resulting in somatic EPSPs that
tic strength may be normalized for amplitude of individual synapses to be rise and decay more slowly if they are
distance simply by virtue of the mor- increased or decreased by synaptic plas- generated in distal dendrites rather
phology and passive membrane proper- ticity. Notably, some of the smaller EPSPs, than proximal dendrites 3. A corollary
ties of a neuron 6 . In short, distal which could easily be detected in the den- of this is that EPSPs arising from more
synapses that contact small, high- drite, could not be resolved in the soma. distal synapses should summate at the
impedance dendrites will generate larg- This finding implies that many silent soma over longer time courses than
er EPSPs, unless the synaptic current is synapses, often studied as targets for EPSPs generated in more proximal
very fast. Working against passive nor- potentiation via insertion of postsynaptic dendrites. Magee reported that this
malization in CA1 neurons, however, glutamate receptors7, might actually be does not happen. Instead, a gradient of
are the large, low-impedance apical den- whisperingtoo soft-spoken to be heard hyperpolarization-activated ion chan-
drites and the rapid time course of at the soma. Studies with dendritic nels, whose density increases with dis-
EPSCs6. By substituting dendritic cur- recording and clever tricks for evoking tance from the soma, accelerates the
rent injection for synaptic activation, unitary EPSPs, like those used by Magee decay of each EPSP, rendering the
Magee and Cook showed directly that and Cook, are sure to be exploited in decay and summation of EPSPs at the
passive normalization cannot account future studies of LTP. soma comparable, regardless of synap-
for their observations in CA1 neurons. The new findings presented by tic location (Fig. 2).
Furthermore, the authors used dendrit- Magee and Cook indicate that dendrit- Both modern optical cables and
ic voltage-clamp recordings to show that ic Na+ and Ca2+ channels do not ampli- old-fashioned electrical cables com-
EPSCs increased with distance from the fy small EPSPs in CA1 neurons. The pensate for signal attenuation using
soma, arguing convincingly for scaling finding, about three years ago, that repeaters, spaced tens of kilometers
of synaptic conductance. CA1 dendrites also contain a high den- apart, that amplify the signals. Because
The discovery that synapse strength sity of low-threshold, inactivating K + these devices are expensive (and hard
is increased for distal synapses raises the channels is likely to explain why; these to fix when they lie on the ocean
question of whether this mechanism is K+ channels are activated together with floor), engineers have recently begun

news and views
to develop technology that amplifies unravel the mechanisms of learning 6. Jaffe, D. B. & Carnevale, N. T. J. Neurophysiol.
signals at the launch end 14 . Thus, it and memory in the hippocampus, we 82, 32683285 (1999).
seems that telecommunications engi- now know that excitator y synapses 7. Malenka, R. C. & Nicoll, R. A. Neuron 19,
neers have begun, unwittingly, to should be treated equally regarding 473476 (1997).
mimic a trick used by neurons to com- their strength and summation in the 8. Hoffman, D. A., Magee, J. C., Colbert, C.
M. & Johnston, D. Nature 387, 869875
pensate for attenuation. Neurons, soma, regardless of their position on (1997).
however, are faced with an additional the dendritic tree. 9. Cash, S. & Yuste, R. Neuron 22, 383394
problem (not an issue for modern (1999).
1. Magee, J. C. & Cook, E. P. Nat. Neurosci. 3,
optical cables), which is that signals 895903 (2000). 10. Stuart, G., Spruston, N., Sakmann, B. &
are filtered as they propagate. The high 2. Bergano, N. S. Optics Photonics News 11,
Husser, M. Trends Neurosci. 20, 125131
density of hyperpolarization-activated (1997).
2030 (2000).
channels in CA1 dendrites provides an 3. Rall, W. in Handbook of Physiology: The
11. Linden, D. J. Neuron 22, 661666 (1999).
elegant solution to this problem by Nervous System. Cellular Biology on Neurons 12. Spruston, N., Stuart, G. & Husser, M. in
compensating for this distance-depen- Vol. 1 3997 (Am. Phys. Soc., Bethesda, Dendrites (eds. Stuart, G., Spruston, N. &
Maryland, 1977). Husser, M.) 231270 (Oxford Univ. Press,
dent filtering. These findings provide Oxford, 1999).
great insight into the process of synap- 4. Magee, J. C. in Dendrites (eds. Stuart, G.,
Spruston, N., Husser, M.) 139160 13. Magee, J. C. Nat. Neurosci. 2, 508514
tic integration and dramatically nar- (Oxford Univ. Press, Oxford, 1999). (1999).
row the focus of how we think about
5. Pettit, D. L. & Augustine, G. J. 14. Hansen, P. Laser Focus World 34, 7988
neural network function. As we seek to J. Neurophysiol. 84, 2838 (2000). (1998).
Evidence for this elusive CNTF II

CNTF II, I presume? was given a strong boost by a study
comparing the phenotypic conse-
Steven S. Lesser and Donald C. Lo quences of disrupting CNTF versus
CNTR receptor subunit (CNTFR)
Elson and colleagues report the identification of a new, genes4. Although CNTF/ mice appear
secreted ligand for the ciliary neurotrophic factor receptor, largely normal, CNTFR gene disrup-
which is likely to be important during development. tion is devastating, with deletions in
motor nuclei causing perinatal death
from the inability to suckle. In this issue,
Elson and colleagues report the charac-
The biological role of ciliary neu- at advanced ages 3 . These and other terization of what promises to be this
rotrophic factor (CNTF) remains large- observations suggested a rather restrict- long-sought second ligand for the CNTF
ly an enigma. It was first identified as a ed role for CNTF, perhaps as a trophic receptor5.
trophic factor in chick eye and nerve factor released after traumatic injury to The relationship between neural
extracts that supported the survival of peripheral nerves, and suggested that a cytokines like CNTF and hematopoiet-
ciliary ganglionic neurons in vitro; CNTF II must exist that normally sup- ic cytokines (typified by interleukin-6
mammalian CNTF was subsequently ports the important biological functions and granulocyte colony stimulating fac-
purified and cloned 1,2 . Intriguingly, of the CNTF signaling pathway1,2. tor6,7) helps to set the stage for under-
CNTF was also found to support the
survival of neurons in the CNS, includ-
ing clinically important dopaminergic
neurons in the substantia nigra as well
as motor neurons; these observations
generated hope that CNTF could be
used as a therapeutic agent in neurode-
generative diseases. The existenceand
importanceof another CNTF-like fac-
tor, however, was suggested because
CNTF itself is not a secreted protein
and, furthermore, because 2.3% of the
Japanese population is homozygous for
a null mutation of CNTF but neverthe-
less shows no neurological deficits even
The authors are in the Department of

Neurobiology, Box 3209, Duke University Fig. 1. Two principal transmembrane receptor subunits, gp130 and LIFR, form the core of a
Medical Center, Durham, North Carolina, broad range of receptor complexes for neural and hematopoietic cytokines. The addition of a
27710, USA. third subunit confers ligand selectivity to receptors for interleukin-6 (IL-6), CNTF and perhaps
email: lo@neuro.duke.edu or CT-1 (upper row), whereas an additional component that is part receptorpart ligand is required
ssless@neuro.duke.edu for factor secretion and receptor activation in the cases of IL-12 and CLCCLF (lower row).

news and views
standing the new ligand and its recep- ously, Elson and colleagues cloned a lated as NR6) results in a phenotype
tor. A key event for understanding the novel cytokine receptor with homology where mice are unable to suckle15, rem-
parallels between these signaling mole- to type-I cytokine receptors, which they iniscent of CNTFR knockouts. It will
cules was the discovery that cholinergic termed cytokine-like factor-1 (CLF)12. be interesting to determine if CLF and
differentiation factor (CDF) is identical Intriguingly, they found that CLF was a CNTFR gene disruption produce sim-
to leukemia inhibitory factor (LIF). secreted protein, and therefore predict- ilar underlying deficits in the motor
CDF was originally identified as a ed that it might act more like a ligand neuron populations in the brainstem
secreted activity that could induce phe- than a receptor. They proposed a func- and spinal cord. Finally, it will be
notypic switching in sympathetic neu- tional analogy to CNTFR, which is important to positively identify
rons, whereas LIF was identified as a known (upon its release from the plas- CLCCLF as CNTF IIand to exclude
factor inducing the terminal differenti- ma membrane by phospholipase cleav- the possibility of yet more ligands for
ation of myeloid leukemia cells6. Such age of its GPI linkage) to confer CNTF the CNTF receptorby demonstrating
functional parallels and overlaps responsiveness to cells that express only that a double knockout for CNTF and
between neural cytokines and the two transmembrane components of CLC/CLF encompasses the phenotype
hematopoietic factors arise from their the CNTF receptor. The ligand in this resulting from CNTFR gene disrup-
structural similarities at both the factor case was proposed to be formed by a tion.
and receptor levels. CNTF turned out to complex of CNTF together with the Perhaps the discovery of this new fac-
be a member of a broad family of soluble form of CNTFR13. tor complex will rekindle efforts to find
hematopoietic factors related in tertiary Based on this idea of a secreted therapeutic applications for CNTF
structure 7 , and its primary receptor, receptor bound to its ligand, Elson and receptor signaling in nervous system
CNTFR, is structurally related to the colleagues demonstrate that a complex diseases and disorders. Thus far,
subunit of the receptor for inter- of CLC and CLF forms a ligand for the attempts to use CNTF as a therapeutic
leukin-6 (ref. 8). Signal transduction CNTF receptor. First, they show that agent in the treatment of disorders such
studies showed that the receptor com- CLC is normally not released from cells, as amyotropic lateral sclerosis have not
plex for CNTF not only closely parallels but that coexpression of the receptor- met with success, although CNTF is now
that for interleukin-6, but also contains like CLF efficiently shepherds CLC being used in clinical trials again for one
two membrane-spanning subunits through the secretion process. They of its side effects, weight loss (Regeneron
(gp130 and LIFR) that are shared by a then show that CLC and CLF are secret- Pharmaceuticals, http://graphics.regen-
host of receptor complexes for ed as a stable complex, and that this het- eron.com/research/researchAxokine.htm).
hematopoietic cytokines 1 (Fig. 1). eromer binds to CNTFR. Critically, Interestingly enough, this biological
Whereas CNTFR confers selectivity for each of the subunits of the CNTF receptor action of CNTF probably arises from
CNTF upon the receptor complex, complexgp130, LIFR and CNTFR parallels in its signal transduction cas-
gp130 and LIFR transduce ligand is required to support CLCCLF bind- cade with yet another cytokine signaling
binding into cellular responses via acti- ing to the receptor complex and to system, that for leptin, but that is the
vation of the JAK-STAT signaling path- activate downstream signal transduction subject for a completely different News
way 1,2 . CNTFR itself is not a events, including phosphorylation of and Views
transmembrane protein, but is anchored STAT3. Finally, and importantly, they
1. Ip, N. Y. & Yancopoulos, G. D. Annu. Rev.
extracellularly to the plasma membrane show that CLCCLF competes against Neurosci. 19, 491515 (1996).
via a glycosylphosphatidylinositol (GPI) CNTF for binding to the CNTF recep-
2. Segal, R. A. & Greenberg, M. E. Annu. Rev.
linkage. Without CNTFR, the remain- tor and can promote the survival of Neurosci. 19, 463489 (1996).
ing two subunits constitute a receptor embryonic motor neurons in vitro, a 3. Takahashi, R. et al. Nat. Genet. 7, 7984
for LIF, whereas two gp130 subunits hallmark of the biological activity of (1994).
plus the interleukin-6 receptor form CNTF. 4. DeChiara, T. M. et al. Cell 83, 313322
the functional receptor complex for The final ligandreceptor complex, (1995).
interleukin-6. This scheme of mixing containing the CLCCLF heteromeric 5. Elson, G. C. A. et al. Nat. Neurosci. 3,
and matching various and subunits ligand and the CNTFR, gp130 and 867872 (2000).
produces a range of receptor complex- LIFR receptor complex, bears striking 6. Patterson, P. H. & Nawa, H. Neuron 10
es from a relatively limited set of recep- similarity to the interleukin-12 factor (Suppl.), 123137 (1993).
tor components (Fig. 1). Because the and receptor complex, a configuration 7. Bazan, J. F. Neuron 7, 197208 (1991).
transmembrane subunits are common that previously had no known parallel 8. Davis, S. et al. Science 253, 5963 (1991).
to many receptor complexes, the down- in a neurobiological context. In the case 9. Shi, Y. et al. Biochem. Biophys. Res. Comm.
stream effects of receptor activation by of interleukin-12, a hematopoetic fac- 262, 132138 (1999).
different ligands are often similar if not tor involved in the differentiation and 10. Pennica, D. et al. Proc. Natl. Acad. Sci. USA
identical1,2. regulation of T-helper cells, the p35 and 92, 11421146 (1995).
The work reported by Elson and col- p40 components of the ligand complex 11. Senaldi, G. et al. Proc. Natl. Acad. Sci. USA
96, 1145811463 (1999).
leagues 5 in this issue stems from the also require coexpression for secretion,
12. Elson, G. C. et al. J. Immunol. 161,
identification last year of a new mem- and are further stabilized as a het- 13711379 (1998).
ber of the IL-6 class of cytokines, which erodimer by a disulfide bond14 (Fig. 1).
13. Davis, S. et al. Science 259, 17361739
was termed cardiotrophin-like cytokine These elegant experiments show that (1993).
(CLC) based on its homology to car- CLCCLF is an excellent candidate for 14. Gately, M. K. et al. Annu. Rev. Immun. 16,
diotrophin-1, itself a relative of CNTF9,10; the elusive CNTF II. Indeed, an earlier 495521 (1998).
CLC was independently identified as study has already shown that genetic 15. Alexander, W. S. et al. Curr. Biol. 9, 605608
novel neurotrophin-1 (NNT-1)11. Previ- disruption of CLF (independently iso- (1999).

news and views
throughout the animal kingdom7 sug-

Probing the olfactory code gested that the identity of an odor may
also be coded in the synchronous timing
C. Giovanni Galizia and Randolf Menzel of action potentials in different neurons8.
These oscillations arise from the activity
Multielectrode recording arrays in the moth antennal lobe of many neurons, and are driven by
indicate that the relative timing of action potentials may local neurons in the antennal lobe.
convey information about odor concentration and mixture. Intracellular recordings from pairs of
olfactory output neurons show stimulus-
characteristic patterns of synchronized
activity, which are in tune with oscilla-
The olfactory system has a particularly defined by the complement of olfactory tions of a local field potential; single
difficult task: unlike vision or sound, the receptors that they express. In most cases, spikes ride on top of the oscillations.
number of elementary stimulithe a receptor cell only expresses a single However, spike timing could also code for
odorsseems to be almost infinite. Fur- olfactory receptor gene; each class is selec- odor concentration or for a particular
thermore, airborne odors come in plumes tive for a number of chemical substances blend1, and the identity of the odor could
that reach the olfactory organ intermit- (odors), and the receptors respond in a be coded in the identity of the activated
tently and unpredictably. A single plume dose-dependent manner. These proper- glomeruli9. To understand the relation-
may be very short, and successive plumes ties define the cells molecular receptive ship between the spatial and the temporal
may come at a high frequency or after range, which may overlap for different properties of output neuron responses,
long intervals. The odor concentration in receptor cells. Consequently, each odor we need simultaneous recordings from
different plumes is highly variable and elicits a pattern of activated receptor cell many neurons within the antennal lobe
contains only scarce information about axons in the antennal lobe or olfactory or olfactory bulb. The identity of the
the odor source. To complicate things, bulb. In all cases studied so far, glomeruli glomeruli that they innervate must be
behaviorally relevant odors are often receive inputs from a single or at most a known, and different concentrations as
blends of several substances, for which the few classes of receptor cells24. What is the well as odor mixtures need to be tested.
composition and relative proportion of nature of the output from these struc- In pioneering work, Christensen et
the blends components are important. tures, and how are odor concentrations, al.1 now lay the foundations for such
Odors must also be recognized against a specific blends and the temporal struc- detailed studies. They took advantage of
background of confounding odors, which ture of the odor stimuli encoded? the extensively studied sexual-
are always present in the natural environ- Within the antennal lobe or olfactory pheromone system in the moth
ment, and context variables (such as time bulb, the chemical identity of an odor Manduca sexta, and the technique of
of day, developmental stage and motiva- (odor quality) is represented by patterns multiunit extracellular recording, which
tional conditions) may change the impor- of active glomeruli and their output neu- has been developed for the mammalian
tance and meaning of an odor. It is not rons, providing a spatial, combinatorial brain10,11. The application of multielec-
surprising, then, that both neuroscientists neural code5,6. The discovery of odor- trode techniques to the insect brain and
and behavioral zoologists are greatly inter- induced oscillations in olfactory systems the olfactory system represents a new
ested in how the brain processes olfactory
information and ultimately identifies a
relevant odor. However, this olfactory to other brain regions
code is far from being understood. What
we know for sure is that many neurons are
involved, and that to understand how
odor information is encoded, we have to
understand the relationship between var-
ious levels of spatial and temporal infor-
mation. In this issue, Christensen et al.1 PN PN PN
take a big step forward by simultaneously
from other IN IN IN IN IN IN
measuring responses from several neurons glomeruli to other
in the olfactory system of a living insect. IN IN IN
glomeruli
The first brain structure to process
olfactory information is the olfactory
bulb in vertebrates, or its functional and
structural analogue in insects, the anten-
nal lobe (Fig. 1). The receptor cells that
project to the antennal lobe come in dif- axons from receptor cells from the antenna
ferent functional classes, which are
Fig. 1. A schematic diagram of the insect antennal lobe (analogous to the olfactory bulb in verte-
The authors are at the Institut fr Biologie, brates). Axons from olfactory receptor cells enter the structure and branch within small balls of
Neurobiologie Freie Universitt Berlin, Knigin- axons and dendrites called olfactory glomeruli, which line the outside of the antennal lobe. A
Luise Str. 28-30, D-14195 Berlin, Germany. selection of the numerous connections within and between the glomeruli are illustrated (IN, local
email: galizia@zedat.fu-berlin.de or interneurons). Projection neurons (PN, analogous to mitral/tufted cells in vertebrates) relay the
menzel@neurobiologie.fu-berlin.de information to higher brain areas.

news and views
analyzed the temporal coinci- oscillatory circuit, but may have no rele-
dence of the action potentials vance for the olfactory code.
Odor A
recorded across the different The highly specialized pheromone-
electrodes, revealing two inter- sensitive macroglomerular complex,
esting novelties. First, no oscilla- designed to optimize speed and sensitiv-
tory responses were apparent. ity, may thus use different strategies than
Second, synchronicity between the ordinary olfactory glomeruli, in
units was not a function of the encoding odors without the need for
odor, but rather of the odor con- oscillations. This finding, therefore, does
Odor A at high concentration
centration and the specific not contradict the hypothesis that oscil-
blend (Fig. 2). Consequently, lations may be useful in improving odor
spatial and temporal patterns identification for general odors. Indeed,
may encode different aspects of when the GABAergic neurons responsi-
an olfactory stimulus. Odor ble for synchronization are pharmaco-
quality may be encoded in the logically blocked, honeybees lose their
identity of the active output ability to distinguish between similar
Odor A + B fibers (thus in the spatial com- odors15. This exemplifies the idea that,
ponent of the odor code), despite the architectural similarity of dif-
whereas stimulus properties ferent olfactory systems, results from

such as concentration and blend one model cannot be directly extrapo-
composition could be encoded lated to other systems. The beauty of the
in the relative timing of output new technique also lies in its suitability
neuron action potentials. With for investigating ordinary glomeruli as
the new approach reported well as glomeruli devoted to processing
Fig. 2. The relative timing of action potentials may convey here1, a thorough investigation sexual pheromones, in both insects and
information about odor concentration and mixture.
of these hypotheses seems to be mammals alike. This will tremendously
Christensen and colleagues presented a moth (Manduca
sexta) with a pulse of pheromone (odor A, E10,Z12-hexa- within reach. Because of our improve our understanding of glomeru-
decadienal) at two different concentrations, as well as a considerable knowledge about lar organization. Multiunit recordings
mixture of two pheromones (odor A + odor B, the insect olfactory system, it are now joining intracellular recordings
E10,E12,Z14-hexadecadienal), while recording simultane- will be possible to apply this and optical imaging in the quest to deci-
ously from multiple neurons in the antenna lobe (five traces method to neurons for which pher the olfactory code.
shown here; data are schematic). Coincident spikes are the innervated glomeruli are
labeled in red. Note that in the three different conditions, identified. Furthermore, multi- 1. Christensen, T. A., Pawlowski, V. M., Lei, H. &
different pairs of neurons share coincident spiking events. neuron recordings may be Hildebrand, J. G. Nat. Neurosci. 3, 927931
(2000).
applied to animals performing
learning tasks. Perhaps the elec- 2. Hansson, B. S., Ljungberg, H., Hallberg, E. &
Lfstedt, C. Science 256, 13131315 (1992).
trodes could even be miniatur-
3. Mombaerts, P. Curr. Opin. Neurobiol. 6,
and exciting development. Because the ized and chronically implanted in a 481486 (1996).
single electrodes in the array are 150 m free-moving and behaving insect. This
4. Gao, Q., Yuan, B. & Chess, A. Nat. Neurosci. 3,
apart, they record activity from different would lead the way to investigating the 780785 (2000).
glomeruli, and thus permit analysis of olfactory code under natural conditions. 5. Galizia, C. G., Sachse, S., Rappert, A. &
the contribution of different glomeruli Although some of the reported find- Menzel, R. Nat. Neurosci. 2, 473478 (1999).
to the odor response. Furthermore, the ings seem to contradict previous studies, 6. Mori, K., Nagao, H. & Yoshihara, Y. Science
electrode tip design allows for isolation these discrepancies are more likely a 286, 711715 (1999).
of multiple neurons within a given measure of our ignorance of the olfactory 7. Gelperin, A. J. Exp. Biol. 202, 18551864
glomerulus as well. system. Christensen et al. did not find (1999).
The study concentrated on the oscillations when they stimulated with 8. Laurent, G. Curr. Opin. Neurobiol. 7, 547553
prominent macroglomerular complex short odor pulses1. In the sexual (1997).
(MGC) in the antennal lobe of the male pheromone system, single odor plumes 9. Laurent, G. Science 286, 723728 (1999).
moth, which is devoted to processing may be very short, because the continous 10. Nicolelis, M. A., Ghazanfar, A. A., Faggin,
sexual pheromone information. The filament is easily broken down into small B. M., Votaw, S. & Oliveira, L. M. Neuron 18,
529537 (1997).
female moth attracts the male over long plumes by air turbulences. In some
11. Sutherland, G. R. & McNaughton, B. Curr.
distances by releasing a pheromone species, females even release the Opin. Neurobiol. 10, 180186 (2000).
blend, and the MGC contains highly pheromones as individual pulses. This 12. Hansson, B. S. & Christensen, T. A. in Insect
specialized glomeruli that are selective would make it necessary for the olfactory Olfaction (ed. Hansson, B. S.) 125161
for individual substances of the system to identify an odor without the (Springer, Berlin, 1999).
blend12,13. Christensen and colleagues1 need for oscillatory activity. In the same 13. Heinbockel, T., Christensen, T. A. &
now confirm that individual antennal system, however, oscillations appear with Hildebrand, J. G. J. Comp. Neurol. 409, 112
(1999).
lobe neurons have differing response longer odor stimuli14. In these instances,
profiles in that they responded differ- oscillations may arise as a consequence of 14. Heinbockel, T., Kloppenburg, P. &
Hildebrand, J. G. J. Comp. Physiol. A 182,
ently to different odors and to the blend. the wiring within the antennal lobe, 703714 (1998).
By simultaneously recording up to seven because several recurrent excitatory and 15. Stopfer, M., Bhagavan, S., Smith, B. H. &
neurons in each animal, the authors also inhibitory connections easily form an Laurent, G. Nature 390, 7074 (1997).

news and views
into one of nine categories: drawings or

Coding for visual categories photographs of famous people, emotion-
al expressions on unknown actors, house-
in the human brain hold objects, cars, animals, food, spatial
layout (scenes) and abstract patterns (Fig.
Charles G. Gross 1). Several examples from each category
were shown several times during the
Single-neuron recordings in the human hippocampus, entorhinal recording session with each neuron. To
insure that the stimuli engaged the
cortex and amygdala demonstrate that cells in these areas can
patients attention, they had to indicate
respond selectively to particular categories of visual stimuli. whether the picture was a face or not by
pressing a button. This type of experiment
is more restricted in humans than in
For the last 50 years, the microelectrode be specialized for representing visual stim- monkeys because the electrodes cannot be
has been one of the most powerful tools uli or visual categories, and are more com- moved around to sample different neu-
for revealing the neuronal bases of per- monly associated with memory. rons, and the time available to study the
ception and cognition. Electrophysiology The subjects were patients with epilep- activity of each is severely limited.
has identified individual neurons that sy who were resistant to drug treatment. Kreiman and colleagues4 found that
respond selectively to highly complex and To find an epileptogenic focus that could 18% of hippocampal, 16% of entorhinal
abstract visual stimuli. The first discovered be surgically excised, the surgeons and 9% of amygdala neurons responded
and most intensively studied are the face- implanted patients with intracranial depth selectively to one or, more rarely, to two of
selective neurons in inferior temporal (IT) electrodes for several weeks. The elec- the visual categories. For example, some
cortex of the macaque monkey 1. Face- trodes were in portions of the medial tem- neurons fired more in response to pictures
selective cells, which are a small propor- poral lobe, including the hippocampus, of animals than to stimuli in any other cat-
tion of IT cells, respond best or only to the entorhinal cortex and amygdala. (A pos- egory. However, no particular animal
sight of faces. Although virtually all faces sible problem with such studies is that the elicited more firing than any other animal.
activate these cells, different faces produce recordings are made from diseased brains, Other neurons fired more in response to
varying response magnitudes. Thus, the and indeed quite near known sites of mal- both photographs and drawings of famous
response of one such cell cannot represent function.) While action potentials were faces but not to any other categories; again,
a specific face, but the pattern of firing recorded from single neurons, patients there were no differences in responses to
across a set of face-selective cells would be viewed complex visual stimuli that fell different face stimuli. Thus, these neurons
unique for a particular face. As these cells
usually do not code for individual faces,
they are best described as face-category
cells. The brain may represent individual
faces by the pattern of activity over a pop-
ulation of such face-category cells, in what
is termed population or coarse coding2,3.
Almost all fundamental questions in
neuroscience derive from human experi-
ence, and yet, for ethical and practical rea-
sons, they can usually be investigated only
in nonhuman animals. Thus, despite our
Darwinian faith in the continuity of
species, it is always a relief to confirm in
humans a basic finding from other ani-
mals. In a bold and imaginative study in
this issue, Kreiman and colleagues4 con-
firm and extend results from monkeys.
They report single neurons selective for
faces and other visual categories in the
medial temporal lobe of humans. This
study is interesting and important for two
reasons. First, the authors found neurons
selective for a variety of visual categories
in addition to faces. Second, these neurons
were found in areas of the temporal lobe,
which unlike IT cortex are not known to
The author is at the Department of Psychology,

Princeton University, Green Hall, Princeton,
New Jersey 08544, USA. Fig. 1. Examples from three of the nine categories of visual stimuli used by Kreiman and col-
e-mail: cggross@princeton.edu leagues4. Top row, animals; middle row, food; bottom row, spatial layout (scenes).

news and views
may be described as category selective. pus (by way of entorhinal cortex), and the is an appropriately modest one. Howev-
Neurons in human hippocampus and hippocampus is thought to be involved in er, describing the stimulus or even the cat-
amygdala are known to be selective for consolidation of short-term memories into egory selectivity of temporal neurons is
faces, objects or letters5,6, but the current long-term ones. Thus, monkey studies on still very far from understanding how the
results 4 more extensively demonstrate hippocampal and entorhinal cells have circuits they compose underlie the com-
selectivity for visual categories. almost exclusively been concerned with plexity and subtlety of perception and the
The existence of category-selective cells questions of supra-modal recognition mysteries and vagaries of memory.
fits nicely with other studies of human memory, short-term memory and similar
brain function. After brain damage, par- mnemonic matters, rather than selectivity 1. Gross, C. G. Cereb. Cortex 5, 455469 (1994).
ticularly to the temporal lobe, humans to visual stimuli and their categorization. 2. Desimone, R. J. Cogn. Neurosci. 3, 18 (1991).
may have selective deficits in visually rec- What are the implications of Kreiman 3. Gross, C. G. Phil. Trans. R. Soc. Lond. B Biol.
ognizing specific categories of objects. For and colleagues demonstration4 of cate- Sci. 335, 310 (1992).
example, in spite of otherwise normal per- gory-selective cells in medial temporal 4. Kreiman, G., Koch, C. & Fried, I.
ception, patients may have selective diffi- areas? One is that the hippocampus car- Nat. Neurosci. 3, 946953 (2000).
culty in recognizing only faces, only ries more than just relational or spatial 5. Heit, G., Smith, M. E. & Halgren, E. Nature
333, 773775 (1988).
animals or only man-made objects7. Sim- information (although the hippocampus
6. Fried, I., MacDonald, K. A. & Wilson, C. L.
ilarly, PET and fMRI reveal localized had the largest proportion of cells selec- Neuron 18, 753765 (1997).
regions of the temporal lobe that are dif- tive for the category of spatial scenes). A
7. Warrington, E. K. & McCarthy, R. A. Brain

ferentially active in response to specific related implication is that the hippocam- 110, 12731296 (1987).
categories of visual stimuli such as faces, pus has more than just a modulation or 8. Martin, A., Ungerleider, L. G. & Haxby, J. V. in
words, houses or chairs8. consolidation effect on cortex; instead, it The New Cognitive Neurosciences (ed.
Kreiman and colleagues found that carries complex visual information. Gazzaniga, M. S.) 10231036 (MIT Press,
Cambridge, Massachusetts, 2000).
hippocampal cells were more selective for The evidence that medial temporal cells
spatial scenes than for any other visual are selective for visual categories implies 9. OKeefe, J. & Nadel, L. The Hippocampus as a
Cognitive Map (Oxford Univ. Press, 1978).
categories, which was not the case for that these cells may be involved in visual
10. Maguire, E. A., Frackowiak, R. S. J. &
either amygdala or entorhinal neurons. categorization. In addition, and perhaps Frith, C. D. J. Neurosci. 17, 71037110 (1997).
This is consistent with long-standing evi- more parsimoniously, the category-selec-
11. Gross, C. G., Rocha-Miranda, C. E. &
dence for specialization of the rodent hip- tive cells may be members of an ensemble Bender, D. B. J. Neurophysiol. 35, 96111
pocampus for spatial processing9, which that represents individual category mem- (1972).
seems to be true in humans as well10. bers by population or sparse coding. For 12. Hasselmo, M. E., Rolls, E. T. & Baylis, G. C.
In monkeys, neurons in IT cortex are example, the cross-fiber pattern of firing of Behav. Brain Res. 32, 203218 (1989).
selective not only for faces, but also for the ensemble of cells selective for animals 13. Logothetis, N. K. & Sheinberg, D. L. Annu.
other natural categories of visual stimuli, may be the code for individual animals, just Rev. Neurosci. 19, 577621 (1996).
such as hands11 and facial expressions12. as face cells seem to form ensembles for 14. Fantz, R. in Behavior of Non-Human Primates
Vol. II (eds. Schrier, A. M., Harlow, H. F. &
Other studies show selectivity for arbitrary encoding individual faces. Stollnitz, F.) 365403 (Academic, New York,
visual categories after explicit training to The authors conclusion that their 1965).
distinguish those categories. For example, results may be relevant in the represen- 15. Rodman, H. R., Scalaidhe, S. P. & Gross. C. G.
Logothetis and colleagues showed that tation and retrieval of visual information J. Neurophysiol. 70, 11151136 (1993).
monkeys trained to discriminate wire fig-
ures had IT neurons selective for such fig-
ures but not for spheroidal objects, and The mechanics behind spines on the move
the converse was true for monkeys trained
on spheroidal objects 13. Thus neurons Dendritic spinesthe main sites of excitatory synaptic contacts in the CNShave been
selective for visual categories are found in in the spotlight recently. Their curious motility has been linked to synapse formation as
both humans and monkeys, for natural well as plasticity in response to sensory experience. Now, Matus and colleagues (pages
and arbitrary categories. Although most 887894, this issue) have directly addressed the mechanism by which neural activity
of these categories are presumably and spine motility may be linked. The authors made time-lapse videos of spine motility
learned, some, such as selectivity for faces in GFP-actin transfected hippocampal neurons as well as slice cultures from transgenic
and facial expression, may well be present mice expressing GFP-tagged actin. A major finding was that actin dynamics are rapidly
at birth in both humans and monkeys14,15. and reversibly inhibited following activation of AMPA receptorsspines became more
One difference between human and stable and assumed a more regular appearance. Furthermore, inhibition of motility via
monkey studies of category-selective neu- AMPA receptors required postsynaptic membrane depolarization and the influx of
rons is in the recording sites. Virtually all calcium. In combination with previous work, the results suggest that spines initially
systematic analyses of stimulus selectivity formed by NMDA receptor activation are subsequently stabilized by AMPA receptors.
in temporal lobe neurons in the monkey Although the results may seem paradoxical because the quite different processes of
has been on IT neurons rather than on spine formation and stabilization both require the influx of calcium, the authors point
hippocampal, entorhinal or amygdala cells. out that there are interesting parallels with growth cone motility, where calcium
Perhaps this is because IT cortex is activation at different stages of synapse
assumed to be the last exclusively visual formation can have opposite effects.
processing station in the ventral cortical
pathway, which analyzes information about John E. Spiro
visual identity. IT cortex is believed to send 10m
control +AMPA
the result of its analysis to the hippocam-

Fast synaptic fatigue in (Fig. 1c), several orders of magnitude faster than minimum esti-
mates of synaptic vesicle recycling times7.
shibire mutants reveals a The above results suggest a requirement for dynamin in addi-
tion to its role in generating recycled synaptic vesicles. Howev-
rapid requirement for er, an alternative possibility is that the presynaptic terminal was
depleted of reserve vesicles before the stimulus train. Although
dynamin in synaptic vesicle we considered this unlikely on the basis of previous work (see
below), we addressed this issue directly by ultrastructural analy-
membrane trafficking sis of neuromuscular synapses of the intracoxal lateral levator
muscle (ICLM). This preparation is well-suited to ultrastructural
studies of synaptic vesicle trafficking, and electrophysiological
Fumiko Kawasaki, Missy Hazen and Richard W. Ordway
recordings confirmed that shi mutant ICLM synapses show the
fast synaptic fatigue phenotype described above (Fig. 2a).
Department of Biology, 208 Mueller Laboratory, Penn State University,
University Park, Pennsylvania, 16802, USA Initial ultrastructural analysis was done on preparations in
Correspondence should be addressed to R.W.O. (rwo4@psu.edu)
which the ICLM motor axons were not cut, as in the electro-
physiology experiments, but rather left intact to allow endoge-
The GTPase dynamin is involved in endocytosis in many cell nous CNS activity to drive synaptic transmission. Similar
types1, as first revealed by temperature-sensitive paralytic muta- experiments in other preparations2,3 show a severe depletion of
tions in the Drosophila dynamin gene, shibire (shi), which dis- synaptic vesicles at restrictive temperatures. This prototypical shi
rupt synaptic vesicle endocytosis and deplete synaptic terminals phenotype was also observed at ICLM synapses, which were near-
of vesicles26. Here we report that shi synapses exhibit a fast synap- ly devoid of synaptic vesicles at 33C (Fig. 2b). Consistent with
tic fatigue phenotype within 20 ms of repetitive stimulation, earlier findings3, removing endogenous neural activity by cut-
which cannot be explained by vesicle depletion, as we confirmed ting the ICLM motor axons prevented synaptic vesicle depletion
by electron microscopy. These results suggest that, in addition to in shiTS1 at 33C (Fig. 2c). The number of free synaptic vesicles
its well-characterized role in synaptic vesicle recycling, dynamin within 200 nm of the presynaptic dense body was 23.9 1.7
may be required for short-term maintenance of the readily (n = 13) in wild type and 21.1 1.1 (n = 15) in shiTS1. Thus rapid
releasable pool of synaptic vesicles. onset of the shi phenotype is not caused by release from a deplet-
A newly formed synaptic vesicle requires 1530 seconds to enter ed synaptic vesicle pool.
the releasable pool7. If dynamin acts exclusively in recycling vesicles, Taken together, these observations indicate that shi mutant
then onset of the shi synaptic phenotype during repetitive stimu- synapses with a full complement of synaptic vesicles show a fast
lation should require a similar time period. However, previous synaptic fatigue phenotype within 20 ms of repetitive stimula-
observations in shi suggest that this may not be the case4. To further tion. Consistent with previous work, the first stimulus produced
address this issue, we used voltage-clamp analysis of synaptic cur- a wild-type synaptic current in shi (Figs. 1a and 2a), indicating
rents to compare the time and activity dependence of wild-type that calcium signaling, vesicle fusion and postsynaptic events are
and shi synapses at a restrictive (paralytic) temperature for shi. At normal before vesicle depletion. In contrast, the rapid and strict-
wild-type dorsal longitudinal muscle (DLM) neuromuscular ly activity-dependent reduction of the synaptic current in shi
synapses, 1 Hz stimulation produced constant-amplitude synaptic suggests a defect in maintaining the releasable pool of synaptic
currents (Fig. 1a). In contrast, shiTS1 exhibited a clear activity- vesicles. Refilling the releasable pool involves mobilization and
dependent reduction in synaptic current under the same condi- priming of vesicles over several seconds8; however, faster, calci-
tions (Fig. 1a). Consistent with the conditional
nature of this mutant, stimulation of shi TS1 a b c
synapses at 20C (1 or 50 Hz) produced wild-
type synaptic currents (data not shown).
To further investigate the time required for
onset of the shi phenotype, we stimulated wild-
type and shi synapses at a higher frequency of
50 Hz. Under these conditions, wild-type
synapses are fatigued, showing activity-depen-
dent reduction in synaptic current amplitude
(Fig. 1a and b). The shi synapses showed a
more rapid and pronounced activity-depen-
dent reduction (Fig. 1a and b), with a
68 1.9% (n = 5) decrease within the first Fig. 1. Fast synaptic fatigue in shi. (a) Two-electrode voltage-clamp recordings of DLM synap-
200 ms of stimulation, in contrast to a tic currents in wild type (WT) and shi at 33C. Synaptic currents were evoked by 1 Hz or 50 Hz
28 2.0% (n = 4) reduction in wild type stimulation of the cut DLM motor axon. The first ten responses and representative traces
(Fig. 1a and b). Similar results were obtained throughout the remainder of the first 100 responses are superimposed. Arrows indicate stim-
in shiTS2 mutants (data not shown), indicating ulation of the motor axons; stimulation artifacts were removed. Holding potential, 80 mV.
that the observed phenotype is a general con- (b) Peak amplitude measurements of DLM excitatory postsynaptic currents (EPSCs) are plot-
ted as a function of time for 50-Hz stimulation at 33C. The zero time point is the beginning of
sequence of disrupting dynamin function. Sur-
the stimulus train. (c) Expanded view of the initial period of 50-Hz stimulation in (b). Asterisks,
prisingly, a clear difference between wild-type statistically significant differences from wild type. In shi, but not wild type, a slower component
and shi synapses was observed in the second of activity-dependent reduction reduced or eliminated the remaining synaptic current over
response to a 50-Hz stimulus train. Thus onset 5 seconds of 50-Hz stimulation. Wild-type flies were Canton S. DLM synaptic currents were
of the shi phenotype occurs within 20 ms recorded and analyzed as described14.

a um-dependent refilling is reported as well9. One possibility is

that fast refilling is occluded in shi by accumulation of endocyt-
ic intermediates at release sites. A role for dynamin in rapid clear-
ance of these intermediates during repetitive stimulation would
complement fast, calcium-dependent refilling and contribute to
short-term maintenance of the releasable pool. Relevant findings
in adrenal chromaffin cells indicate that retrieval of vesicle mem-
branes after intense stimulation includes a calcium- and
dynamin-dependent component with fast kinetics10,11. This may
b reflect a kiss-and-run mechanism12, in which exocytosis occurs
without collapse of the vesicle into the plasma membrane. Addi-
tional work will address the mechanism of fast synaptic fatigue
in shi in the context of the above working model.
Finally, dynamin activity is regulated by phosphorylation as
well as by interactions with its binding partners1,13, and these
mechanisms are proposed to regulate synaptic function. The pre-
sent finding that shi synapses exhibit rapid synaptic fatigue sup-
ports a role for dynamin in short-term synaptic plasticity.
ACKNOWLEDGEMENTS
c Electron microscopy was done in the Penn State Electron Microscopy facility. The
shiTS2 mutant was provided by the Bloomington Stock Center. Supported by the
National Science Foundation.
RECEIVED 25 APRIL; ACCEPTED 28 JULY 2000
1. Schmid, S. L., McNiven, M. A. & De Camilli, P. Curr. Opin. Cell Biol. 10,
504512 (1998).
2. Kosaka, T. & Ikeda, K. J. Neurobiol. 14, 207225 (1983).
3. Poodry, C. A. & Edgar, L. J. Cell Biol. 81, 520527 (1979).
Fig. 2. ICLM synapses show a shi synaptic phenotype, but are not 4. Salkoff, L. & Kelly, L. Nature 273, 156158 (1978).
depleted of vesicles before stimulation. (a) Single-electrode voltage- 5. Costello, W. J. & Salkoff, L. B. J. Neurosci. 6, 36343639 (1986).
clamp recordings at ICLM synapses from wild-type and shiTS1 flies, dis- 6. Ramaswami, M., Krishnan, K. S. & Kelly, R. B. Neuron 13, 363375 (1994).
played as in Fig. 1a. Synaptic currents were elicited by 50-Hz stimulation 7. Betz, W. J. & Wu, L.-G. Curr. Biol. 5, 10981101 (1995).
of the cut ICLM motor axon. Holding potential, 50 mV. Similar results 8. von Gersdorff, H. & Matthews, G. Annu. Rev. Physiol. 61, 725752 (1999).
9. Wang, L.-Y. & Kaczmarek, L. K. Nature 394, 384388 (1998).
were obtained in six WT and four shiTS1 experiments. (b) Transmission 10. Artalejo, C. R., Henley, J. R., McNiven, M. A. & Palfrey, H. C. Proc. Natl. Acad.
electron microscopy (TEM) image from a shiTS1 preparation in which the Sci. USA 92, 83288332 (1995).
ICLM motor axons were left intact (uncut). Arrows point to a few 11. Smith, C. & Neher, E. J. Cell Biol. 139, 885894 (1997).
remaining synaptic vesicles. Arrowheads indicate endocytic intermedi- 12. Als, E. et al. Nat. Cell Biol. 1, 4044 (1999).
13. Robinson, P. J., Liu, J.-P., Powell, K. A., Fykse, E. M. & Sdhof, T. C. Trends
ate structures as defined previously2. (c) TEM images from WT and Neurosci. 17, 348353 (1994).
shiTS1 preparations in which the ICLM motor axons were cut. ICLM 14. Kawasaki, F., Mattiuz, A. M. & Ordway, R. W. J. Neurosci. 18, 1024110249
synaptic currents were recorded as described15. TEM was done using (1998).
conventional methods, essentially as described14. 15. Kawasaki, F., Felling, R. & Ordway, R. W. J. Neurosci. 20, 48854889 (2000).
Motor timing learned specific perceptual learning, we found that training on a percep-
tual task showed significant transfer to a motor task. This result
without motor training provides evidence for a common neural architecture underlying
analysis of sensory input and control of motor output, and sug-
gests a potential role for perception in motor development and
Daniel V. Meegan1,2, Richard N. Aslin1 and rehabilitation.
Robert A. Jacobs1 Both tasks required processing of time. The perception task
was to discriminate temporal intervals denoted by brief auditory
1 Department of Brain and Cognitive Sciences, Meliora Hall, University of stimuli, and the motor task was to produce successive finger
Rochester, Rochester, New York 14627, USA
movements separated by a target temporal interval. Evidence sug-
2 Department of Psychology, University of Guelph, Guelph, Ontario, N1G 2W1, gests a common neural substrate for time perception and motor
Canada
timing2. Thus we tested the possibility that plastic modifications
Correspondence should be addressed to D.V.M. (meegan@psy.uoguelph.ca) of such a substrate, induced by perceptual training, could affect
motor timing.
Improvements due to perceptual training are often specific to the Twelve right-handed adults participated in seven experimen-
trained task and do not generalize to similar perceptual tasks1. tal sessions, with no more than two days between successive ses-
Surprisingly, given this history of highly constrained, context- sions. During the training period (sessions 26), each participant

a b
Percent improvement in mean motor timing accuracy

were the result of perceptual training, then it
would be present only for the perceptually
Percent reduction in motor timing variability
trained interval. This prediction was sup-
(pre-training to post-training)
ported by a significant interaction between
(pre-training to post-training)
trained perceptual interval and target motor

interval (Fig. 1a; F1,10 = 24.57, p < 0.001).
Specifically, participants trained at 300 ms
showed greater reduction of motor timing
variability in the 300-ms task than the
500-ms task (t5 = 2.69, p < 0.05), whereas par-
ticipants trained at 500 ms showed greater
reduction of motor timing variability in the
500-ms task than the 300-ms task (t5 = 5.36,
Standard perceptual interval Standard perceptual interval
(training) (training) p < 0.005). In other words, participants
Fig. 1. Change in motor performance as a function of perceptual training condition. The motor showed more motor improvement when the
tasks involved attempting to produce successive thumb presses separated by a target interval of temporal requirements of the motor task
time. Motor timing variability (a) was measured by the standard deviation of the interpress inter- matched the temporal characteristics of their
val, and mean motor timing accuracy (b) was measured by the difference between mean IPI and perceptual training.
the target IPI; note that mean IPI can correspond well to target IPI even when performance is We suggest that this motor learning was a
highly variable (as in the pre-training session). Participants showed a greater reduction of vari-
byproduct of an enhanced representation of
ability (a) when the target interval matched the interval used in an auditory discrimination task
during the training phase. Consistent with the measure of variability (a), participants showed a
a particular temporal interval, induced by
tendency toward greater improvement in mean accuracy (b) when the target interval matched auditory training, in a plastic network shared
the trained interval, although the interaction between perceptual and motor interval was not by sensory and motor systems. Alternatively,
significant (F1, 10 = 2.66). The lack of a significant effect was not surprising because mean IPI was one might suggest that unintended motor
close to the target IPI in the pre-training session; in other words, there was little room for training occurred during the training period,
improvement in mean accuracy. or that auditory feedback aided motor per-
formance in the post-training session. Some
methodological constraints were implement-
ed in anticipation of such concerns. For
made 2500 discrimination judgments. On each trial, subjects example, in auditory training, the onset of tones was unpre-
indicated which of two successive temporal intervals was per- dictable, eliminating the possibility that subjects used rhythmic
ceived to be longer in duration. The beginning and end of each bodily movement to aid in auditory discrimination. Also, par-
interval were marked by auditory tones of constant duration ticipants listened to white noise while performing the motor
(25 ms), frequency (1 kHz) and amplitude. The shorter of the tasks, thereby eliminating any auditory feedback from the move-
two intervals (the standard interval) was 300 ms on every trial ment (such as button depression). Despite the white noise, par-
for six of the participants, and 500 ms on every trial for the other ticipants possibly could have deliberately used auditory memory
six participants. The duration of the longer interval (the com- during the post-training session to aid motor timing at the
parison interval) was determined by a weighted updown trained interval. However, during debriefing after the experi-
method3, which estimated the temporal threshold at which the ment, they were surprised to learn of the temporal relationship
standard and comparison intervals could be discriminated with between their training task and one of the motor tasks.
75% accuracy. Consistent with a previous report4, participants Behavioral evidence for a common sensory and motor timer
had lower thresholds after the training period than before; an has been indirect6,7. For example, performance declines similar-
ANOVA with trained interval (300 or 500ms) and session (pre- ly at increasing temporal intervals for perception and motor
training and post-training) as factors yielded only a main effect of tasks6. Such correlational evidence has left open the possibility
session (F1,10 = 12.94, p < 0.005). that sensory and motor systems merely represent time in a sim-
The purpose of the auditory training was to enhance the neur- ilar fashion, rather than sharing a common neural mechanism.
al representation of the standard interval. To assess whether this The anatomical correlate of this view might have sensory and
training affected performance on a motor task, participants per- motor timers located in sensory and motor cortices, respective-
formed two motor tasks before (session 1) and after training (ses- ly. A compromise view2 is that temporal representations for per-
sion 7). Only one of the two motor tasks involved a temporal ception and motor control, although anatomically distinct, are
interval matched to the standard used during the perceptual located more proximally (for example, adjacent regions of the
training task. In both tasks, participants used the right thumb, cerebellum8). There is mounting evidence for a cerebellar role in
which was covered from view, to press a button twice in succes- both sensory and motor timing9,10, particularly for the range of
sion. Following the second press, the interpress interval (IPI) was durations used in the present study11. The cerebellum also is
displayed on a monitor as feedback. Participants attempted to important in learning skills that require precise timing12, such as
produce a target IPI of 300 ms in one task and 500 ms in the the coordination of the individual components of multi-com-
other. There were 3 blocks of 50 trials in each task, and the ponent movements13, or the anticipation of temporally modu-
6 blocks occurred in random order. lated sensory stimuli 14 . Therefore it seems reasonable to
As is customary in studies of discrete motor timing5, we used hypothesize a cerebellar contribution to the generalized learn-
variability (standard deviation of IPI) as the dependent measure; ing reported here. Whatever the anatomical loci of sensory and
we also report the correspondence between mean IPI and target motor timers, our results suggest that they are closely intercon-
IPI (Fig. 1b). Motor learning was gauged by the reduction in tim- nected, such that plastic modifications of sensory temporal rep-
ing variability from before to after training. If this motor learning resentations automatically affect motor temporal representations.

1. Karni, A. & Bertini, G. Curr. Opin. Neurobiol. 7, 530535 (1997).

The finding that motor learning can occur without motor train- 2. Ivry, R. B. Curr. Opin. Neurobiol. 6, 851857 (1996).
ing has interesting implications for development and rehabilitation. 3. Kaernbach, C. Percept. Psychophys. 49, 227229 (1991).
In cases where the neural or musculoskeletal control of movement is 4. Wright, B. A., Buonomano, D. V., Mahncke, H. W. & Merzenich, M. M.
J. Neurosci. 17, 39563963 (1997).
underdeveloped or incapacitated, exposure to sensory input matched 5. Wing, A. & Kristofferson, A. Percept. Psychophys. 14, 512 (1973).
temporally to some future motor goal might accelerate the attain- 6. Ivry, R. B. & Hazeltine, R. E. J. Exp. Psychol. Hum. Percept. Perform. 21, 318
ment of that goal once motor production becomes possible. One (1995).
7. Treisman, M., Faulkner, A. & Naish, P. L. Q. J. Exp. Psychol. A 45, 235263
example is speech development, where speech perception could con- (1992).
ceivably train the neural control of speech production before the 8. Middleton, F. A. & Strick, P. L. Trends Cogn. Sci. 2, 348354 (1998).
vocal apparatus reaches the requisite level of maturity. 9. Ivry, R. B. & Keele, S. W. J. Cogn. Neurosci. 1, 136152 (1989).
10. Jueptner, M. et al. Neurology 45, 15401545 (1995).
11. Clarke, S., Ivry, R., Grinband, J., Roberts, S. & Shimizu, N. in Time, Internal
ACKNOWLEDGEMENTS Clocks, and Movement (eds. Pastor, M. A. & Artieda, J.) 257280 (Elsevier,
This work was supported by NIH grants T32-MH19942 and R29- MH54770, New York, 1996).
12. Cordo, P. J., Bell, C. C. & Harnad, S. (eds.) Motor Learning and Synaptic
and NSF research grant SBR-9873477. Plasticity in the Cerebellum (Cambridge Univ. Press, Cambridge, 1997).
13. Welsh, J. P. & Llins, R. Prog. Brain Res. 114, 449461 (1997).
RECEIVED 5 JUNE; ACCEPTED 3 AUGUST 2000 14. Buonomano, D. V. & Mauk, M. D. Neural Comput. 6, 3855 (1994).

commentary
Should the neuroscience community make a

paradigm shift to sharing primary data?
Stephen H. Koslow
The author outlines the pros and cons of data sharing for neuroscientists and argues that
continued progress in the field will depend on a cultural shift toward making primary data freely
available. He argues in favor of distributed databases to maximize the efficient use of data.
New discoveries about the nervous system and, to some extent, neuroimaging), only a now contains almost 13,000 three-dimen-
are being made at an enormous rate. There few representative examples are shown. sional structures, including 800 distinct pro-
are now about 50,000 to 65,000 neurosci- Either way, there is little prospect of re-ana- tein folds, providing a rich resource for
entists worldwide, publishing their results lyzing the data in any meaningful way. Har- structural insights in biology and medicine,
each month in more than 300 journals. Not nessing the capabilities of information and leading to discussions of a structural
only is this community growing steadily, but technology (IT) could in principle provide genomic initiative1. The field of bioinfor-
technological advances have meant that the a solution to this problem, allowing for the matics continues to advance rapidly; the
quantity of data per individual researcher immediate storage of experimental data and databases are growing daily, and new algo-
has also grown, in some cases by orders of background information in a standardized rithms, even new companies, are being cre-
magnitude. Functional imaging techniques digital form, so that it is available both to the ated to mine their contents in ever more
allow the activity of the human brain to be individual investigator andafter publica- sophisticated ways. Neuroscience is fre-
imaged in an almost endless variety of states tionto other researchers. quently the beneficiary of data-sharing prac-
and tasks. Dense electrode arrays allow the The fields of genomics and proteomics tices in other fieldsto cite one recent
activities of hundreds of cells to be record- have embraced IT much more effectively example, the identification of bitter taste
ed simultaneously with millisecond preci- than neuroscience. High-throughput receptors2,3 was made possible by the ability
sion. Confocal microscopy, along with new genome sequencing has revolutionized our to search through large amounts of genom-
indicator dyes, allows an unprecedented scientific thinking, and this would have been ic sequence databut there has been little
degree of visualization at the single-cell level. impossible without IT. Individual scientists progress toward implementing similar
Finally, the tools of molecular genetics now routinely contribute their sequence data to arrangements within the neuroscience com-
allow neuroscientists to monitor the expres- centralized databases at the time of publica- munity itself. If neuroscience is to continue
sion of thousands of genes within the ner- tion, and as a result, there are now some 8.6 as a major growth area within the biomed-
vous system, and to manipulate them in a billion bases of sequence data available in the ical sciences, I believe that we must do bet-
variety of experimental animals. public domain (www.ncbi.nlm.nih.gov/Gen- ter than this. In this commentary, I argue
The result of these advances is an explo- bank/GenbankOverview.html). Similarly, that the time is now propitious for the field
sion of data that threatens to outstrip our the Protein Data Bank (www.rcsb.org/pdb/) to formulate a federation of distributed data-
capacity to analyze and absorb it. Maintain-
ing our knowledge of even our own special- Table 1. Pros and cons of sharing primary data.
ized areas of research has become a major Against sharing primary data Response
task, despite the availability of on-line jour- No one else can understand the complexity This can be overcome by including the
nals and indexing systems. Moreover, as our of my data. relevant experimental conditions and
ability to gather data continues to increase, variables in the database.
the proportion that is analyzed, let alone If someone else analyzes my data, they may The true answer is what we are pursuing, and
come up with a different answer, disproving by considering different perspectives on the
published, becomes ever-smaller. A typical my perspective. same data set, we will come closer to reality.
article in a scientific journal contains a sum-
Someone else may find something new in my Finding something new in an existing data set
mary of the experiments and their implica- data that I did not see. will increase our scientific knowledge without
tions, but only a tiny fraction of the data on the unnecessary effort and cost of repeating
which they are based is ever presented. In the entire experiment.
some cases, data are summarized as mean It is my data that I worked very hard to Publication of the study already implies that
values along with error estimates; in others collect, and no one else has the right to it. its results and conclusions are to be shared. If
these are to be evaluated in detail, readers
(particularly neurophysiology, morphology should have access to the primary data on
which they are based. In most cases, sharing
primary data also serves the interests of the
Stephen Koslow is at the National Institute of organizations that fund the research.
Mental Health, National Institutes of Health, I have not finished analyzing my data, and I A published paper suggests that the
6001 Executive Blvd., Room 6167, Bethesda, will make it available once my analysis is experimental data have been substantially
Maryland 20892-9613, USA complete. analyzed; thus sharing at this point would
e-mail: koz@helix.nih.gov seem appropriate.
This is not an official policy document of the I cannot trust or understand the data Currently we all try to understand data from
National Institute of Mental Health (NIMH), produced in another laboratory. other laboratories whenever we read the
scientific literature; this influences our own
National Institutes of Health (NIH) or the U.S. future experimental and theoretical pursuits.
federal government. It reflects the opinions of the Having the complete data set available for re-
author. analysis would increase experimental efficiency.

commentary
Table 2. Organizations supporting the Human Brain Project. and compared, as well as related to data
Federal agencies NIH institutes emerging from other disciplines such as
genomics and proteomics. In sum, by trans-
National Institutes of Health (NIH) National Institute of Mental Health
National Science Foundation National Institute on Drug Abuse forming neuroscience from a cottage indus-
National Aeronautics and Space National Institute on Aging try to big science status, we would
Administration National Institute of Child Health and Human maximize the value obtained from research
Department of Energy Development
National Institute on Deafness and other funding, thereby allowing the field to reach
Communication Disorders its full potential.
National Library of Medicine Clearly, these benefits will only be real-
National Heart, Lung, and Blood Institute
National Institute of Dental and Craniofacial ized if individual researchers are willing to
Research share their data with their peers. I, like many
National Institute on Alcohol Abuse and others, believe that scientists ought to make
Alcoholism
National Institute of Neurological Disorders freely available the data on which their
and Stroke papers are based, and that this should be
Fogarty International Center considered an intrinsic part of scientific
National Cancer Institute publication. This is particularly so in cases
where the research has been funded with
bases and analytical tools to maximize its species or different levels of analysis. Such public money; regardless of the source of
efficient use of experimental data. a workbench already exists for analyzing funding, however, I believe science and
The creation of a centralized neuro- genomic data4; the neuroscience work- humanity are best served if we share our
science database is a formidable task. bench would be far more complex in its data, to maximize progress for the public
Whereas genomic sequence data is rela- capabilities, but the concept is similar. The good. What are the counter-arguments?
tively simple in its structure, the datasets final step in this process would be to link Some of the most frequently stated reasons
generated by neuroscience research are far the databases to the published literature, are listed in Table 1. As this table indicates,
more diverse and complex. The relevant by encouraging researchers to submit their however, I believe that each of these argu-
variables may include morphology, func- primary data at the time of publication. ments can be answered, and although others
tional connectivity, neurophysiology, The advantages of having all these data could doubtless be added to the list, none is
chemistry, molecular biology, genomics, available a few clicks away would be enor- sufficient, in my view, to outweigh the ben-
brain imaging and behavior. These vari- mous. It would allow researchers to make efits of sharing data.
ables may change over time scales ranging meaningful comparisons between their own If the vision of a shared set of databas-
from milliseconds to years, and may be results and those of others. It would allow es is to become a reality, there will need to
subject to diverse experimental manipula- them to answer what if questions and to be a paradigm shift in how neuroscience
tions. Many other factors may also con- carry out pilot studies without repeating research is conducted. To create the neces-
tribute to the context of an experiment and experiments; this alone would lead to more sary technical infrastructure, neuroscien-
be essential for its interpretation. Clearly it efficiency in terms of time, money and use tists and information scientists will need to
is not realistic at present to create a single of experimental animals. Databases of learn each others languages and under-
database that can store all these types of empirical data would be an invaluable stand their questions and approaches. But
information. A more practical approach is resource for modelers seeking to build and there will also need to be a change in the
to establish a variety of smaller-scale ini- test theories based on realistic biological scientific culture, if researchers are to vol-
tiatives, in which individual neuroscientists assumptions, and would also be useful to untarily share their data with other
work with information scientists to create students, clinicians, educators, industry and researchers, including their competitors.
databases relevant to their own research. the general public. More generally, the cre- Although funding organizations such as
Over time, these will evolve into a federa- ation of a federated set of databases would NIH might in principle require their grant
tion of distributed databases and analyti- lead to a more integrated view of brain func- recipients to deposit primary data in public
cal tools that encompass the breadth of tion, in which results obtained through dif- databases, I believe that this would be coun-
neuroscience research. These databases and ferent levels of analysis could be combined terproductive; the new paradigm will only
tools need to be web-based, with appro-
priate security to protect the data from cor- Table 3. Countries participating in the Organization for Economic Cooperation
ruption, yet be available to scientists on and Development.
request. The request itself could come Member countries Observers
through a neuroscience portal using a
Australia Luxembourg Israel
smart neuroscience browser instructed to Austria Mexico Russia
look for a particular variable or set of vari- Belgium The Netherlands Slovak Republic
ables and import the data back to the users Canada New Zealand
Czech Republic Norway
computer. The data could then be analyzed Denmark Poland
using a neuroscience workbench, a set of Finland Portugal
programs and tools into which a wide vari- France South Africa
ety of data could be imported from vari- Germany Spain
Greece Sweden
ous sites and then manipulated and Hungary Switzerland
analyzed. This would allow the user to Iceland Turkey
search and reanalyze data, to use it for sim- Ireland United Kingdom
Italy United States
ulations and models, and to compare data Japan
from different experiments, different Korea European Commission

commentary
work if it evolves from a shared consensus engineers, computer scientists, mathe- namely to provide, first, a resource or
within the worldwide scientific communi- maticians and physicists). The neuro- clearinghouse for neuroinformatics tools
ty. Journals could become involved by science components of HBP-funded and databases, second, guidelines for
requiring their authors to release data as a projects range from genomics to systems interoperability, organization and quality
precondition of publication, but again this neuroscience to human and animal behav- control for these tools and databases, and,
will only be effective if the standards have ioral studies. The informatics components third, a portal to serve as an internet-
broad community support. To bring about include creating new database and storage based knowledge repository for neuroin-
the necessary paradigm shift, we will need capabilities, ways to query and retrieve data formatics.
to rethink career incentives for individual from databases, methods for analysis, pre- Progress in neuroscience, as in other
researchers as well as the responsibilities and sentation and visualization of data, data disciplines, depends on recognizing the
opportunities offered by universities. In integration and synthesis (models and sim- value of new discoveries, integrating the
particular, we will need to alter the way in ulations) as well as methods for electronic new information into existing hypotheses
which we evaluate individual scientific pro- collaboration. Detailed information about or formulating new ones, and then carry-
ductivity, to include not only research the HBP and summaries of funded grants ing out additional experiments to test
papers but also creation of databases and (with links to their individual web sites) these hypotheses and thus extend our
associated tools, contributions of data to can be found at www.nimh.nih.gov/neu- knowledge. Neuroscience is inherently
these databases, and mining of their con- roinformatics/index.cfm. interdisciplinary, and so this process can
tents. Universities will need to provide The creation of a comprehensive set only work efficiently if researchers can
opportunities for training across disciplines, of neuroscience databases will necessari- integrate their own results with those
and means for interdisciplinary scientists ly be an international project, and to this obtained from other disciplines. At pre-
to work together while maintaining exper- end the HBP has held consultative meet- sent, progress in the field is being impeded
tise in their primary discipline. At the same ings with the Medical Research Council by the lack of a knowledge-management
time, it will be important to ensure that of the UK and with the Medical Research infrastructure for comparing and com-
intellectual property rights, including copy- Council of the European Science Foun- bining results within and across disci-
right and patents, are not compromised; dation. Both these organizations have plines. Fortunately, efforts are now
many groups are concerned with resolving gone on to host a variety of meetings to underway in many countries to address
these complex issues in a way that will allow discuss future activities in neuroinfor- this situation, and to lay the basic ground-
the protection of commercial interests with- matics. Japan has also recently funded its work for a paradigm shift in which pri-
out impeding scientific progress5,6. first program in neuroinformatics, at Toy- mary research data are openly shared
To facilitate this paradigm shift and to ohashi University of Technology. Several within the worldwide neuroscience com-
create the necessary technical infrastructure, major multi-national initiatives are also munity. The potential impact will be very
in 1993 the US government initiated the underway. A joint USEuropean Com- great, but this potential will only be real-
Human Brain Project (HBP; Table 2), of munity (EC) neuroinformatics steering ized if researchers come to believe that the
which I chair the coordinating committee. committee has been established under the incentives for cooperation and sharing are
The HBP was initiated following a report7 auspices of the USEC Task Force on greater than the costs. If and when this
from the Institute of Medicine (part of the Biotechnology Research; this has led to happens, neuroscience will enter a new era
National Academy of Sciences), and the development of a grant funding pro- of growth, analogous to that which has
although its ultimate goal is to understand gram for neuroinformatics in the Fifth transformed the genetics community. The
the human brain, it encompasses work on Framework program of the EC to support result will be a new depth of understand-
all species, as well as research in informat- training and planning activities as well as ing of how the nervous system works in
ics and biological computation. The HBP the creation of databases (www.cordis.lu). both health and disease.
budget for fiscal year 2000 is $10.6 million. At a more global level, a biological infor-
It uses these funds to support feasibility matics working group met from 1. Burley, S. K. et al. Structural genomics: beyond
the human genome project. Nat. Genet. 23,
studies (Phase I) and studies that call for 19961999 within the Megascience 151157 (1999).
sharing of resources on-line, providing com- Forum (MSF) of the Organization for 2. Adler, E. et al. A novel family of mammalian
putational modeling capabilities, beta test- Economic Cooperation and Development taste receptors. Cell 100, 693702 (2000).
ing, documentation and plans for (OECD) in Paris. The Forum brings 3. Matsunami, H., Montmayeur, J.-P. & Buck, L. B.
maintenance of these resources (Phase II). together science policy officials from A family of candidate taste receptors in human
In addition, it supports curriculum devel- OECD Member governments (Table 3) and mouse. Nature 404, 601604 (2000).
opment and short-term training in neu- to discuss ways of strengthening interna- 4. Subramaniam, S. The biology workbencha
seamless database and analysis environment for
roinformatics, including grants to tional cooperation on very large scientif- the biologist. Prot. Struct. Funct. Genet. 32, 12
universities to train young researchers under ic projects and programs. The 1999 (1998).
the mentorship of an established investiga- report of the neuroinformatics subgroup 5. A Question of Balance: Private Rights and the
tor. The HBP annual meeting provides (http://www.oecd.org/dsti/mega) pro- Public Interest in Scientific and Technical
opportunities for grantees to share their vided a defining report for the field. This Databases (National Research Council, National
Academy Press, Washington, D.C., 1999).
research results and to interact with each year, the Global Science Forum (GSF; a
6. Burk, D. L. in Electronic Collaboration in
other, with non-HBP scientists and with continuation of the Megascience Forum Science (eds. Koslow, S. H. & Huerta, M. F.)
NIH science administrators. under a new five-year mandate) ratified 1544 (Lawrence Erlbaum, Mahwah, New
By supporting these activities, we aim a proposal from the US and Italy to estab- Jersey, 2000).
to set the stage for the new paradigm of lish a neuroinformatics working group to 7. Pechura, C. M. & Martin, J. B. Mapping the
interdisciplinary collaboration, requiring follow up on some of the recommenda- Brain and its Functions. Integrating Enabling
Technologies into Neuroscience Research
a joint research effort between neuroscien- tions from this defining report. The GSF (National Academy Press, Washington, D.C.,
tists and information scientists (including working group has three major goals, 1991).

articles
CLF associates with CLC to form a

functional heteromeric ligand for
the CNTF receptor complex
Greg C. A. Elson1, Eric Lelivre2, Catherine Guillet2, Sylvie Chevalier2, Hlne Plun-Favreau2,
Josy Froger2, Isabelle Suard2, Amlie Benoit de Coignac1, Yves Delneste1, Jean-Yves Bonnefoy1,
Jean-Franois Gauchat1 and Hugues Gascan2
1 Centre dImmunologie Pierre Fabre, 5 Avenue Napoleon III, 74164 St. Julien-en-Genevois, France
2 INSERM E 9928, CHU Angers, Batiment Monteclair, 4 Rue Larrey, 49033 Angers, France
The first three authors contributed equally to this work
Correspondence should be addressed to J.-F.G. (jean.francois.gauchat@pierre-fabre.com)
Ciliary neurotrophic factor (CNTF) is a cytokine supporting the differentiation and survival of various
cell types in the peripheral and central nervous systems. Its receptor complex consists of a non-
signaling chain, CNTFR, and two signaling chains, gp130 and the leukemia inhibitory factor
receptor (LIFR). Striking phenotypic differences between CNTF- and CNTFR-deficient mice suggest
that CNTFR serves as a receptor for a second, developmentally important ligand. We have identified
this factor as a stable secreted complex of cardiotrophin-like cytokine (CLC) and the soluble receptor
cytokine-like factor-1 (CLF). CLF expression was required for CLC secretion, and the complex acted
only on cells expressing functional CNTF receptors. The CLF/CLC complex activated gp130, LIFR and
signal transducer and activator of transcription 3 (STAT3) and supported motor neuron survival. Our
results indicate that the CLF/CLC complex is a second ligand for CNTFR with potentially important
implications in nervous system development.
The cytokine CNTF seems to be released only by damaged cells1,2. ronal cell types following injury. In striking contrast, newborn
It promotes the differentiation and survival of a wide range of mice deficient in CNTFR fail to begin feeding, die shortly after
cell types in the peripheral and central nervous systems2 and sus- birth and show a substantial loss of motor neuron populations18.
tains the survival of motor neurons in vitro3 and in vivo4. CNTF These contrasting phenotypes point to the existence of a second,
prevents the degeneration of axotomized motor neurons5 and developmentally important ligand for CNTFR. In addition,
attenuates the motor deficits found in different strains of mice CNTFR is strongly expressed in the developing embryo, where-
with neuromuscular deficiencies6,7. In addition to its activities in as CNTF expression is barely detectable19. LIFR- and CNTFR-
the nervous system, CNTF is trophic for denervated skeletal mus- deficient mice have overlapping phenotypes, suggesting that LIFR
cle8 and regulates muscular strength in aging9. is also a signaling subunit for this second CNTF-like ligand20.
CNTF belongs to the IL-6 family of cytokines, which also con- We identified an orphan cytokine receptor, CLF, which shares
tains IL-6, IL-11, LIF, oncostatin M (OSM) cardiotrophin-1 a significant level of homology with the receptor subunits of the
(CT-1) and CLC, the latter also known as novel neurotrophin-1/B IL-6 cytokine family but lacks transmembrane and cytoplasmic
cell stimulating factor-3 (refs. 10, 11). These cytokines share one domains21. The CLF gene was independently cloned by a second
or both of the receptor signal transducing subunits gp130 and LIFR group and inactivated in mice22. The gene deletion leads to peri-
in their respective receptor complexes, resulting in some functional natal death due to a suckling defect, similar to the phenotype
redundancy throughout the family12. In addition to these shared observed in CNTFR-deficient mice. This led us to hypothesize
subunits, the CNTF receptor complex contains a specific binding that CLF could be involved in the formation of a second ligand
subunit, CNTFR13, anchored to the cell membrane through a gly- for CNTFR and therefore prompted us to identify proteins capa-
cosylphosphatidylinositol (GPI) linkage. The association between ble of interacting with CLF. Here we show that CLF forms a stable
CNTF and CNTFR leads to recruitment and dimerization of gp130 secreted complex with the novel IL-6-type cytokine CLC. The
and LIFR14, which in turn induces downstream signaling events, heterocomplex acts only on those cells expressing the functional
including activation of the Janus kinase 1 (JAK1)/STAT3 pathway15. CNTF receptor (CNTFR, gp130 and LIFR) and therefore is like-
CNTF-deficient mice have a mild phenotype involving motor ly to be the developmentally important second ligand for CNTFR.
neuron degeneration that is only prevalent postnatally16. Fur-
thermore, a small proportion of Japanese people (2.3 %) are RESULTS
homozygous for a mutation inactivating the CNTF gene with- CLF and CLC form a secreted heterodimer
out any sign of neurological defect, even in old age17. It therefore We had speculated that CLF was part of a functional heteromer-
seems that the main role of CNTF is maintenance of certain neu- ic complex similar in structure to IL-12, a covalently linked het-

articles
Fig. 1. CLF and CLC form a stable secreted complex. (a) CLC requires
co-expression of CLF for secretion. COS cells were transfected as indi- a
cated. Top, CLCmyc and p35myc in cell lysates. Bottom, CLCmyc secre-
tion following coexpression with CLF. (b) CLF-Fc and CLCmyc are
secreted as a stable complex. COS cells were transfected as indicated.
Top, CLCmyc and p35myc in cell lysates. Middle, Fc derivatives present
in the culture media. Bottom, CLCmyc coprecipitation with CLF-Fc. IP,
immunoprecipitation; WB, western blot.
erodimer composed of two polypeptide chains, p35 and p40

(ref. 23). IL-12 p35 contains a putative signal peptide and shares b
homology with IL-6-type cytokines, whereas IL-12 p40, like CLF,
is a soluble protein homologous to type-I cytokine receptors.
Cells efficiently secrete p35 only when it is coexpressed with p40.
To identify subunits interacting with CLF, we expressed an epi-
tope-tagged form of CLC (CLCmyc) in COS cells. Although we
detected the protein in cell lysates, it was not secreted into the cul-
ture medium (data not shown). CLC contains a putative signal pep-
tide, suggesting that it enters into the classical secretory pathway.
In this respect, CLC seems to resemble IL-12 p35, which also con-
tains a signal peptide but is poorly secreted. We therefore cotrans-
fected CLCmyc with CLF in COS cells. The culture medium was
immunoprecipitated with a tag-specific monoclonal antibody, and induce secretion of the cytokine. To test whether CLC and CLF were
the purified fraction was analyzed by western blotting. CLCmyc released as a stable heterocomplex, we coexpressed CLCmyc with
was present in the culture medium of cells cotransfected with a CLF-Fc fusion protein. Proteins forming a complex with CLF-Fc
cDNAs for both CLCmyc and CLF but not those from cells trans- in the culture medium were immunoprecipitated with protein G-
fected with CLCmyc cDNA alone (Fig. 1a), suggesting that CLC sepharose. The purified fraction was assayed by western blotting
release depends on its coexpression with CLF. We failed to detect with an anti-c-myc monoclonal antibody, which identified CLCmyc,
an epitope-tagged form of IL-12 p35 (p35myc) in the culture medi- demonstrating the formation of a stable secreted complex (Fig. 1b).
um when it was coexpressed with CLF. Likewise, coexpression of Neither IL-12 p35myc and CLF-Fc nor CLCmyc and a sIL-6R Fc
the soluble IL-6 receptor chain (sIL-6R) with CLCmyc failed to fusion derivative (sIL-6R-Fc) formed stable complexes secreted
from cells coexpressing the two proteins, suggesting that CLF shows
a specificity in its interaction with CLC and vice versa. Identical obser-
vations were made following transfections in a human embryonic
kidney fibroblast cell line (data not shown). These results show that
CLC secretion is controlled by CLF. SDS-polyacrylamide gel elec-
trophoresis of the complex under non-reducing conditions, fol-
lowed by western blotting, indicated that CLC was not covalently
linked to CLF by a disulfide bridge (data not shown).
The tripartite CNTF receptor is the CLF/CLC receptor

To investigate the functional activity of the CLF/CLC hetero-
complex, we used derivatives of the IL-3-dependent Ba/F3 cell
Fig. 2. CLF/CLC uses the functional CNTF receptor. (a) Proliferation of

transfected Ba/F3 cells to CLF/CLC. Ba/F3 cells (transfected as indi-
cated), were incubated with 3-fold serial dilutions of the control
cytokine (open circles, 10 ng/ml highest concentration) or culture media
from COS cells (1/5 dilution at highest concentration) transfected with
CLC (open triangles), CLF (filled triangles), CLF/CLC (filled circles) or
c empty vector (open squares). Proliferation was assayed as described24.
b Vertical bars, s.e.m. (b) Inhibition of CLF/CLC-induced Ba/F3 cell prolif-
eration with anti-gp130 and LIFR neutralizing antibodies. Ba/F3 cells
transfected with cDNAs for gp130, LIFR and CNTFR were incubated
with purified CLF/CLC (10ng/ml) alone (none) or with 50g/ml of an
anti-gp130, anti-LIFR (AN-HH1 and AN-F1 respectively; H.G. generated
in the laboratory) or an isotype-matched control Ab, as indicated.
Proliferation was assayed as described in (a). Horizontal bars, s.e.m.
(c) Induction of IL-6 production in CNTFR-transfected KB cells by
CLF/CLC. KB cells, transfected as indicated, were exposed to medium
alone, CLF/CLCmycprotC, CNTF, LIF or IL-2 (40 ng/ml). IL-6 produc-
tion was measured as described24. Horizontal bars, s.e.m.

articles
a b
Fig. 3. Physical interaction between CLF/CLC and CNTFR. (a) Co-immunoprecipitation of CLF/CLC and CNTFR. Left, Mock or CLF-Fc/CLCmyc
culture media were immunoprecipitated with protein G-sepharose following the addition of buffer (-CNTFR) or 1 g/ml sCNTFR (+CNTFR).
CLCmyc, CLF-Fc and CNTFR were detected in the immunopurified fraction with an anti-c-myc monoclonal antibody, an anti-human IgG antibody and
an anti-CNTFR monoclonal antibody (AN-E4, C.G. unpublished data), respectively. Center and right, co-immunoprecipitation of CNTFR with CLF-
Fc/CLCmyc from the surface of BAF GLC and IMR 32 cells. Proteins were immunoprecipitated from lysates with protein G-sepharose and revealed
as in the left panel. (b) CLF/CLC binding to transfected cells expressing CNTFR, detected with the anti-c-myc monoclonal antibody.
line, rendered responsive to cytokines of the IL-6 family by trans- blotting revealed the presence of CLCmyc, CLF-Fc and sCNTFR
fection with cDNAs encoding the appropriate receptor chains24. only in the purified fraction of culture medium containing CLF-
COS culture medium containing the CLF/CLC complex only Fc/CLCmyc. Furthermore, CNTFR coprecipitated with the CLF-
induced proliferation of Ba/F3 cells (termed BAF GLC) express- Fc/CLC heteromer when either IMR 32 or BAF GLC cells were
ing the functional CNTF receptor (gp130, LIFR and CNTFR) on incubated with the CLF-Fc/CLC complex and the correspond-
their surface. Those cells expressing only gp130 and LIFR (BAF ing cell lysates analyzed by immunoprecipitation and western
GL) failed to proliferate in the presence of the complex. Culture blotting as above (Fig. 3a). These results strongly argue in favor
media from control transfections did not produce significant sig- of direct contact between the heteromeric cytokine and CNTFR.
nals (Fig. 2a). This demonstrates that both CNTF and CLF/CLC This is further supported by flow cytometric analysis showing
require the same cell-surface receptor components to induce that BAF GL and L 929 cells acquire the ability to bind CLF/CLC
Ba/F3 proliferation. Both anti-gp130 and anti-LIFR neutralizing following transfection with CNTFR cDNA (Fig. 3b). Cell sur-
antibodies inhibited the proliferative response to CLF/CLC, impli- face CLF/CLCmyc binding was also strictly correlated with sur-
cating both these signaling subunits in the functional receptor face expression of CNTFR for a number of cell lines (Table 1).
complex for CLF/CLC (Fig. 2b). Cytokines signaling through gp130/LIFR can usually com-
CLF/CLC was also inactive in conventional LIF-responsive pete, at least to some extent, for binding to the same receptor
assays using cell lines expressing gp130 and LIFR (the two com- complex25,26. We therefore examined whether CLF/CLC and
ponents of the functional LIF receptor), but not CNTFR, on their CNTF could be mutually displaced from the cell membrane.
surface. The heterocomplex failed to trigger the proliferation of CLF/CLCmyc binding to L 929 cells transfected with CNTFR
TF1 erythroleukemia cells, activation of HepG2 hepatoma cells (L-CNTFR) was monitored by flow cytometry in the presence
or IL-6 synthesis in KB epidermoid carcinoma cells. KB cells were of increasing concentrations of CNTF. There was a dose-depen-
responsive to CLF/CLC or CNTF only
when transfected with membrane- Table 1. Cell surface binding of CLF/CLC correlates with expression of CNTFR.
bound CNTFR (Fig. 2c). CLF/CLC and Tissue origin Cell line CNTFR expression CLFCLC binding
CNTF also induced STAT3 tyrosine Neuroblastoma SK-N-FL ++ ++
phosphorylation in HepG2 cells trans-
SK-N-GP +++ +++
fected with CNTFR, but were inactive on
SK-N-SH ++ ++
both non-transfected and mock trans-
fected cells (data not shown). These SY5Y ++ ++
results demonstrate that CLF/CLC IMR-32 +++ +++
specifically requires CNTFR for biologi- Glioblastoma T98G
cal activity. SW1088
We also demonstrated direct contact GO-G-UVM
between CLF/CLC and CNTFR by GO-G-CCM
adding soluble CNTFR (sCNTFR) to Ewing sarcoma EW1 + +
culture medium containing the CLF-
Epidermoid carcinoma KB
Fc/CLCmyc heterocomplex or a mock
Hepatoma HepG2
control, followed by immunopurifica-
tion with protein G-sepharose. Western Osteosarcoma MG63

articles
Fig. 4. CLF/CLC and CNTF com-

pete for the CNTFR. (a) CNTF dis-
a b
places CLF/CLC from L-CNTFR.
Cells were incubated with
CLF/CLCmyc culture medium and
the indicated concentration of either
CNTF or IL-4. Binding was detected
with the anti-c-myc monoclonal anti-
body. (b) CLF/CLC displaces CNTF
binding to L-CNTFR and IMR 32
cells. Radiolabeled CNTF was added
to cells in the presence of buffer
(marked as 0), CNTF (100 ng), IL-4
(100 ng), a 1/3 dilution of
CLF/CLCmyc culture medium or a
mock control. Vertical bars, s.e.m.
dent competition for receptor binding, with no effect on bind- expressing the functional CNTF receptor complex also provides
ing with IL-4 in a control experiment (Fig. 4a). Conversely, radi- clear evidence that two non-signaling so-called receptor chains
olabeled CNTF binding to L-CNTFR or IMR32 cells was reduced (CLF and CNTFR) can be present in the same IL-6-type receptor
by approximately 50% by addition of COS cell culture medium complex, and suggests that CLC has independent binding sites
containing CLF/CLC (Fig. 4b). It remains to be determined for CLF and CNTFR. These findings also indicate that receptor
whether this partial competition is related to a low concentra- molecules are involved in the secretion of cytokines independent
tion of the CLF/CLC in the culture medium, or to a difference of their ability to confer cell specificity. Indeed, our findings sug-
between CNTF and CLF/CLC in their respective affinity for the gest that the actions of a cytokine (in this case CLC) may be sub-
receptor subunits, as observed for LIF and CT-1 (refs. 25, 26). ject to a dual control, whereby the cytokine must recruit two
We also studied CLF/CLC-induced signal transduction. The receptor chains, with complementary but clearly contrasting
heterocomplex failed to induce tyrosine phosphorylation of roles, to exert its biological effects. In the first instance, CLC must
gp130, LIFR and STAT3 in HepG2 cells (gp130+, LIFR+, interact with CLF in the producer cell to be released. A subse-
CNTFR) but readily phosphorylated these proteins in SK-N- quent interaction with CNTFR is then required to define target
GP neuroblastoma cells (gp130+, LIFR+, CNTFR+). Tyrosine cell specificity.
phosphorylation was induced in both cell lines following LIF We also demonstrated that CNTF and the CLF/CLC hetero-
stimulation, with CNTF only affecting SK-N-GP cells (Fig. 5). complex act through the same receptor complex to generate sim-
Furthermore, we consistently observed an association between ilar biological functions (Fig. 6). The existence of a second,
gp130, LIFR and a third phosphorylated protein (with the same developmentally regulated ligand for CNTFR has been postulat-
molecular weight as JAK1), which were all co-immunoprecipi- ed18,20. In contrast to CNTF, CNTFR is highly expressed in the
tated. Similar results were also obtained with a second CNTFR- developing embryo19, and we are therefore investigating embry-
expressing neuroblastoma cell line, IMR 32 (data not shown). onic expression of both CLF and CLC. However, CLF transcripts
We then demonstrated that CNTF and CLF/CLC had over- are indeed present in the developing CNS22. That CLF/CLC is
lapping biological properties. CNTF can sustain the in vitro sur- secreted from cells distinguishes it from CNTF, which seems
vival of embryonic motor neurons 27. Rat embryonic motor mainly to be released from cells during trauma. This would sug-
neurons were therefore purified to homogeneity based on their gest that CLF/CLC acts as a physiological regulator in non-trau-
surface expression of the p75 low-affinity neurotrophin recep- matic situations. Taken together, these findings suggest that this
tor. A 48-hour cell survival assay was performed with serial dilu- heterocomplex has the properties of a second CNTFR ligand
tions of either CLF/CLC, CNTF, LIF or IL-4. LIF, CNTF and important during development. This is further supported by the
CLF/CLC all sustained in vitro motor neuron survival, whereas overlapping phenotypes observed between CNTFR- and CLF-
no effect was observed with IL-4 (Fig. 6). deficient mice18,22.
DISCUSSION
This study reveals a novel release mechanism for cytokines in the
IL-6 family, whereby the interaction between a non-secreted
cytokine and a soluble cytokine receptor in the same cell leads to
the release of a functional heterodimer. This raises the possibili-
ty that the secretion of other as-yet unidentified cytokine-like
molecules is controlled in a similar fashion by either CLF or other
soluble cytokine receptors. The finding that CLF can co-immuno-
precipitate with CLC and CNTFR from the surface of cells
Fig. 5. Tyrosine phosphorylation of gp130, LIFR and STAT 3 in SK-N-GP

and HepG2 cells. Following exposure to PBS (marked as 0) LIF, CNTF or
CLF/CLCmyc, cells were lysed and analyzed by immunoprecipitation (IP)
and western blotting (WB) as indicated. P-Y, phosphotyrosine.

articles
Manassas, Virginia) for 4 h at 4oC and subsequently incubated with

protein G-sepharose beads for 1 h at 4 oC. Fc fusion proteins were
directly immunoprecipitated with protein G-sepharose beads for 2 h at
4oC. Following extensive washing, beads were boiled in Tris-Glycine-
SDS sample buffer under reducing conditions, and samples were sep-
arated by SDS-polyacrylamide gel electrophoresis before being
transferred to a PVDF membrane. Membranes were subsequently
probed with the relevant antibody and processed for enhanced chemi-
luminescence detection (ECL, Amersham Pharmacia Biotech, Upp-
sala, Sweden).
Flow cytometric analysis. Cells were successively incubated for 30 min

at 4 oC with the appropriate primary antibody (10 g/ml) and a phy-
coerythrin-conjugated anti-mouse antibody. Fluorescence was subse-
Fig. 6. CLF/CLC supports the survival of rat embryonic motor neurons quently analyzed on a FACScan flow cytometer (Becton Dickinson,
in vitro. Cells were incubated with indicated concentrations of recombi- Erembodegen, Belgium).
nant IL-4 (filled triangles), CNTF (open triangles), LIF (open squares) or
CLF/CLCmycprotC (filled squares). Cell survival was monitored by Radiolabeling and binding assays. CNTF was iodinated as described26.
reduction of MTT. Vertical bars, s.e.m. One ng of iodinated CNTF (specific activity 3.2 104 cpm/ng) was added
to 3 106 cells in the presence of the relevant control or cytokine. Fol-
lowing a 2 h incubation at 4oC, bound radioactivity was separated from

the unbound fraction and determined as described26.
Because of its well characterized neuroprotective effects both
Tyrosine phosphorylation analysis. Following an overnight incubation
in vitro and in vivo, CNTF has been tested in preclinical and clin- in serum-free medium, SK-N-GP or HepG2 cells were stimulated for
ical trials for neurodegenerative disorders, such as Huntingtons 10 min with the relevant control or cytokine (50 ng/ml). Protein tyrosine
disease and amyotrophic lateral sclerosis28,29. Progress to date has phosphorylation was subsequently analyzed essentially as described24,25.
been impaired largely by the poor tolerance to this cytokine when
administered systemically. More promising results hve been Protein immunoaffinity purification. Following cotransfection with
obtained in pre-clinical and clinical trials with CNTF targeted to CLF and CLCmycprotC, COS cells were cultured for 72 h in serum-
the CNS by direct intrathecal injection or intrathecal implanta- free medium. Cell culture medium was harvested, concentrated
tion of encapsulated transfected cells3032. As CLF/CLC represents approximately 20-fold using Centricon-80 concentrators (Millipore,
a naturally secreted ligand for CNTFR, it would be of interest to Bedford, Massachusetts) and incubated overnight at 4oC with an anti-
protein C affinity matrix (Roche Diagnostics). Following extensive
examine the ability of this heteromer to protect neurons in sim-
washing (100 mM phosphate buffer, 1 M NaCl, 1 mM CaCl2), pro-
ilar preclinical and clinical studies. teins were eluted from the affinity matrix by chelating calcium ions.
Protein concentration was determined by SDS-polyacryamide gel elec-
METHODS trophoresis and silver staining using a bovine serum albumin protein
Cloning. All cDNAs were cloned into pCDNA3 (Invitrogen, Carlsbad, standard.
California) with the exception of the CNTFR cDNA, which was cloned
in the episomal expression vector pEBS-PL. The cDNA encoding CLF Motor neuron survival. Rat embryonic motor neurons (embryonic
(AF059293) was cloned by PCR from human fetal lung cDNA; cDNA day 14.5) were purified essentially as described for the mouse33 except
encoding the extracellular portion of IL-6R (sIL-6R) was amplified by that the metrizamide cushion centrifugation step was replaced by
PCR from an expression construct encoding the membrane-bound form filtration through a metal filter. Immunomagnetic separation was
of the protein. The cDNAs encoding CLC and IL-12 p35 were amplified by done with the anti-p75 NGF receptor monoclonal antibody MC192
PCR from human lymph node cDNA. The CLC (AF172854) and IL-12 (ATCC). A purity of over 98% was determined for the cell popula-
p35 expression constructs were modified so that both proteins contained tion by flow cytometry. Cells were cultured in neurobasal medium
the c-myc peptide epitope (EQKLISEEDL), or the c-myc peptide epitope (Life Technologies, Rockville, Maryland) supplemented with L-glu-
followed by the protein C peptide epitope (EDQVDPRLIDGK) in the case tamine, B7 supplement, 5 10 -5 M 2--ME and 2% FCS. Fifteen
of CLC (CLCmycprotC), at their C termini. The CLF-Fc expression con- hundred neurons were seeded in triplicate in 96-well plates coated
struct contained cDNA encoding the complete CLF amino acid sequence with polyornithin and laminin, and cytokines at the appropriate con-
(as above) fused to the amino acid sequence of the Fc portion of human centrations were added to the cultures. After a 48 hour incubation
IgG1. The sIL-6R-Fc expression construct contained cDNA encoding the period, survival of cultured motor neurons was monitored by reduc-
extracellular region of IL-6R fused to the amino acid sequence of the Fc tion of a tetrazolium salt (MTT). Optical densities were measured
portion of human IgG1. at 570 nm.
Cell culture and transfection. COS cells were cultured in Dulbeccos

modified Eagles medium (DMEM) supplemented with 10% FCS. HepG2 ACKNOWLEDGEMENTS
hepatoma cells and SK-N-GP and IMR 32 neuroblastoma cells were E.L. is supported by an Association Franaise contre les Myopathies (AFM)
maintained in RPMI 1640 medium supplemented with 10% FCS. Ba/F3 grant. The authors thank K.J. Kallen at the University of Mainz, Germany
cells were maintained as described24. The lipid reagent Fugene 6 (Roche for providing transfected Ba/F3 cell lines, V. Steimle at the Max-Planck
Diagnostics, Meylan, France) was used for all transfections. Cells were Institute for Immunobiology, Freiburg, Germany for providing the pEBS-PL
cultured for 4872 hours after transfection before analysis. expression vector and C. Lacheny and F. Derouet for preparation of the
anti-c-myc antibody.
Immunoprecipitation and western blotting. Where appropriate, cells
were lysed in 1% Brij 96 buffer (50 mM Tris-HCl, pH 7.5, 150 mM
ACCEPTED 19 JULY 2000
NaCl, 1% Brij 96) containing protease inhibitors25,26. A centrifugation
step removed cell debris from both lysates and cell culture media. The
c-myc epitope-tagged proteins were immunoprecipitated from sam- 1. Stckli, K. A. et al. Molecular cloning, expression and regional distribution of
ples with the c-myc monoclonal antibody 9E10 (10 g/ml; ATCC, rat ciliary neurotrophic factor. Nature 342, 920923 (1989).

articles
2. Sendtner, M., Carroll, P., Holtmann, B., Hughes, R. A. & Thoenen, H. Ciliary 19. Ip, N. Y. et al. The alpha component of the CNTF receptor is required for
neurotrophic factor. J. Neurobiol. 25, 14361453 (1994). signaling and defines potential CNTF targets in the adult and during
3. Arakawa, Y., Sendtner, M. & Thoenen, H. Survival effect of ciliary development. Neuron 10, 89102 (1993).
neurotrophic factor (CNTF) on chick embryonic motoneurons in culture: 20. Li, M., Sendtner, M. & Smith, A. Essential function of LIF receptor in motor
comparison with other neurotrophic factors and cytokines. J. Neurosci. 10, neurons. Nature 378, 724727 (1995).
35073515 (1990). 21. Elson, G. C. et al. Cytokine-like factor-1, a novel soluble protein, shares
4. Oppenheim, R. W., Prevette, D., Qin-Wei, Y., Collins, F. & MacDonald, J. homology with members of the cytokine type I receptor family. J. Immunol.
Control of embryonic motoneuron survival in vivo by ciliary neurotrophic 161, 13711379 (1998).
factor. Science 251, 16161618 (1991). 22. Alexander, W. S. et al. Suckling defect in mice lacking the soluble
5. Curtis, R. et al. Retrograde axonal transport of ciliary neurotrophic factor is haemopoietin receptor NR6. Curr. Biol. 9, 605608 (1999).
increased by peripheral nerve injury. Nature 365, 253255 (1993). 23. Trinchieri, G. Interleukin-12: a proinflammatory cytokine with
6. Sendtner, M. et al. Ciliary neurotrophic factor prevents degeneration of immunoregulatory functions that bridge innate resistance and antigen-
motor neurons in mouse mutant progressive motor neuronopathy. Nature specific adaptive immunity. Annu. Rev. Immunol. 13, 251276 (1995).
358, 502504 (1992). 24. Kallen, K.-J. et al. Receptor recognition sites of cytokines are organized as
7. Mitsumoto, H. et al. Arrest of motor neurone disease in wobbler mice exchangeable modules. Transfer of the leukemia inhibitory factor receptor-
cotreated with CNTF and BDNF. Science 265, 11071110 (1994). binding site from ciliary neurotrophic factor to interleukin-6. J. Biol. Chem.
8. Helgren, M. E. et al. Trophic effect of ciliary neurotrophic factor on 274, 1185911867 (1999).
denervated skeletal muscle. Cell 76, 493504 (1994). 25. Robledo, O. et al. Signalling of cardiotrophin-1 receptor: evidence for a third
9. Guillet, C., Auguste, P., Mayo, W., Kreher, P. & Gascan, H. Ciliary receptor component. J. Biol. Chem. 272, 48554863 (1997).
neurotrophic factor is a regulator of muscular strength in aging. J. Neurosci. 26. Robledo, O. et al. Binding interactions of LIF and CNTF with the different
19, 12571262 (1999). subunits of their high affinity receptors. J. Neurochem. 66, 13911399
10. Shi, Y. et al. Computational EST database analysis identifies a novel member (1996).
of the neuropoietic cytokine family. Biochem. Biophys. Res. Comm. 262, 27. Zurn, A. D., Winkel, L., Menoud, A., Djabali, K. & Aebischer, P. Combined
132138 (1999). effects of GDNF, BDNF and CNTF on motoneuron differentiation in vitro.
11. Senaldi, G. et al. Novel neurotrophin-1/B cell stimulating factor-3: J. Neurosci. Res. 44, 133141 (1996).
a cytokine of the IL-6 family. Proc. Natl. Acad. Sci. USA 96, 1145811463 28. Anderson, K. D., Panayotatos, N., Corcoran, T. L., Lindsay, R. M. & Wiegand, S. J.
(1999). Ciliary neurotrophic factor protects striatal output neurons in an animal
12. Taga, T. & Kishimoto, T. Gp130 and the interleukin-6 family of cytokines. model of Huntington disease. Proc. Natl. Acad. Sci. USA 93, 73467351 (1996).
Annu. Rev. Immunol. 15, 797819 (1997). 29. Miller, R. G. et al. A placebo-controlled trial of recombinant human ciliary
13. Davis, S. et al. The receptor for ciliary neurotrophic factor. Science 253, 5963 neurotrophic (rhCNTF) factor in amyotrophic lateral sclerosis. RhCNTF ALS
(1991). Study Group. Ann. Neurol. 39, 256260 (1996).
14. Davis, S. et al. LIFR beta and gp130 as heterodimerizing signal transducers of 30. Penn, R. D., Kroin, J. S., York, M. M. & Cedarbaum, J. M. Intrathecal ciliary
the tripartite CNTF receptor. Science 260, 18051808 (1993). neurotrophic factor delivery for treatment of amyotrophic lateral sclerosis
15. Heinrich, P. C., Behrmann, I., Muller-Newen, G., Schaper, F. & Graeve, L. (phase I trial). Neurosurgery 40, 9499 (1997).
Interleukin-6 type cytokine signalling through the gp130/Jak/STAT pathway. 31. Aebischer, P. et al. Intrathecal delivery of CNTF using encapsulated
Biochem. J. 334, 297314 (1998). genetically modified xenogeneic cells in amyotrophic lateral sclerosis
16. Masu, Y. et al. Disruption of the CNTF gene results in motor neuron patients. Nature Med. 2, 696699 (1996).
degeneration. Nature 365, 2732 (1993). 32. Mittoux, V. et al. Restoration of cognitive and motor functions by ciliary
17. Takahashi, R. et al. A null mutation in the human CNTF gene is not causally neurotrophic factor in a primate model of Huntingtons disease. Hum. Gene
related to neurological diseases. Nat. Genet. 7, 7984 (1994). Ther. 11, 11771187 (2000).
18. DeChiara, T. M. et al. Mice lacking the CNTF receptor, unlike mice lacking 33. Arce, V. et al. Cardiotrophin-1 requires LIFRbeta to promote survival of
CNTF, exhibit profound motor neuron deficits at birth. Cell 83, 313322 mouse motoneurons purified by a novel technique. J. Neurosci. Res. 55,
(1995). 119126 (1999).

articles
Control of Mller glial cell

proliferation and activation
following retinal injury
Michael A. Dyer and Constance L. Cepko
Department of Genetics and Howard Hughes Medical Institute, Harvard Medical School, 200 Longwood Avenue, Boston, Massachusetts 02115, USA
Correspondence should be addressed to C.L.C. (cepko@rascal.med.harvard.edu)
Mller glial cells are the major support cell for neurons in the vertebrate retina. Following neuronal
damage, Mller cells undergo reactive gliosis, which is characterized by proliferation and changes in
gene expression. We have found that downregulation of the tumor supressor protein p27Kip1 and re-
entry into the cell cycle occurs within the first 24 hours after retinal injury. Shortly thereafter, Mller
glial cells upregulate genes typical of gliosis and then downregulate cyclin D3, in concert with an
exit from mitosis. Mice lacking p27Kip1 showed a constitutive form of reactive gliosis, which leads to
retinal dysplasia and vascular abnormalities reminiscent of diabetic retinopathy. We conclude that
p27Kip1 regulates Mller glial cell proliferation during reactive gliosis.
Glial cells are found ubiquitously throughout the central nervous E2F-regulated genes and entry into S-phase2123. Cyclin kinase
system (CNS), where their primary role is to maintain neuronal inhibitors (CKIs) of the Cip/Kip family regulate cell-cycle pro-
health. When CNS homeostasis is perturbed as a result of trauma, gression by forming a ternary complex with the cyclin/CDK com-
neurodegeneration, inflammatory disease or excitotoxicity, glial plex, thereby blocking phosphorylation of Rb24.
cells become activated and undergo reactive gliosis 16. This We have found that the p27Kip1 cyclin kinase inhibitor interacts
process is characterized by proliferation and changes in gene with cyclin D3 in Mller glial cells of the adult mouse retina. Fol-
expression that are believed to be important for the protection lowing retinal injury, p27Kip1 protein levels decrease, and Mller
or repair of neurons15. In addition to the benefits ascribed to glial cells re-enter the cell cycle and upregulate GFAP. Shortly
glialneuronal interactions, glial cell activation is proposed to be thereafter, cyclin D3 is downregulated. This latter event most
harmful under certain circumstances, and to contribute to dis- likely is what prevents deregulated glial cell proliferation follow-
ease progression1,4. ing acute or chronic retinal injury. The changes in p27Kip1 expres-
The vertebrate retina contains a specialized type of glia, the sion are not simply correlated with reactive gliosis. Examination
Mller glia, not found elsewhere in the CNS. Like other glial cells of a p27Kip1 knockout mouse revealed gratuitous Mller glial cell
of the CNS, Mller cells undergo reactive gliosis following acute activation and GFAP upregulation shortly after these cells were
retinal injury or chronic neuronal stress79. Mller cell gliosis is formed during retinal development. The p27Kip1-deficient mice
characterized by proliferation8, changes in cell shape due to alter- also exhibited retinal dysplasia, which is suggested25 to be due to
ations in intermediate filament production10, changes in ion outer limiting membrane disruptions that most likely result from
transport properties11, and secretion of signaling molecules such widespread glial cell activation during development, rather than
as vascular endothelial growth factor12,13. Ostensibly, gliosis is from photoreceptor proliferation. In addition, dramatic alter-
important for the protection and repair of retinal neurons, yet ations in the retinal vasculature were observed in retinae con-
some pathologies such as diabetic retinopathy may be exacer- taining Mller glial cells undergoing reactive gliosis. This finding
bated by reactive gliosis1113. Although a great deal is known about supports the hypothesis that Mller glial cell activation is impor-
the environmental factors that can induce glial cell activation and tant in the progression of diabetic retinopathy.
the changes in these cells during gliosis, very little is known about
how the process is regulated. We were particularly interested in RESULTS
the regulation of Mller glial cell proliferation and the changes Cyclin D3 and p27Kip1 are expressed in Mller glia
in cell shape that result from the rapid upregulation of glial fib- Mller glial cell bodies lie in a narrow band in the middle of the
rillary acidic protein (GFAP) following acute retinal injury inner nuclear layer of the adult retina (Fig. 1a). Their processes
because these are the earliest hallmarks of reactive gliosis8,14, and span all cellular and plexiform layers of the retina, forming
they have been documented in a variety of neuropathological microvilli at the apical surface26. We found that the p27Kip1 cyclin
conditions of the CNS1519. kinase inhibitor was expressed in a restricted row of nuclei in the
The molecular events that lead to cell cycle re-entry from a middle of the inner nuclear layer, consistent with Mller glial cell
quiescent G0 state are extensively characterized in tumor cells and localization (Fig. 1b). To directly demonstrate that Mller glia
fibroblasts2022. Phosphorylation of the retinoblastoma protein expressed p27Kip1, we did co-immunolocalization using an anti-
(Rb) or related family members by a cyclin/cyclin-dependent- body directed against cellular retinaldehyde-binding protein
kinase (CDK) complex leads to de-repression of transcription of (CRALBP; Fig. 1c). Nuclei that were immunoreactive for p27Kip1

articles
Fig. 1. Expression of p27Kip1 and cyclin D3 in the

adult mouse retina. (a) Drawing of a Mller glial cell
spanning all three nuclear layers of the vertebrate reti-
nae. Mller microvilli make up the outer limiting mem-
brane (arrow). (be) p27Kip1 and CRALBP
immunofluorescence in the adult retina. (b) An anti-
body specific for p27Kip1 stains a uniform row of nuclei
in the inner nuclear layer whose shape and position
are consistent with Mller glial cells. (c) CRALBP is
expressed in the cytoplasm of Mller glial cells.
(d) The p27Kip1 immunoreactive nuclei precisely local-
ize to the CRALBP immunoreactive Mller glial cell
bodies. (e) A high-magnification view of CRALBP and
p27Kip1 colocalization. (fh) p27Kip1 and cyclin D3
immunofluorescence in the adult retina. (f) A high-
magnification view of the p27Kip1 immunoreactive
nuclei of Mller glial cells. The same field of cells
shown in (f) also exhibit cyclin D3 immunoreactivity
(g), and these two proteins colocalize to the nuclei of
Mller glial cells (h). (i) Co-immunoprecipitation of
p27Kip1 with cyclin D3. Cyclin D3 was immunoprecipi-

tated from crude retinal lysate. The starting material (lane 1), washes (lanes 24), control IgG immunoprecipitation (lane 5) and bound fractions (lane 6)
were separted by SDSPAGE and immunoblotted. The starting material contained p27Kip1, and it co-immunoprecipitated with cyclin D3 (arrow).
CRALBP, cellular retinaldehyde binding protein; OS, photoreceptor outer segments; olm, outer limiting membrane; ONL, outer nuclear layer; OPL,
outer plexiform layer; INL, inner nuclear layer; IPL, inner plexiform layer; GCL, ganglion cell layer; Mr, protein markers, relative molecular mass from
bottom, 26, 36, 42, 66, 97, 116, 158 kDa. Scale bars, 100 m (bd), 10 m (eh).
precisely colocalized within the CRALBP-immunoreactive Mller ed with the injected neurotoxins, GFAP immunoreactive Mller
glial cell bodies (Fig. 1d and e). Cyclin D3 is expressed in the glial cells were occasionally found to enter S-phase and incorpo-
nuclei of Mller glial cells (C. Ma and C.L.C., unpublished data). rate BrdU (Fig. 2e).
Co-immunolocalization using antibodies directed against cyclin To quantify the effects of such treatment on glial cell acti-
D3 and p27Kip1 revealed that Mller glial cell nuclei contain both vation, we induced reactive gliosis in vitro and labeled cells
proteins (Fig. 1fh). with [3H]thymidine. Adult retinae were explant cultured for
Biochemical studies indicate that cyclin kinase inhibitors can 48 hours in the presence or absence of ouabain (Methods).
form stable complexes with cyclin/CDK dimers in vitro and in After the first 24 hours in culture, [3H]thymidine was added
cultured cells24,27,28. To determine whether cyclin D3 and p27Kip1 to label mitotic cells. Retinae were then dissociated, stained
are part of a complex in Mller glial cells, we immunoprecipi- with antibodies specific for cell types or cell cycle proteins, and
tated cyclin D3 from an adult retinal lysate. Proteins that co- processed for autoradiography. Although the magnitude of
immunoprecipitated with cyclin D3 were separated by induction of gliosis as measured by the proportion of GFAP-
SDSpolyacrylamide gel electrophoresis (PAGE) and immunoreactive cells was similar to that observed in vivo
immunoblotted with an antibody directed against p27Kip1. Sig- (1015 fold; Table 1), the proportion of GFAP-immunoreac-
nificant amounts of p27Kip1 co-immunoprecipitated with cyclin tive cells in the untreated samples was slightly higher in vitro
D3 from adult retinae (Fig. 1i). (0.4% as compared to 0.1% in vivo). All cells that incorporat-
ed [3H]thymidine were CRALBP-immunoreactive Mller glial
p27Kip1 is downregulated following retinal injury cells (Fig. 2f). We never observed [3H]thymidine-labeled rod
To test whether the expression levels of p27Kip1 and/or cyclin D3 photoreceptors (anti-rhodopsin), bipolar interneurons (anti-
changed during the proliferation response associated with reac- Chx10) or horizontal/amacrine cells (anti-syntaxin-1; data not
tive gliosis in the retina, we examined the expression of these pro- shown). Significantly, all the [ 3 H]thymidine-labeled cells
teins following introduction of the neurotoxins ouabain and expressed GFAP, as expected for reactive gliosis (Fig. 2g; Table 1).
domoic acid. These agents induce widespread glial cell activation Among the GFAP-immunoreactive glial cells, only a subset
across the retina8. Adult mice received an intraocular injection (5 0.3%, 13 of 256) entered S-phase in the last 24-hour peri-
of ouabain, domoic acid or saline (PBS)8. Retinae were harvest- od of culture in the presence of ouabain (Table 1). These data
ed 48 hours after injection and stained with antibodies against are consistent with previous in vivo analyses8. To determine
Mller glial cell markers. Half the retinal tissue from each eye whether p27Kip1 downregulation correlated with the activation
was dissociated for single-cell analysis, and the other half was of glial cell proliferation, we examined the expression of p27Kip1
embedded for cryosectioning. Retinae from eyes that received an in [3H]thymidine-labeled Mller glial cells (Fig. 2h). From sev-
intraocular injection of PBS showed no glial cell activation, as eral independent retinae, we found no [3H]thymidine-labeled
indicated by the very small amount of GFAP immunoreactivity in cells among 37 cells that expressed p27Kip1 (Table 1). In addi-
Mller glia (Fig. 2a) and the absence of detectable proliferation tion, p27Kip1-immunoreactive cells, which were readily detect-
(data not shown and below). In contrast, retinae exposed to ed (Fig. 2i), never incorporated [3H]thymidine. When cyclin
ouabain or domoic acid exhibited a dramatic increase (1020 fold) D3 expression was examined, among the [ 3 H]thymidine-
in the proportion of GFAP-immunoreactive cells (Fig. 2b and c; labeled Mller glial cells, a small number (5 of 31, 16%) were
Table 1). Cells upregulating GFAP appeared to downregulate positive (Fig. 2j) but a significant proportion had downregu-
p27Kip1 (Fig. 2d). When bromo-deoxyuridine (BrdU) was includ- lated cyclin D3 (26 of 31, 84%).

articles
Table 1. Analysis of protein expression in proliferating Mller glial cells following retinal injury.
Treatment CRALBP+/total GFAP+/total GFAP+,[3H]thy+/GFAP+ p27Kip1+,[3H]thy+/[3H]thy+
(counts; mean % s.d.)1 (counts; mean % s.d.) (counts; mean % s .d.)2 (counts; mean % s.d.)3
Ouabain 32/500, 44/500, 36/500, 21/500, 24/500, 21/500, n.a.4 n.a.
(in vivo) 47/500 18/500
(7.9 1.4) (4.2 0.5)
PBS 27/500, 34/500, 28/500, 1/500, 1/500, 0/500, n.a. n.a.

(in vivo) 26/500 0/500
(5.7 0.7) (0.1 0.1)
Ouabain 23/500, 29/500, 31/500 35/500, 37/500, 38/500 8/147, 5/109 0/20, 0/17
(in vitro) (5.5 0.8) (7.2 0.4) (5.2 0.3) (0)
PBS 22/500, 19/500, 35/500 3/500, 4/500, 2/500 0/9, 0/11 n.a.5
(in vitro) (5.1 1.7) (0.6 0.2) (0)
1Adult (6 week old) mice received a 0.51.0 microliter injection of ouabain in the left eye and PBS in the right eye. Forty-eight hours after injection, retinae
were harvested, dispersed, plated and stained by immunohistochemistry. For in vitro analysis, freshly dissected retinae were explant cultured in the presence
or absence of ouabain for 48 hours. Four independent animals were analyzed in vivo, and three animals were analyzed in vitro. 2Due to the relative scarcity of
GFAP-immunoreactive cells, the fields of dispersed cells were searched for GFAP immunoreactivity and then scored with respect to [3H]thymidine incorpora-
tion. Therefore, the total number of cells scored is not presented but can be inferred from the total proportion of GFAP-immunoreactive cells. The number
of silver grains in labeled cells varied from 23 to 65 (mean, 34 17) and the number of silver grains in unlabeled cells ranged from 0 to 6 (mean, 2 1.4). 3Due
to the sparse distribution of [3H]thymidine-labeled cells, the fields of dispersed cells were searched for [3H]thymidine labeling and then scored for p27Kip1
immunoreactivity. Therefore, the total number of cells scored is not presented. 4Data on the [3H]thymidine labeling of cells from in vivo treatment is not avail-
able because [3H]thymidine was not injected into the eyes of adult animals. 5Data on the [3H]thymidine labeling of cells from in vitro treatment with PBS is not
available because adult retinal cells do not proliferate.
To delineate the order of the changes in p27Kip1, GFAP and ately or following a culture period of 4, 8, 16, 24 or 48 hours in
cyclin D3 expression in proliferating Mller glial cells during the absence of [3H]thymidine. Immediately after [3H]thymidine
reactive gliosis, we did a [3H]thymidine pulse-labeling experi- labeling, all [ 3H]thymidine-labeled Mller glial cells lacked
ment. Retinae were cultured for 24 hours in the presence of expression of p27Kip1 and maintained expression of cyclin D3
ouabain, and then labeled for 2 hours with [3H]thymidine to (Table 1). Few of these cells (1 of 29, 2%) had upregulated GFAP
mark cells in S-phase. Tissue was then either harvested immedi- (Fig. 2k). By 4 hours after labeling, GFAP immunoreactivity was
Fig. 2. Changes in gene expression in Mller glial cells following reti-

nal injury. (ae) Analysis of GFAP expression, p27Kip1 expression, and
BrdU incorporation following injection of neurotoxins. (a) Forty-
eight hours after an intraocular injection of 0.5 microliters of PBS,
GFAP expression was only detected in the rare astrocytes (As) along
the vitreal surface of the retina. (b) GFAP immunoreactivity was
observed in Mller glial cells (M) after neuronal toxicity brought on
by intraocular injection of ouabain or domoic acid (not shown). (c) A
high-magnification view of a GFAP-immunoreactive Mller glial cell.
(d) Immunohistochemical staining with antibodies directed against
GFAP (red fluorescence) and p27Kip1 (arrow, green fluorescence)
indicated that the Mller glial cells that expressed GFAP had down-
regulated p27Kip1. (e) When BrdU was injected along with ouabain,
ocassionally a GFAP immunoreactive (red fluorescence) Mller cell
had entered S-phase as measured by BrdU immunofluorescence
(arrow, green fluorescence). (fi) Analysis of the gene expression
kinetics in explanted retinae treated with neurotoxins. Gene expres-
sion was assayed by single-cell immunofluorescence (FITC).
Proliferation was measured by the uptake of [3H]thymidine (bright
field) in the nuclei (DAPI) of individual cells expressing a given protein.
(f) CRALBP immunoreactive Mller glial cells re-entered the cell cycle k
Percent [3H]+, immunopositive
after treatment with ouabain. (g) Mller glial cells that re-entered the
cell cycle upregulated GFAP. Mller glial cells that re-entered the cell
cycle never expressed p27Kip1 (h), and those that continued to
express p27Kip1 were never found to incorporate [3H]thymidine (i).
(j) Forty-eight hours after treatment with ouabain, a small fraction of
mitotic Mller glial cells expressed cyclin D3. (k) Kinetics of cyclin D3
and GFAP expression following S-phase entry. Immediately after pulse
labeling with [3H]thymidine, most of the cells that incorporated
[3H]thymidine expressed cyclin D3 and showed no GFAP expression.
Several hours later, the fraction of cells expressing cyclin D3 was
reduced, whereas the proportion of cells expressing GFAP increased. Time after labeling (hours)
PBS, phosphate buffered saline; GFAP, glial fibrillary acidic protein; INL, inner nuclear layer; ONL, outer nuclear layer; GCL, ganglion cell layer;
CRALBP, cellular retinaldehyde binding protein; BrdU, bromo-deoxyuridine. Scale bars, 100 m (ad), 10 m (ei).

articles
Fig. 3. Characterization of Mller glial cells in the retinae of p27Kip1-

deficient mice. (a, b) GFAP immunoflurescence in p27/ and p27+/ reti-
nae. Retinae were harvested from six-week-old mice lacking the p27Kip1
gene or their heterozygous littermates, and cryosections were incu-
bated with an antibody specific for GFAP. (a) Extensive GFAP
immunoreactivity was observed in Mller glial cells of the p27Kip1-defi-
cient retinae. (b) Heterozygous and wild-type (data not shown) litter-
mates had no GFAP immunoreactivity beyond the astrocyte staining
observed normally along the vitreal surface. (c, d) Single-cell staining for
Mller glial cell markers. (c) A single GFAP-immunoreactive Mller glial
cell from a p27Kip1-deficient retina, which maintained its distinctive mor-
phology even after the tissue was dissociated. (d) Retinae from p27Kip1-
Percent immunopositive cells

deficient, heterozygous or wild-type littermates were dispersed and
plated, and individual cells were examined for the expression of
CRALBP or GFAP. The proportion of Mller glial cells was similar in
mice lacking p27Kip1 as measured by CRALBP immunoreactivity.
However, the proportion of GFAP-immunoreactive Mller glial cells was
substantially increased in the p27Kip1-deficient retinae. Each bar repre-
sents the average of 1000 cells scored for 24 independent retinae.
*Some samples had so few GFAP immunoreactive cells (01) that the
bars are not visible on the graph. (eg) Cylin D3 expression in the reti-
nae of p27Kip1-deficient mice. (e) Retinae from wild-type mice show the
normal pattern of Mller glial nuclear immunoreactivity.
(f) Heterozygous littermates exhibited a slight disorganization of the
Mller glial cell layer, but cyclin D3 expression persisted. (g) Mller glial
cells in the retinae from mice lacking p27Kip1 lacked cyclin D3 expres-
sion. GFAP, glial fibrillary acidic protein; CRALBP, cellular retinaldehyde
binding protein; ONL, outer nuclear layer; INL, inner nuclear layer;
GCL, ganglion cell layer. Scale bars, 100 m (a, b, eg), 10 m (c).
detected in significantly more [3H]thymidine-labeled Mller glial testing this hypothesis, we examined the retinal dysplasia in
cells (14 of 30,47%), and this proportion reached a maximum greater detail using antibodies to specific retinal cell types
24 hours after labeling (Fig. 2k). During this same time interval, and/or structures.
cyclin D3 was downregulated in Mller glia that were in S-phase CD44 is expressed at high levels in the apical microvilli of
at the time of labeling (Fig. 2k). Mller glial cells, which make up the outer limiting membrane
of the retina31. Using an antibody directed against CD44, we
Reactive gliosis in the retinae of p27Kip1-deficient mice found the outer limiting membrane to be disrupted in retinae
To determine if loss of p27Kip1 expression was sufficient for Mller from p27Kip1-deficient mice at the sites of retinal dysplasia (Fig. 4e
glial cell induction, we examined retinae from p27Kip1-deficient and i). In each region where the outer limiting membrane was
mice25,29,30. Retinae from adult p27Kip1/ mice and their het- disrupted, the processes of activated (GFAP immunoreactive)
erozygous or wild-type littermates were dispersed or cryosec- Mller glial cells were present and, in many cases, extended
tioned and used for immunohistochemical studies. beyond the boundary of the outer limiting membrane (Fig. 4f
GFAP-immunoreactive glial cells were found throughout the reti- and j). By using antibodies directed against a variety of cell-type-
nae of p27Kip1-deficient mice (Fig. 3a), but not in their wild-type specific markers, we found that most of the cells that made up
or heterozygous littermates (Fig. 3b; data not shown). Quanti- the retinal dysplasia were rhodopsin-immunoreactive rod pho-
tative analysis of dispersed retinal cells revealed a 1020-fold toreceptors (Fig. 4g and k). Occasionally, a calbindin/neurofila-
induction of GFAP immunoreactive cells with distinctive Mller ment-immunoreactive horizontal cell was found outside the outer
glial cell morphology (Fig. 3c and d). To determine whether cyclin limiting membrane (Fig. 4h and l), displaced from its normal
D3 was downregulated in the retinae from p27Kip1-deficient mice, position at the INL/ONL boundary.
we did immunohistochemical staining using an antibody direct- To test the possibility that retinal dysplasia was the result of
ed against cyclin D3. Mller glial cells from wild-type and ectopic proliferation as proposed25, we did in vivo lineage analysis
p27Kip1+/ retinae expressed cyclin D3 (Fig. 3e and f), whereas reti- on the knockout retinae. If retinal dysplasia resulted from dereg-
nae from the p27Kip1-deficient mice showed no detectable cyclin ulated proliferation, we would expect to see abnormally large
D3 expression (Fig. 3g). These results are similar to the data clones associated with the dysplastic lesions in p27Kip1/ retinae.
obtained on retinae treated with neurotoxins (above). Newborn mouse pups from a mating of p27Kip1+/ animals received
intraocular injections of the replication-incompetent retrovirus
Reactive gliosis leads to retinal dysplasia LIA encoding alkaline phosphatase32. Following complete retinal
One of the hallmarks of retinae from p27Kip1-deficient mice is development (3 weeks), retinae were harvested, stained for alka-
dysplasia (Fig. 4ad). The presence of p27Kip1 and cyclin D3 in line phosphatase expression and sectioned, and clones of cells
Mller glial cells, combined with the suggestion that the reti- derived from single infected progenitor cells were scored for cell
nae from mice lacking p27Kip1 have defects in their outer limit- number and cell type. In the p27Kip1-deficient retinae, 17 clones
ing membranes25, led us to test whether inappropriate Mller were found within or immediately adjacent to regions where reti-
glial cell activation during development could have caused the nal dysplasia was apparent. Most of these were single-cell (10 of
retinal dysplasia of p27Kip1-deficient mice. As a first step toward 17; Fig. 4m) or two-cell (6 of 17; Fig. 4n) clones; none of them

articles
Fig. 4. Characterization of
the retinal dysplasia found
in p27Kip1-deficient mice.
(ad) Retinal dysplasia in
p27Kip1-deficient retinae. DAPI
staining (a) and DIC
microscopy (b) reveal cell
bodies displaced beyond the
outer limiting membrane into
the region normally occupied
by photoreceptor outer seg-
ments. (c, d) DAPI staining
and DIC microscopy of a
retina from the wild-type lit-
termate to the p27Kip1-defi-
cient animal shown above.
(el) Detailed characteriza-
tion of retinal dysplasia found
in p27Kip1-deficient mice.
Photographs (eh) and trac-
ings (il) of immunostaining

using antibodies to specific
retinal structures or cell types.
(e, i) The outer limiting mem-
brane was disrupted in regions
of retinal dysplasia (arrow) as
indicated by an alteration in
the CD44 immunoreactivity. (f, j) GFAP-immunoreactive Mller glial cell processes were found wherever the outer limiting membrane was disrupted
and retinal dysplasia formed. (g, k) Most cells that breeched the outer limiting membrane were rhodopsin-immunoreactive photoreceptors.
(h, l) Occasionally a neurofilament/calbindin-immunoreactive horizontal cell was found displaced from its normal position at the ONL/INL boundary to the
outer segment layer. (m, n) In vivo lineage analysis in p27Kip1-deficient retinae. Replication-incompetent retroviruses were injected into the eyes of newborn
mice from a cross of p27Kip1 heterozygous parents. (m) A single alkaline-phosphatase-expressing cell in the region of retinal dysplasia (arrows) of the p27Kip1-
deficient mouse. (n) An example of a two-cell clone containing a bipolar interneuron in the INL. Dashed lines indicate the boundary between the ONL and
photoreceptor outer segments. DIC, differential interference contrast; OS, photoreceptor outer segments; OLM, outer limiting membrane; ONL, outer
nuclear layer; INL, inner nuclear layer; GFAP, glial fibrillary acidic protein; Rho4D2, anti-rhodopsin antibody. Scale bars, 50 m (ad, m,n), 20 m (el).
contained more than three cells. Approximately 100 clones for seen in the p27Kip1-deficient retinae (M.A.D. and C.L.C., unpub-
each group (p27Kip1/, p27Kip1+/, p27Kip1+/+) were found through- lished data). Two weeks after the injection, when retinal devel-
out the retina (data not shown). The average size and composi- opment was complete, retinae were harvested, sectioned and
tion of clones from mice lacking one or both alleles of p27Kip1 were stained with antibodies to GFAP, p27Kip1, cyclin D3 and the outer
similar to those identified in their wild-type littermates (data not limiting membrane marker, CD44. We observed retinal dysplasia
shown) and was consistent with previous data33,34. in the retinae treated with ouabain that was similar to that found
To test whether the induction of reactive gliosis during devel- in p27Kip1-deficient mice (Fig. 5). Retinae from eyes that received
opment of wild-type animals can lead to retinal dysplasia, and an intra-ocular injection of saline never showed retinal dyspla-
whether there is a correlation with an alteration in the outer lim- sia (data not shown). In the dysplastic regions of ouabain-inject-
iting membrane, we administered an intraocular dose of ouabain ed eyes, CD44 immunoreactivity was absent, indicating
to mouse pups at postnatal day 10.5. This is the stage when disruption of the outer limiting membrane (Fig. 5a and e). Fur-
Mller glial cells are differentiating and retinal dysplasia is first thermore, Mller cells initiated a program of reactive gliosis at
Fig. 5. Induction of reactive gliosis during reti-
nal development. Intraocular injection of
oubain at postnatal day 10.5 led to retinal dys-
plasia (arrows delineate the boundaries) simi-
lar to that found in the retinae from
p27Kip1-deficient mice. (ad) Photographs of
immunostaining using antibodies to specific
retinal structures or Mller glial cell markers in
the regions of retinal dysplasia (eh). (a) A
decrease in CD44 immunofluorescence indi-
cates disruption of the outer limiting mem-
brane at the site of retinal dysplasia following
intra-ocular injection of ouabain at P10.5.
(b) GFAP immunofluorescence was also con-
centrated at the sites of retinal dysplasia.
(c, d) Both p27Kip1 (c) and cyclin D3 (d) were
downregulated in the nuclei of Mller cells undergoing reactive gliosis (open arrowheads). olm, outer limiting membrane; ONL, outer nuclear layer;
INL, inner nuclear layer; GCL, ganglion cell layer; GFAP, glial fibrillary acidic protein. Scale bars, 100 m.

articles
Fig. 6. Examination of the vasculature in retinae under-

going reactive gliosis. (ac) Normal vascular structure in
the adult retina. (a) Transverse section of a normal retina
stained with an anti-PECAM-1 antibody. Vascular
endothelial structures are indicated by arrows. (b) These
vessels are found primarily in the INL, the plexiform lay-
ers, and adjacent to the ganglion cell layer of the verte-
brate retina as shown by this DIC image of the same
retinal section shown in (a). (c) Whole-mount staining
for PECAM-1 in a normal adult retina photographed
from the vitreal side of the tissue. The optic nerve is
approximately 1 mm to the left. (df) PECAM-1 staining
on retinae following an intra-ocular injection of ouabain
at P10.5. (d) PECAM-1 immunoreactivity was more
extensive in these retinae. (e) DIC image of the retinal
section shown in (d). (f) Whole-mount staining verified
the increases in retinal vasculature. As in (c), the optic
nerve lies approximately 1 mm to the left. (gi) PECAM-1
staining on retinae of p27Kip1-deficient mice. (g) The reti-
nae from six-week-old p27Kip1-deficient mice showed
extensive vasodialation and neovascularization. The

outer nuclear layer is not visible in this photograph.
(h) High-magnification view of a particularly large
PECAM-1 immunoreactive vessel, which contained large
numbers of red blood cells (i) within the vessel and in the
surrounding tissue (arrows). ONL, outer nuclear layer;
opl, outer plexiform layer; ipl, inner plexiform layer; INL,
inner nuclear layer; GCL, ganglion cell layer. Scale bars,
100 m (a, b, d, e, g, h), 0.5 mm (c, f), 10 m (i).
the sites of retinal dysplasia; GFAP was upregulated (Fig. 5b and retinal injury, Mller glial cells rapidly downregulated p27Kip1
f), p27Kip1 was downregulated (Fig. 5c and g), and cyclin D3 was protein levels and re-entered the cell cycle. Shortly thereafter,
downregulated (Fig. 5d and h). These molecular events precise- the levels of the intermediate filament protein, GFAP, were
ly mimic those observed in p27Kip1-deficient retinae and in adult increased. Subsequently, additional rounds of cell division were
retinae treated with neurotoxins. not observed, most likely because of the observed downregu-
lation of cyclin D3 protein in the activated Mller glial cells.
Vascular defects associated with reactive gliosis The causal role of p27Kip1 in reactive gliosis was demonstrated by
Previous work on human retinopathies associated with diabetes observations on p27Kip1-deficient mice. Retinae from mice lack-
and on animal models suggested that Mller glial cells are involved ing p27Kip1 exhibited reactive gliosis throughout the retina. We
in the very earliest steps of diabetic retinopathy12,13,35,36. To test propose that the retinal dysplasia reported previously in
whether glial cell activation might lead to alterations in retinal vas- p27Kip1/ animals results from disruptions in the outer limiting
culature, we did immunohistochemical staining using an antibody membrane due to reactive gliosis during development. We also
directed against the vascular endothelial cell marker, PECAM-1. found alterations in the vasculature of retinae from p27Kip1-defi-
Normally, the retinal vasculature is found in the plexiform layers, cient mice or retinae of wild-type mice treated with excitatory
the GCL and INL, and along the vitreal surface of the retina. Trans- amino acids. Taken together, our experiments indicate that the
verse sections through these structures generally give cross sections regulation of Mller glial cell proliferation occurs via p27Kip1,
of the retinal vasculature with branched vessels sometimes appear- and that cyclin D3 regulation ultimately contributes to keep-
ing parallel to the plane of the section in the inner nuclear layer ing Mller cell proliferation in check. In addition, these data
(INL; Fig. 6ac). Two major alterations in the vascular structures of show the crucial role of Mller cells in the formation of retinal
wild-type retinae with reactive gliosis or p27Kip1-deficient retinae dysplasia and support the notion that Mller cells are key reg-
were observed. First, more extensive INL vascularization was seen ulators of retinal vasculature.
in regions of retinal dysplasia and reactive gliosis (Fig. 6dg). Sec- We found that p27Kip1 downregulation is the earliest indicator
ond, an enlargement of the vessels was observed in retinae from of Mller glial cell activation identified to date. GFAP is upregulat-
p27Kip1-deficient mice, and blood cells were often seen outside of ed quickly in Mller glial cells following retinal injury8,9. Here we
these structures (Fig. 6h and i and data not shown). report that p27Kip1 downregulation preceeds GFAP upregulation.
Why might p27Kip1 regulate both proliferation and GFAP levels,
DISCUSSION with proliferation preceeding GFAP upregulation? Intermediate fil-
Nearly every major disease of the retina, including retinitis pig- aments are broken down and reassembled when cells undergo mito-
mentosa14,37,38, macular degeneration3942 and diabetic retinopa- sis43. It is possible that Mller glial cells coordinate mitosis with
thy13,35,36, is associated with reactive gliosis involving Mller GFAP upregulation to efficiently incorporate these molecules into
glial cells. Despite the importance of this association, no genes the cytoskeleton during reassembly following cytokinesis, perhaps
that regulate this process have been reported. Our experiments explaining why cell cycle re-entry precedes GFAP upregulation.
demonstrate that the level of p27Kip1 expression in Mller glial Although proliferation occurs during reactive gliosis, glial
cells is critical for regulation of reactive gliosis. Following acute tumors are not common, implying that some mechanism(s) to

articles
limit proliferation must be present. We found that cyclin D3 relies on an intact outer limiting membrane during develop-
was downregulated approximately 1624 hours after Mller ment. Perturbations in this structure result in retinal dyspla-
glial cells initiated reactive gliosis. Without an interacting cyclin, sia and a significant reduction in phototransduction25.
the cyclin-dependent kinase(s) in Mller glial cells would be Mller glial cells undergoing reactive gliosis are believed to
unable to phosphorylate the retinoblastoma protein (or its relat- alter the local environment for neurons and to provide additional
ed family members), and Mller glial cells would enter a qui- structural integrity to the retina at the site of injury by the afore-
escent G 0 state 24 . It is intriguing that cyclin D3 was mentioned changes in intermediate filament constituents. How-
downregulated following cell division rather than p27Kip1 lev- ever, when reactive gliosis is deregulated to the point where it
els returning to normal. Presumably, the signals from damaged results in so-called massive gliosis, or under conditions of chron-
neurons that led to Mller glial cell activation and p27Kip1 down- ic Mller glial cell activation such as ischemia associated with
regulation in the first place were still present after the first diabetes, severe neuronal damage may lead to blindness. In the
48 hours in culture. Therefore, perhaps to prevent further cell case of diabetic retinopathy, the prolonged secretion of vascular
division, cyclin D3 was downregulated. endothelial growth factor, a potent vasodialator and vascular
If p27Kip1 levels determine the entry of Mller cells into reac- endothelial mitogen, by Mller glial cells could contribute to the
tive gliosis, as opposed to simply being correlated with it, then progression of this debilitating disease12,13. Evidence from the
the p27Kip1-deficient retinae should exhibit reactive gliosis with- retinae of p27Kip1-deficient mice and injured retinae lend sup-
out induction by any of the known stimulants of reactive gliosis. port to this hypothesis. Our initial characterization suggests that
This was indeed the case. At postnatal day 10.5 when Mller glial Mller glial cell activation results in vasodilation and vessel
cells are normally differentiating, we observed inappropriate S- growth or reorganization. These results demonstrate that Mller
phase entry and apoptosis in p27Kip1-deficient retinae (M.A.D. glial cell activation per se (targeted disruption of p27Kip1 and exci-
and C.L.C., unpublished data). Notably, GFAP expression in tatoxicity) can lead to vascular changes in the vertebrate retina.
p27Kip1/ Mller glial cells was detected as early as P10.5 and per- These findings may be of considerable clinical relevance if the
sisted into the adult stages44 (M.A.D. and C.L.C., unpublished same molecules are found to regulate Mller glial cell prolifera-
data). However, proliferation ceased shortly after Mller glia were tion and reactive gliosis in the human retina.
formed, in concert with the downregulation of cyclin D3 in the
Mller glial cells of the p27Kip1-deficient mice (M.A.D. and C.L.C.,
unpublished data). We did not observe an increase in the total METHODS
number of Mller glial cells as measured by CRALBP immunore- Animals. C57BL/6 and ICR mice were purchased from Taconic Farms
(Germantown, New York). The p27Kip1-deficient mice30 were crossed to
activity in the adult retina (Fig. 3d), which may reflect limited
ICR or C57BL/6 mice with equivalent results. Genotypes were deter-
proliferation or elimination by apoptosis in the p27Kip1-deficient mined by PCR amplification of the wild-type and mutant alleles from
retina (M.A.D. and C.L.C., unpublished data). Overall, the results tail DNA30.
from p27Kip1-deficient mice are consistent with a model in which
p27Kip1 levels control the initiation of reactive gliosis in the retina, Immunohistochemistry and antibodies. Immunohistochemical stain-
and, most likely, cyclin D3 downregulation prevents uncontrolled ing of retinal cryosections or dissociated cells was done as described46.
Mller cell proliferation. Many of the antibodies used for these studies have been described46.
In the p27 Kip1/ retina, dysplasia does not seem to result Other antibodies were anti-CD31 (PECAM-1), MEC 13.3 (rat mono-
from photoreceptor proliferation as previously reported25. The clonal, 1:100; Pharmingen) and anti-CD44, 5D2-27(rat monoclonal,
hypothesis of photoreceptor proliferation was based on 1:100; Developmental Studies Hybridoma Bank). Detailed protocols can
also be found at http://axon.med.harvard.edu/cepko/protocol/mike/.
immunohistochemical staining using an antibody specific for
p27Kip1, which demonstrated immunoreactivity in the outer [3H] thymidine labeling, retinal explant cultures and dissociation. To
segments of photoreceptors25. Restriction to outer segments is label retinal progenitor cells in S-phase, we incubated retinae in 1 ml
not consistent with the identification of a putative nuclear explant culture medium46 containing [3H] thymidine (NEN, 5 Ci/ml;
localization signal near the carboxy terminus of p27 Kip1 89 Ci/mmol) for 1 hour at 37C. Autoradiography was done as
(ref. 45), and we could find no evidence for p27Kip1 expression described47. The procedure for explant culturing of retinae has been
in the nuclei or outer segments of photoreceptors. Therefore, it described46 and is available at http://axon.med.harvard.edu/cepko/pro-
is likely that the previously reported expression of p27Kip1 in tocol/mike/. Extensive characterization has demonstrated that retinal
photoreceptor outer segments was the result of non-specific proliferation and differentiation are normal using this explant culture
system46. Tissue dissociation was done as described47.
antibody binding to the outer segments, which is a common
artifact of retinal immunostaining. Beyond this disparity in Replication-incompetent retroviral vector constructs and viral pro-
the expression pattern, in vivo lineage analysis, BrdU labeling of duction. To prepare high-titer retroviral stocks, the plasmid construct
adult retinae, and examination of dysplastic regions in older pLIA-E (M.A.D. & C.L.C., unpublished data), which is similar to pLIA48,
animals argue against deregulated photoreceptor cell prolifer- was transiently transfected into a 293T ecotropic producer cell line
ation leading to the observed retinal dysplasia in p27Kip1-defi- (Phoenix-E) by calcium phosphate co-precipitation as described48 .
cient mice. Instead, we found that when reactive gliosis was Supernatant containing the viral particles was harvested at 48 hours after
induced by administration of an intraocular injection of transfection, and viral titer was determined on NIH-3T3 cells48. In vivo
ouabain before the completion of retinal maturation (P10.5), lineage analysis was done as described33,34.
we observed retinal dysplasia that was very similar to that seen
Co-immunoprecipitation and immunoblotting. For cyclin D3 co-
in retinae from p27 Kip1 -deficient mice. It is likely that the
immunoprecipitation, 10 adult (3 week) mouse retinae were briefly son-
changes in the shape and structure of Mller glial cells associ- icated in 2 ml RIPA buffer (PBS, 1% NP-40, 0.5% sodium deoxycholate,
ated with alterations in intermediate filament components 0.1% SDS) containing 1 mM PMSF, a cocktail of protease inhibitors
(GFAP) led to a disruption in the outer limiting membrane (Sigma), and phosphatase inhibitors (1 mM Levamisole, 2 mM Na2VO3,
and subsequent dysplasia in that region. Furthermore, these 1 mM NaF). Anti-cyclin D3 (C-16, rabbit polyclonal, Santa Cruz Biotech,
data indicate that the structural integrity of the mature retina Santa Cruz, California) antibody was incubated with the crude retinal

articles
lysate and immunoprecipitated with protein-G agarose according to the 16. Nork, T. M., Ghobrial, M. W., Peyman, G. A. & Tso, M. O. Massive retinal
manufacturers instructions (Santa Cruz Biotech). SDSPAGE and gliosis. A reactive proliferation of Muller cells. Arch. Ophthalmol. 104,
immunoblotting were done as described46. 13831389 (1986).
17. Cogan, D. G. Congenital anomalies of the retina. Birth Defects Orig. Artic. Ser.
7, 4151 (1971).
Induction of reactive gliosis. For in vitro induction of reactive gliosis, 18. Berger, B., Peyman, G. A., Juarez, C., Mason, G. & Raichand, M. Massive
retinae from mature (3 week old) mice (c57BL/6) were dissected away retinal gliosis simulating choroidal melanoma. Can. J. Ophthalmol. 14,
from the surrounding tissue, and the lens was removed. Only retinae 285290 (1979).
19. Dithmar, S., Holz, F. G. & Volcker, H. E. [Massive reactive gliosis of the
lacking holes or tears, which could result in reactive gliosis, were used retina.] Klin. Monatsbl. Augenheilkd 211, 338341 (1997).
for this procedure. Retinae were then explant cultured46 in the pres- 20. Elledge, S. J. Cell cycle checkpoints: preventing an identity crisis. Science 274,
ence of domoic acid (Sigma, 20 M) or ouabain (Sigma, 70 M) for 16641672 (1996).
48 hours49. For analysis of mitotic cells, [3H]thymidine (5 Ci/ml) was 21. Sherr, C. J. & Roberts, J. M. Inhibitors of mammalian G1 cyclin-dependent
added after the initial 24 hours in culture. Following this culture peri- kinases. Genes Dev. 9, 11491163 (1995).
22. Sanchez, I. & Dynlacht, B. D. Transcriptional control of the cell cycle. Curr.
od, retinae were processed for cryosectioning, dissociation and/or Opin. Cell Biol. 8, 318324 (1996).
autoradiography as described above. For in vivo induction of reactive 23. Yee, A. S., Shih, H. H. & Tevosian, S. G. New perspectives on retinoblastoma
gliosis, 0.51.0 microliters of neurotoxins (domoic acid 2 mM, ouabain family functions in differentiation. Front. Biosci. 3, D532547 (1998).
7 mM) were injected into the eyes of 36 week-old C57BL/6 mice anes- 24. Hengst, L. & Reed, S. I. Inhibitors of the Cip/Kip family. Curr. Top. Microbiol.
thetized by ether inhalation as described49. For labeling of mitotic cells Immunol. 227, 2541 (1998).
25. Nakayama, K. et al. Mice lacking p27(Kip1) display increased body size,
in vivo, BrdU (1 mM) was injected along with the aforementioned multiple organ hyperplasia, retinal dysplasia, and pituitary tumors. Cell 85,
compounds. Retinae were harvested 48 hours after injection. Induc- 707720 (1996).
tion of reactive gliosis in P10.5 retinae was achieved by injecting 0.51.0 26. Dowling, J. E. The RetinaAn Approachable Part of the Brain (Harvard Univ.
microliters of ouabain at 0.7 mM. Similar results were obtained from Press, Cambridge, Massachusetts, 1987).
27. LaBaer, J. et al. New functional activities for the p21 family of CDK inhibitors.
retinae exposed to neurotoxins in vivo and in vitro; however, back- Genes Dev. 11, 847862 (1997).
ground induction of reactive gliosis was slightly higher in vitro due to 28. Cheng, M. et al. The p21(Cip1) and p27(Kip1) CDK inhibitors are essential
punctures or microtears. activators of cyclin D-dependent kinases in murine fibroblasts. EMBO J. 18,
15711583 (1999).
29. Kiyokawa, H. et al. Enhanced growth of mice lacking the cyclin-dependent
ACKNOWLEDGEMENTS kinase inhibitor function of p27(Kip1). Cell 85, 721732 (1996).
30. Fero, M. L. et al. A syndrome of multiorgan hyperplasia with features of
We thank M.H. Baron for discussion and support throughout this project, gigantism, tumorigenesis, and female sterility in p27(Kip1)-deficient mice.
S. Elledge, W. Harper and P. Zhang for cDNAs; J. Roberts and L.H. Tsai for Cell 85, 733744 (1996).
knockout mice, and J. Zitz, M. Peters and L. Rose for technical support. 31. Rich, K. A., Figueroa, S. L., Zhan, Y. & Blanks, J. C. Effects of Muller cell
disruption on mouse photoreceptor cell development. Exp. Eye Res. 61,
M.A. Dyer was supported by NRSA fellowship # EY06803-02 and the 235248 (1995).
Charles H. Revson Foundation Fellowship for Biomedical Research. This work 32. Cepko, C. L. et al. Lineage analysis using retroviral vectors. Methods 14,
393406 (1998).
was supported by National Institutes of Health Grant # EY0-8064.
33. Turner, D. L. & Cepko, C. L. A common progenitor for neurons and glia
persists in rat retina late in development. Nature 328, 131136 (1987).
34. Fields-Berry, S. C., Halliday, A. L. & Cepko, C. L. A recombinant retrovirus
RECEIVED 26 JUNE; ACCEPTED 28 JULY 2000 encoding alkaline phosphatase confirms clonal boundary assignment in
lineage analysis of murine retina. Proc. Natl. Acad. Sci. USA 89, 693697 (1992).
35. Nork, T. M., Wallow, I. H., Sramek, S. J. & Anderson, G. Mullers cell
1. Ridet, J. L., Malhotra, S. K., Privat, A. & Gage, F. H. Reactive astrocytes: involvement in proliferative diabetic retinopathy. Arch. Ophthalmol. 105,
cellular and molecular cues to biological function. Trends Neurosci 20, 14241429 (1987).
570577 (1997). [erratum appears in Trends Neurosci. 21, 80, 1998] 36. Robison, W. G. Jr., Tillis, T. N., Laver, N. & Kinoshita, J. H. Diabetes-related
2. Wilson, J. X. Antioxidant defense of the brain: a role for astrocytes. Can. J. histopathologies of the rat retina prevented with an aldose reductase
Physiol. Pharmacol. 75, 11491163 (1997). inhibitor. Exp. Eye Res. 50, 355366 (1990).
3. Ide, C. F. et al. Cellular and molecular correlates to plasticity during recovery 37. Fariss, R. N., Li, Z. Y. & Milam, A. H. Abnormalities in rod photoreceptors,
from injury in the developing mammalian brain. Prog. Brain Res. 108, amacrine cells, and horizontal cells in human retinas with retinitis
365377 (1996). pigmentosa. Am. J. Ophthalmol. 129, 215223 (2000).
4. Streit, W. J., Walter, S. A. & Pennell, N. A. Reactive microgliosis. Prog. 38. Li, Z. Y., Possin, D. E. & Milam, A. H. Histopathology of bone spicule
Neurobiol. 57, 563581 (1999). pigmentation in retinitis pigmentosa. Ophthalmology 102, 805816 (1995).
5. Unger, J. W. Glial reaction in aging and Alzheimers disease. Microsc. Res. 39. Kimura, H. et al. Cellular response in subretinal neovascularization induced
Tech. 43, 2428 (1998). by bFGF-impregnated microspheres. Invest. Ophthalmol. Vis. Sci. 40,
6. Streit, W. J. Microglial response to brain injury: a brief synopsis. Toxicol. 524528 (1999).
Pathol. 28, 2830 (2000). 40. Kuhrt, H. et al. Changes in CD44 and ApoE immunoreactivities due to retinal
7. MacLaren, R. E. Development and role of retinal glia in regeneration of pathology of man and rat. J. Hirnforsch. 38, 223229 (1997).
ganglion cells following retinal injury. Br. J. Ophthalmol. 80, 458464 41. Birnbach, C. D., Jarvelainen, M., Possin, D. E. & Milam, A. H. Histopathology
(1996). and immunocytochemistry of the neurosensory retina in fundus
8. Sahel, J. A., Albert, D. M. & Lessell, S. [Proliferation of retinal glia and flavimaculatus. Ophthalmology 101, 12111219 (1994).
excitatory amino acids]. Ophtalmologie 4, 1316 (1990). 42. Madigan, M. C., Penfold, P. L., Provis, J. M., Balind, T. K. & Billson, F. A.
9. Humphrey, M. F., Constable, I. J., Chu, Y. & Wiffen, S. A quantitative study of Intermediate filament expression in human retinal macroglia.
the lateral spread of Muller cell responses to retinal lesions in the rabbit. Histopathologic changes associated with age-related macular degeneration.
J. Comp. Neurol. 334, 545558 (1993). Retina 14, 6574 (1994).
10. Rutka, J. T. & Smith, S. L. Transfection of human astrocytoma cells with glial 43. Foisner, R. Dynamic organisation of intermediate filaments and associated
fibrillary acidic protein complementary DNA: analysis of expression, proteins during the cell cycle. Bioessays 19, 297305 (1997).
proliferation, and tumorigenicity. Cancer Res. 53, 36243631 (1993). 44. Levine, E. M., Close, J., Fero, M., Ostrovsky, A. & Reh, T. A. p27(Kip1)
11. Reichenbach, A. et al. The Muller (glial) cell in normal and diseased retina: regulates cell cycle withdrawal of late multipotent progenitor cells in the
a case for single-cell electrophysiology. Ophthalmic Res. 29, 326340 mammalian retina. Dev. Biol. 219, 299314 (2000).
(1997). 45. Matsuoka, S. et al. p57KIP2, a structurally distinct member of the p21CIP1
12. Sueishi, K. et al. Endothelial and glial cell interaction in diabetic retinopathy Cdk inhibitor family, is a candidate tumor suppressor gene. Genes Dev. 9,
via the function of vascular endothelial growth factor (VEGF). Pol. J. 650662 (1995).
Pharmacol. 48, 307316 (1996). 46. Dyer, M. A. & Cepko, C. L. p57 regulates progenitor cell proliferation and
13. Amin, R. H. et al. Vascular endothelial growth factor is present in glial cells of amacrine interneuron development in the mouse retina. Development (in press).
the retina and optic nerve of human subjects with nonproliferative diabetic 47. Morrow, E. M., Belliveau, M. J. & Cepko, C. L. Two phases of rod
retinopathy. Invest. Ophthalmol. Vis. Sci. 38, 3647 (1997). photoreceptor differentiation during rat retinal development. J. Neurosci. 18,
14. Taomoto, M. et al. Retinal degeneration induced by N-methyl-N-nitrosourea 37383748 (1998).
in Syrian golden hamsters. Graefes Arch. Clin. Exp. Ophthalmol. 236, 688695 48. Cepko, C. L., Fields-Berry, S., Ryder, E., Austin, C. & Golden, J. Lineage
(1998). analysis using retroviral vectors. Curr. Top. Dev. Biol. 36, 5174 (1998).
15. Hjelmeland, L. E. & Harvey, A. K. in Progress in Retinal Research, Vol. 7 (eds. 49. Sahel, J. A. et al. Mitogenic effects of excitatory amino acids in the adult rat
Osborn, N. & Chader, G.) 259281 (Pergamon, New York, 1988). retina. Exp. Eye Res. 53, 657664 (1991).

articles
Molecular memory by reversible

translocation of calcium/calmodulin-
dependent protein kinase II
K. Shen1, M. N. Teruel1,2, J. H. Connor3, S. Shenolikar3 and T. Meyer1,2
1 Department of Cell Biology, Duke University Medical Center, Durham, North Carolina 27710, USA
2 Department of Molecular Pharmacology, 269 Campus Drive, Rm #3215, Stanford University Medical School, Stanford, California 94305, USA
3 Department of Pharmacology and Cancer Biology, Duke University Medical Center, Durham, North Carolina 27710, USA
Correspondence should be addressed to T.M. (tobias1@stanford.edu)
Synaptic plasticity is thought to be a key process for learning, memory and other cognitive functions
of the nervous system. The initial events of plasticity require the conversion of brief electrical signals
into alterations of the biochemical properties of synapses that last for much longer than the initial
stimuli. Here we show that a regulator of synaptic plasticity, calcium/calmodulin-dependent protein
kinase II (CaMKII), sequentially translocates to postsynaptic sites, undergoes autophosphorylation
and gets trapped for several minutes until its dissociation is induced by secondary autophosphoryla-
tion and phosphatase 1 action. Once dissociated, CaMKII shows facilitated translocation for several
minutes. This suggests that trapping of CaMKII by its targets and priming of CaMKII translocation
may function as biochemical memory mechanisms that change the signaling capacity of synapses.
A broad range of pharmacological, biochemical and genetic stud- neurons in which CaMKII was blocked using peptide blockers
ies establish that the serine/threonine kinase CaMKII is a key reg- and pharmacological agents5. This role was confirmed by the
ulator of long-term potentiation as well as of other forms of introduction of constitutively active CaMKII into neurons by
neuronal plasticity. CaMKII was initially discovered as one of the purified protein or viral transfection6,7. Furthermore, CaMKII-
most abundant neuronal serine/threonine kinases with an activ- deficient mice show impaired hippocampal LTP and behavioral
ity that is induced by binding calciumcalmodulin (Ca2+CaM)1. defects in spatial learning and memory8,9. A different mouse
The kinase is enriched at the postsynaptic density (PSD), a model, in which a constitutively active CaMKII is expressed in
cytoskeletal structure beneath the postsynaptic membrane that hippocampal and other neurons, shows a shift in the frequency
contains many structural and signaling proteins. In adult mam- dependence of long-term potentiation and defects in spatial
malian central neurons, the predominantly expressed CaMKIIs learning10-12. A mutant mouse in which the endogenous CaMKII
are the and isoforms1,2, with lower expression of the and can be activated but not autophosphorylated showed that the
isoforms. autophosphorylation site at Thr286 is necessary for hippocampal
The isoform of CaMKII is a multimeric enzyme that con- LTP and spatial learning13,14. Taken together, these experiments
sists of approximately 12 subunits per holoenzyme3. After acti- strongly supported the functional importance of CaMKII and
vation by Ca 2+ CaM, CaMKII undergoes a characteristic its autophosphorylation in hippocampal LTP and in spatial learn-
trans-subunit autophosphorylation on Thr286. This autophos- ing and memory.
phorylation requires CaM binding to two neighboring subunits As biochemical studies show that CaMKII can phosphorylate
and renders the kinase partially Ca 2+ CaM independent 3 . AMPA and NMDA glutamate receptor subtypes3 and as signifi-
Autophosphorylation also leads to a several-hundredfold cant amounts of CaMKII can exist at PSDs1, it is proposed that
increase in Ca2+CaM binding affinity4. This partial activity is CaMKII may exert its function at synapses by regulating postsy-
sustained until Thr286 is dephosphorylated, presumably by naptic AMPA and/or NMDA receptors. CaMKII also translocates
phosphatase 1 action. In addition to phosphorylating itself, from the cytosol to this postsynaptic region in response to stim-
CaMKII has a wide spectrum of substrates in vitro, including uli that activate NMDA receptors15, and biochemical studies sug-
AMPA receptors and NMDA receptors, which are key compo- gest that this translocation is mediated by the binding of CaMKII
nents of the postsynaptic membrane. The broad interest in the to postsynaptic density-localized NMDA receptors1618. Here we
neuronal role of CaMKII stems in part from the activation- used confocal imaging of green fluorescent protein
induced autonomous activity of the kinase. This property of (GFP)CaMKII in cultured hippocampal neurons to identify the
the kinase led to the suggestion that autophosphorylation may key signaling steps in the translocation cycle of CaMKII. Partic-
function as a biochemical memory process that can prolong ularly, we were interested in whether trapping of CaMKII at its
a sufficiently strong but brief calcium signal into a long-lasting postsynaptic sites (target trapping) and facilitation of the translo-
change in CaMKII activity. cation of CaMKII by previous activity (translocation priming)
The first direct evidence for a role of CaMKII in synaptic plas- are potential prolongation mechanisms that could generate
ticity came from electrophysiological studies in hippocampal altered biochemical memory states of the synapse.

articles
Fig. 1. Field stimulation induces a synaptically a

trapped state of CaMKII. (a) Cultured hip-
pocampal CA1CA3 pyramidal neurons
(7 days in vitro) were transfected with
GFPCaMKII. Transfected neurons were
imaged by confocal microscopy and stimulated
by extracellular electrodes that were placed
adjacent to the neurons. A single 100 Hz pulse,
lasting for 1 s, was delivered at time 0. Time-
lapse confocal images were taken to monitor
the reversible translocation of GFPCaMKII
(typical experiment shown; n = 12 neurons).
(b) A translocation time course is analyzed
from measuring the local fluorescence inten-
sity at individual synaptic sites in each of the
images as a function of time. (c) Average rela-
tive fluorescence intensity of the different
synaptic sites in the neuron shown in (a) plot- b
ted against time. Fluo-3-am-loaded neurons
c
were stimulated with the same protocol, and
the calcium signal was measured and plotted
Rel. syn. fluor. int.

for comparison (right). The field-induced
synaptic translocation of GFPCaMKII lasts
much longer than the brief calcium transient.
Scale bar in (a), 10 m.
RESULTS
Stimulation of hippocampal neurons with
series of electrical field pulses (10100 Hz
for one second) triggered a transient translocation of of the cytosol. Such a local mechanism is likely, as two Ca2+CaM
GFPCaMKII from cytosol to postsynaptic sites. Similar electri- molecules have to bind to adjacent subunits of CaMKII for
cal stimulation protocols induce synaptic plasticity614. The stim- autophosphorylation to occur20. NMDA receptor activation is
ulus needed for maximal translocation was variable with both necessary and sufficient for the synaptic translocation of
individual neurons but typically required between 20 and 50 field GFPCaMKII15. When we activated NMDA receptors in a den-
pulses. We acquired a series of sequential confocal images of the dritic region by applying brief puffs of glutamate using
transient synaptic translocation of GFPCaMKII (Fig. 1a). The micropipettes, a rapid and transient synaptic translocation of
time course of translocation was obtained by measuring the flu- GFPCaMKII was observed (Fig. 3a, panels II and III). With
orescence intensity at each synaptic site in a series of images lower-amplitude puffs, the duration of synaptic localization of
(Fig. 1b) and plotting the average fluorescence intensity as a func- GFPCaMKII was brief, similar to the brief translocation
tion of time. Because CaMKII is activated by calcium signals, we observed for the Thr286Ala mutant. With higher-amplitude glu-
used the fluorescent indicator fluo-3 to measure the calcium tamate puffs, a marked transition occurred, and GFPCaMKII
response for the same stimulation protocol. The synaptic local- was trapped for prolonged periods within the synapse (Fig. 3a,
ization of GFPCaMKII persisted for many minutes beyond the panels IV and V). The striking difference in the translocation
second long increase in calcium concentration trig-
gered by the field stimulus (Fig. 1c). a
Because neuronal stimulation induces autophos-
phorylation of CaMKII19, and because autophos-
phorylation is required for its functions in synaptic
plasticity and learning and memory13,14, we tested
whether the prolonged synaptic localization of
CaMKII could be observed for a mutant
GFPCaMKII with an Ala residue instead of a Thr at
position 286. Although this autophosphorylation-
deficient mutant still translocated to synaptic sites,
the translocation was slightly weaker and much more
transient than that of the wild-type enzyme (Fig. 2).
This suggests that autophosphorylation per se is not
required for inducing the translocation of CaMKII to b
Fig. 2. Absence of a trapped state of CaMKII in a Thr286
synaptic sites15, but it is required to induce the synap- mutant of CaMKII. (a) Neurons transfected with the
tically trapped state. Thr286Ala mutant of GFPCaMKII were stimulated with the
It is conceivable that autophosphorylation occurs same protocol as in Fig. 1. A typical time series of images is
only after CaMKII is bound to NMDA receptors, shown during and after field stimulation (n = 6 neurons). Scale
which can position the enzyme near an area of high bar, 10 m. (b) Time course of stimulus-induced translocation
calcium concentration that is not found in the rest of the GFPCaMKII Thr286Ala mutant.

articles
Fig. 3. Transition from a rapidly reversible

state to a synaptically trapped state of CaMKII a
by increased amplitude of local glutamate stim-
uli. (a) GFPCaMKII-transfected pyramidal
neurons were stimulated by glutamate puffs
delivered from a micropipette placed near the
neurons. Panels IVI are time-lapse confocal
images of an individual neuron taken at 0 s, 20 s,
40 s, 140 s, 320 s and 480 s. The position of the
micropipette is indicated in panel I. A weak
stimulus (WS, 25 psi injection pressure, 50 ms
injection duration) was applied at t = 0 s. A
strong stimulus (SS, 25 psi injection pressure,
300 ms injection duration) was applied at t = 90 s. b c d
Scale bar, 10 m. (b) The average relative fluo-
rescence intensity of the synaptic sites in Rel. syn. fluor. int.
(a) plotted against time. The stimulation time
points WS and SS are labeled, as well as the
time points when the images in (a) were taken
(typical experiment shown; n = 4 neurons).
(c) The relative cytosolic Ca2+ signal was mea-

sured for the same protocol using the calcium
indicator fluo-3. The same weak and strong
stimuli as in (b) were applied at the time points
indicated in the graph. (d) Comparison of the
average synaptic dissociation half-times (T1/2)
following either a strong puff (n = 31 synapses from 4 neurons) or a weak glutamate puff (n = 30 synapses from 4 neurons). The same comparison is
shown for the dissociation times following a 1 s, 100 Hz stimulus for wild type (n = 76 synapses from 12 neurons) and a 1 s, 100 Hz stimulus for the
Thr286Ala mutant (n = 38 synapses from 5 neurons). T1/2 was defined as the time required for the synaptic fluorescence intensity of GFPCaMKII to
decrease to 50% of its maximal value.
time course is seen in fluorescence intensity traces recorded from on the dissociation time (data not shown), consistent with earlier
synaptic regions (Fig. 3b). The trapped localization state large- biochemical data that identified protein phosphatase-1 (PP1) as
ly results from an increase in the amplitude but not duration of a CaMKII Thr286 phosphatase in the synapse24,25. In CA1 neu-
the calcium signals triggered by the two stimuli (Fig. 3c). Thus, rons, it is suggested that dephosphorylation of CaMKII can be
different synaptic stimuli induce two markedly different states regulated indirectly by cAMP, because PP1 activity can be con-
of CaMKII, a short transiently localized state that lacks mem- trolled by PKA-mediated phosphorylation of inhibitor-1
ory for unphosphorylated CaMKII and a prolonged trapped (ref. 26). We therefore analyzed the effect of PKA activation on
state that results from the autophosphorylation of CaMKII at GFPCaMKII dissociation by treating hippocampal neurons with
Thr286. The duration of the two states was significantly different 8-bromo-cAMP. Indeed, consistent with the hypothesis that the
(Fig. 3d). cAMP pathway regulates CaMKII function, activation of PKA
How is the duration of this memory state controlled? We significantly prolonged the synaptically localized memory state
measured the dissociation time of synaptically localized of GFPCaMKII (Fig. 4c).
GFPCaMKII by first inducing maximal synaptic translocation To further dissect the mechanism for CaMKII dissociation,
with bath-applied glutamate and then replacing the extracellu- we investigated the importance of secondary autophosphoryla-
lar glutamate buffer with low-calcium, glutamate-free medium. tion sites of CaMKII (Thr305/Thr306), which inhibit further
This protocol rapidly lowered intracellular calcium and thereby binding of calmodulin after one or both of the neighboring sites
enabled comparative measurements of synaptic dissociation times are phosphorylated2729. Although the physiological occurrence
(Fig. 4a). Whereas wild-type GFPCaMKII dissociated from and significance of these phosphorylation events has not yet been
synapses over a few minutes15, we observed a tenfold prolongation established, we were interested in testing the hypothesis that tran-
of the dissociation time for mutant GFPCaMKII with an aspar- sient phosphorylation at these sites is involved in the dissocia-
tic acid residue at the 286 position (Fig. 4b), which mimics the tion of CaMKII from its postsynaptic sites. Indeed, a
autophosphorylated state of CaMKII21,22. For comparison, the GFPCaMKII mutant containing Ala substitutions in place of
dissociation of a previously characterized Ala286 mutant is also both Thr305 and Thr306 showed markedly slower synaptic dis-
shown15. Because wild-type CaMKII dissociates more rapidly sociation (Fig. 4d), suggesting that secondary autophosphoryla-
than the Asp286 mutant but more slowly than the Ala286 mutant, tion at these sites is required for effective synaptic dissociation
the dissociation of wild-type CaMKII must involve at least a par- of CaMKII. Together, these studies show that the duration of
tial dephosphorylation of the Thr286 residues in the CaMKII postsynaptic CaMKII localization is tightly regulated by two
oligomer. autophosphorylation events that render the kinase autonomous
We tested this hypothesis by incubating the neuron with and/or incapable of binding Ca2+CaM and that the duration of
4 nM calyculin A, a cell-permeable phosphatase inhibitor23. The the stimulatory Thr286 autophosphorylation is regulated by the
synaptic dissociation time of GFPCaMKII was significantly slow- PKA/PP1 signaling pathway.
er in the presence of calyculin A (Fig. 4c). Similar concentrations Given that, following its activation, the postsynaptic CaMKII
of the less-specific inhibitor okadaic acid had no measurable effect undergoes autophosphorylation at the stimulatory site (Thr286)

articles
Fig. 4. The duration of the trapped state of

CaMKII is regulated by reversible autophospho- a
rylation at Thr286 and Thr305/Thr306.
(a) Measurement of the synaptic dissociation of
CaMKII after removal of calcium and glutamate.
A series of images is shown that was used for the
dissociation analysis. (b) Time course of synaptic
dissociation of wild-type GFPCaMKII (n = 55
synapses from 8 neurons) and of a Thr286Asp
mutant of CaMKII (n = 34 synapses from 5 neu-
rons), compared with the dissociation of the pre-
viously measured Ala286 mutant15. Exponential
fits are shown for each of the mutants (exponen-
tial time constants, 1.8 102 Hz for wild type,
3.8 104 Hz for Asp286 mutant and > 0.1 Hz
for Ala286 mutant; standard errors calculated for
each time point). (c) Time courses for the disso-
ciation of wild-type GFPCaMKII, in the absence
b c d
or presence of 4 nM calyculin A (n = 39 synapses

from 6 neurons) for 10 min before the experi-
ment. Also shown is a trace in which the wild

type was treated with 100 M 8-bromo cAMP
(n = 32 synapses from 5 neurons) for 15 min. For
the dissociation measurements, the glutamate
solution was exchanged for a low-Ca2+ solution
without glutamate at t = 0. Relative synaptic fluorescence intensity as a function of time during the dissociation process. Standard errors are shown
for each time point and condition. Exponential fits through the data are shown (exponential time constants, 1.8 102 Hz for control, 9.7 104 Hz
for calyculin A and 2.9 103 Hz 8-Br-cAMP). (d) Time course of synaptic dissociation of GFPCaMKII wild type (WT; n = 55 synapses from 8 neu-
rons), compared to a GFPCaMKII mutant with the inhibitory phosphorylation sites disabled (Thr305Ala + Thr306Ala mutant; n = 29 synapses
from 5 neurons; exponential time constants, 1.8 103 Hz for Ala305/306 double mutant).
as well as inhibitory sites (Thr305/Thr306), it is likely that the ization of the kinase at its postsynaptic site. The decision between
dissociated cytosolic CaMKII remains autophosphorylated at one a rapidly reversible and the trapped state of the enzyme depends
or both of these sites in some of its approximately 12 subunits. on stimulus intensity and requires autophosphorylation of the
Such residual autophosphorylation is likely to alter the translo- kinase at Thr286. The term target trapping was introduced to
cation of CaMKII in response to subsequent stimuli. We tested describe this transition from an initially rapidly reversible bound
for changes in translocation sensitivity using a protocol with three state to a tightly bound state. By inducing this high-affinity state,
sequential electrical stimuli: first an initial submaximal electri- the autophosphorylated and presumably partially active kinase
cal stimulus, second a stimulus that induced maximal transloca- can remain near its targets for a significant period of time after
tion of GFPCaMKII and third a submaximal stimulus identical the calcium signal is terminated. Surprisingly, our studies also
to the first one (Fig. 5a). We monitored the time course of the showed that the dissociated CaMKII remains in a cytosolic facil-
synaptic translocation events (Fig. 5b) or of the changes in free itated state for at least minutes. The term primed state was intro-
calcium concentration (Fig. 5c) induced by this protocol. The duced for the dissociated CaMKII because it is characterized by a
same 20 Hz stimulus induced markedly more GFPCaMKII significantly lower stimulus requirement for translocation
translocation when applied within minutes after its dissociation induced by a subsequent stimulus.
from synaptic sites. Thus, cytosolic CaMKII can exist after its dis- Together with previous biochemical studies3,16-18, our mea-
sociation in a primed or sensitized state that is much more surements suggest that CaMKII is activated in a defined cycle
responsive for synaptic translocation. that involves enzymatic as well as translocation and localization
Is this primed state dependent on autophosphorylation of steps. In this model, the inactive CaMKII is cytosolic or weakly
CaMKII? We tested the same paired-pulse stimulation protocol actin bound30 if present as a hetero-oligomer with isoforms.
using the Thr286Ala mutant of GFPCaMKII. No priming of Electrical stimuli that trigger a dendritic Ca2+ increase lead to the
translocation was observed with this mutant, suggesting that binding of Ca2+CaM to relatively low-affinity Ca2+CaM bind-
residual autophosphorylation of CaMKII at Thr 286 is respon- ing sites on the CaMKII oligomer (50 nM)3. Because the diffu-
sible for the primed state (Fig. 6a). Priming was significant for sion coefficient of CaMKII is approximately 1 m2/s (ref. 30),
wild-type CaMKII, but not for the mutant CaMKII (Fig. 6b). the cytosolic Ca2+CaMCaMKII complex diffuses within a lim-
The primed cytosolic state returned to baseline sensitivity sever- ited region until it binds to NMDA receptors and possibly other
al minutes after the synaptic dissociation of GFPCaMKII postsynaptic targets. If the calcium stimulus is relatively weak,
(Fig. 6c), suggesting that the Thr286 residue of CaMKII becomes this binding is rapidly reversible. However, once the
largely dephosphorylated once the kinase is back in the cytosol Ca2+CaMCaMKII complex is bound to its postsynaptic sites,
for several minutes. autophosphorylation can occur for sufficiently strong Ca2+CaM
stimuli. The need for a stronger stimulus for trapping compared
DISCUSSION to translocation likely results from the biochemical requirement
Our studies show that electrically or glutamate-induced translo- that Ca2+CaM has to be bound to two neighboring subunits for
cation of CaMKII can lead to either transient or prolonged local- autophosphorylation of Thr286 to occur. This corresponds to

articles
a
Fig. 5. Identification of a cytosolic state of CaMKII primed for synaptic

translocation. (a) Time-lapse confocal images of a GFPCaMKII trans- b c
fected pyramidal neuron stimulated by three electrical field stimuli. A

20 Hz, 1 s stimulus was applied at t = 20 s. A 100 Hz, 1 s stimulus was
applied at t = 140 s. Another 20 Hz, 1 s stimulus was applied at t = 340 s.
(b) Local synaptic fluorescence intensity plotted against time to show the
repetitive translocation of GFPCaMKII to synapses. The time points for
the three stimuli and the time when the images in (a) were taken are marked
with thin and thick bars, respectively. The second 20 Hz stimulus induced
markedly more translocation than the first one. (c) Relative Ca2+ concentra-
tion measured in a neuron stimulated with the same protocol as in (b).
approximately 8 Ca2+ ions and two CaM molecules for inducing targets, at least on the time scale of minutes, requires the dephos-
the trans-phosphorylation process, ensuring that the induction of phorylation of the autonomous phosphorylation site. A previously
the high-affinity partially autonomous state of the kinase is tight- proposed role of phosphatase-1 in this process is supported by the
ly dependent on the amplitude and duration of the local Ca2+ observed inhibition of the dissociation by the PP1 inhibitors caly-
signal. culin A and okadaic acid. Interestingly, this release seems to require
The autophosphorylation therefore has two roles, to gener- not only dephosphorylation by phosphatase-1 but also secondary
ate an enzyme that is partially active in the absence of phosphorylation at inhibitory phosphorylation site(s) of CaMKII
Ca 2+CaM and to generate a trapped state of CaMKII that that suppress subsequent binding of Ca2+CaM binding. In vitro,
enables the kinase to remain near its targets even after the cal- Thr 305 and 306 are phosphorylated as burst autophosphoryla-
cium stimulus is terminated. From a biochemical perspective, tion events that can occur secondary to phosphorylation of the
the postsynaptically trapped state of CaMKII is likely an indirect Thr286 sites. A critical role for inhibitory autophosphorylation at
result of the requirement of bound Ca2+CaM for the interac- these sites for the CaMKII dissociation is suggested by the delayed
tion of CaMKII with its postsynaptic sites 15 and of the bio- dissociation of CaMKII mutants that have both threonines replaced
chemically measured several-hundred-fold increase in by alanines. Finally, the observed PKA-induced delay in dissocia-
Ca2+CaM binding affinity induced by the autophosphorylation of CaMKII supports a previous hypothesis that PKA may reg-
tion of the kinase4. ulate CaMKII function indirectly by phosphorylating inhibitor-1,
The observed persistent postsynaptic binding of the Asp286 thereby suppressing PP1 activity and prolonging the time that
mutant of CaMKII suggests that the release of the kinase from its CaMKII spends at its sites of action.
Fig. 6. Control measurements showing the same
translocation analysis using neurons transfected with the
a b c
autophosphorylation-deficient mutant GFPCaMKII

(Thr286Ala). (a) For the Thr286Ala mutant, the second
20 Hz-induced translocation is equally strong as the first
20 Hz-induced translocation, suggesting that autophos-
phorylation of Thr286 is necessary for the cytosolic
state of CaMKII that is primed for synaptic translocation.
(b) Statistical analysis of the enhanced translocation of
the wild-type GFPCaMKII (n = 36 synapses from 8 neu-
rons) and lack of an enhancement for the Thr286Ala
mutant (n = 18 synapses from 4 neurons). The ratio2/1 value is a measure for the potentiation or priming of the translocation response. It was calculated as
the ratio of the second postsynaptic translocation amplitude (the 20 Hz response after the 100 Hz stimulus) over the first 20 Hz translocation amplitude.
Standard errors are shown. (c) Analysis of the duration of the translocation primed state of CaMKII. For time course measurements, different time inter-
vals were analyzed between the synaptic dissociation of CaMKII and the application of the 20 Hz stimulus. Each point is the average of the ratio2/1 value of
different synapses in a single neuron. Standard errors are shown for each of the measured time points.

articles
In addition, we found that CaMKII remains for several min- 3. Braun, A. P. & Schulman, H. The multifunctional calcium/calmodulin-
utes in a facilitated or primed state after dissociating from its dependent protein kinase: from form to function. Annu. Rev. Physiol. 57,
417445 (1995).
postsynaptic targets. Such an enhanced capability for transloca- 4. Meyer, T., Hanson, P. I., Stryer, L. & Schulman, H. Calmodulin trapping by
tion is possibly a result of residual autophosphorylation of at least calcium-calmodulin-dependent protein kinase. Science 256, 11991202
a few of the Thr286 autophosphorylation sites of the oligomer. (1992).
5. Malinow, R., Madison D.V.. & Tsien, R. W. Persistent protein kinase activity
This is supported by the finding that the Ala286 mutant kinase underlying long-term potentiation. Nature 335, 820824 (1988).
does not show facilitated translocation. However, the priming of 6. Pettit, D. L., Perlman, S. & Malinow, R. Potentiated transmission and
translocation is ultimately reversible, suggesting that CaMKII prevention of further LTP by increased CaMKII activity in postsynaptic
hippocampal slice neurons. Science 266, 18811885 (1994).
ultimately loses most of the residual phosphate moieties in an 7. Lledo, P. M. et al. Calcium/calmodulin-dependent kinase II and long-term
inactive neuron. potentiation enhance synaptic transmission by the same mechanism. Proc.
Similar to other studies with cultured neurons, we have used in Natl. Acad. Sci. USA 92, 1117511179 (1995).
8. Silva, A. et al. Deficient hippocampal long-term potentiation in alpha-
our study relatively young neurons with a small number of synaps- calcium-calmodulin kinase II mutant mice. Science 257, 201206 (1992).
es that are not yet fully mature. Although neurons that are more 9. Silva, A. et al. Impaired spatial learning in alpha-calcium-calmodulin kinase
than 10 days in culture are more difficult to transfect, we found in II mutant mice. Science 257, 206211 (1992).
10. Mayford, M. et al. Control of memory formation through regulated
several measurements that CaMKII can still translocate in cultures expression of a CaMKII transgene. Science 274, 16781683 (1996).
that are greater than 10 days old and have more developed synaps- 11. Mayford, M. et al. CaMKII regulates the frequency-response function of
hippocampal synapses for the production of both LTD and LTP. Cell 81,
es. Because CaMKII also associates with the PSD and NMDA recep- 891904 (1995).
tors in biochemical assays18, it is likely that the PSD translocation
12. Bach, M. E. et al. Impairment of spatial but not contextual memory in

of CaMKII described here is a general mechanism for the regula- CaMKII mutant mice with a selective loss of hippocampal LTP in the range of
the theta frequency. Cell 81, 905915 (1995).
tion of synapses in developing as well as mature neurons. 13. Giese, K. P. et al. Autophosphorylation at Thr286 of the alpha calcium-
In summary, our studies have uncovered two distinct states calmodulin kinase II in LTP and learning. Science 279, 870873 (1998).
of molecular memory associated with CaMKII, a prolonged local- 14. Cho, Y. H. et al. Abnormal hippocampal spatial representations in
alphaCaMKIIT286A and CREBalphaDelta-mice. Science 279, 867869
ization of CaMKII near its synaptic targets (target trapping) and (1998).
a transient sensitization of CaMKII for future translocation 15. Shen, K. & Meyer, T. Dynamic control of CaMKII translocation and
(translocation priming). The regulation of the trapped and localization in hippocampal neurons by NMDA receptor stimulation. Science
284, 162167 (1999).
primed states of CaMKII by primary and secondary autophos- 16. Gardoni, F. et al. AlphaCaMKII binding to the C-terminal tail of NMDA
phorylation, as well as by PP1 and PKA, enables a high degree of receptor subunit NR2A and its modulation by autophosphorylation. FEBS
control by additional neuronal signaling pathways. By integrating Lett. 456, 394398 (1999).
17. Leonard, A. S. et al. Calcium/calmodulin-dependent protein kinase II is
our live cell imaging studies with previous biochemical and phys- associated with the N-methyl-D-aspartate receptor. Proc. Natl. Acad. Sci. USA
iological results, we suggest that CaMKII may regulate an initial 96, 32393244 (1999).
phase of neuronal plasticity by a biochemical memory cycle that 18. Strack, S. & Colbran, R. J. Autophosphorylation-dependent targeting of
calcium/calmodulin-dependent protein kinase II by the NR2B subunit of
includes the key steps of translocation, autonomous activity, tar- the N-methyl-D-aspartate receptor. J. Biol. Chem. 273, 2068920692
get trapping and translocation priming. (1998).
19. Ouyang, Y. et al. Visualization of the distribution of autophosphorylated
METHODS calcium/calmodulin-dependent protein kinase II after tetanic
Cell culture and transfections. CA3CA1 hippocampal cultures (post- stimulation in the CA1 area of the hippocampus. J. Neurosci. 17,
54165427 (1997).
natal days 2 to 4) were prepared as described31. Neurons, 6 to 12 days in 20. Hanson, P. I., Meyer, T., Stryer, L. & Schulman, H. Dual role of calmodulin in
culture, were transfected by microporation of the respective DNA con- autophosphorylation of multifunctional CaM kinase may underlie decoding
structs31. Two electrical pulses (22 V/mm, 20 s apart) were applied to of calcium signals. Neuron 12, 943956 (1994).
neurons immediately after DNA application. Localization and translo- 21. Fong, Y.-L. et al. Studies of the regulatory mechanism of Ca2+/calmodulin-
cation studies were done 8 to 16 hours after transfection. dependent protein kinase II. Mutation of threonine 286 to alanine and
aspartate. J. Biol. Chem. 264, 1675916763 (1989).
22. Waldmann, R., Hanson, P. I. & Schulman, H. Multifunctional
Microscopy and electrical and local glutamate stimulation. All fluores- Ca2+/calmodulin-dependent protein kinase made Ca2+ independent for
cent confocal images were taken on a Zeiss LSM 410 (Carl Zeiss, Yena, functional studies. Biochemistry 29, 16791684 (1990).
Germany) with the pinhole typically set at 30. Electrical stimuli were gen- 23. Shenolikar, S. Protein phosphatasesnew avenues for cell regulation. Annu.
Rev. Cell Biol. 10, 5586 (1994).
erated with a Grass S48 stimulator (Grass Instruments, Quincy, 24. Strack, S. et al. Differential inactivation of postsynaptic density-associated
Massachusetts). Trains of 0.5-ms square pulses were delivered at different and soluble Ca2+/calmodulin-dependent protein kinase II by protein
frequency using constant current mode through glass micropipettes that phosphatases 1 and 2A. J. Neurochem. 68, 21192128 (1997).
were placed near the imaged neuron. Currents were 3070 A. For local 25. Shields, S. M., Ingebritsen, T. S. & Kelly, P. T. Identification of protein
glutamate stimulation experiments, glass micropipettes containing 100 phosphatase 1 in synaptic junctions: dephosphorylation of endogenous
calmodulin-dependent kinase II and synapse-enriched phosphoproteins.
mM glutamate and 10 mM glycine were placed 1030 m away from the J. Neurosci. 5, 34143422 (1985).
dendritic arbors. Injection duration was typically 50500 ms. Injection 26. Blitzer, R. D. et al. Gating of CaMKII activity by cAMP-regulated phosphatase
pressure was 1530 psi. during LTP. Science 28, 19401943 (1998).
27. Colbran, R. J. & Soderling, T. R. Calcium/calmodulin-independent
ACKNOWLEDGEMENTS autophosphorylation sites of calcium/calmodulin-dependent protein kinase II.
J. Biol. Chem. 265, 1121311219 (1990).
We thank G. Augustine, L. Katz, A. Means and C. Fink, Duke University, for 28. Hanson, P. I. & Schulman, H. Inhibitory autophosphorylation of
comments. This work was supported by grant RO1-GM48113 from the National multifunctional Ca2+/calmodulin-dependent protein kinase analyzed by site-
Institute of General Medicine. directed mutagenesis. J. Biol. Chem. 267, 1721617224 (1992).
29. Patton, B. L., Miller, S. G. & Kennedy, M. B. Activation of type II
calcium/calmodulin-dependent protein kinase by Ca2+/calmodulin is
RECEIVED 14 APRIL; ACCEPTED 26 JULY 2000 inhibited by autophosphorylation of threonine within the calmodulin-
binding domain. J. Biol. Chem. 265, 1120411212 (1990).
1. Bennett, M. K. et al. Purification and characterization of a calmodulin- 30. Shen, K, Teruel, M. N., Subramanian, K. & Meyer, T. CaMKIIbeta functions
dependent protein kinase that is highly concentrated in brain. J. Biol. Chem. as an F-actin targeting module that localizes CaMKIIalpha/beta
258, 1273512744 (1983). heterooligomers to dendritic spines. Neuron 21, 593606 (1998).
2. Erondu, N. E. & Kennedy, M. B. Regional distribution of type II 31. Teruel, M. N., Blanpied, T., Shen, K., Augustine, G. J. & Meyer, T. A versatile
Ca2+/calmodulin-dependent protein kinase in rat brain. J. Neurosci. 5, microporation technique for the transfection of CNS neurons. J. Neurosci.
32703277 (1985). Methods 93, 3748 (1999).

articles
Glutamate receptors regulate actin-

based plasticity in dendritic spines
Maria Fischer, Stefanie Kaech, Uta Wagner, Heike Brinkhaus and Andrew Matus
Friedrich Miescher Institute, P.O. Box 2543, 4002 Basel, Switzerland

The first two authors contributed equally to this work
Correspondence should be addressed to A.M. (matus@fmi.ch)
Dendritic spines at excitatory synapses undergo rapid, actin-dependent shape changes which may
contribute to plasticity in brain circuits. Here we show that actin dynamics in spines are potently
inhibited by activation of either AMPA or NMDA subtype glutamate receptors. Activation of either
receptor type inhibited actin-based protrusive activity from the spine head. This blockade of
motility caused spines to round up so that spine morphology became both more stable and more
regular. Inhibition of spine motility by AMPA receptors was dependent on postsynaptic membrane
depolarization and influx of Ca2+ through voltage-activated channels. In combination with previous
studies, our results suggest a two-step process in which spines initially formed in response to NMDA
receptor activation are subsequently stabilized by AMPA receptors.
The adaptive properties of brain circuitry require that the tran- If these events have a meaningful role in synaptic plasticity,
sient events of synaptic transmission be converted into lasting they should be regulated by synaptic transmission. Here we
changes in neuronal connectivity. However, the cellular events examined the influence of AMPA and NMDA glutamate recep-
underlying this transitionespecially those that operate in adult tors on actin dynamics in dendritic spines of excitatory synaps-
circuitryare poorly understood. Experience-dependent plas- es, where glutamate is the major neurotransmitter. Our results
ticity is associated particularly with glutamatergic excitatory reveal a sequence of events, triggered by glutamate receptor acti-
synapses1, the vast majority of which are made onto dendritic vation, that lead to the inhibition of actin motility and the sta-
spines2. Abundant evidence indicating that spines can undergo bilization of spine morphology.
activity-dependent changes in shape and number has raised wide-
spread interest in them as a possible cellular substrate for synap- RESULTS
tic plasticity in the brain37. Dynamic imaging studies on living Glutamate receptor inhibition of spine actin dynamics
neurons confirm that spines are motile structures814 and suggest We studied the influence of glutamate receptors on spine motili-
that motile filopodial spine precursors may initiate synapse forty by making video time-lapse recordings of GFP-actin dynam-
mation9. Changes in spine shape occur in the living brain and, ics in dendrites of transfected neurons (Fig. 1a). Dendritic spines
particularly during early postnatal development, levels of spine in the brain have many morphologies18,19, and this is reflected in
motility are responsive to alterations in sensory experience14. our images of spines of living hippocampal neurons in cell cul-
Other studies have begun to suggest how activity in neu- ture. To ensure that our results would incorporate any variations
ronal circuits might influence spine morphology. Experiments in dynamic behavior associated with this structural heterogeneity,
with hippocampal slice cultures suggest that maintenance of we selected well-separated dendritic segments long enough to
spine morphology requires continual low-level activation of contain 30 to over 100 individual spines for recording (Fig. 1b).
AMPA receptors by spontaneously released glutamate15, where- To demonstrate the dynamic changes in GFPactin captured
as other studies demonstrate de novo spine formation follow- in time-lapse recording, video sequences were processed using a
ing stimulation protocols that lead to NMDA computer algorithm that subtracts gray-scale values between pix-
receptor-dependent long term potentiation (LTP)11,12,16. These els in neighboring video frames and displays the summed dif-
studies implicate AMPA and NMDA receptors in different ferences. Dark areas indicative of high motility were particularly
aspects of spine plasticity but do not address the nature of the associated with dendritic spines (Fig. 1b). The pixel densities in
cellular mechanism that converts receptor activation into mor- these difference images also reflect levels of motility. This was
phological change. Some observations suggest that actin-based used to assess the influence of glutamate receptor activation on
motility may be important in this process. Cytoplasmic actin spine actin dynamics. The example shown in Fig. 1b is one of
isoforms, associated with shape changes in motile cells, accu- many experiments (n > 50) in which treating hippocampal neu-
mulate at high concentrations in dendritic spines, whereas skele- rons with 100 M glutamate strongly inhibited actin-based
tal muscle actins do not17. Also, time-lapse studies of neurons dynamics (compare Fig. 1b panels 2 and 4). Glutamate receptor
expressing actin tagged with green fluorescent protein (GFP), activation exerted a profound effect, completely blocking actin
either alone or fused to actin (GFPactin), reveal rapid changes dynamics in all spines on a given segment of dendrite. These
in spine shape that are sensitive to blockers of actin dynamics effects occurred without gross changes in dendritic morpholo-
such as cytochalsin D10,13. gy (compare panels 1 and 3 of Fig. 1b).

articles
a a
a
control 10m
motility
20m AMPA
b
motility
10m
control washout
motility
b
motility b
control
motility
glutamate
AMPA
motility 10m
motility
Fig. 1. Glutamate inhibits actin dynamics in dendritic spines. Fig. 2. AMPA receptor stimulation inhibits actin-based dendritic spine
(a) Transfected hippocampal neuron expressing GFPactin throughout motility. (a) Individual fluorescence images (upper panels) and summed
the dendritic tree after 3 weeks in low-density culture (phase contrast, difference images (lower panels) for the same section of dendrite under
left; GFP fluorescence, right). (b) A 90 m segment of dendrite used for control conditions (top), in the presence of 1 M AMPA (middle) and
time-lapse recording of actin dynamics. Panels 1 and 3, individual frames after AMPA wash-out (bottom). The decrease in the summed image dif-
showing GFPactin fluorescence recorded under control conditions and ferences in the presence of AMPA demonstrates the inhibition of spine
after the addition of glutamate (100 M). The presence of glutamate motility, visible in the supplementary video data available on the Nature
does not disturb dendritic structure or actin distribution when NMDA Neuroscience web site. Even at this low magnification, the rounding effect
receptors are blocked as shown here by 100 M APV. Panels 2 and 4, on spine morphology produced by AMPA receptor activation can be
corresponding summed difference images showing that actin motility is seen in many of the spines. (b) Another experiment documenting the
primarily associated with dendritic spines (panel 2) and is strongly sup- effect of AMPA (2 m) on spine motility. In this case, NMDA receptors
pressed by the addition of glutamate (panel 4). were blocked with 100 M APV. See supplementary video data
Fig2b.mov for an original time-lapse recording of an experiment with
AMPA and APV, available on the Nature Neuroscience web site.
Spine motility is sensitive to AMPA receptor activation

To eliminate NMDA receptor-induced excitotoxic effects that
occur when hippocampal neurons are exposed to glutamate20,21, rapidly and was readily reversable when AMPA was removed
the experiments described above were conducted with the NMDA (Fig. 2). The level of AMPA required to inhibit spine motility was
receptor antagonist APV (100 M). This indicates that glutamate determined by applying increasing concentrations to cultures
can block actin-based spine motility independent of NMDA (n = 15). Inhibition was detectable at AMPA concentrations from
receptors and suggests the involvement of AMPA receptors in 500 nM (n = 4), with complete blockade of motility occurring
regulating actin dynamics in spines. Consequently we examined between 1 M (n = 4) and 2 M (n = 7). These effects, and sim-
the effect of exposing cells to AMPA and found that it too had a ilar ones produced by 100 M kainate (n = 6), were completely
strong inhibitory effect on spine motility. This effect developed blocked by the selective AMPA receptor antagonists CNQX (20

articles
complete inhibition of spine motility.

Despite bleaching of the GFPactin sig-
nal (produced by the long periods of
imaging necessary to record the effects
of drug treatment in slice cultures),
recovery of actin dynamics was still
observable after AMPA was removed
10m recovery
control +AMPA from the perfusing medium (Fig. 3,
right).
To assess the possible contribution
of NMDA receptors, we recorded actin
dynamics in spines exposed to 1 M
NMDA while AMPA receptors were
blocked with 100 M DNQX. In
recordings made at physiological salt
concentrations, NMDA had no
detectable effect on spine motility
control +AMPA recovery (n = 4; example in Fig. 4, compare pan-
Fig. 3. AMPA inhibits spine motility in organotypic slice cultures from hippocampus. Top panels show els 2 and 4). This might result from the
individual frames in a time-lapse recording of GFP fluorescence taken by confocal microscopy from a blockade of NMDA receptor activity by
living hippocampal slice established from a transgenic mouse expressing GFP-tagged actin and cul- Mg2+ that occurs at resting membrane
tured for four weeks. The labels identify the phase of the experiment from which the frames and their potential22,23. To test this possibility, we
corresponding difference images (bottom) were taken. The difference images are shown using a false examined the effect of NMDA on spine
color scale (left) and correspond to the area outlined by the red box in the top left panel. Actin motility when the Mg2+ concentration
dynamics during each phase were recorded in 60 frames over 15 min. AMPA was introduced into the
in the medium was reduced from
medium 54 min after the beginning of recording and was washed out 34 min before the recording
shown in the recovery phase. See the three parts of Fig3_.mov in the supplementary video data for
0.5 mM to 0.1 mM (n = 5). Under these
the original time-lapse data for this experiment. conditions, NMDA inhibited spine
motility in all 5 experiments (Fig. 4,
compare panels 4 and 6).
M, n = 6 for AMPA, n = 3 for kainate) and AMOA (15 M, AMPA receptor activation alters spine shape
n = 3 for AMPA, not determined for kainate). In addition to blocking spine motility, AMPA receptor activa-
As a further test for the receptor subtype involved, we tion also produced a distinct change of actin configuration in the
recorded the influence of AMPA on spine actin when NMDA spine, in which the irregular profile characteristic of motile spines
receptors were blocked by selective antagonists (Fig. 2b;
n = 17). Despite the high concentration of NMDA receptor
antagonist, AMPA completely blocked spine motility at either
1 M (n = 7) or 2 M (n = 10). Similar results were obtained in
experiments in which cells were treated with AMPA in the pres-
control
ence of the open-channel NMDA receptor blocker MK801 at
10 M (n = 6). In all experiments, control recordings were
made in the presence of APV or MK801 alone, before the addi-
tion of AMPA. Neither drug had any detectable effect on spine
shape or motility. motility
To ensure that these effects were not limited to neurons
growing in dispersed cultures, we also examined the influence
of AMPA on dendritic spine actin in confocal time-lapse
recordings of hippocampal slice cultures established from
NMDA
transgenic mice expressing GFP-tagged actin. A total of 8
experiments with increasing concentrations of AMPA (0.5 M,
n = 1; 1 M, n = 5, Fig. 3, left and center; 2 M, n = 2) showed
motility
Fig. 4. Inhibition of spine motility by NMDA requires lowering of the
external Mg2+ concentration. Panels 1, 3 and 5 show individual frames
taken from a time-lapse recording of GFPactin fluorescence under con-
trol conditions (panel 1), after the addition of 1 M NMDA (panel 3) and
with 1 M NMDA in low (0.1 mM) Mg2+ (panel 5). The corresponding 2+
NMDA/low Mg
difference images (motility) show that at normal concentrations of Mg2+
(0.5 mM) adding NMDA did not reduce spine motility (compare panels 2
and 4), but when Mg2+ was reduced to 0.1 mM, NMDA spine motility
was decreased significantly (panel 6). AMPA receptors were blocked with
100 M DNQX throughout the experiment. Scale bar, 10 m. See 10m
motility
Fig4.mov in the supplementary video data for the original time-lapse data.

articles
a 1 for varying degrees of irregularity (Methods). Changes in spine

morphology during a time-lapse recording can then be followed
by plotting the shape factor values against time (Fig. 5c).
The example in Fig. 5b shows outline profile data for a single
spine in successive frames of a time-lapse recording taken 15 sec-
onds apart. Following an initial control period when the spine
changed shape continuously (top row), its morphology became
stable after AMPA was added (middle row). When the drug was
washed out, shape changes recommenced (bottom row). This
sequence of profiles also demonstrates the rounding-up of spines
in the presence of AMPA (middle row), an effect that consistently
bb accompanied the inhibition of actin motility by AMPAreceptor
activation. Inspection of time-lapse recordings showed that
AMPA-induced spine rounding resulted from the cessation of
protrusive activity from the surface of the spine head and the
apparent collapse of actin into the core of the spine head
(Fig. 5a). This effect is most clearly seen in the original video
available as supplementary data on the Nature Neuroscience web
site (http://www.nature.com/neuro/web_specials). The existence

of a stable core of actin filaments in the spine head implied by
these results is in agreement with previous evidence demon-
strating that dendritic spines contain a set of actin filaments that
are resistant to actin-depolymerizing drugs24.
Both the inhibition of motility and the rounding up of spines
c are strikingly demonstrated in shape factor plots (two examples
in Fig. 5c). In both cases, the wide fluctuations in the shape fac-
tor that occurred under control conditions were blocked with
2 M AMPA. At the same time, the shape factor value moved
closer to 1, indicating that spine morphology had become more
stable and that spine shape had become rounder. When AMPA
was subsequently washed out, the shape factor began to fluctuate
again, indicating the resumption of spine shape changes (Fig. 5c,
right).
Motility inhibition depends on membrane depolarization

AMPA receptor activation causes the influx of Na+ ions through
receptor-associated channels and postsynaptic membrane depo-
larization. To test whether these events were involved in gluta-
mate receptor-induced inhibition of spine actin dynamics, we
examined the effect of AMPA on spine motility when Na+ was
Fig. 5. AMPA receptor activation leads to spine rounding. (a) Live cell removed from the medium (n = 4). All four experiments gave
images of GFPactin in a dendrite showing the effect of AMPA (1 M) on the same result (Fig. 6). Panel 1 shows shape factor data from
spine shape. Motile spines with irregular outlines (arrowheads, left) are an initial control recording made to establish the level of spine
round and immobile 15 min after addition of AMPA (right). (b) AMPA- motility. The medium was then changed to one in which Na+
induced spine shape change as revealed by profile outlines produced ions were removed by replacing NaCl with choline chloride.
from thresholded images of spine heads. Each row shows a single fluo- This had no significant effect on spine motility, which contin-
rescence image (left) and derived spine profiles taken 15 s apart. The
data are from a continuous recording of a single dendritic spine in con-
ued unabated (Fig. 6a, panel 2). Next, AMPA (2 M) was added
trol conditions (top row), in the presence of 1 M AMPA (middle) and in the medium lacking Na+. Under these conditions, AMPA did
after AMPA washout (bottom). The round profile of the spine during not inhibit spine motility (Fig. 6a, panel 3). Finally, when Na+
AMPA blockade of spine motility is evident. (c) Data for two further was re-introduced into the medium together with 2 M AMPA,
spines, taken from a separate experiment, displayed using shape factor spine motility was blocked, and the spines adopted the round-
analysis. The spine rounding and blockade of spine motility induced by ed profile typically found when AMPA receptors are activated
AMPA are shown by the increase in value of the shape factor and flat- (Fig. 6a, panel 4). These results indicate that the effects of AMPA
tening of the shape factor plots (see text). receptor activation on spine motility depend on the influx of
Na+ ions into the spine cytoplasm. Treatment with 500 nM ver-
atridine, which depolarizes the cell membrane by provoking
the opening of voltage-dependent Na+ channels, also blocked
was replaced by a more rounded appearance of the spine head spine motility and produced spine rounding (n = 4; Fig. 6b).
(Fig. 5a). To analyze this effect in more detail, we derived out- However, blocking voltage-dependent Na+ channels with 1 M
line profiles of individual spines (n = 9, Fig. 5b) using data in tetrodotoxin (TTX) did not suppress the inhibitory effect of
three separate experiments with independently established cul- AMPA on spine motility (n = 4, Fig6_TTX.mov, supplemen-
tures. These profiles were used to calculate a shape factor that tary data on the Nature Neuroscience web site). This suggests
returns a value of 1 for a perfectly round profile and values below that Na+ ions that mediate the effects of AMPA receptor acti-

articles
vation enter the spine aa b

b 1
sf 1
cytoplasm through chan-
nels associated with the 0.75
0.75
receptor itself.
To determine whether
the effects of Na + were 0.5 0.5
mediated by depolarizing
the cell membrane, we 0.25
0.25
raised extracellular K +
ctl 1 min
from 2 to 8 mM (n = 5). 0
ctl 1min
0
This caused an immediate 1
blockade of spine motility 1
accompanied by spine 0.75
rounding (Fig. 7a), which 0.75

recovered spontaneously 0.5
as the cell adjusted to the

0.5
altered external K + con- 0.25
centration (Fig. 7b).

Increasing external K+ also no Na+ 0.25

0
inhibited spine motility 1
when AMPA receptors veratridine
0
were blocked with 100 M 0.75
DNQX (n = 2), when
Fig. 6. Inhibition of spine motility by AMPA depends on
NMDA receptors were 0.5 extracellular Na+. (a) Shape factor plots for individual den-
blocked with 100 M APV dritic spines recorded in a single experiment. Spine shape
(n = 2), or when both 0.25 changes seen under control conditions (ctl) continued when
blockers were present Na+ was removed from the medium (no Na+). When 2 M
(n = 2). This suggests that no Na+/AMPA AMPA was added in the absence of Na+, shape changes con-
membrane depolarization 0
tinued undiminished (no Na+/AMPA). After Na+ was reintro-
inhibits spine motility 1 duced into the medium, AMPA effectively blocked shape
downstream of glutamate changes and induced spines to round up, shown by the
receptor activation. Alto- 0.75 increase in the shape factor value toward 1 (Na+/AMPA).
See supplemental video data in Fig6.mov for the original
gether these results sug-
time-lapse recording. Other experiments showed that
gest that depolarization of 0.5
blocking voltage-dependent Na+ channels with tetrodotoxin
the postsynaptic mem- did not affect the influence of AMPA on spine motility (see
brane is both necessary 0.25
Fig6_ttx.mov for time-lapse recording). (b) Shape factor
and sufficient for the plots for 6 spines showing the effect of veratridine-stimu-
Na+/AMPA
inhibitory effect of gluta- 0 lated Na+ influx on actin motility and spine shape.
mate receptors on spine
actin dynamics.
Calcium mediates the effect of receptor activation at 1 M, n = 4) and P/Q channels (-agatoxin TK at 50 nM,
The most likely mechanism by which glutamate receptor-depen- n = 1; 250 nM, n = 1; FTX 3.3 at 1 M, n = 4). None of these
dent membrane depolarization might influence spine motility detectably reduced the inhibition of spine motility by 2 M
is via Ca2+, whose levels can be regulated in individual dendrit- AMPA. In contrast, blockers selective for low-voltage-activated
ic spines by synaptic transmission2527 and which can influence Ca2+ channels were more effective. Ni2+ (10 M) reduced AMPA
spine actin through a variety of pathways 21,28,29. The AMPA inhibition of spine motility by 66 4% (n = 3), whereas 10 M
receptors of hippocampal pyramidal cells contain GluR2/GluRB mibefradil produced a 38 9% reduction (n = 4). This distinc-
subunits, which render their associated ion channels Ca2+ imper- tion between the two groups of agents is clearer in the original
meable30,31. This leaves voltage-activated Ca2+ channels (VACs) as time-lapse recordings, examples of which are shown in the sup-
the most likely source of glutamate receptor-mediated Ca 2+ plementary video data available on the web site. These results
influx into spines, a conclusion consistent with the effectiveness suggest that low- rather than high-voltage-activated Ca2+ chan-
of K+-induced depolarization in blocking spine actin dynamics. nels mediate the inhibitory effects of AMPA receptors on post-
We examined this possibility by testing the ability of different synaptic actin dynamics.
VAC antagonists3234 to suppress the inhibitory effect of AMPA
on spine motility. The results were quantified by summing pixel DISCUSSION
densities obtained by subtracting the gray scale values of pixels Our experiments show that activation of glutamate receptors
between frames of time-lapse sequences. Among blockers of can regulate synaptic plasticity by influencing actin dynamics
high-voltage-activated Ca2+ channels, Cd2+ (500 M) suppressed in dendritic spines. The direction of this effect is significant:
the effects of AMPA by 6 2 % (n = 4), nifedipine (20 M) by actin motility was inhibited when glutamate receptors were
21 10% (n = 2) and nimodipine (20 M) by 24 9% (n = 3). stimulated, implying that spine morphology is stabilized by
We also tested antagonists selective for N-type channels (-cono- signal transmission at glutamatergic synapses. These effects
toxin MVIIA at 2 M, n = 3), Q channels (-conotoxin MVIIC occurred in hippocampal neurons after 34 weeks in culture,

articles
aa
is consistent with the failure of AMPA to block spine actin
dynamics in the absence of extracellular Na+, the normal charge
carrier for AMPA receptor ion channels, and with the lack of
effect of the voltage-dependent Na+ channel blocker TTX on
spine motility. The effects of glutamate receptor activation seem
to depend on depolarization of the postsynaptic membrane
because they were mimicked by raising extracellular K+. This
treatment was effective when K+ was increased in the presence
of AMPA or NMDA receptor antagonists, indicating that the
intermediary depolarization event lies downstream of the recep-
tors. The effects of depolarization in turn seem to be mediated
by influx of Ca2+ through VACs, known to be activated follow-
ing membrane depolarization induced by Na+ influx38,39. The
ctrl
comparative effects of selective Ca 2+ channel antagonists in
reducing AMPA-induced inhibition of spine motility favor the
involvement of low-voltage T-type channels. However, because
we lack specific antagonists for this channel type, a definitive
identification is not yet possible.
KCl Distinct roles of AMPA and NMDA receptors

Spine outgrowth from dendrites following NMDA receptor-
1
dependent stimulation protocols11,12,16 is commonly accompa-
bb nied by motile activity of the nascent spine9,11. In contrast, spines
at established synapses require continual activation of AMPA
0.75
receptors for their maintenance15. Taken together with our present
data, these results suggest the existence of two distinct modes of
KCl morphological plasticity in dendritic spines, one in which the
0.5
formation of new spines is initiated by LTP-like stimulation oper-
ating through NMDA receptors, and a second in which spine
0.25
morphology at established synapses is stabilized by AMPA recep-
tor activation. This interpretation is consistent with evidence that
many newly formed synapses exhibit only NMDA receptor-medi-
0
2min ated currents with AMPA receptor-based responses emerging
only later by the physical acquisition of AMPA receptors4042. The
Fig. 7. Membrane depolarization inhibits spine motility. (a) A single association of NMDA receptor-dependent LTP with the emer-
frame (top panel) and two difference images showing spine motility gence of new spines11,12,16 and of AMPA receptor activation with
under control conditions and when K+ concentration in the medium was the stabilization of established synapses shown here therefore
raised from 2 mM (control) to 8 mM (KCl). Scale bar, 10 m. (b) Another parallels the physiological maturation of these same glutamater-
experiment showing the effect of KCl on spine motility documented gic synapses.
using the shape factor plot. At the point indicated by the arrow, medium Glutamatergic transmission leading to an increase of postsy-
containing elevated K+ (8 mM KCl) was introduced into the chamber, naptic Ca2+ levels can thus have opposite effects on spine mor-
producing a transitory inhibition of spine motility accompanied by round-
phology. A precedent for this seemingly paradoxical situation is
ing of the spine profile. This effect reversed spontaneously and fluctua-
tions in spine shape resumed. See Fig7.mov, available as supplementary
shown in growth cones, where Ca2+ can promote or suppress
information, for the original time-lapse recording to Fig. 7b. actin-based filopodial activity depending on extracellular signals
and ancillary circumstances4345. In dendritic spines, the con-
trasting effects of Ca2+ elevation at different stages of synapse for-
mation may be related to the different conditions appropriate for
making synapses compared to those required for maintaining
when spines are contacted by presynaptic terminals10 and are them. Forging new connections is appropriate when afferent
part of ultrastructurally intact synaptic contacts35,36. The asso- stimulation is significantly above background, conditions that
ciation of this stabilizing influence of glutamate receptors with are represented by the repetitive, high-frequency stimulation
established synapses is further supported by the persistence of needed for NMDA receptor-dependent induction of new spines.
AMPA receptor-regulated spine motility in brain slices after Conversely, to be useful for information storage, synapses should
several months in culture, long after spine synapses have been be maintained by a mechanism that is relatively stimulus insen-
established37. sitive, conditions that are met by the sensitive AMPA receptor-
Actin dynamics in dendritic spines could be blocked by acti- dependent mechanism for regulating postsynaptic actin dynamics
vation of both AMPA and NMDA receptors. However, AMPA described here.
was equally effective when NMDA receptors were blocked by
selective antagonists, indicating that AMPA receptors can sta- METHODS
bilize spine motility independent of NMDA receptors. In addi- Materials. -cytoplasmic actin tagged with green fluorescent protein was
tion to the effects of NMDA receptors shown in previous expressed from a eukaryotic expression plasmid containing a chicken
studies11,12,16, these results implicate AMPA receptors in the reg- -actin promoter10. Active compounds were obtained from the follow-
ulation of spine plasticity. The involvement of AMPA receptors ing sources: amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid

articles
(AMPA), D(-)-2-amino-5-phosphonopentanoic acid (APV) , N-methyl- ACKNOWLEDGEMENTS

D-aspartate (NMDA), 2-amino-3-[3- (carboxymethoxy)-5-methyl-isox- We thank Beat Ludin for assistance with image analysis and John Kemp for sup-
azol-4-yl]proprionic acid (AMOA), (5R, 10S)-(+)-5-methyl-10, plying mibefradil.
11-dihydro-5H-dibenzo[a,d]cyclohepten-5,10-imine hydrogen maleate
(MK-801), 6,7-dinitroquinoxaline-2,3(1h,4h)-dione (DNQX) and
kainate from RBI, Natick, Massachusetts; tetrodotoxin (TTX), -cono- RECEIVED 2 JUNE; ACCEPTED 11 JULY 2000
toxin MVIIA, -conotoxin MVIIC, -agatoxin TK and FTX 3.3 (from
Agelenopsis aperta) from Alomone Laboratories, Jerusalem, Israel. 1. Feldman, D. E. & Knudsen, E. I. Experience-dependent plasticity and the
Mibefradil was supplied by J. Kemp, HoffmannLa Roche, Basel, maturation of glutamatergic synapses. Neuron 20, 10671071 (1998).
Switzerland. 2. Gray, E. G. Electron microscopy of synaptic contacts on dendritic spines of
the cerebral cortex. Nature 183, 15921593 (1959).
3. Carlin, R. K. & Siekevitz, P. Plasticity in the central nervous system: do
Cell culture and microscopy. E19 primary hippocampal neurons were synapses divide? Proc. Natl. Acad. Sci. USA 80, 35173521 (1983).
transfected with GFPactin as described10. Cultures of dispersed cells46 4. Harris, K. M. & Kater, S. B. Dendritic spines: cellular specializations
were maintained in glia-conditioned, serum-free medium for 34 imparting both stability and flexibility to synaptic function. Annu. Rev.
Neurosci. 17, 341371 (1994).
weeks, until the pyramidal cells developed spines of mature appear- 5. Frotscher, M., Mannsfeld, B. & Wenzel, J. Environmentally affected
ance. Spines were observed at 37C in purpose-built observation differentiation of dendrite spines of pyramidal neurons in rat hippocampus
chambers (Life Imaging Services, Olten, Switzerland) using a GFP- (CA1). J. Hirnforsch 16, 443450 (1975).
optimized filter set (Chroma Technologies, Brattleboro, Vermont). 6. Moser, M. B., Trommald, M. & Andersen, P. An increase in dendritic spine
Illumination intensity was adjusted using neutral-density filters and density on hippocampal CA1 pyramidal cells following spatial learning in
adult rats suggests the formation of new synapses. Proc. Natl. Acad. Sci. USA
images taken every 15 s with a 2 s exposure using a MicroMax cooled
91, 1267312675 (1994).

CCD camera (Princeton Instruments, Trenton, New Jersey) and Meta- 7. Muller, D. Ultrastructural plasticity of excitatory synapses. Rev. Neurosci. 8,
morph 3.0 imaging software (Universal Imaging Corporation, West 7793 (1997).
Chester, Pennsylvania). Drugs were either added directly to the imag- 8. Dailey, M. E. & Smith, S. J. The dynamics of dendritic structure in developing
hippocampal slices. J. Neurosci. 16, 29832994 (1996).
ing chamber or introduced by gravity-feed perfusion. Similar results
9. Ziv, N. E. & Smith, S. J. Evidence for a role of dendritic filopodia in
were obtained using glia-conditioned medium or in Tyrodes solu- synaptogenesis and spine formation. Neuron 17, 91102 (1996).
tion (Life Technologies, Basel, Switzerland) containing glucose 10. Fischer, M., Kaech, S., Knutti, D. & Matus, A. Rapid actin-based plasticity in
(11 mM) and supplemented with 10 mM glycine and 0.2 g/l Na2HPO4 dendritic spines. Neuron 20, 847854 (1998).
adjusted to pH 7.2 to buffer the medium during prolonged imaging 11. Maletic-Savatic, M., Malinow, R. & Svoboda, K. Rapid dendritic
morphogenesis in CA1 hippocampal dendrites induced by synaptic activity.
without CO 2 gassing. Low Ca 2+ medium consisted of this supple- Science 283, 19231927 (1999).
mented Tyrodes solution containing 0.1 mM instead of 2 mM Ca2+ 12. Engert, F. & Bonhoeffer, T. Dendritic spine changes associated with
and 2 mM Mg2+ instead of 0.5 mM. In some experiments, Mg2+ was hippocampal long-term synaptic plasticity. Nature 399, 6670 (1999).
reduced to 0.1 mM without further changes to the medium. For 13. Dunaevsky, A., Tashiro, A., Majewska, A., Mason, C. & Yuste, R.
experiments in Na+-free conditions, NaCl was replaced by 110 mM Developmental regulation of spine motility in the mammalian central
nervous system. Proc. Natl. Acad. Sci. USA 96, 1343813443 (1999).
choline chloride. 14. Lendvai, B., Stern, E. A., Chen, B. & Svoboda, K. Experience-dependent
plasticity of dendritic spines in the developing rat barrel cortex in vivo.
Organotypic cultures from transgenic mice. Transgenic mice were Nature 404, 876881 (2000).
15. McKinney, R. A., Capogna, M., Durr, R., Gahwiler, B. H. & Thompson, S. M.
generated using a 5.5 kb Nde1/EcoR1-fragment containing C-terminal Miniature synaptic events maintain dendritic spines via AMPA receptor
GFP-tagged -cytoplasmic actin10 injected into oocytes of B6CF1 mice. activation. Nat. Neurosci. 2, 4449 (1999).
Animals were analyzed by Southern blotting or direct inspection using 16. Toni, N., Buchs, P. A., Nikonenko, I., Bron, C. R. & Muller, D. LTP promotes
a Leitz MZ 12 dissection microscope fitted with fluorescence optics. formation of multiple spine synapses between a single axon terminal and a
Organotypic slice cultures from hippocampus were prepared accord- dendrite. Nature 402, 421425 (1999).
17. Kaech, S., Fisher, M., Doll, T. & Matus, A. Isoform specificity in the
ing to described procedures47. After a minimum of 4 weeks, slice cul- relationship of actin to dendritic spines. J. Neurosci. 17, 95659572 (1997).
tures were observed at 37C under perfusion by gravity feed with 18. Marin-Padilla, M. Number and distribution of apical dendritic spines of the
artificial cerebrospinal fluid (124 mM NaCl, 2.5 mM KCl, 2.0 mM layer V pyramidal cells in man. J. Comp. Neurol. 131, 475490 (1967).
MgSO4, 1.25 mM KH2PO4, 26 mM NaHCO3, 10 mM glucose, 4 mM 19. Spacek, J. & Hartmann, M. Three-dimensional analysis of dendritic spines. I.
sucrose, 2.5 mM CaCl2 saturated with 95%O2/5% CO2) or Tyrodes Quantitative observations related to dendritic spine and synaptic
morphology in cerebral and cerebellar cortices. Anat. Embryol. (Berl.) 167,
buffer (Gibco, Basel, Switzerland). AMPA was added at 1 M. 289310 (1983).
Dynamic imaging of actin-GFP in slice cultures were performed using 20. Choi, D. W. Glutamate receptors and the induction of excitotoxic neuronal
a Leica DM-IRBE inverted microscope fitted with a Yokogawa microlens death. Prog. Brain. Res. 100, 4751 (1994).
Nipkow confocal system (Life Science Resources, Cambridge, UK). 21. Halpain, S., Hipolito, A. & Saffer, L. Regulation of F-actin stability in
dendritic spines by glutamate receptors and calcineurin. J. Neurosci. 18,
Summed difference images were derived from time-lapse recordings using 98359844 (1998).
MetaMorph image analysis software. 22. Mayer, M. L., Westbrook, G. L. & Guthrie, P. B. Voltage-dependent block by
Mg2+ of NMDA responses in spinal cord neurones. Nature 309, 261263
Image analysis. Outlines of spines were created from thresholded (1984).
23. Nowak, L., Bregestovski, P., Ascher, P., Herbet, A. & Prochiantz, A.
images using edge-detection functions of Metamorph 3.0. Changes in Magnesium gates glutamate-activated channels in mouse central neurones.
spine shape were then assessed using the shape factor routine Nature 307, 462465 (1984).
(sf = 4A/p2) in the Metamorph program, which is calculated from 24. Allison, D. W., Gelfand, V. I., Spector, I. & Craig, A. M. Role of actin in
the perimeter p and the area A of the object (spine head). Values close anchoring postsynaptic receptors in cultured hippocampal neurons:
differential attachment of NMDA versus AMPA receptors. J. Neurosci. 18,
to 0 represent elongated or ruffled shapes, whereas a value of 1 denotes 24232436 (1998).
a perfect circle. Data were directly recorded, analyzed and displayed 25. Guthrie, P. B., Segal, M. & Kater, S. B. Independent regulation of calcium
using Microsoft Excel. To visualize drug effects, arithmetical pixel- revealed by imaging dendritic spines. Nature 354, 7680 (1991).
intensity differences of the same range of consecutive frames before, 26. Denk, W., Sugimori, M. & Llinas, R. Two types of calcium response limited to
during and after drug treatment were summed. The resulting images single spines in cerebellar Purkinje cells. Proc. Natl. Acad. Sci. USA 92,
82798282 (1995).
were combined and scaled to display large differences between the 27. Koester, H. J. & Sakmann, B. Calcium dynamics in single spines during
original images as dark pixels (gray scale) or using a pseudo-color scale coincident pre- and postsynaptic activity depend on relative timing of
as illustrated. back-propagating action potentials and subthreshold excitatory
postsynaptic potentials. Proc. Natl. Acad. Sci. USA 95, 95969601 (1998).
28. Furukawa, K. et al. The actin-severing protein gelsolin modulates calcium
Note: Supplementary video data can be found on the Nature Neuroscience channel and NMDA receptor activities and vulnerability to excitotoxicity in
web site (http://www.nature.com/neuro/web_specials/). hippocampal neurons. J. Neurosci. 17, 81788186 (1997).

articles
29. Matus, A. Postsynaptic actin and neuronal plasticity. Curr. Opin. Neurobiol. 38. Magee, J., Hoffman, D., Colbert, C. & Johnston, D. Electrical and calcium
9, 561565 (1999). signaling in dendrites of hippocampal pyramidal neurons. Annu. Rev. Physiol.
30. Hollmann, M., Hartley, M. & Heinemann, S. Ca2+ permeability of KA- 60, 327346 (1998).
AMPA-gated glutamate receptor channels depends on subunit composition. 39. Svoboda, K. & Mainen, Z. F. Synaptic [Ca2+]: intracellular stores spill their
Science 252, 851853 (1991). guts. Neuron 22, 427430 (1999).
31. Geiger, J. R. et al. Relative abundance of subunit mRNAs determines gating 40. Liao, D., Zhang, X., OBrien, R., Ehlers, M. D. & Huganir, R. L. Regulation of
and Ca2+ permeability of AMPA receptors in principal neurons and morphological postsynaptic silent synapses in developing hippocampal
interneurons in rat CNS. Neuron 15, 193204 (1995). neurons. Nat. Neurosci. 2, 3743 (1999).
32. Markram, H. & Sakmann, B. Calcium transients in dendrites of neocortical 41. Petralia, R. S. et al. Selective acquisition of AMPA receptors over postnatal
neurons evoked by single subthreshold excitatory postsynaptic potentials via development suggests a molecular basis for silent synapses. Nat. Neurosci. 2,
low-voltage-activated calcium channels. Proc. Natl. Acad. Sci. USA 91, 3136 (1999).
52075211 (1994). 42. Shi, S. H. et al. Rapid spine delivery and redistribution of AMPA receptors
33. Avery, R. B. & Johnston, D. Multiple channel types contribute to the low- after synaptic NMDA receptor activation. Science 284, 18111816 (1999).
voltage-activated calcium current in hippocampal CA3 pyramidal neurons. 43. Kater, S. B. & Mills, L. R. Regulation of growth cone behavior by calcium.
J. Neurosci. 16, 55675582 (1996). J. Neurosci. 11, 891899 (1991).
34. McDonough, S. I. & Bean, B. P. Mibefradil inhibition of T-type calcium 44. Hong, K., Nishiyama, M., Henley, J., Tessier-Lavigne, M. & Poo, M. Calcium
channels in cerebellar purkinje neurons. Mol. Pharmacol. 54, 10801087 (1998). signalling in the guidance of nerve growth by netrin-1. Nature 403, 9398 (2000).
35. Bartlett, W. P. & Banker, G. A. An electron microscopic study of the 45. Zheng, J. Q. Turning of nerve growth cones induced by localized increases in
development of axons and dendrites by hippocampal neurons in culture. II. intracellular calcium ions. Nature 403, 8993 (2000).
Synaptic relationships. J. Neurosci. 4, 19541965 (1984). 46. Banker, G. & Goslin, K. in Culturing Nerve Cells (eds. Banker, G. & Goslin, K.)
36. Boyer, C., Schikorski, T. & Stevens, C. F. Comparison of hippocampal 251281 (Bradford Books, MIT Press, Cambridge, 1991).
dendritic spines in culture and in brain. J. Neurosci. 18, 52945300 (1998). 47. Gahwiler, B. H., Thompson, S. M., Audinat, E. & Robertson, R. T. in
37. Gahwiler, B. H. Development of the hippocampus in vitro: cell types, Culturing Nerve Cells (eds. Banker, G. & Goslin, K.) 379411 (Bradford
synapses and receptors. Neuroscience 11, 751760 (1984). Books, MIT Press, Cambridge, Massachusetts, 1991).

articles
Somatic EPSP amplitude is

independent of synapse location in
hippocampal pyramidal neurons
Jeffrey C. Magee1 and Erik P. Cook2
1 Neuroscience Center, Louisiana State University Medical Center, 2020 Gravier St., New Orleans, Louisiana 70112, USA
2 Howard Hughes Medical Institute, Baylor College of Medicine, One Baylor Plaza, Houston, Texas 77050, USA
Correspondence should be addressed to J.C.M. (jmagee@lsumc.edu)
Most neurons receive thousands of synaptic inputs onto widely spread dendrites. Because of
dendritic filtering, distant synapses should have less efficacy than proximal ones. To investigate this,
we characterized the amplitude and kinetics of excitatory synaptic input across the apical dendrites
of CA1 pyramidal neurons using dual whole-cell recordings. We found that dendritic EPSP amplitude
increases with distance from the soma, counterbalancing the filtering effects of the dendrites and
reducing the location dependence of somatic EPSP amplitude. Dendritic current injections and a
multi-compartmental computer model demonstrated that dendritic membrane properties have only
a minor role in elevating the local EPSP. Instead a progressive increase in synaptic conductance
seems to be primarily responsible for normalizing the amplitudes of individual inputs.
Like many other neurons in the CNS, hippocampal CA1 pyra- of both neocortical and hippocampal pyramidal neurons are
midal neurons receive tens of thousands of excitatory synaptic more sensitive to glutamate than the proximal dendrites 14
contacts over several hundred microns of their apical dendritic (A. Frick, W. Zieglgansberger & H.U. Dodt, Soc. Neurosci. Abstr.
arborizations1. Dendritic filtering reduces the amplitude and 24, 325, 1998). This may indicate an increase in local synaptic
slows the kinetics of synaptic input. The amount of filtering amplitude that, with distance, could counterbalance the ampli-
directly depends on the electrotonic distance of the synapse from tude-filtering effects of the dendrites. Others have proposed that
the final integration site (in this case, the soma or proximal an increase in synaptic conductance with distance from the soma
axon)25. Because of the widespread spatial distribution of synap- reduces the location dependence of synaptic activity1518. We set
tic input onto these neurons, individual inputs receive widely out to directly examine this idea by determining the location
varying amounts of filtering depending on synapse location. This dependence of AMPA receptor-mediated excitatory synaptic
location-dependent synaptic variability would result in distal syn- input in hippocampal CA1 pyramidal neurons.
apses having less impact on the firing state of the neuron than
their proximal counterparts. There are, however, many potential RESULTS
mechanisms available to reduce location-dependent synaptic Using dual whole-cell recordings from hippocampal CA1 pyra-
variability in central neurons. These include passive and active midal neurons, we simultaneously measured the amplitude of
postsynaptic membrane properties as well as various pre- and putative unitary (single terminal) EPSPs both at the dendritic
postsynaptic properties of the synapses themselves. input site and at the soma (Fig. 1a and b). The location of the
The dendrites of CA1 pyramidal neurons contain a variety of synaptic input and the dendritic recording pipette were varied
voltage-gated ion channels that could counter some of the filter- together from cell to cell, which insured a recording directly from
ing effects of the arborization (reviewed in ref. 6). However, in the site of input paired with another at the soma. Unitary events
CA1 cells, the main impact of these ion channels on synaptic were evoked primarily using the localized application of high
input seems to be a reduction in the kinetic distortion rather than (600 mOsM) osmolarity external solution to the apical dendrite
a boosting of the distal synaptic amplitude79 (but see refs. 10, around the dendritic recording pipette. Minimal stimulation
11). Neuron morphology can also counter location-dependent techniques were also used (Methods).
synaptic variability. The changes in input impedance and capac- For each cell, all EPSPs recorded at the soma were used to
itance that occur in distal dendritic compartments (impedance calculate an average somatic EPSP, and all EPSPs simultane-
increases, whereas capacitance decreases) could act to spatially ously recorded at the input site were used to calculate an aver-
normalize EPSP amplitude and kinetics to some degree5,12,13. It age dendritic EPSP (Fig. 1d). To determine the location
is unlikely, however, that the passive geometry of pyramidal neu- dependence of unitary input, we plotted the amplitudes of the
rons can substantially reduce the location-dependent variability average somatic and dendritic EPSPs against synapse location
of relatively high-frequency signals like synaptic inputs. for the entire group of cells (Fig. 1e). The amplitude of the
If dendritic mechanisms for reducing location-dependent local dendritic EPSP, recorded at the site of input, increased
amplitude variability do exist, they are most likely to be found nearly fourfold with distance of the synapse from the soma
at the synapse itself. In support of this, the distal apical dendrites (0.25 to 0.8 mV from 50 to 325 m; Fig. 1e). On the other

articles
Fig. 1. Synaptically evoked EPSP amplitude at the

soma is independent of synapse location.
(a) Dendritic membrane potential just before the a
localized application of high osmolar external solu-
tion (600 mOsM) to a distal dendritic location
(300 m, upper trace, baseline). Application of
high osmolar external solution results in the spon-
taneous occurrence of EPSPs that are simultane-
ously recorded at both the dendritic input site (d)
and at the soma (s; lower traces, +sucrose). (b) In
another neuron, EPSPs are evoked more proxi- b
mally (50 m) and recorded from the dendritic
site (d) and the soma (s). (c) Scatterplot of den-
dritic and simultaneously recorded somatic EPSP
amplitude from the cell shown in (a).(d) The aver-
age of all EPSPs simultaneously recorded both at
the site of input (dendrite) and at the soma for the
neuron receiving distal input (distal; average of 93;
same cell shown in a) and for the neuron receiving
proximal input (prox; average of 122; same cell c d Average EPSP
shown in b). Dendritic amplitude is largest in the

neuron receiving more distal input, whereas
somatic EPSP amplitude is similar for both neu-
rons. (e) Mean EPSP amplitude for all cells plotted
as a function of input distance from the soma
for both dendritic and somatic recordings.
(f) Cumulative amplitude histograms for the above
recordings showing that the distribution of distal
dendritic EPSPs (light solid line) is skewed to the
right compared to more proximal dendritic EPSPs
(dark solid line). Also the distributions of EPSPs
from both input locations are similar once they e f
reach the soma (distal, light dashed line; proximal,
dark dashed line). Traces were digitally smoothed
Peak amplitude (mV)
as in Methods. All EPSPs shown were evoked by

high osmolar solution.
hand, EPSPs recorded simultaneously at the

soma exhibited an average amplitude that
was virtually independent of synapse loca-
tion (Fig. 1e; 0.2 mV from 50 to 325 m;
EPSC amplitudes ranged from 4.3 to 5.6 pA
for four recordings from 50290 m). The values for somatic EPSC-shaped currents (30 pA, r = 0.1 ms, d = 3 ms) directly
EPSP/Cs are similar to those reported for single synaptic or into the dendrites. Local dendritic and propagated somatic
quantal events recorded at the soma1723. voltage transients (EPSP i ) were simultaneously measured
Cumulative histograms of EPSP amplitudes were generated (Fig. 2a). In contrast to the synaptically generated EPSPs, the
for both local dendritic and propagated somatic EPSPs (Fig. 1f). EPSP is were location dependent: somatic EPSP i amplitude
The amplitude distribution of distal events recorded at the den- decreased to less than half as the input location was moved dis-
dritic site of input was shifted toward larger amplitudes when tally (0.5 to 0.19 mV from 50 to 325 m; Fig. 2b). The ampli-
compared to proximal events recorded at the input site (Fig. 1f, tude of the EPSPi at the dendritic injection site did increase
light solid line versus dark solid line; range of distal, 0.1 to 4 mV; slightly with distance from the soma (50%, Fig. 2b) but much
proximal, 0.1 to 1.5 mV). Following propagation to the soma, less than the synaptically evoked EPSP (Fig. 1e). The level of
however, the amplitude distributions of both proximal and distal amplitude attenuation from the site of input to the soma was
inputs were similarly shaped and showed a smaller range of similar for synaptically evoked EPSPs and EPSPi (from 80% at
amplitudes than those recorded at the dendrite (Fig. 1f, dashed the most distal to 20% at the most proximal; Fig. 2c), demon-
lines). These data suggest that the amplitude of a local dendritic strating that a difference in propagation was not responsible
EPSP is increased sufficiently to counter the electrotonic filter- for the location dependence of the EPSPi. Together these data
ing it will experience as it propagates to the soma. The final result indicate that dendritic membrane mechanisms are not suffi-
is that the mean amplitude of all synaptic potentials at the soma cient to increase the local dendritic amplitude of the EPSP and
does not depend on synapse location. We next examined the dis- counter dendritic filtering. Therefore, as current injections of
tinct roles of active and passive postsynaptic membrane proper- proper size and shape could not reproduce the spatial EPSP
ties, as well as the properties of the Schaffer collateral synapses profile observed with actual synaptic input (< 50%; Fig. 1d),
themselves, in producing this effect. we conclude that the passive structure and active properties of
To examine the contribution of postsynaptic membrane CA1 dendrites cannot account for the reduction in the loca-
properties, we bypassed the synaptic machinery by injecting tion dependence of synaptic input.

articles
Fig. 2. The somatic amplitude of EPSPs generated by uniform current

injections depends on input location. (a) Traces are the average voltage a
transients (EPSPI, average of 5 sweeps) recorded in the dendrite (den-
drite) and simultaneously at the soma in response to EPSC-shaped cur-
rent injections into the dendrites. These recordings are from the same
neurons shown in Fig. 1 (distal, Fig. 1a; proximal, Fig. 1b). The den-
dritic EPSPi amplitude is slightly increased in the more distal dendritic
recording, whereas the somatic EPSPi is much larger for proximal input.
(b) The average EPSPI amplitudes are plotted as a function of input dis- Distal Prox
tance from the soma for both dendritic () and somatic () recordings.
Note that dendritic EPSPi amplitude does not increase as much as
synaptic EPSP amplitude (Fig. 1c) and that somatic EPSPI amplitude
decreases with distance. (c) EPSP attenuation is plotted as a function of
input distance from the soma for both synaptic EPSPs () and current- b c
injection EPSPI (). Attenuation is expressed as the decrease in EPSP
Peak amplitude (mV)
(dend soma)/dend
amplitude (dend - soma) relative to the initial amplitude (dend). Traces
were digitally smoothed as described in Methods.
The location and amplitude dependence of EPSP kinetics can

reveal the forces shaping EPSP propagation. Therefore both the
actual synaptic events as well as EPSPi were fit by an exponential
function that provided time constants for the rising (r) and
Distance (m from soma) Distance (m from soma)
decaying (d) phases of the synaptic potentials. We then exam-
ined their dependence on synapse location (Fig. 3ad) or ampli-
tude (Fig. 3eh). For local (dendritic) events, the rise time of reported in these cells and is presumably the result of both active
EPSPs of similar amplitude (0.20.6 mV) did not seem to depend and passive dendritic membrane mechanisms8,9. The duration
on location (Fig. 3a). On the other hand, both the EPSP and of the synaptic conductance is unlikely to depend on location
EPSPi decay time constants decreased substantially with distance because there was a similar change in the decay time constant of
from the soma (Fig. 3b). A similar location dependence has been both the local EPSP and the EPSPi (Fig. 3b).
e f Local decay
a Local rise b Local decay Local rise
Tau (ms)
Tau (ms)
Tau (ms)
Tau (ms)
Amplitude (mV) Amplitude (mV)

Distance (m) Distance (m)
c Soma rise d Soma decay

g Local rise h Local decay
Tau (ms)
Tau (ms)
Tau (ms)
Tau (ms)
Distance (m) Distance (m)

Amplitude (mV) Amplitude (mV)
Fig. 3. Location and amplitude dependence of EPSP kinetics. Rise (a) and decay (b) times for average EPSPs recorded at the dendritic input site as a
function of distance from the soma (time constants from fit by an exponential function; fits are constrained to the first 56 ms of the EPSP decay phase
for decay times). There is no location dependence to EPSP rise times, whereas decay times drop sharply with distance. Rise (c) and decay (d) times for
average EPSPs recorded at the soma, as a function of input distance from the soma (propagated EPSPs simultaneously recorded with the dendritic EPSPs
plotted above). Somatic EPSP rise times show a more pronounced location dependence than do somatic EPSP decay times. EPSPs evoked by high osmo-
lar solution (, solid line fits) show similar location dependence as EPSPi produced by dendritic current injections (, dashed line fits). (eh) Rise and
decay times for all EPSPs recorded at a distal (e, f) or proximal (g, h) input site as a function of EPSP amplitude. The rise times of both distal and prox-
imal EPSPs (e, g, , solid line fits) increase with amplitude, whereas the rise times of EPSPi do not (e, g, , dashed line fits). Decay times for distal events
do not show any amplitude dependence (f). Proximal events show a slight increase of decay time as amplitude increased (h). Data points were fit by
either a linear or single exponential function. Kinetics from EPSPs evoked by high osmolar stimulation.

articles
Fig. 4. An increase in synaptic conductance can account for

the normalization of somatic EPSP amplitude in a passive a b
computer model. (a) Reconstructed CA1 pyramidal neuron
used in multi-compartmental model. (b) EPSPs recorded at
the site of input (dendrite) and simultaneously at the soma
(soma) for a proximal (proximal; 100 m from the soma) and
more distal (distal, 300 m from the soma) synaptic input of
uniform conductance (872 pS). Also shown are EPSPs for the
more distal input with the synaptic conductance increased
approximately twofold (distal X2, 1600 pS). Inset, proximal
100 m
and distal dendritic EPSPs with amplitude normalized to
show differences in decay kinetics. (c) EPSP amplitudes plot-
ted as a function of input distance from the soma for both
dendritic () and somatic () recordings for the 872 pS
synaptic input. Note that the dendritic EPSP amplitude is not
sufficient to remove the decrease in somatic EPSP amplitude
with distance. (d) Synaptic conductance required to produce
a location-independent 0.2 mV EPSP at the soma for a variety
of different membrane parameters. (Each line is the result of
a quadratic fit.) Note that an increase in synaptic conduc- c d
tance is required in all conditions. m, specific membrane

resistance Rm; a, axial resistance Ra.
Peak amplitude (mV)
The propagated events recorded simultaneously

at the soma showed the expected location-dependent
rise time, with the time constant increasing over
threefold across the range of input locations (Fig. 3c).
The location dependence of the decay time for prop-
agated EPSPs was reduced because the local decrease
in decay time counters the filtering effects of the den-
drites to some degree9 (Fig. 3d). Overall, however,
there were no significant differences between the kinetics of the tance was moved farther into the dendrites, the local dendritic
true synaptic events and the voltage transients induced by cur- EPSP amplitude and decay rate increased (Fig. 4c), whereas the
rent injection. These data indicate that the somatic EPSP rise EPSP amplitude at the soma decreased as a function of distance
time is a more accurate indicator of synapse location than EPSP (Fig. 4c). A progressive increase in the conductance of the
duration and further demonstrate the role of dendritic mem- synapses with distance from the soma could remove the loca-
brane properties in shaping EPSPs. tion dependence for a realistic range of passive parameters
We next examined the amplitude dependence of EPSP kinetics (Fig. 4d). (An approximately 2-fold increase was required for
for both distal and proximal local dendritic events (Fig. 3eh). There a synapse located 300 m from the soma.) In each case, the
was a direct relationship between the EPSP rise time and ampli- synaptic input produced a 0.2 mV EPSP at the soma (example
tude, with the rise time increasing nearly threefold for distal and in Fig. 4b), regardless of its dendritic location. These simula-
about twofold for proximal synapses across the range of amplitudes tions, together with the above experimental observations, sug-
(slopes, proximal, 109 11 ms/mV; distal, 91 8 ms/mV; Fig. 3e gest that an increase in synaptic conductance with distance from
and g). There was no such observable amplitude-dependent the soma, and not postsynaptic morphology or voltage-depen-
increase in the rise time of the EPSPi (Fig. 3e and g), suggesting a dent mechanisms, account for the elimination of somatic EPSP
synaptic mechanism for the increase in EPSP rise time with location-dependent variability.
amplitude. There was very little amplitude dependence to the If an increase in AMPA receptor-mediated synaptic con-
decay time constants other than a slight increase that was occa- ductance is indeed the mechanism for eliminating the effect
sionally observed at proximal sites (both EPSP and EPSPi; Fig. of distance on the synaptic response, then it should be possi-
3f and h). As discussed further below, the amplitude-dependent ble to directly measure this using a dendritic voltage clamp.
increase in EPSP rise time suggests a mechanism underlying the We measured EPSCs at the site of synaptic input across the
increase of distal EPSP amplitude. entire axis of the apical dendrite receiving Schaffer collateral
The experimental evidence thus far suggests the mechanism input using four different techniques for stimulating localized
for eliminating location dependence is centered at the synapse. single synaptic activity: minimal electrical stimulation, local
We therefore sought to determine if a simple increase in synap- application of high osmolar external solution (Fig. 5a), low
tic conductance could theoretically account for this effect, using electrical stimulation in locally applied Sr2+-containing exter-
a realistic passive multi-compartmental computer model of a nal solution, and local application of adenosine A1 receptor
CA1 pyramidal neuron (Fig. 4a; Methods). To determine the antagonist. As with the EPSPs, mean EPSC amplitude in-
extent of location dependence in the model, we placed synaptic creased from 8 pA for synaptic locations near 100 m to
conductances in the apical dendrites at varying distances from 24 pA for synapses around 300 m from the soma (Fig. 5b
the soma. The conductance amplitude was adjusted such that and c). All four stimulating techniques produced EPSCs of
a proximal synapse (located 50 m from the soma) produced similar mean amplitude, with EPSCs evoked in Sr2+ having a
a somatic EPSP of 0.2 mV (Fig. 4b). As this synaptic conduc- slightly lower mean amplitude across the dendritic range. This

articles
Fig. 5. Dendritic EPSC amplitude increases with synapse dis-

tance. (a) Dendritically recorded EPSC activity induced by a
local application of high-osmolarity external solution onto a
proximal (100 m; top trace) and a distal (280 m; bot-
tom trace) dendrite of two different neurons. (b) Average of
all EPSCs recorded from the two neurons shown in (a).
(c) Mean EPSC amplitude as a function of input distance from
the soma for EPSCs induced by high osmolarity () or
EPSCs evoked in normal () or Sr2+-bcontaining external
solution (). (d) Normalized EPSC amplitude histograms for
the recordings in (a). The distribution of distal EPSCs (open
histogram) contains many more large-amplitude events, and
a larger mean, than the histogram for proximal EPSCs (filled b c
histogram). Inset, cumulative histogram (distal, light line;
proximal, dark line). (e) Synaptic charge, normalized to the
value of the proximal synapses, as a function of distance from
the soma. Data were grouped into proximal (75 25 m),
middle (175 25 m) and distal (275 25 m) regions.
Integrated EPSCs were induced by high osmolarity or evoked
in Sr2+-containing external solution. *p < 0.05; **p < 0.02.
may indicate that there is a relationship between

EPSC amplitude and probability of release, which
should be somewhat lower when Sr2+ is used as the
main divalent 24. Amplitude distributions demon- d e
strated that mean EPSC amplitude of distal synaps-
es is increased because of the presence of
larger-amplitude events that are not found in prox-
imal synapses (Fig. 5d). Currentvoltage relation-
ships showed that there was no difference in the
reversal potentials, and therefore driving forces,
between the proximal and distal synapses (proximal,
7.3 1.3 mV; distal, 6.0 2.3 mV; n = 3 in both, not
corrected for junction potential).
Because imperfect dendritic voltage-clamp condi-
tions would allow location-dependent filtering to affect
the amplitude and shape of recorded EPSCs (Fig. 2b;
Discussion), we also compared synaptic charge. As synaptic The similarity of the location dependence of the actual EPSC
charge is less affected by filtering, comparing the integral of the decay and that of the EPSCi (Fig. 3b) suggest that the faster
EPSC should provide a relatively filtering-independent assess- decay of the distal events results from dendritic morphology
ment of synaptic conductance. The synaptic charge measured and an imperfect voltage clamp and not from any kinetic dif-
from the EPSC integrals increased more than twofold from prox- ferences in synaptic conductance. A similar conclusion was
imal to distal Schaffer collateral synapses (Fig. 5e), further indi- reached above for the apparent location-dependent differences
cating an increase in synaptic conductance with distance. in EPSP decay rates (Fig. 3b).
To directly compare the extent of voltage clamp in the dis- Although the rise times of similarly sized EPSCs were inde-
tal and proximal dendritic regions, we placed two recording pendent of location (77.8 4.8 s, n = 9; 79.4 3.9, n = 10, for
pipettes within 20 m of each other on the same dendrite. 10 pA proximal and distal EPSCs, respectively), a marked
EPSC-shaped current injections were delivered via the most amplitude dependence of the EPSC rise was noted (Fig. 6c). In
distal electrode and the voltage deflection recorded either with nearly all cells, the rise time constant increased with EPSC
both pipettes in current-clamp mode or with the more proxi- amplitude in an approximately linear fashion (proximal slope,
mal pipette in voltage-clamp mode (Fig. 6a). The level of volt- 9.7 1.1 s/pA, n = 9; distal slope, 6.3 1.4 s/pA, n = 8). The
age clamp, quantified as the ratio of the EPSPi integrals, was difference in slope results from an apparent saturation of the
similar for both proximal and distal recordings (voltage rise times at larger amplitudes that are only present in the dis-
clamp/no clamp, 9% proximal, 11% distal; n = 2 for both tal recordings (notice exponential fit in Fig. 6c). There was no
groups). Although the clamp was similar, and quite good, for amplitude dependence to the EPSCi even for a twofold larger
both proximal and distal recordings, it was nevertheless imper- range of current amplitudes (Fig. 6c, left traces). These data
fect. As a result, the amplitude and decay rate of the EPSCi (the suggest that the slowing of EPSC rise time with increasing
EPSCs resulting from current injections) were slightly location amplitude is the result of an actual kinetic change in the synap-
dependent, with distal events having a somewhat larger ampli- tic conductance and not of voltage-clamp imperfections.
tude and faster decay rate (Fig. 6a). EPSCi rise time was inde- The amplitude distributions of both proximal and distal
pendent of location. EPSP/Cs exhibited a large amount of variability and a promi-
A similar decrease in decay time constant with distance nent peak at smaller amplitudes (Fig. 7ac). This peak could
from the soma was observed in the synaptic EPSCs (Fig. 6b). be observed when the distributions were fit by a single Gauss-

articles
Fig. 6. Location and amplitude dependence of EPSC

kinetics. (a) Paired recordings from either proximal (left a
traces, Prox) or distal (right traces, Distal) dendritic
regions from different neurons. Large EPSPi (no clamp)
evoked by current injection with both electrodes in cur-
rent-clamp mode. Smaller EPSPi (clamp) evoked by cur-
rent injection through one electrode in current-clamp
mode and the other in voltage-clamp mode. EPSCi
recorded by the voltage-clamp electrode (lower traces, I)
in response to the current injection. The EPSPi labeled
clamp is the escape voltage of the dendritic region and is b Synaptic EPSC Decay tau (ms)
similar between the regions. (b) The left traces are the
exponential fits of the EPSCis in (a) with their amplitudes
normalized. The middle traces are exponential fits of
synaptically evoked EPSCs (from recordings in Fig. 5).
Note that the distal EPSCs always decay faster than the
proximal even with the EPSCi that were evoked by a uni-
form current injection. Plot is decay time constant versus
input location for synaptically evoked EPSCs () and cur-
rent-injection EPSCi (), showing that the decay times of
both events decrease similarly with distance. (c) Left

traces, EPSCi (same recordings as in a) for a range of cur-
rent injection amplitudes, showing that rise time remains c Synaptic EPSC EPSC rise tau
constant. Middle traces, selective averages of synaptically
evoked EPSCs (from another dendrite 290 m). EPSCs
with amplitudes 37 pA, 812 pA, 1317 pA and 1822 pA
were averaged, respectively. Note that the rise time of
these events slows as the amplitude increases. Plot is rise
time constant verses EPSC amplitude for all synaptically
evoked EPSCs (from same dendrite), showing that the
rise times slow with EPSC amplitude. Number of points
in the traces in (c) have been reduced to one third for
clarity. Traces are digitally smoothed. All EPSCs shown
were evoked by high osmolarity.
ian function that was constrained to the smaller amplitudes DISCUSSION

(09 pA range) or as the first of several peaks when fit by a We have characterized the amplitude and kinetic properties of
multi-Gaussian function. The mean of this first peak was sim- AMPA receptor-mediated synaptic input across a wide range of
ilar across the entire range of dendritic synapse locations the apical dendrites in CA1 pyramidal neurons. The main find-
(56 pA and 0.150.20 mV; < 50% increase from 50 to 300 m; ing is that the mean amplitude of the synaptic conductance
Fig. 7d). These data suggest that there may be an increase in the increases with distance from the soma and that this increase
number of quanta released (quantal content) from synapses counterbalances dendritic filtering, dramatically reducing the
with distance from the soma. However, the coefficient of varia- location dependence of synaptic amplitude. The use of realistic
tion (CV), which should directly reflect quantal content, does dendritic current injections and a passive multicompartmental
not increase with distance (Fig. 7e). Instead there is a greater computer model demonstrate that the passive and, most likely,
than twofold increase in the variance divided by the mean the active properties of CA1 dendrites have only a minor role in
(2/x), which suggests that mean quantal size, not quantal con- elevating the local EPSP amplitude of distal unitary inputs. Volt-
tent, increases with distance (Fig. 7e). Together these data imply age-gated ion channels, however, significantly shape the ampli-
that there is a population of larger-amplitude, slower-rising, tude of larger EPSPs 6,7,10,11. Instead of dendritic membrane
unitary events increasingly present at more distal Schaffer col- properties elevating the amplitude of distal dendritic EPSPs, it
lateral synapses. seems that a progressive increase in synaptic conductance, with
Following propagation to the soma, the amplitude of the distance from the soma, is primarily responsible. Both in the hip-
smallest synaptic EPSPs and EPSP is showed a pronounced pocampus and in other CNS regions, location dependence of
dependence on input location (open circles; Fig. 7d). For input synaptic efficacy is minimal, with several authors suggesting an
located more distal than 125150 m, the amplitude of the small- increase in synaptic conductance as a mechanism1518,24,25. Our
est EPSPs at the soma was lower than the noise for single sweeps data, recorded directly from the site of input, fit well with these
(< 0.05 mV) and would normally have been missed (counted as other studies.
failure) if not for the dendritic recording (Fig. 7c). Beyond about What is the mechanism of the increased synaptic conductance
200 m, the events seem to lose all ability to significantly change at distal synapses? There are several possibilities: increases in
the voltage at the soma, even in selected averages. This demon- AMPA receptor number or density, agonist affinity, single-chan-
strates the difficulties of using only somatic recordings to char- nel conductance, quantal glutamate concentration, or number
acterize synaptic events and indicates that the use of such of quanta released per terminal. Increases in cleft glutamate or
recordings to determine failure rates or the presence of silent AMPA receptor density, affinity or single-channel conductance
synapses should be limited to only the most proximal locations seem unlikely to be responsible. In all of these cases, we should
(within 150 m of the soma). see an increase in the amplitude of the smallest events with dis-

articles
Fig. 7. Amplitude of smallest dendritic events shows little dependence

on synapse location. Amplitude distributions for dendritically recorded
a Proximal b Distal
EPSCs from proximal (a, 50 m) or distal (b, 300 m) locations.

Distributions were fit by a multiple-Gaussian function. Mean values of
Count
Count
the first peaks are shown. EPSCs in (a) were evoked by high osmolarity. peak 1 = 4.7 pA peak 1 = 5.8 pA
EPSCs in (b) were electrically evoked in Sr2+. (c) Distribution of den-
dritically recorded EPSPs for a distal input location. Inset, distribution of
EPSPs simultaneously recorded at the soma of the same cell as in (c).
Note large number of apparent failures (arrow). Distribution was fit
(constrained to 00.6 mV) by a single Gaussian function. (d) Amplitude
of the smallest dendritically recorded events (EPSPs, or EPSCs, ) EPSC amplitude (pA) EPSC amplitude (pA)
plotted as a function of synapse location. Amplitude of the smallest
somatically recorded events (EPSPs and voltage transients) plotted as a
function of synapse location (open circles). (e) Coefficient of variation c EPSP d Small event amplitude
() and the 2/x () as a function of location.
Count
tance from the soma, and rise times should not increase with
amplitude22,25,27. Therefore, it seems most likely that the increase
in distal synaptic conductance results from an increased number

of AMPA receptors and/or quanta released.
EPSC amplitude (mV) Synaptic location
Terminals capable of releasing multiple quanta from multiple
release sites are observed in several cell types, including Schaf-
fer-collateral synapses of hippocampal CA1 pyramidal neu-
rons 22,23,25,2830. Theoretically, the number of multisynaptic e
boutons or the number of release sites per bouton could increase
with distance from the soma. Also, Schaffer collateral synapses
have a range of areas, and the number of AMPA receptors at
these synapses is directly related to the area31,32. Thus it is pos-
sible that quantal release at a larger synapse would open more
AMPA receptors at lower cleft glutamate concentrations. Such
a situation could produce larger-amplitude currents with a slow- Synaptic location
er rise time33. Either of these two mechanisms should correlate
with an elevated dendritic glutamate sensitivity 14 (A. Frick, ly regulated by common neuromodulators (such as acetyl-
W. Zieglgansberger & H.U. Dodt, Soc. Neurosci. Abstr. 24, 325, choline, norepinephrine and serotonin)8,9,35. The release of these
1998). Given the complexity of these synapses, however, a thor- neuromodulators at a localized region of the dendrite could
ough determination of the exact mechanisms involved in the lower threshold, change temporal summation and produce a
increase in distal synaptic conductance as well as the processes very nonlinear form of integration36,37, perhaps even resulting in
involved in producing and maintaining this gradient will require the local initiation of dendritic spikes7,9,38. Such a regional, non-
more direct experimentation. linear type of integration would be transposed on top of the
Several lines of data now indicate that, contrary to theoreti- otherwise global, spatially normalized input discussed above.
cal expectations, there is minimal location dependence to mul- Therefore by removing the location dependence of synaptic
tiple components of dendritic integration, including unitary EPSP input, the neuron has extended the range of synaptic integra-
amplitude and temporal summation. Also, the threshold for tion possible in the same dendritic arborization from a very
action potential generation increases with distance from the soma global non-varying type of processing to one that is perhaps
at a rate that is very similar to that of the increase in local synap- regional and highly nonlinear.
tic amplitude6,7. Because of this, all synaptic input should be
equally distant from threshold no matter where it is in the cell.
In this situation, all synapses will have the same ability to initi- METHODS
Hippocampal slices (400 m) were prepared from 612 week-old
ate action potentials and to induce long-term synaptic plasticity
SpragueDawley rats using standard procedures as described8. Individ-
regardless of their location in the dendritic arborization6. This ual neurons were visualized with a Zeiss Axioskop fit with differential
elimination of location from the synaptic weight is critical for interference contrast (DIC) optics using infrared illumination. All neu-
Hebbian-type processes, where the weight of any given synapse at rons had resting membrane potentials between 63 and 75 mV. Whole-
the final integration site primarily depends on its history of use cell patch-clamp recordings were made using two Dagan BVC-700 or a
and not on other factors such as synapse location34. Furthermore, combination of one Dagan and an Axopatch 200B amplifier in active
normalization of synaptic input can reduce the variability of bridge mode. All synaptic currents were recorded with an Axopatch
action potential firing in single neurons and improve the syn- 200B. Data were acquired at 50 kHz and filtered at 3 or 5 kHz. The nor-
chrony of neuronal population activity9. mal external recording solution contained 124 mM NaCl, 2.5 mM KCl,
1.2 mM NaH2PO4, 25 mM NaHCO3, 2.0 mM CaCl2, 1.5 mM MgCl2 and
CA1 pyramidal neurons have equalized the impact of synap-
10 mM dextrose, bubbled with 95% O2 and 5% CO2 at 35 C (pH 7.4).
tic input while preserving the location independence of synap- In most cases, the external recording solution used during high osmolar-
tic integration and plasticity through a beautifully intricate ity stimulation contained 0.5 mM CaCl2, 7.0 mM MgCl2 and 0.51 M
interplay between dendritic membrane excitability and single- tetrodotoxin (TTX). Whole-cell recording pipettes (somatic, 2 to 4 M;
synapse properties. However, this is only half the story, as all dendritic, 3.5 to 7 M), were pulled from borosilicate glass. The internal
the ion channels responsible for this balanced interplay are high- pipette solution consisted of 120 mM KMeSO4, 20 mM KCl, 10 mM

articles
HEPES, 0.05 mM EGTA, 4.0 mM Mg2ATP, 0.3 mM Tris2GTP, 14 mM D.B. Jaffe). All synaptic input was modeled as an AMPA receptor-like
phosphocreatine and 4 mM NaCl (pH 7.25 with KOH). A cesium-based conductance change using an alpha function (time to peak, 1 ms) and a
internal solution (120 mM Cs-gluconate replaced KMeSO4, and 12 mM reversal of 0 mV. The passive parameters Rm and Ra were varied as indi-
QX314 was added) was used for the generation of currentvoltage plots. cated, with Cm always set to 1 uF/cm2.
Series resistance for somatic recordings was 6 to 20 M, whereas that
for dendritic recordings was 15 to 40 M for voltage recording or 12 to ACKNOWLEDGEMENTS
25 M for current recordings. Dendritic pipettes were coated with syl- We thank M. Carruth for technical assistance, M. Vollrath for comments on the
gard. Voltages have not been corrected for the theoretical liquid junction manuscript and D. Johnston for discussions throughout the study. This work was
potential (7 mV). Final concentrations of bicuculine methiodide (10 supported by National Institute of Health grants NS35865 and NS39458 and by
M), APV (50 M), TTX (0.51 M, RBI) and A1 antagonist 8- the Alfred P. Sloan Foundation.
cyclopentyl-1,3-dipropylxanthine (DPCPX, 2 M) were made daily from
stock solutions dissolved in water. Error bars represent s.e.m., and the RECEIVED 17 MAY; ACCEPTED 25 JULY 2000
number of cells (n) is given. ANOVA with Fishers post-hoc test was used
for statistical comparison.
The stimulation techniques used in this study were intended to 1. Bannister, N. J. & Larkman, A. U. Dendritic morphology of CA1 pyramidal
stimulate release from single terminals that were localized to an neurones from the rat hippocampus: II. Spine distributions. J. Comp. Neurol.
approximate area of the dendrite (25 m diameter). Bicuculine 360,161171 (1995).
2. Rall, W. Theory of physiological properties of dendrites. Ann. NY Acad. Sci.
methiodide (10 M) and APV (50 M) were present in the external 96, 10711079 (1962).
solutions during all recordings. When minimal stimulation techniques 3. Jack, J. J. & Redman, S. J. The propagation of transient potentials in some
were used21,22, a stimulating electrode was placed within 510 m of linear cable structures. J. Physiol. (Lond.) 215, 283320 (1971).
the dendrite under study (whole-cell pipettes coated with sylgard or 4. Mainen, Z. F., Carnevale, N. T., Zador, A. M., Claiborne, B. J. & Brown, T. H.
Electrotonic architecture of hippocampal CA1 pyramidal neurons based on
tungsten bipolar electrodes; A-M Systems, Carlsborg, Washington). three-demensional reconstruction. J. Neurophysiol. 76, 19041923 (1996).
The close proximity of the stimulating electrodes made it very diffi- 5. Jaffe, D. B. & Carnevale, N. T. Passive normalization of synaptic
cult, however, to find a stimulating spot where a wide range of stim- integration influenced by dendritic architecture. J. Neurophysiol. 82,
ulus currents could be used without a resulting increase in amplitude. 32683285 (1999).
We used a template (composed of the rising and decaying exponential 6. Magee, J. C., Hoffman, D., Colbert, C. & Johnston, D. Electrical and calcium
signaling in dendrites of hippocampal pyramidal neurons. Annu. Rev. Physiol.
equation shown in Methods) to isolate EPSCs from the recorded 60, 327346 (1998).
region of the dendrite, and all selected EPSCs had similar synaptic 7. Hoffman, D. A., Magee, J. C., Colbert, C. M. & Johnston, D. K+ channel
delay, rise and decay time. High osmolar external solution was pres- regulation of signal propagation in dendrites of hippocampal pyramidal
sure applied using a computer controlled pneumatic pump (Medical neurons. Nature 387, 869875 (1997).
8. Magee, J. C. Dendritic hyperpolarization-activated currents modify the
Systems, Greenvale, New York) to a spot on the dendrite. (The pres- integrative properties of hippocampal CA1 pyramidal neurons. J. Neurosci.
sure wave had a diameter of < 20 m.)39,40. High osmolar external 18, 76137624 (1998).
solution consisted of normal external solution with the addition of 9. Magee, J. C. Dendritic Ih normalizes temporal summation in hippocampal
300 mM sucrose and 0.51 M TTX and HEPES replacing NaHCO3. CA1 neurons. Nat. Neurosci. 2, 508514 (1999).
Based on previous studies, we estimate that the final concentration 10. Gillessen, T. & Alzheimer, C. Amplification of EPSPs by low Ni2+- and
amiloride-sensitive Ca2+ channels in apical dendrites of rat CA1 pyramidal
of external solution reached approximately 450 mOsM after mixing40. neurons. J. Neurophysiol. 77, 16391643 (1997).
A combination of these two techniques was used to stimulate in Sr2+- 11. Lipowsky, R., Gillessen, T. & Alzheimer, C. Dendritic Na+ channels amplify
containing solutions24. First, the external bath solution was changed EPSPs in hippocampal CA1 pyramidal cells. J. Neurophysiol. 76, 21812191
to one with 0.5 mM CaCl2 and 7 mM MgCl2, and a solution contain- (1996).
ing 0 mM added CaCl2, 10 mM SrCl2 and HEPES replacing NaHCO3 12. Rall, W. Distinguishing theoretical synaptic potentials computed for different
soma-dendritic distributions of synaptic input. J. Neurophysiol. 30,
was pressure applied as above. The application pipette was used to 11381168 (1967).
electrically stimulate at low stimulus intensities. In a few cases, Sr2+- 13. Rinzel, J. & Rall, W. Transient response in a dendritic neuron model for
containing external solution was bath applied (NaHCO3 as buffer). current injected at one branch. Biophys. J. 14, 759790 (1974).
For analysis of EPSP/Cs, records were first smoothed (5 points aver- 14. Pettit, D. L. & Augustine, G. J. Distribution of functional glutamate and
GABA receptors on hippocampal pyramidal cells and interneurons
aged with a type of moving average algorithm, Igor, Wavemetrics) and J. Neurophysiol. 84, 2838 (2000).
a threshold-crossing protocol (2 pA or 0.1 mV; 2 noise) was used to 15. Inasek, R. & Redman, S. J. The amplitude, time course and charge of unitary
automatically select events. The non-smoothed versions of these events excitatory post-synaptic potentials evoked in spinal motoneurone dendrites.
were then individually inspected and fit by the function (t) = a(1 J. Physiol. (Lond.) 234, 665688 (1973).
exp(t/r))5exp(t/d), where a is a constant and r and d are the rise 16. Andersen, P., Silfvenius, H., Sundberg, S. H. & Sveen, O. A comparison of
distal and proximal dendrite synapses on CA1 pyramids in guinea pig
and decay time constants. All events with a rise time slower than 0.8 and hippocampal slices in vitro. J. Physiol. (Lond.) 307, 273299 (1980).
0.4 ms for EPSP/Cs, respectively, were discarded as being too slow to have 17. Jack, J. J., Redman, S. J. & Wong, K. The components of synaptic potentials
occurred near the recording site (Fig. 3a). The spontaneous release rate evoked in cat spinal motoneurones by impulses in single group Ia afferents.
is very low at physiological temperatures in slices, so the number of these J. Physiol. (Lond.) 321, 6596 (1981).
18. Stricker, C., Field, A. C. & Redman S. J. Statistical analysis of amplitude
contaminating events was very small (< 5%). To minimize the impact of fluctuations in EPSCs evoked in rat CA1 pyramidal neurones in vitro.
even these very infrequent spontaneous events, we triggered the averag- J. Physiol. (Lond.) 490, 419441 (1996).
ing of somatic EPSPs only after the dendritic EPSP crossed the voltage 19. Larkman, A., Stratford, K. & Jack, J. Quantal analysis of excitatory synaptic
threshold during the dual recordings. Events were binned at 1 pA or 0.1 action and depression in hippocampal slices. Nature 350, 344347 (1991).
mV for amplitude distributions. Although there seemed to be some evi- 20. Kullmann, D. M. & Nicoll, R. A. Long-term potentiation is associated with
increases in quantal content and quantal amplitude. Nature 357, 240243
dence of evenly spaced peaks in some distributions, insufficient num- (1992).
bers of events were collected (usually < 100) to allow an appropriate 21. Liao, D., Jones, A. & Malinow, R. Direct measurement of quantal changes
determination24. The mean event rate (determined from interval distri- underlying long-term potentiation in CA1 hippocampus. Neuron 9,
butions) was less than 10 Hz for high osmolar solution application and 10891094 (1992).
22. Bolshakov, V. Y., Golan, H., Kandel, E. R. & Siegelbaum, S. A. Recruitment
less than 5 Hz for events evoked in Sr2+. Using these rates, the probabil-
of new sites of synaptic transmission during the cAMP-dependent late
ity of two events occurring within 1 ms of each other is 5 106 for high phase of LTP at CA3CA1 synapses in the hippocampus. Neuron 19,
osmolar solution, and 2.5 106 for stimulation in Sr2+ according to 635646 (1997).
Poisson statistics41. In most figures, the data were arbitrarily fit by either 23. Larkman, A. U., Jack, J. J. & Stratford, K. J. Quantal analysis of excitatory
an exponential or linear function. If they could not be fit by one of these synapses in rat hippocampal CA1 in vitro during low-frequency depression.
J. Physiol. (Lond.) 505, 457472 (1997).
functions, then they were arbitrarily fit by a polynomial function. 24. Bekkers, J. M. & Clements, J. D. Quantal amplitude and quantal variance of
The neurophysiological modeling program NEURON was used to strontium-induced asynchronous EPSCs in rat dentate granule neurons.
simulate a reconstructed CA1 cell (provided by B.J. Claiborne and J. Physiol. (Lond.) 516, 227248 (1999).

articles
25. Korn, H., Bausela, F., Charpier, S. & Faber, D. S. Synaptic noise and 33. Nusser, Z. Cull-Candy, S. & Farrant, M. Differences in synaptic GABA(A)
multiquantal release at dendritic synapses. J. Neurophysiol. 70, 12491253 receptor number underlie variation in GABA mini amplitude. Neuron 19,
(1993). 697709 (1997).
26. Lim, R., Alvarez, F. J. & Walmsley, B. Quantal size is correlated with receptor 34. Cook, E. P. & Johnston, D. Active dendrites reduce location-dependent
cluster area at glycinergic synapses in the rat brainstem. J. Physiol. (Lond.) variability of synaptic input trains. J. Neurophysiol. 78, 21162128
516, 505520 (1999). (1997).
27. Liu, G., Choi, S. & Tsien, R. W. Variability of neurotransmitter concentration 35. Johnston, D., Hoffman, D. A., Colbert, C. M. & Magee, J. C. Regulation of
and nonsaturation of postsynaptic AMPA receptors at synapses in back-propagating action potentials in hippocampal neurons. Curr. Opin.
hippocampal cultures and slices. Neuron 22, 395402 (1999). Neurobiol. 9, 288292 (1999).
28. Prange, O. & Murphy, T. H. Analysis of multiquantal transmitter release from 36. Mel, B. Synaptic integration in a excitable dendritic tree. J. Neurophysiol. 70,
single cultured cortical neuron terminals. J. Neurophysiol. 81, 18101818 10861101 (1993).
(1999). 37. Archie, K. A. & Mel, B. W. A model for intradendritic computation of
29. Bykhovskaia, M., Hackett, J. T. & Worden, M. K. Asynchrony of quantal binocular disparity. Nat. Neurosci. 3, 5463 (2000).
events in evoked multiquantal responses indicates presynaptic quantal 38. Golding, N. L. & Spruston, N. Dendritic sodium spikes are variable triggers
interaction. J. Neurophysiol. 81, 22342242 (1999). of axonal action potentials in hippocampal CA1 pyramidal neurons. Neuron
30. Sorra, K. E. & Harris, K. M. Occurrence and three dimensional structure of 21, 11891200 (1998).
multiple synapses between individual radiatum axons and their target 39. Bekkers, J. M. & Stevens, C. F. Cable properties of cultured hippocampal
pyramidal cells in hippocampal area CA1. J. Neurosci. 13, 37363747 (1993). neurons determined from sucrose-evoked miniature EPSCs. J. Neurophysiol.
31. Nusser, Z. et al. Cell type and pathway dependence of synaptic AMPA 75, 12501255 (1995).
receptor number and variability in the hippocampus. Neuron 21, 545559 40. Magee, J. C. & Carruth, M. Dendritic voltage-gated ion channels regulate the
(1998). firing mode of CA1 pyramidal neurons. J. Neurophysiol. 82, 18951901
32. Takumi, Y., Ramrez-Len, V., Laake, P., Rinvik, E. & Ottersen, O. P. Different (1999).
modes of expression of AMPA and NMDA receptors in hippocampal synapse. 41. Johnston, D. & Wu, S. M. in Foundations of Cellular Neurophysiology 294301
Nat. Neurosci. 2, 618624 (1999). (MIT Press, Cambridge, Massachusetts, 1995).

articles
A network of electrically coupled

interneurons drives synchronized
inhibition in neocortex
Michael Beierlein, Jay R. Gibson and Barry W. Connors
Department of Neuroscience, Division of Biology & Medicine, Box 1953, Brown University, Providence, Rhode Island 02912, USA
The first two authors contributed equally to this work
Correspondence should be addressed to B.W.C. (bwc@brown.edu)
The neocortex has at least two different networks of electrically coupled inhibitory interneurons: fast-
spiking (FS) and low-threshold-spiking (LTS) cells. Agonists of metabotropic glutamate or acetylcholine
receptors induced synchronized spiking and membrane fluctuations, with irregular or rhythmic
patterns, in networks of LTS cells. LTS activity was closely correlated with inhibitory postsynaptic
potentials in neighboring FS interneurons and excitatory neurons. Synchronized LTS activity required
electrical synapses, but not fast chemical synapses. Tetanic stimulation of local circuitry induced effects
similar to those of metabotropic agonists. We conclude that an electrically coupled network of LTS
interneurons can mediate synchronized inhibition when activated by modulatory neurotransmitters.
In the awake neocortex, the timing of spikes is often strikingly work was independent of chemical synapses, and arose from elec-
irregular1, and, when induced by sensory stimuli2,3 and move- trical coupling between the LTS interneurons.
ments4, spiking is often correlated across neurons. Synaptic input
is the likely source of both the variability5,6 and the synchrony of RESULTS
spiking7. Some synchronized cortical activity is also rhythmic, Whole-cell recordings, usually from pairs of neurons, were made
and a wide variety of forebrain oscillations accompany states of in layer 4. All neurons were classified as one of three types based
sensory perception, motor performance, arousal and sleep8. The on their intrinsic membrane properties16,18 and synaptic con-
neuronal mechanisms that generate irregular, rhythmic and syn- nections18. Regular-spiking (RS) neurons generated adapting
chronous synaptic inputs in the neocortex are poorly understood. spike patterns when stimulated with current pulses, and formed
Studies in vitro suggest that synchronous oscillations can be gen- excitatory synaptic connections with all cell types. FS cells pro-
erated by networks of inhibitory, GABAergic interneurons911. duced brief action potentials, with little or no adaptation, and
Hippocampal rhythms ranging from less than 1 Hz (ref. 12) to formed inhibitory synaptic connections with all cell types. LTS
70 Hz (ref. 13) can be induced by cholinergic and glutamatergic cells produced adapting spiking patterns, generated rebound
agonists or potassium channel blockers. Computational models spikes after a hyperpolarizing current pulse, and made frequent
imply that chemical inhibitory synapses14 or electrical synapses15 inhibitory synaptic connections with RS and FS cells, but rarely
between interneurons may be critical for maintaining synchrony inhibited other LTS cells. Electrical coupling between neuron
or stabilizing rhythmic activity in hippocampus. pairs was assessed as described18. For inter-somatic distances
The neocortex has diverse types of inhibitory neurons16,17, but under 100 m, coupling was observed in 4 of 7 FSFS cell pairs,
their individual functions are obscure. The synaptic interconnec- and 9 of 16 LTSLTS cell pairs. Only 2 of 15 FSLTS pairs were
tions of two major types of inhibitory interneurons (Fig. 1) sug- coupled, and RS cells were never coupled (n = 43 pairs). For dis-
gest they may be able to serve as independent synchronizing tances over 100 m, no electrical coupling was observed in 5
networks. First, fast-spiking cells, but not low threshold-spiking FSFS cell pairs, 6 FSLTS pairs or 46 pairs with at least one RS
cells, commonly make inhibitory synapses with cells of the same cell; only 1 of 9 such LTSLTS cell pairs was coupled.
type18. Second, both FS cells1820 and LTS cells18 are frequently
interconnected to neurons of their own type by electrical synaps- Modulator-induced synchrony in LTS network
es; these synapses are strong enough to synchronize spiking of In control medium, the frequency of spontaneous synaptic
coupled cell pairs. Third, FS neurons selectively mediate feedfor- potentials was low, and there was no spontaneous spiking. When
ward inhibition in thalamic recipient layers of the neocortex18, the metabotropic glutamate receptor (mGluR) agonist ACPD
whereas LTS cells seem specialized for mediating a facilitating (50100 M) was added to the bath, LTS cells quickly depolar-
form of local inhibition18,21,22. Fourth, FS and LTS cells are differ- ized by about 10 4 mV (n = 38), and more than 90% of them
entially sensitive to cholinergic and noradrenergic agonists2325. began spiking with peak firing rates of 2050 Hz. In the presence
We found that two neuromodulators selectively induced the of tetrodotoxin (TTX; 2 M) to block spikes, ACPD depolarized
network of LTS interneurons to generate both irregular and LTS cells by an average of 18.4 mV (n = 9). In control medium,
rhythmic patterns of synchronized inhibitory postsynaptic poten- ACPD induced spikes in only 12% of FS cells, and depolarized
tials (IPSPs) in neighboring neurons. Synchrony in the LTS net- them by 8 3 mV (n = 41); ACPD hyperpolarized RS cells by

articles
of their cross-correlograms were close to 0 ms. Cross-power spec-

tra of the correlograms showed peaks within the range of 314
Hz (mean 6.1 Hz, n = 6 pairs).
Blocking ionotropic glutamate or -aminobutyric acid
(GABA) receptors did not affect ACPD-induced synchrony in
LTS cells. Application of ACPD in the presence of APV (to block
NMDA receptors), DNQX (to block AMPA receptors) and bicu-
culline or picrotoxin (to block GABA A receptors) produced
strong, synchronous membrane fluctuations (Fig. 3b; n = 4
LTSLTS pairs). The strength and frequencies of cross-correlo-
grams obtained in antagonists were indistinguishable from those
of control LTSLTS pairs, suggesting that chemical EPSPs and
IPSPs were not involved in synchronized LTS oscillations. Fur-
thermore, synchronous subthreshold oscillations in LTS cells did
not depend on action potentials. In the presence of TTX to block
sodium channels, plus APV, DNQX and picrotoxin, membrane
Fig. 1. Two networks of inhibitory interneurons in neocortex. FS cells
oscillations (range 47 Hz) occurred when ACPD was applied to
and LTS cells make inhibitory synapses (round terminals), whereas RS
16 of 26 single LTS cells, and when muscarine was applied to 5
cells make excitatory synapses (rectangular terminals). Electrical

synapses (zigzags) interconnect interneurons of the same type. Axons of 6 single LTS cells. LTS cells never oscillated in the absence of
from relay cells in the thalamus synapse mainly onto FS cells and RS cells. ACPD, with or without TTX. Epochs of TTX-resistant oscillations
were synchronized in pairs of LTS cells that were electrically cou-
pled, during application of ACPD (n = 4 pairs; Fig. 3c and d) or
muscarine (n = 2 pairs). ACPD in the presence of TTX also
induced rhythms in another six pairs of LTS cells that were not
4 3 mV (n = 47), and did not induce spikes in any of the 69 electrically coupled, but in each case oscillations occurred in only
cells tested. Simultaneous recordings from pairs of LTS cells in one cell at a time. Irregular patterns of subthreshold fluctuations
ACPD revealed that spiking patterns were complex and closely were not evoked by ACPD in the presence of TTX.
correlated. Typically, 10-second periods of aperiodic spiking activ- Together, these results strongly imply that the close synchrony
ity alternated with 25-second periods of rhythmic bursts. (Inter- of activity in neighboring LTS neurons is mediated by electrical
burst frequencies were usually 36 Hz; Fig. 2a.) Alternation synapses. They do not rule out the possibility that chemical
between irregular spiking and rhythmic bursts continued for two synapses or action potentials augment or modify oscillations in
to five cycles, lasting several minutes in the continued presence the intact circuit. Pre-incubating slices with octanol (12 mM),
of ACPD. Thereafter, LTS cells remained strongly depolarized, which blocks electrical coupling between interneurons18, strong-
rhythmic bursting ceased, and spiking continued irregularly and ly attenuated the ACPD-induced depolarization of LTS cells
more slowly. Cross-correlograms showed that spiking was strong-
ly synchronous in three of four LTSLTS pairs during both non- a
rhythmic (Fig. 2b) and rhythmic (Fig. 2c) periods. Correlograms
had a central peak near zero ms and an average magnitude 14
baseline (bin width, 1 ms) or 3.5 baseline (bin width, 15 ms).
Synchronous LTS pairs were less than 50 m apart and coupled
electrically; the uncorrelated LTS cells were 160 m apart and
uncoupled.
When ACPD-treated LTS cells were hyperpolarized to pre-
vent spiking, sizeable fluctuations of the subthreshold membrane
potential were revealed (28 of 38 neurons; Fig. 3a, top). These
fluctuations included sharp, 24 mV spikelets, but no obvious
IPSPs. Subthreshold activity was often periodic, with peak fre-
quencies averaging 7.7 4.9 Hz (n = 28), and tightly correlated in
cell pairs, with peak normalized correlations of 0.7 (n = 6 b c
LTSLTS pairs, all less than 200 m apart; Fig. 3a, bottom). For
closely spaced pairs of LTS neurons (< 100 m), the central peaks
Fig. 2. Synchronous activity in LTS cells. (a) Bath application of ACPD

(100 M) caused depolarization and action potentials in a pair of electri-
cally coupled LTS cells. Close synchrony of action potentials during peri-
ods of irregular (left) and rhythmic (right) spiking is apparent. The periods
marked 1 and 2 on top traces are expanded in time in the lower traces.
Action potential amplitudes are truncated. (b, c) Cross-correlations of
spikes for the two cells in (a) show sharp synchrony during both non-
rhythmic (b) and rhythmic (c) periods of spiking (upper graphs, 15 ms
time bins; lower graphs, 1 ms time bins). Dashed lines represent the cor-
relations expected if the two cells were firing independently. All cross-
correlations in the figures use the upper trace as the reference (cell A).

articles
Fig. 3. Synchronized subthreshold oscillations of LTS cells did not

require chemical synaptic transmission or action potentials. (a) Cross- a c
correlation of subthreshold activity during ACPD application from the
same cells shown in Fig. 2a. Membrane potentials were held just below
threshold with injected current. (b) ACPD-induced subthreshold oscilla-
tions in two electrically coupled LTS cells (10 m apart) held below firing
threshold, with fast chemical synapses blocked by APV (50 M), DNQX
(20 M) and BMI (50 M). (c) ACPD-evoked synchronized oscillations in
two moderately coupled LTS cells (45 m apart), with fast chemical
synapses and action potentials blocked by APV (50 M), DNQX (20 M),
picrotoxin (100 M) and TTX (2 M). (d) Cross-correlation of rhythmic
activity shown in (c).
(mean of 3 mV; n = 3 cells) and prevented oscillations, suggesting

that octanol impaired the mGluR-mediated depolarization of
LTS cell membranes. d
b
Synchronous inhibition driven by LTS network

LTS cells are GABAergic inhibitory interneurons16,18. Because
ACPD induces synchronized, periodic spiking in LTS cells, it
should also induce synchronous, periodic IPSPs in neurons that
are postsynaptic to LTS cells. Indeed, simultaneous recordings
showed that each burst of spikes in an LTS cell was coincident
with a cluster of IPSPs in a neighboring FS cell (Fig. 4a). Similar
clusters of IPSPs were observed in about 70% of all FS and RS
cells recorded in ACPD. The subthreshold activity underlying
spikes in LTS cells also strongly predicted the occurrence of IPSPs described above. DHPG (100200 M), a specific group I
in the FS cells (Fig. 4b). Analysis of the two membrane poten- mGluR agonist, caused ACPD-like effects in all tested cells (3
tials revealed strongly anticorrelated activity with zero phase lag pairs of FS/RS cells; 1 of 2 LTS cells in TTX). The selective
(Fig. 4c); depolarizations of the LTS cell were mirrored by hyper- mGluR1a antagonist LY367385 (30200 M) prevented ACPD-
polarizations of the FS cell (n = 7 of 8 LTSFS pairs). Such strong induced spiking and oscillations in 8 LTS cells. Only 1 of 11 FS
synchrony was independent of whether the LTS and FS cells in and RS cells showed ACPD-induced rhythmic inhibition, and
each pair were monosynaptically interconnected. LTSRS cell none of 7 pairs of FS/RS cells showed synchronous inhibition in
pairs exposed to ACPD produced very similar anticorrelated the presence of 30 M LY367385. In contrast, the selective
activity (3 of 3 pairs).
Paired recordings from closely spaced (< 100 m) RS and/or
FS cells showed that the ACPD-induced inhibition was profoundly
synchronized (Fig. 4d). Both cell types received very similar pat- a
terns of synchronized IPSPs; cross-correlation amplitudes were c
quite high and had zero phase lag, for all RSRS (Fig. 4e; n = 12),
FSFS (n = 4) and RSFS (n = 4) pairs. Similar to the ACPD-
induced activity of LTS cells, large, rhythmic IPSPs occurred in
RS/FS pairs at about 5 Hz, in 24-second epochs, often separated
by periods of correlated smaller, aperiodic IPSPs. Synchronous
IPSPs were not blocked by the combination of APV plus DNQX
(n = 24 neuron pairs). We conclude that local FS and RS neurons
b
are synchronously inhibited by the ACPD-induced activity of the
synchronized LTS network.
Metabotropic receptors drive synchronous LTS activity e

We tested receptor-selective drugs on the rhythmic inhibition
(in FS and RS cells) or subthreshold oscillations (in LTS cells)
d
Fig. 4. Synchronous, rhythmic IPSPs induced by ACPD in FS and RS

cells. (a) ACPD evoked bursts of IPSPs in an FS cell that were coincident
with action potentials in an LTS cell. Cells were 20 m apart. (b) In the
same neuron pair, subthreshold activity in the LTS cell was highly anti-
correlated with IPSPs in the FS cell, as shown by cross-correlogram (c).
(d) Rhythmic IPSPs in a pair of RS cells were highly synchronous in
ACPD, as shown by cross-correlogram (e). The bath contained APV,
DNQX and 6 mM K+ for these recordings.

articles
Fig. 5. Muscarine-evoked synchrony. (a) Muscarine (50 M)

triggered action potentials in LTS cell that were synchronous a b
with IPSPs in RS cell. (b) LTS spike-triggered average of RS
cell potential from experiment in (a). Spikes in LTS cell pre-
dicted IPSPs in RS cell, even though cells were not monosy-
naptically connected. (c) Muscarine-induced subthreshold
fluctuations of LTS cell were anti-correlated with activity of
RS cell (upper traces); cross-correlogram (lower) quantifies
the data (same cell pair as shown in a). (d) RS cell pair with
tightly synchronized IPSPs in muscarine (50 M). Close-up c
shows that IPSPs were correlated even during high-fre-
quency, aperiodic activity. The bath contained APV, DNQX
and 6 mM K+ for these recordings, and all cell pairs had
somata less than 50 m apart.
d
mGluR5 antagonist MPEP (5200 M) did not impair

the usual effects of ACPD (n = 2 LTS cells, 9 RS/FS cells
and 5 pairs of various neuron combinations). We con-
clude that the synchronizing effects of ACPD are medi-

ated via mGluR1a on LTS cells.
LTS cell membranes are particularly sensitive to
muscarinic acetylcholine receptor (mAChR) ago-
nists23. Muscarine (50 M) indeed depolarized LTS
cells and caused some spiking, but was less potent
than ACPD and usually did not induce clear syn-
chrony in normal medium. However, when the bath
K+ concentration was raised to 6 mM, 50 M mus-
carine induced LTS spiking, subthreshold fluctuations and syn- cells showed that IPSPs were closely synchronous during both
chrony similar to that evoked by ACPD (n = 2 LTSLTS pairs). rhythmic and irregular phases of inhibition (Fig. 5d; n = 2
The antagonist pirenzepine (510 M) blocked muscarinic pairs). These results suggest that mAChRs and mGluRs acti-
effects on LTS cells (n = 2) and muscarine-induced rhythmic vate the LTS network similarly.
IPSPs (n = 4 FS cells), suggesting they were mediated
by M1/M4 receptors. Bursts of muscarine-induced LTS a
spikes were coincident with clustered IPSPs in an RS
cell (Fig. 5a). An average of the RS recording, triggered
by the spikes in the LTS cell, shows that the onset of the
IPSP in the RS cell led the spikes in the recorded LTS
cell by several milliseconds (Fig. 5b). The phase-lead
and slow rise time of the RS-IPSP suggest that it was
generated by convergent input from spikes in several
LTS cells, some of them firing a few milliseconds before
the recorded LTS cell. Similar to ACPD-induced activ-
ity, muscarine-induced depolarizations of LTS cells were
strongly anticorrelated with IPSPs in RS cells (Fig. 5c;
n = 2 pairs). Simultaneous recordings from RS and FS
Fig. 6. Synchrony declines with distance. (a) Sliding cross-correl-

ograms for three different ACPD-treated neuron pairs (all
RSRS) with increasing intersomatic distance. Horizontal axis,
time at the center of the 1 s sliding window; vertical axis, time
from the center of the window. Color codes the correlation b c
amplitude, and the corresponding intracellular voltage records
are shown below each graph. Note that the most closely spaced
pair is strongly correlated during both rhythmic and aperiodic
activity. (b) Absolute value of peak cross-correlations (measured
over a 2-s rhythmic epoch) as a function of intersomatic distance,
for all combinations of cell types (non-LTS include FS and RS
cells). The line shows the linear regression (r = 0.59, p < 0.0001,
n = 73 pairs). (c) Average temporal overlap (as percentage) of the
rhythmic, 35 Hz epochs recorded from cell pairs, as a function
of intersomatic distance. Only pairs in which the limits of the
epochs could be clearly determined were used. The line shows
the best-fit linear regression (r = 0.89, p < 0.0001, n = 35).

articles
Fig. 7. Endogenous neuromodulators also trig-

a c ger synchronous oscillations. (a) Recordings
from an LTS and an RS cell (20 m apart); an
extracellular tetanic stimulus (100 Hz, 20 pulses)
evoked prolonged LTS depolarization and oscilla-
tion, coincident with hyperpolarization and IPSPs
in the RS cell. The LTS cell was held below
threshold with injected current. (b) Sliding cross-
correlogram calculated from the traces in (a)
shows rhythmic anti-correlated activity.
(c) Synchronous IPSPs in an FSRS cell pair were
evoked by a tetanic stimulus in control solution,
slightly attenuated by atropine (30 M), blocked
by MCPG (1 mM), and recovered during wash in
b control solution. The bath contained APV,
DNQX and 6 mM K+ for these recordings.
2.8 0.6 Hz (n = 6 cells; Fig. 7a). The same

tetani evoked prolonged hyperpolarizations

in RS cells (2.7 3.5 mV, n = 31 cells;
Fig. 7a and c) and relatively small depolar-
izations in FS cells (6.6 4.0 mV, n = 22 cells)
that rarely triggered spikes. Most important-
ly, tetani induced rhythmic IPSPs in both RS
and FS cells (4.1 1.7 Hz, n = 18 cells;
Fig. 7a), but not in LTS cells. Tetanus-induced
activity was synchronous in all tested cell
pairs, with a strength similar to that produced
Spatial extent of LTS-mediated synchrony by bath-applied agonists (Fig. 7b; mean peak cross-correlation,
To estimate the spatial extent of synchronized LTS networks, we 0.58 0.22; n = 18 pairs of various combinations of cell types).
recorded from pairs of ACPD-activated neurons with increasing Ionotropic glutamate receptors were not necessary for tetanus-
horizontal separation within layer 4, and calculated their cross- induced synchrony, as it occurred in 15 of 20 cell pairs tested with
correlations. Sliding cross-correlograms (unfolding in time) were APV and DNQX in the bath. Atropine (30 M), an antagonist of
calculated for pairs of neurons at different distances (Fig. 6a). mAChRs, when applied alone reduced the number of tetanus-
Closely spaced (< 100 m) pairs of RS and/or FS cells tended to induced synchronized events by an average of 29% (from 7.3 to
be strongly correlated during the epochs of rhythmic IPSPs 5.2 cycles of synchronized LTS depolarizations and RSFS IPSPs
(Fig. 6a, left, time less than 4 s; n = 34 pairs) and during inter- per tetanus, n = 9 cell pairs; Fig. 7c). MCPG (1 mM), an antagonist
vening irregular IPSP activity as well (Fig. 6a, left, time greater of groups I and II mGluRs, when applied alone reliably blocked
than 4 s). In pairs of cells 100400 m apart, rhythmic activity the sustained depolarizations of LTS cells and reduced the num-
was less well correlated, and the intervening nonrhythmic activ- ber of tetanus-evoked synchronized rhythms by an average of 95%
ity was uncorrelated (n = 23 pairs; Fig. 6a, middle). In pairs over (from 8.6 to 0.4 cycles/tetanus, n = 11 cell pairs; Fig. 7c). We con-
400 m apart, peak correlations tended to be quite low even dur- clude that brief activation of local cortical circuitry can induce syn-
ing rhythmic events (Fig. 6a, right; n = 16). Peak rhythmic cor- chronous inhibition, largely via activation of mGluRs on LTS cells.
relations fell with intercellular distance (Fig. 6b; p < 0.0001). We
also measured the average temporal overlap of epochs of rhyth- DISCUSSION
mic IPSPs in cell pairs. Up to 9 such epochs appeared during each The synchronous inhibitory activity we describe here is excep-
application of ACPD (mean 3.2 epochs, n = 35 pairs). The onset tional in several ways. First, it was generated by a specific type of
and decay of these epochs were coincident in most closely spaced inhibitory interneuronthe LTS cellsthat previously had not
pairs, but independent in pairs over 400 m apart (Fig. 6a, com- been implicated in synchronous cortical activity. Second, syn-
pare left versus right voltage traces; Fig. 6c). These data indicate chrony could occur independent of chemical synapses, and appar-
that LTS networks can mediate synchronous inhibition over hor- ently required electrical synapses between LTS cells. Third,
izontal separations of roughly 400 m, with rhythmic synchrony synchronous rhythms in many LTS cells were TTX insensitive.
extending over longer distances than irregular synchrony. Fourth, the LTS network drove synchronized inhibition, with
both rhythmic and irregular patterns, in the larger circuit of exci-
LTS network is activated by endogenous neuromodulators tatory neurons and in a second type of inhibitory neuron, the FS
We tested whether endogenous factors could activate LTS-mediat- cells. The results suggest that a network of LTS neurons, when
ed synchrony. A brief tetanic stimulus (100 Hz for 200 ms) was activated by metabotropic receptors under certain behavioral
applied to layer 4 with extracellular electrodes placed near the states, can coordinate the firing patterns of most cortical neu-
recorded neurons. The bath K+ concentration was 6 mM for these rons over a distance of several hundred microns.
experiments. Following a tetanus, LTS cells depolarized for about LTS cells are very likely to be the pacemaker for the synchro-
1520 s, peaking at about 8 3 mV (n = 11 cells). This depolariza- nized inhibition we observed because they were strongly and
tion evoked repetitive, often rhythmic spiking. LTS cells held below selectively excited by the metabotropic agonists tested, they fired
spike threshold generated membrane potential oscillations at in close synchrony with each other, they form abundant GABAer-

articles
gic synapses onto RS and FS cells18, and their spiking and sub- inhibitory connections (Fig. 1). The emergent behavior of an
threshold depolarizations were highly correlated with popula- activated FS network, with its chemical and electrical synapses,
tion IPSPs in RS and FS cells. No other cell type we encountered is likely to be quite distinct from that of neurons in an LTS net-
had any of these properties. work, which interact mainly by electrical synapses. Furthermore,
The ACPD effects on LTS cells were likely mediated by the FS and LTS interneuron networks are reciprocally interconnect-
mGluR1a receptor subtype. Many LTS cells are immunoreactive ed by inhibitory synapses (Fig. 1), suggesting the potential for
for somatostatin16,18,22, consistent with structural evidence that unique activity patterns arising from their interaction.
neurons containing GABA and somatostatin are particularly LTS network synchrony is triggered by agonists of mGluRs and
enriched in mGluR1a (ref. 26). Thus, inositol triphosphate (IP3) mAChRs. Endogenous ligands for these receptors are probably at
production and an increase in intracellular Ca2+ concentration27 their highest levels during behavioral states such as alert wakeful-
may be involved in the excitation of LTS cells. Somatostatin-con- ness and rapid eye movement stages of sleep, which are associat-
taining interneurons are also particularly sensitive to mAChR ed with strong cortical activity and increased input from the
agonists23. The ionic mechanisms of mGluR- and mAChR-medi- cholinergic basal forebrain8. The function of a synchronized LTS
ated effects in LTS cells are unknown, but they probably involve network may be to coordinate the timing of action potentials
effects on one or more TTX-insensitive, voltage- or Ca2+-gated across a spatially localized subpopulation of neurons. Phasic inhi-
intrinsic membrane currents. Interestingly, activation of group bition is an effective way to regulate spike timing37, particularly
I mGluRs in immature neocortex leads to very slow oscillations at around 47 Hz (ref. 38). There is growing evidence that the
of intracellular Ca2+ concentration28 that are sometimes syn- precise timing of spikes1 and the correlated spiking of groups of
chronized across clusters of dye-coupled neurons29, perhaps by neurons24,39,40 may encode information within cortical circuits.
the passage of IP3 through gap junctions30. Such intercellular bio- Neurons in the intact, awake neocortex often fire with highly
chemical signaling is too slow to synchronize the 47 Hz rhythms irregular patterns1,5. This irregularity does not seem to arise with-
of LTS cells, but it could account for the slower modulation of in the intrinsic spike-generating process of cortical neurons41,
rhythms observed in LTS cells treated with TTX (Fig. 3c). but is probably a function of complex and partially synchronous
What synchronizes the LTS network? Theoretical10,31,32 and patterns of synaptic input6,7. The sources of such irregular, syn-
experimental9,13,33 studies suggest that GABAergic synapses alone chronized synaptic activity are unknown. Spontaneous inhibito-
can effectively synchronize a mutually interconnected network of ry input can greatly increase the variability of spiking in cerebellar
interneurons. However inhibitory synapses cannot be critical for neurons42. We found that electrically coupled LTS networks could
synchronizing LTS cells in neocortex, because LTS-to-LTS inhibito- induce both irregular and rhythmic patterns of synchronized
ry synapses are very sparse18, and blocking GABAA receptors did inhibitory output, suggesting that LTS cells serve as an effective,
not impair the synchrony of LTS cell pairs. Synchronous mem- rapidly regulated mechanism for coordinating the activity of local
brane oscillations persisted even when ionotropic glutamate, assemblies of cortical neurons.
GABA and voltage-sensitive sodium channels were blocked, so
chemical synapses cannot be essential for LTS synchrony. On the METHODS
other hand, electrical synapses abundantly interconnect LTS neu- Thalamocortical slices 43 (250450 m thick) were obtained from
rons18, they can closely synchronize spiking and subthreshold SpragueDawley rats aged P1421, incubated at 32C for 1 h, and kept
membrane fluctuations, and they are not blocked by transmitter at room temperature until they were transferred to a 32C submersion-
receptor antagonists. The simplest explanation is that electrical type recording chamber. The bathing solution contained 126 mM NaCl,
synapses mediate the synchronization of the LTS network. 3 mM KCl, 1.25 mM NaH 2 PO 4 , 2 mM MgSO 4 , 26 mM NaHCO 3 ,
10 mM dextrose and 2 mM CaCl2, saturated with 95% O2/5% CO2. Dur-
Electrical synapses are a common mediator of synchronized
ing some experiments, noted in the text, the bath K+ concentration was
neural activity in neural networks34, and they are particularly raised to 6 mM; the most robust rhythmic activity was generated by
effective in synchronizing subthreshold oscillations35. Electrical 50 M ACPD in 6 mM K+. Micropipettes were filled with 135 mM K-
coupling alone, or in conjunction with inhibitory synaptic con- gluconate, 4 mM KCl, 2 mM NaCl, 10 mM HEPES, 0.2 mM EGTA, 4
nections, can mediate synchronous firing of hippocampal9,36 or mM ATP-Mg, 0.3 mM GTP-Tris and 7 mM phosphocreatine-Tris
neocortical11 inhibitory neurons during activity induced by potas- (pH 7.25, 295 mOsm). Recordings were made in current-clamp mode
sium channel blockers. A model of such activity implies that elec- (with amplifiers from Axon Instruments, Foster City, California). IR-
trical synapses serve to generate and stabilize very slow, DIC visualization was done using a Zeiss Axioskop and a CCD camera
synchronized bursting15. Electrical coupling between LTS neu- (Hamamatsu, Hamamatsu City, Japan). Extracellular tetanic stimuli
were applied through an ultrasmall concentric bipolar electrode (Fred-
rons is most effective at transmitting lower-frequency signals18,
erick Haer, Bowdoinham, Maine) with the tip placed 30100 m from
such as those dominating the rhythmic fluctuations described the recorded cells. Tetani consisted of 20 pulses (each 200 s long) at
here, but it can also mediate close synchrony of single action 100 Hz, using intensities of 100300 A.
potentials (Fig. 2a and b). Irregular synchronous activity was The following drugs were used: the N-methyl-D-aspartate (NMDA)
probably mediated by relatively small groups of neighboring, receptor antagonist D,L-2-amino-5-phosphopentanoic acid (AP5; Sigma,
strongly coupled LTS cells, whereas slower rhythmic activity arose St. Louis, Missouri), the AMPA/kainate receptor antagonist 6,7-dini-
from much larger populations of LTS cells that spanned a greater troquinoxaline-2,3-dione (DNQX; Sigma), the GABAA receptor antag-
cortical volume. We suggest that electrical synapses are necessary onists bicuculline methiodide (BMI; Sigma) and picrotoxin (Sigma),
for coordinating the timing of spikes in LTS cells over time scales the mGluR (groups I and II) agonist (1S,3R)-1-aminocyclopentane-
from a few milliseconds to hundreds of milliseconds. 1,3-dicarboxylic acid ((1S,3R)-ACPD; Tocris, Bristol, UK), the selec-
tive mGluR group I agonist 3,5-dihydroxyphenylglycine (DHPG;
The importance of electrical synapses for synchronizing the
Tocris), the mGluR (groups I and II) antagonist (S)--methyl-4-car-
LTS network may not apply to other networks of interneurons boxyphenylglycine ((S)-MCGP; Tocris), the selective mGluR1a antag-
in the cerebral cortex. Almost all previous studies of synchro- onist (+)--amino-4-carboxy-2-methylbenzeneacetic acid (LY367385;
nization by cortical interneurons have focused on FS-like cells1114, Tocris), the selective mGluR5 antagonist 2-methyl-6-
which are indeed electrically coupled in neocortex1820. However, (phenylethynyl)pyridine (MPEP; Tocris), the mAChR agonist (+)-mus-
unlike LTS cells, FS cells are also interconnected by abundant carine chloride (Sigma), the mAChR antagonist atropine sulfate

articles
(Sigma), and the selective M1/M4 antagonist pirenzepine dihydrochlo- 12. Zhang, Y. et al. Slow oscillations (1 Hz) mediated by GABAergic
ride (Tocris). Tetrodotoxin was obtained from Sigma. interneuronal networks in rat hippocampus. J. Neurosci. 18, 92569268
Data were analyzed with auto- and cross-correlograms and power (1998).
13. Whittington, M. A., Traub, R. D. & Jefferys, J. G. Synchronized oscillations in
spectra using Labview software (National Instruments, Austin, Texas). interneuron networks driven by metabotropic glutamate receptor activation.
Subthreshold membrane potentials were band-pass filtered (1.570 Hz). Nature 373, 612615 (1995).
The correlograms were first averaged, then normalized by the product 14. Traub, R. D., Jefferys, J. G. R. & Whittington, M. A. Fast Oscillations in
of the standard deviation of the signal in each recording: Cortical Circuits (MIT Press, Cambridge, Massachusetts, 1999).
15. Skinner, F. K., Zhang, L., Velazquez, J. L. & Carlen, P. L. Bursting in inhibitory
interneuronal networks: A role for gap-junctional coupling. J. Neurophysiol.
average [A(t) B(t t)] 81, 12741283 (1999).
normalized correlation (t) = 16. Kawaguchi, Y. & Kubota, Y. GABAergic cell subtypes and their synaptic
s.d.(A(t)) s.d.(B(t)) connections in rat frontal cortex. Cereb. Cortex. 7, 476486 (1997).
17. Somogyi, P., Tamas, G., Lujan, R. & Buhl, E. H. Salient features of synaptic
organisation in the cerebral cortex. Brain Res. Rev. 26, 113135 (1998).
where A(t) and B(t) represent the two voltage records, t is the time vari- 18. Gibson, J. R., Beierlein, M. & Connors, B. W. Two networks of electrically
able for the actual traces, t is the time variable for the correlation, and coupled inhibitory neurons in neocortex. Nature, 402, 7579 (1999).
s.d. is the standard deviation9. Thus, if the analysis were an autocorrelo- 19. Galarreta, M. & Hestrin, S. A network of fast-spiking cells in the neocortex
gram, the normalization factor would be the variance, and the value at connected by electrical synapses. Nature 402, 7275 (1999).
time zero would always be 1. Sliding cross-correlograms 20. Tams, G., Buhl, E. H., Lrincz, A. & Somogyi, P. Proximally targeted
GABAergic synapses and gap junctions synchronize cortical interneurons.
(Figs. 6 and 7) were calculated by averaging signals over a window of 1 s, Nat. Neurosci. 3, 366371 (2000).
and stepping the window in 100-ms increments. 21. Thomson, A. M. Activity-dependent properties of synaptic transmission at
Discrete cross-correlations were determined on action potentials by two classes of connections made by rat neocortical pyramidal axons in vitro.
calculating the time intervals between each spike in the reference trace J. Physiol (Lond.) 502, 131147 (1997).
22. Reyes, A. et al. Target-cell-specific facilitation and depression in neocortical
and all spikes in the nonreference trace, binning these intervals (at circuits. Nat. Neurosci. 1, 279285 (1998).
either 1 or 15 ms bin widths), generating a histogram of spike counts, 23. Kawaguchi, Y. Selective cholinergic modulation of cortical GABAergic cell
and then normalizing by the number of spikes in the reference trace. subtypes. J. Neurophysiol. 78, 17431747 (1997).
The units in this cross-correlogram equal the average number of spikes 24. Kawaguchi, Y. & Shindou, T. Noradrenergic excitation and inhibition of
that occur in the nonreference trace given one spike in the reference GABAergic cell types in rat frontal cortex. J. Neurosci. 18, 69636976
(1998).
trace. Thus, if the two spike trains were identical, time zero would have 25. Xiang, Z., Huguenard, J. R. & Prince, D. A. Cholinergic switching within
a value of one. No filtering or sample reduction was done before action neocortical inhibitory networks. Science 281, 985988 (1998).
potential correlation. The expected level of correlation between two 26. Baude, A. et al. The metabotropic glutamate receptor (mGluR1 alpha) is
independently firing cells with this method is determined by the fol- concentrated at perisynaptic membrane of neuronal subpopulations as
detected by immunogold reaction. Neuron 11, 771787 (1993).
lowing equation: 27. Schoepp, D. D. & Conn, P. J. Metabotropic glutamate receptors in brain
function and pathology. Trends Pharmacol. Sci. 14, 1320 (1993).
(total spike count of nonreference cell) 28. Flint, A. C., Dammerman, R. S. & Kriegstein, A. R. Endogenous activation of
(bin width) metabotropic glutamate receptors in neocortical development causes
(analysis window duration) neuronal calcium oscillations. Proc. Natl. Acad. Sci. USA 96, 1214412149
(1999).
29. Yuste, R., Peinado, A. & Katz, L. C. Neuronal domains in developing
ACKNOWLEDGEMENTS neocortex. Science 257, 665669 (1992).
30. Kandler, K. & Katz, L. C. Coordination of neuronal activity in developing
We thank Yael Amitai and Michael Paradiso for comments and Saundra Patrick visual cortex by gap junction-mediated biochemical communication.
for technical assistance. This work was supported by a fellowship to M.B. from J. Neurosci. 18, 14191427 (1998).
the Burroughs-Wellcome Trust, a fellowship to J.R.G. from NIH (NS10478) and 31. Wang, X. J. & Rinzel, J. Spindle rhythmicity in the reticularis thalami nucleus:
synchronization among mutually inhibitory neurons. Neuroscience 53,
grants to B.W.C. from NIH (NS25983 and DA12500). 899904 (1993).
32. Van Vreeswijk, C., Abbott, L. F. & Ermentrout, G. B. When inhibition not
excitation synchronizes neural firing. J. Comput. Neurosci. 1, 313321
RECEIVED 3 MAY; ACCEPTED 11 JULY 2000 (1994).
33. Buhl, E. H., Tamas, G. & Fisahn, A. Cholinergic activation and tonic
excitation induce persistent gamma oscillations in mouse somatosensory
1. Holt, G. R., Softky, W. R., Koch, C. & Douglas, R. J. Comparison of discharge cortex in vitro. J. Physiol. (Lond.) 513, 117126 (1998).
variability in vitro and in vivo in cat visual cortex neurons. J. Neurophysiol. 34. Bennett, M. V. Gap junctions as electrical synapses. J. Neurocytol. 26, 349366
75, 18061814 (1996). (1997).
2. Gray, C. M. The temporal correlation hypothesis of visual feature integration: 35. Manor, Y., Rinzel, J., Segev, I. & Yarom, Y. Low-amplitude oscillations in the
still alive and well. Neuron 24, 3147 (1999). inferior olive: a model based on electrical coupling of neurons with
3. deCharms, R. C. & Merzenich, M. M. Primary cortical representation of heterogeneous channel densities. J. Neurophysiol. 77, 27362752 (1997).
sounds by the coordination of action-potential timing. Nature 381, 610613 36. Michelson, H. B. & Wong, R. K. Synchronization of inhibitory neurones
(1996). in the guinea-pig hippocampus in vitro. J. Physiol. (Lond.) 477, 3545
4. Donoghue, J. P., Sanes, J. N., Hatsopoulos, N. G. & Gaal, G. Neural discharge (1994).
and local field potential oscillations in primate motor cortex during 37. Lytton, W. W. & Sejnowski, T. J. Simulations of cortical pyramidal neurons
voluntary movements. J. Neurophysiol. 79, 159173 (1998). synchronized by inhibitory interneurons. J. Neurophysiol. 66, 10591079
5. Azouz, R. & Gray, C. M. Cellular mechanisms contributing to response (1991).
variability of cortical neurons in vivo. J. Neurosci. 19, 22092223 (1999). 38. Cobb, S. R., Buhl, E. H., Halasy, K., Paulsen, O. & Somogyi, P.
6. Stevens, C. F. & Zador, A. M. Input synchrony and the irregular firing of Synchronization of neuronal activity in hippocampus by individual
cortical neurons. Nat. Neurosci. 1, 210217 (1998). GABAergic interneurons. Nature 378, 7578 (1995).
7. Lampl, I., Reichova, I. & Ferster, D. Synchronous membrane potential 39. Engel, A. K., Roelfsema, P. R., Fries, P., Brecht, M. & Singer, W. Role of the
fluctuations in neurons of the cat visual cortex. Neuron 22, 361374 (1999). temporal domain for response selection and perceptual binding. Cereb.
8. Steriade, M. Synchronized activities of coupled oscillators in the cerebral Cortex 7, 571582 (1997).
cortex and thalamus at different levels of vigilance. Cereb. Cortex 7, 583604 40. Hatsopoulos, N. G., Ojakangas, C. L., Paninski, L. & Donoghue, J. P.
(1997). Information about movement direction obtained from synchronous activity
9. Fisahn, A., Pike, F. G., Buhl, E. H. & Paulsen, O. Cholinergic induction of of motor cortical neurons. Proc. Natl. Acad. Sci. USA 95, 1570615711
network oscillations at 40 Hz in the hippocampus in vitro. Nature 394, (1998).
186189 (1998). 41. Mainen, Z. F. & Sejnowski, T. J. Reliability of spike timing in neocortical
10. Buzsaki, G. & Chrobak, J. J. Temporal structure in spatially organized neurons. Science 268, 15031506 (1995).
neuronal ensembles: a role for interneuronal networks. Curr. Opin. 42. Husser, M. & Clark, B. A. Tonic synaptic inhibition modulates neuronal
Neurobiol. 5, 504510 (1995). output pattern and spatiotemporal synaptic integration. Neuron 19, 665678
11. Benardo, L. S. Recruitment of GABAergic inhibition and synchronization of (1997).
inhibitory interneurons in rat neocortex. J. Neurophysiol. 77, 31343144 43. Agmon, A. & Connors, B. W. Thalamocortical responses of mouse
(1997). somatosensory (barrel) cortex in vitro. Neuroscience 41, 365379 (1991).

articles
The role of -CaMKII

autophosphorylation in neocortical
experience-dependent plasticity
S. Glazewski1, K. P. Giese2, A. Silva3 and K. Fox1
1 Cardiff School of Biosciences, Cardiff University, Museum Avenue, Cardiff CF10 3US, UK
2 Department of Anatomy and Developmental Biology, University College London, London WC1E 6BT, UK
3 Departments of Neurobiology, Psychology and Psychiatry, University of California, Los Angeles, 695 Young Drive South, Room 2357, Los Angeles, California 90095, USA
Correspondence should be addressed to K.F. (foxkd@cardiff.ac.uk)
Calcium/calmodulin kinase type II (CaMKII) is a major postsynaptic density protein. CaMKII is postu-
lated to act as a molecular switch, which, when triggered by a transient rise in calcium influx,
becomes active for prolonged periods because of its ability to autophosphorylate. We studied expe-
rience-dependent plasticity in the barrel cortex of mice carrying a point mutation of the -CaMKII
gene (T286A), which abolishes this enzymes ability to autophosphorylate. Plasticity was prevented
in adult and adolescent mice homozygous for the mutation, but was normal in heterozygotes and
wild-type littermates. These results provide evidence that the molecular switch hypothesis is valid
for neocortical experience-dependent plasticity.
The ability of neocortical neurons to remodel their receptive field plasticity. The barrel cortex is the primary cortical representa-
characteristics in response to changes in sensory experience is tion of the whiskers, which are a highly developed tactile senso-
known as experience-dependent plasticity. Visual, auditory, ry system in rodents. Because barrel cortex plasticity can be
motor and somatosensory cortex all show experience-dependent studied in rodents, questions can be asked that involve genetic
plasticity that may depend on similar forms of synaptic plastici- manipulation. Plasticity measured in the barrel cortex involves
ty. Synaptic potentiation depends critically on postsynaptic cal- cortical but not subcortical pathways1517, and depends on cor-
cium influx triggering several second messenger systems1,2. tical activity (H. Wallace and K.F., J. Physiol. (Lond.) 515P, 51,
Although the signaling pathways are likely to be complex, involv- 1999) and on -CAMKII based on studies on -CaMKII null
ing several interacting sets of molecules3, there is substantial evi- mutants18. However, why -CaMKII is necessary for plasticity is
dence that calcium/calmodulin kinase type II (CaMKII), a major still unknown. For example, the -CaMKII null mutation may
component of the postsynaptic density4,5, is involved in the induc- affect the expression of genes involved in plasticity or be required
tion of synaptic potentiation. For example, CaMKII inhibitors for a general kinase action in the neuron. Alternatively, the rele-
or a null mutation of the -CaMKII gene prevent long-term vant attribute of -CaMKII may be its ability to act as molecular
potentiation (LTP) in area CA1 of the hippocampus2,6 and impair switch for synaptic plasticity9,19. We therefore tested whether
LTP in an age-dependent manner in visual cortex7. autophosphorylation at T286 of -CaMKII is necessary for expe-
Autophosphorylation of the autonomy site of -CaMKII (thre- rience-dependent plasticity by measuring plasticity in mice
onine-286, T286) is crucial for synaptic plasticity8. The activity homozygous for a targeted point mutation (T286A), which ren-
of the T286-phosphorylated kinase is independent of the pres- ders CaMKII unable to phosphorylate at T286 (ref. 8). The results
ence of calcium, and is thought to overcome phosphatase activi- show that experience-dependent potentiation is prevented by the
ty by phosphorylating other subunits within the holoenzyme5,811. point mutation and therefore that autophosphorylation of
Phosphorylation at T286 is triggered during the induction of LTP CaMKII is required for experience-dependent plasticity.
and lasts for at least 30 minutes1214. Because CaMKII remains
active beyond the presence of the original trigger, it can be thought RESULTS
of as a molecular switch. Evidence suggests that autophosphory- Barrel cortex plasticity in wild-type littermates
lation of -CaMKII at T286 is essential for the induction of To induce plasticity, we deprived all vibrissae except D1 for 18
NMDA-receptor-dependent LTP in the hippocampus8, but it is days before recording from barrel cortex (Methods). In wild-
not known whether it is involved in neocortical LTP or experi- type mice, this caused potentiation of the spared D1 vibrissa
ence-dependent plasticity in vivo. Testing whether autophospho- response. In barrels surrounding the D1 barrel, responses of
rylation is required for experience-dependent plasticity is therefore layer 2/3 cells to stimulation of the spared (D1) vibrissa
important, for assessing both whether the molecular switch is increased in adults (Table 1) and in adolescents (Table 2) com-
involved in experience-dependent plasticity, and whether synap- pared to D1 responses in undeprived littermates (F1,23 = 19.8,
tic plasticity itself underlies experience-dependent plasticity. p < 0.0002). Similarly, responses to D1 stimulation increased
The barrel cortex is an excellent system in which to study this relative to responses to stimulation of other surrounding
type of question, as it shows quantifiable experience-dependent whiskers in the same animals (F1,30 = 21.6, p < 0.0001).

articles
Table 1. Results for adult animals (> 6 months). mutants, the spared vibrissa response
-CaMKII T286A was indistinguishable from the deprived
D1 Surround n D1 Surround n
surround vibrissa response recorded in
the same animals (t11 = 1.4, p = 0.18),
Wild type 0.37 0.11 0.41 0.09 7 0.34 0.10 0.31 0.05 7
undeprived whereas in wild-type mice, spared and
deprived surround responses were clear-
Wild type 1.10 0.18 0.26 0.06 7 0.86 0.11 0.28 0.05 13 ly different (t22 = 4.9, p < 0.0001).
deprived Two further measures confirmed
Heterozygotes 0.55 0.15 0.63 0.09 7 0.48 0.04 0.41 0.07 5 that plasticity was absent in the T286A
undeprived mutants. First, the area of cortex
responding at a high level to the D1
Heterozygotes 0.75 0.14 0.54 0.04 6 1.04 0.17 0.43 0.07 6 vibrissa is normally restricted to the
deprived
D1 barrel, but expands into the sur-
Homozygotes 0.31 0.26 2 0.44 0.06 0.28 0.02 3 rounding barrels following a period of
undeprived single-vibrissa experience (Fig. 1).
Responses in wild-type mice are
Homozygotes 0.46 0.12 0.46 0.17 5 0.47 0.09 0.32 0.05 7
deprived increased to a distance of at least
300 m (Fig. 1c). Because the recep-
The average response to stimulation of the D1 vibrissa (D1) and the surround receptive field whisker (surround) tive field properties of neurons within
are shown for adult -CaMKII null mutants (CaMKII) and T286A point mutants (T286A). Numbers are group
mean responses (spikes per stimulus). In the case of undeprived animals, all these whiskers are intact from birth the barrel field are closely related to
to recording. In the case of deprived animals, the principal and surround vibrissae are deprived for 18 days and their position within the anatomical
regrow for 811 days before recording, and the D1 whisker remains intact throughout. Averages shown are barrel map in undeprived animals, a
group means; errors are standard errors; n is number of animals in each group.
deviation in response for a given posi-
tion is a sensitive index of plasticity.
Plasticity is impaired in adult T286A mutants The spared vibrissa domain did not expand in T286A mutants
The T286A mutation prevented potentiation of spared vibrissa (Fig. 1), indicating a lack of plasticity. Second, the weighted vib-
responses in adult mice. The ratio between the deprived and unde- rissae dominance index (WVDI), a measure of the response of
prived D1 vibrissa response was unchanged at 1.07 (Table 1). A each cell relative to the principal whisker, normally increases in
two-way ANOVA showed that deprivation affected the spared vib- wild-type animals (Methods). In this case, each cell is its own
rissa response in general (F1,33 = 16.58, p < 0.0003) but also indi- control; the spared whisker response is measured relative to the
cated an interaction between deprivation and genotype (F1,33 = principal vibrissa response, which, being unaffected by depri-
5.8, p < 0.03). Post-hoc tests showed that the interaction term was vation in adults, acts as a reference (Fig. 2). The vibrissae dom-
due to a lack of D1 potentiation in the T286A mutants. For the inance index shifted substantially for wild-type mice (deprived
Fig. 1. The spared D1 vibrissa domain

expands in adult wild-type mice but a b c
not in T286A littermates following
deprivation. Position of each penetra-
tion relative to the barrel field is
shown. The color of each circle repre-
sents the intensity of response to the
spared D1 vibrissa at each location
(black, > 50 spikes per stimulus; gray,
25 < D1 < 50; white, D1 < 25).
Responses are calculated from pene-
trations containing at least four cells
and are responses of layer 2/3 cells
only. (a) In undeprived adult wild-type
mice, most of the responses outside d e f
the D1 barrel itself (dark shading) to
D1 stimulation are small (white cir-
cles). (b) In adult wild-type mice,
responses to spared vibrissa stimula-
tion increase following deprivation, so
that most penetration sites respond at
the highest level (black circles).
(c) Average response as a function of
distance from the edge of the D1 bar-
rel (0 m) out across the surrounding
barrels for deprived (black diamonds)
and undeprived (white squares) wild-
type mice. Note that the deprived curve indicates larger response over at least 300 m. Asterisks indicate significant differences (at = 0.005). (d) Undeprived adult
T286A mutants and wild-type littermates show similar responses to D1 stimulation. Note the barrel nomenclature. (e) Adult T286A mutants show no clear pat-
tern of D1 domain expansion following deprivation. Three penetration sites in D2 respond at high levels, but this is not repeated at any other location. (f) Average
response as a function of distance for deprived (black circles) and undeprived (white squares) homozygotes. Note that the curves overlap completely at all dis-
tances, and no case is significantly different (at = 0.05). All figures show summary data pooled from all animals.

articles
Fig. 2. Vibrissae dominance histograms shift because of

potentiation in wild-type mice but do not shift in T286A
homozygotes. (a) Distribution of vibrissae dominance values
(Methods) for cells in undeprived (black) and deprived
a b
(hatched) adult wild-type mice. The distribution is predomi-
Percent of cells in sample

nantly biased to the left for undeprived animals because of
the dominance of the principal vibrissa. However, after 18
days deprivation, the distribution shifts rightward because of
increased response to the spared D1 vibrissa. Neuronal sam-
pling to construct histograms was approximately 21 cells per
animal for undeprived mice (5.2 cells per penetration) and 34
cells per animal for deprived mice (6.2 cells per penetration).
(b) The cumulative distribution function (CDF) does not
change following deprivation (gray diamonds, undeprived
responses; open squares, deprived (regrown) whisker
responses), indicating that the VDH shift is not due to
depression of the principal whisker response. (c) The lower
histograms show the effect of deprivation on vibrissae domi-
nance in T286A mutants. The distributions are very similar, c d
indicating lack of plasticity. Neuronal sampling was approxi-
mately 30 cells per animal for undeprived T286A mutants Percent of cells in sample

(5.6 cells per penetration) and 32 for deprived mice (5.3 cells
per penetration). Sampling radius includes all barrels sur-
rounding D1. (d) As for wild-type mice, the CDFs for the
adult T286A mutants do not show signs of depression follow-
ing deprivation, unlike adolescents (Fig. 4).
mean, 0.41; undeprived, 0.17; t17 = 3.5, p < 0.003) but

did not change for homozygous mutant mice
(deprived mean, 0.21; undeprived, 0.23; t7 = 0.422, p > 0.68).
In adults, the shift in the VDH is due to potentiation of the mals18,20. We therefore examined plasticity in adolescent T286A
spared whisker response and not to a decrease in the deprived mutants, and found that it was severely impaired. The spared D1
principal whisker response. The cumulative distribution func- vibrissa response did not increase following deprivation relative
tions for responses to principal whisker stimulation in deprived to undeprived responses (t6 = 0.52, p > 0.6) or relative to deprived
and undeprived animals (Fig. 2) were indistinguishable for wild- surround receptive field responses recorded in the same animals
type and mutant mice (wild type, Dmax = 0.10, p > 0.05 ; T286A (t8 = 1.9, p > 0.08; Table 2). Similarly, the spared whisker domain
Dmax = 0.15, p > 0.05, KolmogorovSmirnov 2-sample test), indi- increased only marginally in the cortex (Fig. 3f).
cating that principal vibrissae responses are not depressed by The T286A mutation, although it affected potentiation, did not
deprivation in adults. prevent depression of deprived whisker responses in adolescents. The
median principal whisker response decreased from 1.07 spikes per
Plasticity is impaired in adolescent T286A mutants stimulus to 0.6 following deprivation in wild-type mice (Fig. 4b).
We previously found that mutations affecting adults do not nec- Similarly, in T286A homozygotes, the median response decreased
essarily affect plasticity in 12 month-old (adolescent) ani- significantly from 1.16 spikes per stimulus to 0.74 (Fig. 4d; also see
supplementary figure at http://neu-
Table 2. Results for adolescent animals (12 months). rosci.nature.com/web_specials), and both
-CaMKII T286A effects were significant (wild type,
D1 Surround n D1 Surround n
Dmax = 0.26, p < 0.05; T286A, Dmax = 0.31,
p < 0.05, KolmogorovSmirnov).
Wild type 0.44 0.08 0.48 0.06 5 0.12 0.15 0.21 0.05 4
undeprived Inspection of the VDH for homozy-
gous mutants suggests some plasticity,
Wild type 0.95 0.09 0.28 0.07 8 1.29 0.47 0.31 0.05 4 presumably due to depression rather
deprived than potentiation, as potentiation was
Heterozygotes 0.27 0.07 0.38 0.06 3 0.57 0.11 0.44 0.06 3 blocked (Fig. 4c). However, the shift did
undeprived not reach significance in T286A animals
(undeprived mean WVDI = 0.18;
Heterozygotes 1.06 0.14 0.37 0.12 4 0.95 0.11 0.37 0.07 4 deprived 0.29; t5 = 1.87, p > 0.1), where-
deprived
as it was highly significant in wild-type
Homozygotes 0 0.42 0.08 0.40 0.10 3 mice (undeprived mean WVDI= 0.07;
undeprived deprived 0.61; t6 = 5.1, p < 0.003). This
implies that potentiation and depression
Homozygotes 1.60 0.29 0.39 0.09 4 0.45 0.01 0.32 0.06 5
deprived cause the VDH shift in wild-type mice
but, with potentiation blocked by the
Responses to stimulation of the D1 whisker and surround whiskers. Numbers are group mean responses T286A mutation, depression alone is
(spikes per stimulus). Conventions same as Table 1.
insufficient to cause a significant shift.

articles
Fig. 3. The spared vibrissa domain

does not expand for deprived adoles-
cent T286A mutants. (a) In unde- a b c
prived adolescent wild-type animals,
cells in layers 2/3 respond at low lev-
els to D1 stimulation outside the D1
barrel. (b) In adolescent wild-type
mice, following deprivation, the D1
domain expands, as evidenced by the
increased incidence of high-level
responses to D1 outside the D1 bar-
rel (black circles). (c) The D1
response as a function of distance
from the edge of the D1 barrel (con-
ventions as in Fig. 2). Note that the d e f
response is greater in the deprived
wild-type mice (black diamonds) over
a distance of at least 500 m. Asterisks
denote significance at = 0.005.
(d) Undeprived adolescent T286A
mutants. (e) With the exception of

one penetration located close to the
D1 barrel in D2, cells do not respond
at high levels to D1 vibrissa stimula-
tion outside the D1 barrel following
deprivation of adolescent T286A
mutants. (f) Response levels do not
increase systematically at distances away from the D1 barrel following deprivation in T286A mutants (black diamonds) compared to undeprived
(white squares). One point is different at = 0.05, but the response level is very low compared to the equivalent distance in deprived wild-type mice
(c). All figures show summary data pooled from all adolescent animals.
Plasticity is rescued in T286A heterozygotes in adolescents (Tables 1 and 2; Fig. 5c). An ANOVA revealed that
Heterozygous mice have one normal CaMKII gene and one age had no effect for heterozygotes (F1,36 = 0.15, p = 0.69), so we
mutant CaMKII gene, which means that the neuron contains a analyzed adults and adolescents together. The absolute magni-
mixture of phosphorylatable and unphosphorylatable -CaMKII. tude of response to D1 vibrissa stimulation was significantly dif-
To study whether the amount of switchable -CaMKII
is important for neocortical plasticity, we analyzed het-
erozygous mutants. For T286A heterozygotes, the degree
of spared whisker potentiation seemed normal at 2.2 a b
times greater than control in adults and 1.7-fold greater
Fig. 4. Vibrissae dominance and principal whisker response

depression in adolescents (conventions same as in Fig. 2).
(a) The distribution is predominantly biased to the left for unde-
prived animals because of dominance of the principal vibrissa, but
shifts rightward because of the increased response to the spared
D1 vibrissa. Neuronal sampling to construct histograms was
approximately 28 cells per animal for undeprived mice (5.2 cells
per penetration) and 24 cells per animal for deprived mice (5.1
cells per penetration). (b) Part of the shift in VDH is due to
depression of the deprived principal whisker response, which
shifts left (open squares) relative to undeprived animals (gray dia- c d
monds). (c) The lower histograms show the effect of deprivation
on vibrissae dominance in T286A mutants. The shift is far smaller

than occurs in the wild-type littermates and not statistically signif-
icant. Neuronal sampling was approximately 34 cells per animal
for undeprived T286A mutants (5.2 cells per penetration) and 20
for deprived mice (4.1 cells per penetration). Sampling radius
includes all barrels surrounding D1. (d) As with wild-type mice,
CDFs for adolescent T286A mutants show depression following
deprivation. The deprived principal whisker responses are shifted
left at lower values compared to the undeprived responses, indi-
cating depression. The slight, though statistically insignificant, shift
in the VDH can therefore be explained by depression of the prin-
cipal whisker response (Results).

articles
Fig. 5. Absolute magnitude of response to stimulation of surround

receptive field whiskers in -CaMKII mutants and their wild-type litter-
mates. Responses to the spared D1 vibrissa (mean s.e.; gray bars,
-CaMKII null mutants; black bars, T286A mutants; hatched bars, equiv-
a
alent deprived but regrown surround receptive field vibrissa for each
case). Responses for undeprived animals (U) are shown next to
responses for animals deprived for 18 days of all but the D1 vibrissa (D).
(a) In adult wild-type mice (+/+), the spared D1 response is greater fol-
lowing deprivation, whereas in adult homozygous (/) -CaMKII
knockouts (left) and T286A mutants (right), the spared D1 whisker
response is indistinguishable from undeprived levels. (b) Average
responses for adolescent mice. Again, in wild-type mice, the D1 whisker
response is potentiated. However, although potentiation occurs in
-CaMKII adolescent knockouts (/, left), it does not occur in adoles-
cent T286A point mutants (/, right). (c) Susceptibility is reversed in
the heterozygotes. Adolescent (young) mice show potentiation for both
mutations. However, potentiation is reduced to insignificant levels in the
-CaMKII heterozygote (adult, gray bars), although potentiation occurs b
normally in the heterozygote T286A mutant (adult, black bars).
Responses are shown for 50 stimuli. All cells are located outside the D1
barrel. ND, no data available for this group.
ferent between deprived and undeprived animals (t 23 = 4.1,

p < 0.0005). The vibrissa dominance shift in T286A mutants was
comparable to that observed in wild-type animals at 0.41 for
deprived T286A adults and 0.57 for adolescents and was again
significantly shifted compared to undeprived T286A heterozy-
gotes (t15 = 4.3, p < 0.0006).
-CaMKII knockouts versus T286A point mutants

Comparisons between -CaMKII knockouts and T286A point
mutants can be grouped into three categories: between adult c
homozygotes (Fig. 5a), adolescent homozygotes (Fig. 5b) and
heterozygotes (Fig. 5c). In adult homozygotes, the effect of the
T286A point mutation was very similar to the effect of the
homozygous -CaMKII knockout (Fig. 5a). The average response
to D1 stimulation was potentiated in adult wild-type mice but
not in homozygous -CaMKII knockouts or T286A mutants
(Table 1).
In adolescent homozygotes, potentiation occurred in the
-CaMKII null mutants (Fig. 5b, left) but not in the point
mutants (Fig. 5b, right). The spared whisker response was sig-
nificantly greater than the surround whisker response in the
null mutants (t3 = 3.98 p < 0.025) but not in the point mutants
(Table 2, above). In adolescent mice, the point mutation there- CaMKII null mutation or the T286A point mutation. Further-
fore had a greater effect on plasticity than the -CaMKII more, the relationship between receptive fields and the anatom-
knockout. ical location of cells within the barrel map was normal. The
However, in adult heterozygotes, the knockout affected plas- principal vibrissa, independent of age, gave the greatest response
ticity, whereas the point mutation did not. Plasticity seemed to in undeprived animals and the shortest-latency response within
be present but reduced in adult heterozygous -CaMKII mutants layer 2/3 (18 of 18 penetrations in undeprived homozygotes, 20
(Fig. 5c, left), whereas it was normal in T286A heterozygotes of 21 in undeprived heterozygotes). This relationship does not
(Fig. 5c, right). Although the vibrissae dominance shift occurred hold in the case of developmental abnormalities or neonatal
in the knockout mice (WVDI = 0.35, t13 = 2.77, p < 0.02), their NMDA receptor blockade21,22.
spared whisker responses potentiated less than for wild-type mice The receptive field structure was also normal in CaMKII
(or heterozygous T286A mutants) at 1.36 times the undeprived mutants. The average response levels were normal for principal
levels, and the spared whisker response was not significantly dif- vibrissa (F1,13 = 1.01, p > 0.33) and surround receptive field
ferent from undeprived controls (F1,14 = 1.63, p = 0.22). In con- whiskers (F1,30 = 4.03, p > 0.05) in T286A point mutants, as for
trast, plasticity was normal in adult T286A heterozygotes -CaMKII mutants18. There was no difference in receptive field
(see above). size of neurons in layers 2/3 of adult T286A homozygous animals
(T1 = 2.19, p = 0.14, KruskalWallis test) and a small difference in
Receptive field development in CaMKII mutants adolescents of one whisker on average (average receptive field size,
The gross topology of the barrel field viewed after cytochrome 2.1 vibrissae in wild type, 3.4 in homozygotes, T1 = 4.5, p < 0.04).
oxidase staining was normal in mice homozygous for the - Finally, the response latencies to principal vibrissa stimulation were

articles
normal for cells located in layer 2/3 both in adolescent (wild-type due to depression of closed eye input. Here we found depression
latency to principal vibrissa stimulation, 13.2 1.1 ms; homozy- of deprived whisker input to be unaffected by either form of
gotes, 15.6 1.8 ms; T1 = 1.1, p = 0.29) and adult T286A point CaMKII mutation. In this way, CaMKII mutations would affect
mutants (wild type, 10.8 1.2 ms; homozygotes, 8.7 0.5 ms; potentiation processes such as LTP, spared-vibrissa potentiation
T1 = 1.0, p = 0.31), as were those for the -CaMKII mutants18. and open-eye potentiation in ocular-dominance plasticity but
not deprived-whisker or closed-eye depression.
DISCUSSION The present experiments coupled with knowledge of the dis-
The finding that two different mutations of the -CaMKII gene tribution of CaMKII in the cortex further demonstrate that exci-
both affect experience-dependent potentiation in the barrel cor- tatory pathways are involved in the expression of neocortical
tex makes it highly unlikely that either mutation is acting in a non- plasticity. We have shown that -CaMKII is required for plastic-
specific manner and conversely make it likely that -CaMKII itself ity, and this molecule is found in excitatory cells and excitatory
is specifically required for potentiation. Furthermore, the idea synapses onto excitatory cells3135. Conversely, it is not found in
that CaMKII acts as molecular switch for neocortical plasticity is inhibitory cells3134, nor in the postsynaptic densities of inhibito-
strongly supported given that the T286A mutants show no poten- ry synapses onto excitatory cells35. Other evidence also points to
tiation in either adult or adolescent animals. This restricts great- excitatory pathways being necessary for expression of cortical plas-
ly the number of ways in which the mutation could be affecting ticity. Barrel-cortex plasticity involves an increase in the lateral
plasticity because CaMKII is still capable of phosphorylation in spread of excitation beyond the principal column associated with
the T286A animals, it is just that it cannot reach an autonomous the whisker and into neighboring columns36. Lesions located in
state in which it becomes independent of the continued presence the septum between two barrels prevent lateral transmission of
of calcium. This raises the question of why the autonomous state excitation between the two barrels and abolish expression of plas-
should be important for neocortical potentiation. ticity once it has been induced15, thereby implicating lateral exci-
The autonomous state for LTP is required probably because tatory pathways in plasticity.
it allows the kinase to preserve its phosphorylating ability for an One criticism of experiments with gene-targeted mice is that
extended period of time. To overcome dephosphorylation by pro- the mutation may affect development. However, it is highly
tein phosphatase 1 (PP1)3,23, the kinase reaction has to run much unlikely that the results of the present studies are explained by
faster than the dephosphorylation reaction9. Autophosphoryla- developmental defects because, in the somatosensory cortex, the
tion achieves this end by increasing the affinity of the kinase for expression of -CaMKII begins after the major developmental
calcium/calmodulin, by removing the need for calcium to be pre- events have occurred37. The small difference in receptive field size
sent continuously and by creating a critical mass of phosphory- of one whisker in adolescent homozgotes is unlikely to be signif-
lated subunits within the holoenzyme10,19. It is suggested that icant, as the normal range of whisker receptive fields encoun-
autophosphorylation itself may not be enough in this regard and tered in any given animal is from 1 to 5 whiskers, and in any case
that direct inhibition of PP1 by the cAMP/PKA pathway is also the effect is not sustained into adulthood. A second criticism of
required to induce stable synaptic plasticity3. The extended peri- gene-targeting experiments is that they may yield false negative
od of CaMKII activation allowed by autophosphorylation pro- results because of compensatory mechanisms. There may be a
longs its association with the postsynaptic density protein PSD-95 case for this view based on the results of the present studies
following calmodulin-dependent translocation to the mem- because plasticity was not prevented by the null mutation in ado-
brane24. This allows the kinase to remain in contact with possible lescents (indeed it may even have been enhanced), whereas it was
substrates, such as AMPA receptors, for a longer period of time completely blocked by the T286A mutation in the same age
and thereby enhance synaptic transmission2527. group. This result could be explained if upregulation of a com-
The present data further support the idea that experience- pensating kinase occurs in the null mutants but not in the T286A
dependent plasticity is based on an LTP-like mechanism similar point mutants. Overcompensation could explain the greater plas-
to that observed in the Schaeffer collateralCA1 pathway in the ticity observed in adolescent -CAMKII knockouts. The com-
hippocampus because LTP is blocked in the hippocampus of the pensation would have to be age dependent to explain why it did
same point-mutant mice8. Interestingly, stimuli that trigger LTP not affect adult null mutants. In any case, this finding modifies
not only cause the autophosphorylation of -CaMKII13 but also our previous view that plasticity might be different between the
lead to synthesis of new -CaMKII14. In the same way, upregu- adults and adolescents and suggests instead that the same plas-
lation of -CaMKII is observed in a related form of plasticity fol- ticity mechanisms are involved from late development through
lowing monocular deprivation in adult monkey visual cortex28. to adulthood.
Therefore, two stages of CaMKII activation that occur during In conclusion, we have found that autophosphorylation of
LTP, namely CaMKII autophosphorylation and new CaMKII syn- CaMKII is essential for experience-dependent potentiation but
thesis, also occur during experience-dependent plasticity. not depression in the barrel cortex of the mouse. The results sup-
On the other side of the argument, CaMKII may not be crit- port the concept that neocortical plasticity is based not only on
ically involved in experience-dependent plasticity because knock- synaptic plasticity, but in particular on a specific mechanism of
out of the -CaMKII gene produces a variable (bimodal) synaptic plasticity requiring the bistability of a major postsy-
impairment of ocular dominance plasticity in visual cortex29, naptic density enzyme, CaMKII5,9.
which cannot be explained by a unimodal impairment of LTP in
visual cortex by the same mutation 7: LTP is also unimodally
impaired in hippocampus30. We checked to see whether there was
METHODS
Subjects. Recordings were made from 942 neurons in layers 2/3 of the
any evidence for a bimodal impairment in the -CaMKII knock- barrel cortex of 24 undeprived and 34 deprived mice that were either
outs and the T286A animals in the present studies and found that -CaMKII null mutants or their wild-type littermates. The null mutants
the results were normally distributed with a single mode close to were backcrossed 4 or 5 times into C57/BL/6 from the initial 129Ola
the mean. One possible resolution of the findings is that the plas- genetic background. Recordings were made from 1,587 neurons in layers
ticity observed in visual cortex of some -CaMKII knockouts is 2/3 of 25 undeprived and 39 deprived T286A mice or their wild-type lit-

articles
termates. The T286A mutants were backcrossed 4 times into a C57/BL/6 Note: Some additional data on plasticity in adolescent T286A
background from the original 129sv source. Adolescent animals were mutants can be found on the Nature Neuroscience website
about two months old, and adults were at least 6 months old. Animals (http://www.nature.com/neuro/web_specials/).
were genotyped by PCR as described8. We did not observe any seizure
activity in the T286A mutants either behaviorally or while recording elec-
trical activity. All experiments were done blind to genotype. ACKNOWLEDGEMENTS
We thank Paul Chapman and Frank Sengpiel for critical reading of the text and
Deprivation. To induce plasticity, we spared the D1 vibrissa and deprived Mervyn McKenna for histology. This work was supported by grants from NIH
all other vibrissae for 1719 days. After recording, deprived vibrissae were NS27759 and MRC(UK) to K.F. K.P.G. was supported by the German Research
allowed to regrow for 811 days. Litters were housed in standard cages Council (DFG).
throughout deprivation. Cages were changed every week. The depriva-
tion technique and its effect have been described extensively before36 and RECEIVED 18 MAY; ACCEPTED 31 JULY 2000
is known not to affect vibrissae follicle innervation and to act by alter-
ing activity patterns38 (H. Wallace and K.F., J. Physiology 515P, 51, 1999).
1. Malinow, R., Schulman, H. & Tsien, R. W. Inhibition of postsynaptic PKC or
Anesthesia and recording. Anesthesia was induced with metofane and CaMKII blocks induction but not expression of LTP. Science 245, 862866 (1989).
2. Malenka, R. C., Kauer, J. A., Zucker, R. S. & Nicoll, R. A. Postsynaptic calcium
maintained with urethane (1.5 g/kg body weight). Anesthetic depth was is sufficient for potentiation of hippocampal synaptic transmission. Science
monitored throughout the experiment, and supplements of urethane 242, 8184 (1988).
were injected if necessary18. For recording, the skull was thinned 24 mm 3. Bhalla, U. S. & Iyengar, R. Emergent properties of networks of biological
lateral to midline and 03 mm caudal to bregma. A small hole was made signaling pathways. Science 283, 381387 (1999).
in the skull for each penetration using a hypodermic needle to introduce 4. Kennedy, M. B., Bennett, M. K. & Erondu, N. E. Biochemical and
immunochemical evidence that the major postsynaptic density protein is a
a glass-insulated carbon fiber microelectrode into the neocortex for sin- subunit of a calmodulin-dependent protein kinase. Proc. Natl. Acad. Sci. USA
gle-neuron recording. All experiments were compliant with the U.K. 80, 73577361 (1983).
1986 Scientific Procedures Act. 5. Miller, S. G. & Kennedy, M. B. Distinct forebrain and cerebellar isozymes of
type II Ca2+/calmodulin-dependent protein kinase associate differently with
the postsynaptic density fraction. J. Biol. Chem. 260, 90399046 (1985).
Stimulation and analysis. Vibrissae were stimulated using a computer-con-
6. Silva, A. J., Stevens, C. F., Tonegawa, S. & Wang, Y. Deficient hippocampal
trolled piezoelectric stimulator. Fifty 1 degree stimuli were delivered at 1 long-term potentiation in alpha-calcium-calmodulin kinase II mutant mice.
Hz to every whisker that might lie in the receptive field for each neuron. Science 257, 201206 (1992).
Two parameters of the response were measured: the average magnitude of 7. Kirkwood, A., Silva, A. & Bear, M. F. Age-dependent decrease of synaptic
response (that is, the number of spikes counted in the 5- to 50-ms interval plasticity in the neocortex of alphaCaMKII mutant mice. Proc. Natl. Acad.
Sci. USA 94, 33803383 (1997).
after the stimulus minus spontaneous activity) and the modal latency (from 8. Giese, K. P., Fedorov, N. B., Filipkowski, R. K. & Silva, A. J.
whisker stimulation to first spike). The average magnitudes of response Autophosphorylation at Thr286 of the alpha calcium-calmodulin kinase II in
(recorded in the barrel column immediately adjacent to the D1 barrel col- LTP and learning. Science 279, 870873 (1998).
umn) to stimulation of principal (P), spared (D1) and surround whiskers 9. Lisman, J. A mechanism for memory storage insensitive to molecular
were calculated for each animal. The average for each group of animals test- turnover: a bistable autophosphorylating kinase. Proc. Natl. Acad. Sci. USA
82, 30553057 (1985).
ed was then calculated from the average from each individual in the group. 10. Lisman, J. & Goldring, M. Evaluation of a model of long-term memory based
The relative response to stimulation of the spared vibrissae versus the on the properties of the Ca2+/calmodulin-dependent protein kinase. Journal
response to the principal whisker (denoted F) was calculated for each de Physiologie 83, 187197 (1988).
cell, where (F) = D1/(D1 + P), and D1 and P are averaged D1 vibrissa 11. Lisman, J. A mechanism for the Hebb and the anti-Hebb processes underlying
learning and memory. Proc. Natl. Acad. Sci. USA 86, 95749578 (1989).
response and principal vibrissa response respectively for that cell. A vib- 12. Fukunaga, K., Muller, D. & Miyamoto, E. Increased phosphorylation of
rissa dominance histogram (VDH) could then be composed from the Ca2+/calmodulin-dependent protein kinase II and its endogenous substrates in
frequency of occurrence of each class of relative response (F). VDHs were the induction of long-term potentiation. J. Biol. Chem. 270, 61196124 (1995).
compared using the weighted vibrissae dominance index (WVDI) cal- 13. Ouyang, Y., Kantor, D., Harris, K. M., Schuman, E. M. & Kennedy, M. B.
culated for each animal and averaged for treatment groups. WVDI = (0F0 Visualization of the distribution of autophosphorylated calcium/calmodulin-
dependent protein kinase II after tetanic stimulation in the CA1 area of the
+ 1F1 + ...9F9)/9N where, F0 is the number of cells in 00.099 band, F1 hippocampus. J. Neurosci. 17, 54165427 (1997).
is the number of cells in 0.10.199 band, etc., and N is total number of 14. Ouyang, Y., Rosenstein, A., Gabriel, K., Schuman, E. R. & Kennedy, M. B.
cells in particular sample. An ANOVA was used to check for effects and Tetanic stimulation leads to increased accumulation of Ca2+/calmodulin-
interactions between variables, and t-tests were used for post-hoc analy- dependent protein kinase II via dendritic protein synthesis in hippocampal
neurons J. Neurosci. 19, 78237833 (1999).
sis. As described before18, these parameters (D1, WVDI) are distributed 15. Fox, K. The cortical component of experience-dependent synaptic plasticity
normally. Where distributions were not normal (deprived principal in rat barrel cortex. J. Neurosci. 14, 76657679 (1994).
whisker responses, receptive field sizes and latency of response) a 2 or 16. Wallace, H. & Fox, K. The effect of chessboard deprivation pattern on
KruskalWallis test was used. Cumulative distribution functions (CDFs) potentiation and depression of vibrissae responses in rat barrel cortex.
were used to analyze the skewed distribution of principal whisker respons- Somatosens. Mot. Res. 16, 122138 (1999).
17. Glazewski, S., McKenna, M., Jacquin, M. & Fox, K. Experience-dependent
es caused by deprivation in adolescents. Differences in CDFs were assessed depression of vibrissae responses in rat barrel cortex. Eur. J. Neurosci. 10,
using a KolmogorovSmirnov two-sample test. 21072116 (1998).
18. Glazewski, S., Chen, C.-H., Silva, A. & Fox, K. The requirement for CAMKII in
Histological identification. Recording locations were identified in sec- experience-dependent plasticity of the barrel cortex. Science 272, 421423 (1996).
tions of the barrel field stained with cytochrome oxidase39 from lesions 19. Miller, S. G. & Kennedy, M. B. Regulation of brain type II Ca2+/calmodulin-
dependent protein kinase by autophosphorylation: a Ca2+-triggered
made at the end of each penetration (1 A, 10 s tip negative). All record- molecular switch. Cell 44, 861870 (1986).
ings made from layers 2 and 3 were estimated as being between 30 to 20. Glazewski, S. et al. Impaired experience-dependent plasticity in barrel cortex
270 m down from the pia during recording and corrected if necessary of mice lacking the alpha and delta isoforms of CREB. Cereb. Cortex 9,
from the histology. The distribution of recorded penetrations within the 249256 (1999).
21. Welker, E. et al. Altered sensory processing in the somatosensory cortex of the
barrel field was found to be equivalent for deprived and undeprived wild- mouse mutant barrelless. Science 27, 18641867 (1996).
type, heterozygous and homozygous adolescent mice (p > 0.1, 2) and 22. Fox, K., Schlaggar, B. L., Glazewski, S. & OLeary, D. M. M. Glutamate
deprived and undeprived wild type and heterozygous mice in adults blockade at cortical synapses disrupts development of thalamocortical and
(p > 0.1, 2). The distributions for adult deprived and undeprived columnar organization in somatosensory cortex. Proc. Natl. Acad. Sci. USA
homozygotes differed by a small though significant amount (average dis- 93, 55845589 (1996).
23. Strack, S., Barban, M. A., Wadzinski, B. E. & Colbran, R. J. Differential
tance of penetration from edge of D1 barrel, 118 49 m, undeprived; activation of post-synaptic density associated and soluble Ca2+/calmodulin-
181 80 m, deprived). Therefore, as a precaution, all data were ana- dependent protein kinase II by protein phosphatase 1 and 2A. J. Neurochem.
lyzed as a function of distance (Figs. 1 and 3; Results). 68, 21192128 (1997).

articles
24. Shen, K. & Meyer, T. Dynamic control of CaMKII translocation and localization in kinase gene expression in the rat central nervous system. Neuroscience 46,
hippocampal neurons by NMDA receptor stimulation. Science 284, 162166 (1999). 825849 (1992).
25. Barria, A., Muller, D., Derkach, V., Griffith, L. C. & Soderling, T. R. 33. Jones, E. G., Huntley, G. W. & Benson, D. L. Alpha calcium/calmodulin-
Regulatory phosphorylation of AMPA-type glutamate receptors by CaM-KII dependent protein kinase II selectively expressed in a subpopulation of
during long-term potentiation Science 276, 20422045 (1997). excitatory neurons in monkey sensory-motor cortex: comparison with GAD-
26. Mammen, A. L., Kameyama, K., Roche, K. W. & Huganir, R. L. 67 expression. J. Neurosci. 14, 611629 (1994).
Phosphorylation of the alpha-amino-3-hydroxy-5-methylisoxazole4- 34. Sik, A., Hajos, N., Gulacsi, A., Mody, I. & Freund, T. F. The absence of a major
propionic acid receptor GluR1 subunit by calcium/calmodulin-dependent Ca2+ signaling pathway in GABAergic neurons of the hippocampus.
kinase II. J. Biol. Chem. 272, 3252832533 (1997). Proc. Natl. Acad. Sci. USA 95, 32453250 (1998).
27. Derkach, V., Barria, A. & Soderling, T. R. Ca2+/calmodulin-kinase II 35. Liu, X. B. & Jones, E. G. Localization of alpha type II calcium calmodulin-
enhances channel conductance of alpha-amino-3-hydroxy-5-methyl-4- dependent protein kinase at glutamatergic but not gamma-aminobutyric
isoxazolepropionate type glutamate receptors. Proc. Natl Acad Sci. USA 96, acid (GABAergic) synapses in thalamus and cerebral cortex. Proc. Natl. Acad.
32693274 (1999). Sci. USA 93, 73327336 (1996).
28. Hendry, S. H. & Kennedy, M. B. Immunoreactivity for a calmodulin- 36. Fox, K. A critical period for experience dependent plasticity in
dependent protein kinase is selectively increased in macaque striate cortex somatosensory cortex. J. Neurosci. 12, 18261838 (1992).
after monocular deprivation. Proc. Natl. Acad. Sci. USA 83, 15361541 (1986). 37. Burgin, K. E. et al. In situ hybridization histochemistry of Ca2+/calmodulin-
29. Gordon, J. A., Cioffi, D., Silva, A. J. & Stryker, M. P. Deficient plasticity in the dependent protein kinase in developing rat brain. J. Neurosci. 10, 17881798
primary visual cortex of alpha-calcium/calmodulin-dependent protein (1990).
kinase II mutant mice. Neuron 17, 491499 (1996). 38. Li, X., Glazewski, S., Lin, X., Elde, R. & Fox, K. The effect of vibrissae
30. Hinds, H. L., Tonegawa, S. & Malinow, R. CA1 long-term potentiation is deprivation on follicle innervation, neuropeptide synthesis in the trigeminal
diminished but present in hippocampal slices from alpha-CaMKII mutant ganglion and S1 barrel field cortical plasticity. J. Comp. Neurol. 357, 465481
mice. Learn. Mem. 5, 344354 (1998). (1995).
31. Benson, D. L., Isackson, P. J., Hendry, S. H. & Jones E. G. Differential gene 39. Wong-Riley, M. Changes in visual system of monocularly sutured or
expression for glutamic acid decarboxylase and type II calcium-calmodulin- enucleated cats demonstrable with cytochrome oxidase histochemistry. Brain
dependent protein kinase in basal ganglia, thalamus, and hypothalamus of Res. 171, 1128 (1979).
the monkey. J. Neurosci. 11, 15401564 (1991). 40. Glazewski, S. & Fox, K. The time-course of experience-dependent synaptic
32. Benson, D. L., Isackson, P. J., Gall, C. M. & Jones E. G. Contrasting patterns in potentiation and depression in barrel cortex of adolescent rats.
the localization of glutamic acid decarboxylase and Ca2+/calmodulin protein J. Neurophysiol. 75, 17141729 (1996).

articles
Competitive Hebbian learning

through spike-timing-dependent
synaptic plasticity
Sen Song1, Kenneth D. Miller2 and L. F. Abbott1
1 Volen Center for Complex Systems and Department of Biology, Brandeis University, Waltham, Massachusetts 02254-9110, USA
2 Departments of Physiology and Otolaryngology, Neuroscience Graduate Program, W.M. Keck Center for Integrative Neuroscience, Sloan Center for Theoretical
Neurobiology, University of California, San Francisco, California 94143-0444, USA
Correspondence should be addressed to L.F.A. (abbott@brandeis.edu)
Hebbian models of development and learning require both activity-dependent synaptic plasticity
and a mechanism that induces competition between different synapses. One form of experimentally
observed long-term synaptic plasticity, which we call spike-timing-dependent plasticity (STDP),
depends on the relative timing of pre- and postsynaptic action potentials. In modeling studies, we
find that this form of synaptic modification can automatically balance synaptic strengths to make
postsynaptic firing irregular but more sensitive to presynaptic spike timing. It has been argued that
neurons in vivo operate in such a balanced regime. Synapses modifiable by STDP compete for
control of the timing of postsynaptic action potentials. Inputs that fire the postsynaptic neuron with
short latency or that act in correlated groups are able to compete most successfully and develop
strong synapses, while synapses of longer-latency or less-effective inputs are weakened.
Hebbian learning, the development of neural circuits based on dence of synaptic modification on spike timing provides a mech-
correlated activity, relies on two critical mechanisms. The best anism that can lead to competitive Hebbian learning without
known of these is activity-dependent synaptic modification along requiring global intracellular signaling, or preset activity or
the lines proposed by Hebb1. Equally important is a mechanism synaptic efficacy levels.
that forces different synapses to compete with one another so Experimental evidence from several different preparations1120
that when some synapses to a given postsynaptic neuron are suggests that both the sign and degree of synaptic modification
strengthened, others are weakened2,3. For example, correlation- arising from repeated pairing of pre- and postsynaptic action
based rules of synaptic modification can provide a reasonable potentials depend on their relative timing. In neocortical slices14,
account of many aspects of development in visual cortex, but hippocampal slice17 and cell18 cultures, and tadpole tectum in
only when they are combined with constraints introduced to vivo19, long-term strengthening of synapses occurs if presynaptic
ensure competition4. Although Hebbian synaptic modification action potentials precede postsynaptic firing by no more than
has received support from experiments on long-term potentia- about 50 ms. Presynaptic action potentials that follow postsy-
tion and depression5, much less is known about the mechanisms naptic spikes produce long-term weakening of synapses. The
that generate competition between synapses. largest changes in synaptic efficacy occur when the time differ-
At first, it might seem that any mechanism that imposes com- ence between pre- and postsynaptic action potentials is small,
petition among synapses must involve a global intracellular sig- and there is a sharp transition from strengthening to weakening
nal that reflects the state of many synapses. The constraints used as this time difference passes through zero. We call this form of
in many models of Hebbian learning6, although not biophysi- synaptic modification spike-timing-dependent plasticity (STDP).
cally realistic, are based on this idea. Typically these constraints Synaptic modification by STDP-like rules has been studied in
limit the sum of synaptic strengths received by a cell, or the mean models of temporal pattern recognition21, temporal sequence
activity of the cell. Competition can also arise locally if the learning2224, coincidence detection25,26, navigation2729 and direc-
processes that modify synaptic strengths equilibrate at a preset tion selectivity30,31 (N.J. Buchs, J. Reutimann & W. Senn, Soc. Neu-
level of total synaptic innervation or postsynaptic activity. This rosci. Abstr. 25, 2259, 1999). We focus instead on the competitive
can be achieved through static mechanisms such as thresholds and stabilizing properties of STDP. The competitive nature of
and negative input correlations6, dynamic mechanisms involv- STDP has been noted19,25, but not studied in detail. Stability of
ing non-Hebbian synaptic growth or decay terms such as synap- an STDP-like rule in combination with non-Hebbian plasticity
tic scaling79, or shifts in the synaptic modification rule itself as in has been studied in a linear, stochastically spiking neuron model32,
the sliding threshold of the BCM model10. Here we explore an but we find qualitatively new behavior when the intrinsic non-
entirely different mechanism suggested by experimental results linearity of the spike-generation mechanism is taken into account.
on the effect of spike timing on long-term synaptic modifica- We find that STDP alone can lead to stable distributions of synap-
tion1120, in which different synapses compete for control of the tic conductances, subject only to a limit on the strengths of indi-
timing of postsynaptic action potentials. We show that the depen- vidual synapses. The synaptic conductance distributions produced

articles
fication requires the integral of the function F to be negative,

which ensures that uncorrelated pre- and postsynaptic spikes
produce an overall weakening of synapses. A negative integral of
F requires A-- > A++. The data are mixed on this issue. The
results that report similar time scales for synaptic strengthening
and weakening14,18,19 indicate rough equality between the two
effects and, in some cases, even suggest a slight dominance of
strengthening over weakening. The data showing a longer tem-
poral window for synaptic weakening17,20 support the dominance
of synaptic weakening over strengthening by STDP. In our sim-
ulations, we use A-/A+ = 1.05, except for Fig. 2d, where A-/A+
varies.
In the model we study, ga denotes the peak synaptic conduc-
tance (the synaptic conductance immediately after an isolated
Fig. 1. The STDP modification function. The change of the peak con- presynaptic spike) due to an excitatory synapse labeled by the
ductance at a synapse due to a single pre- and postsynaptic action
integer a (with a = 1,2,...,N). This conductance must always be
potential pair is F(t) times the maximum value gmax with t the time of
the presynaptic spike minus the time of the postsynaptic spike. In this positive, and is not allowed to exceed a maximum value gmax. A
pre- and postsynaptic spike pair separated by a time interval t
figure, F is expressed as a percentage.

modifies the peak synaptic conductance by an amount gmaxF(t).
The value A+ = 0.005 thus corresponds to a change of 0.5% of
the maximum synaptic strength per spike pair. If this modifica-
by STDP force the postsynaptic neuron into a balanced, irregu- tion rule would push the peak synaptic conductance beyond the
larly firing regime3342 in which it is sensitive to the timing of the allowed range 0 ga gmax, ga is set to the appropriate limiting
presynaptic action potentials it receives. Such sensitivity leads to value. A scheme for implementing this modification rule is pre-
competition among inputs for the control of postsynaptic spike sented in Methods.
timing. This allows STDP to selectively strengthen synapses of We examine how STDP acts on the excitatory synapses dri-
inputs with relatively shorter latencies or stronger mutual corre- ving an integrate-and-fire model neuron with N = 1000 excitatory
lations, while weakening the remaining synapses. and 200 inhibitory synapses (Methods). The excitatory synaps-
es are activated by various types of spike trains: uncorrelated spike
RESULTS trains generated by independent Poisson processes at various
The modeling studies we present are based on a spike-timing- rates, bursts of action potentials with different latencies, and par-
dependent synaptic plasticity rule in which a function F(t) tially correlated spike trains. The model neuron also receives
determines the amount of synaptic modification arising from a inhibitory input consisting of Poisson spike trains with a fixed
single pair of pre- and postsynaptic spikes separated by a time rate of 10 Hz. In the simulations, excitatory synapses are modified
t. The function (Fig. 1) based on their pre- and postsynaptic spike timing, whereas
inhibitory synapses are held fixed.
) if t < 0
F(t) = { AA exp(t/
+ +
exp(t/ ) if t 0
- - Balanced excitation
To function properly, a neuron must establish and maintain an
provides a reasonable approximation of the dependence of synap- appropriate level of excitation so that it can respond to its inputs
tic modification on spike timing observed experimentally. The by firing action potentials at reasonable rates. Response variabil-
parameters + and - determine the ranges of pre-to-postsynap- ity also provides a constraint on the synaptic inputs to a neuron.
tic interspike intervals over which synaptic strengthening and The responses of an integrate-and-fire model receiving many
weakening occur. A+ and A-, which are both positive, determine independent presynaptic inputs can be considerably less variable
the maximum amounts of synaptic modification, which occur than responses observed in vivo33. Correlations of input spike
when t is close to zero. timing, such as synchronization, can contribute to increased vari-
Experimental results suggest a value for + in the range of tens of ability34. However, many authors have noted that a high degree
milliseconds and, in the examples we present, we use + = 20 ms. of variability can also arise if the excitatory inputs to a neuron
Data from some preparations indicate that the temporal window are balanced relative to the inhibitory synaptic and membrane
for synaptic weakening is roughly the same as that for synaptic currents3541. The critical condition is that the mean input to the
strengthening14,18,19, whereas other results reveal a larger window neuron should only be sufficient to raise the membrane potential
for synaptic weakening17,20. We have run simulations under both to a point below, or slightly above, the threshold for action poten-
conditions. For the results we report here, we do not see a signif- tial generation, so that spike times are determined primarily by
icant difference between the two cases, and we use - = + = 20 positive fluctuations in the total level of input. As we will show,
ms throughout. STDP provides a mechanism by which this balance can be estab-
We determine the parameters A+ and A- by dividing the total lished and maintained over a wide range of input firing rates42.
modification measured experimentally for multiple spike pairs This results in a state in which presynaptic action potentials can
by the number of pairs. This assumes that the effects of individ- control the timing of postsynaptic spikes, and competition among
ual spike pairs sum linearly (see Discussion). In our simulations, synapses can arise.
A+ = 0.005, except in Fig. 2f, where A+ = 0.02. To set the value of To study the equilibrium distribution of synaptic strengths aris-
A-, we make the important assumption that synaptic weakening ing from STDP, we initially set the peak conductances of all exci-
through STDP is, overall, a slightly larger effect than synaptic tatory synapses of the model neuron to gmax, which produces a
strengthening26. Specifically, stable competitive synaptic modi- high firing rate. All the excitatory synapses to the model neuron

articles
received independent Poisson spike train inputs with the same to develop, so STDP only regulates the long-term average firing
average rate. After a period of adjustment, a steady-state condition rate, and the neuron remains highly sensitive to transient
was achieved in which the firing rate of the postsynaptic neuron changes of input firing rates.
and the distribution of peak synaptic conductances remained con- The coefficient of variation (CV) of the postsynaptic spike
stant. Although all the peak synaptic conductances started with train, which is the standard deviation of the interspike intervals
the same value, there is no stable equilibrium state with a uniform divided by their mean, is fairly large and remarkably indepen-
distribution. Instead, most of the peak synaptic conductances are dent of the input firing rate (Fig. 2c) when the distribution of
pushed toward the limiting values of zero or gmax(Fig. 2a and 2b). synaptic strengths due to STDP has equilibrated. This suggests
For low input rates, more synapses approach the upper limit that STDP regulates the variability of the postsynaptic response.
(Fig. 2a), and for high input rates, more are pushed toward zero The high degree of firing variability is primarily due to an over-
(Fig. 2b). This has the effect of keeping the total synaptic input to all balance between inhibitory and excitatory conductances in
the neuron roughly constant, independent of the presynaptic firing the model. A reasonable measure of this balance is the ratio of
rates. The split between strong and weak synapses is also affected by total inhibitory to excitatory currents when the membrane poten-
the values of gmax (fewer strong synapses develop for larger gmax) tial is at the action-potential threshold. STDP adjusts this ratio
and A-/A+. The initial distribution of synaptic strengths has no to be slightly greater than one over the entire range of presynap-
effect on the final steady-state distribution as long as the postsy- tic firing rates considered (Fig. 2d). This indicates a balanced
naptic neuron is initially firing action potentials. condition in which, on average, inhibitory effects are slightly
STDP has a strong regulatory effect on the steady-state fir- dominant at threshold.
ing rate of the postsynaptic neuron, which, for the equilibrium An additional contribution to firing variability comes from
distribution of synaptic strengths, increases by only about 1 Hz the reduction in the number of strong synapses for high input
for each 5 Hz increase in the input firing rate (Fig. 2c). In con- rates. Figure 2d shows the number of strong synapses (those with
trast, if the peak synaptic conductances are held fixed in this g 0.8gmax) for different presynaptic firing rates. For the value
model, the firing rate increase is over 100 Hz for a 5 Hz increase of gmax we used, roughly half the synapses are strong for a 10 Hz
in the input firing rates. Synaptic changes due to STDP take time presynaptic rate. The number of strong synapses drops to 10%
when the presynaptic rates are set to 40 Hz. In all
cases, the balance between inhibition and excita-
a b tion is the dominant source of variability, but the
reduction in the number of strong inputs also con-
tributes when the presynaptic firing rates are high.
Both the firing rate and the coefficient of vari-
ation of the postsynaptic neuron depend on the
ratio A-/A+ (Fig. 2e). If this ratio is slightly larger
than one, the firing rate of the postsynaptic neuron
is maintained in a reasonable range, and the CV is
close to one, indicating an irregular postsynaptic
spike train.
Synaptic conductances tend to be pushed close
to the upper and lower limits of their allowed range
c d
Fig. 2. Balanced excitation and irregular firing produced
by STDP. (a) Histogram of the fraction of synapses taking
different peak conductance values ranging from zero to
g
max. For an input rate of 10 Hz, the peak synaptic conduc-
tances tend to the limiting values, but more are near gmax
than near zero. (b) Same as (a), but for an input rate of 40
Hz. Now more peak conductances are near zero than
near gmax. (c) The postsynaptic firing rate and CV (stan-
dard deviation divided by the mean) of the postsynaptic
interspike intervals for different input firing rates. (d) The
ratio of total inhibitory to excitatory currents that would
flow if the membrane potential were clamped at threshold,
e f and the percentage of strong synapses (g ) for dif-
0.8g
max
ferent presynaptic firing rates. The leak conductance is
included as a contributor to the inhibitory current in this
ratio because it acts to hyperpolarize the neuron. (e) The
postsynaptic firing rate and CV of the postsynaptic inter-
spike intervals for input firing rates of 10 Hz but different
values of A-/A+, the ratio of the amplitudes of maximal
synaptic weakening and strengthening. (f) Same as (a), but
with gmax 2.33 times larger and the synaptic modification
= 0.035, A = 0.020,
per spike pair four times larger (g max +
A- = 0.021). The larger value of gmax forces more synapses
to lower conductance values, whereas the higher modifi-
cation rate fills in the distribution.

articles
Fig. 3. Correlation between pre- and postsynaptic action potentials

before and after STDP. The solid curves indicate the relative probability a
of any presynaptic spike occurring at time tpre when a postsynaptic spike
occurs at time tpost. A correlation of one is the value due solely to chance
occurrences of such pairs. The dashed curves show the STDP modifica-
tion function from Fig. 1. The time-integral of the product of the synaptic
modification curve and the correlation function determines whether, on
average, synapses are strengthened or weakened. (a) At the beginning of
our simulations, when all the peak synaptic conductances are set to their
maximal value, there is only a small excess of presynaptic spikes before a
postsynaptic action potential. (b) At the end of the simulations, when
STDP has established a steady-state distribution of conductances, there
is a larger excess of presynaptic spikes before a postsynaptic action
potential. In the steady state, this excess compensates for the asymmetry
b
in the STDP modification curve, that is, for the fact that A-/A+ > 1.
by the STDP modification rule we are using (Fig. 2a and 2b).

This results in a bimodal distribution. A more continuous dis-

tribution arises if the degree of synaptic modification per spike
pair is increased. In an example in which the equilibrium distri-
bution of synaptic conductances is roughly exponential except
for a small excess near g = gmax (Fig. 2f), the model maintains the is greater than the area under the strengthening part. Initially in
basic features of STDP, regulation of the postsynaptic firing rate our simulations, there is an overall weakening of the excitatory
and CV, but the synaptic conductance distribution more closely synapses because the small excess of presynaptic spikes occur-
matches the experimentally observed distribution of spontaneous ring before postsynaptic action potentials is not large enough to
synaptic (mini) potentials43, which provides one estimate of the overcome the excess of synaptic weakening imposed by the STDP
distribution of synaptic strengths. rule (Fig. 3a).
The reason that STDP achieves a balanced state can be under- As the excitatory synapses are weakened by STDP, the post-
stood from basic response characteristics of a neuron integrat- synaptic neuron enters a balanced mode of operation in which
ing many inputs. Such a neuron can operate in two different it generates a more irregular sequence of action potentials that
modes with distinct spike-train statistics and inputoutput cor- are more tightly correlated with the presynaptic spikes that evoke
relations38,39,42. When excitation is strong, as at the beginning of them. The total synaptic input in the balanced mode is, on aver-
our simulations, so that the mean input to the neuron would age, near or below threshold, so the postsynaptic neuron fires
bring it well above threshold if action potentials were blocked, irregularly, primarily in response to statistical fluctuations in the
the neuron operates in an input-averaging or regular-firing mode. total input. Because action potentials occur preferentially after a
The postsynaptic action potential sequences produced in this random fluctuation, there tend to be more excitatory presynap-
mode are significantly more regular than the presynaptic spike tic spikes before than after a postsynaptic response 38,39,42
trains that evoke them. The interspike intervals of the postsy- (Fig. 3b). The STDP rule achieves a steady-state distribution of
naptic response depend on the total synaptic input, but the peak synaptic conductances when the excess of presynaptic action
absolute timing of individual postsynaptic
action potentials is fairly insensitive to a b
presynaptic spike times. As a result, there
are roughly equal numbers of presynaptic
action potentials before and after each
postsynaptic spike39,42 (Fig. 3a). As we have
noted, the area under the synaptic weak-
ening portion of the STDP curve (Fig. 1)
Fig. 4. Reduction of latency by STDP. (a) The

initial peak synaptic conductances as a function c d
of the relative latency of their synaptic inputs.
(b) The initial postsynaptic response to a barrage
of excitatory input with burst onset for each
synapse occurring at the time of its relative
latency. (c) The steady-state peak synaptic con-
ductances plotted as a function of the relative
latency of the synaptic input. Short-latency
synapses have been strengthened, and long-
latency synapses have been weakened. (d) The
response of the postsynaptic neuron to the same
input barrage as in (b), but after STDP has modi-
fied the peak synaptic conductances as in (c).

articles
a b latency excitatory inputs while weakening those with

longer latencies. The ultimate effect of this synaptic
modification is to make the postsynaptic neuron
respond more quickly.
To illustrate this latency reduction, we considered
a model neuron receiving inputs that are silent except
for isolated events represented by bursts of spikes
with a Poisson distribution at 100 Hz for 20 ms. Dif-
ferent synapses are not activated precisely synchro-
nously during these events. Instead, each synapse is
assigned a relative latency chosen randomly from a
c d Gaussian distribution with a mean of zero and a stan-
dard deviation of 15 ms. The burst of action poten-
tials at a given synapse occurs at a time given by the
sum of its relative latency and the absolute latency
associated with the event.
Initially, all the synapses were set to the same
strength of 0.2gmax (Fig. 4a). This produces a response
in the postsynaptic cell that begins shortly after the

time marked zero, which indicates the mean input
latency, and lasts for about 25 ms (Fig. 4b). The input
events are then repeated periodically until the STDP
Fig. 5. Effects of input correlation, firing rate or variability on steady-state peak synaptic rule has established a fixed distribution of peak
conductances. Each input parameter is divided into 20 bins. The histograms show the aver-
age peak synaptic conductances for all inputs within a given bin. These values are the results
synaptic conductances. STDP strengthens short-
of averaging bimodal distributions of synaptic strengths within each bin. (a) The synaptic latency inputs to the maximum allowed level, gmax,
inputs have correlation parameters ca ranging from zero to 0.2 (ca = 0.2(a 1)/(N 1)) and and weakens synapses with longer latencies to zero
have been binned on this basis. The degree of correlation between any two inputs is deter- (Fig. 4c). This produces a quicker response in the
mined by the product of their correlation parameters. The correlation time constant is postsynaptic neuron, which fires almost 20 ms earli-
20 ms. The degree of correlation of a synapse has a strong effect on its peak conductance. er than it did originally (Fig. 4d).
(b) Same as (a), but with a correlation time constant of 200 ms. No effect of correlation
on synaptic strength is observed. (c) The synaptic inputs have different firing rates ra rang- Correlation-based Hebbian modification
ing from 10 to 40 Hz (ra = 10 + 30(a 1)/(N 1) Hz), and this range has been binned. No Factors that enhance the ability of a given synapse to
strong effect of rate on synaptic strength is observed. (d) The synaptic inputs have distrib-
rapidly evoke a postsynaptic response lead to its
utions of input firing rates with different standard deviations (labeled input variability)
ranging from 0 to 0.5 in units of the mean rate (a = 0.5(a 1)/(N 1)). No effect of vari-
strengthening through STDP. Correlating different
ability on synaptic strength is observed. In this example, c = 20 ms, as in (a). synaptic inputs so they are more likely to arrive
together in a cluster is an effective way of increasing
their ability to evoke postsynaptic action potentials.
By cooperatively generating action potentials, such a
potentials before postsynaptic firing compensates for the asym- cluster of synapses can grow stronger, while weakening other
metry in areas under the positive and negative portions of the synapses that are not part of the cluster. To study this effect, we
STDP modification curve42 (Fig. 3b). If the total excitatory drive generated input spike trains at rates that were correlated across
were weaker than that provided by this distribution, stronger synapses (see Methods) and examined the effect of STDP.
fluctuations of the total input would be required to cause post- Presynaptic firing rates were generated to have a correlation
synaptic spikes. This would create an even greater excess of presyn function that decayed exponentially with a time constant c and
aptic action potentials before postsynaptic firing, which would varied in amplitude across the population of synapses (see Meth-
cause an increase in synaptic strengths, driving the system back to ods). Specifically, the correlation between two cells a and b is cacb
the steady-state distribution. STDP thus modifies excitatory with ca and cb, which we call correlation parameters, varying from
synaptic strengths until there is a sufficiently, but not excessively, zero to 0.2 uniformly across the 1000 excitatory synapses. When
high probability of a presynaptic action potential occurring before the correlations decay rapidly (c = 20 ms, Fig. 5a), more-corre-
a postsynaptic spike. This causes the neuronal response to be sen- lated synapses become stronger, and less-correlated synapses
sitive to the timing of input fluctuations. weaken (compare Fig. 5a and 5b). This trend disappears for larg-
er correlation times (c = 200 ms, Fig. 5b). To be strengthened, a
Latency reduction group of inputs must fire together long enough to generate a
For uncorrelated stochastic presynaptic spike trains, chance deter- postsynaptic action potential, but must then stop firing so they
mines whether a given synapse will ultimately become weak or are not subsequently weakened. As a result, correlations have a
strong through STDP. When the presynaptic inputs are correlat- large effect when the correlation time constant is approximately
ed in various ways, the fate of individual synapses is controlled equal to the time constants + and - that govern the time scales
in a more systematic manner. STDP strengthens synapses that for STDP32.
fire before a postsynaptic spike and weakens those that fire later. Although the degree of strengthening produced by STDP is
Suppose, for example, that a stimulus causes a barrage of presy- sensitive to correlations, it is not strongly affected by other prop-
naptic inputs to fire with varying latencies, and that these laten- erties of the presynaptic spike trains. When input firing rates are
cies extend over a period longer than that required to evoke time independent and uncorrelated but vary uniformly across
postsynaptic spiking. In this case, STDP will strengthen shorter- the population of synapses, there is little tendency for synapses

articles
firing at either faster or slower rates to be preferentially strength- gesting that the resulting reduction of latency in the postsynaptic
ened or weakened by STDP (Fig. 5c). Higher firing rates increase response occurs in vivo. A phenomenon analogous to the reduc-
the speed at which synaptic modification occurs, but they do not tion of latency discussed here predicts that, when a rat moves
otherwise affect the final equilibrium distribution of maximal through a particular region of space, place cells active for that
synaptic conductance values produced by STDP. region should fire earlier after the rat has repeatedly traversed
The steady-state peak synaptic conductances are also insen- the area23,27. This effect has been observed experimentally29,45.
sitive to the degree of variability of the presynaptic input. When A key assumption in our model is that synaptic weakening by
we arranged the input firing rates to have a standard deviation STDP dominates over synaptic strengthening. This is critical for
that varied uniformly across the inputs, we found no tendency stability. If this assumption is not true, the results we have report-
for synapses with either more or less variable firing rates to be ed might nevertheless arise from a combination of STDP and
preferentially strengthened or weakened by STDP (Fig. 5d). homosynaptic long-term depression (weakening of presynaptic
The basic result of these studies is that STDP is insensitive to inputs that fire in the absence of a postsynaptic spike5). As long as
the average rate or degree of variability of a given synaptic input. It STDP strengthens causally effective inputs, while STDP and/or
is, however, strongly affected by correlations between different other forms of long-term plasticity more strongly weaken causal-
inputs, provided that they decay rapidly enough as a function of ly ineffective inputs, the basic results found here should apply.
time. Synapses with strong, rapidly decaying temporal correlations Our model of STDP involves two additional assumptions. We
are strengthened as a cluster and suppress other synapses that are assumed that the effects of spike pairs sum linearly. At least one
uncorrelated or have temporal correlations that last over longer contradictory effect has been reported, a dependence of synaptic
time periods. STDP thus shows the basic feature of Hebbian learn- strengthening on pairing frequency, including a threshold effect
ing, the strengthening of correlated groups of synapses, while dis- and frequency-dependent saturation14. Our model does not incor-
playing the desirable features of firing-rate independence and porate this finding, but we maintain presynaptic rates above the
stability and a novel dependence on correlation decay time. reported threshold frequency for synaptic strengthening14. We
also assumed that we could ignore delays of several minutes
DISCUSSION between pairing of pre- and postsynaptic spikes and the resultant
Although Hebbian synaptic plasticity is a powerful concept, it suf- induction of synaptic modification that are suggested by experi-
fers from a number of problems. First, synapses are modified ments14. If the effect is merely a delay, this has no impact on our
whenever correlated pre- and postsynaptic activity occurs. Such results. If, on the other hand, the process acts as a low-pass filter on
correlated activity can occur purely by chance, rather than reflect- the temporal dynamics of weight change (averaging the effects of
ing a causal relationship that should be learned. To correct for this, STDP over a long period of time and changing weights accord-
neural network models often use a covariance rather than corre- ing to this average), this could have a more significant impact. We
lation-based synaptic modification rule44. However, such a rule have re-run our simulations assuming such a low-pass filtering
cannot, in general, achieve competition between synapses6. This effect. We observed no changes in our results except for the case of
brings up a second problem of purely Hebbian modification; it is Fig. 2f, in which individual spike pairings caused larger changes
not competitive, so constraints must be added to obtain interesting than in the other examples. In this case, the impact of these larg-
results. STDP seems to solve both of these problems. Accidental, er changes is damped by the long-term averaging.
non-causal coincidences weaken synapses if, as we have assumed, STDP may modify the short-term synaptic plasticity proper-
the integral of the synaptic modification function is negative. Com- ties of a synapse as well as its efficacy, an effect that has been called
petition arises in a new way, not due to a global signaling or growth synaptic redistribution46. We have run simulations in which we
factor, or to an artificially imposed balance of nonspecific synaptic coupled the strengthening and weakening of synapses through
decay and growth terms, but rather through competition for con- STDP to the degree of depression exhibited by the synapse, in a
trol of the timing of postsynaptic action potentials. Inputs that manner consistent with synaptic redistribution. Although this
consistently are best at predicting a postsynaptic response become does not change the results we report, it does reveal an interest-
the strongest inputs to the neuron. Causality is a key element of ing interplay between STDP and short-term plasticity. The most
STDP. As Hebb suggested1, synapses are only strengthened if their effective way to strengthen a synapse under STDP is to have it
presynaptic action potentials precede, and thus could have con- release transmitter before a postsynaptic spike and then stop
tributed to, the firing of the postsynaptic neuron. releasing so that it does not get weakened by subsequent releases
STDP automatically leads to a balanced, irregular firing state occurring after postsynaptic activity. A high degree of synaptic
in which pre- and postsynaptic spike times are causally correlat- depression, which is a feature of strong synapses in the redistri-
ed. This result depends crucially on the nonlinearity of the spike- bution scheme46, assures this. STDP that acts to modify release
generation process. In a model in which the probability of spiking probability and change the degree of synaptic depression is thus
depends linearly on membrane voltage32, the correlation between an extremely effective and competitive mechanism for driving
pre- and postsynaptic firing does not change shape with overall individual synapses to strong or weak limits.
changes in synaptic efficacy, as it does in Fig. 3. Nonlinear effects, STDP, although it makes an important contribution to com-
which make causal inputoutput correlations grow relative to petition, probably cannot be the sole source of plasticity in Heb-
acausal correlations as overall synaptic efficacy decreases, are cru- bian learning situations. Like any other Hebbian modification
cial for producing the stabilizing and competitive effects of STDP rule, STDP cannot strengthen synapses without postsynaptic fir-
that we have discussed. STPD regulates both the rate and the ing. If, for some reason, the excitatory synapses to a neuron are
coefficient of variation of postsynaptic firing over a wide range too weak to make it fire, STDP cannot rescue them. A non-Heb-
of input rates. This represents a homeostatic regulatory function bian mechanism, such as synaptic scaling79, may serve this func-
of STDP, which is surprising given that, like the Hebb rule, it is tion instead. In the model of STDP we use, two sets of inputs that
destabilizing at individual synapses. fire at times separated by more than about 100 ms generate STDP
STDP can differentially strengthen the shortest-latency inputs independently and thus do not compete. Experiments suggest
evoked by a stimulus. There is some experimental evidence sug- that competition can nevertheless occur under these conditions47.

articles
Such a result could arise if the STDP temporal window for synap- fires an action potential at time t, ga is modified according to gaga +
tic weakening has a long enough tail, or if STDP is supplement- Pa(t)gmax. If this makes ga > gmax, ga is set to gmax. These modifications are
ed by sufficiently strong heterosynaptic long-term depression48 made after the jump in conductance described in the previous paragraph,
or competition induced by synaptic scaling79. but changing this order does not modify our results.
The presynaptic firing rates in Fig. 5a and 5b have the correlation
The size of the temporal windows over which synaptic
function ra(t)rb(t) = r2 + r2(2ab + (1 ab)cacb)exp(tt/c),
strengthening and weakening occur is critical in determining the where the angle brackets represent an average over the ensemble of rates,
effects of STDP. It would seem highly advantageous for window r = 10 Hz, and = 0.5. To generate such rates, we chose intervals of time
sizes to be different in various brain regions, to be modified dur- from an exponential distribution with mean interval c. For every inter-
ing stages of development, and perhaps to be dynamically val, we generated N + 1 random numbers, y and xa for a = 1,2,...,N, from
adjustable over shorter time scales as well. This would allow STDP Gaussian distributions with zero mean and standard deviation one and
to stay compatible with relevant input correlations. STDP seems a respectively, where a2 = 2 ca2. At the start of each interval, the fir-
to depend on NMDA receptors1419, and NMDA receptor sub- ing rate for synapse a was set to ra = r(1 + xa + cay), and it was held at
unit substitution might provide a mechanism for adjusting its this value until the start of the next interval. The correlation function for
time course. For example, the developmental transition from a Fig. 5d is ra(t)rb(t) = r2 + r2a2abexp(tt/c), and the rates were
computed using a similar procedure but with ca = 0 and a variable a.
predominance of NR2B to NR2A subunits leads to a faster decay
To ensure that our results do not depend on initial conditions, we
time of NMDA-receptor-mediated currents49. This might be asso- ran multiple trials of the simulations starting from different ran-
ciated with a reduction in the width of the STDP window50. STDP domly generated sets of initial synaptic weights. There were no
also depends on postsynaptic back-propagating action poten-
detectable changes. After convergence, the variability in CV and out-

tials14,15, and modification of their waveforms might also change put rate between trials was indistinguishable from that observed in
the timing requirements for synaptic plasticity. Finally, the results measurements within a trial. There is always a small degree of vari-
we report are sensitive to the ratio of the areas under the strength- ability over time after a simulation has converged because statistics
ening and weakening parts of the STDP curve (Fig. 1) and would are gathered over a finite time, inputs are stochastic, and individual
be more robust if this ratio were under the dynamic control of synapses continually change their values, although their overall dis-
tribution does not change significantly. We consider the synaptic dis-
the average postsynaptic firing rate. It will be interesting to see if
tributions to have converged when the output firing rate stops
future experiments reveal evidence of developmental or activi- changing in a systematic manner. This occurs in about 100 seconds of
ty-dependent meta-plasticity in either the amplitudes or decay simulated time. Stability has been checked in some simulations for as
times of STDP modification curves. long as 100 hours of simulated time, and we have never seen apprecia-
ble changes in output rate or CV once convergence is reached. To be
assured of convergence, all presented data were collected only after
METHODS 1000 seconds of simulated time.
The membrane potential of the integrate-and-fire model neuron we
use is determined by ACKNOWLEDGEMENTS
Research supported by the Sloan Center for Theoretical Neurobiology at
dV
m = Vrest V + gex (t)(Eex V) + gin (t)(Ein V) Brandeis University, the National Science Foundation (IBN-9817194), the
dt
National Institute of Mental Health (MH58754) and the W.M. Keck
with m = 20 ms, Vrest = 70 mV, Eex = 0 mV, and Ein = 70 mV. In addi- Foundation (L.F.A.); a Howard Hughes Predoctoral Fellowship (S.S.); and by
tion, when the membrane potential reaches a threshold value of 54 mV, R01-EY11001 from the National Eye Institute and an Alfred P. Sloan Research
the neuron fires an action potential, and the membrane potential is
Fellowship (K.D.M.). We thank Todd Troyer for discussions.
reset to 60 mV (parameters take from ref. 37). The synaptic conduc-
tances, gin and gex, and their related peak conductances (see below) are
measured in units of the leakage conductance of the neuron and are RECEIVED 1 APRIL; ACCEPTED 18 JULY 2000
thus dimensionless.
On arrival of a presynaptic action potential at excitatory synapse a, 1. Hebb, D. O. The Organization of Behavior: A Neuropsychological Theory
gex(t)gex(t) + ga, and when an action potential arrives at an inhibitory (Wiley, New York, 1949).
synapse, gin(t)gin(t) + gin, where ga and gin are the peak synaptic con- 2. Guillery, R. W. Binocular competition in the control of geniculate cell
ductances. Otherwise, both excitatory and inhibitory synaptic conduc- growth. J. Comp. Neurol. 144, 117130 (1972).
3. Miller, K. D. Synaptic economics: competition and cooperation in synaptic
tances decay exponentially,
plasticity. Neuron 17, 371374 (1996).
dgex dgin 4. Miller, K. D. in Models of Neural Networks, III (eds. Domany, E., van
ex = gex and in dt = gin. Hemmen, J. L. & Schulten, K.) 5578 (Springer, New York, 1996).
dt 5. Bear, M. F. & Malenka, R. C. Synaptic plasticity: LTP and LTD. Curr. Opin.
We have taken ex = in = 5 ms, gin = 0.05, and 0 ga gmax with Neurobiol. 4, 389399 (1994).
6. Miller, K. D. & MacKay, D. J. C. The role of constraints in Hebbian learning.
gmax = 0.015 (except for Fig. 4, where gmax = 0.02 and Fig. 2d, where Neural Comput. 6, 100126 (1994).
gmax = 0.035). For a 100 M input resistance, gmax = 0.015 corresponds 7. Turrigiano, G. G., Leslie K. R., Desai, N. S., Rutherford L. C. & Nelson, S. B.
to a peak synaptic conductance of 150 pS. Activity-dependent scaling of quantal amplitude in neocortical neurons.
Synaptic modification is generated in the model through N + 1 func- Nature 391, 892896 (1998).
8. Davis, G. W. & Goodman, C. S. Synapse-specific control of synaptic efficacy
tions, M(t) and Pa(t), for a = 1,2,...,N. These decay exponentially, at the terminals of a single neuron. Nature 392, 8286 (1998).
9. OBrien, R. J., et. al. Activity-dependent modulation of synaptic AMPA
dM dPa
- = M and + = Pa. receptor accumulations. Neuron 21, 10671078 (1998).
dt dt 10. Bienenstock, E. L., Cooper, L. N. & Munro, P. W. Theory for the development
of neuron selectivity: orientation specificity and binocular interaction in
Every time the postsynaptic neuron fires an action potential, M(t) is visual cortex. J. Neurosci. 2, 3248 (1982).
decremented by an amount A-, and every time synapse a receives an 11. Levy, W. B. & Steward, D. Temporal contiguity requirements for long-term
action potential, Pa(t) is incremented by an amount A+. M(t) is used to associative potentiation/depression in the hippocampus. Neuroscience 8,
decrease synaptic strength. If synapse a receives a presynaptic action 791797 (1983).
potential at time t, its maximal conductance parameter is modified 12. Gustafsson, B., Wigstrom, H., Abraham, W. C. & Huang, Y.-Y. Long-term
potentiation in the hippocampus using depolarizing current pulses as the
according to gaga + M(t)gmax. If this makes ga < 0, ga is set to zero. Pa(t) conditioning stimulus to single volley synaptic potentials. J. Neurosci. 7,
is used to increase the strength of synapse a. If the postsynaptic neuron 774780 (1987) .

articles
13. Debanne, D., Gahwiler, B. H. & Thompson, S. M. Asynchronous pre- and 31. Mehta, M. R. & Wilson, M. From hippocampus to V1: Effect of LTP on
postsynaptic activity induces associative long-term depression in area CA1 spatiotemporal dynamics of receptive fields. Neurocomputing 32, 905911
of the rat hippocampus in vitro. Proc. Natl. Acad. Sci. USA 91, 11481152 (2000).
(1994). 32. Kempter, R., Gerstner, W. & van Hemmen, J. L. Hebbian learning and spiking
14. Markram, H., Lubke, J., Frotscher, M. & Sakmann, B. Regulation of synaptic neurons. Phys. Rev. E59, 44984514 (1999).
efficacy by coincidence of postsynaptic APs and EPSPs. Science 275, 213215 33. Softky, W. R. & Koch, C. The highly irregular firing of cortical cells is
(1997). inconsistent with temporal integration of random EPSPs. J. Neurosci. 13,
15. Magee, J. C. & Johnston, D. A synaptically controlled, associative signal for 334350 (1994).
Hebbian plasticity in hippocampal neurons. Science 275, 209213 (1997). 34. Stevens, C. F. & Zador, A. M. Input synchrony and the irregular firing of
16. Bell, C. C., Han, V. Z., Sugawara, Y. & Grant, K. Synaptic plasticity in a cortical neurons. Nat. Neurosci. 1, 210217 (1998).
cerebellum-like structure depends on temporal order. Nature 387, 278281 35. Shadlen, M. N. & Newsome, W. T. Noise, neural codes and cortical
(1997). organization. Curr. Opin Neurobiol. 4, 569579 (1994).
17. Debanne, D., Gahwiler, B. H. & Thompson, S. M. Long-term synaptic 36. Tsodyks, M. & Sejnowski, T. J. Rapid switching in balanced cortical network
plasticity between pairs of individual CA3 pyramidal cells in rat hippocampal models. Network 6, 114 (1995).
slice cultures. J. Physiol. (Lond.) 507, 237247 (1998). 37. Troyer, T. W. & Miller, K. D. Physiological gain leads to high ISI variability in a
18. Bi, G.-q. & Poo, M.-m. Activity-induced synaptic modifications in simple model of a cortical regular spiking cell. Neural Comp. 9, 971983 (1997).
hippocampal culture: dependence on spike timing, synaptic strength and cell 38. Troyer, T. W. & Miller, K. D. in Computational Neuroscience, Trends in
type. J. Neurosci. 18, 1046410472 (1998). Research (ed. Bower, J.) 197201 (Plenum, New York, 1997).
19. Zhang, L. I., Tao, H. W., Holt C. E., Harris W. A. & Poo M.-m. A critical 39. Bugmann, G., Christodoulou, C. & Taylor, J. G. Role of temporal integration
window for cooperation and competition among developing retinotectal and fluctuation detection in the highly irregular firing of a leaky integrator
synapses. Nature 395, 3744 (1998). neuron model with partial reset. Neural Comput. 9, 9851000 (1997).
20. Feldman, D. E. Timing-based LTP and LTD at vertical inputs to layer II/III 40. Amit, D. J. & Brunel, N. Global spontaneous activity and local structured
pyramidal cells in rat barrel cortex. Neuron 27, 4556 (2000). (learned) delay activity in cortex. Cereb. Cortex 7, 237252 (1997).
21. Gerstner, W., Ritz, R. & van Hemmen, J. L. Why spikes? Hebbian learning and 41. van Vreeswijk, C. & Sompolinsky, H. Chaotic balanced state in a model of
retrieval of time-resolved excitation patterns. Biol. Cybern. 69, 503515 cortical circuits. Neural Comput. 10, 13211327 (1998).
(1993). 42. Abbott, L. F. & Song, S. in Advances in Neural Information Processing Systems
22. Minai, A. A. & Levy, W. B. Sequence learning in a single trial. INNS World 11 (eds. Kearns, M. S., Solla, S. A. & Cohn, D. A.) 6975 (MIT Press,
Congress of Neural Networks II 505508 (1993). Cambridge, Massachusetts, 1999).
23. Abbott, L. F. & Blum, K. I. Functional significance of long-term potentiation 43. Bekkers, J. M. & Stevens, C. F. J. Cable properties of cultured hippocampal
for sequence learning and prediction. Cereb. Cortex 6, 406416 (1996). neurons determined from sucrose-evoked miniature EPSCs. Neurophysiology
24. Roberts, P. D. Computational consequences of temporally asymmetric 75, 12501255 (1996).
learning rules: I. Differential Hebbian learning. J. Comput. Neurosci. 7, 44. Sejnowski, T. J. Storing covariance with nonlinearly interacting neurons.
235246 (1999) . J. Math. Biol. 4, 303321 (1977).
25. Gerstner, W., Kempter, R., van Hemmen, J. L. & Wagner, H. A neuronal 45. Mehta, M. R., Barnes, C. A. & McNaughton, B. L. Experience-dependent,
learning rule for sub-millisecond temporal coding. Nature 383, 7678 asymmetric expansion of hippocampal place fields. Proc. Natl. Acad. Sci. USA
(1996). 94, 89188921 (1997).
26. Gerstner, W., Kempter, R., van Hemmen, J. L. & Wagner, H. in Computational 46. Markram, H. & Tsodyks, M. V. Redistribution of synaptic efficacy between
Neuroscience (ed. Bower, J.) 665669 (Plenum, New York, 1997). neocortical pyramidal neurones. Nature 382, 807809 (1996).
27. Blum, K. I. & Abbott, L. F. A model of spatial map formation in the 47. Stryker, M. P. in The Biology of Change in Otolaryngology (eds. Ruben, R. J.,
hippocampus of the rat. Neural Comput. 8, 8593 (1996). Van De Water, T. R. & Rubel, E. W.) 211224 (Elsevier, Amsterdam, 1986).
28. Gerstner, W. & Abbott, L. F. Learning navigational maps through 48. Scanziani, M., Malenka, R. C. & Nicoll, R. A. Role of intercellular interactions
potentiation and modulation of hippocampal place cells. J. Comput. Neurosci. in heterosynaptic long-term depression. Nature 380, 446450 (1996).
4, 7994 (1997). 49. Tang, Y.-P. et al. Genetic enhancement of learning and memory in mice.
29. Mehta, M. R., Quirk, M. C. & Wilson, M. Experience dependent asymmetric Nature 401, 6369 (1999).
shape of hippocampal receptive fields. Neuron 25, 707715 (2000). 50. Yuste, R., Majewska, A., Cash, S. S. & Denk, W. Mechanisms of calcium influx
30. Rao, R. & Sejnowski, T. J. in Advances in Neural Information Processing into hippocampal spines: heterogeneity among spines, coincidence detection
Systems 12 (eds. Solla, S. A., Leen, T. K. & Muller K.-b.) 164171 (MIT Press, by NMDA receptors, and optical quantal analysis. J. Neurosci. 19, 19761987
Cambridge, Massachusetts, 2000). (1999).

articles
Multi-unit recordings reveal context-

dependent modulation of synchrony
in odor-specific neural ensembles
Thomas A. Christensen, Vincent M. Pawlowski, Hong Lei and John G. Hildebrand
Arizona Research Laboratories, Division of Neurobiology, University of Arizona, PO Box 210077, Tucson, Arizona 85721-0077, USA
Correspondence should be addressed to T.A.C. (tc@neurobio.arizona.edu)
We used neural ensemble recording to examine odor-evoked ensemble patterns in the moth antennal
(olfactory) lobe. Different odors are thought to evoke unique spatiotemporal patterns of glomerular
activity, but little is known about the population dynamics underlying formation of these patterns.
Using a silicon multielectrode array, we observed dynamic network interactions within and between
glomeruli. Whereas brief odor pulses repeatedly triggered activity in the same coding ensemble, the
temporal pattern of synchronous activity superimposed on the ensemble was neither oscillatory nor
odor specific. Rather, synchrony strongly depended on contextual variables such as odor intensity and
intermittency. Also, because of emergent inhibitory circuit interactions, odor blends evoked temporal
ensemble patterns that could not be predicted from the responses to the individual odorants. Thus
even at this early stage of information processing, the timing of odor-evoked neural representations
is modulated by key stimulus factors unrelated to the molecular identity of the odor.
The insect antennal lobe (AL), a structural and functional ana- defined and behaviorally relevant odor input (Fig. 1)2,8,12. Exper-
logue of the vertebrate olfactory bulb, receives input from axons iments in progress confirm that many of the complex odor-
of antennal olfactory receptor cells that converge and synapse in evoked neural interactions described here are also observed in
discrete modules of condensed neuropil called glomeruli non-MGC glomeruli of both male and female moths13.
(reviewed in refs. 1, 2). According to a long-standing hypothe-
sis, each glomerulus is functionally unique, and different odorants Intra- and inter-glomerular neural interactions
are represented in the brain by the coordinated activity of differ- In 12 recording sessions lasting at least 3 hours each, we were
ent combinations of glomeruli1,2. Activity-labeling studies in able to isolate spiking activity from over 50 AL neurons. Up to
insects reveal such odor-evoked spatial activity patterns35, but seven units were separated unambiguously in each experiment,
these methods cannot resolve the rapid changes in neuronal fir- and multi-unit activity was recorded simultaneously from up to
ing that underlie odor recognition, and occur on a time scale of eight separate recording sites distributed across a 75 300 m
milliseconds6,7. Similarly, electrophysiological studies show that layer of olfactory neuropil (Fig. 1c). In accord with previous find-
output or projection neurons (PNs) arborizing in the same ings, different odors evoked patterns of activity in overlapping
glomerulus respond to the same odorant8, but to date, simulta- yet distinct populations of AL neurons15 (Fig. 1d).
neous recordings and temporal analysis of spatially distributed Multi-unit recordings also confirmed the results of our ear-
glomerular activity are limited to only a few neurons 9,10. A lier single-neuron intracellular studies, showing that the popu-
method that combines the spatial resolution of activity labeling lation of PNs innervating a single glomerulus is physiologically
with the temporal resolution of a microelectrode would there- heterogeneous with respect to fine odor tuning7,8. Some PNs are
fore fill an important gap in helping to decipher the brains olfac- narrowly tuned to one odor, whereas the activity of others is
tory code. We find that a method originally developed to monitor modulated by other, often chemically similar odors12. Moreover,
distributed multi-neuron activity in the mammalian brain11 can some PNs are unresponsive to certain stimuli, whereas the ongo-
be used to explore the dynamic, spatially organized neural activ- ing activity of others is suppressed by odor stimulation14. This
ity patterns evoked by different odors across anatomically and range of complex network interactions is also a prominent feature
functionally defined glomeruli in the insect olfactory system. of mammalian olfactory circuits15.
Here we show that multi-channel silicon microprobes revealed In accordance with these findings, examination of odor-evoked
dynamic, non-linear and unpredictable interactions within ensemble responses revealed that different neurons, although
ensembles of AL neurons associated with specific, identified olfac- localized to the same glomerulus and responsive to the same odor,
tory glomeruli12. were easily discriminated by their unique spike characteristics
(Fig. 2a and b). Thus it is possible to identify different subtypes
RESULTS of neurons in ensemble recordings based on specific features of
We selected the macroglomerular complex (MGC) in male moths their spike waveforms10. We also detected the activities of different
(Manduca sexta) for these studies because the MGC is a well char- physiological subtypes of neurons at the same recording site
acterized array of identified olfactory glomeruli with chemically (Fig. 2b). Although this method does not allow us to determine

articles
aa multichannel cc dd the presence of a particular blend

odor A (1 ng)
microprobe or ratio of odorants1,2,7,8, or may fire
array Probe Map
only in other environmental or
ALs 1 1
2
behavioral contexts (see below).
OL 900.12
a b c 3 3
bb odor B (1 ng) 4 4
Neuron-specific firing patterns
b c
Another important feature of these
1
2
3
MGC population responses is that the fir-
MGC 35
c
4
ing patterns of different neurons
A+B (1 ng each)
spikes/s
b
150m within the odor-encoding ensem-
units
a
i
ble were specific (Fig. 3). The dis-
ii 0 charge patterns observed across all
iii
trial1 trial10 neurons in our sample were thus
divided into three broadly defined
Fig. 1. Ensemble recordings show the dynamics of population responses evoked by odor. (a) Experimental classes: fast, moderate, and dif-
set-up and position of the multi-channel probes in one of the antennal lobes (ALs). OL, optic lobe. Scale bar,
fusely spiking neurons (Fig. 3).
1 mm. (b) Identification of the positions of the three tracks of a microprobe in the AL, from sections imaged
in the confocal microscope (frontal view). The probe tracks were mapped to three locations previously iden- When the firing patterns of differ-
tified in three-dimensional reconstructions of the M. sexta AL17: track a was in the coarse central neuropil, b ent neurons recorded within a
in the array of typical glomeruli, and c in the dorsal neuropil of the ring-shaped toroid in the macroglomerular 25-m radius (that is, probably
complex (MGC). Scale bar, 250 m. (c) Horizontal reconstruction of probe tracks, and spatial mapping of within the same glomerulus) were
neural activity to specific recording sites (site separation, 25 m). Here, spike waveforms from three neurons compared, it was clear that
were clearly discriminated (1012 superimposed sweeps each), and examples of these spikes recorded simul- although these neurons responded
taneously from tracks b and c illustrate their isolation to different sites. Only the two neurons recorded from to the same stimulus, they did not
the MGC (green and orange) responded to sex pheromone, reflecting the spatial segregation of different odor always exhibit the same firing
signals to different glomeruli. (d) Surface plots depicting the spatiotemporal patterns of odor-evoked activity
dynamics (Fig. 3). Conversely, neu-
recorded across two probe tracks, both in the MGC. The color scale represents the response magnitudes
(maximum instantaneous frequency) of four neurons (three additional neurons were inactive), each localized
rons recorded at different sites, and
to a different electrode site. The response to each 100-ms odor pulse was averaged over 100 ms at the responsive to different odors, could
response peak. Each odor evoked synchronous activity in different, but overlapping neuron sets, and these pat- display the same firing dynamics
terns were reproducible over repeated trials. Note also that the odor blend (A + B) evoked a strong response (for example, neurons 1 and 2 in
from the neuron at site c-1, while simultaneously suppressing activity at all other sites. Fig. 2). In contrast to findings in
mammals18, these results demon-
strate that odor-coding neurons
localized to the same glomerulus do
the exact origins of these extracellularly recorded signals, it is clear not necessarily display the same firing patterns. This is signifi-
that spikes recorded at one electrode site are not necessarily prop- cant because temporal synchronization of firing between neu-
agated to another site only 50 m away (Fig. 1c). Because we have rons with the same discharge pattern could be involved in either
observed clear spatial segregation and differential attenuation of intra- or inter-glomerular integration of olfactory information19.
many of these signals across recording sites, we believe that dif- If functionally related classes of AL neurons synchronize their
ferent unit signals detected at the same site reflect the spike activ- activity in response to a specific odor blend, for example, this
ity of different neurons that are in close proximity to that could be an effective mechanism for binding these multiple sig-
electrode16. The recording sites were furthermore localized histo- nal streams into a more coherent representation of the blend
logically to specific, identified AL glomeruli17 (Fig. 1c). stimulus in the brain19,20.
Additionally, unresponsive neurons (for example, neuron 5
in Fig. 2b) were sometimes detected along with responsive neu- Temporal binding is context dependent
rons at the same recording site. This finding suggests that certain We tested this hypothesis by computing the extent of synchro-
types of neurons may participate in the coding ensemble only in nous firing across odor-coding AL ensembles (Fig. 4). For exam-
aa 1 2
Neuron#
3 4 5
Fig. 2. Simultaneous recordings from multiple neurons show functional dif-
ferences even among closely spaced cells in the ensemble. (a) Average bb 10
spike waveforms from five neurons discriminated from a single probe track
Control
(averages of 1264,631 events). These are aligned in columns with their
odor-evoked responses in (b). The position of the recording site yielding 0
10
the maximum spike amplitude for each neuron is shown next to each
Odor A
waveform in (a). Note that spike undershoot was key to distinguishing one
neuron from another. (b) The five neurons in this ensemble differed in their 0
responses to different odors. Peristimulus histograms show the summed Odor B
10
responses to multiple pulses (n = 20 for each row) with four different stim-
uli: air blank (control), 0.1 ng bombykal (odor A), 0.1 ng C15 (odor B), and 0
the blend of the two (A + B). Pulse duration was 100 ms; pulse interval was 10
1 s (odor onset at time 0). Note that the response profiles of the neurons A+B
were strongly heterogeneous, even in this relatively small population. The

0
maximum evoked response (number of spikes per 25-ms bin) for each neu- 0 0.5 0.5 0.5 0.5 0.5
Time (s)
ron is indicated by an asterisk.

articles
Fig. 3. Different neurons in olfactory glomeruli can be distinguished by

their different discharge patterns. Interspike interval (ISI) histograms
aa
600
and rasters (insets) reveal the three distinct temporal patterns of spiking
observed across all ensembles: (a) fast spiking (sporadic bursts of high- 400
frequency activity); (b) moderate spiking (slower, but broader distribu- 200 1s
tion of spike frequencies); (c) diffuse spiking (irregular firing). Between
983 and 4,375 events were used to calculate each histogram. These dif- 0
ferent patterns were frequently visible at the same recording site. bb 80
Counts/bin
40
ple, neurons 1 and 2 in Fig. 2 were both classified as moderate 0

spikers and showed a preference for different odors. If the binding cc
hypothesis is correct, we would expect to observe greater coactiv- 20
ity between these neurons when the odor blend is presented than
when either of the blend components is presented alone. What 10
we found was surprising: the odor blend indeed resulted in coac-
0
tivity between neurons 1 and 2, but only when presented at a low
0 0.05 0.1 0.15 0.2
dosage (Fig. 4b). At a higher dosage, the blend had the opposite
effect: complete desynchronization of spiking activity between the ISI (s)

two neurons. Importantly, however, this relationship did not hold
for all pairs of cells in the ensemble. Blend-evoked synchronization Further examination of ensemble patterns yielded clear evi-
between neurons 1 and 4, for example, was not apparent at the dence for complex and multiple blend effects within a single odor-
lower stimulus concentration, but was well developed at the high- encoding assembly. When compared to the responses evoked by
er dosage (Fig. 4b). The absence of synchronization between these odor B alone, the low-dosage odor blend evoked significantly greater
two neurons at the low dosage may have been a consequence of firing rates in neurons 2 and 3, whereas a significant blend-induced
mixture suppression, which was clearly evident in neuron 4. These suppression of activity was observed in neuron 4 (first column in
data therefore provide support for the temporal binding hypoth- Fig. 4b). This odor blend also led to increased coactivity in some
esis, but they also indicate that the expression of binding is a con- neuron pairs (neurons 1 and 2, 2 and 3, 2 and 4), leaving the other
text-dependent phenomenon, contingent on such stimulus pairs completely desynchronized. The higher-dosage odor blend
variables as concentration (Fig. 4b). also led to increased coactivity between neurons 2 and 3, but unlike
a 1s
Neuron # b 0.1 ng 1.0 ng
a 1
b
Neuron # Neuron #
2 1 2 3 4 5 1 2 3 4 5
c 5 5
3 o
Neuron #
n 4 4
4 t 3 3 100%
5 r 2 2
o 1 1
l 0
1 2 3 4 5 1 2 3 4 5
Mean rates of individual neurons
stim
1s
Fig. 4. Odor blends evoked network dynamics that could not be pre-
dicted from the responses to the individual odors in the blend. o
(a) Responses from the five-neuron ensemble (Fig. 2) are shown as d
o
rasters. Five 100-ms odor-blend pulses were delivered at 1 Hz to simulate r
naturally brief and intermittent stimulus conditions (stim). Peri-event his- A
togram (below rasters) illustrates how the overall activity of the ensemble
is time-locked to the stimulus pulses. Note, however, that some neurons
(such as #2) were more tightly stimulus-locked than others. Each odor
pulse evoked a discrete burst of spikes across the population, followed o
immediately by strong suppression (arrows) that helped to further syn- d
chronize the population response to the stimulus. (b) The complex pat- o
r
terns of neuronal synchrony evoked by two different odors (A, B) and
their blend (A+B) at two dosages are represented as color-coded coactiv- B
ity matrices. The number of synchronous events (10- or 15-ms window;
Fig. 5) evoked by each odor pulse was calculated for 0.5 s from stimulus
onset, and averaged over 20 trials (color scale ranges from 03.8 coinci-
dent spikes per stimulus; two-way repeated measures ANOVA; F = 18.28,
A
p < 0.005). The horizontal display beneath each matrix shows the mean +
firing rate for each neuron individually (color scale ranges from 05.5 B
spikes per stimulus; F = 13.79, p < 0.005). Responses over 20 consecutive
trials showed neither significant changes in intensity with time (F = 1.54,
p > 0.1) nor significant changes in synchrony with time (F = 1.47, p > 0.1),
in contrast to results reported in the locust AL27.

articles
)
(s
ce
ll2 1.
0
1.
0
1.
0 learn (even after a single trial) to associate an odor
e-
Ti
m
5 5 5
stimulus with a food reward24,25. By combining
se 0. 0. 0.
on behavioral conditioning with chronic ensemble
esp
R
0 0 0 recording, we now can monitor odor-evoked
0 0 0
ensemble activity before, during and after olfacto-
R
ry reinforcement training, and compare these

es
po
0. 0. 0.
ns
5 5 5
responses over time in the same animal.
eT
im
e-
(K.C. Daly, T.A.C., V.M.P., B.H. Smith & J.G.H.,

ce
ll1
1. 1. 1.
0 0 0
(s
Assoc. Chemoreception Sci. Abstr. 35, 2000.) This

)
approach will be used to probe how olfactory expe-

1 2 3
Odor-pulse train (1-s interval)
rience affects odor coding in the brain, and will
Fig. 5. Time course of odor-evoked coactivity within the ensemble. Dynamic correlation
facilitate the analysis of neural interactions under-
analysis revealed that the temporal structure of the ensemble response was modulated by stim- lying the acquisition, retention and extinction of
ulus dynamics. Each surface plot shows how the cross correlations between two neurons these learned associations.
changed as a function of time. The odor stimulus (odor B, 1 ng) was presented at time 0 in each Odor-modulated oscillatory dynamics are a
panel, and both neurons responded to repeated stimulus pulses after an expected latency. prominent feature of numerous olfactory systems,
(Only three stimulusresponse epochs are shown here for detail.) An incidence of high corre- including that of M. sexta26, and olfactory infor-
lation (spikes in the two neurons synchronized to within 15 ms) is shown as a prominent peak mation processing in both vertebrates and inver-
along the diagonal (indicated by the dashed line). Although there is some variability in the time tebrates is thought to involve oscillatory
course of the response from trial to trial, note how the period of maximum coactivity in each synchronization of ensemble activity9,20,21,27. How-
correlogram is time-locked to the stimulus pulse. This peak furthermore is followed consis-
ever, whether different odors are each encoded by
tently by a period of suppression and spike desynchronization until the next stimulus pulse.
Note also the absence of any regular periodicity in the responses to these brief, 100-ms odor
a unique oscillatory pattern of ensemble activity
pulses. remains controversial, because most studies have
not addressed how the brain discriminates one
odor from another when the stimulus is suffi-
ciently brief and/or intermittent that network oscil-
the lower-dosage blend, it also enhanced synchrony between neu- lations are unable to develop22. Using brief, intermittent odor
rons 1 and 3 (second column in Fig. 4b). At the same time, the high- pulses in the present study, we find that the temporal patterning of
er-dosage blend resulted in a significant reduction in synchrony odor-evoked ensemble responses is consistent with our previous
between other pairs (neurons 1 and 2, 2 and 4, 3 and 4). single-neuron data: that is, the firing patterns of glomerular PNs
Finally, we wanted to know whether the patterns of synchrony depend not only on the chemistry of the odor but also on the
we observed were governed by oscillatory network dynamics, as physical context in which the odor is delivered22. Our new data
in other invertebrate and vertebrate olfactory systems (reviewed also show that, because of emergent properties of glomerular net-
in ref. 21), or whether coactivity among neurons in an ensemble works, the patterns of synchrony among different members of an
is linked to other factors. We examined the spike trains from 17 odor-encoding ensemble are not the same for different concen-
pairs of AL neurons (tentatively identified as PNs) for evidence of trations of the same odor. Furthermore, the responses to odor
an underlying oscillation. Dynamic cross-correlation analysis blends cannot necessarily be predicted from the responses to the
revealed that brief odor pulses consistently failed to evoke rhyth- individual odors in the blend. We therefore propose that ensem-
mic spike discharge in single neurons, and furthermore failed to bles of olfactory PNs must use multiple and overlapping coding
evoke a rhythmic pattern of coactivity between neurons (Fig. 5). strategies to process olfactory information, and that these strate-
Instead, odor-responsive neurons were coactive only during a gies are matched to the particular circumstances surrounding odor
brief time window following the stimulus, and this transient syn- presentation. Through their complex spiking patterns, they must
chrony was followed by a pronounced period of suppression and report primary information about both the chemical identity of
return to baseline activity until the next odor pulse (Fig. 5). These the stimulus (through their specific spatial association with one
results collectively show that the formation of patterned odor or more glomeruli), as well as information regarding stimulus
representations in these glomerular networks involves both exci- intensity and the spatiotemporal distribution of odor in space
tatory and inhibitory cellular interactions, and that the tempo- (information that is ecologically relevant to locating the odor
ral structure of these ensemble responses is neither oscillatory source)22 . Our results also suggest that in some stimulus contexts,
nor stimulus specific, but strongly context dependent. oscillations are not necessary for synchronizing odor-encoding
neural ensembles at the first stage of glomerular processing. For
DISCUSSION animals that must continually monitor changes in a dynamic odor
The application of ensemble recording methods to the analysis of plume, a lack of dependence on oscillations is advantageousthe
olfactory-system function opens up many possible avenues for necessary information about stimulus dynamics could be easily
testing established hypotheses about the functional roles of olfac- corrupted or lost completely if an underlying periodicity were to
tory glomeruli in odor recognition and discrimination. Ensem- modulate the network response to odor.
ble data add a necessary dimension to the analysis of the complex
intercellular relationships that emerge when glomerular circuits
in the olfactory system process information under naturalistic
METHODS
Preparation. Manduca sexta (L.; Lepidoptera, Sphingidae) were reared
conditions, when stimulus presentation is intermittent and unpre- on an artificial diet under a long-day (14L:10D) photoperiod. Adult
22
dictable . In addition, this recording strategy enables us to explore moths, one to four days after emergence, were used in all experiments.
central coding of odor information in another important con- In preparation for recording, the moth was restrained in a plastic tube
textlearning. Behavioral studies using M. sexta, modeled after with its head and antennae fully exposed. The labial palps, proboscis and
proboscis-extension training in honeybees23, show that moths can cibarial musculature were then removed to allow frontal access to the

articles
brain7. The head was secured to the plastic tube with wax, and then the comments, and Heather Stein and A.A. Osman for technical assistance.
tube was fixed to the recording table and situated with the ALs facing Supported by grants and contracts from NIH/NIDCD and DARPA/CBS.
upward. The tracheae overlying one AL were then removed with fine for-
ceps. The brain was superfused slowly with physiological saline solution RECEIVED 18 APRIL; ACCEPTED 17 JULY 2000
(150 mM NaCl, 3 mM CaCl2, 3 mM KCl, 10 mM TES buffer and 25 mM
sucrose, pH 6.9) for the duration of the experiment.
1. Hildebrand, J. G. & Shepherd, G. M. Mechanisms of olfactory
discrimination: converging evidence for common principles across phyla.
Ensemble recording and data analysis. To study population respons- Annu. Rev. Neurosci. 20, 595631 (1997).
es across glomeruli, we used multi-channel silicon microprobe arrays, 2. Christensen, T. A. & White, J. in The Neurobiology of Taste and Smell Vol. 2
supplied by the University of Michigans Center for Neural Commu- (eds. Finger, T. E., Silver, W. L. & Restrepo, D.) (Wiley, New York, in press).
nication Technology. We selected a probe design that matched the spa- 3. Rodrigues, V. Spatial coding of olfactory information in the antennal lobe of
Drosophila melanogaster. Brain Res. 453, 299307 (1988).
tial dimensions of the MGC in the AL of male M. sexta (Fig. 1). 4. Joerges, J., Kttner, A., Galizia, G. & Menzel, R. Representations of odours
Fork-like probes with three tines, with four recording sites per tine and odour mixtures visualized in the honeybee brain. Nature 387, 285288
(eight were active) were inserted into the MGC and other glomeruli (1997).
using several key landmarks. The final depth of the probes was noted, 5. Distler, P. G., Bausenwein, B. & Boeckh, J. Localization of odor-induced
neural activity in the antennal lobes of the blowfly Calliphora vicina: a
and the exact locations of the three tines and their recording sites were [3H]2-deoxyglucose labeling study. Brain Res. 805, 263266 (1998).
verified histologically (Fig. 1b). Recording sites measured 9 12 m, 6. Murlis, J. in Insect Pheromone Research New Directions (eds. Carde, R. T. &
and were separated by 25 m; thus multiple sites could be situated in Minks, A. K.) 221231 (Chapman & Hall, New York, 1997).
a single glomerulus. 7. Christensen, T. A. & Hildebrand, J. G. Coincident stimulation with
Multi-neuron activity was recorded simultaneously on eight chan- pheromone components improves temporal pattern resolution in central
olfactory neurons. J. Neurophysiol. 77, 775781 (1997).

nels (Lynx-8 amplifier, Neuralynx, Tucson, Arizona). Spike data were 8. Hansson, B. S. & Christensen, T. A. in Insect Olfaction (ed. Hansson, B. S.)
extracted from the recorded signals and digitized at 20 kHz per chan- 125161 (Springer, Berlin, 1999).
nel using Data Waves Discovery acquisition software and Data Trans- 9. Stopfer, M., Wehr, M., MacLeod, K. & Laurent, G. in Insect Olfaction (ed.
lations 2821-G 16SE A/D board on a Pentium III platform (Data Wave Hansson, B. S.) 163180 (Springer, Berlin, 1999).
10. Christensen, T. A., Waldrop, B. R., Harrow, I. D. & Hildebrand, J. G. Local
Technologies, Longmont, Colorado). Filter settings (typically 0.63 kHz)
interneurons and information processing in the olfactory glomeruli of the
and system gains (typically 520 K) were software-adjustable on each moth Manduca sexta. J. Comp. Physiol. A 173, 385399 (1993).
channel. Multiple spike waveform parameters were extracted and used 11. Nicolelis, M. A. L. Methods for Neural Ensemble Recordings (CRC, New York,
to classify different units based on multi-dimensional cluster analysis. 1999).
Only those spike clusters that were verified statistically by multivariate 12. Heinbockel, T., Christensen, T. A. & Hildebrand, J. G. Temporal tuning of
odor responses in pheromone-sensitive projection neurons in the brain of
discriminant analysis (Wilks lambda test) were used for further analy- the sphinx moth Manduca sexta. J. Comp. Neurol. 409, 112 (1999).
sis. Each spike in each separated cluster was time stamped, and these 13. King, J. R., Christensen, T. A. & Hildebrand, J. G. Response characteristics of
data were used to calculate peristimulus histograms, inter-spike inter- an identified, sexually dimorphic olfactory glomerulus. J. Neurosci. 20,
val histograms, mean-rate values, and the number of nearly coincident 23912399 (2000).
firings between all neurons in each ensemble (NeuroExplorer, Nex Tech- 14. Christensen, T. A. & Hildebrand, J. G. Male-specific, sex pheromone-
selective projection neurons in the antennal lobes of the moth Manduca
nologies, Winston-Salem, North Carolina). All data were normalized to sexta. J. Comp. Physiol. A 160, 553569 (1987).
correct for components of the correlation caused directly by modula- 15. Mori, K. & Shepherd, G. M. Emerging principles of molecular signal
tion of the stimulus28. processing by mitral/tufted cells in the olfactory bulb. Seminars Cell Biol. 5,
6574 (1994).
16. Buonviso, N. & Chaput, M. A. Response similarity to odors in olfactory bulb
Odor stimulation. The solenoid-driven device for odor delivery has been output cells presumed to be connected to the same glomerulus:
described7,10,12. For most trials, several stimulus patterns were present- electrophysiological study using simultaneous single-unit recordings.
ed to the antenna: a single, 200-ms odor pulse was the most common J. Neurophysiol. 63, 447454 (1990).
stimulus, but 520 sequential 50 or 100-ms pulses were also presented 17. Rospars, J. P & Hildebrand, J. G. Anatomical identification of glomeruli in
to examine changes over repeated trials. The odorants used are the two the antennal lobes of the male moth Manduca sexta. Cell Tissue Res. 270,
205227 (1992).
essential components of the sex pheromone of M. sexta, bombykal 18. Buonviso, N., Chaput, M. A. & Berthommer, F. Temporal pattern analyses in
(E10,Z12-hexadecadienal; odor A) and E11,Z13-pentadecadienal (C15, pairs of neighboring mitral cells. J. Neurophysiol. 68, 417424 (1992).
a mimic of the second key component, E10,E12,Z14-hexadecadienal; 19. Rieke, F., Warland, D., de Ruyter van Steveninck & Bialek, W. Spikes:
odor B)12, as well as binary blends of the two. Exploring the Neural Code (MIT Press, Cambridge, Massachusetts, 1997).
20. Kashiwadani, H., Sasaki, Y. F., Uchida, N. & Mori, K. Synchronized
oscillatory discharges of mitral/tufted cells with different molecular receptive
Histological identification of recording sites. Immediately following a ranges in the rabbit olfactory bulb. J. Neurophysiol. 82, 17861792 (1999).
recording session, the brain was excised and immersed in 12% glu- 21. Gelperin, A. Oscillatory dynamics and information processing in olfactory
taraldehyde in 0.1 M phosphate buffer to increase contrast. Brains were systems. J. Exp. Biol. 202, 18551864 (1999).
fixed for at least one hour, then dehydrated by passage through a graded 22. Christensen, T. A., Waldrop, B. R. & Hildebrand, J. G. Multitasking in the
olfactory system: context-dependent responses to odors reveal dual GABA-
ethanol series, and finally embedded in Spurrs resin (Electron regulated coding mechanisms in single olfactory projection neurons.
Microscopy Sciences, Ft. Washington, Pennsylvania) in preparation for J. Neurosci. 18, 59996008 (1998).
serial sectioning. Sections were cut at 48 m on a sliding microtome, and 23. Faber, T., Joerges, J. & Menzel, R. Associative learning modifies neural
2-m optical sections were collected with a laser-scanning confocal representations of odors in the insect brain. Nat. Neurosci. 2, 7478 (1999).
microscope (Bio-Rad MRC-600, Cambridge, Massachusetts, equipped 24. Daly, K. C. & Smith, B. H. Associative olfactory learning in the moth
Manduca sexta. J. Exp. Biol. (in press).
with a Nikon Optiphot-2 microscope). This method reliably revealed the 25. Hartlieb, E. Olfactory conditioning in the moth Heliothis virescens.
location of each of the three tines of the silicon microprobes in the AL Naturwissenschaften 83, 8788 (1996).
without the need for tissue staining (Fig. 1b). 26. Heinbockel, T., Kloppenburg, P. & Hildebrand, J. G. Pheromone-evoked
potentials and oscillations in the antennal lobes of the sphinx moth Manduca
sexta. J. Comp. Physiol. A. 182, 703714 (1998).
ACKNOWLEDGEMENTS 27. Stopfer, M. & Laurent G. Short-term memory in olfactory network
We are grateful to David Anderson and coworkers for providing microprobes and dynamics. Nature 402, 664668 (1999).
28. Aertsen, A. M. H. J., Gerstein, G. L., Habib, M. K. & Palm, G. Dynamics of
technical support, and we thank Carol Barnes and Bruce McNaughton for neuronal firing correlation: modulation of effective connectivity.
advice. We also thank Kevin Daly and Brian Smith for discussions and J. Neurophysiol. 61, 900917 (1989).

articles
Fixation neurons in the superior

colliculus encode distance between
current and desired gaze positions
Andr Bergeron and Daniel Guitton
Department of Neurology and Neurosurgery, and Montreal Neurological Institute, McGill University, 3801 University St., Montreal, Qubec H3A2B4, Canada
Correspondence should be addressed to D.G. (dguitt@mni.mcgill.ca)
A visual scene is scrutinized during sequential periods of steady fixation, connected by saccades
that shift the visual axis (gaze) to new positions. During such exploratory scan paths, gaze
frequently strays from and then returns to salient features. How the brain keeps track of major
end-goals and intermediate subgoals is not understood. We studied the discharge of fixation
neurons of the brainstems superior colliculus during multiple-step gaze shifts composed of a
sequence of saccades made in the dark and separated by short periods of steady fixation. Cells
were initially silent. As sequential gaze saccades approached the goal, firing began; its
frequency increased progressively and peaked when gaze was on the remembered target
location. We conclude that these fixation neurons encode the error between desired and actual
gaze positions, irrespective of trajectory characteristics.
Neurons that discharge tonically when an animal fixates a RESULTS

salient target are found in the parietal and frontal lobes, thal- When the cat fixated the food target in the light, SCFNs dis-
amus, subthalamic nucleus, superior colliculus and brainstem charged tonically (Fig. 1b). When ambient illumination was
(reviewed in refs. 1, 2). The role of each of these nodes and extinguished, and the animal was suddenly in complete dark-
how they interact with the motor circuits that generate sac- ness, but gaze was still aligned on the now invisible target, tonic
cades is not understood. From the motor perspective, fixation discharge in all cells continued for at least 400 ms. In some cells
is the act of not moving the visual axis. In the superior col- (Fig. 1b), the rate in the dark during this protocol was similar to
liculus of alert cats37 and monkeys1,811, anatomically separate that in light. In others, the rate in the dark decreased, but never
neural systems (Fig. 1a) generate and suppress saccadic move- below 25 spikes per second in our sample. In general, our obser-
ments of the visual axis (gaze; Methods). Saccadic gaze shifts vations fully conformed to previous findings in the head-fixed
are generated via burst-like discharges of neurons in the supe- monkey1,10 that have been taken as evidence that the tonic dis-
rior colliculuss motor map4,12. At least some of these may be charge of SCFNs during attentive fixation is not solely visually
tectoreticular neurons whose axons leave the superior collicu- induced.
lus and project to eyehead motor circuits of the reticular for- The cat, with head unrestrained, has eye motion restricted to
mation4,13,14. A fixation zone in the superior colliculuss rostral about 15 degrees from center4,19. Thus, head motion in this
pole1,310,15 is thought to contribute to suppressing saccadic gaze animal is indispensable for the generation of orienting gaze shifts
shifts by inhibiting the motor map16 in concert with projec- over 15, which makes the cat an ideal model to study eyehead
tions of superior colliculus fixation neurons (SCFNs) to brain- gaze control. In a large, single-step gaze shift (Fig. 1c), the eye
stem omnipause neurons (OPNs)5,11,17,18. OPNs in turn inhibit saccade may be followed by a period of relative eye immobility,
the gaze saccade generator. called an eye-position plateau19, during which gaze motion is
SCFNs are also tectoreticular neurons4, but their discharges assured by the head. The end of the eye saccade does not corre-
are different than those of the motor map: SCFNs discharge ton- spond to the end of the gaze shift, and this permits a dissocia-
ically during steady fixation of a behaviorally significant target, tion between eye and gaze saccade trajectories that is useful for
and they pause during gaze and eye saccades1,3,4,7,8,10,11. It is pro- analyzing SCFN discharges.
posed3,4,810,15,16 that the tonic activity of SCFNs is important for In our protocol, before a gaze shift, the cat faced an opaque
preventing saccade generation, via projections onto omnipause barrier on which there was no target of significance. Monkey fix-
neurons, and that the pause permits motor tectoreticular neu- ation neurons have a low tonic discharge when there is no tar-
rons to excite the saccade generator, thereby generating a gaze get to be fixated1. So do cats. Indeed, in the period preceding
saccade. In support of this model, electrical stimulation of either target presentation, the cat undoubtedly anticipated target
the omnipause area or the rostral superior colliculus suppresses appearance to one or the other side of the barrier4,15, and a typ-
saccades5,6,11,18, and deactivation of the rostral superior collicu- ical neuron had a characteristically low tonic discharge1 (Fig. 1c),
lus leads to difficulty in maintaining fixation9. For SCFN tonic lower than during attentive fixation in the dark. This is due per-
activity, our results point to a role more complex than that of haps either or both to the lack of a significant fixation target and
simply suppressing saccades. to motor preparation for the next orienting gaze shift1,15. The

articles
Fig. 1. Overview of connectivity and discharge patterns of SCFNs.

(a) Large oval, dorsal view of the SC. Dotted circle in the rostral supe- a b
rior colliculus identifies a fixation zone whose precise boundaries are
unknown. The vertical and horizontal lines indicate locations where,
respectively, horizontal and vertical saccades are encoded. Small dark
circles with curved lines indicate tectoreticular neurons, whose axons
project to the brainstem reticular formation. Those emanating from
the fixation zone are SCFNs that project onto OPNs. Those emanating
from the remaining part of the motor map project to brainstem gaze-
saccade-generating circuits. (b) Cat steadily fixated a food target; at the
vertical dotted line, the ambient lighting was extinguished, and the cat
was in total darkness, as also indicated by the thick horizontal dark bar
(bottom). This procedure was repeated six times. Top, the six horizon-
tal gaze traces (eye-in-space (G) = eye-in-head (E) + head-in-space (H));
below, the associated discharges of SCFN, cell N26 in right superior
colliculus. Each dot represents an action potential, and each line of
action potentials corresponds to the equivalently placed gaze trace
above. Below the action potential traces is the averaged discharge, the c d
spike-density histogram. (c) A 45 single-step gaze (G) shift and associ-
ated discharge of a SCFN (cell M54). The eye, after the rapid saccade,
was relatively immobile in an eye-position plateau. For this and subse-

quent panels, the line of vertical tic marks below the eye position trace
(E) shows action potentials; vertical tic mark on G trace indicates when
target was presented to animal; dark circle indicates when ambient
lighting was extinguished; open circle when it was re-illuminated. (d) A
45 multiple-step gaze shift and associated discharges of cell N42a.
Gaze, in the dark, is on the remembered target location at the right
, H
, gaze and head velocity. In all traces, down
vertical dotted line. G
means leftward movement. Cells N42a and M54 in right superior col-
liculus. Gaze shifts generated in barrier protocol (Methods).
pause in discharge continued, in the dark, beyond the end of the

eye saccade, into the eye-position plateau, and activity resumed,
not in relation to a particular feature of the eye movement but, as
we discuss below, at a particular point in the gaze trajectory. The
firing rate continued to increase after the end of the gaze step, as
reported in monkey SCFNs1.
During leftward single-step gaze shifts for cell N40 (Fig. 2),
the cat oriented, in the dark, to a spatially fixed target situated at
distances varying from about 30 (top trace) to 10 (bottom trace) shift end. After the pause, the cell began discharging before gaze
from different initial fixation positions. The gaze trajectories, and arrived on the remembered target location in the dark, and firing
associated cell discharges, were aligned on gaze shift onset in frequency peaked when gaze arrives on target.
Fig. 2a. Cell N40 paused before gaze shift onset, irrespective of The gaze-related pause in all our SCFNs, during single-step
amplitude. In Fig. 2b, the same trials have been aligned on gaze gaze shifts, have been reported for the head-unrestrained cat 3,4
and are not surprising because the motor map of the cats supe-
rior colliculus encodes gaze shifts4,6,20. Furthermore, omnipause
neurons, which receive projections from SCFNs5,11,17, pause dur-
a b ing gaze shifts21. Here we focus on what determines when SCFN
firing resumes after the pause, by studying the multiple-step gaze
shifts (Fig. 1d) that are frequently generated by cats making large
shifts. In this example, the target, at about 50 from initial gaze
position, was attained using a sequence of three gaze saccades,
made in the dark and separated by short periods (gaze-position
plateaus) where gaze was moving at much lower velocity com-
pared to peak gaze saccade velocity. The eye-position trace shows
Fig. 2. Discharges of cell N40 (right superior colliculus) during 12 single-step

gaze saccades of different amplitudes. Cat initiated gaze shifts from different
initial positions to a fixed remembered target location. Trajectories aligned on
the start (a) and end (b), respectively, of the movements. As in Fig. 1b, the
respective 12 action potential traces are shown below the gaze traces. The
cumulative discharge, expressed as a spike density histogram, is shown below
the action potentials. Same symbols and conventions as in Fig. 1.

articles
Fig. 3. Discharges of two SCFNs during multiple-

step gaze shifts of different amplitudes. (ad) Cell
N40. (eh) Cell N42a. For each cell, the same trials a b c d
are shown in all panels, with traces aligned on differ-
ent features of the gaze trajectory. (a, d) Traces
aligned on onset of multiple-step sequences (vertical
dotted line). (b, e) Traces aligned on end of the first
fixation plateau that follows the first gaze step.
(c, f) Traces aligned on end of multiple-step sequence.
(d, h) Traces aligned on onset of ambient illumination.
Same symbols and conventions as in Figs. 1 and 2.
that, during gaze plateaus, the eye was rotating in

a direction opposite to head motion, presumably
due to the action of the vestibulo-ocular reflex. In
this trial, the head accelerated in synchrony with
each eye saccade, and gaze velocity during some e f g h
plateaus attained zero. In other trials, head veloc-

ity was smooth and fairly constant over a large
portion of the multiple steps.
For this study, we exploited a feature of gaze
trajectories: for a given overall gaze displacement
amplitude, the structure of the multiple-step gaze
shiftsnumber of steps and their amplitude and
duration, as well as gaze-plateau durationsvar-
ied from one trial to the next. As a result, multi-
ple-step gaze shift trajectories were very variable, in
contrast to the stereotyped eye-saccade trajecto-
ries usually considered in the literature.
A series of multiple-step gaze shifts (Fig. 3a)
was made by a cat as it oriented, in the dark, to a
spatially fixed target situated at distances varying
from about 50 (top gaze trace) to 25 (bottom
trace) from different initial fixation positions. The movements even though the visual axis was moving very slowly compared
are aligned on the onset of the first gaze step. The tonic discharge to saccade velocities, and in some trials it was immobile. This
of this typical SCFN, cell N40, ceased shortly after the target was contrasts strongly with the activity of SCFNs, at the end of single-
presented, and silence continued, in the dark, during and beyond step gaze shifts, during steady fixations of a remembered target
the first gaze saccade. In Fig. 3b, the same trajectories are aligned location (Figs. 1c and 2). In Fig. 3c, the movements are aligned on
on the end of the first gaze-position plateau, which followed the the end of the multiple-step trajectories. The activity of the cell
first gaze saccade. The SCFN was silent during these plateaus, had fully resumed by then and notably, in contrast to the pause at
the onset of multiple-step sequences, the cell was active during
a b the last gaze saccade and, in some trials, the gaze plateau that
preceded it and the second-to-last gaze saccade. For cell N40, at
s.d. of first spike (ms)
least part of the increase in firing beyond the end of the gaze shift
is due to a visual response, as shown by the visually induced burst
discharge in Fig. 3d, in which the traces have been aligned on the
Fig. 4. Calculation of minimum standard deviation (s.d.) of time of onset

of first spike. (a) For cells N40 and N42a, plot of s.d. when gaze traces
are aligned (left) on different values of gaze-position error (GPE) or
(right) on beginning (open symbols) or end (full symbols) of step whose
c d number is indicated on abscissa. Short horizontal line through each gaze
trace indicates value of s.d. above which s.d.s are significantly different
than that at the minimum of the curve. Thus, aligning on the end of the
multiple steps (GPE = 0) produces an s.d. that is significantly greater
than the minimum. (b) Same as in (a), but for single-step gaze shifts.
Note that s.d. is essentially invariant, no matter what GPE the first spike
is referred to. (c) Similar plots for 5 SCFNs. Identification of traces from
top to bottom along ordinate (GPE = 0) is as follows. Cat (number of
trajectories used in calculation), N(13); M(17); N(20); P(8); N(25). The
latter is for cell N27a. (d) Similar plots for 5 other SCFNs. Identification
of traces, P(6); P(19); T(6); N(7); M(61).

articles
a b c of the scale of this variability, see the cell

discharges in Fig. 3c and g.) As the traces
were aligned on increasing values of GPE,
the s.d. gradually decreased and reached a
minimum at GPE of about 9 and 18 for
these two cells, respectively.
By comparison, the curves for single-
step gaze shifts (Fig. 4b) are essentially flat,
thereby indicating that, for single steps,
this technique does not work for detect-
ing the optimal GPE for onset of the first
spike. This is because these gaze shifts are
too stereotyped, and each part of a trajec-
tory is correlated in time to another. Thus,
no part of a trajectory is better than anoth-
er for minimizing GPE.
In 10 other cells, 4 recorded in cat P, 3
in cat N, 2 in cat M and 1 in cat T (Fig. 4c
and d), each curve is stopped at the GPE

corresponding to the average at which the
first spike occurred. In some examples
(dashed line in Fig. 4d), the curves mini-
mum is reached at a GPE considerably
lower than that corresponding to the
appearance of the first spike. This is
because the first spike in these trials
occurred during the first gaze saccade, and
(as in Fig. 4b), for single-step gaze sac-
cades, the s.d. of the first spike remains
approximately constant no matter what
Fig. 5. Timing of first spike referred to preferred GPE. (a, c) For cells N40 and N42a during multi-
ple-step gaze shifts, the first spike in activity after the pause ended on average when the visual axis point in the saccades trajectory it is
was at GPE = 9 4 and 18 5 from target respectively. The example traces for each cell are referred to.
aligned on their respective values of preferred GPE. (b) Same cell and preferred GPE as in (a), but The reduction in the scatter of dis-
for single-step gaze shifts. Same symbols and conventions as in Figs. 1 and 2. charge onset provided by aligning multi-
ple-step gaze shifts on a cells optimal
value of GPE is apparent by comparing
Fig. 5a and c to Fig. 3c and g. This pro-
onset of ambient illumination. The discharge pattern (Fig. 3eh) cedure also slightly reduces the s.d. for single-step gaze shifts
of another typical cell (N42a) was very similar to that of cell N40 (compare Figs. 5b and 2b). However, this reduction is not sig-
except that its visual response on average was weaker (Fig. 3h). nificant (Fig. 4b).
After the pause, the firing frequency of this cell reached its max- Comparing the timing of the first spike between preferred
imum just at the end of the multiple-step sequence. GPE and end of gaze (Fig. 6a) provides further evidence that the
In these multiple-step gaze shifts, why were SCFNs silent dur- scatter in timing of discharge onset after the pause is greater
ing the first gaze saccade and subsequent gaze-position plateau, when the first spike is referred to the end of the multiple-step
but often active during the second-to-last saccade, the last plateau sequencewhen peak firing frequency occurscompared to
and the subsequent gaze saccade that brought gaze onto the tar- when it is referred to 9. For single-step gaze shifts, as expected
get? In other words, what determined the end of the pause and based on Fig. 4b, the scatter for alignments on 9 is about the
the appearance of the first spike in SCFN activity? We first same as for alignments on the end (Fig. 6b). The conclusions
explored this question by aligning, for a given cell, all gaze tra- for cell N42a are the same as for cell N40 (Fig. 6c and d). Note
jectories on the first spike that followed the pause in neural dis- again that, for single-step gaze shifts, the scatter in timing of the
charge. This procedure suggested that the first spike occurs when first spike relative to the optimal GPE is small, and similar to the
gaze arrives at a specific distance from the target, that is, at a spe- scatter relative to the end of the gaze shift. Studying single-step
cific gaze position error (GPE). Put another way, the minimum gaze shifts is not useful for distinguishing whether spatial or
variability in the timing of the onset of discharge should occur temporal characteristics of a gaze trajectory best describe when
when the first spike after the pause is referenced to a specific GPE. the first spike occurs (Fig. 6b and d). This is because these move-
To verify this hypothesis quantitatively, for each cell we deter- ments are too similar from one trial to the next, unlike for mul-
mined which point in a gaze trajectory was associated with the tiple-step gaze shifts.
minimal variabilitythat is, the lowest standard deviation Note that the minimum in the curve of s.d. versus GPE
(s.d.)in timing of the first spike. When all traces were aligned (Fig. 4) cannot be lowered by referring the first spike to differ-
on the end of the multiple-step sequences, when gaze was on tar- ent gross features of a multiple-step gaze shift sequence (Fig. 4a,
get (GPE of 0), the time of occurrence of the first spike relative right). For example, if for cells N40 and N42a, all traces were
to gaze-shift end had a standard deviation of about 120 ms and aligned on the beginning and end of the first, second or third
135 ms for cells N40 and N42a, respectively (Fig. 4a). (For a sense steps, the s.d. was significantly greater than at the optimal GPE.

articles
Fig. 6. Comparison of tim-

ing of first spike between
preferred GPE and end of a c
gaze sequence. (a) Com
parison, for cell N40, of dis- e
tribution of time of first
spike after the pause rela-
s.d. re best GPE (ms)

tive to preferred GPE
(= 9; ordinate) and end
of multiple-step sequence
(abscissa). Each point repre-
sents one multiple-step
sequence. Bar graphs along b d
each axis show distributions
for alignment on that GPE.
(b) Same cell and alignment
conditions as in (a), but for
single-step gaze shifts. Each
s.d. re end (ms)
point represents one single-
step gaze shift. (c, d). Same
as for (a) and (b), respec-

tively, but for cell N42a,
whose preferred GPE was 18. (e) For all cells, comparison of s.d. for alignment on optimal GPE versus alignment on end of multiple-step sequence. Each
point represents one cell. Closed symbols, contralateral gaze shifts; open symbols, ipsilateral gaze shifts. Triangle, cat H; circle, cat N; cross, cat M; diamond,
cat P; square, cat T. For all points outside the dotted contour, the s.d. is significantly less when alignment is made on the preferred GPE.
Furthermore, a cells preferred GPE was independent of the size jectory, such as the start or end of each gaze step. Note that there
of the overall multiple-step sequence. If the amplitude of the over- was no difference between cells that paused preferentially for ipsi-
all multiple-step gaze displacement was less than the cells pre- lateral or contralateral gaze shifts (Methods).
ferred GPE, the cell did not pause, but continued to discharge After the first spike, a cells firing frequency did not immedi-
during the entire movement (data not shown). ately reach peak value (Fig. 5). This property is well illustrated
For 15 of 18 cells, alignment on the optimal GPE was statisti- for cell N27a, for which we were fortunate to have at least 10 mul-
cally better for minimizing s.d. than alignment on the end of the tiple-step sequences in which the last gaze position plateau in
multiple-step sequences (Fig. 6e). For the 3 points that lie close to each trace ended at about the same distance from the target
the 45 diagonal and that are enclosed by the dotted contour, (Fig. 7a). The spike-density histogram shows that during the last
there was no statistical difference between the 2 s.d.s. The pre- plateau, where GPE was about 5, the firing frequency was lower
ferred GPE of these cells was close to zero, the end of the multi- compared to when GPE was 0. A similar, complementary result
ple-step trajectory. For all cells, a result similar to that shown in was found for cell N42a, recorded during multiple-step gaze shifts
Fig. 4a (right) was found when the s.d. at optimal GPE was com- that had gaze plateaus in many different positions relative to final
pared to s.d.s for alignments on different features of a gaze tra- gaze position. There are two segments to the relationship between
Fig. 7. Firing frequency
during gaze plateaus a b c
depends on GPE. (a) Ten
multiple-step gaze shifts
Mean firing frequency in plateau (sp/s)
with amplitude of last sac-

cade about 5. Traces
aligned on beginning of last
saccade. Cell N27a shows
lower discharge when G is
~5 from goal than when it
is at final position where
GPE = 0. (b) For cell N42a,
mean firing frequency dur-
ing plateau (filled circles)
increases as GPE of plateau
decreases. Open circles on
abscissa represent plateau
positions for which cell did
not discharge. The triangle
on the zero of the abscissa
indicates average GPE at
which first spike occurred
(18) for this cell.
Regression line was forced through this point. (c) Population activity in relation to gaze trajectory. For cat N, we compare (bottom) the average
summed (or population) activity (p) of 10 cells with the average trajectory (t) during this summed activity. The average trajectory was obtained from
100 multiple-step gaze trajectories (top), 10 for each cell, each spanning a range of about 2555.

articles
firing frequency and GPE (Fig. 7b): one through the open circles GPE tends to zero? First we emphasize that our cats were fully
where the cell did not discharge, the other through the black cir- alert and were orienting to food targets, and that multiple-step
cles. Clearly, after the preferred GPE was reached (black trian- gaze shifts were interspersed with single-step gaze shifts. We have
gle), the firing frequency during plateaus increased monotonically proposed a gaze position error criterion. Alternatively, could the
as GPE approached zero. Regression lines with significant corre- time of the first spike be related to the velocity characteristics of
lation coefficients were obtained in 10 of 18 cells. In the remain- a trajectory? We believe not. Figure 1d shows that both the gaze
ing cells, it was not possible to calculate a significant regression and head velocity profiles can be very bumpy: the same gaze
line because, first, the multiple-step gaze trajectories had gaze- velocity value can occur at six different times during a gaze shift.
plateau positions that did not cover a sufficiently large range of The same is also true for the head velocity profile. Thus, s.d. can-
GPEs, and, second, the estimate of mean firing frequency dur- not be minimized by referring the first spike to a particular gaze
ing plateaus that are close to the preferred GPE is inherently very or head velocity.
noisy because of low firing frequency and comparatively short It should be noted that in the cat, head and gaze position tra-
plateau duration. jectories are highly correlated19, and we cannot absolutely rule
In our sample, the optimal GPE at which the discharge out the possibility that our SCFNs were reactivated at a specific
resumed varied across cells in all cats from about 4 to 26. The head position error. We believe this is unlikely for two reasons.
average for all cells was 13. For cat N, the optimal GPE of the 10 First, evidence indicates the superior colliculus encodes gaze para-
SCFNs we recorded in this animal varied between about 4 and meters4,6,12,15,20. Second, SCFNs are reactivated at the end of the
25, averaging 11. Given that, for the SCFNs of a particular cat, multiple-step gaze trajectories (Fig. 7c), whereas head trajecto-
there seems to be a continuum of preferred GPEs, we sought to ries terminate later.

obtain an approximate measure of the population response. To In monkey, the error signal might reflect a planned eye sac-
do this, we summed for cat N the discharges of all 10 cells during cade22. Our results rule out a motor interpretation of fixation
10 trajectories for each cell, a total of 100 gaze shifts (Fig. 7c, top), cell discharges, at least until peak discharge occurs at the end of
all made in the dark and having a large range of amplitudes and a multiple-step sequence. For SCFNs, a multiple-step gaze tra-
trajectories. The population discharge, obtained by dividing the jectory could reach the threshold GPE for first spike genera-
summed firing frequency by 100, closely resembles the GPE of tion during a saccade (for example, Fig. 5a, top and bottom
the average gaze trajectory whose amplitude was about 30 traces) or at the start of a long plateau (Fig. 5a, second trace
(Fig. 7c, bottom). This suggests that GPE, below a certain thresh- from top). As a result, the time of the first spike relative to a
old value (about 30 in our sample for cat N), is continuously specific feature of a multiple-step gaze trajectorysuch as, for
encoded in the population discharge of SCFNs. Note also that example, before the last gaze saccadecould be highly vari-
the population activity peaks at a GPE of zero. able and depend on the number of gaze saccades and their
amplitudes and durations, as well as the durations of intersac-
DISCUSSION cadic gaze plateaus. Contrary to concepts derived from stud-
These observations lead to conclusions that are different than ies of saccadic eye movements in head-fixed monkeys1,8 and
those based on observations of SCFN discharges during classical single-step gaze shifts in head-free cats3,4, we found that tonic
single-step eye saccades in head-fixed monkey1,8. In these stud- activity and pauses in activity were not necessarily associated
ies, SCFN pause onset and offset times were linked to temporal with periods of fixation and the occurrence of gaze saccades,
characteristics of the eye trajectory, notably saccade beginning respectively. Recall that at the start of a multiple-step sequence,
and end respectively. Eye saccades have a very stereotyped tra- activity in SCFNs was not a necessary condition for assuring
jectory, and a specific time relative to saccade end corresponds that gaze was immobile because during the initial interstep
closely to a specific GPE. Thus, using head-fixed saccades, it is gaze plateaus, gaze was immobile, and yet SCFNs were totally
virtually impossible to distinguish between these two variables silent (Fig. 3b and f) unless their preferred GPE had been
(Figs. 4b and 6b and d). By comparison, multiple-step gaze shifts reached (Fig. 3f, bottom trace). Furthermore, the gaze saccades
permit a dissociation between position and time because these that occurred when gaze approached the salient target were
trajectories are not stereotyped: target acquisition is accomplished generated in spite of ongoing SCFN activity, and there was no
using a variable sequence of gaze saccades and plateaus, which saccade-related pause (Fig. 7a).
implies that, for a given overall gaze shift amplitude, the rela- For multiple-step gaze shifts larger than the preferred GPE,
tionship between gaze position and time in the trajectory is high- the silence of SCFNs for all the first gaze saccades indicates that
ly variable. reactivation of SCFN activity toward the end of a multiple-step
If SCFNs were all reactivated at strictly the end of a gaze-step sequence is not simply motor activity due to activation of a cells
sequence (that is, at GPE = 0), then their activity would encode movement field10,12 to generate a corrective gaze saccade to the
nothing else but that. However we found that, for each cell, the target. Put another way, if a SCFNs discharge before the final
first spike occurred at its own preferred GPE. Across cells, the saccade in a multiple-step sequence were motor, then the same
preferred GPEs spanned a fairly large range of angles; for cat N, cell should discharge for an equivalent saccade at the beginning
in which we held the largest number of cells, the range was 425; of a sequence. All our data (Figs. 1, 3, 5 and 7) show this is not
mean, 11. This suggests (Fig. 7c) that, in any one cat, the popu- the case if the first saccade begins and ends before the preferred
lation activity may continuously encode the current GPE after a GPE is reached.
threshold is reached (about 30 in cat N). Our results agree with Both SCFNs and omnipause neurons (Fig. 1a) share simi-
observations on monkey that SCFNs report mismatches between lar discharge properties in the head-fixed monkey and head-
eye and target positions22. However, in monkey, the encoded unrestrained cat; they discharge tonically during fixation and
mean error is 1, much smaller than in cat. pause during eye and single-step gaze saccades, respective-
Are there alternative interpretations of what determines the ly1,3,4,7,8,10,11,18,21. Because SCFNs project to OPNs17, it has been
time of onset of the first spike and the closely associated obser- natural to hypothesize that SCFNs have a major influence on
vation that, after the first spike, the firing frequency increases as OPN discharge characteristics and as such provide a higher-

articles
level gate on saccade generation. However, the tonic activity in brainstems superior colliculus1,35,8,9,11. Eye and head movements were
SCFNs can be modulated differently than that in OPNs8. We recorded using a search coil in a magnetic field27. All surgical and exper-
do not yet understand the mechanisms underlying the decou- imental protocols, described elsewhere46, were approved by the Animal
pling between SCFN and OPN discharges. We have shown that Care Committee of the Montreal Neurological Institute and complied
OPNs are active during gaze plateaus in multiple-step gaze with the Canadian Council on Animal Care policy on the use of labora-
tory animals. We restricted our study to horizontal gaze shifts. The ani-
shifts21, but observations of OPNs in head-fixed cat23 suggest mals faced an opaque barrier (barrier protocol4) of variable width
some OPNs, like our SCFNs, may also pause during head-free (1060) directly in front of them at a distance of about 35 cm. The head
multiple-step gaze shifts. Thus OPN discharges during multi- of each cat was attached, via two universal joints, to a vertical shaft rotat-
ple-steps gaze shifts in the head-unrestrained animal need to ing in low-friction bearings. This minimized constraint on the animals
be studied further. orienting behavior. Two cells in cat M were recorded with the animals
During a multiple-step gaze shift, the tonic activity or pause head not attached to the vertical shaft and therefore totally unrestrained.
in activity of SCFNs does not indicate whether gaze is in the This did not alter cell discharge characteristics.
saccade or fixation modes. However, peak discharge occurred A small food target was hidden behind the barrier and suddenly appeared
consistently at the end of a multiple-step sequence. This indi- on one of the two vertical sides. The hungry cats naturally redirected their
visual axis to the target and were fed. The amplitude of the overall gaze shift
cates that the SCs fixation zone knows that the remembered
required to fixate the target depended on the initial gaze position, which
target location has been reached. We suggest that SCFN dis- varied from trial to trial. With practice, the timing of a cats orienting behav-
charges may not affect downstream motor circuits until the ior became very reliable. A fluorescent light with very fast decay time
time of peak discharge when a multiple-step gaze shift is over.
(~100 ms) provided ambient lighting. Some trials were performed in the
Perhaps it is only thenat overall gaze shift endthat SCFN light; in others, the light was extinguished for 1 s beginning 120 ms after
discharges are truly motor. The barrage of impulses from target appearance, such that gaze shifts were made in complete darkness.
SCFNs onto OPNs (Fig. 1a) at the end of a gaze-step sequence Before recording neurons, we deduced the superior colliculus motor
may suppress further movements away from the target when maps organization based on the amplitude and direction of the gaze shifts
it is reached. Consideration of the short efferent delay4 (about evoked at different electrode positions6,20. The location of the rostral SCs
10 ms) between cell activity and its effect on movement char- fixation zone corresponds to the location where stimulation does not
evoke gaze shifts but rather stops ongoing ones. A histological recon-
acteristics does not alter this conclusion. struction of the recording site was made in two of the cats. We also placed
We originally proposed for the cats collicular gaze motor stimulating electrodes in the predorsal bundle of all cats to determine by
map3,4,6,15 that, first, during a single-step gaze shift, feedback antidromic stimulation whether an SCFN was a tectoreticular neuron.
to the superior colliculus displaces neural activity tangen- When head-free cats orient to a target, they do not necessarily utilize
tially along the map toward the fixation zone and away from multiple steps. To get a sufficient number of these, it is generally necessary
the initial site of activation, which encodes the desired gaze to record an SCFN for more than two hours. This is very difficult to do in
displacement vector. Second, when activity reaches and the head-unrestrained condition because head movements cause small
invades the fixation zone, the desired gaze displacement displacements of the brainstem relative to the microelectrode tip, and con-
stops. Third, activity on the motor map suppresses activity tact with the cell is easily lost. Nevertheless, we recorded in 5 cats a total
of 18 cells in the rostral superior colliculus that we classified as SCFNs
in the fixation zone. To extend these mechanisms to multi-
according to criteria that conform to prior descriptions of SCFNs in cat
ple-step gaze shifts, we suggest first a bell-shaped hill of and monkey (Fig. 1bd)1,3,4,7,8,10,11,15. Ten, three, three, one and one cell(s)
activity4 that begins centered at a location on the SCs motor were recorded in cats N, M, P, H and T, respectively. Six of these cells were
map, appropriate for the overall amplitude of the multiple- identified as tectoreticular neurons. The lack of identification of the remain-
step gaze shift; second, that this hill is large and spans per- ing SCFNs does not imply that they were not tectoreticular neurons; with
haps a third of the SCs length 4 ; and third that this hill is time, the reliability of the stimulating electrodes waned. For each SCFN
displaced progressively toward the fixation zone with each that we analyzed, we had between 6 and 61 (average about 20) multiple-
step in the sequence. This mechanism, in combination with step gaze shifts from which we could reliably assess a cells discharge pattern.
a different threshold value for each SCFNs reactivation when Monkey SCFNs have omnidirectional pauses in their tonic activity1, where-
the leading edge of the hill arrives, would explain each as, for SCFNs in cat, the pause may be more pronounced for gaze shifts
directed either ipsilaterally or contralaterally, depending on the cell4. In
SCFNs preferred GPE and the plot of population activity our sample, 14 of 18 cells paused preferentially for contralaterally directed
versus gaze trajectory (Fig. 7c), which suggests that the pop- gaze shifts, 4 of 18 cells for ipsilaterally directed gaze shifts. In the non-
ulation discharge of SCFNs encodes GPE. Whatever the optimal direction, a pause was not discernable on spike-density histograms,
mechanism, the present results support the hypothesis that and consequently identification of first spike was not possible.
there is feedback to the cats SC, at least to the fixation zone,
and that this feedback generates a GPE signal. Data analysis. Gaze (G) = eye position in space = eye position in head
The distinction between end-goals and intermediate-goals in (E) + head position in space (H). The eye and head coils gave G and H.
motor behavior generally, and scan paths in particular2426, are We recorded also target position and light onset and offset. Data were
of interest. We do not know what determines whether a gaze shift stored on DAT tape. In off-line analysis, E was calculated from G and
H. Action potentials were converted to logic pulses via a time-ampli-
to a target will be either a single-step or a multiple-step sequence.
tude window discriminator. Movement and pulse traces were digitized
However, irrespective of the number of waypointsor gaze fix- at 2 kHz with data acquisition software and analyzed with Matlab (Math
ationsin the trajectory to a final goal, our results indicate that Works, Natick, Massachusetts). Spike-density histograms were gener-
the fixation zone of the superior colliculus keeps track of the final ated by substituting for each spike a Gaussian function with a width of
intended gaze destination: as a population, SCFNs encode how 10 ms28,29 and then summing all of the Gaussians together to generate a
close gaze is to the end goal, not the details of the trajectory that continuous function in time.
achieves this goal.
ACKNOWLEDGEMENTS
METHODS Funded by the Medical Research Council of Canada. We thank J. Murphy for
Five cats (H, N, M, P, T) were prepared for chronic single-unit record- developing some of the electrode technology and W.Y. Choi, D. Crawford,
ings of fixation neurons, which are located in the rostral extent of the K. Cullen and T. Herter for reading earlier versions of the manuscript.

articles
RECEIVED 8 JUNE; ACCEPTED 18 JULY 2000 14. Grantyn, A. & Grantyn, R. Axonal patterns and sites of termination of cat
superior colliculus neurons projecting in the tecto-bulbo-spinal tract. Exp.
Brain Res. 46, 243256 (1982).
1. Munoz, D. P. & Wurtz, R. H. Fixation cells in monkey superior colliculus. 15. Munoz, D. P. & Guitton, D. Control of orienting gaze shifts by the
I. Characteristics of cell discharge. J. Neurophysiol. 70, 559575 (1993). tectoreticulospinal system in the head-free cat. II. Sustained discharges
2. Hanes, D. P., Patterson, W. F. & Schall J. D. Role of frontal eye fields in during motor preparation and fixation. J. Neurophysiol. 66, 16241641
countermanding saccades: visual, movement, and fixation activity. (1991).
J. Neurophysiol. 79, 817834 (1998). 16. Munoz, D. P. & Istvan, P. J. Lateral inhibitory interactions in the intermediate
3. Munoz, D. P. & Guitton, D. Fixation and orientation control by the tecto- layers of the monkey superior colliculus. J. Neurophysiol. 79, 11931209
reticulo-spinal system in the cat whose head is unrestrained. Rev. Neurol. (1998).
(Paris) 145, 567579 (1989). 17. Bttner-Ennever, J. A., Horn, A. K. E., Henn, V. & Cohen, B. Projections from
4. Munoz, D. P., Guitton, D. & Pelisson, D. Control of orienting gaze shifts by the superior colliculus motor map to omnipause neurons. J. Comp. Neurol.
the tectoreticulospinal system in the head-free cat. III. Spatiotemporal 413, 5567 (1999).
characteristics of phasic motor discharges. J. Neurophysiol. 66, 16421666 18. Evinger, C., Kaneko, C. R. & Fuchs, A. F. Activity of omnipause neurons in
(1991). alert cats during saccadic eye movements and visual stimuli. J. Neurophysiol.
5. Par, M. & Guitton, D. The fixation area of the cat superior colliculus: effects 47, 827844 (1982).
of electrical stimulation and direct connection with brainstem omnipause 19. Guitton, D., Douglas, R. M. & Volle, M. Coordinated eye head movements in
neurons. Exp. Brain Res. 101, 109122 (1994). the cat. J. Neurophysiol. 52, 10301050 (1984).
6. Par, M., Crommelinck, M. & Guitton, D. Gaze shifts evoked by stimulation 20. Roucoux, A., Guitton, D., & Crommelinck, M. Stimulation of the superior
of the superior colliculus in the head-free cat conform to the motor map but colliculus in the alert cat. II. Eye and head movements evoked when the head
also depend on stimulus strength and fixation activity. Exp. Brain Res. 101, is unrestrained. Exp. Brain Res. 39, 7585 (1980).
123139 (1994). 21. Par, M. & Guitton, D. Brain stem omnipause neurons and the control of
7. Peck, C. K. & Baro, J. A. Discharge patterns of neurons in the rostral superior combined eyehead gaze saccades in the alert cat. J. Neurophysiol. 79,
colliculus of cat: activity related to fixation of visual and auditory targets. Exp. 30603076 (1998).
Brain Res. 113, 291302 (1997). 22. Krauzlis, R. J., Basso, M. A., & Wurtz, R. H. Shared motor error for multiple
8. Everling, S., Par, M., Dorris, M. C. & Munoz, D. P. Comparison of the eye movements. Science 276, 16931695 (1997).
discharge characteristics of brain stem omnipause neurons and superior 23. Petit, J., Klam, F., Grantyn, A. & Berthoz, A. Saccades and multisaccadic gaze
colliculus fixation neurons in monkey: implications for control of fixation shifts are gated by different pontine omnipause neurons in head-fixed cats.
and saccade behavior. J. Neurophysiol. 79, 511528 (1998). Exp. Brain Res. 125, 287301 (1999).
9. Munoz, D. P. & Wurtz, R. H. Fixation cells in monkey superior colliculus. 24. Jeannerod, M., Gerin, P. & Pernier, J. Dplacement et fixations du regard dans
II. Reversible activation and deactivation. J. Neurophysiol. 70, 576589 (1993). lexploration libre dune scne visuelle. Vision Res. 8, 8197 (1968).
10. Munoz, D. P. & Wurtz, R. H. Saccade-related activity in monkey superior 25. Yarbus, A.L. Eye Movement and Vision (Plenum, New York, 1967).
colliculus. II. Spread of activity during saccades. J. Neurophysiol. 73, 26. Zusne, L. & Michels, K. M. Nonrepresentational shapes and eye movements.
23342348 (1995). Perceptual Motor Skills 18, 1120 (1964).
11. Gandhi, N. J. & Keller, E. L. Spatial distribution and discharge characteristics 27. Robinson, D. A. A method of measuring eye movements using a scleral
of superior colliculus neurons antidromically activated from the omnipause search coil in a magnetic field. IEEE Trans. Biomed. Eng. 10, 137145 (1963).
region in monkey. J. Neurophysiol. 78, 22212225 (1997). 28. MacPherson, J. M. & Aldridge, J. W. A quantitative method of computer
12. Freedman, E. G. & Sparks, D. L. Activity of cells in the deeper layers of the analysis of spike train data collected from behaving animals. Brain Res. 175,
superior colliculus of the rhesus monkeyevidence for a gaze displacement 183187 (1979).
command. J. Neurophysiol. 78, 16691690 (1997). 29. Richmond, B. J., Optican, L. M., Podell, M. & Spitzer, H. Temporal encoding
13. Grantyn, A. & Berthoz, A. Burst activity of identified tecto-reticulo-spinal of two-dimensional patterns by single units in primate inferior temporal
neurons in the alert cat. Exp. Brain Res. 57, 417421 (1985). cortex. I. Response characteristics. J. Neurophysiol. 57, 132146 (1987).

articles
Activity in primary visual cortex

predicts performance in a visual
detection task
David Ress, Benjamin T. Backus and David J. Heeger
Stanford University, Dept. of Psychology, 450 Serra Mall, Bldg. 420/400, Stanford, California 94305-2130, USA
Correspondence should be addressed to D.J.H. (heeger@stanford.edu)
Visual attention can affect both neural activity and behavior in humans. To quantify possible links
between the two, we measured activity in early visual cortex (V1, V2 and V3) during a challenging
pattern-detection task. Activity was dominated by a large response that was independent of the
presence or absence of the stimulus pattern. The measured activity quantitatively predicted the sub-
jects pattern-detection performance: when activity was greater, the subject was more likely to
correctly discern the presence or absence of the pattern. This stimulus-independent activity had sev-
eral characteristics of visual attention, suggesting that attentional mechanisms modulate activity in
early visual cortex, and that this attention-related activity strongly influences performance.
Our ability to perform visual discrimination tasks is improved the fixation mark; on the other trials, no pattern was presented,
when we are instructed in advance when and where to attend to and the display remained gray. Subjects pressed one of two but-
a visual stimulus13. Neuronal activity in several areas of visual tons to indicate whether they believed the pattern was present.
cortex, including primary visual cortex (V1), can be modified by Before commencing fMRI scanning sessions, subjects practiced
attention420. For example, in macaque monkeys, stimulus-evoked the task extensively until their performance stabilized; this
electrophysiological responses in extrastriate areas are enhanced involved at least a dozen half-hour sessions over a three-week
by attention4,1114, and similar results are observed in human sub- period. The contrasts used during the fMRI experiments were
jects with neuroimaging610. These neuronal correlates of atten- individually chosen so that each subject would perform with an
tion are spatially selective710 and depend on task difficulty20. accuracy of ~75% correct (d ~1). The fMRI data were collect-
Most notably, baseline firing rates in monkey extrastriate visual ed during several hundred trials for each of three subjects. Tri-
cortex are elevated by attention even without visual stimulation13, als were categorized according to whether the stimulus pattern
and analogous phenomena are observed in humans5,6,10. It has was present or absent, and whether the subject responded cor-
not yet been determined, however, if the attention-related increase rectly or incorrectly. The fMRI data were analyzed separately for
in baseline activity reported in these studies is actually relevant each of these four trial types.
to behavior. Cortical activity in V1 was dominated by a large, stimulus-
The main goal of the experiments reported here was to estab- independent base response (Fig. 1c). This response was very sim-
lish a relationship between V1 activity and behavioral perfor- ilar for pattern-present (filled symbols) and pattern-absent (open
mance. Event-related functional magnetic resonance imaging symbols) trials. The base response was readily evident and high-
(fMRI)2124 was used to measure cortical activity while subjects ly significant (p 0) in all three subjects (Z-scores, DJH, 41.3;
performed a challenging pattern-detection task at contrast thresh- BTB, 16.1; DBR, 38.6). Based on the current understanding of
old. We observed a large, stimulus-independent response in V1, fMRI signals2527, we infer that the underlying neuronal activity
V2 and V3, which we call the base response. We infer that this associated with the base response was largely confined to the first
base response corresponds to an increase in the baseline firing one or two seconds of each trial, and that the slow time course
rates of a large number of neurons. Remarkably, the trial-to-trial of the fMRI measurements reflects a sluggish change in local cere-
variability in the base response in all these areas predicted behav- bral blood flow and oxygenation. We hypothesize that the base
ioral performance on the task; performance was best (more cor- response reflects an attention-related increase in the baseline fir-
rect judgments, largest d) when the base response was large and ing rates of a large number of visual cortex neurons.
worst when the base response was small. Under the hypothesis that the base response reflects atten-
tional mechanisms, one would expect it to be correlated with
RESULTS performance in the threshold detection task; performance should
Subjects viewed a uniform gray field and continuously fixated a be best when the attention-related base response is large and
small, high-contrast mark at its center (Fig. 1a and b) while lying worst when the base response is small. Indeed, the base response
in the bore of the magnetic resonance scanner. Once every 20 was highly predictive of behavioral performance (Fig. 2). The
seconds, a short auditory tone cued subjects that a new trial was linear regression showed a statistically significant positive slope
beginning. On half the trials (randomly interleaved), a low-con- (r = 0.90; p = 0.005). This result was robust with respect to the
trast pattern was presented briefly in a peripheral annulus around binning of the data (Methods); the regression was statistically

articles
Fig. 1. Experimental protocol and typical result. (a) Subjects viewed a

uniform gray field (27.3 cd/m2) while fixating a small, high-contrast
a
mark. On each trial, either the display remained blank or a plaid pattern
was presented briefly. The pattern, when present, was the sum of two
sinusoidal gratings (0.5 cycles/degree), contrast-reversing (4 Hz), and
confined to an annulus (36 radius). In the threshold detection experi-
ments, the pattern was presented at contrasts of 0.80.9%, much lower
than illustrated. (b) Trials began with an auditory warning tone. An audi-
tory click at 1 s marked the onset of the stimulus period, which had a
duration of 0.75 s. Another click (0.1 s later) signaled subjects to
respond yes or no by pressing one of two buttons. Subjects then
waited quietly for the next trial while maintaining fixation. No (cor-
rect/incorrect) feedback was provided to the subject. (c) V1 responses
were large both for pattern-present (filled symbols) and pattern-absent
(open symbols) trials. Each curve represents the average time course of
the fMRI signal, averaged across many trials (163 pattern-present trials b
and 133 pattern-absent trials) and averaged throughout the region of
cortical gray matter corresponding to the V1 representation of the stim-
ulus annulus. Error bar indicates standard error at one time; other time
points have similar standard errors.
significant (p < 0.05) for 525 bins. This result does not derive c Typical V1 time series
from individual differences between the subjects; the range of d
values and the range of relative response amplitudes were similar
in all three subjects.
The contingency between V1 activity and behavioral perfor-
mance can be tested more rigorously by a logistic regression. We
fit a logistic function to the subjects performance on each trial as
a function of the relative fMRI response amplitude on that trial
(Methods). The slope of the best-fitting logistic function measures
the strength of the predictive effect. For all subjects combined, the
logistic regression was statistically significant, that is, the best-fit
slope was reliably greater than zero (p = 0.002). The logistic regres-
sion was also statistically significant in two of three individual sub-
jects (DJH, p = 0.026; BTB, p = 0.060; DBR, p = 0.013).
A third way to examine the relationship between V1 activity
and behavior is to compare the base responses corresponding to eral of the retinotopically organized visual areas, it was essen-
correct and incorrect trials. For all subjects combined, the relative tially absent in a lateral region of the occipital lobe that is adjacent
response amplitudes were larger for correct trials than for incor- to these visual areas (data not shown). Hence, the base response
rect trials (p = 0.002). This difference was also statistically sig- was not an artifact of increased respiration or pulse during task
nificant in two of three individual subjects (DJH, p = 0.014; BTB, performance, nor was it due to nonspecific arousal.
p = 0.052; DBR, p = 0.018). The base response depended on task difficulty, with activity
A stimulus-independent base response was also evident in decreasing as the task was made easier. In a separate series of
extrastriate visual areas V2 and V3 (Fig. 3a and b). The respons- experiments, we measured V1 activity while subjects performed
es in these areas were similar in shape and amplitude to those comparatively easier tasks, detecting patterns with higher con-
observed in area V1 (compare with Fig. 1c). We used the slope trasts (Fig. 6). Subjects were instructed to perform the task as
parameter of the logistic regression to quantitatively compare the accurately as possible, but without exerting attention unneces-
three visual areas using data combined across all subjects
(Fig. 3c). The base responses were predictive of behavioral per-
formance in both extrastriate areas (V2, p = 0.007; V3, p < 0.001).
The downward trend from V1 through V3 was not statistically
significant. Individual subjects also showed significant effects in
extrastriate areas (DJH, V2; BTB, V3; DBR, V2, V3).
Our interpretation of the contingency between cortical activ-
ity and behavioral performance relies on the hypothesis that the
base response reflects attentional mechanisms. Indeed, the base
response exhibited two critical characteristics of visual attention:
it was spatially selective, and it depended on task difficulty. The
base response was smaller in the subregions of V1 correspond-
ing to the central visual field and the far periphery than in sub- Fig. 2. Trial-to-trial variability in V1 activity was highly predictive of
regions corresponding to the stimulus annulus (Fig. 4). A similar behavioral performance in the threshold detection task. Performance,
but weaker pattern of spatial selectivity was also evident in V2 as measured by d, is highly correlated with the amplitude of V1 activity
and V3 (Fig. 5). Although the base response was evident in sev- (r = 0.92; p < 0.001). Solid line is the linear regression.

articles
a lus-evoked activity in the minority of neurons that are selective

c for the pattern. When the task is difficult and the sensory input
small, measured cortical activity is dominated by small respons-
es in each of a large number of neurons. When the task is easy
and the sensory input large, measured cortical activity is domi-
nated by large responses in each of a small number of neurons.
Activity in areas V1, V2 and V3 were all observed to predict
behavioral performance, suggesting that these visual areas perform
b computations that are relevant to the contrast-detection task. We
cannot distinguish whether the extrastriate base response was inher-
ited from their V1 afferents, or if the V1 base response was caused
by feedback from extrastriate cortex, or perhaps both.
Our main result, the contingency between attention-related
cortical activity and behavioral performance, has been hinted at
previously. Parietal lobe activity is correlated with performance
(activity on correct-response trials greater than activity on incor-
rect-response trials) during a motion-detection task32, although
Fig. 3. Activity in extrastriate areas V2 and V3 predicted performance. this result was not analyzed for statistical significance. It is pos-
(a, b) Typical time series (subject DBR) of activity in areas V2 and V3. sible that the attention-related signals that we have observed in
Standard errors are similar to the size of the plot symbols. The amplitudes
early visual cortex derive from attentional control mechanisms
of the fMRI responses and the standard errors were similar to those in V1
(compare with Fig. 1c). (c) Slope of the best-fitting logistic function for in the parietal lobe, but our slice prescription did not allow us to
each of the three visual areas. All three visual areas showed a reliable con- examine such activity.
tingency between cortical activity and behavioral performance. Eye movements pose a potential confound for interpretation
of our results, but it is likely that subjects maintained steady fix-
ation while performing the detection task. If subjects had moved
their eyes, the shift of the fixation mark would have evoked activ-
sarily. At higher contrasts, performance was 100% correct, and ity in the foveal representation of V1; we see relatively little activ-
the base responses were small. The absolute amplitudes of the ity in this region (Figs. 4 and 5). Second, in control experiments,
pattern-absent V1 responses were considerably smaller at 50% performance was not improved by allowing subjects to break fix-
contrast (Fig. 6e) than at contrast threshold (Fig. 6a); this dif- ation; best performance was obtained when the eyes were held
ference in V1 activity was statistically significant for each of the steady on the fixation mark.
three subjects individually (p < 0.001, ~5:1 geometric mean ratio The interpretation of fMRI data depends on the sequence
across the three subjects). A similar dependence on task difficul- of events from the neuronal response to the fMRI signal, which
ty was also observed for the base responses in areas V2 and V3 is only partially understood25,26. For example, the fMRI signal
(Fig. 7). The stimulus-evoked increment in the pattern-present might reflect not only neuronal firing rates but also subthresh-
responses increased with stimulus contrast (Figs. 6 and 7), con- old synaptic activity (due to simultaneous excitation and inhi-
sistent with previous measurements of activity in human and bition), which would be invisible to the extracellular electrode.
monkey visual cortex2731. However, the available data suggest that the fMRI signal is
DISCUSSION a
Our results lead us to hypothesize that attention, cortical activi-
ty and behavioral performance are linked. The base response
probably reflects the activity of attentional mechanisms. This
hypothesis is supported by the dependence on task difficulty and
by the spatial selectivity of the base response, two defining char-
acteristics of visual attention. In macaque monkeys, attention
can cause an increase in the baseline firing rates of neurons in
V2 and V4 (ref. 13), secondary visual areas that receive projec-
tions from V1. We suggest, therefore, that the base response that
we observed reflects a small increase in the baseline firing rates b
of a very large number of neurons whose receptive fields overlap
the stimulus annulus. Increasing baseline firing rates may poten-
tiate an enhancement in the stimulus-evoked responses, which
could improve the signal-to-noise ratio in the sensory represen-
tation of the stimulus by biasing neurons into a more sensitive
segment of their operating range12. We attribute the contingency
between V1 activity and performance to trial-to-trial fluctua-
tions in attention (for example, lapses in attention or spatial
uncertainty). We suggest that variability in the observers atten-
tional state causes variability in the baseline firing rates, which, in Fig. 4. V1 activity was spatially selective. A larger response was observed
turn, causes variability in performance. in regions corresponding to the stimulus annulus than in regions corre-
The pattern-present response increments evident at higher sponding to the fovea or the periphery. (a) Sample fMRI time series (sub-
contrasts (Fig. 6), on the other hand, probably reflect the stimu- ject DJH). (b) Relative fMRI response amplitudes for all subjects.

articles
ly significant results. Therefore, these effects may be better

a revealed by fMRI measurements, which reflect the activity of
a large population of neurons. Second, as mentioned above,
the fMRI measurement may reflect subthreshold activity that
would be invisible to extracellular electrodes. Third, our data
suggest that reliable and robust increases in baseline firing rates
may be evident only when performing a particularly demand-
ing task. The studies that failed to find baseline increases may
have been using tasks that did not place sufficiently high
demands on spatial attention. Fourth, the monkeys in those
b experiments may have been so highly trained that the usual
attentional mechanisms were no longer needed to perform the
task; training can have a critical effect on attentional signals38.
Fifth, small shifts in eye position, equal in size to the V1 recep-
tive fields, present a difficulty for the electrophysiology exper-
iments. If eye position is systematically correlated with shifts
in spatial attention, then the responses of individual V1 neu-
rons will modulate as the receptive fields are shifted toward
and away from the stimulus. These biases can be avoided by

carefully accounting for eye position14, but perhaps at the cost
Fig. 5. Extrastriate activity was also spatially selective. (a) V2; (b) V3. of underestimating the magnitude of the attentional effects.
Small shifts in eye position do not present a difficulty in the
fMRI experiments because they have a negligible effect on mea-
surements of pooled neuronal activity. Sixth, there may be a
roughly proportional to average firing rates3337, even under genuine species difference.
conditions that seem to involve synaptic inhibition33. Moreover, We conclude that during threshold pattern detection, when
the strong correlation that we observed between our brain and sensory input is small and attentional demands are large, activi-
behavior measurements is striking: the fMRI signal is measur- ty in early visual cortex is dominated by a stimulus-independent
ing what seems to be behaviorally relevant cortical activity. response that seems to be related to visual attention. This activ-
Many electrophysiological studies report that baseline fir- ity is strongly correlated with behavioral performance. Our results
ing rates do not increase with attention (for example, ref. 14). suggest that attention-related activity in primary visual cortex
Moreover, the studies that do find baseline increases in extras- affects performance accuracy during a visual task.
triate cortex fail to find them in V1 (for exam-
ple, ref 13). There are at least six possible a b
explanations for the discrepancy between those
results and our results (along with ref. 10),
which do indicate baseline increases in V1.
First, because cortical neurons generally have
low baseline firing rates, the responses of many
neurons must be recorded to obtain statistical-
Fig. 6. V1 activity depended on task difficulty. c d

(a) Typical fMRI time series for threshold-contrast
stimuli (same as Fig. 1c, subject DBR). (b) Relative
fMRI response amplitudes for threshold-contrast
stimuli. Error bars represent 1 standard error (subject
DBR, 0.85% stimulus contrast, n = 656 trials; DJH,
0.9% contrast, n = 456; BTB, 0.85% contrast, n = 400).
(c) Example fMRI time series (subject DBR) for
roughly twice the threshold stimulus contrast, so that
performance was 100% correct. The base response
was smaller than that measured at threshold (com-
pare open symbols with those in a). (d) For contrasts e f
roughly twice threshold, the pattern-present
responses were larger than the pattern-absent
responses, although this was statistically significant in
only one subject (DBR, stimulus contrast, 2%, n = 36
trials, p < 0.01; DJH, 1.5% contrast, n = 96, p = 0.07;
BTB, 2% contrast, n = 36, p = 0.23; one-tailed t-tests).
(e) Example fMRI time series (subject DBR) for 50%
stimulus contrast. (f) At high contrast, the base
responses were small for all three subjects, and the
pattern-present responses were much larger than the
pattern-absent responses (p < 0.001 for each subject).

articles
METHODS
Magnetic resonance (MR) imaging was done on a standard clinical GE 1.5
Tesla Signa scanner with a custom designed dual surface coil. The exper-
iments were undertaken with the written consent of each subject, and in
compliance with the safety guidelines for MR research. Each subject par-
ticipated in several MR scanning sessions: one to obtain a high-resolu-
tion anatomical volume; one to functionally define the early, retinotopic
visual areas, including V1, V2 and V3; one to functionally define the sub-
regions of these visual areas that correspond to the stimulus annulus,
and several sessions (9 for DBR, 6 for DJH and 6 for BTB) to measure
fMRI responses in the various experimental conditions. For each sub-
ject, at least four sessions were specifically devoted to the threshold-detec-
tion task, whereas two sessions were devoted to examining the effects of
higher-contrast stimuli. In these latter sessions, four-minute scanning
blocks of higher-contrast stimuli were interleaved with blocks of thresh-
old-contrast stimuli.
Stimuli were presented on a flat-panel display (NEC, multisynch LCD
2000) placed within a Faraday box with a conducting glass front, posi-
tioned near the subjects feet. Subjects lay on their backs in the bore of
the MR scanner and viewed the display through binoculars with a pair
of angled mirrors attached just beyond the two objective lenses. Each Fig. 7. Extrastriate activity also depended on task difficulty.
subject had normal or corrected-to-normal vision. (a, b) V2; (c, d) V3.
Each MR scanning session began with the acquisition of a set of
anatomical images using a T1-weighted fast spin-echo pulse sequence in
the same slices as the functional images (TR = 500 ms, TE = 15 ms, echo- ry, gradient-echo pulse sequence44,45. For our particular scanner hard-
train length of 2, FOV = 240 mm, 4-mm slice thickness). The eight slices ware, spiral fMRI pulse sequences compare favorably with echo-planar
were arranged obliquely, perpendicular to the calcarine sulcus, with the imaging in terms of sensitivity, and spatial and temporal sampling res-
most caudal slice approximately tangential to the occipital pole olution (A.M. Sawyer-Glover & G.H. Glover, SMRT Seventh Annual
These in-plane anatomical images were aligned to a high-resolution Meeting 69, Sydney, 1998). Pulse sequence parameters for the fMRI
anatomical volume (acquired using a three-dimensional SPGR pulse scans were TE = 40 ms, TR = 500 (or 750) ms, FA = 55 (or 65),
sequence and a head coil) of each subjects brain so that all MR images FOV = 240 mm, effective in-plane pixel size of 3.2 3.2 mm, slice
(across multiple scanning sessions) from a given subject were coregis- thickness of 4 mm. These pulse sequence parameters were chosen to
tered. The alignment was done using an intensity-based algorithm that optimize signal-to-noise ratio: the minimum number of slices to cover
was specifically designed to deal with the variable contrast and intensity the desired brain regions, fast TR to acquire more temporal frames,
profiles produced by the different pulse sequences and different coils, to and the largest voxel size that allowed for accurate and reliable local-
an accuracy of ~1 mm (ref. 39). The alignment algorithm computed a ization of the retinotopic ROIs.
rigid-body coordinate transformation that maps the coordinates of each A bite bar stabilized the subjects heads. The time series of fMRI images
in-plane voxel to the corresponding coordinates in the high-resolution from each scan were visually inspected for head movements; 5 of approx-
anatomical volume. Given this coordinate transformation, we used a imately 260 total scans showed evidence of head movements and were
standard image resampling technique (backward mapping with nearest- excluded from further analysis.
neighbor, point sampling) to transform the functional imaging data from Data from the first six seconds of each fMRI scan (before the first pat-
the in-plane images to the volume. The end result was a time-series of tern-detection trial) were discarded to minimize transient effects of mag-
data for each gray matter voxel from each fMRI scan. netic saturation. The fMRI time series were preprocessed by high-pass
The fMRI data were analyzed in each of three regions of interest filtering them at each voxel to compensate for the slow signal drift we
(ROIs) corresponding to the V1, V2 and V3 representations of the stim- typically observe in our fMRI signals46, and by dividing each voxels time
ulus annulus. These ROIs were defined, separately for each subject, in series by its mean intensity to convert the data from arbitrary image
two steps. First, the retinotopically organized visual areas were identi- intensity units to percent signal modulation and to compensate for the
fied, following well-established methods4043, by measuring the polar decrease in mean image intensity with distance from the surface coil.
angle component of the cortical retinotopic map. Second, subregions of The resulting time series were averaged throughout the region of cortical
these visual areas that corresponded to the cortical representation of gray matter corresponding to the V1 representation of the stimulus annu-
the stimulus annulus were further delimited based on a separate series lus (and likewise for V2 and V3).
of fMRI measurements. During these scans, subjects held fixation while Relative fMRI response amplitudes were computed from these spa-
the display alternated every 20 s between a contrast-reversing, high- tially averaged time series as follows. Regard each trials time series as a
contrast, plaid pattern within the 36 radius annulus, surrounded by vector, Ri, where i is the trial index. All N trials of a particular fMRI scan-
a uniform (luminance-matched) gray field, and its geometric comple- ning session were averaged together, as in Fig. 1c, but regardless of trial
ment, a flickering plaid pattern everywhere except the annulus. Data type, to create a session-mean time series,
were averaged across 68 repeated scans, each with 6 cycles of alterna- N
R .
tion between the two stimulus conditions. The final ROIs were then R = 1 i+1 i
defined as those subregions of V1, V2 and V3 that were strongly cor- N
related with the stimulus alternations (V1, r > 0.7; V2, r > 0.5; V3, Next, we computed a normalized relative amplitude for each trial, Ai by
r > 0.4; all regions, 010 s time lag). projecting (inner product) the time series from each individual trial onto
Although we also defined additional retinotopic visual areas (V4v and the session-mean time series and dividing by the squared norm of the
V3a), the data were analyzed only in V1, V2 and V3 because V4v was too session-mean time series:
far anterior to be fully included in the slice prescription for all subjects,
RiR
and noise levels were too high to reliably resolve the V3a representation Ai = 2.
of the stimulus annulus. R
During each pattern-detection fMRI scan, subjects performed 12 These calculations were done separately for each scanning session because
consecutive trials (20- or 21-s duration) while a time series of volumes of substantial session-to-session variations in the hemodynamic
were acquired (every 1 or 1.5 s) using a T2*-sensitive, spiral-trajecto- response47. Thus, we forced the average amplitude for all trials within a

articles
session to be unity. Finally, we sorted the trials by type (pattern-present, 14. McAdams, C. J. & Maunsell, J. H. Effects of attention on the reliability of
pattern-absent) and computed the mean and standard error of the rel- individual neurons in monkey visual cortex. Neuron 23, 765773 (1999).
ative amplitudes across multiple scanning sessions. 15. Colby, C. L., Duhamel, J. R. & Goldberg, M. E. Visual, presaccadic, and
cognitive activation of single neurons in monkey lateral intraparietal area.
Three schemes were used to test the contingency between behavioral J. Neurophysiol. 76, 28412852 (1996).
performance and cortical activity. First, the relative fMRI response ampli- 16. Motter, B. C. Focal attention produces spatially selective processing in visual
tudes, which fluctuated greatly from trial to trial, were sorted into ascend- cortical areas V1, V2, and V4 in the presence of competing stimuli.
ing order, then divided into bins, each containing the same number of J. Neurophysiol. 70, 909919 (1993).
17. Roelfsema, P. R., Lamme, V. A. & Spekreijse, H. Object-based attention in the
trials. Each trial within a bin is associated both with a relative fMRI primary visual cortex of the macaque monkey. Nature 395, 376381 (1998).
response amplitude and with the task-related behavioral data (that is, 18. Haenny, P. E. & Schiller, P. H. State dependent activity in monkey visual
whether the stimulus pattern was absent or present, and whether the sub- cortex. I. Single cell activity in V1 and V4 on visual tasks. Exp. Brain Res. 69,
jects button press was correct or incorrect). For each bin, we computed 225244 (1988).
the average fMRI amplitude, and we computed d, a measure of the 19. Haenny, P. E., Maunsell, J. H. & Schiller, P. H. State dependent activity in
monkey visual cortex. II. Retinal and extraretinal factors in V4. Exp. Brain
behavioral performance calculated from the percentage of hits, misses, Res. 69, 245259 (1988).
false alarms and correct rejections48. In the second scheme, a logistic 20. Spitzer, H., Desimone, R. & Moran, J. Increased attention enhances both
function was fit to the binary (correct/incorrect) performance data for behavioral and neuronal performance. Science 240, 338340 (1988).
all individual trials, as a function of the relative fMRI response ampli- 21. Ogawa, S., Lee, T. M., Kay, A. R. & Tank, D. W. Brain magnetic resonance
tude. The form of the function was imaging with contrast dependent on blood oxygenation. Proc. Natl. Acad. Sci.
USA 87, 98689872 (1990).
0.5 + exp(a + bA) 22. Kwong, K. K. et al. Dynamic magnetic resonance imaging of human brain
P(A) = activity during primary sensory stimulation. Proc. Natl. Acad. Sci. USA 89,
1 + exp(a + bA) 56755679 (1992).
23. Ogawa, S. et al. Intrinsic signal changes accompanying sensory stimulation:

functional brain mapping with magnetic resonance imaging. Proc. Natl.
where A is the relative fMRI response amplitude for a given trial, P(A) Acad. Sci. USA 89, 59515955 (1992).
is the predicted probability that the subject was correct on that trial, and 24. Belliveau, J. W. et al. Functional mapping of the human visual cortex by
a and b are the parameters of the fit. The parameters a and b were adjust- magnetic resonance imaging. Science 254, 716719 (1991).
ed by an optimization routine to match the measured probability of sub- 25. Malonek, D. & Grinvald, A. Interactions between electrical activity and
jects responses across all of the trials. The slope parameter, b, is a cortical microcirculation revealed by imaging spectroscopy: implications for
functional brain mapping. Science 272, 551554 (1996).
quantitative measure of the contingency between the fMRI amplitudes 26. Buxton, R. B., Wong, E. C. & Frank, L. R. Dynamics of blood flow and
and the subjects performance. The reliability of the estimate of b was oxygenation changes during brain activation: the balloon model. Magn.
determined using a statistical bootstrap procedure49. The third scheme Reson. Med. 39, 855864 (1998).
was a straightforward comparison of the ensemble of relative response 27. Boynton, G. M., Engel, S. A., Glover, G. H. & Heeger, D. J. Linear systems
analysis of functional magnetic resonance imaging in human V1. J. Neurosci.
amplitudes corresponding to correct (hits, correct rejections) and erro- 16, 42074221 (1996).
neous (misses, false alarms) trials. We used a t-test on the hypothesis that 28. Tootell, R. B. et al. Functional analysis of human MT and related visual cortical
the mean amplitude of correct trials was greater than the mean ampli- areas using magnetic resonance imaging. J. Neurosci. 15, 32153230 (1995).
tude for erroneous trials. 29. Boynton, G.M., Demb, J. B., Glover, G. H. & Heeger, D. J. Neural basis of
contrast discrimination. Vision Res. 39, 257269 (1999).
30. Carandini, M., Heeger, D. J. & Movshon, J. A. Linearity and normalization in
ACKNOWLEDGEMENTS simple cells of the macaque primary visual cortex. J. Neurosci. 17, 86218644
We thank W.T. Newsome, B.A. Wandell and M. Carandini for comments. This (1997).
31. Demb, J., Boynton, G. M. & Heeger, D. J. Functional magnetic resonance
research was supported by an NEI grant (R01-EY11794), a grant from the imaging of early visual pathways in dyslexia. J. Neurosci. 17, 86218644 (1997).
Human Frontier Science Program, and two NRSA postdoctoral research fellow- 32. Shulman, G. L. et al. Areas involved in encoding and applying directional
ships (F32-EY06899 and F32-EY06952). expectations to moving objects. J. Neurosci. 19, 94809496 (1999).
33. Heeger, D. J., Boynton, G. M., Demb, J. B., Seidemann, E. & Newsome, W. T.
Motion opponency in visual cortex. J. Neurosci. 19, 71627174 (1999).
RECEIVED 12 APRIL; ACCEPTED 10 JULY 2000 34. Wandell, B. A. et al. Color signals in human motion-selective cortex. Neuron
24, 901909 (1999).
35. Seidemann, E., Poirson, A. B., Wandell, B. A. & Newsome, W. T. Color signals
1. Pashler, H. E. The Psychology of Attention (MIT Press, Cambridge, in area MT of the macaque monkey. Neuron 24, 911917 (1999).
Massachusetts, 1998). 36. Rees, G., Friston, K. & Koch, C. A direct quantitative relationship between the
2. Posner, M. I., Snyder, C. R. & Davidson, B. J. Attention and detection of functional properties of human and macaque V5. Nat. Neurosci. 3, 716723
signals. J. Exp. Psychol. 109, 160174 (1980). (2000).
3. Graham, N. Visual Pattern Analyzers (Oxford Univ. Press, New York, 1989). 37. Heeger, D. J., Huk, A. C., Geisler, W. S. & Albrecht, D. G. Spikes versus BOLD:
4. Desimone, R. & Duncan, J. Neural mechanisms of selective visual attention. What does neuroimaging tell us about neuronal activity? Nat. Neurosci. 3,
Annu. Rev. Neurosci. 18, 193222 (1995). 631633 (2000).
5. Martnez, A. et al. Involvement of striate and extrastriate visual cortical areas 38. Ito, M. & Gilbert, C. D. Attention modulates contextual influences in the
in spatial attention. Nat. Neurosci. 2, 364369 (1999). primary visual cortex of alert monkeys. Neuron 22, 593604 (1999).
6. Chawla, D., Rees, G. & Friston, K. J. The physiological basis of attentional 39. Nestares, O. & Heeger, D. J. Robust multiresolution alignment of MRI brain
modulation in extrastriate visual areas. Nat. Neurosci. 2, 671676 (1999). volumes. Magn. Reson. Med. 43, 705715 (2000).
7. Brefczynski, J. A. & DeYoe, E. A. A physiological correlate of the spotlight of 40. Engel, S. A. et al. fMRI of human visual cortex. Nature 369, 525 (1994).
visual attention. Nat. Neurosci. 2, 370374 (1999). 41. Engel, S. A., Glover, G. H. & Wandell, B. A. Retinotopic organization in
8. Tootell, R. B. et al. The retinotopy of visual spatial attention. Neuron 21, human visual cortex and the spatial precision of functional MRI. Cereb.
14091422 (1998). Cortex 7, 181192 (1997).
9. Gandhi, S. P., Heeger, D. J. & Boynton, G. M. Spatial attention affects brain 42. Sereno, M. I. et al. Borders of multiple visual areas in humans revealed by
activity in human primary visual cortex. Proc. Natl. Acad. Sci. USA 96, functional magnetic resonance imaging. Science 268, 889893 (1995).
33143319 (1999). 43. DeYoe, E. A. et al. Mapping striate and extrastriate visual areas in human
10. Kastner, S., Pinsk, M. A., De Weerd, P., Desimone, R. & Ungerleider, L. G. cerebral cortex. Proc. Natl. Acad. Sci. USA 93, 23822386 (1996).
Increased activity in human visual cortex during directed attention in the 44. Glover, G. H. & Lai, S. Self-navigated spiral fMRI: interleaved versus single-
absence of visual stimulation. Neuron 22, 751761 (1999). shot. Magn. Reson. Med. 39, 361368 (1998).
11. Hillyard, S. A., Vogel, E. K. & Luck, S. J. Sensory gain control (amplification) 45. Glover, G. H. Simple analytic spiral K-space algorithm. Magn. Reson. Med.
as a mechanism of selective attention: electrophysiological and 42, 412415 (1999).
neuroimaging evidence. Phil. Trans. R. Soc. Lond. B Biol. Sci. 353, 12571270 46. Smith, A. M. et al. Investigation of low frequency drift in fMRI signal.
(1998). Neuroimage 9, 526533 (1999).
12. Reynolds, J. H., Chelazzi, L. & Desimone, R. Competitive mechanisms 47. Aguirre, G. K., Zarahn, E. & DEsposito, M. The variability of human, BOLD
subserve attention in macaque areas V2 and V4. J. Neurosci. 19, 17361753 hemodynamic responses. Neuroimage 8, 360369 (1998).
(1999). 48. Green, D. M. & Swets, J. A. Signal Detection Theory and Psychophysics
13. Luck, S. J., Chelazzi, L., Hillyard, S. A. & Desimone, R. Neural mechanisms of (Peninsula Publishing, Los Altos, California, 1988).
spatial selective attention in areas V1, V2, and V4 of macaque visual cortex. 49. Efron, B. & Tibshirani, R. J. An Introduction to the Bootstrap (Chapman &
J. Neurophysiol. 77, 2442 (1997). Hall, New York, 1993).

articles
Category-specific visual responses

of single neurons in the human
medial temporal lobe
Gabriel Kreiman1, Christof Koch1 and Itzhak Fried2
1 Computation and Neural Systems Program, California Institute of Technology, 139-74, Pasadena, California 91125, USA
2 Division of Neurosurgery and Department of Psychiatry and Biobehavioral Sciences, Box 957039, UCLA School of Medicine, 740 Westwood Plaza, Los Angeles,
California 90095-7039, USA
Correspondence should be addressed to I.F. (ifried@mednet.ucla.edu)
The hippocampus, amygdala and entorhinal cortex receive convergent input from temporal neocorti-
cal regions specialized for processing complex visual stimuli and are important in the representation
and recognition of visual images. Recording from 427 single neurons in the human hippocampus,
entorhinal cortex and amygdala, we found a remarkable degree of category-specific firing of individ-
ual neurons on a trial-by-trial basis. Of the recorded neurons, 14% responded selectively to visual
stimuli from different categories, including faces, natural scenes and houses, famous people and
animals. Based on the firing rate of individual neurons, stimulus category could be predicted with a
mean probability of error of 0.24. In the hippocampus, the proportion of neurons responding to spa-
tial layouts was greater than to other categories. Our data provide direct support for the role of
human medial temporal regions in the representation of different categories of visual stimuli.
Visual recognition of objects is a key function of the primate electrode probes, containing several microwires each, were placed
brain. There is a progression in the complexity of the represen- in medial temporal lobe targets bilaterally (Table 1). Based on
tation of the visual scene by single neurons. Neurons in early MRI confirmation (Fig. 1), 149 of the neurons recorded were in
visual areas in monkeys are tuned to simple features of the stim- the amygdala, 153 neurons in the entorhinal cortex and 125 neu-
uli, such as the orientation of bars in area V1 or direction of rons in the hippocampus. (Eighty-five percent of the sites were
motion in area V5. In the monkey inferotemporal cortex (IT), in the anterior segment of the hippocampus.) Most of the elec-
neurons respond to complex stimuli, including faces and hands, trodes we recorded from were in the right temporal lobe (79%).
but also abstract patterns or common, everyday objects13. There During single-neuron recording, subjects were presented with
are strong projections from IT to higher association areas in the visual stimuli (Fig. 2) and performed a simple discrimination
temporal lobe, including the parahippocampal gyrus, perirhinal task indicating whether the picture was a human face or not. Of
cortex, entorhinal cortex, hippocampus and amygdala4,5. Single the 427 neurons, 85 (20%) showed changes in firing rate during
neurons in these polymodal areas in monkeys also show visual presentation of the visual stimuli, and 61 (14%) showed visual-
selectivity for complex stimuli68. ly selective responses that were category specific.
Temporal lobe lesions lead to profound category-specific
deficits in visual recognition in both macaques and humans912. Visual responses
There is evidence from electrical stimulation studies in epileptic Most neurons showed maintained firing rates below 10 spikes
patients that current injection in the temporal lobe can interfere per second (Table 1). No significant response differences were
with visual recognition13 and elicit visual memories and halluci- observed between the right and left hemispheres, and there-
nations13,14. Functional brain imaging and event-related poten- fore the data were pooled. The average overall firing rate was
tials (ERP) also show a correlation between brain activity and 3.6 5.6 spikes per second, similar to observations in rats and
visual recognition of specific categories of stimuli such as human monkeys68,22.
faces and spatial layouts or places1520. We studied the responses of each neuron to the 1000-ms pre-
We reported that neurons in the human medial temporal lobe sentation of the visual stimuli by averaging the activity for all
discriminate objects from faces21. Here we further investigated images within each category. For each neuron and each category,
visual response properties and showed that single neurons a post-stimulus time histogram was computed, showing the neu-
respond selectively to different stimulus categories. ronal response starting 1000 ms before stimulus onset and end-
ing 1000 ms after the stimulus disappeared.
RESULTS A neuron was considered visually selective for a specific cat-
We recorded the activity of 427 single neurons in 11 patients with egory if the activity during stimulus presentation for that cate-
pharmacologically resistant epilepsy who had intracranial depth gory was significantly different from the baseline activity and
electrodes implanted to determine the location of the seizure from the responses to other categories of stimuli (Methods).
focus for possible surgical resection. Based on clinical criteria, For example, a visually selective neuron in the entorhinal cor-

articles
b
Fig. 1. Electrode placement. The trajectory of an electrode placed in the Fig. 2. Sample of stimuli presented in each category. Figures (mostly
hippocampus is depicted in axial (a) and coronal (detail, b) structural MR color) were drawn from a group of nine categories that included faces
images (1.5 Tesla scanner). Post-operative CT and MRI were used to con- denoting emotional expressions by unknown actors21, household
firm the location of the electrode. The CT was co-registered with MRI objects, spatial layouts (including house exteriors, interiors and natural
structural information for anatomic verification. The distal end of the scenes), animals, cars, drawings of famous people or cartoon characters,
electrode included platinum-iridium microwires from which single neu- photographs of famous people, food items and abstract patterns. Stimuli
rons were recorded. The microwires extended about 4 mm from the tip, were presented for 1000 ms. Subjects had to indicate by pressing a but-
lying on a cone with an opening angle of less than 45 degrees. ton whether the image was a human face or not.
tex had increased firing rate in response to pictures of animals tures of animals using an analysis of variance (ANOVA). Both
(Fig. 3a). The activity of this neuron during the 1001000 ms parametric and non-parametric tests failed to show differences
interval after stimulus onset was different from baseline for ani- among individual stimuli (p > 0.4).
mal stimuli (p < 104, Wilcoxon rank-sum test), but not for the One visually selective neuron in the anterior hippocampus
other stimulus categories (p > 0.1). A one-way ANOVA com- (Fig. 4) showed an increased firing rate over baseline in response
paring firing rates between categories yielded p < 0.001, and to drawings of famous people as well as, to a lesser degree, to
comparing the activity for animals to all other categories using photos of famous people (p < 0.001). A one-way ANOVA yield-
a pair-wise non-parametric Wilcoxon test yielded statistically ed p < 0.001, and subsequent across-categories, pair-wise com-
significant differences (p < 0.001). The latency of response for parisons also showed that the activity during stimulus
this neuron was 219 ms, and the duration of increased response presentation was significantly higher for these two categories.
over baseline was 752 ms. Although the peak response was larger for drawings than for
How specific was the response of the neuron within the photographs (13.9 spikes/s versus 9.6 spikes/s), the average
selective category? If the average
increased response to animals were Table 1. Number of neurons and response properties.
due to enhanced firing for only a few Amygdala Entorhinal cortex Hippocampus Total
pictures of animals, one might
expect to observe a bimodal (or mul- Left Right Left Right Left Right Left Right
timodal) distribution of firing rates. Number of channels 21 57 13 58 12 54 215
However, the distribution of firing Number of neurons 50 99 24 129 44 81 427
rates for this neuron during presen- f (spikes/s) 2.1 3.5 4.7 5.4 4.0 7.2
tation of different pictures of ani- Responsive neurons 6 (12) 12 (12) 6 (25) 29 (22) 16 (36) 16 (20) 28 (24) 57 (18)
mals did not show any clear signs of Selective neurons 5 (10) 9 (9) 5 (21) 20 (16) 13 (30) 9 (11) 23 (19) 38 (12)
multimodality (Fig. 3b). Although Latency (ms) 240 145 209 119 239 132
there was variability in the response
Durations (ms) 459 322 507 356 568 325
of the neuron to individual instances
of animals, the neuron responded Number of recorded channels and isolated neurons in each location in each hemisphere from the 11 patients.
above baseline for all pictures of ani- On average, we recorded approximately two neurons per microwire. For each neuron, we computed the
mean firing rate (f) over the entire experimental session. The firing rates in spikes/s ranged from 0.03 to 27 in
mals (Fig. 3c; p < 0.005). We also the amygdala, 0.02 to 38 in the entorhinal cortex and 0.03 to 29 in the hippocampus. The criteria used to clas-
compared the variability for differ- sify a neuron as visually responsive or selective are described in the text. The numbers in parenthesis indicate
ent presentations of the same animal the percentages with respect to the total number of neurons in each location. Neurons with late responses
were not included in the mean latency computation (Methods).
to the variability across different pic-

articles
Fig. 3. Visually selective neuron in

the entorhinal cortex. (a) Post-stim- a
ulus time histogram (PSTH) of the
responses of a neuron in the right
entorhinal cortex. The rasters and
histograms are aligned to the onset Emotional face (100) Object (51) Spatial (105) Animal (46)
of the stimulus. The stimulus was pre- 30 30 30 30
sented between t = 0 and t = 1000 ms
Spikes/s
(indicated by dashed vertical lines in 20 20 20 20
each histogram). Responses were
10 10 10 10
averaged for all stimuli within a given
category using a bin size of 200 ms. 0 0 0 0
The dashed horizontal line indicates -1 0 1 2 -1 0 1 2 -1 0 1 2 -1 0 1 2
the mean firing rate over the whole
experiment (10.3 spikes/s). The cate-
gory and the number of stimuli pre-
sented in each category are indicated
at the top of each histogram. The fir-
ing rate in the 1001000 ms interval
upon presentation of a picture of an Car (17) Face drawing (134) Famous face (100) Pattern (54)
animal was significantly different from 30 30 30 30

that in the 1000 to 0 ms baseline
Spikes/s
preceding the stimulus onset (p < 104). 20 20 20 20

The probability of error (pe) from the
10 10 10 10
ROC analysis (Fig. 6) was 0.21.
There were only five presentations of 0 0 0 0
food items in this experiment, and -1 0 1 2 -1 0 1 2 -1 0 1 2 -1 0 1 2
they were not included in the graph. b c Time (s)
The neuron did not respond to food
15 30 30 30 30
items based on these five repetitions.
20 20 20 20
(b) Distribution of firing rates during
10 10 10 10
presentation of pictures of animals.
10
Proportion (%)
spikes/s
0 0 0 0
Histogram distribution of mean firing 0 1 0 1 0 1 0 1
rate of response in each trial, bin size 30 30 30
of 1.5 spikes/s. There is no clear sign 5 20 20 20
of bimodality in the distribution.
10 10 10
(c) PSTHs showing the responses of
0 0 0
this neuron to each individual stimu- 0 0 1 0 1 0 1
lus within the category of animals. 0 10 20 30

Although the responses vary from FR (spikes/s)
one stimulus to another, the neuron
responds to all stimuli within this category (comparison with baseline, p < 0.01). An analysis of variance comparing the responses to different indi-
vidual animals did not yield significance (p > 0.4). The scale is the same as in (a).
activity was not significantly different (Wilcoxon rank-sum test, dala, 10% in the entorhinal cortex and 25% in the hippocam-
p > 0.05). This neuron did not respond to faces per se, as indi- pus. Most of these neurons responded to two categories.
cated by the lack of change in the activity for emotional faces We also observed neurons that showed significant but non-
of unknown actors. The distribution of firing rates in response selective changes in firing rate during stimulus presentation (Meth-
to photos of famous people for this neuron did not show clear ods). The proportion of nonselective neurons, relative to the total
signs of displaying more than one mode (Fig. 4b). Variability number of responsive neurons, was 28% in the hippocampus, 29%
in the responses to distinct individual photos (Fig. 4c) was not in the entorhinal cortex and 17% in the amygdala.
higher than variability across different presentations of the same Although most of the visually selective neurons showed
photograph (ANOVA, p > 0.2). Similar results held for draw- increases in the firing rate on presentation of visual stimuli, some
ings of famous people. The response to all individual stimuli cells had reduced firing rate from baseline (three neurons in the
within the selective categories was significantly different from amygdala, two in the entorhinal cortex and two in the hip-
baseline (p < 0.01). pocampus). Decreases were also observed in the visually respon-
Although we used a significance criterion of 0.05 in the sta- sive but nonselective neurons (one neuron in the amygdala, five
tistical analysis, most of the actual p values were below 0.01. For in entorhinal cortex and three in hippocampus).
the visually selective neurons, 69% of the p values were less than Most of the neurons responded during the stimulus-pre-
0.01 (average p value, 0.01 0.02). A 2 test rejected the hypoth- sentation period, but there were some that responded when the
esis that these responses could be due to chance (p < 0.001). stimulus was removed. To address this, we computed for all
Some of the neurons showed changes in firing rate in response neurons, within the 2000 ms after stimulus onset, the number
to more than one of the categories (for example, Fig. 4). The pro- of spikes in a 600-ms interval centered on the peak of the
portion of neurons, relative to the number of selective neurons, response and statistically analyzed the responses as described
that responded to more than one category was 21% in the amyg- above. There were eight selective neurons (four in hippocam-

articles
Fig. 4. Visually selective neuron in the

hippocampus. (a) PSTH of the a
responses of a neuron in the right
anterior hippocampus. The notation
and symbols are the same as in the
previous figure. The firing rate in the
Emotional face (100) Object (51) Spatial (105) Animal (46)
1001000 ms interval on presentation 15 15 15 15
of a drawing or a photo of a famous
face was significantly different from 10 10 10 10
Spikes/s
that in the 1000 to 0 ms baseline
preceding stimulus onset (p < 0.001). 5 5 5 5
Note that this neuron does not
0 0 0 0
respond to just any face, as it fails to -1 0 1 2 -1 0 1 2 -1 0 1 2 -1 0 1 2
change its activity to the unknown
actors depicting emotional expres-
sions (top, left histogram). The mean
firing rate over the whole experiment
was 3.1 spikes/s, and the pe was 0.19.
There were only five presentations of
food items in this experiment, and Car (17) Face drawing (134) Famous face (100) Pattern (54)
they were not included in the graph. 15 15 15 15

The neuron did not respond to food
items based on these five repetitions. 10 10 10 10
Spikes/s
(b) Distribution of firing rates during

5 5 5 5
presentation of famous faces.
Histogram distribution of mean firing 0 0 0 0
rate of responses in each trial, bin size -1 0 1 2 -1 0 1 2 -1 0 1 2 -1 0 1 2
of 1.5 spikes/s. There is no clear sign Time (s)
of bimodality in the distribution.
(c) PSTHs showing the responses of
b c 15 15 15 15
10 10 10 10
this neuron to each individual stimulus 20 5 5 5 5
within the category of photos of 0
0 1
0
0 1
0
0 1
0
0 1
Proportion
famous faces. Although the responses 15 15 15 15 15

spikes/s
vary from one stimulus to another, 10 10 10 10

5 5 5 5
the neuron responds to all stimuli 10
0 0 0 0
0 1 0 1 0 1 0 1
within this category (comparison with
15 15 15
baseline, p < 0.05). An ANOVA com- 5 10 10 10
paring the responses to different indi- 5 5 5
vidual famous faces did not yield 0 0

0 1
0
0 1
0
0 1
0 5 10 15 20
significance (p > 0.2). The scale is the
FR (spikes/s)
same as in (a).
pus, three in entorhinal cortex and one in the amygdala) that We computed the latency and the duration of the evoked
showed a statistically significant late response and were not activities for all neurons showing a visual response (Methods).
detected with the previous analysis. The latencies ranged from 52 to 695 ms and the durations from
Under the assumption that there is no preference in the 53 to 1190 ms (Table 1). There was no significant difference
prevalence of selectivity for any of the nine different stimulus among the amygdala, entorhinal cortex and hippocampus in
categories, the number of selective neurons within any one area either of these two variables (ANOVA, p > 0.1).
should be uniformly distributed among these categories. We
tested this hypothesis using a 2 test23. We obtained 2 values Classification by an ideal observer
of 6.0, 9.6 and 29.5 for the amygdala, entorhinal cortex and We assessed how well an ideal observer could discriminate, based
hippocampus respectively (Fig. 5). The p values were over 0.25 on the response of an individual neuron, whether a stimulus
for the amygdala and entorhinal cortex but less than 103 for belonged to the category for which the neuron was selective or
the hippocampus. In the hippocampus, we observed a small not. We computed pe, the probability of misclassifying a stimulus
relative proportion of responses to animals, food items and based on the firing rate, using a classical optimal decision pro-
patterns and, interestingly, a relatively high number of neu- cedure (ROC analysis; Methods and Fig. 6ac). The value of pe
rons responding to spatial layouts (Fig. 5). No significant dif- can range from 0 to 0.5, with pe = 0.5 indicating chance perfor-
ference was observed in a direct comparison of the response mance and pe = 0 indicating perfect classification. Our pe values
of any of these neurons to house facades versus natural scenes ranged from 0.13 to 0.32 (0.22 0.06, mean s.d.) in the amyg-
(Wilcoxon, p > 0.05). dala, 0.04 to 0.44 (0.23 0.10) in the entorhinal cortex and 0.08
Among the selective neurons, there were 2 that yielded p < 0.05 to 0.47 (0.23 0.10) in the hippocampus (Fig. 6df).
(and 3 more with a p value between 0.05 and 0.1) in the ANOVA
analysis of specificity to individual stimuli within the selective cat- DISCUSSION
egory. These neurons responded more strongly to one to three of Increasingly complex stimulus attributes are represented from
the individual stimuli within the selective category. These were the retina to the higher visual areas. Evidence from neurolo-
excluded from the number of selective neurons in Table 1. gy11,12,24, functional brain imaging1517,20 and evoked-potential

articles
Emotional faces Objects Spatial Animals Cars

Emotional faces Objects Spatial Animals Cars
the activity of multiple neurons, it is likely that an even higher
30 30 30 30 30
level of accuracy can be achieved. Such category-specific pro-
25 25 25 25 25 cessing may be important not only in object recognition, but also
20 20 20 20 20
in the representation and retrieval processes that have been close-
ly linked with the medial temporal lobe3436.
15 15 15 15 15
Twenty percent of the neurons that we recorded showed a
10 10 10 10 10 visual response and 14% a visually selective one. These percent-
ages are comparable to those reported in monkeys. In the
units
5 5 5 5 5
entorhinal cortex, 11% of the neurons show selective visual
responsive units
responses7, compared to 16% in our study. In the monkey amyg-

ofofresponsive
dala, 12% of the neurons show visual responses and about 33%
of those are selective for faces6, compared to 12% and 20%,
Face
Face drawings
drawings Famous
Famous Food
Food Patterns
Patterns Non-selective
Non-selective
respectively, in our data.
Percentage
30 30 30 30 30
Most of the selective neurons responded to only a single
Percentage
25 25 25 25 25
stimulus category, rather than weakly responding to a large frac-
20 20 20 20 20 tion of all stimuli. Our data thus support the existence of sparse
coding in the medial temporal lobe. Sparsely coded represen-
15 15 15 15 15
tation has been suggested for information processing in the
10 10 10 10 10 rodent and primate hippocampus3739 and for processing of

faces and objects in IT40,41.
5 5 5 5 5
Whereas a significant proportion of neurons are selective for
faces in the superior temporal sulcus13, responses in the entorhi-
nal cortex7, hippocampus8 and amygdala6 are much more var-
Amygdala
Amygdala Entorhinal
Entorhinal Hippocampus
Hippocampus
ied. Responses were also diverse in our sample of entorhinal
Fig. 5. Distribution of selective neurons for each category. The per- cortex and amygdala neurons (Fig. 5). However, in the hip-
centage of neurons selective for each category of stimuli and the per- pocampus, we observed more responses to images showing spa-
centage of nonselective neurons are shown for each location (black, tial layouts, including houses, natural scenes and interiors. The rat
amygdala; gray, entorhinal cortex; white, hippocampus). The percent- hippocampus contains place cells that respond selectively to the
ages are based on the total number of responsive neurons in each loca-
position of a rat while it is navigating a maze22,42. Neurons in the
tion. Data from the right and left hemisphere have been pooled. A
neuron was considered to be visually selective if the activity during stim-
monkey hippocampus respond selectively depending on the posi-
ulus presentation for a given category was significantly different from the tion of the stimulus in a conditioned spatial response task8. Func-
baseline and from the neuronal response to other categories. If the tional MRI studies report parahippocampal43 and hippocampal44
response was different from baseline but not among the different cate- activation associated with navigational tasks as well as while
gories, the neuron was defined as nonselective. observing images similar to the ones shown in our study27,28. This
area is posterior to the medial temporal recording sites in our
study, and likely projects to the hippocampus.
It is possible that some neurons may change their activity
studies in humans18, and from single-neuron electrophysiology13 more specifically than to the broad categories used in our study,
and lesions in monkeys9,10,25 suggests a fundamental role for the responding, for instance, only to one specific example that we
medial temporal lobe in visual object recognition. did not present 45. Our experimental setting did not enable
Category-specific knowledge deficits occur in which neuro- investigation of this issue for several reasons. First, because of
logical patients show impairments in identifying living things, clinical considerations, an electrode location remained fixed
objects, food items or faces11,12,24. Functional imaging studies show once placed, and we did not change electrode location in search
activation of different areas in the temporal lobe that correlates for an optimal stimulus, as is commonly done in animal exper-
with subjects observing pictures belonging to different cate- iments. Also, because the data for all the channels were ana-
gories1517. In particular, there are areas specialized for faces20,26, lyzed off-line, we could not determine, via immediate feedback,
spatial layouts27,28, objects and animals17. Specific changes in activ- the optimal stimulus for any of the cells. For most of the selec-
ity on presentation of faces, objects and letter strings can also be tive neurons (61 of 63), the analysis of variance showed that
observed by evoked potentials in humans2931. variability for different stimuli within the selective category
Single inferotemporal cortex (IT) neurons in monkeys respond was comparable to variability due to different presentations of
to complex visual stimuli, including faces, objects and abstract the same stimulus.
patterns13. Neurons in human temporal neocortex respond to Data from very different experiments and using distinct tech-
faces and words32,33. Information from these neocortical neurons niques are converging to show an important role for the human
is conveyed to polymodal association areas in the temporal lobe4,5. medial temporal lobe in visual object recognition. This study
We showed that in the relatively small matrix of hippocam- establishes that single neurons in humans explicitly respond to
pus, entorhinal cortex and amygdala, there is a remarkable degree specific categories of stimuli, which may be relevant to the rep-
of segregation of categories at the level of single neurons. Neu- resentation and retrieval of visual information.
rons in these regions show visual object discrimination among
at least nine stimulus categories. Based on the firing rate of indi-
METHODS
vidual neurons, it was possible to predict with a mean probabil- Patients. Subjects were patients with pharmacologically resistant epilep-
ity of error of 0.24 whether the preferred stimulus category was sy. Extensive non-invasive evaluation did not yield concordant data
presented or not (Fig. 6). This, by itself, shows a striking degree corresponding to a single resectable epileptogenic focus, and therefore
of category-specific firing on a trial-by-trial basis. By combining the patients were stereotactically implanted with up to 12 chronic

articles
intracranial depth electrodes for one to two a Pref. b c

15 1
weeks to determine the focus of their seizures 0.5
for possible surgical resection21,46. Through 10
0.8
the lumen of the electrodes, up to 8 5 0.4
microwires (40 m diameter) were inserted.
Proportion (%)
error
0 0.6
The surgeries were performed by I.F. All stud- 0.3
CD
P
ies described here conformed with the guide-
P
15 0.4
lines of the Medical Institutional Review 0.2

Non-Pref.
10 pe
Board at UCLA. The present study describes 0.2 0.1
data from 11 subjects (6 males and 5 females; 5
8 right-handed and 3 left-handed; 24 to 48 0 0 0
years old). 0 10 20 30 0 0.5 1 0 0.5 1
The sites of implantation of the electrodes Firing rate (spikes/s) P
FA
P
FA
were based exclusively on clinical criteria. The
location of the electrodes was verified by d Amygdala e Entorhinal f Hippocampus
25
structural MR images obtained before remov- 30 20
ing the electrodes (Fig. 1). Individual
20
microwires extended approximately 4 mm 25
15
from the tip, lying in a cone with an opening
Proportion (%)
20 15
angle of less than 45 degrees.
We report here the activity of neurons for 15 10

10
all the probes located in the medial temporal 10
lobe. Neurons from anterior (85%), middle 5 5
(10%) and posterior (5%) parts of the hip- 5
pocampus were pooled together as hip- 0 0 0
pocampus neurons. Most neurons in the 0 0.2 0.4 0 0.2 0.4 0 0.2 0.4
amygdala were in the basolateral nuclear com- pe
plex. Our MR resolution did not allow us to
accurately determine in which CA fields the Fig. 6. ROC analysis and pe distribution. (a) Distribution of firing rates for the neuron shown in
hippocampal probes were placed or what layer Fig. 3 for the preferred category (top, animals) and non-preferred category (bottom, all other cate-
of the entorhinal cortex we recorded from. gories). Bin size, 1.5 spikes/s. These represent the conditional probability distributions P(f | stim pre-
The information recorded during seizures ferred category) and P(f | stim preferred category), where f denotes the firing response computed in
from the depth electrodes was used to localize the 1001000 ms interval and stim indicates the stimulus. (b) ROC analysis. Each stimulus is classified
the seizure focus 46 . Eighty-five percent of into the preferred category if the firing rate during the 1001000 ms interval is above a given thresh-
recorded neurons were outside the clinically old T. The probability of correctly classifying a stimulus (PCD) is plotted as a function of the probability
determined zone of seizure onset (that is, either of making a false alarm (PFA). This was calculated for successive values of the threshold T by integrating
in the other hemisphere or in a different brain the tails of the two distributions: P(f > T | stim preferred cat.) and P(f > T | stim preferred cat.). The
area on the same side). Ninety-four percent of dashed line indicates chance performance (PCD = PFA). The dots indicate the actual data using numer-
responsive neurons were outside the seizure ical integration, whereas the continuous line indicates the values after assuming a Gaussian distribu-
focus. Because we did not observe any differ- tion of firing rates. (c) The overall probability of error, perror(T) = 1/2 PFA (T) + 1/2 (1PCD(T)) is plotted
ences in their waveforms, firing rates, interspike as a function of the probability of false alarm. The classification performance of the neuron based on
interval distributions or response properties, the firing rate was characterized by the minimum in this curve, pe (indicated by an arrow).
all neurons were included in Table 1. The per- (df) Distribution of pe for neurons in each of the three locations. Bin size, 0.03.
centage of responsive neurons would increase
from 20% to 22% if we excluded neurons with-
in the seizure focus.
Experimental protocol. A series of images was key experiments. We analyzed all spikes from all neurons we detected.
shown on a monitor at an approximate size of 5 degrees of visual angle Spikes from single neurons were discriminated from the extracellular
(Fig. 2). Each picture was repeated 410 times (depending on time con- recordings based on the height, width, peak voltage and other parameters
straints), and there was a total of up to 600 presentations; the order of of the waveforms using a manual cluster-cutting method implemented
presentation of the stimuli was random. The number of different indi- in Datawave. For each isolated neuron, we determined the fraction of all
vidual stimuli per category ranged from 3 to 25 (7.2 4.6, mean s.d.). spikes that were within 2 ms of each other. If this fraction exceeded 2%,
Each picture was presented for 1000 ms. In the first two patients, stim- the data were discarded because of possible contamination by firing from
uli from only three different categories were presented (emotional faces, more than one neuron.
objects and spatial layouts). Immediately after the picture disappeared, In some cases, we recorded from the same microwires on separate days
there was a tone that indicated that the subject had to respond whether (presenting different sets of pictures from the same categories). Because
the picture was a human face or not by pressing a button. This was done of various constraints, we often could not be sure that a neuron record-
to engage the subjects attention and to verify that he was seeing the ed on one day was identical to a neuron recorded on another day. Of our
pictures. Trials in which subjects made an error (mean percentage cor- 427 neurons, 128 (30%) were recorded on consecutive days. If we assume
rect 97 1%) or in which the subjects moved were discarded from sub- that all 128 neurons are identical across days, the total number of neu-
sequent analysis. The behavioral response occurred on average 472 rons reduces to 299, and the percentage of responsive neurons increas-
201 ms after stimulus offset. es from 20% (85 of 427) to 23% (70 of 299).
Recordings. Data from each of the recorded microwires were amplified Data analysis. For each neuron, we computed histograms locked to the
and high-pass filtered (with a corner frequency of 300 Hz), A/D con- stimulus presentation times (post-stimulus time histograms, PSTHs) by
verted and stored for off-line spike sorting using Experiment Workbench averaging the neuronal responses for all stimuli within each category. We
data acquisition software (Datawave, Denver, Colorado). Note that also compared all color versus non-color pictures; we found only two
because the microelectrodes were chronically implanted, we could not neurons with a statistically significant enhanced response to color stim-
further select neurons by moving the electrodes, as is common in mon- uli. To assess the significance of the response to different categories of

articles
visual stimuli, we ran a series of statistical comparisons. For a neuron to 2 in the amygdala) for which the activity in the 1001000 ms interval for
be considered visually selective for a specific category, it had to meet three all face stimuli was significantly larger than the baseline activity and than
requirements. First, the neuronal response pooled over all stimuli belong- the response for non-face stimuli (1-way ANOVA, p < 0.05; pairwise
ing to the same category had to show a different firing rate, that is, lower comparisons, p < 0.05). However, for some of these neurons (for exam-
or higher, during the time of presentation of the image than during the ple, Fig. 4a), the response was selective for some faces and not others.
preceding baseline. The firing rate during image presentation was com- For those neurons that showed visual selectivity for stimuli within
puted in the interval 100 ms t < 1000 ms. (The lower boundary of specific categories, we also addressed the question of how well an ideal
100 ms was chosen because most neurons showed latencies above observer could estimate which category the stimulus belonged to by
100 ms.) The baseline before stimulus presentation was computed in the observing the firing rate. To quantify the classification performance of
interval 1000 ms t < 0 ms. Significance was assessed both by a two- the neurons, we used an ROC analysis as used in signal-detection theo-
tailed t-test and a non-parametric Wilcoxon rank-sum test23. Because ry and psychophysics experiments50. For each visually selective neuron,
the results from both tests were very similar, the p values reported we computed the distribution of firing rates for the preferred stimuli and
throughout the text for pairwise comparisons correspond to those from the non-preferred stimuli (the remaining categories; Fig. 6a). From the
the non-parametric test. Second, a one-way analysis of variance (ANOVA) distribution of firing rates we evaluated, by sliding a threshold T over
during the 1001000 ms interval after stimulus onset, assessing whether the whole range of firing rates, the probability of correct detection (PCD)
there were differences in firing rate among the different categories, had to and the probability of false alarm (PFA). Assuming chance performance,
show a significant p-value (< 0.05). In this analysis, stimuli were pooled PCD = PFA (Fig. 6b, dashed line). The departure from the diagonal shows
according to the category they belonged to. Third, subsequent pairwise the possibility of discriminating between the preferred and non-pre-
comparisons between the activity during this interval for the putative ferred categories. The probability of misclassification plotted against the
selective category and the rest of the categories had to show a statistical- probability of false alarm can be obtained as
ly significant change (Wilcoxon rank sum test, < 0.05).

Although we used a significance criterion of 0.05, most responsive perror (T) = 1/2 PFA (T) + 1/2 (1 PCD(T))
neurons showed p values that were less than 0.01 (Results). Our analy-
sis encompassed both increases and decreases in firing rate. The overall probability of error, pe, is then defined as the minimum
If the ANOVA test failed to indicate a significant difference between value of this function (arrow in Fig. 6c). A value of pe = 1/2 corresponds
categories, but the activity for 75% of the stimulus categories was sig- to chance performance, whereas a value of pe = 0 indicates that it is pos-
nificantly different from the baseline activity (that is, the first criterion sible to predict with perfect accuracy based on the number of spikes
was met but not the second), we labeled the neuron as visually responsive whether the specified category was presented or not.
but nonselective. Throughout the manuscript, values are given as mean standard devi-
For the neurons that showed visual selectivity, we further studied the ation (s.d.).
degree of specificity of the responses. We computed the distribution of
firing rates for the selective category to assess whether it was a polymodal
distribution. We also computed an analysis of variance (both a para- ACKNOWLEDGEMENTS
metric analysis23 as well as a non-parametric analysis using a bootstrap This work was supported by grants from NIH (to I.F.), the Keck Foundation (to
procedure47) to compare the variance to different individual stimuli with- C.K.) and the Center for Consciousness Studies, University of Arizona (to I.F.
in the same category to the variance to repeated presentations of the same and C.K.). We would like to thank D. Rozenfarb and M. Zirlinger for
stimulus. suggestions and reading the manuscript and Peter Steinmetz for general discus-
It is reported that neurons can respond to visual stimuli with a long sion. We wish to acknowledge Tony Fields, Jack Morrow, Eve Isham, Charles
delay even after the stimulus disappears21,48,49. To study these cases, we
Wilson, Rick Staba, Eric Behnke and Anatol Bragin for help with the recordings.
analyzed the responses in an interval of 600 ms centered on the response
peak. To estimate the time of occurrence of the peak, we computed an
estimation of the spike density function (sdf) for each category of stim- RECEIVED 16 JUNE; ACCEPTED 18 JULY 2000
uli by convolving the spike trains with a Gaussian of fixed width of 100 ms
and then averaging over repetitions.
1. Logothetis, N. K. & Sheinberg, D. L. Visual object recognition. Annu. Rev.
We computed the latency and duration of the responses from the sdf Neurosci. 19, 577621 (1996).
for each category. The latency was defined as the first time value on which 2. Tanaka, K. Inferotemporal cortex and object vision. Annu. Rev. Neurosci. 19,
five consecutive bins of the sdf yielded a value that was beyond two s.d. of 109139 (1996).
the mean response before stimulus presentation. In analogous fashion, 3. Gross, C. G. How inferior temporal cortex became a visual area. Cereb. Cortex
the end of the response was defined as the first time point at which five 5, 455-469 (1994).
4. Suzuki, W. A. Neuroanatomy of the monkey entorhinal, perirhinal and
consecutive bins of the sdf yielded a value not different from baseline parahippocampal cortices: Organization of cortical inputs and
after the latency value. interconnections with amygdala and striatum. Sem. Neurosci. 8, 312 (1996).
To discriminate whether the neurons showed significant activity relat- 5. Saleem, K. S. & Tanaka, K. Divergent projections from the anterior
ed to the behavioral response, we did two additional analyses. First, we inferotemporal area TE to the perirhinal and entorhinal cortices in the
macaque monkey. J. Neurosci. 16, 47574775 (1996).
computed histograms of all the neuronal responses locked to the time 6. Leonard, C. M., Rolls, E. T., Wilson, F. A. W. & Baylis, G. C. Neurons in the
at which the subjects pressed the button by averaging over all stimuli. amygdala of the monkey with responses selective for faces. Behav. Brain. Res
We then computed whether there was a significant difference in the 15, 159176 (1985).
response in the interval 300 ms before and 300 ms after the button was 7. Suzuki, W., Miller, E. & Desimone, R. Object and place memory in the
pressed. This interval was chosen so that it would not overlap on average macaque entorhinal cortex. J. Neurophysiol. 78, 10621081 (1997).
8. Miyashita, Y., Rolls, E. T., Cahusac, P. M. B., Niki, H. & Feigenbaum, J. D.
with the periods of visual presentation. We found four neurons (three in Activity of hippocampal formation neurons in the monkey related to a
the amygdala, one in the hippocampus) for which the response before conditional spatial response task. J. Neurophysiol. 61, 669678 (1989).
the button press was significantly different from that after the button 9. Mishkin, M. Memory in monkeys severely impaired by combined but not
press and also from that during the baseline 1000 to 0 ms interval separate removal of amygdala and hippocampus. Nature 273, 297298
(1-way ANOVA, p < 0.05; pairwise comparisons, p < 0.05). None of (1978).
10. Meunier, M., Bachevaoer, J., Mishkin, M. & Murray, E. A. Effects on visual
these neurons was visually responsive. We also found 2 neurons that recognition of combined and separate ablations of the entorhinal and
showed an increased activity in a 600-ms window around the response perirhinal cortex in rhesus monkeys. J. Neurosci. 13, 54185432 (1993).
compared to the baseline (Wilcoxon test, p < 0.05). Neither of these 11. Tranel, D., Damasio, H. & Damasio, A. Face agnosia and the neural substrates
neurons was visually responsive. of memory. Annu. Rev. Neurosci. 13, 89109 (1990).
12. Warrington, E. & McCarthy, R. Categories of knowledge Further
When computing the average neuron activity in response to all human fractionations and an attempted integration. Brain 110, 12731296 (1987).
faces (emotional faces, drawings and photos of famous faces), we 13. Fried, I., Mateer, C., Ojemann, G., Wohns, R. & Fedio, P. Organization of
observed 10 neurons (5 in the hippocampus, 3 in the entorhinal cortex, visuospatial functions in human cortex. Brain 105, 349371 (1982).

articles
14. Penfield, W. & Jasper, H. Epilepsy and the Functional Anatomy of the Human 32. Ojemann, J. G., Ojemann, G. A. & Lettich, E. Neuronal activity related to
Brain (Little, Brown, Boston, 1954). faces and matching in human right nondominant temporal cortex. Brain
15. Treisman, A. M. & Kanwisher, N. G. Perceiving visually presented objects: 115, 113 (1992).
recognition, awareness, and modularity. Curr. Opin. Neurobiol. 8, 218226 33. Ojemann, G. A., Ojemann, S. G. & Fried, I. Lessons from the human brain:
(1998). neuronal activity related to cognition. The Neuroscientist 4, 285300 (1998).
16. Ishai, A., Ungerleider, L. G., Martin, A., Schouten, J. L. & Haxby, J. V. 34. Scoville, W. B. & Milner, B. Loss of recent memory after bilateral
Distributed representation of objects in the human ventral visual pathway. hippocampal lesions. J. Neurol. Neurosurg. Psychiatry. 20, 1121 (1957).
Proc. Natl. Acad. Sci. USA 96, 93799384 (1999). 35. Penfield, W. & Milner, B. Memory deficit produced bilateral lesions in the
17. Martin, A., Wiggs, C. L., Ungerleider, L. G. & Haxby, J. V. Neural correlates of hippocampal zone. Arch. Neurol. Psychol. 79, 475479 (1958).
category-specific knowledge. Nature 379, 649652 (1996). 36. Zola-Morgan, S. & Squire, L. R. Neuroanatomy of memory. Annu. Rev.
18. Allison, T. et al. Face recognition in human extrastriate cortex. J. Neurosci. 16, 547563 (1993).
Neurophysiol. 71, 821825 (1994). 37. McNaughton, B. L. & Nadel, L. in Neuroscience and Connectionist Theory (ed.
19. Nobre, A., Allison, T. & McCarthy, G. Word recognition in the human Seifert, W.) 164 (Academic, New York, 1990).
inferior temporal lobe. Nature 372, 260263 (1994). 38. Alvarez, P. & Squire, L. R. Memory consolidation and the medial temporal
20. Tong, F., Nakayama, K., Moscovitch, M., Weinrib, O. & Kanwisher, N. lobe: a simple network model. Proc. Natl. Acad. Sci. USA 91, 70417045 (1994).
Response properties of the human fusiform face area. Cognit. Neuropsychol. 39. Treves, A. & Rolls, E. T. Computational analysis of the role of the
17, 257279 (2000). hippocampus in memory. Hippocampus 4, 374391 (1994).
21. Fried, I., MacDonald, K. A. & Wilson, C. Single neuron activity in human 40. Young, M. P. & Yamane, S. Sparse population coding of faces in the inferior
hippocampus and amygdala during recognition of faces and objects. Neuron temporal cortex. Science 256, 13271331 (1992).
18, 753765 (1997). 41. Logothetis, N. K. & Pauls, J. Psychophysical and physiological evidence for
22. Wilson, M. A. & McNaughton, B. L. Dynamics of the hippocampal ensemble viewer-centered object representations in the primate. Cereb. Cortex 3,
code for space. Science 261, 10551058 (1993). 270288 (1995).
23. Keeping, E. S. Introduction to Statistical Inference (Dover, New York, 1995). 42. Muller, R. A quarter of a century of place cells. Cell 17, 813822 (1996).
24. Adolphs, R., Tranel, D., Damasio, H. & Damasio, A. Impaired recognition of 43. Aguirre, G. K., Detre, J. A., Alsop, D. C. & DEsposito, M. The
emotion in facial expressions following bilateral damage to the amygdala. parahippocampus subserves topographical learning in man. Cereb. Cortex 6,
Nature 372, 669672 (1994). 823829 (1996).
25. Buckley, M. J. & Gaffan, D. Perirhinal cortex ablation impairs object 44. Maguire, E. A. et al. Knowing where and getting there: a human navigation
identification. J. Neurosci. 18, 22682275 (1998). network. Science 280, 921924 (1998).
26. Kanwisher, N., McDermott, J. & Chun, M. M. The fusiform face area: a 45. Heit, G., Smith, M. E. & Halgren, E. Neural encoding of individual words
module in human extrastriate cortex specialized for face perception. and faces by the human hippocampus and amygdala. Nature 333, 773775
J. Neurosci. 17, 43024311 (1997). (1988).
27. Epstein, R. & Kanwisher, N. A cortical representation of the local visual 46. Fried, I. et al. Cerebral microdialysis combined with single-neuron and
environment. Nature 392, 598601 (1998). electroencephalographic recording in neurosurgical patients. J. Neurosurg.
28. Epstein, R., Harris, A., Stanley, D. & Kanwisher, N. The parahippocampal 91, 697705 (1999).
place area: recognition, navigation, or encoding? Neuron 23, 115125 (1999). 47. Efron, B. & Tibshirani, R. J. An Introduction to the Bootstrap (Chapman &
29. Allison, T., Puce, A., Spencer, D. & McCarthy, G. Electrophysiological studies Hall/CRC, London, 1993).
of human face perception. I: Potentials generated in occipitotemporal cortex 48. Yakovlev, V., Fusi, S., Berman, E. & Zohary, E. Inter-trial neuronal activity in
by face and non-face stimuli. Cereb. Cortex 9, 415430 (1999). inferior temporal cortex: a putative vehicle to generate long-term visual
30. Klopp, J., Halgren, E., Marinkovic, K. & Nenov, V. Face-selective spectral associations. Nat. Neurosci. 1, 310317 (1998).
changes in the human fusiform gyrus. Clin. Neurophysiol. 110, 676682 (1999). 49. Miyashita, Y. Neuronal correlate of visual associative long-term memory in
31. Halgren, E., Raij, T., Marinkovic, K., Jousmaki, V. & Hari, R. Cognitive the primate temporal cortex. Nature 335, 817820 (1988).
response profile of the human fusiform face area as determined by MEG. 50. Green, D. & Swets, J. Signal Detection Theory and Psychophysics (Wiley, New
Cereb. Cortex 10, 6981 (2000). York, 1966).

articles
Motion distorts visual space:

shifting the perceived position of
remote stationary objects
David Whitney and Patrick Cavanagh
Vision Sciences Laboratory, Harvard University, 33 Kirkland Street, Cambridge, Massachusetts 02138, USA
Correspondence should be addressed to D.W. (whitney@wjh.harvard.edu)
To perceive the relative positions of objects in the visual field, the visual system must assign locations
to each stimulus. This assignment is determined by the objects retinal position, the direction of
gaze, eye movements, and the motion of the object itself. Here we show that perceived location is
also influenced by motion signals that originate in distant regions of the visual field. When a pair of
stationary lines are flashed, straddling but not overlapping a rotating radial grating, the lines appear
displaced in a direction consistent with that of the gratings motion, even when the lines are a
substantial distance from the grating. The results indicate that motions influence on position is not
restricted to the moving object itself, and that even the positions of stationary objects are coded by
mechanisms that receive input from motion-sensitive neurons.
The visual system integrates information from multiple sources RESULTS

to judge the relative positions of objects in the visual field. These The first experiment examined the distance over which the
sources include the position and movement of the eye14, the reti- motion of a grating influenced the perceived positions of sta-
nal location and motion of the object513, and vestibular or pro- tionary flashed lines. Two physically aligned flashes that strad-
prioceptive signals14. For example, flashing a stimulus just before dled a rotating radial grating (Fig. 1a) appeared to be misaligned
the initiation of a saccade leads to presaccadic mislocalization1, in a direction consistent with that of the nearest motion. To avoid
bidirectional spatial compression2, and relative position shifts3 adaptation, the grating reversed direction every 2.25 0.5 sec-
of stationary stimuli. onds (Fig. 1b). The flashes were presented within 1 second of
In addition to eye movements, the movement of the object the reversal.
itself strongly influences its perceived position. For example, A vertical misalignment was perceived between the two
the initial5 or final15 position of a moving object can appear flashed lines (Fig. 2a) when the configuration in Fig. 1b was pre-
shifted in the direction of motion, the position of a moving sented. The misalignment is plotted as a function of the differ-
stimulus can appear to lead a stationary flashed target79, and a ence in time (interstimulus asynchrony or ISA) between the
stationary patch filled with moving texture appears not only to reversal of the moving grating and the presentation of the flashes.
move, but also to be positionally shifted in the direction of the The separation between the inner edge of the flashes and the
apparent motion10,11. In all these motion-based phenomena, nearest outer edge of the radial grating was three degrees. When
the perceived position of an object is shifted in the direction of a sigmoid was fit to these data (y = a((exp(b(x + c)) 1)/
its apparent motion, which suggests that the locations assigned (exp(b(x+ c)) + 1)) + d, where the parameter a estimates the height
to stimuli interact with their motion signals. Many explana- of the function along the ordinate), the overall misalignment (2a)
tions argue that the locations of moving and stationary stimuli was found to be about 20 minutes of arc for both subjects.
are processed differently511,1620. However, if the motion of one Note that although the flashes appeared misaligned in a
object shifts the position of another, stationary object, then the direction consistent with that of the nearest motion, a misalign-
position shifts found for moving stimuli may actually reflect a ment was present at 0 ISA, when the grating was physically sta-
more fundamental and generalized position-coding mecha- tionary (Fig. 2a). This would arise if there were a longer delay
nism that analyzes moving and stationary stimuli alike. for assigning positions to the flashes than for registering the
Here we show that when a moving stimulus is presented in gratings motion16,17.
one region of visual space, stationary flashes that are briefly pre- There was a noticeable effect of grating size on the misalign-
sented concurrently in remote areas of the visual field appear to ment, but almost no effect of flash eccentricity on the magni-
be displaced in the direction of motion. Unlike previously tude of the illusion (Fig. 2b). The misalignment remained
reported phenomena1013, the displacement occurs even when roughly constant with increasing flash-to-grating separation.
the flashes do not appear to move and are physically separated This clearly demonstrates that motion in one region of the
from the motion. This provides direct evidence that the local- visual field can directly influence the perceived positions of sta-
ization of a physically and perceptually stationary stimulus tionary objects at distant locations.
depends on motion-processing mechanisms that are active even Consistent with previous studies21, the threshold flash mis-
at some distance from the stimulus. alignment (vernier) rose with increasing eccentricity of the

articles
a This misalignment, produced at such a distance from the

motion of the grating, might suggest that rotational eye move-
ments (for example, torsion) are responsible for the effect22.
Small torsional eye movements can be induced by rotary
motion23,24, such as the grating in the first experiment, so it is
possible that the misalignment observed between the two
flashed lines is simply a product of small compensatory tor-
sional adjustments25. In a second experiment, we rejected this
explanation by presenting two pairs of linear gratings that
b moved in opposite directions (Fig. 4a). Three flashed lines were
then presented: two straddled the outer gratings, and one was
superimposed on the fixation point (Methods). If eye move-
ments of any kind were responsible for the misalignment illu-
sion, we would then expect to find no perceived misalignment
between the two outer flashes and the central flash in Fig. 4a,
because the eye cannot move in two directions simultaneously;
we verified the results both monocularly and binocularly. A
misalignment was still observed between the central and two
outer flashed lines, in contrast to the eye-movement hypothesis

(Fig. 4b). Similar results were also obtained when the experi-
ment was replicated using two radial gratings that rotated in
opposite directions.
The results of the second experiment also addressed an
Fig. 1. A schematic view of the stimulus configuration and perception. alternative explanation for the misalignment based on a frame
(a) An illusory misalignment between two physically aligned flashes of reference effect26,27. For example, if the radial grating in the
occurred when they were presented on either side of a rotating radial first experiment created an apparent tilt in some frame of refer-
grating. The apparent misalignment was consistent with the direction of
ence, such as the monitor, then the flashes could have appeared
motion and occurred even when the flashed lines were distantly sepa-
rated from the rotating grating. (b) The stimulus configuration used in displaced relative to this tilted frame. However, because the
the first experiment. The two flashed lines were presented at various gratings moved in opposite directions in the second experiment
interstimulus asynchronies (ISA) before or after the radial grating and still produced the apparent misalignment, the frame-of-
reversed direction. See http://visionlab.harvard.edu/ for demonstrations. reference explanation was effectively ruled out.
Given the remote nature of motions effect on position judg-
ments, as found in experiments 1 and 2, we were interested in
examining the possibility that higher-level motion processes
flashes (Fig. 3). This trend contrasts with the largely constant might be contributing to precise perceptual localization. In a
misalignment we measured across eccentricities. One conse- third experiment, we tested this by using a dichoptic display:
quence of the difference in these functions is that the visibility of the radial grating from the first experiment was presented to
the misalignment caused by the motion of the grating was below one eye while the flashed lines were presented to the other eye.
threshold when the flashed lines were at large eccentricities. If the motion of the grating in one eye affects the perception of
Fig. 2. Experiment 1 results for subjects DW and a

EV. (a) The ordinate shows the perceived mis-
alignment between the two flashes, as measured
by the number of arc minutes that the flashes had
to be displaced to make them appear aligned
(Methods). The abscissa shows the time (ISA)
between the presentation of the test flashes and
the gratings reversal. (Negative ISA indicates that
the flashes were presented before the reversal of
motion.) Data have been merged so that the
motion of the grating is clockwise, then counter-
clockwise, as indicted along the abscissa. Each of
the seven data points represents a single psycho- b
metric function (not shown). The separation
between the inner edge of each flash and the near-
est outer edge of the grating was three degrees. A
sigmoid (see text) was fit to the data to measure
the overall misalignment observed (the height of
the sigmoid). Error bars, s.e.m. (b) Overall mis-
alignment for three gratings of different sizes and
various flash eccentricities. Insets, outer edges of
the radial gratings. The distance between each of
the data points and the horizontally adjacent inset
indicates the flash-to-grating separation. Each of the data points on these graphs represents the height of a sigmoid like that in (a). The error bars show
representative s.e.m for each flash-to-grating separation. Data for a third subject were similar.

articles
Fig. 3. Thresholds from

at flash durations where there was still a misalignment present.
experiment 1 for sub-
jects DW and EV. This pattern of results was consistent across all grating sizes and
Threshold misalignments flash-to-grating separations.
(minutes of arc) are plot- Reported displacements in position are always accompanied
ted for each of the by perceived motion of the target whose position is shifting13. The
points on Fig. 2b as a present misalignment effect, however, was seen for stimuli that are
function of the flash neither physically nor perceptually moving. Furthermore, the
eccentricity. Thresholds magnitude of the misalignment observed in this experiment was
were calculated as half much larger than that for many previous phenomena1013. The
the distance between
misalignment also remained constant with increasing eccentric-
the 25th and 75th per-
centiles on each psycho-
ity, which contrasts with the only previous comparable measure-
metric function, then ment showing a position shift of an apparently moving stimulus
averaged across ISA. The that increases substantially with eccentricity11.
threshold misalignments The finding that the misalignment was observed between
were similar whether or flashes that were stationary rules out anticipatory retinal
not the radial grating responses20 and latency variations1618 as explanations for the
was presented. mislocalization. In addition, because the flashes in our experi-
ments were not moving physically or perceptually, they could
not trigger an extrapolation mechanism that would shift the

apparent locations of moving stimuli in the direction of motion
flash alignment from the other eye, it is because motion infor- to compensate for the lag in their perceived positions caused by
mation is reaching cortical areas with binocularly driven neu- unavoidable neural delays7,911,20.
rons before influencing the processing of the flashes. The results We next asked whether the misalignment was due to a local
with dichoptic presentation (Fig. 5) were very similar to those mechanism that acts on each of the two flashes independently
of the first experiment (Fig. 2a), indicating that the binocularly or on a configural process that requires a comparison between
driven cortical neurons that responded to the moving grating positions that span across a moving stimulus. In this fifth
also influenced the assignment of the flashes relative positions. experiment, we tested whether the flashes were mislocalized in
We next investigated the relationship of this flash misalign- a local manner by presenting one of the flashes from Fig. 1b
ment to previously reported phenomena, such as the position
shifts of perceptibly moving stimuli. When a moving pattern is
viewed through a stationary aperture, the aperture not only a
seems to move, but also appears displaced in the direction of
motion (when the mean luminance inside and outside the aper-
ture are equated)10,11. This phenomenon has been interpreted as
possible evidence of an extrapolation10,11 or motion-capture
mechanism10,28. In both these cases, motion signals are assigned
to the aperture. This illusory motion of the aperture may then
cause it to appear shifted in the direction of motion in the same
way that the above motion-based position displacements occur.
To investigate whether the underlying mechanism for our
misalignment effect is also mediated by an illusory motion, we
measured the perceived speed of the stationary flashes in a
fourth experiment. If the misalignment is produced even
though the flashes appear stationary, this would argue against
b
explanations that have been proposed for previously reported
phenomena, as these explanations require the stimulus to move
to appear displaced. The flashes appeared stationary when they
were presented for longer than 120 ms (Fig. 6), and yet these
perceptually and physically stationary flashes still appeared
substantially misaligned. The perceived flash speed increased at
shorter durations, where judgments are very noisy29. At longer
durations, the perceived speed of the flashes was judged more
accurately, and converged convincingly to a stationary percept
Fig. 4. Experiment 2. (a) Schematic view of experiment 2 stimulus and

perception. When three flashed lines were presented in physical align-
ment, the flashes appeared misaligned consistent with the direction of
the nearest motion (white flashes). (b) Results for subjects DW and EV
(squares). (For comparison, the circles indicate the results for a compa-
rable stimulus in the first experiment.) The ordinate shows the per-
ceived misalignment between the inner and outer flashes as measured
by a nulling procedure (Methods). The abscissa shows the time between
the flashes and the motion reversal (ISA). Error bars, s.e.m.

articles
Fig. 5. Experiment 3
flashes and gratings increased (Fig. 7). These results were con-
results for sub
jects DW and EV sistent with the previous experiment in that a misalignment
(squares). (The cir- depended largely on the presence of motion in the space
cles show results for spanned by the comparison flashes.
a comparable stimu- We manipulated the gratings temporal frequency in a sev-
lus in the first enth experiment to measure the velocity dependence of the
experiment, where flash misalignment and to test whether it is due to attentional
viewing was binocu- tracking30 of the moving gratings. The misalignment was
lar.) The ordinate roughly band-pass, which resembles the temporal contrast sen-
shows the perceived
sitivity function31 (Fig. 8). Interestingly, the misalignment does
misalignment
between the two
not appear similar to the low-pass characteristics of the motion
flashes (minutes of sensitivity function32. Because the misalignment occurred even
arc). The abscissa at temporal frequencies above 12 Hz, the mislocalization was
shows the time not due to attentional tracking of the grating, as the ability to
between the presen- track positions on a rotating grating falls off dramatically at
tation of the flashes around 7 Hz33.
and the radial grat-
ings reversal (ISA). DISCUSSION
Note that the flashes We showed that motion information in one region of the visu-
were presented to
al field influenced the perceived positions of apparently sta-
one eye while the radial grating was presented to the other eye. The
data indicate that there was still a perceived misalignment between the tionary objects, even when those objects were located in distant
two flashes, and that it was comparable in magnitude to the first two areas of the visual scene. The perceived misalignment was not
experiments. Error bars, s.e.m. due to eye movements, a frame-of-reference effect or attentional
tracking. The magnitude of the misalignment depended on the
size of the moving stimulus and the duration of the flashes, but
it was always in the direction of the nearest motion. These
twice at the same location (that is, on only one side of the radial experiments suggest the involvement of higher cortical areas
grating), once before and once after the reversal of the motion; that are binocularly driven and specialized in the processing of
the interval between flashes was varied. If the misalignment motion. In contrast to previous results713,20, the misalignment
illusion were due to a local mechanism that acts on a single flash occurred even when the flashes appeared stationary, showing
independently, then the two successive flashes should appear that position displacement does not depend upon physical or
displaced from each other as they are shifted first one way and even apparent motion of the test stimulus, although, of course,
then the other. For example, following the data for the medium motion must be present in the display.
grating from experiment 1, if a single flash is presented 900 ms The remote effect of motion on localization, as reported
before and 300 ms after the reversal, the two flashes should here, indicated that assigning positions to brief stimuli depends
appear displaced from each other by about 30 to 35 minutes of on the configuration of motion signals throughout the visual
arc (derived from the overall misalignment divided by 2 in field. These long-distance interactions were similar in range to
Fig. 2b). This did not occur, however. The local shift in this sin- that for motion capture28 and induced motion34 (motion assim-
gle location test was 4 and 6 minutes for subjects DW and EV, ilation and contrast), where motion at one location can influ-
respectively, about one-sixth of the comparable values in exper- ence perceived motion at some distance3440. Yet, motion
iment 1, which suggests that much of the misalignment was capture and induced motion differ from the position shifts
produced only when two locations that straddled a moving observed in this study on at least two factors. First, the induced
stimulus were compared. motion phenomena generally fall off with increasing separation
To explore this further, we conducted a sixth experiment, in between the moving inducer and the induced target40, a pattern
which a pair of flashes were always presented on the same hori- that did not hold for the misalignment reported here. Second,
zontal plane as the fixation point, while a pair of linear gratings and, more importantly, the misalignment we report occurred
were vertically offset above or below this plane. The misalign- even when the flashes appeared stationarywhen no motion
ment dropped off rapidly as the vertical distance between the assimilation or contrast had occurred.
Position shifts of apparently moving stimuli are well docu-
mented and have garnered a number of explanations including
extrapolation7,9, attention shifts8, differential latencies1618, inte-
gration or interpolation of the moving objects trajectory19, and
anticipatory retinal responses20. Yet, none of these diverse mod-
els can explain how an apparently stationary stimulus could be
displaced in position by motion in a remote area of the visual
field. This is primarily because these models are intended to
address the question, How is a moving stimulus coded differ-
ently from a stationary one? Because apparently stationary
Fig. 6. Experiment 4 results. (a) The overall flash misalignment (as in
Fig. 2b) is plotted as a function of flash duration for DW and EV. (b) stimuli can be shifted in position, however, this may not be the
The perceived speed of the flashes is plotted as a function of flash dura- right question to ask. The misalignment reported here could
tion. To measure the perceived speed of the flashes, they were physically reflect a more basic mechanism that underlies or contributes to
moved at velocities that nulled any apparent motion. Error bars, s.e.m, many of the motion-based position displacement phenomena
shown only when they are larger than the symbols. mentioned earlier. The issue, then, is not the dissociation

articles
for the size43 and motion of a stimulus4146. Given the strong

feedback connections from these areas to V1 (ref. 47), where
retinotopic localization is very precise11,48, it is possible that the
misalignment reflects a re-entrant mechanism13,47,49 by which
motion information influences position judgments of moving
and stationary stimuli. This mechanism subserves both station-
ary and moving stimuli and may therefore underlie a number
of visual phenomena that involve localizing a stimulus in the
presence of motion.
METHODS
Stimuli were presented on a high-resolution CRT monitor (832 624
pixels, 75 Hz refresh) controlled by a computer (Apple Power Macin-
tosh, Cupertino, California). Subjects were immobilized with a chin rest
12 cm from the visual display, unless otherwise noted. A fixation bulls-
eye was provided at the center of the screen. DW and two well practiced,
naive subjects participated in the experiments. Subjects had normal or
corrected-to-normal visual acuity.
Experiment 1. The radial grating subtended 8.9, 23.6 or 47.2 in diam-

eter, with a 2.05 hole in the center for the fixation bulls-eye. The grat-
ing had a sinusoidal luminance modulation of 8 cycles per rotation at
98.5% contrast on a dark (0.01 cd/m2) background. Each experimental
trial consisted of clockwise or counterclockwise rotation of the grating
(1.4 Hz) for 2.25 0.5 seconds, followed by an equivalent rotation in the
opposite direction. At varying periods of time before or after the gratings
reversal (ISA), two flashed lines were presented simultaneously for 60 ms
on either side of the grating. Each flash (34.5 cd/m2) was 6 0.85. The
flashes were vertically offset from one another, and subjects judged
whether the flash on the right appeared above or below the flash on the left
(method of constant stimuli, two-alternative, forced-choice task). A psy-
chometric function was fit to the data for each of the various ISAs, and
Fig. 7. Experiment 6 results. The ordinate shows the overall perceived
the physical misalignment between the flashes that created an apparent
misalignment between the two flashes (as in Fig. 2b). The abscissa shows
alignment was measured. The initial direction of motion was random-
the vertical separation between the flashes and the linear grating, as mea-
ized across trials in either a clockwise or counterclockwise direction. The
sured from the center of the grating to the horizontal plane on which the
data from the two directions were then merged. The separation between
flashes were presented. A schematic view of the linear gratings is shown
the inner edge of each flash and the nearest outer edge of the radial grat-
along the abscissa (though the linear gratings were vertically, not horizon-
ing was varied across experimental sessions between 1.3 and 58.
tally, offset in the experiment). The linear gratings were centered between
the flashes in the 0.0 deg separation condition. A single error bar is shown
Experiment 2. Four linear gratings were presented behind rectangular aper-
for one data point in each plot to indicate a representative s.e.m.
tures, two on either side of a fixation point. The gratings had sinusoidal
luminance modulations of 0.2 cycles per degree visual angle at 99.5% con-
trast on a dark (0.01 cd/m2) background. Each linear grating was 20.5 high
and 3.6 wide. The outer two gratings were centered 8.98 left and right of
between the coding of stationary and moving stimuli, but how the fixation point, which was at the center of the visual display. The inner
the configuration of motion in the visual field influences the two gratings were centered 3.85 left and right of the fixation point. The
localization of both moving and stationary stimuli. inner two gratings always translated in a vertical direction opposite the
A potential criticism of this interpretation is that the magni-
tude of the misalignment reported here is different from that
reported for other phenomena. For example, the flash-lag mis-
alignment (in which a moving object appears to lead a flash
presented at an adjacent location), is often much greater than
the flash-to-flash misalignment we measured for the same
flash-to-motion separation. Indeed, given that the mislocaliza-
tion of the flash that occurred in our stimulus should also occur
in the flash-lag studies, it should reduce the flash-lag phenome-
non by shifting the flash in the direction of the moving stimu-
lus. The actual mislocalization of the moving bar in flash-lag
experiments may therefore be greater than those experiments
revealed. Further, if the mislocalization mechanism that we
report applies to moving objects as well as to stationary ones,
then it must shift the perceived position of moving stimuli by
an even greater amount than the flashed stimuli.
The experiments reported here suggest the involvement of Fig. 8. Experiment 7 results. The overall perceived misalignment
cortical areas, such as MT or MST, whose neurons have between the two flashes is plotted as a function of the gratings tempo-
large4143 binocularly driven receptive fields41,44 that are selective ral frequency (Hz). Error bars, s.e.m.

articles
outer two gratings, and the initial direction of the inner gratings was ran- 12. Snowden, R. J. Shifts in perceived position following adaptation to visual
domized across trials (upward or downward). As in the first experiment, motion. Curr. Biol. 8, 13431345 (1998).
13. Nishida, S. & Johnston, A. Influence of motion signals on the perceived
the gratings translated (2.67 Hz) for 2.25 0.5 seconds before reversing position of spatial pattern. Nature 397, 610612 (1999).
direction and translating for an equal period of time. Two flashed lines 14. Schlag, J., Cai, R., Dorfman, A., Mohempour, A. & Schlag-Rey, M.
(2.05 0.25) were presented straddling the outer gratings at various ISAs Extrapolating movement without retinal motion. Nature 403, 3839 (2000).
before or after the reversal of direction. A third flash (1.025 0.125) was 15. Freyd, J. J. & Finke, R. A. Representational momentum. J. Exp. Psychol. 19,
presented superimposed on the fixation point. All flashes were presented 369401 (1987).
16. Purushothaman, G., Patel, S. S., Bedell, H. E. & Ogmen, H. Moving ahead
simultaneously for 60 ms. For the two outer flashes, the separation between through differential visual latency. Nature 396, 424 (1998).
the inner edge of each flash and the outer edge of the nearest grating was 17. Whitney, D. & Murakami, I. Latency difference, not spatial extrapolation.
1.54. The two outer flashes were vertically misaligned relative to the fixed Nat. Neurosci. 1, 656657 (1998).
position of the center flash, and subjects were asked to judge the direction of 18. Whitney, D., Murakami, I. & Cavanagh, P. Illusory spatial offset of a flash
relative to a moving stimulus is caused by differential latencies for moving
misalignment. As in the first experiment, the physical flash misalignment and flashed stimuli. Vision Res. 40, 137149 (2000).
that created an apparent alignment was measured. The viewing distance 19. Eagleman, D. M. & Sejnowski, T. J. Motion integration and postdiction in
was 40 cm. visual awareness. Science 287, 20362038 (2000).
20. Berry, M. J. II, Brivanlou, I. H., Jordan, T. A. & Meister, M. Anticipation of
Experiment 3. A single radial grating, identical to the 8.9 grating from moving stimuli by the retina. Nature 398, 334338 (1999).
21. Levi, D. M. & Klein, S. A. The role of separation and eccentricity in encoding
the first experiment, was presented on the left side of the monitor with a position. Vision Res. 30, 557585 (1990).
fixation bulls-eye at its center. On the right side of the monitor another 22. Dichgans, J., Held, R., Young, L. R. & Brandt, T. Moving visual scenes
fixation bulls-eye was presented. A haploscope was provided to assist influence the apparent direction of gravity. Science 178, 12171219 (1972).
subjects in fusing the two fixation points. The procedure was identical 23. Kertesz, A. & Jones, R. The effect of angular velocity of stimulus on human
to the first experiment except that the flashes (each 2.7 0.27) were torsional eye movements. Vision Res. 9, 995998 (1969).
24. Merker, B. H. & Held, R. Eye torsion and the apparent horizon under head tilt
presented on the right side of the monitor, centered on the right fixation and visual field rotation. Vision Res. 21, 543547 (1981).
bulls-eye. The flashes were separated by 12.83. When the two fixation 25. Hughes, P. C., Brecher, G. & Fishkin, S. Effects of rotating backgrounds upon
points were fused, the inner edge of each flash was 1.3 from the nearest the perception of verticality. Percept. Psychophys. 11, 135138 (1972).
outer edge of the radial grating. The procedure was identical to that in 26. Witkin, H. & Asch, S. Studies in space orientation. IV. Further experiments
on perception of the up-right with displaced visual fields. J. Exp. Psychol. 38,
the first experiment, except that only one flash-to-grating separation was 762782 (1948).
tested. The viewing distance was 40 cm. 27. Rock, I. in The Legacy of Solomon Asch (ed. Rock, I.) 243268 (Lawrence
Erlbaum, Hillsdale, New Jersey, 1990).
Experiment 4. A pair of linear gratings (each 17.4 5.8) translated in 28. Ramachandran, V. S. Interaction between colour and motion in human
opposite directions at the center of the display. A pair of flashes (each 6 vision. Nature 328, 645647 (1987).
29. Cropper, S. J. & Derrington, A. M. Motion of chromatic stimuli: first-order or
0.85) were presented straddling the gratings for various durations second-order? Vision Res. 34, 4958 (1994).
between 30 and 300 ms. The inner edge of each flash was 9.9 from the 30. Cavanagh, P. Attention-based motion perception. Science 257, 15631565
nearest linear gratings outer edge. The procedure for measuring the (1992).
apparent flash misalignment was identical to that in the first experiment. 31. Kelly, D. H. Motion and vision. II. Stabilized spatio-temporal threshold
surface. J. Opt. Soc. Am. 69, 13401349 (1979).
To measure the perceived speed of the flashes, we physically moved the 32. Nakayama, K. Biological image motion processing: a review. Vision Res. 25,
flashes in opposite vertical directions. Subjects judged whether the flash- 625660 (1985).
es moved clockwise or counterclockwise (although they only moved ver- 33. Verstraten, F., Cavanagh, P. & Labianca, A. Limits of attentive tracking reveal
tically and did not rotate). As in the first experiment, a psychometric temporal properties of atention. Vision Res. (in press).
function was fit to the data, and the point of subjective equality (where 34. Duncker, K. in Source Book of Gestalt Psychology (ed. and trans. Ellis, W. D.)
161172 (Routledge & Kegan Paul, London, 1938). [Reprinted from
the flashes appeared stationary) was measured. Psychologische Forschung 12, 180259, 1929].
35. Anstis, S. M. & Reinhardt-Rutland, A. H. Interactions between motion
Experiment 6. The stimulus was identical to that in experiment 4 (flash aftereffects and induced movement. Vision Res. 16, 13911394 (1976).
duration, 60 ms). The flashes were always presented on a horizontal plane 36. Nakayama, K. & Silverman, G. Temporal and spatial characteristics of the upper
displacement limit for motion in random dots. Vision Res. 24, 293300 (1984).
with the fixation. The pair of linear gratings were vertically displaced
37. Snowden, R. J. & Braddick, O. J. Extension of displacement limits in multiple-
either above or below this plane randomly across trials. The vertical sep- exposure sequences of apparent motion. Vision Res. 29, 17771787 (1989).
aration between the center of the linear gratings and the flashes varied 38. Nawrot, M. & Sekuler, R. Assimilation and contrast in motion perception:
from 0.0 to 35. The procedure for measuring the apparent flash mis- explorations in cooperativity. Vision Res. 30, 14391451 (1990).
alignment was identical to that in the first experiment. 39. Reinhardt-Rutland, A. H. Induced movement in the visual modality: An
overview. Psychol. Bull. 103, 5771 (1988).
40. Gogel, W. C. & Koslow, M. The adjacency principle and induced movement.
Percept. Psychophys. 11, 309314 (1972).
RECEIVED 19 MAY; ACCEPTED 19 JULY 2000 41. Felleman, D. J. & Kaas, J. H. Receptive-field properties of neurons in middle
temporal area (MT) of owl monkeys. J. Neurophysiol. 52, 488513 (1984).
42. Mikami, A., Newsome, W. T. & Wurtz, R. H. Motion selectivity in macaque
1. Matin, L. in Handbook of Sensory Physiology VII/4 (eds. Jameson, D. & visual cortex. II. Spatiotemporal range of directional interactions in MT and
Hurvich, L.) 331380 (Springer, Berlin, 1972). VI. J. Neurophysiol. 55, 13281339 (1986).
2. Ross, J., Morrone, C. & Burr, D. C. Compression of visual space before 43. Tanaka, K. & Saito, H. Analysis of motion of the visual field by direction,
saccades. Nature 386, 598601 (1997). expansion/contraction, and rotation cells clustered in the dorsal part of the
3. Cai, R. H., Pouget, A., Schlag-Rey, M. & Schlag, J. Perceived geometrical medial superior temporal area of the macaque monkey. J. Neurophysiol. 62,
relationships affected by eye-movement signals. Nature 386, 601604 (1997). 626641 (1989).
4. Deubel, H., Schneider, W. X. & Bridgeman, B. Postsaccadic target blanking 44. Maunsell, J. H. & Van Essen, D. C. Functional properties of neurons in
prevents saccadic suppression of image displacement. Vision Res. 36, 985996 middle temporal visual area of the macaque monkey. II. Binocular
(1996). interactions and sensitivity to binocular disparity. J. Neurophysiol. 49,
5. Frohlich, F. W. Die Empfindungszeit (Verlag von Gustav Fischer, Jena, 1929). 11481167 (1983).
6. Matin, L., Boff, K. & Pola, J. Vernier offset produced by rotary target motion. 45. Maunsell, J. H. & Van Essen, D. C. Functional properties of neurons in middle
Percept. Psychophys. 20, 138142 (1976). temporal visual area of the macaque monkey. I. Selectivity for stimulus
7. Nijhawan, R. Motion extrapolation in catching. Nature 370, 256257 (1994). direction, speed, and orientation. J. Neurophysiol. 49, 11271147 (1983).
8. Baldo, M. V. & Klein, S. A. Extrapolation or attention shift? Nature 378, 46. Tootell, R. B. et al. Visual motion aftereffect in human cortical area MT revealed
565566 (1995). by functional magnetic resonance imaging. Nature 375, 139141 (1995).
9. Khurana, B. & Nijhawan, R. Extrapolation or attention shift? reply. Nature 47. Zeki, S. in Coding and Efficiency (ed Blakemore, C.) 321345 (Cambridge
378, 566 (1995). Univ Press, Cambridge, 1990).
10. Ramachandran, V. S. & Anstis, S. M. Illusory displacement of equiluminous 48. Daniel, P. M. & Whitteridge, D. The representation of the visual field on the
kinetic edges. Perception 19, 611616 (1990). cerebral cortex in monkeys. J. Physiol. (Lond.) 159, 203221 (1961).
11. DeValois, R. L. & DeValois, K. K. Vernier acuity with stationary moving 49. Edelman, G. M. in The Mindful Brain (eds. Edelman, G. M. & Mountcastle, V. B.)
gabors. Vision Res. 31, 16191626 (1991). 55100 (MIT Press, Cambridge, Massachusetts, 1978).

Nature Neuroscience September 2000 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Nature Neuroscience September 2000 PDF

Uploaded by

Copyright:

Available Formats

2000 Nature America Inc. http://neurosci.nature.

volume 3 no 9 september 2000

The cover shows the dendritic

CNTF II, I presume?. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 851

Probing the olfactory code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853

The mechanics behind spines on the move . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 856

Glutamate receptors and actin Thinking about feeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 857

REVIEWED BY ZACHARY F. MAINEN

nature neuroscience volume 3 no 9 september 2000 i

A Benoit de Coignac, Y Delneste, J-Y Bonnefoy, J-F Gauchat and H Gascan

Somatic EPSP amplitude is independent of synapse location in

A network of electrically coupled interneurons drives synchronized

Fixation neurons in the superior colliculus encode distance between

Motion distorts visual space: shifting the perceived position of

nature neuroscience volume 3 no 9 september 2000 ii

A debate over fMRI data sharing

nature neuroscience volume 3 no 9 september 2000 845

846 nature neuroscience volume 3 no 9 september 2000

letters to the editor

Does a stretch-inactivated cation channel

nature neuroscience volume 3 no 9 september 2000 847

letters to the editor

848 nature neuroscience volume 3 no 9 september 2000

news and views

Distant synapses raise their voices

nature neuroscience volume 3 no 9 september 2000 849

news and views

common to all neu- Na+ and Ca2+ channels, essentially can-

850 nature neuroscience volume 3 no 9 september 2000

news and views

Evidence for this elusive CNTF II

The authors are in the Department of

nature neuroscience volume 3 no 9 september 2000 851

news and views

852 nature neuroscience volume 3 no 9 september 2000

news and views

throughout the animal kingdom7 sug-

nature neuroscience volume 3 no 9 september 2000 853

news and views

whereas stimulus properties ferent olfactory systems, results from

854 nature neuroscience volume 3 no 9 september 2000

news and views

into one of nine categories: drawings or

The author is at the Department of Psychology,

nature neuroscience volume 3 no 9 september 2000 855

news and views

7. Warrington, E. K. & McCarthy, R. A. Brain

856 nature neuroscience volume 3 no 9 september 2000

nature neuroscience volume 3 no 9 september 2000 859

a um-dependent refilling is reported as well9. One possibility is

RECEIVED 25 APRIL; ACCEPTED 28 JULY 2000

860 nature neuroscience volume 3 no 9 september 2000

Percent improvement in mean motor timing accuracy

trained interval. This prediction was sup-

trained perceptual interval and target motor

nature neuroscience volume 3 no 9 september 2000 861

1. Karni, A. & Bertini, G. Curr. Opin. Neurobiol. 7, 530535 (1997).

862 nature neuroscience volume 3 no 9 september 2000

Should the neuroscience community make a

nature neuroscience volume 3 no 9 september 2000 863

864 nature neuroscience volume 3 no 9 september 2000

nature neuroscience volume 3 no 9 september 2000 865

CLF associates with CLC to form a