DOI: 10.1126/science.1139597
, 1609 (2007);316 Science
et al.Thilo Womelsdorf
Synchronization
Modulation of Neuronal Interactions Through Neuronal
This copy is for your personal, non-commercial use only.
clicking here.colleagues, clients, or customers by
, you can order high-quality copies for yourIf you wish to distribute this article to others
here.following the guidelines
can be obtained byPermission to republish or repurpose articles or portions of articles
): November 13, 2012 www.sciencemag.org (this information is current as of
The following resources related to this article are available online at
http://www.sciencemag.org/content/316/5831/1609.full.html
version of this article at:
including high-resolution figures, can be found in the onlineUpdated information and services,
http://www.sciencemag.org/content/suppl/2007/06/12/316.5831.1609.DC1.html
can be found at: Supporting Online Material
http://www.sciencemag.org/content/316/5831/1609.full.html#related
found at:
can berelated to this article A list of selected additional articles on the Science Web sites
http://www.sciencemag.org/content/316/5831/1609.full.html#ref-list-1
, 12 of which can be accessed free:cites 27 articlesThis article
112 article(s) on the ISI Web of Sciencecited by This article has been
http://www.sciencemag.org/content/316/5831/1609.full.html#related-urls
79 articles hosted by HighWire Press; see:cited by This article has been
http://www.sciencemag.org/cgi/collection/neuroscience
Neuroscience
subject collections:This article appears in the following
registered trademark of AAAS.
is aScience2007 by the American Association for the Advancement of Science; all rights reserved. The title
CopyrightAmerican Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005.
(print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by theScience
on November 13, 2012www.sciencemag.orgDownloaded from
Beringia 2005 expeditions funded by the Polar Research
Secretariat at the Royal Swedish Academy of Sciences.
The main work was funded by grant 150322/720 to
Brochmann from the Research Council of Norway.
Additional grants to Westergaard were obtained from
K. and H. Jakobsens Fund, King Haakon VII Educational
Fund, Roald Amundsen’s Centre for Arctic Research,
KOMETEN, and Tromsø University Museum.
Supporting Online Material
www.sciencemag.org/cgi/content/full/316/5831/1606/DC1
Materials and Methods
Tables S1 and S2
References
22 December 2006; accepted 2 May 2007
10.1126/science.1139178
Modulation of Neuronal Interactions
Through Neuronal Synchronization
Thilo Womelsdorf,
1
*† Jan-Mathijs Schoffelen,
1
*† Robert Oostenveld,
1
Wolf Singer,
2,3
Robert Desimone,
4,5
Andreas K. Engel,
6
Pascal Fries
1,7
Brain processing depends on the interactions between neuronal groups. Those interactions
are governed by the pattern of anatomical connections and by yet unknown mechanisms that
modulate the effective strength of a given connection. We found that the mutual influence among
neuronal groups depends on the phase relation between rhythmic activities within the groups.
Phase relations supporting interactions between the groups preceded those interactions by a few
milliseconds, consistent with a mechanistic role. These effects were specific in time, frequency,
and space, and we therefore propose that the pattern of synchronization flexibly determines the
pattern of neuronal interactions.
G
roups of activated neurons synchronize
in the gamma-frequency band (30 to
100 Hz), and previous studies have re-
lated gamma-band synchronization to several
cognitive functions (1–6). Yet, if gamma-band
synchronization subserves those functions, it
must have mechanistic consequences for neuro-
nal processing (7). It has been shown that the
precise timing of pre- and postsynaptic activation
determines long-term changes in synaptic
strength (8–10) and that gamma-band synchro-
nization of sy naptic inputs directly enhances their
effective synaptic strength (11–13).
Synchronization between two groups of
neurons is also likely to facilitate interactions
between them (Fig. 1A) (6, 14). Gamma-band
synchronization entails rhythmic inhibition of
the local network (15–17), and the periods
between inhibition provide temporal windows
for neuronal interaction. Two groups of neurons
will therefore probably have a greater influence
on each other when their temporal interaction
windows open at the same times, i.e., when the
rhythmic synchronization within the groups is
also synchronized between the groups. By the
same token, the interaction is probably curtailed
if the temporal interaction windows open either
in an uncorrelated way or consistently out of
phase with each other.
We analyzed four data sets: (i) one from
awake cat area 17, (ii) one combining awake cat
area 18 with area 21a recordings, (iii) one from
awakemonkeyareaV1,and(iv)onefrommon-
key area V4. [Data from two of the three area 17
data sets have been used in (18, 19); the V4 data
set has been used in (3, 20).] In all cases, we
recorded multiunit activity (MUA) and local
field potentials (LFPs) simultaneously from
four to eight electrodes while the neurons were
visually stimulated with moving gratings. From
each data set, we used trials with identical visual
stimulation and behavioral tasks and based our
analysis on the natural fluctuation of neuronal
gamma-band synchronization. For each pair of
neuronal groups, we quantified synchronization
by means of the MUA-MUA phase-coherence
spectrum (Fig. 1B) and the MUA-LFP phase-
coherence spectrum (Fig. 1C) (21).
Phase-coherence spectra showed a peak in
the gamma-frequency band, indicating that phase
relations between signals were not random.
However , phase coherence was far from perfect
(a value of 1.0), but it assumed average peak
values of 0.14 and 0.27 for MUA-MUA and
MUA-LFP combinations, respectively. The
phase relations at 60 Hz in one example MUA-
MUA pair are shown for 708 trials of 250-ms
length (phase-coherence value of 0.06) (Fig. 1D).
The spread of phase relations around their
mean might just be irrelevant noise. Here, how-
ever , we used this spread to actually test for its
potential physiological consequences. W e hypoth-
esized that the mutual influence between two
neuronal groups was a function of their ph ase
relation (Fig. 1A). Phase relations are meaning-
fully defined per frequency, and we hypothesized
that the phase relation at a given frequency
should modulate the interaction among the local
rhythmic activities specifically at that frequency .
We investigated this hypothesis for the ex-
ample pair of recordings sites. We sorted the trials
into six bins according to the 60-Hz phase
relation between the two MUAs (Fig. 1D). For
each phase-relation bin separately, we then quan-
tified the two MUAs’ mutual influence as the
Spearman rank correlation coefficient between
the two MUAs’ 60-Hz power , across the trials in
the bin (Fig. 1E). Fluctuations of 60-Hz power
were most strongly correlated when the 60-Hz
phase relation was close to its mean across the
trials. Specifically, when the gamma-band rhythm
in group A led the one in group B by 2.1 ms
(mean phase relation at 45.8°), the correlation
between each group’s gamma-band power was
four times as strong as when the rhythms were
separated by 10.5 ms (phase relation at 225.8°).
The example pair illustrates this for a case with a
nonzero mean phase to demonstrate that the
effect cannot be ascribed to external artifacts or
volume conduction, but the mean phase rela-
tions across our sample distributed closely
around zero (Fig. 1B).
We performed the same analysis after
replacing one of the MUAs by the LFP recorded
through the same electrode. The mean MUA-
LFP phase relations clustered around 141°
(Fig. 1C), and power correlations were again
substantially enhanced around the mean phase
relation (Fig. 1, F and G). Across our sample,
good phase relations mostly distributed close to
the respective mean phase relations for both
MUA-MUA and MUA-LFP pairs (fig. S1). We
correspondingly dubbed the mean phase rela-
tion as “good” and the opposite phase relation
as “bad,” and we aligned the trial binning to the
good phase relation.
The observed effect was consistent across the
four data sets (Fig. 2 and fig. S2) (140, 86, 111,
and 1 11 MUA-MUA pairs from area 17, areas
18×21a, area V1, and area V4, respectively, and
280, 172, 228, and 237 MUA-LFP pairs from
the same areas). MUA-LFP pairs showed qual-
itatively the same effect as MUA-MUA pairs but
with higher signal-to-noise ratios (Fig. 2B). We
therefore focused our further analyses on MUA-
LFP pairs (recorded from separate electrodes).
The effect was also present for pairs of LFP
and single-unit recordings (fig. S3). The effect
generalized to long-range interactions, because
the analysis of the data set combining cat area
1
F. C. Donders Centre for Cognitive Neuroimaging, Radboud
University Nijmegen, 6525 EN Nijmegen, Netherlands.
2
Department of Neurophysiology, Max Planck Institute
for Brain Research, 60528 Frankfurt, Germany.
3
Frankfurt
Institute for Advanced Studies, Johann Wolfgang Goethe
University, 60438 Frankfurt, Germany.
4
Laboratory of
Neuropsychology, National Institute of Mental Health,
National Institutes of Health, Bethesda, MD 20892, USA.
5
McGovern Institute for Brain Research, Massachusetts
Institute of Technology, Cambridge, MA 02139, USA.
6
Department of Neurophysiology and Pathophysiology,
University Medical Center Hamburg-Eppendorf, 20246
Hamburg, Germany.
7
Department of Biophysics, Radboud
University Nijmegen, 6525 EZ Nijmegen, Netherlands.
*These authors contributed equally to this work.
†To whom correspondence should be addressed. E-mail:
thilo.womelsdorf@fcdonders.ru.nl (T.W.); jan.schoffelen@
fcdonders.ru.nl (J.-M.S.)
www.sciencemag.org SCIENCE VOL 316 15 JUNE 2007 1609
REPORTS
on November 13, 2012www.sciencemag.orgDownloaded from
18 recordings with area 21a recordings was
restricted to interareal pairs of recording sites
(Fig. 2D).
We did not attempt to relate the two signals’
power correlation to their coherence, because
the coherence measure confounds phase syn-
chronization and power correlation. In contrast,
we related the correlation between the two
signals' power across trials directly to their rela-
tive phase. The presence of coherence (Fig. 1H)
does not necessarily result in a phase-relation–
dependent power correlation (Fig. 1I) and thus,
the demonstration of phase-relation–dependent
power correlation goes beyond the demonstra-
tion of coherence.
Fig. 1. Precise timing between rhythmic
neuronal activities determines the strength of
their mutual influence. (A)Sketchofthree
groups of neurons, each rhythmically active
(LFP oscillations with spikes in troughs). Time
windows for effective communication are
either aligned (red and blue group) or not
aligned (red and gray group). (B and C)
Average phase-coherence spectrum across all
(B) MUA-MUA and (C) MUA-LFP pairs (area 17
data) and corresponding distributions of mean
phase relations at 60 Hz. (D) Trialwise phase
relations from an example MUA-MUA pair.
Phase relations were sorted into bins (light and
dark gray ring segments) aligned to the mean
phase relation (red line). (E) Spearman rank
correlation coefficients between the two MUAs’
60-Hz power as a function of their phase
relation. (The solid line indicates a cosine fit.)
(F and G) Same as (D) and (E), but with one
MUA substituted by the respective LFP. (H and
I) E xample MUA-LFP pair from the area 17
data set demonstrating that coherence does
not necessarily result in phase-relation–
dependent power correlations. (H) Coherence
with a clear peak around 60 Hz. (I) Power
correlations as a function of phase relations,
showing no consistent relation.
Fig. 2. Phase-relation–dependent modulation of power correlations is fre-
quency specific. (A) Average power correlation as a function of phase relation
(x axis) and frequency (y axis) for MUA-MUA pairs recorded in cat area 17. (B)
Same as (A), but for MUA-LFP pairs. (C) Modulation depth of the cosine function
fitted to the phase-relation–dependent power correlations. Gray bars indicate
significant frequencies (P < 0.05, multiple comparisons corrected). (Right)
Average phase-relation–dependent power correlation at 60 Hz. (D and E)Same
as (C), but for (D) cat area 18×21a and (E) monkey area V1.
15 JUNE 2007 VOL 316 SCIENCE www.sciencemag.org
1610
REPORTS
on November 13, 2012www.sciencemag.orgDownloaded from
One concern is that the effect might be due
to rhythmic common input that fluctuates in
strength across trials. Common input would bias
phase relations and impose correlated power on
both groups. Clues about the actual causal chain
of events might be gained from the relative
timing between phase relation, on one hand, and
power correlation, on the other hand. If power
correlation covaried with phase relation because
of common input, then there should be no delay
between the two. However, if good phase re-
lations were actually the mechanistic cause of
strong power correlations, then good phase re-
lations should precede strong power correlations
by a few milliseconds, incurred by axonal,
synaptic, and intracellular delays. We therefore
compiled time-resolved estimates of the “good-
ness” of phase relations and of the strength of
power correlations, and then we determined the
cross-correlation as a function of time lag
betweenthetwotimeseries(21). Figure 3
shows this analysis pooled across all four data
sets and demonstrates that good phase relations
preceded strong power correlations by 5 ms. We
observed this temporal precedence in each of the
four data sets. Additional observations arguing
against a common input explanation are given in
the supporting online material text.
The modulation of two neuronal groups’
interaction by their phase relation would be
particularly interesting if it was spatially specific
(Fig. 1A). We therefore investigated recording
triplets A, B, and C from three separate electrodes
(in which A was an LFP and B and C were
MUAs). Figure 4A shows, for one example
triplet, a scatter plot in which each dot cor-
responds to one trial and the x and y values give
the “goodness” of the A-B and A-C phase
relations, respectively (22). W e sorted the trials
according to the quadrants of the plot. In
quadrants Q1 and Q2, the phase relation between
A and B was good, whereas in Q3 and Q4 it was
bad. We contrasted the A-B power correlation for
Q1 and Q2 with that for Q3 and Q4 (red line in
Fig. 4, B to D). Orthogonally to this, the phase
relation between A and C was good in Q1 and
Q3, whereas it was bad in Q2 and Q4, and we
contrasted the A-B power correlation for Q1
andQ3withthatforQ2andQ4(bluelineinFig.
4, B to D). The A-B power correlation depended
significantly more on the A-B phase relation
than on the A-C phase relation. Thus, the effect
had a spatial resolution that was at least as high
as the spatial resolution of our recordings (down
to 0.65 mm in the monkey data sets).
We provided evidence suggesting that neu-
ronal interactions mechanistically depend on the
phase relation between rhythmic activities. The
most likely reason for this dependence is that
rhythmic activities modulate the gain of incom-
ing synaptic input rhythmically. Effective con-
nectivity can thus be maximized or minimized
through synchronization at a good or bad phase
relation. The impact of pyramidal cells could be
enhanced, for example, if their firing phase rel-
ative to interneurons were advanced (15, 23, 24)
or if interneuronal firing were delayed through
inhibition or reduced excitation (25). Such mech-
anisms might be invoked directly by cognitive
top-down control.
Effective connectivity would diminish when
synchronization is less precise, because then syn-
aptic input is more likely to arrive at random
phases. This mechanism has the advantage that
within a sufficiently wide frequency band, mul-
tiple groups can be desynchronized, with respect
to a given target group, without being necessarily
synchronized to each other . Periods of putative
interactions between distant neuronal groups are
marked by an increased precision of synchroni-
zation (1, 4, 6, 26–30).
We propose that the pattern of synchroniza-
tion (its precision, phase, or both) weights the
anatomical-connection infrastructure with a gain
pattern, resulting in an effective interaction pat-
tern (14). Such a mechanism would have several
Fig. 3. Good phase re-
lations precede strong
power correlations. (A)
Spearman rank correla-
tion coefficient (y axis)
between the power cor-
relation and the “good-
ness” of the phase relation
across all MUA-LFP pairs
of all data sets for relative
time lags (x axis) be-
tween –200 and 200 ms. (B) Detail from (A), demonstrating the peak of the cross-correlation function at
–5 ms. A latency of the peak outside the gray shaded area is significant at P < 0.05.
Fig. 4. Spatial selectivity of phase-
relation–de pendent power correlation.
(A) Scatterplot shows the distribution of
trialwise phase relations between groups
AandB(y axis) and between groups A
and C (x axis) for an example triplet at
60 Hz. Equations define how A-B power
correlations from each quadrant were
combined for the results shown in (B to
D). In the equations, c(AB
q
) denotes the
A-B power correlation across trials in
quadrant q (where q is 1, 2, 3, or 4). (B)
A-B power correlation as a function of
the A-B phase relation [irrespective of
the A-C phase relation (red line)] and
as a function of the A-C phase relation
[irrespective of the A-B phase relation
(blue line)]. Gray bars indicate frequen -
cies with significant differences (P < 0.05,
multiple comparisons corrected). The y
axis denotes the differences in power
correlations according to the equators
shown in (A). (C and D) Same as (B), but
for (C) monkey area V1 and (D) monkey
area V4.
www.sciencemag.org SCIENCE VOL 316 15 JUNE 2007 1611
REPORTS
on November 13, 2012www.sciencemag.orgDownloaded from
interesting features. First, the effective interac-
tion pattern could be modified very dynamically.
Second, the mechanism would act connection-
wise. Third, a transient interaction would lead to
spike-time–dependent plast icity (8– 10)and
thus, a long-term trace. And fourth, synchroni-
zation might emerge in a self-organized manner
between “matching” neuronal groups. In the
visual cortex, synchronization is stronger among
neurons activated by the same visual stimulus
(1). This principle might generalize to the hand-
shaking between cognitive top-down control and
matching sensory bottom-up information, in
which case consecutive synchronization could
contribute to the selective routing of sensory
information t o behavioral control (13, 14, 25).
Our results suggest that synchronization has
consequences for neuronal interactions, pro-
viding a putative mechanism through which syn-
chronization contributes to cognitive functions.
References and Notes
1. C. M. Gray, P. König, A. K. Engel, W. Singer, Nature 338,
334 (1989).
2. A. Bragin et al., J. Neurosci. 15, 47 (1995).
3. P. Fries, J. H. Reynolds, A. E. Rorie, R. Desimone, Science
291, 1560 (2001).
4. J. Fell et al., Nat. Neurosci. 4, 1259 (2001).
5. B. Pesaran, J. S. Pezaris, M. Sahani, P. P. Mitra, R. A.
Andersen, Nat. Neurosci. 5, 805 (2002).
6. J.-M. Schoffelen, R. Oostenveld, P. Fries, Science 308,
111 (2005).
7. G. Buzsáki, Rhythms of the Brain (Oxford Univ. Press,
Oxford, 2006).
8. H. Markram, J. Lübke, M. Frotscher, B. Sakmann, Science
275, 213 (1997).
9. G. Q. Bi, M. M. Poo, J. Neurosci. 18, 10464 (1998).
10. V. Wespatat, F. Tennigkeit, W. Singer, J. Neurosci. 24,
9067 (2004).
11. P. König, A. K. Engel, W. Singer, Trends Neurosci. 19,
130 (1996).
12. R. Azouz, C. M. Gray, Neuron 37, 513 (2003).
13. E. Salinas, T. J. Sejnowski, Nat. Rev. Neurosci. 2, 539
(2001).
14. P. Fries, Trends Cognit. Sci. 9, 474 (2005).
15. A. Hasenstaub et al., Neuron 47, 423 (2005).
16. I. Vida, M. Bartos, P. Jonas, Neuron 49,107
(2006).
17. M. Bartos, I. Vida, P. Jonas, Nat. Rev. Neurosci. 8,45
(2007).
18. P. Fries, P. R. Roelfsema, A. K. Engel, P. König, W. Singer,
Proc. Natl. Acad. Sci. U.S.A. 94, 12699 (1997).
19. P. Fries, J. H. Schröder, P. R. Roelfsema, W. Singer, A. K.
Engel, J. Neurosci. 22, 3739 (2002).
20. T. Womelsdorf, P. Fries, P. P. Mitra, R. Desimone, Nature
439, 733 (2006).
21. Materials and methods are available as supporting
material on Science Online.
22. Positive and negative departures from the good phase
relation toward the bad phase relation were pooled
because our previous analyses had demonstrated
symmetric effects in both directions.
23. P. König, A. K. Engel, P. R. Roelfsema, W. Singer,
Neural Comput. 7, 469 (1995).
24. J. Csicsvari, B. Jamieson, K. D. Wise, G. Buzsáki, Neuron
37, 311 (2003).
25. J. Mishra, J. M. Fellous, T. J. Sejnowski, Neural Networks
19, 1329 (2006).
26. A. K. Engel, A. K. Kreiter, P. König, W. Singer, Proc. Natl.
Acad. Sci. U.S.A. 88, 6048 (1991).
27. G. Buzsáki, Z. Horváth, R. Urioste, J. Hetke, K. Wise,
Science 256, 1025 (1992).
28. P. R. Roelfsema, A. K. Engel, P. König, W. Singer, Nature
385, 157 (1997).
29. C. Tallon-Baudry, S. Mandon, W. A. Freiwald, A. K. Kreiter,
Cereb. Cortex 14, 713 (2004).
30. J. Gross et al., Proc. Natl. Acad. Sci. U.S.A. 101, 13050
(2004).
31. We thank P. König, P. R. Roelfsema, and J. H. Schröder
for help during the cat experiments and J. H. Reynolds,
A. E. Rorie, and A. F. Rossi for help during the monkey
experiments. This work was supported by the Human
Frontier Science Program and the Netherlands
Organization for Scientific Research (P.F.); the
Volkswagen Foundation (A.K.E. and P.F); the Max Planck
Society (P.F., A.K.E., and W.S.); European Union grants
EU IST-027268 and EU NEST-043457 (A.K.E.); and
NIH grant R01EY017292 and the National Institute of
Mental Health–Intramural Research Program (R.D.).
Supporting Online Material
www.sciencemag.org/cgi/content/full/316/5831/1609/DC1
Materials and Methods
SOM Text
Figs. S1 to S6
References
5 January 2007; accepted 1 May 2007
10.1126/science.1139597
Neural Mechanisms of Visual
Attention: How Top-Down Feedback
Highlights Relevant Locations
Yuri B. Saalmann,
1
Ivan N. Pigarev,
2
Trichur R. Vidyasagar
1
*
Attention helps us process potentially important objects by selectively increasing the activity of
sensory neurons that represent the relevant locations and features of our environment. This
selection process requires top-down feedback about what is important in our environment. We
investigated how parietal cortical output influences neural activity in early sensory areas. Neural
recordings were made simultaneously from the posterior parietal cortex and an earlier area in the
visual pathway, the medial temporal area, of macaques performing a visual matching task. When
the monkey selectively attended to a location, the timing of activities in the two regions became
synchronized, with the parietal cortex leading the medial temporal area. Parietal neurons may thus
selectively increase activity in earlier sensory areas to enable focused spatial attention.
A
ttention allows us to engage with our en-
vironment by selecting information rel-
evant for behavior (1–3). This enables
preferential processing of particular locations in
the visual field or specific features of objects.
Attentionmaintainedonalocationisusually
referred to as spatial attention and that on a
feature as feature-based attention (4, 5). Both
types of attention manifest in visual cortical areas
as increased activity of neurons representing the
attended location or feature and reduced activity
of other neurons (6–14). This may require top-
down feedback about what is relevant in the
environment; however , such feedback has not
been empirically demonstrated. There is evidence
(6, 11, 15–17) that the posterior parietal cortex
(PPC) is critical for spatial attention. The PPC is a
higher-order structure along the dorsal stream of
visual areas (Fig. 1A), which are particularly
concerned with spatial aspects of a scene. It has
been suggested that the spatial information about
a scene extracted by the PPC forms the basis for
feedback signals to earli er levels of the visual
pathway, highlighting spatial locations of poten-
tial interest (18, 19) and gating responses depend-
ing upon the state of attention.
We simultaneously measured the activity in
a part of the macaque PPC called the lateral
intraparietal area (LIP), and the immediately
earlier stage of the dorsal pathway, the medial
temporal area (MT; Fig. 1A). We tested whether
LIP feedback increases MT responses to attended
visual stimuli. The monkeys performed a delayed
match-to-sample (DMS) task, which manipu-
lated both where they were attending and what
stimulus feature they were attending to (Fig. 1B)
(20). We recorded 29 pairs of neurons from MT
and LIP, each pair having overlapping receptive
fields (RFs) and the same preferred orientation.
Local field potentials (LFPs) from the two sites
were also recorded.
We first tested whether our paradigm resulted
in increased responses of MT neurons to attended
stimuli. Figure 2A shows the effect of spatial at-
tention on a single MT neuron and Fig. 2B, the
average population response. For both data sets,
the MT response to the second stimulus was sig-
nificantly increased in the “spatial and feature-
based attention” and “spatial attention” condi-
tions, compared to the “neutral” control. When
“attention was elsewhere,” theMTresponseto
the second stimulus was significantly reduced
(Holm’s controlled Wilco xon test, P <0.05).These
attentional effects on MT neurons are consistent
with those reported in other types of cognitive
1
Department of Optometry and Vision Sciences, The University
of Melbourne, Parkville 3010, Australia.
2
Institute for In-
formation Transmission Problems, Russian Academy of
Sciences, Bol’shoy Karetniy 19, 127994 Moscow, Russia.
*To whom correspondence should be addressed. E-mail:
trv@unimelb.edu.au
15 JUNE 2007 VOL 316 SCIENCE www.sciencemag.org1612
REPORTS
on November 13, 2012www.sciencemag.orgDownloaded from