Internally Generated Cell Assembly Sequences in

the Rat Hippocampus

Eva Pastalkova, et al.
Science 321, 1322 (2008);
DOI: 10.1126/science.1159775

The following resources related to this article are available online at

www.sciencemag.org (this information is current as of September 5, 2008 ):

Updated information and services, including high-resolution figures, can be found in the online
version of this article at:
http://www.sciencemag.org/cgi/content/full/321/5894/1322

Downloaded from www.sciencemag.org on September 5, 2008

Supporting Online Material can be found at:
http://www.sciencemag.org/cgi/content/full/321/5894/1322/DC1

This article cites 36 articles, 9 of which can be accessed for free:

http://www.sciencemag.org/cgi/content/full/321/5894/1322#otherarticles

This article appears in the following subject collections:

Neuroscience
http://www.sciencemag.org/cgi/collection/neuroscience

Information about obtaining reprints of this article or about obtaining permission to reproduce
this article in whole or in part can be found at:
http://www.sciencemag.org/about/permissions.dtl

Science (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by the
American Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. Copyright
2008 by the American Association for the Advancement of Science; all rights reserved. The title Science is a
registered trademark of AAAS.
 RESEARCH ARTICLES
strates are processed in mFAS by individual full
sets of active sites, according to the path of ACP
described above. However, these studies have
also shown that a minority of substrates can be
shuttled between the two sets of active sites,
either by ACP serving both MAT domains or by
direct interaction of ACP with both KS domains
(6, 60–62). In light of the large 135 Å distance
between the ACP anchor point located in one
catalytic cleft and the MAT in the other, the most
plausible explanation for the minor mode-of-
domain interaction is a large-scale rotation of the
upper portion of mFAS, relative to the lower
portion (fig. S4).
The molecular description of active sites in
mFAS should stimulate the development of
improved inhibitors as anticancer drug candi-
dates. As demonstrated by structural homology,
29. I. Molnar et al., Chem. Biol. 7, 97 (2000).
30. D. J. Edwards et al., Chem. Biol. 11, 817 (2004).
31. A. Witkowski, A. K. Joshi, S. Smith, Biochemistry 36,
16338 (1997).
32. L. Serre, E. C. Verbree, Z. Dauter, A. R. Stuitje,
Z. S. Derewenda, J. Biol. Chem. 270, 12961 (1995).
33. C. Oefner, H. Schulz, A. D'Arcy, G. E. Dale, Acta
Crystallogr. D Biol. Crystallogr. 62, 613 (2006).
34. V. S. Rangan, S. Smith, J. Biol. Chem. 272, 11975 (1997).
35. B. Persson, Y. Kallberg, U. Oppermann, H. Jornvall,
Chem. Biol. Interact. 143-144, 271 (2003).
36. A. C. Price, Y. M. Zhang, C. O. Rock, S. W. White,
Structure 12, 417 (2004).
37. D. A. Rozwarski, C. Vilcheze, M. Sugantino, R. Bittman,
J. C. Sacchettini, J. Biol. Chem. 274, 15582 (1999).
38. T. J. Sullivan et al., Am. Chem. Soc. Chem. Biol. 1, 43
(2006).
39. M. Leesong, B. S. Henderson, J. R. Gillig, J. M. Schwab,
J. L. Smith, Structure 4, 253 (1996).
40. M. S. Kimber et al., J. Biol. Chem. 279, 52593 (2004).
41. A. K. Joshi, S. Smith, J. Biol. Chem. 268, 22508 (1993).
42. S. Pasta, A. Witkowski, A. K. Joshi, S. Smith, Chem. Biol.
54. K. J. Weissman, ChemBioChem 7, 485 (2006).
55. R. N. Perham, Biochemistry 30, 8501 (1991).
56. A. Roujeinikova et al., Structure 10, 825 (2002).
57. A. Roujeinikova et al., J. Mol. Biol. 365, 135 (2007).
58. G. A. Zornetzer, B. G. Fox, J. L. Markley, Biochemistry 45,
5217 (2006).
59. S. Jones, J. M. Thornton, Proc. Natl. Acad. Sci. U.S.A. 93,
13 (1996).
60. S. Smith, A. Witkowski, A. K. Joshi, Prog. Lipid Res. 42,
289 (2003).
61. A. K. Joshi, V. S. Rangan, A. Witkowski, S. Smith, Chem.
Biol. 10, 169 (2003).
62. V. S. Rangan, A. K. Joshi, S. Smith, Biochemistry 40,
10792 (2001).
63. All data were collected at the Swiss Light Source (SLS,
Paul Scherrer Institute, Villigen). We thank
C. Schulze-Briese, S. Gutmann, R. Bingel-Erlenmeyer,
S. Russo, A. Pauluhn, and T. Tomizaki for their
outstanding support at the SLS; S. Jenni and M. Sutter for
critically reading the manuscript and all members of the
Ban laboratory for suggestions and discussions;

Downloaded from www.sciencemag.org on September 5, 2008

this structure is also a good template for the
organization of PKS modules; it agrees with and
extends present theoretical models of PKS
architecture (19, 22). Furthermore, the structure
of mFAS paves the way for structure-based ex-
periments to answer remaining questions on the
dynamics and substrate shuttling mechanism in
megasynthases.
References and Notes
1. S. W. White, J. Zheng, Y. M. Zhang, C. O. Rock, Annu.
Rev. Biochem. 74, 791 (2005).
2. S. Jenni et al., Science 316, 254 (2007).
3. S. Jenni, M. Leibundgut, T. Maier, N. Ban, Science 311,
1263 (2006).
4. I. B. Lomakin, Y. Xiong, T. A. Steitz, Cell 129, 319 (2007).
5. T. Maier, S. Jenni, N. Ban, Science 311, 1258 (2006).
6. S. Smith, S. C. Tsai, Nat. Prod. Rep. 24, 1041 (2007).
7. F. Lynen, Eur. J. Biochem. 112, 431 (1980).
8. E. Schweizer, J. Hofmann, Microbiol. Mol. Biol. Rev. 68,
501 (2004).
9. F. P. Kuhajda et al., Proc. Natl. Acad. Sci. U.S.A. 91,
6379 (1994).
10. J. A. Menendez, R. Lupu, Nat. Rev. Cancer 7, 763 (2007).
11. F. P. Kuhajda, Cancer Res. 66, 5977 (2006).
12. H. Orita et al., Clin. Cancer Res. 13, 7139 (2007).
13. M. Leibundgut, S. Jenni, C. Frick, N. Ban, Science 316,
288 (2007).
14. E. Ploskon et al., J. Biol. Chem. 283, 518 (2008).
15. G. Bunkoczi et al., Chem. Biol. 14, 1243 (2007).
16. B. Chakravarty, Z. Gu, S. S. Chirala, S. J. Wakil,
F. A. Quiocho, Proc. Natl. Acad. Sci. U.S.A. 101, 15567
(2004).
17. C. W. Pemble IV, L. C. Johnson, S. J. Kridel, W. T. Lowther,
Nat. Struct. Mol. Biol. 14, 704 (2007).
18. Y. Tang, C.-Y. Kim, I. I. Mathews, D. E. Cane, C. Khosla,
Proc. Natl. Acad. Sci. U.S.A. 103, 11124 (2006).
19. A. T. Keatinge-Clay, R. M. Stroud, Structure 14, 737
(2006).
20. A. T. Keatinge-Clay, Chem. Biol. 14, 898 (2007).
21. Y. Tang, A. Y. Chen, C.-Y. Kim, D. E. Cane, C. Khosla,
Chem. Biol. 14, 931 (2007).
22. C. Khosla, Y. Tang, A. Y. Chen, N. A. Schnarr, D. E. Cane,
Annu. Rev. Biochem. 76, 195 (2007).
23. A. K. Joshi, A. Witkowski, H. A. Berman, L. Zhang, S. Smith,
Biochemistry 44, 4100 (2005).
24. Materials and methods are available as supporting
material on Science Online.
R. Grosse-Kunstleve, P. Afonine, and P. Adams for
14, 1377 (2007).
support with the PHENIX software; and A. Jones for
43. C. Baldock, J. B. Rafferty, A. R. Stuitje, A. R. Slabas,
D. W. Rice, J. Mol. Biol. 284, 1529 (1998).
support with the program O. This work was supported by
the Swiss National Science Foundation (SNSF) and the
44. R. J. Heath, N. Su, C. K. Murphy, C. O. Rock, J. Biol.
National Center of Excellence in Research Structural
Chem. 275, 40128 (2000).
Biology program of the SNSF. Structure factors and
45. R. P. Massengo-Tiasse, J. E. Cronan, J. Biol. Chem. 283,
atomic coordinates of the porcine FAS in the apo- and
1308 (2008).
NADP+-bound form have been deposited in the Protein
46. J. Saito et al., Protein Sci. 17, 691 (2008).
Data Bank with accession codes 2vz8 and 2vz9.
47. E. Nordling, H. Jornvall, B. Persson, Eur. J. Biochem. 269,
4267 (2002).
48. T. T. Airenne et al., J. Mol. Biol. 327, 47 (2003). Supporting Online Material
49. T. Hori et al., J. Biol. Chem. 279, 22615 (2004). www.sciencemag.org/cgi/content/full/321/5894/1315/DC1
50. Y. Shimomura, Y. Kakuta, K. Fukuyama, J. Bacteriol. 185, Materials and Methods
4211 (2003). Figs. S1 to S15
51. D. A. Hopwood, D. H. Sherman, Annu. Rev. Genet. 24, 37 Tables S1 to S4
(1990). References
52. C. D. Richter et al., FEBS J. 274, 2196 (2007).
53. C. D. Richter, D. Nietlispach, R. W. Broadhurst, 3 June 2008; accepted 31 July 2008
K. J. Weissman, Nat. Chem. Biol. 4, 75 (2008). 10.1126/science.1161269
Internally Generated Cell Assembly
Sequences in the Rat Hippocampus
Eva Pastalkova, Vladimir Itskov,* Asohan Amarasingham, György Buzsáki†
A long-standing conjecture in neuroscience is that aspects of cognition depend on the brain’s ability
to self-generate sequential neuronal activity. We found that reliably and continually changing cell
assemblies in the rat hippocampus appeared not only during spatial navigation but also in the
absence of changing environmental or body-derived inputs. During the delay period of a memory
task, each moment in time was characterized by the activity of a particular assembly of neurons.
Identical initial conditions triggered a similar assembly sequence, whereas different conditions
gave rise to different sequences, thereby predicting behavioral choices, including errors. Such
sequences were not formed in control (nonmemory) tasks. We hypothesize that neuronal
representations, evolved for encoding distance in spatial navigation, also support episodic recall
and the planning of action sequences.
prominent theory states that the hippo- memory and action planning, draw on the activ-
A campal system primarily serves spatial
navigation (1, 2); a component of this
theory is that the place-dependent activity of
ity of hypothetical internally organized cell as-
semblies (8–13).

25. J. G. Olsen, A. Kadziola, P. von Wettstein-Knowles,

M. Siggaard-Andersen, S. Larsen, Structure 9, 233 (2001).
26. J. M. Thorn, J. D. Barton, N. E. Dixon, D. L. Ollis,
K. J. Edwards, J. Mol. Biol. 249, 785 (1995).
27. J. L. Martin, F. M. McMillan, Curr. Opin. Struct. Biol. 12,
783 (2002).
28. I. Fujii, N. Yoshida, S. Shimomaki, H. Oikawa, Y. Ebizuka,
Chem. Biol. 12, 1301 (2005).
neurons [place cells (1, 2)] in the hippocampus
arises from external serially ordered environ-
mental stimuli (3–7). Place cells are thought to
embody the representation of a cognitive map,
enabling flexible navigation. However, neural
theories of other cognitive processes that may
depend on the hippocampus, such as episodic
Center for Molecular and Behavioral Neuroscience, Rutgers,
The State University of New Jersey, 197 University Avenue,
Newark, NJ 07102, USA.
*Present address: Center for Neurobiology and Behavior,
Columbia University, 1051 Riverside Drive, New York, NY
10032, USA.
†To whom correspondence should be addressed. E-mail:
buzsaki@axon.rutgers.edu

1322 5 SEPTEMBER 2008 VOL 321 SCIENCE www.sciencemag.org


Several observations have refined the navi-
gation theory. Hippocampal neurons can predict
where the animal is coming from, or its desti-
nation (14–17); the sequential activity of place
cells during locomotion is replicated within
single cycles of the theta oscillation (8 to 12 Hz)
(18–20); furthermore, the temporal recruitment of
active neurons in the population bursts of rest and
sleep also reflects, again on a faster time scale,
their sequential activity as place cells during
locomotion (21–23). Thus, the sequential activa-
tion of hippocampal neurons can be disengaged
from external landmarks (24, 25). However, in-
ternally generated assembly sequences operat-
ing at the time scale of behavior have not yet
been reported.
The frameworks of environment-controlled
versus internally generated assembly sequences
give rise to distinct predictions. Imagine that a rat
is frozen in position during its travel (and yet the
theta oscillation is maintained). According to the
navigation theory, a subset of landmark-controlled
place cells should then display sustained activity,
and other neurons would remain suppressed (2–6).
In contrast, if assembly sequences were gener-
ated by internal mechanisms, neurons might
rather display continually changing activity. We
tested these predictions by examining the activity
RESEARCH ARTICLES
of hippocampal neurons while a rat was running
in a wheel at a relatively constant speed (26, 27)
during the delay of a hippocampus-dependent
alternation memory task.
Internally generated cell-assembly sequences.
Rats were trained to alternate between the left
and right arms of a figure-eight maze [Fig. 1A
and supporting online material (SOM) text].
During the delay period between maze runs (10 s
for rat 1; 20 s each for rats 2 and 3), the animals
were trained to run steadily in the same direction
in a wheel (Fig. 1A). To confront the predictions
of the navigation theory with those of the internal
sequence-generation hypothesis, we compared

Fig. 1. Episode fields in A B C

the wheel and place

Downloaded from www.sciencemag.org on September 5, 2008

Water port

Firing rate - maze (Hz)

Running wheel

% of ne urons a c tive
fields in the maze are 80 5
similar. (A) Color-coded 4
60
spikes (dots) of simulta-
neously recorded hippo- 3
40
campal CA1 pyramidal 2
neurons. The rat was re- 20
quired to run in the 1
wheel facing to the left 0
Animal's trajectory 1 2 3 4 5
during the delay be-
Firing rate - wheel (Hz)
tween the runs in the
maze. (B) Percent of D E
25 1 1
neurons firing >0.2 Hz 4 17 26

Firing rate (normalized)

within each pixel. The

Neurons (sorted)
highest percentage of 0.8
neurons was active when
rats were running in the 0.6
Trials

1
wheel. (C) Relationship 25 18
5 8
between firing rate of 0.4
neurons active in rats
running the wheel and 0.2
the maze (rs = –0.3, P <
0.0001, 681 neurons, 1 30 0
5 10 15 5 10 15 5 10 15 0 5 10 15
three rats, 17 sessions).
(D) Normalized firing Time in wheel (sec) Time in wheel (sec)
rate of six simultane- F G
ously recorded neurons
during wheel running 10
wheel
maze
wheel wheel
(each line shows the 8 1
Time (sec)
N of cells

color-coded activity on
single trials turning to 6 5
the left arm). The epi- 4 2
sode fields occurred at
2
specific segments of the
run. (E) Normalized fir- 0 10 3
0 0.5 1 1.5 2 2.5 5 10 1 2 3
ing rate of 30 simulta- Field width (sec) Time (sec)
neously recorded neurons
1
during wheel running, maze 1
Correlation coeficient

10 wheel maze
ordered by the latency maze
Correl. coef.

8 wheel
of their peak firing rate. 1
0.5
N of cells

(F) Width (top) and peak 6

firing rate (bottom) of
4 2
episode and place fields 0 0
(wheel, n = 135 neurons; 2
maze, n = 162 neurons).
0 3
Arrows indicate medians. 5 10 15 20 25 30
-0.5
0.5 1 1.5 1 2 3
(G) Population vector Peak firing rate (Hz) Time (sec) Time (sec)
cross-correlation matrix
(SOM text). The width of the diagonal stripe indicates the rate at which neuronal assemblies transition. (Lower left) The decay of the population vector correlation
during wheel running and maze traversal. Thin lines, individual sessions; thick lines, group means.

www.sciencemag.org SCIENCE VOL 321 5 SEPTEMBER 2008 1323

 RESEARCH ARTICLES
the firing patterns of CA1 hippocampal neurons
in rats running the wheel and the maze.
We analyzed the activity of ~500 pyramidal
cells recorded in the wheel and ~600 neurons in
the maze (mean firing rate >0.5 Hz) (Fig. 1A).
Pyramidal neurons were transiently active in rats
running both the maze [place cells (1)] and the
wheel. Although the position and direction of the
rat’s head were stationary during wheel running
(fig. S1), the percentage of neurons active in the
pixels occupied by the head during wheel run-
ning was three to four times greater than in any
area of comparable size in the maze (Wilcoxon
rank sum test, P < 0.0001) (Fig. 1B). Thus, if
pyramidal neurons were solely activated by
environmental cues (2–6), this finding would
reflect several-fold–stronger neuronal representa-
tion of the animal’s position within the wheel.
active [firing at least a single spike in 100-ms
windows (averaging over 100-ms windows)]
was similar in the wheel (10.75 T 3.97%) and
the maze (12.56 T 4.32%) (fig. S3).
Pyramidal neurons typically fired transiently,
and reliably in successive trials, at specific times
of wheel running (episode fields), and most cells
had multiple peaks of varying sizes (Fig. 1D).
Typically, and reminiscent of a synfire chain (11),
at least one episode cell was active at every
moment of a wheel run (Fig. 1E).
Were episode cells in rats in the wheel
generated by the same mechanism as place cells
in rats in the maze? We looked for evidence of
differing mechanisms by comparing several mea-
sures of the firing of episode and place cells.
First, we calculated the duration of activity (field
width) (Fig. 1F) of single cells [including only
13.08 and 12.8 Hz, respectively; P = 0.61)
differed significantly between the episode and
place fields (Fig. 1F). Second, to measure the
average lifetime of assembly activity for a pop-
ulation, we determined the maximal time lag at
which the autocorrelation of the population’s ac-
tivity was above 0.5 (29) and again found no
significant difference, with respect to the median,
between the populations of episode and place
cells (medians were 0.83 and 0.75 s, respectively;
P = 0.32) (Fig. 1G). Third, we compared the
relationship between spikes and the local field
potential in episode and place cells. On linear
tracks, sequentially generated spikes of a place
cell gradually shift to earlier and earlier phases of
the theta oscillation as the rat passes through the
place field (phase precession), and there is a sys-
tematic relationship between the phase of spikes

Downloaded from www.sciencemag.org on September 5, 2008

Many individual pyramidal cells were active both
in rats running the wheel and rats running the
maze, but the sequential order of their activation
in rats in the wheel was unrelated to that of rats
in the maze, and their firing rates in these two
areas were inversely correlated [Spearman corre-
lation coefficient (rs) = –0.3, P < 0.0001, n = 681
neurons (Figs. 1C and 4B); contrast this with the
population of interneurons, rs = 0.85, P < 0.0001,
n = 125 interneurons (fig. S2)]. The average
proportion of pyramidal neurons simultaneously
fields with a peak firing rate of ≥6.0 Hz and ≥4.5
SD above the mean firing rate (SOM text)]. The
temporal and spatial extent of the field was
determined as those times and positions at which
firing rates were at least 10% that of the peak
firing rate (in the wheel or maze) (19, 28). By
these criteria, 32% of the neurons recorded in the
wheel and 22% in the maze had at least one field.
Neither the distribution of field widths (medians
were 0.94 and 1.0 s, respectively; Wilcoxon test,
P = 0.44) nor peak firing rates (medians were
and the animal’s position (3, 18–20, 28, 30, 31).
The navigation theory predicts that the phase of
spikes will remain fixed if environmental inputs
do not change (3, 26, 27). In contrast, episode
cells displayed phase precession during wheel
running (Fig. 2A). Similarly to place cells, the
theta frequency oscillation of episode cells was
higher than that of the field theta rhythm (Fig.
2B), and the slope of the phase precession was
inversely related to the length of the episode
field (Fig. 2, A and D) (3, 19, 20, 28, 30, 31).

Fig. 2. Episode neurons A B

Normalized power
0 360 units
in the wheel display 0.8 LFP
theta phase precession
and temporal compres- 180
200 msec
sion. (A) (Top) Unfiltered 0.4
(light gray) and filtered 540 20 F irin g rate (Hz )
(4 to 10 Hz) (dark gray)
traces of LFP and phase 360
Theta phase (deg)

advancement of action 10 6 8 10 12
potentials (dots). (Bottom) 180 Frequency (Hz)
Activity of six example 12
neurons from the same 0 C wheel
Number of cells

12 maze
session. Each dot is an 540 1212
F irin g rate (Hz )

action potential, displayed 8

as a function of theta 360 88
6
phase and time from the 4
beginning of wheel run- 180 44
ning from all trials. One
and a half theta cycles 0
5 10 15 5 10 15 5 10 15 -1 -0.5 0
are shown (y axis). Red Time (sec) Slope (deg/sec)
line, smoothed firing D E
rate. (B) Power spectra 0 0 50
Slope (deg/sec)

of spike trains generated

∆ time (msec)
Slope (deg/m)

during wheel running 25

-400 -400
(n = 283 pyramidal neu-
rons) and the simulta- 0
neously recorded LFP. -800 -800
Faster oscillation of neu- -25
rons occurs relative to -1200 -1200
LFP. (C) Slope of theta -50
1 2 3 1 2 3 -0.5 -0.25 0 0.25 0.5
phase precession within Episode field width (sec) Episode field width (m) ∆ distance (m)
episode fields in the
wheel and within place fields in the maze. (D) Relationship between phase between the peaks of episode fields of neuron pairs in the wheel with
precession slope and episode length (left, r s = 0.46, P < 0.0001) and the temporal offset of the pair's cross-correlogram peaks is shown.
episode field width (right, rs = 0.52, P < 0.0001), respectively. (E) Each dot represents a neuron pair (n = 105 eligible pairs; three rats; rs =
Temporal compression of spikes sequences. Correlation of the distance 0.59; P < 0.0001).

1324 5 SEPTEMBER 2008 VOL 321 SCIENCE www.sciencemag.org


Furthermore, the slopes correlated more strongly
with the length of the episode field (rs = 0.52, P <
0.0001) than with the time it took the rat to run
through the same field (rs = 0.46, P < 0.0001)
(Figs. 2D and 3) because of the variability of the
rat’s running speed (28). The distributions of
phase precession slopes for the episode and place
fields were also similar (medians were –0.6°/s
and –0.6°/s, respectively; P = 0.6) (Fig. 2C).
Finally, we compared the spike timing relation-
ships among neurons. During maze traversals,
the distance between the place-field peaks of a
neuronal pair was correlated with the temporal
offset between its spikes within the theta cycle, a
phenomenon known as distance-time compres-
sion (SOM text) (18, 19). Analogously, the
distance between peaks of the episode fields of
neuron pairs (episode fields with peak firing rate
>5 Hz and >3 SD above the mean firing rate were
included in this analysis; n = 105 pairs) was
correlated with the temporal offsets between the
spikes at the theta time scale (rs = 0.59, P <
0.0001) (Fig. 2E). These findings indicate that
the mechanisms generating place and episode
fields are similar.
Body cues are not sufficient to generate
assembly sequences. It has been suggested that
in addition to generating a cognitive map of the
environment (2), the hippocampus and its asso-
ciated structures integrate self-motion–induced
RESEARCH ARTICLES
information (7, 32, 33). Were the episode cell
sequences generated by idiothetic self-motion
cues? We examined population firing patterns in
two control (nonmemory) tasks. In the first task
(control 1), the animals (rats 3 and 4) were
required to run in the wheel for a water reward
available in an adjacent box (26). In the second
task (control 2), the animals (rats 2 and 3) had
continuous access to a wheel adjacent to their
home cage, and recordings were made during
spontaneous wheel-running episodes. Transient
firing patterns, consistent across trials, were
rarely observed during the control tasks. Rather,
the majority of active neurons exhibited relative-
ly sustained firing throughout the wheel-running

Fig. 3. Firing patterns A Control 1 Control 2

during wheel running

Downloaded from www.sciencemag.org on September 5, 2008

12
depend on the context
Trials

of the task. (A) (Top)

Activity of representative
single neurons (color-
coded) during wheel run-
ning in control tasks 1 and 1
2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10
2 (compare with Fig. 1D).
Theta phase (deg)

540 540
(Bottom) Unit discharges

F ir in g r a te (Hz )
6 6
(dots) from all trials with-
360 360
in a session as a function 4 4
of theta phase, plotted
180 2 180 2
against time from the be-
ginning of a wheel run.
Red line, smoothed mean 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10
firing rate. Relatively Time in wheel (sec) Time in wheel (sec)
steady firing rates and a
steady theta phase occur B C
in both control tasks. (B) Memory Control 1 Control 2
1 1
Cross-correlation matrices memory
control

Correl. coef.
in three different tasks 0.8
2.5 Correl. coef.
Time (sec)

(memory and control 2 0.6

0.5
are from the same rat). In 5
the memory task, trials 0.4
0
with the same future 7.5 0.2
choices [left (L)–trialsn
versus L-trialsn+1 and 10 -0.5
2.5 5 7.5 10 2.5 5 7.5 10 2.5 5 7.5 10 -1 0 1
right (R)–trialsn versus Time (sec)
R-trialsn+1) were cross- Time in wheel (sec)
correlated, whereas in
control tasks trialsn and D Memory E Memory Control F
Control 1

trials n+1 were cross- 1 1 1 25

Normalized power

unit memory
Number of neurons

correlated. Only pixel LFP 20 control

Neuron #

values significantly dif-

ferent from chance are 15
0.5
shown (Spearman rank
Control 2

10
correlation, P < 0.01).
(C) Population-vector cor- 5
relation coefficient values
283 44 0
in the memory task (n = 6 8 10 12 -4 -2 0 2 4 -4 -2 0 2 4 50 100
17 sessions) and control Frequency (Hz) ∆ frequency (unit-LFP; Hz) % spikes within ACG peak
tasks (n = 8 sessions)
(mean T SD). (D) Power spectrum of spike trains of an episode neuron (unit) (black dots, maxima of the cross-correlograms; white line, sum of all
and simultaneously recorded LFP during wheel running in the memory task neurons). There is a significant frequency shift in the memory task (0.44 T
(30). The frequency of unit firing oscillation is higher than the frequency of 0.6 Hz) and a lack of frequency shift in control tasks (combined control 1 and
LFP. (E) Difference between unit and LFP oscillation frequency in the 2, 0.07 T 0.3 Hz). (F) Ratio of spikes in the center and tail of temporal auto-
memory (left) and control (right) tasks. Each line is a color-coded normalized correlograms (SOM text). High values indicate compact episode fields; low
cross-correlogram between power spectra of a pyramidal neuron and values indicate spikes scattered throughout the time of wheel running
simultaneously recorded LFP. A shift of the maximal correlation values to the (memory task, n = 287 neurons; control tasks, n = 85 neurons; rank sum test,
right indicates that unit theta oscillation is faster than LFP theta oscillation P < 0.0001). Arrows indicate medians.

www.sciencemag.org SCIENCE VOL 321 5 SEPTEMBER 2008 1325

 RESEARCH ARTICLES
periods (Fig. 3A and fig. S4) (5, 26–27). During A
runs of opposite direction in the wheel, different 1

Right
populations of neurons were active (fig. S5)

(normalized)
Firing rate
(26), arguing for the importance of distant cues

Trials
(2, 20) and against a critical role of idiothetic 0.5
inputs (26). In addition, the temporal organiza-

Left
tion of cell assemblies in control tasks was less
precise, as reflected by much weaker correla-
tions between temporally adjacent populations 0
0 5 10 15 0 5 10 15 0 5 10 15
during the control tasks than during the memory Time in wheel (sec)
task (Fig. 3, B and C), despite the similarity in
firing rates during all tasks (fig. S6). As another
contrast to the memory task, neurons recorded B Left trials Right trials Stem Significance Stem
during the control tasks fired throughout the Rat 1
1 1 right
trial, with spikes locked to a similar phase of the left
theta cycle (Fig. 3A). Consistent with these
observations, neurons in the rats performing the
0.5
memory task oscillated faster than the local field

Downloaded from www.sciencemag.org on September 5, 2008

potential (LFP) [difference (∆) = 0.44 T 0.6 Hz]
3
(Figs. 2B and 3, D and E), an indication of 0
113

Number of differentiating neurons

phase precession (19, 20, 29, 30), whereas dur- 2 4 6 2 4 6 2 4 6
Neurons (sorted by left trials)

ing the control tasks, the power spectra of the Stem

units and LFP were similar (∆ = 0.07 T 0.3 Hz)
(Fig. 3E). Finally, to quantify differences in
temporal clustering of spikes, we examined an
autocorrelogram of each neuron. We applied
(after filtering, 0.2 to 2 Hz) the same definition
for the peak region boundaries that we used for
the episode field detection boundary of the epi-
sode field (the 10% boundary) and then com-
pared, for each neuron, the ratio of the number
of spikes that fell within the peak region bound-
ary to those that fell outside. These ratios were
significantly larger during the memory task and
reflected the temporal compactness of firing
during the memory task as opposed to the con-
trol tasks (Fig. 3F). Thus, the indicators of tem-
porally precise sequential activity in neuronal
populations were absent during the control
tasks, despite indistinguishable motor character-
istics across all tasks.
Assembly sequences depend on memory
load. What is the behavioral function of in-
ternally generated cell-assembly sequences?
Temporarily inactivating neuronal circuits in the
dorsal hippocampus, we found that performance
in the delayed alternation task depends on the
integrity of the hippocampus (fig. S7) (17). Thus,
we hypothesized that information about choice
behavior is reflected in assembly sequences (34).
All correctly performed trials were sorted accord-
ing to the rat’s future choice of arm (left or right),
and choice-specific firing effects were identified
by comparing the firing patterns of single neu-
rons with those of surrogate spike trains created
by shuffling the left and right labels (Fig. 4, A
and B, and SOM text) (34). Some neurons were
active exclusively before either the left or right
choice, whereas others showed differential firing
rates and/or fired at different times after the
beginning of wheel running (Fig. 4A, figs. S8 to
S10, and movie S1). The largest proportion of
neurons exhibiting choice-predictive activity was
at the beginning of the run; this proportion de-
creased as a function of time during the delay and
Stem
Rat 2
40
1
20
229
4 8 12 4 8 12 4 8 12
Stem
Rat 3 Stem
1 15
10
5
322
4 8 12 4 8 12 4 8 12
Time in wheel (sec)
Fig. 4. Cell-assembly activity in the wheel predicts the future choice of the rat in the maze. (A)
Examples of three neurons that strongly differentiated between wheel-running trials preceding
right and left choices (fig. S7 and movie S1). (B) Normalized firing rate profiles of neurons during
wheel running and in the stem of the maze, ordered by the latency of their peak firing rates during
left trials (each line is a single cell; cells are combined from all sessions). White line, time gap
between the end of wheel running and the initiation of maze stem traversal. (Middle) Normalized
firing rates of the same neurons during right trials. (Right) Time periods of significant differences
(P < 0.05) in firing rates between left and right trials for respective neurons (red line, R > L; blue
line, L > R). Gray line, number of neurons discriminating between left and right trials as a function
of wheel-running time.
in the stem of the maze (Fig. 4B), suggesting a was not possible to disambiguate their influence
critical role for initial conditions in specifying the on neuronal activity. To distinguish such retro-
sequences (fig. S11). In addition, we designed a spective and prospective factors (14–17), we
probabilistic model of the relationship between examined cell-assembly sequences during error
neuronal firing patterns and the animal’s choices trials. Neurons that reliably predicted the behav-
(SOM text). Using this model, the accuracy of ioral choice of the rat on correct trials continued
single-trial prediction, under cross-validation, to predict the choice behavior on error trials (Fig.
varied from low (near 50%) and not significant 5A, fig. S12, and movie S1) (15, 24). Similarly,
to 100% and significant across many sessions population sequences that differentiated correct
(fig. S9). behavioral choices continued to predict behav-
Because the rat was performing an alternation ioral choice errors (Fig. 5, B and C, and fig. S13).
task, past and future choices were deterministi- Although there were only a few error trials, a
cally related on correctly performed trials, and it majority of them could be predicted from the

1326 5 SEPTEMBER 2008 VOL 321 SCIENCE www.sciencemag.org

 RESEARCH ARTICLES

A B Session mean (correct left) Session mean (correct right)

1

(sorted by left trials)

Lleft

Neurons
Trials
Right
Err

43
5 10 15 5 10 15 5 10 15 5 10 15
Time in wheel (sec) Time in wheel (sec)

Fig. 5. Cell-assembly activity during the wheel predicts behavioral errors C Single trial (error left) D

erroneous choices (%)

1 3/3 9/9

(sorted by left trials)

during the maze. (A) Two example neurons from a session with seven left 0.45 100

Correctly predicted
error trials (err). Correct trials are separated into left- and right-turn trials.
(B) Normalized firing rates of 43 neurons simultaneously recorded during

Neurons
9/13
wheel running, ordered by the latency of peak firing rates during correct
left trials (left). (Right) Firing sequence of the same neurons on correct 50
right trials. (C) Firing sequence of neurons in a single error (left) trial.

Downloaded from www.sciencemag.org on September 5, 2008

Neuronal order is the same as in (B). The firing sequence during the error
trial is similar to that of the correct left trials. The correlation coefficient 43
5 10 15 0
between correct and error trial sequences is 0.45 (fig. S13). (D) Percent of Rat1 Rat2 Rat3
Time in wheel (sec)
correctly predicted errors from the neuronal population activity.

firing patterns of neurons during wheel running
(Fig. 5D). Altogether, these observations demon-
strate that a particular sequence of neurons was
activated in a reliable temporal order from the
moment the rat entered the wheel to the time it
reached the reward.
Because running speed, head position, and head
direction during wheel running before left and right
choices were apparently indistinguishable (fig. S1),
the above findings indicate that trial differences
in hippocampal assembly configurations cannot
solely arise from instantaneous environmental
inputs or the integration of motion signals.
Behavioral function of internally gener-
ated cell-assembly sequences. These findings
demonstrate that the rat brain can generate con-
tinually changing assembly sequences. The pat-
terns of the self-evolving neuronal assembly
sequences depend on the initial conditions, and
the particular sequences of cell assemblies are
predictive of behavioral outcome.
Our results offer new insights into the rela-
tionship between hippocampal activity and
navigation (2–7, 14–20, 26–30, 33). Hippocam-
pal firing patterns during maze navigation were
similar to those during wheel running in the
delayed alternation memory task with stationary
environmental and body cues. Therefore, we
suggest that hippocampal networks can produce
sequential firing patterns in two possibly interact-
ing ways: under the influence of environmental/
idiothetic cues or by self-organized internal mech-
anisms. The high-dimensional and largely ran-
dom (nontopographical) connectivity of the CA3
axonal system (35) and its inputs makes the
hippocampus an ideal candidate for internal se-
quence generation (13, 33, 36, 37). The parame-
ters of cell-assembly dynamics (including their
trajectory and lifetimes) are probably affected by
a number of factors, including experience-
dependent and short-term synaptic plasticity
(34, 38); asymmetric inhibition (39); brain state;
and, fundamentally, the character and context of
the input. The evolving trajectory can be ef-
fectively perturbed, or updated, by external inputs
in every theta cycle (40). Because of this
flexibility in the sources of cell-assembly control,
we hypothesize that neuronal algorithms, having
evolved for the computation of distances, can also
support the episodic recall of events and the
planning of action sequences and goals (19).
During learning, the temporal order of external
events is instrumental in specifying and securing
the appropriate neuronal representations, whereas
during recall, imagination (35), or action plan-
ning, the sequence identity is determined by the
intrinsic dynamics of the network.
References and Notes
1. J. O’Keefe, J. Dostrovsky, Brain Res. 34, 171 (1971).
2. J. O’Keefe, L. Nadel, The Hippocampus as a Cognitive
Map (Clarendon, Oxford, UK, 1978).
3. J. Huxter, N. Burgess, J. O’Keefe. Nature 425, 828 (2003).
4. B. L. McNaughton, C. A. Barnes, J. O’Keefe, Exp. Brain
Res. 52, 41 (1983).
5. J. O'Keefe, N. Burgess, Nature 381, 425 (1996).
6. R. U. Muller, J. L. Kubie, J. B. Ranck Jr., J. Neurosci. 7,
1935 (1987).
7. B. L. McNaughton et al., J. Exp. Biol. 199, 173 (1996).
8. D. O. Hebb, The Organization of Behavior:
A Neuropsychological Theory (Wiley, New York, 1949).
9. E. Tulving, Elements of Episodic Memory (Clarendon,
Oxford, UK, 1983).
10. L. R. Squire, Psychol. Rev. 99, 195 (1992).
11. M. Abeles, Corticotronics: Neural Circuits of the Cerebral
Cortex, (Cambridge Univ. Press, New York, 1991).
12. M. W. Howard, M. S. Fotedar, A. V. Datey,
M. E. Hasselmo, Psychol. Rev. 112, 75 (2005).
13. W. B. Levy, A. B. Hocking, X. Wu, Neural Netw. 18, 1242
(2005).
14. L. M. Frank, E. N. Brown, M. Wilson, Neuron 27, 169
(2000).
15. J. Ferbinteanu, M. L. Shapiro, Neuron 40, 1227 (2003).
16. E. R. Wood, P. A. Dudchenko, R. J. Robitsek,
H. Eichenbaum, Neuron 27, 623 (2000).
17. J. A. Ainge, M. A. van der Meer, R. F. Langston,
E. R. Wood, Hippocampus 17, 988 (2007).
18. W. E. Skaggs, B. L. McNaughton, M. A. Wilson,
C. A. Barnes, Hippocampus 6, 149 (1996).
19. G. Dragoi, G. Buzsáki, Neuron 50, 145 (2006).
20. J. R. Huxter, T. J. Senior, K. Allen, J. Csicsvari, Nat.
Neurosci. 11, 587 (2008).
21. G. Buzsáki, Neuroscience 31, 551 (1989).
22. M. A. Wilson, B. L. McNaughton, Science 265, 676 (1994).
23. K. Louie, M. A. Wilson, Neuron 29, 145 (2001).
24. S. A. Deadwyler, T. Bunn, R. E. Hampson, J. Neurosci. 16,
354 (1996).
25. H. Eichenbaum, P. Dudchenko, E. Wood, M. Shapiro,
H. Tanila, Neuron 23, 209 (1999).
26. A. Czurko, H. Hirase, J. Csicsvari, G. Buzsáki, Eur. J.
Neurosci. 11, 344 (1999).
27. H. Hirase, A. Czurko, J. Csicsvari, G. Buzsáki, Eur. J.
Neurosci. 11, 4373 (1999).
28. C. Geisler, D. Robbe, M. Zugaro, A. Sirota, G. Buzsáki,
Proc. Natl. Acad. Sci. U.S.A. 104, 8149 (2007).
29. K. M. Gothard, W. E. Skaggs, B. L. McNaughton, J.
Neurosci. 16, 8027 (1996).
30. J. O’Keefe, M. L. Recce, Hippocampus 3, 317 (1993).
31. A. P. Maurer, S. L. Cowen, S. N. Burke, C. A. Barnes,
B. L. McNaughton, J. Neurosci. 26, 13485 (2006).
32. F. Sargolini et al., Science 312, 758 (2006).
33. B. L. McNaughton, F. P. Battaglia, O. Jensen, E. I. Moser,
M. B. Moser, Nat. Rev. Neurosci. 7, 663 (2006).
34. S. Fujisawa, A. Amarasingham, M. T. Harrison,
G. Buzsáki, Nat. Neurosci. 11, 823 (2008).
35. X. G. Li, P. Somogyi, A. Ylinen, G. Buzsáki, J. Comp.
Neurol. 339, 181 (1994).
36. G. Kreiman, C. Koch, I. Fried, Nature 408, 357 (2000).
37. J. E. Lisman, Neuron 22, 233 (1999).
38. L. F. Abbott, W. G. Regehr, Nature 431, 796 (2004).
39. M. Rabinovich, R. Huerta, G. Laurent, Science 321, 48
(2008).
40. M. B. Zugaro, L. Monconduit, G. Buzsáki, Nat. Neurosci.
8, 67 (2005).
41. We thank H. Hirase for sharing his data and C. Curto,
C. Geisler, S. Ozen, S. Fujisawa, K. Mizuseki, A. Sirota,
D. W. Sullivan, and R. L. Wright for comments.
Supported by NIH (NS34994 and MH54671), NSF (SBE
0542013), the James S. McDonnell Foundation, NSF
(A.A.), the Swartz Foundation (V.I.), and the Robert Leet
and Clara Guthrie Patterson Trust (E.P.).
Supporting Online Material
www.sciencemag.org/cgi/content/full/321/5894/1322/DC1
SOM Text
Figs. S1 to S13
Table S1
Movie S1
References
29 April 2008; accepted 29 July 2008
10.1126/science.1159775

www.sciencemag.org SCIENCE VOL 321 5 SEPTEMBER 2008 1327

 www.sciencemag.org/cgi/content/full/321/5894/1322/DC1

Supporting Online Material for

Internally Generated Cell Assembly Sequences in the Rat Hippocampus
Eva Pastalkova, Vladimir Itskov, Asohan Amarasingham, György Buzsáki*

*To whom correspondence should be addressed. E-mail: buzsaki@axon.rutgers.edu

Published 5 September 2008, Science 321, 1322 (2008)

DOI: 10.1126/science.1159775

This PDF file includes:

SOM Text
Figs. S1 to S13
Table S1
References

Other Supporting Online Material for this manuscript includes the following: (available at
www.sciencemag.org/cgi/content/full/321/5894/1322/DC1)

Movie S1
Supporting Online Material

Surgery and recording

Three male Long-Evans rats (330-400 g) were implanted bilaterally with 32
and/or 64-site silicon probes (4 or 8 shanks 200 µm apart, 8 recording sites per
shank, 20 µm spacing between the sites, 1) under isoflurane anesthesia. Rat 4
was implanted with 4 tetrodes (2). The silicon probes/tetrodes were lowered into
hippocampal CA1 pyramidal layer within two weeks after surgery. After full
recovery (approximately 2 weeks after surgery), access to water was restricted
and the rats were allowed to explore the task maze and identify the position of
the water ports (Fig. 1A). During physiological recording, data from all channels
were filtered (1Hz-5kHz), amplified (gain = 1000) and continuously sampled at 20
kHz on a 128-channel DataMax system (16-bit resolution). Two small light-
emitting diodes (5-cm separation), mounted above the headstage, were recorded
by an overhead digital video camera and sampled (at 30 Hz). For offline spike
sorting, the wideband signals were digitally high-pass filtered (0.8-5 kHz). Spike
sorting was performed semi-automatically, followed by manual adjustment of the
clusters (3-5). We cannot exclude the possibility that some neurons recorded in
different sessions were identical because spikes from each session were sorted
separately. However, the likelihood of recording from the same neuron is small
either because the probes moved spontaneously across recording sessions, as
evidenced by the change in ripple amplitude or the loss/appearance of fast firing
interneurons, or because the probe was moved by the experimenter. All protocols
were approved by the Institutional Animal Care and Use Committee of Rutgers
University.

Behavioral training and recording

Delayed spontaneous alternation task. The maze (100 cm by 120 cm) had a
shape of figure 8 with two water spouts. A delay area was located at the junction
of three arms between the reward sites (Fig. 1A). The delay area was separated
from the two reward sites and the third arm (main stem) by three doors,

1
controlled by the experimenter. The running wheel (10 cm wide, 29.5 cm
diameter) was attached to a wall of the delay area (Fig. 1A). Speed and direction
of the wheel rotation was tracked with an optical sensor (0.72o resolution),
attached to the axis of the wheel rotation. Infrared-beam photo-detectors were
fixed in all arms of the maze and were used to track the movement of an animal
in the maze and to control the delivery of water reward (0.05 ml) after correct
choices.
Each session began by placing the rat into the delay area while all three doors
were closed. During the first day of the training the animal was required to run in
the wheel for at least 3 sec, after which the main-stem door was opened. Rats
were not allowed to return to the main stem after making a choice at the T-
junction of arms. After the consumption of the reward (or no reward), rats walked
back to the delay area and could initiate a new trial by entering the wheel. The
required length of a wheel running was gradually increased from the initial 3 sec
up to 10 sec in rat 1 and up to 20 sec in rats 2 and 3. All rats were required to run
steadily in the wheel, always facing the same direction in each trial (West, Fig. 1,
S1). All rats reached high performance of spontaneous alternation (>90% correct
choices) and steady wheel running within a week. A session was terminated after
the animal failed to initiate a new trial within 2 minutes. After entering the wheel,
the velocity of running increased rapidly and reached a steady level within a few
seconds (Fig. S1). This stereotypic initial acceleration was used as a trigger to
mark the beginning of a trial in offline data processing in both the memory task
and in the control tasks (threshold speed: 20cm/sec).

Control task 1: Wheel running for a water reward. Rats 3 and 4 were trained to
run in a running wheel for a water reward. A wheel, identical to the one used in
the maze, was attached to one side of a rectangular box (30x40x35 cm). A water
port was placed on the wall opposite to the wheel. At the beginning of each
session, a rat was introduced into a box and required to run for a fixed period of
time (rat 3, 20sec; rat 4, 10sec). As in the alternation task, rats were trained to
run with steady speed and heading always in the same direction. All trials were

2
rewarded by water (2, 6).

Control task 2: Spontaneous wheel running. A wheel, identical to the one used in
the maze, was attached to the home cage of the rat (30 x 40 x 35 cm). The rats
(rat 2 and 3) had free access to the wheel and could run in either direction. The
trials were defined as at least 10 second-long periods of wheel running with the
running speed maintained above 50 cm/sec and preceded by at least 2 seconds
of exploration in the home cage. Maze and control task recording sessions were
alternated in rats 2 and 3.

Hippocampal inactivation
Five additional male rats (335 - 400 g) were implanted with stainless steel
injection guide cannulae (27 gauges) above dorsal hippocampi bilaterally (7-8)
under Isoflurane anesthesia. After recovery, the rats were water restricted and
trained in both non-delayed (zero) and delayed (15 sec) versions of left/right
alternation task until their performance reached a steady level (5 days). These
rats were not required to run in the wheel but were simply confined in the delay
area for the duration of the delay. Lidocaine (4%, 0.5µl) or saline were slowly (5
min) injected into hippocampi bilaterally through a 33-gauge cannula connected
to a 10-µl Hamilton syringe through tygon tubing. The injection cannula was left
at its position for 1min before and after an injection. Five minutes after the
injection, the rat was tested in either the non-delayed or delayed version of the
alternation task. The order of lidocaine and saline injections and order of delayed
and non-delayed versions of the task varied across animals.

Data processing
Place and episode fields. On each trial each neuron's spike train was convolved
with a Gaussian (SD = 100msec, sampling rate 1250 Hz). Firing rate profile of
each neuron triggered at the beginning of a wheel or a maze run (‘post-stimulus
time histogram’ with the moment of wheel/maze running initiation at time zero)
was calculated from action potentials generated during trails of the same type (e.

3
g., during wheel or maze runs before left or right turn). In order to distinguish
between place fields on a maze and bouts of increased firing generated by
pyramidal neurons during wheel running, we refer to the areas of an increased
firing rate in a wheel as “episode fields”. Place fields on the maze and episode
fields in the wheel were defined by a minimal peak firing rate of 6.0 Hz or 5Hz
and peak firing rate being at least 4.5-times or 3-times SD above a mean firing
rate, respectively. Using these different criteria yielded similar results. The width
of individual fields was determined by the onset and offset of >10 % rate of the
peak firing rate of that field (9-10).
Neuronal sequences. In each session two firing rate profiles for each neuron
were constructed and normalized by their peak firing rate: one from action
potentials generated during wheel runs before correct left arm choice (“left trials”)
and one from action potentials generated during wheel runs before correct right
arm choice (“right trials”) (e. g., Fig. 1D, E). Mean rate histograms of all neurons
within each trial type (left or right) were ordered according to the latencies of
peak firing rates of individual neurons (e.g., Fig. 1E). To establish an estimate of
statistically reliable firing rate differences between left and right trial types we
shuffled left and right single trial firing rate histograms in a randomized manner
(11). Thus, we examined the null hypothesis whether real firing patterns of
neurons during left and right trials differed significantly, and at which portion of
trials these differences occurred. Note that this approach is more rigorous than
simply comparing firing rate differences between left and right trials and may at
least partially explain the low number of trial discriminating neurons in the stem of
the maze, compared to previous findings (12-17).
Population vector correlation. In the spatial domain, “spatial scale factor” of the
hippocampal population is a measure of distance traveled by a rat over which
temporally overlapping neuronal populations become de-correlated (18-20).
Analogously, we calculated a “temporal scale factor”, which estimates an amount
of time it took for population activity to become de-correlated (‘life-time’ of
population activity). Firing rate histograms of all neurons within each trial type
(see above) were binned into 100 msec bins. Thus, activity of the whole neuronal

4
population within each 100 msec time bin was described by a vector that had a
length of the number of active neurons (population vector). Spearman rank
correlation between all pairs of population vectors characterizing all time bins
was calculated. Values of correlation coefficients were color-coded and plotted
(Fig 1G). The ‘life-time’ of population activity was characterized as a width of a
region along the diagonal of a correlation matrix with increased correlation
coefficient values (19-20).
Correlation of population vectors representing the same time bin is equal to 1
(and thus all values along a diagonal of a correlation matrix are equal to 1) unless
different subsets of trials are used to build population vectors representing the
same time bin (19-20, Fig. 1G). In order to exclude a bias of auto-correlation
values along a correlation matrix diagonal, we constructed population vectors
from activity generated during trialsn and trialsn+1. Due to the trial-by-trial
variability of firing rates, the value of correlation between these population
vectors is typically smaller than 1 (Fig. 3B, 5C, only pixels with significant
correlation coefficient values are shown, p < 0.01). In the memory task, the
correlation matrix was calculated from activity generated during trials with the
same future choices (L-trialsn vs. L-trialsn+1 and R-trialsn vs. R-trialsn+1). In the
control tasks, the correlation matrix was calculated from activity generated during
even and odd trials (trialsn vs. trialsn+1; Fig. S9).
Phase precession. Instantaneous theta phase was derived by Hilbert transform
from filtered (4-10 Hz) LFP trace and a phase value was assigned to each action
potential. Spearman rank correlation was used to estimate relationship between
the phase precession slope and place/episode field size (9).
Frequency of single unit oscillation. All action potentials emitted by a neuron
during each pass through a field were convolved with a Gaussian function (SD =
1 msec), convolved with the first four Slepian functions and the dominant
oscillations frequency was estimated with Fast Fourier Transform, as described
in detail earlier (9). The frequency shift between the unit and LFP power spectra
was estimated by cross-correlation. Only units with significant theta modulation
were included in this analysis.

5
Compression index. Correlation of the distance between place/episode fields of
two neurons in a maze or a wheel with the mean temporal offset of action
potentials of the same pair of neurons within a theta cycle is called ‘compression
index’ (10). Only pairs of pyramidal neurons with robust place/episode fields and
significant cross-correlogram peaks were used for this analysis (10). We obtained
similar compression index values (0.75, 0.62, 0.59, p < 0.0001 for all) after
selecting place fields based on three different sets of parameters: minimal peak
firing rate of 6 Hz, 4.5 Hz or 5 Hz and the peak firing rate 4.5, 3.4 or 3 SDabove
the mean firing rate.
Firing ‘fields’ of neurons. Neurons in the control tasks, strictly speaking, did not
have fields. We compared activity of neurons generated during wheel running in
the memory and control tasks in the temporal domain. An autocorrelogram of
each neuron was filtered between 0.2 Hz and 2 Hz and the local minima of the
filtered trace closest to zero were detected. “Intra-field” spikes were defined as
spikes which occurred within 10% boundaries of the peak rate, and “extra-field”
spikes those that occurred outside of this boundary. This index is high for
compact episode/place fields and low for neurons with sustained or random
spiking patterns.
Inferring behavioral choice from neural activity. In addition to identifying neurons
that reliably discriminate between left and right trials in some portion of wheel
running, we also used a probabilistic model to predict future L/R trial type from
total population activity during the preceding wheel run. To infer the L/R trial type
for a given trial, we fit the model on all other trials in the same session and used it
to predict the remaining trial type. The set of trials on which the model was fit is
called the training set; the trial that is predicted from the model is called the test
set. The percentage of correctly inferred trials was computed across all
training/test set trial combinations, with leave-one-out cross-validation. The
significance of this prediction was determined on a session-by-session basis by
comparing the true values with the distribution of shuffled data. Each ‘shuffled’
data set was constructed by randomly shuffling the left/right labels between the
trial spike trains from each cell across all eligible trials. Only behaviorally correct

6
trials of length ≥ 7s for rat 1, and ≥ 17s for rats 2 and 3 were considered for
model-fitting and for correct trial-type inference. For each training set and each k-
th putative pyramidal cell that had an average firing rate of > 0.1Hz we computed
the “predictability index” Ik from the ratio of minimal and maximal average spike-
counts across left and right trials:
L R
min( N k , N k )
Ik = 1 − L R ,
max( N k , N k )
L
(here N k is the average spike count of the k-th neuron across the considered left
R
trials, and N k is the average spike count of the k-th neuron across the
considered right trials.) The higher the value of Ik the more selective the cell on
the training set (Ik = 1 for the most selective, and Ik = 0 for the least selective.)
While this index is not necessarily indicative of statistical significance of left/right
selectivity of a given cell, it was used for ranking cells in order of their tendency
to be predictive.
Putative pyramidal cells were ordered according to Ik in each training set. Ten
cells with the highest predictability index were selected for prediction in each
G
training set. For each trial we computed a binary population vector u = (u1 ,.., u10 )
as follows. For each trial the cell could be in one of two states: uk=“high” if

N k / N k >= 0.25 , and uk=“low” if N k / N k < 0.25 , where N k is the average spike-

count of the considered cell across all trials. We denote by Pqk (u ) the

probabilities of observing k-th selected neuron (k = 1,…,10) in the state u= “low”

or u=“high” during a trial of type q=L,R:
Pqk (u ) =Prob( k-th neuron is in the state u | given trial type q).

Note that because there are only two states, Pqk (low) = 1 − Pqk (high) . We assume

that the neurons fire independently from each other. This is of course an
incorrect assumption. However the prediction (based on this assumption) can be
only worse than a prediction that would take correlations into account. Under this
G
assumption the probability of observing a binary population vector u = (u1 ,.., uN )

7
during trial type q can be computed as
G N
Pq (u ) = ∏ Pqk (uk )
k =1
G
Given a new test-set trial, with binary population vector u , the model predicts the
trial type q pred (=L or R) via maximum-likelihood:
G G G G G G
q pred (u ) = L if PR (u ) ≤ PL (u ) and q pred (u ) = R if PR (u ) > PL (u ) .

The same analysis was repeated on 100 shuffled data sets. The prediction for a
given session was considered to be significant if the percentage of trials correctly
inferred from the real data was higher than the percentage of trials correctly
inferred from at least 95 of the shuffled data sets.
Error trials. For each session, only those error trials were selected that were
longer than 7s for rat 1, and 17s for rats 2 and 3 (as in the selection of
behaviorally correct trials). In order to predict error trials, the model was fit on all
behaviorally correct trials in the same session. This model was then used to
predict the L/R type of each error trial in a session, as described above.

8
Supplementary figures

Fig. S1. Motor correlates of wheel running A. Top view of the testing apparatus.
Gray lines: trajectory of the animal. Black and green, position of front and rear
LEDs during wheel running (all wheel runs within one recording session). B.
Head position of the rat shown separately for the wheel running trials with future
left and right choices (trials) in the maze. C. Grand mean head-direction angular
vectors for all analyzed sessions (3 rats combined, red arrows: mean of all trials
from one session followed by left turn choices, blue: mean of all trials from one
session followed by right turn choices). D. Examples of single running wheel trial
mean head-direction angles (black) and session mean (red, left trials, blue, right

9
trials) head direction angles. Data from the same rat as shown in A and B. E. top:
Running speed during single wheel running trials (the same color-scheme as in
the bottom panel, the same recording sessions as in Fig. 5 and Fig. S13).
Bottom: Average running speed during correct left, correct right and error trials
(the same data as in the upper panel). F. The significance level of speed
differences between left and right trials was estimated based by the shuffling of
the left and right speed labels (green lines, p<0.025 each, 11). Running speed
during left and right trials did not differ significantly in any of the analyzed
sessions. G. Mean running speed during L and R trials of all sessions of the
memory tasks and during all trials of the control tasks. Note that head position,
head direction and speed are not significantly different between trials followed by
different behavioral choices. These observations indicate that idiothetic cues
alone cannot be responsible for the choice-specific firing patterns of neurons
(Fig. 4; 5, 12, 14).

Fig. S2. Firing rates of putative interneurons in the wheel and in the maze are
correlated. In contrast to pyramidal neurons (episode cells and place cells, Fig.
1C), the firing rates of putative interneurons were correlated significantly in the
wheel and maze. This may occur because interneurons are typically controlled

10
by multiple pyramidal cell assemblies (19, 9).

Fig. S3. Proportion of active neurons during wheel and maze running is similar at
each moment in time. Percentage of neurons firing at least one spike within 100
msec windows during maze (black) and wheel (blue) running (dashed lines: SD).
Note that similar fraction of neurons is active in the wheel and in the maze at
each moment in time, despite the spatial 'over-representation' of the head
position in the wheel by multiple neurons (Fig. 1B). This can occur because
different episode cells in the wheel are activated in a sequence.

11
Fig. S4. Control task 1. Rats were trained to run for water reward. Rat 4, twelve
simultaneously recorded neurons (each panel shows activity of one neuron,
rows: peak-normalized firing rate profiles of a neuron in 21 successive wheel
running trials). Rat 3, 15 simultaneously recorded neurons in a session with 58
trials. Only pyramidal neurons with mean firing rate higher than 0.35 Hz are

12
shown. Note that most neurons fire in a relatively sustained manner (2, 6) and
that firing rate profiles of almost all neurons vary from trial to trial.

Fig. S5. Control task 2. Rats could freely run in the wheel, attached to the wall of
a home cage. No reward was delivered. Each panel shows activity of one
neuron; all 16 neurons were recorded simultaneously (rat 2; single session).
Rows show peak-normalized firing rate profiles of a neuron in successive trials.
Trials are grouped according to the animal’s head-direction/sense of the wheel
rotation during trials (counter clockwise – CCW, clockwise – CW). Note that most
neurons show differential firing rates depending on the rat’s head direction. This
observation emphasizes the dominance of distant cues in this control task (2, 6).

13
Fig. S6. Distribution of mean firing rates of neurons in the memory and control
tasks. Note similar distribution of firing rates in both tasks. Arrows, medians.

Fig. S7. Performance of rats in the delayed alternation task is impaired after
bilateral inactivation of dorsal hippocampi. Five rats (different from rats 1-4 and

14
not equipped with electrodes) were trained in the alternation task with both 0 and
15 sec delays. Bilateral inactivation of dorsal hippocampi (4% lidocaine, 0.5 µl)
impaired performance of the animals during sessions with 15 sec delay between
trials but did not affect performance of the same animals during sessions with no
waiting period between trials (see also ref. 17 for similar results).

15
Fig. S8. Episode fields in the memory task. Activity of individual neurons during
wheel running predicts future behavior of an animal in the maze. Firing patterns

16
of 32 simultaneously recorded neurons in Rat 2 during wheel running in the
delayed alternation task, separated on the basis of future left and right correct
turns in the maze. Each line corresponds to a single trial. Clu, cluster identity.
FR: Firing rate (Hz). Note that many neurons fire only for a short duration (“life
time”) and consistently at the same time after the beginning of wheel running.
Note also robust firing pattern differences between left and right trials in several
neurons.

Fig. S9. Prediction of behavioral choices in the maze from the population activity
during wheel running. In addition to identifying individual neurons with differential
firing rate patterns before left and right choices (Fig. 4), we also examined how
well population activity can predict correct choices. In this analysis, we used the
best 10 discriminating neurons (based on mean spike count during wheel running
before left and right choices), irrespective of whether they individually showed
significant choice differences. In order to infer the left/right trial type for a given
trial, we fit a model on all other trials (excluding the one to be predicted) and
used the model to predict the remaining trial type (leave-one-out cross-validation

17
procedure). Black bars indicate the percentage of correctly predicted trials during
each session. The significance of the prediction was determined by comparing
the original data set with shuffled data from the same session (box plots; see
Methods). Box plots display median, the first and third quartiles and 95%
confidence intervals for model predictions based on shuffled data sets. Numbers
at the bottom of each bar indicate the fraction of neurons that significantly
differentiated between left and right trials on at least a portion of wheel running
trials (Fig. 4B) compared with the total number of pyramidal neurons within a
session. Rat 3 lost the implant after the last recording session.

18
Fig. S10. Firing rates of some interneurons correlate with specific aspects of the
memory task. Each map shows activity of a putative interneuron. Rows, wheel
running trials separated according to the future choices in the maze (left, right).
Neurons from 3 rats (n = 77 neurons). Several interneurons show increase or
decrease of a firing rate at a specific portion of a wheel run (as visualized by the

19
vertical color-stripes). Interneurons, whose firing patterns were correlated with
future left or right choices are highlighted (black frames). These observations are
in register with recent observations reporting that a fraction of interneurons also
show place or other task-related specificity (19, 21, 9).

Fig. S11. Population vector cross-correlation matrices for a single session in

each rat (rows). Matrices were calculated from population vectors representing
activity during trials with either the same (Left: L-trialn vs L-trialn+1; Right: R-trialn+1
vs R-trialn+1, Supplementary Methods) or different future choice (Left vs. Right: L-
trialn vs R-trialn). The trial-to-trial consistency of subsequent trials that were
followed by the same arm choice can be judged from the intensity of the
diagonal. Note the lack of diagonal in the matrices calculated from trials followed

20
by opposite choices (third column).

Fig. S12. Behavioral errors in the maze can be predicted from firing patterns
during wheel running. Two representative sessions from 2 rats (each session
with 9 left and 9 right correct trials and 2 error trials). For each rat, 10 best
left/right discriminating neurons (color-coded dots) from a larger population were
chosen. Note that firing patterns of right error trials and right correct trials are
similar. These findings suggest that assembly sequences of neurons reflect
planned travel path of the rat. During most error trials the rat ran smoothly in the

21
wheel and through the center arm as well as the T-junction in one swift trajectory,
and the commission error became detectable only after the wrong behavioral
turn. Deadwyler et al. (23) referred to this kind of error as “miscoding”, implying
that the previous choice is not correctly interpreted and the rat behaves
indistinguishably from correct trials. The second type of error trials was
associated with overt behavioral changes, such as stopping and premature
departure from the wheel and return to the wheel, grooming and immobility in the
wheel or delay area (analogous to the “delay-dependent” errors in ref. 23). The
neuronal patterns associated with such behaviorally abnormal trials were often
not possible to predict reliably from neuronal patterns, mainly because
comparable correct reference assembly patterns were not available.
Nevertheless, our anecdotal observations indicate that even such late errors
could be distinguished by neurons with perfect left vs right discrimination (e.g.,
error trial in Movie 1). In the presented analyses (Fig. 5) all errors are presented,
since the main goal was to differentiate between the effect of past and future
movement paths of the rat.

22
Fig. S13. Supplementary information for Fig. 5. A. Peak rate-normalized firing

23
rates of 43 neurons simultaneously recorded during wheel running, ordered by
the latency of peak firing rates during correct left trials (left) and correct right trials
(right). Note that while in Fig. 5B neuron order 1 to 43 is the same for both
panels, here the ordering of neurons for LEFT and RIGHT trials was done
separately. B. Activity of the same neurons as in A during seven error trials.
Neuronal order in the left column is the same as in the correct-left matrix in A.
Neuronal order in the right column is the same as in the correct-right matrix in A.
C. Correlation between population vectors from the left-correct and left-error
matrices (left column) and correlation between population vectors from the right-
correct and right-error matrices (right column). Mean correlation coefficient along
the correlation matrix diagonal (± 0.5 sec) is given in the upper right corner of
each correlation matrix. Note that left correct vs. left error mean correlation
coefficient is higher than right correct vs. right error correlation coefficient in all 7
left-error trials.

Table S1. Variability between animals: selected behavioral and

electrophysiological measures for rat 1-3.

24
Movie 1. Single neurons can reliably predict future left and right turns. Running
wheel on the right, the right and left side arms of the maze are on the top and
bottom of the movie, respectively. Color LEDs indicate head orientation. This
strongly discriminating neuron was virtually silent during wheel runs before right
turns. Note that on the error trial (6th), the erroneous turn to the left is predicted
by the increased firing of the cell at the end of the delay period. On another trial
(last one in movie), the rat headed back toward the wheel. Prior to turning back
toward the T junction the neuron became active followed by the neuron-predicted
behavioral choice to the left. Although this neuron reliably discriminated left and
right turns, it alone cannot be responsible for guiding future motor behavior since
it fired transiently at the beginning of wheel running. This information should be

25
passed onto further neuronal assemblies, which continue to discriminate
between left and right movement trajectories during later parts of running in the
wheel and in the stem area (“prospective” neurons, 12-14).

Supplementary References
1. J. Csicsvari, D. A. Henze, B. Jamieson, K. D. Harris, A. Sirota, P. Barthó, K. D.
Wise, G. Buzsáki, J. Neurophysiol. 90, 1314-23 (2003).
2. H. Hirase, A. Czurko, J. Csicsvari, G. Buzsáki, Eur J Neurosci 11, 4373-4380 (1999).
3. K. D. Harris, D. A. Henze, J. Csicsvari, H. Hirase, G. Buzsáki, J. Neurophysiol.
84, 401–414 (2000).
4. J. Csicsvari, H. Hirase, A. Czurko, G. Buzsáki, Neuron 21, 179–189 (1998).
5. L. Hazan, M. Zugaro, G. Buzsáki, J Neurosci Methods 155, 207–216 (2006).
6. A. Czurko, H. Hirase, J. Csicsvari, G. Buzsáki, Eur J Neurosci 11, 344-352 (1999).
7. D. Klement, E. Pastalkova, A. A. Fenton, Hippocampus, 15, 460–471 (2005).
8. E. Pastalkova, P. Serrano, D. Pinkhasova, E. Wallace, A. A. Fenton, T. C. Sacktor,
Science 313, 1141-1144 (2006).
9. C. Geisler, D. Robbe, M. Zugaro, A. Sirota, G. Buzsáki, Proc. Natl. Acad. Sci. U. S. A.
104, 8149-54 (2007).
10. G. Dragoi, G. Buzsáki, Neuron 50, 145-57 (2006).
11. S. Fujisawa, A. Amarasingham, M. T. Harrison, G. Buzsáki, Nature Neurosci 11,
823 - 833 (2008).
12. L. M. Frank, E. N. Brown, M. Wilson, Neuron 27, 169-178 (2000).
13. J. Ferbinteanu, M. L. Shapiro, Neuron 40, 1227-1239, (2003).
14. E. R. Wood, P. A. Dudchenko, R. J. Robitsek, H. Eichenbaum, Neuron 27,
623-633 (2000).
15. M. R. Bower, D. R. Euston, B. L. McNaughton, J Neurosci 25, 1313-23
(2005).
16. J. A. Ainge, M. Tamosiunaite F. Woergoetter, P. A. Dudchenko, J Neurosci
27, 9769-79 (2007).
17. J. A. Ainge, M. A. van der Meer, R. F. Langston, E. R. Wood, Hippocampus

26
17, 988-1002 (2007).
18. K. M. Gothard, W. E. Skaggs, B. L. McNaughton, J Neurosci 16, 8027-8040 (1996).
19. A. P. Maurer, S. L. Cowen, S. N. Burke, C. A. Barnes, B. L. McNaughton, J.
Neurosci. 26, 13485-92 (2006).
20. A. Terrazas, M. Krause, P. Lipa, K. M. Gothard, C. A. Barnes, B. L.
McNaughton, J Neurosci 25, 8085-96 (2005).
21. W. B. Wilent, D. A. Nitz, J. Neurophysiol. 97, 4152-61 (2007).
22. V. Ego-Stengel, M. A. Wilson, Hippocampus, 17, 161-174 (2007).
23. A. Deadwyler, T. Bunn, R. E. Hampson, J Neurosci 16, 354-72 (1996).

27