Nested sequences of hippocampal assemblies during behavior support subsequent sleep replay

See allHide authors and affiliations

Science  09 Nov 2018:
Vol. 362, Issue 6415, pp. 675-679
DOI: 10.1126/science.aat2952

Theta rhythm protects sleep replay

Hippocampal replay of place cell sequences during sleep is critical for memory consolidation in target cortical areas. How is the sequential organization of place cell assemblies maintained across different time scales? Drieu et al. compared periods when a rat either sat passively on a moving train or ran actively on a treadmill on the same train. During the passive movement, the slow behavioral sequence of place cells was still present, but the rapid generation of theta sequences was lost. Active running on the treadmill, however, maintained the theta rhythm. After passive transport, sequence replay during sleep was destroyed, whereas active running protected replay.

Science, this issue p. 675


Consolidation of spatial and episodic memories is thought to rely on replay of neuronal activity sequences during sleep. However, the network dynamics underlying the initial storage of memories during wakefulness have never been tested. Although slow, behavioral time scale sequences have been claimed to sustain sequential memory formation, fast (“theta”) time scale sequences, nested within slow sequences, could be instrumental. We found that in rats traveling passively on a model train, place cells formed behavioral time scale sequences but theta sequences were degraded, resulting in impaired subsequent sleep replay. In contrast, when the rats actively ran on a treadmill while being transported on the train, place cells generated clear theta sequences and accurate trajectory replay during sleep. Our results support the view that nested sequences underlie the initial formation of memory traces subsequently consolidated during sleep.

The sequential activation of neuronal ensembles is a ubiquitous brain coding scheme possibly underlying numerous diverse behaviors (14). Sequential neuronal activity occurs at different time scales, ranging from slow (“behavioral”) time scales, where the dynamics are constrained by stimulus or motor time constants, to fast (“endogenous”) time scales mostly driven by intrinsic network properties. A prominent example is the hippocampus, where place cells code for an animal’s location (5). As the animal explores its environment, different place cell ensembles become successively active along the ongoing trajectory, yielding sequences of neuronal activity at the behavioral time scale. During subsequent slow-wave sleep (SWS), the same sequences are endogenously replayed during sharp-wave ripple complexes at a highly accelerated (20×) time scale (6), mediating memory consolidation during sleep (710). How is the sequential organization of place cell assemblies maintained across these two time scales, expressed at entirely disjoint moments in time and during different brain states?

Sequential information could be readily stored during exploration in the hippocampal network by the sequential activation of place cells at the behavioral time scale, via a recently discovered form of behavioral time scale plasticity (11). An intriguing alternative is that sequential structure is stored via the remarkable ability of the hippocampal network to generate nested sequences of cell assemblies, whereby both slow and fast neural sequences are intermingled in time. This occurs during spatial navigation: Nested within behavioral time scale sequences, the hippocampal network also produces sequences at the theta time scale [one sequence per theta cycle of ~150 ms (1214)], allowing cell assemblies to fire within brief delays [~25 ms (15)] compatible with classical Hebbian plasticity, such as spike timing–dependent plasticity (16). Are these nested sequences of hippocampal cell assemblies required for subsequent sleep replay (12, 13), or are they merely an epiphenomenon deriving from preexisting connectivity within the hippocampal network (1719)?

Contrasting these predictions requires a protocol that selectively disrupts fast theta sequences but preserves slow behavioral sequences of place cell assemblies. Further crucial constraints are the ability to trigger or release theta sequence disruption with temporal precision, as well as the necessity to target the entire hippocampal formation, both in terms of extent (to overcome information redundancy along the septotemporal axis of the hippocampus) and in terms of fields (to overcome compensatory network mechanisms such as pattern completion in CA3, which could restore locally induced impairments). During passive transportation in space, place cells remain spatially selective, but their precise timing relative to theta [“phase precession” (12)] is altered (20) unless the animals actively run on an onboard treadmill (21). We thus transported rats on a model train (fig. S1A) and turned the onboard treadmill off (Passive) or on (Active) to respectively perturb nested sequences or leave them intact. The goal was to determine whether intact nested sequences were required for subsequent replay during SWS. The rats were tested in an entirely novel environment, different from the training room. Hence, hippocampal activity was monitored as the animals learned a novel spatial context, which is known to induce the formation of a novel hippocampal map, increase network coordination, boost replay, and enhance plasticity (2225). Further, because the rats were tested on a single day, this avoided any confounding network changes that could have resulted from previous experience.

We recorded CA1 pyramidal units and local field potentials (LFPs) in five rats. After a baseline sleep session, the rats underwent three travel sessions (Passive 1, Active, Passive 2) interspersed with sleep sessions (Fig. 1A). The train velocity, number of laps, and travel duration were similar in all conditions for all rats (fig. S1, B to D). We first verified the presence of place cell sequences at the behavioral time scale in all three conditions. Pyramidal cells always coded for the location of the animal in space (Fig. 1B and fig. S2). Their fields remained similar in terms of size (Fig. 1C and fig. S3) and together covered the entire train track (fig. S2C). However, peak firing rates and place cell count were somewhat reduced during passive travels [Fig. 1C and fig. S2B (20)]; specific controls for these factors are provided in ensuing analyses (see below). Incidentally, we also confirmed that place fields did not undergo random remapping (Fig. 1, D and E, and fig. S2D). This finding provided further verification that hippocampal dynamics were virtually identical at the slow time scale in all travel sessions.

Fig. 1 Maintenance of behavioral time scale properties.

(A) Behavioral protocol. In a novel environment, rats underwent three successive travel sessions (Passive 1–Active–Passive 2) interspersed with sleep recordings (see also fig. S1). Passive travel was intended to selectively perturb spike timing at the theta time scale but not at the behavioral time scale. (B) Linearized normalized firing curves of place cells recorded from one example rat across travel sessions, showing complete track coverage in all conditions. (C) Place field size was maintained across travel conditions [Kruskal-Wallis (KW) test, P > 0.05]. Peak firing rates were somewhat lower in Passive 1 than in Active (KW test, ***P < 10–3). (D) Firing fields did not remap between Passive 1 and Active. Left: Mean unsigned proportional shift σ (vertical gray line; **P < 0.01) and σ distribution for n = 2000 bootstrapped remapping data sets (black histograms). Inset: Circular distributions of angular differences (in degrees) between place field peak locations on the maze (Rayleigh test, P < 0.001; V-test against 0, P < 10–4). Right: Spatial cross-correlograms of firing fields across successive travel sessions (x axis normalized to field size; black dots represent correlogram modes; black histograms at top, mode distributions). (E) Firing fields also did not remap between Active and Passive 2. Data are displayed as in (D). Left: ***P ~ 0. Inset: Rayleigh test, P < 10–9; V-test against 0, P < 10–10.

Clear theta oscillations with similar frequencies were observed in all conditions (Fig. 2A). However, during passive travel, power was slightly reduced at both the fundamental frequency and the first harmonic, the latter resulting in decreased cycle asymmetry, indicative of a change in the internal structure of theta cycles (Fig. 2, A and B). This was expected to alter the precise spike timing of active cell assemblies. Consistently, place cells continued to oscillate slightly faster than theta (i.e., phase precession) during active travel, whereas this was reduced in passive travel (Fig. 2, C and D, and fig. S4, A and B), indicating an overall degradation of phase precession [Fig. 2, E and F; in phase-precessing neurons, phase range was similar in all conditions (26) (fig. S4, C to E)]. This degradation was noteworthy, because phase precession is thought to be instrumental for the formation of nested sequences: Place cells normally oscillate slightly faster than theta, so they emit spike bursts earlier and earlier in successive theta cycles; this [possibly combined with additional coordinating mechanisms (27, 28)] results in newly activated cells firing after those that have started firing in earlier cycles, effectively resulting in temporal sequences of activity.

Fig. 2 Perturbation of single-cell theta time scale properties.

(A) Theta maintenance across travel sessions. Top row: Example raw LFPs (duration, 1 s). Middle row: Power spectrograms (black dashed line, time of LFP traces shown above; black calibration bar, 15 s). Bottom row: Normalized power spectra (mean ± SEM). (B) Theta frequency was unchanged across conditions [top; repeated-measures analysis of variance (ANOVA), P > 0.90], whereas theta power (middle; repeated-measures ANOVA, P < 10–6) and shape (bottom, theta asymmetry; KW test, P < 10–84) were altered during passive travel (frequency and power, mean ± SEM). **P < 0.0033, ***P < 0.00033 for post hoc comparisons; **P < 0.01, ***P < 0.001 otherwise. (C and D) During passive travel, spike bursts recurred at lower rates, closer to baseline theta frequency. (C) Distributions of spectral modes of spike trains (measured relative to theta frequency, x axis trimmed at ±25% around theta frequency). (D) Spectral modes (***P < 10–6, *P < 0.02). (E and F) Theta-phase precession was perturbed during passive travel. (E) Average normalized phase precession density plots for significantly phase-precessing cells (blue and red indicate minimum and maximum spike density, respectively). (F) Distribution of phase precession slopes for all place cells (Kolmogorov-Smirnov test, **P < 0.008, *P < 0.04; colored dashed vertical lines are medians).

To directly assess how perturbation of phase precession affected theta sequences, we used a Bayesian reconstruction approach (21, 29, 30). Briefly, theta cycles were subdivided into six phase bins, and the sequential structure of reconstructed positions in these bins (“candidate events”) was evaluated using two previously described complementary measures (30): (i) trajectory scores [normalized to compare across animals and conditions (31)], which assess the quality (linearity) of the reconstructed events (i.e., whether the events represent spatially aligned positions versus mere series of random locations); and (ii) slopes, which estimate the speed at which reconstructed events are played and indicate whether the events move through space or remain merely stationary (absence of actual trajectories). Thus, clear theta sequences would be characterized by both high scores and slopes, whereas static representations of current position would result in high scores but low or zero slopes, and random activity would be associated with low scores (but possibly spuriously high slopes). Clear sequential structure was readily visible in individual theta cycles during active, but not passive, travel (Fig. 3A). This was confirmed over all theta cycles for all rats (Fig. 3B and fig. S5). Whereas normalized trajectory scores were significantly better than chance in all conditions (Fig. 3C), slopes were significantly steeper during active travel (Fig. 3D and fig. S5, C and D). Joint analysis of trajectory scores and slopes revealed a much greater proportion of high-value pairs during active travel (Fig. 3E). To confirm and extend these results using an independent measure, we computed the quadrant score (27) of each candidate event, which assessed the overall direction of reconstructed trajectories without assuming constant velocity. Quadrant scores were significantly greater for active travel (clear trajectories) and remained very low for passive travel (degraded trajectories) (Fig. 3F and fig. S5B).

Fig. 3 Perturbation of theta sequences.

(A) Top: Raw LFPs and place cell spikes in six example theta cycles (black dashed lines, theta peaks; place cells are ordered by their place field location on the track) in each travel condition (black calibration bars, 50 ms). Bottom: Bayesian reconstruction of position encoded in the ongoing activity of the hippocampal network (six phase windows per theta cycle; white vertical lines, theta peaks; white dashed lines, actual position of the animal). (B) Average Bayesian reconstruction of position (relative to actual position of the animal) across theta subcycles for all rats. Trajectory score and slope (in cm cycle−1) are indicated above each reconstruction. Two cycles are shown for clarity. (C) Left: Normalized score of theta sequences (KW test, P < 10–23; ***P < 0.00033 for post hoc comparisons). Right: Proportion of significant theta sequences (Passive 1: 7.95%, Active: 13.61%, Passive 2: 9.05%; all proportions are significantly greater than shuffled control proportions, Monte Carlo test, P ~ 0; binomial proportion tests, *P < 0.05, ***P < 0.001). (D) Distributions of significant theta sequence slopes (KS tests; left: Passive 1–Active, ***P < 10–16; right: Active–Passive 2, ***P < 10–18). Colored bands indicate significant differences. (E) Distribution of theta sequence quality assessed by joint trajectory score and slope (for all animals; color code indicates proportion normalized relative to shuffled control data). (F) Quadrant score computed from individual theta cycles (KW test, ***P < 10–24; *P < 0.017, ***P < 0.00033 for post hoc comparisons; **P < 0.01, ***P < 0.001 otherwise). (G) Pairwise bias correlation between awake theta sequences (ordered according to time of occurrence). Sequences were stable only during active travel. Note the absence of correlation during Passive 1 and progressive decay following Active during Passive 2.

Hence, theta sequences were degraded during passive travel. Note that reconstruction quality and quadrant scores were higher in Passive 2 than in Passive 1 (Fig. 3, E and F), implying that sequential structure was slightly less degraded in the second passive travel session (Fig. 3B). This was further supported by higher self-consistency of theta sequences (31) in Passive 2 than in Passive 1, whereas Active sequences were the most self-consistent (Fig. 3G and fig. S6A).

We ensured that our results could not be accounted for by differences in the number of simultaneously recorded place cells (fig. S7) and their firing rates (fig. S8). We also controlled for decoding quality (31) (fig. S9), ruling out a potential bias due to differences in spatial coding, at both the single-cell and population levels. Finally, because small variations in field locations between individual laps could have altered sequence detection, we also ruled out differences in firing variability between travel conditions (31) (fig. S10).

Taken together, the above results show that place cell sequences at the behavioral time scale were present in all three conditions, but theta sequences were disrupted during passive travel, when stationary network activity continued to reflect the ongoing position at the endogenous time scale. How did this affect subsequent activity during SWS? Candidate replay events were defined as transient surges in aggregate firing rate (30) during SWS epochs (fig. S11, A and C), which coincided with ripple events [fig. S11G; results remained unchanged when candidate events were restricted to ripples (fig. S12)]. The average SWS duration, number and rate of candidate events, and ripple occurrence rate were not significantly different across animals (fig. S11, B and D to F). To reconstruct replayed trajectories, we first trained the Bayesian decoder using place cell activity recorded during the preceding travel session, then tested it during SWS on candidate events subdivided into 20-ms nonoverlapping time windows (31).

Candidate replay events were evaluated using the same method as theta sequences, whereby genuine sequences are characterized by both high trajectory scores and slopes. Although candidate events with significant trajectory scores were present in all three sleep sessions, overall the reconstructed trajectories were notably sharper after active travel (Fig. 4A and fig. S13A), and scores were significantly improved relative to baseline (31) only in sleep sessions following Active and Passive 2 (Fig. 4B). In addition, slopes were significantly steeper following Active (Fig. 4C and fig. S13B). Joint analysis of trajectory scores and slopes confirmed a much greater proportion of high-value pairs after active travel (Fig. 4D and fig. S13, C and D).

Fig. 4 Replay is degraded following perturbation of theta sequences.

(A) Examples of significant replay events in sleep sessions (see fig. S16 for more examples). (B) Scores of replay events relative to baseline sleep (KW test, P < 10–17; *P < 0.017, ***P < 0.00033 for post hoc comparisons; ***P < 0.001 otherwise). (C) Absolute slopes of replay events (KW test, P < 10–27; ***P < 0.00033 for post hoc comparisons). (D) Distribution of replay quality assessed by joint trajectory score and slope (for all animals; color denotes proportion normalized relative to shuffled control data). (E) Proportion of significant replay events relative to baseline sleep (binomial proportion tests, **P < 0.01, ***P < 0.001). Sleep replay was boosted after active travel relative to after passive travel. (F) Proportion of forward (darker colors) versus reverse (lighter colors) replay events (binomial proportion tests, ***P < 0.001). (G) Pairwise bias correlation between awake theta sequences and sleep replay (ordered according to time of occurrence).

We next addressed the critical question of whether these trajectories did constitute genuine replay of awake behavior, or merely reflected preexisting connectivity patterns independent of experience. We compared the proportion of significant trajectories in each sleep session relative to baseline sleep—that is, relative to the sleep session preceding any exploration of the environment (as noted above, recordings took place in an entirely novel environment). We did not observe replay during sleep following Passive 1 (Fig. 4E), when theta sequences had been selectively disrupted (Fig. 3). Trajectory replay was then boosted in sleep following Active (Fig. 4E), when nested sequences had remained intact (Fig. 3). Finally, an intermediate but significant level of reactivation was observed following Passive 2 (Fig. 4E), when theta sequences had been perturbed to a lesser degree than during Passive 1 (Fig. 3). Although forward and backward trajectories were found in equal proportions during sleep after passive travel, only active travel resulted in a greater proportion of forward replay actually reflecting awake experience (Fig. 4F). Finally, a direct comparison of theta and replay sequences (31) highlighted a selective correlation between theta sequences in Active and candidate replay events in subsequent sleep (Fig. 4G and fig. S6, B and C).

Our results thus show that during sleep after selective disruption of theta sequences (Passive 1), the proportion of significant trajectories remained at baseline levels observed prior to experience. By contrast, intact nested sequences (Active) resulted in boosted replay during sleep, and trajectories in hippocampal space were preferentially replayed in the same direction as in physical space. In Passive 2, theta sequences were perturbed to a lesser degree than in Passive 1, yielding intermediate levels of trajectory replay during subsequent sleep. This result has three implications. First, repeated experience alone cannot account for the improved replay following Active, because replay was degraded following Passive 2. Second, the improved theta sequences during Passive 2 relative to Passive 1 are consistent with the notion that replay following Active resulted in consolidation (710)—possibly consisting of functional network changes (12, 13)—that carried over to subsequent sessions. Third, during Passive 2, scrambled activation of place cells at the theta time scale appears to have interfered with previously formed and consolidated memory traces (during Active and subsequent sleep), resulting in degraded replay during sleep after Passive 2.

How would nested sequences be altered during passive travel? In the absence of active locomotor signals, spike bursts of pyramidal cells recurred at slightly longer time intervals, consistent with the fact that bursting frequency increases with running speed (12, 32). As predicted by theoretical models (13, 14, 33, 34), this decrease in oscillation frequency resulted in impaired theta-phase precession and prevented the formation of theta sequences. Theta sequences have been related to memory performance (35, 36), although the underlying mechanisms have remained unclear. Our results indicate a causal link between theta sequences and sleep replay for memory consolidation, and suggest that behavioral time scale sequences are insufficient to store sequential information for reactivation during subsequent sleep. Nested sequences emerging from independently phase-precessing place cells (13, 34) enabled hippocampal assemblies to fire dozens of milliseconds apart, which is optimal for classical plasticity mechanisms (16) and can reinforce synaptic connections (13, 14). This would effectively store sequential organization as network connectivity patterns, which can later be replayed during sleep for long-term consolidation (710). Spatiotemporal spike patterns supporting nested sequences have also been reported in the striatum (37) and medial prefrontal cortex (38). This may represent a general neural mechanism to encode and store initial memory traces, and plan future actions (35, 39).

Supplementary Materials

Materials and Methods

Supplementary Text

Figs. S1 to S13

References (4045)

References and Notes

  1. See supplementary materials.
Acknowledgments: We thank D. Robbe for comments on an earlier version of the manuscript, Y. Dupraz for technical support, and A. C. Segú for help in animal training. Funding: Supported by the Agence Nationale de la Recherche (ANR-15-CE16-0001-02), French Ministry of Research (C.D.), Collège de France (C.D.), and a joint grant from École des Neurosciences de Paris Île-de-France and LabEx MemoLife (ANR-10-LABX-54 MEMO LIFE, ANR-10-IDEX-0001-02 PSL*) (R.T.). Author contributions: C.D. and M.Z. designed the study; C.D. performed the experiments; C.D., R.T., and M.Z. designed the analyses; C.D. and R.T. performed the analyses; and C.D. and M.Z. wrote the manuscript. Competing interests: The authors declare no competing interests. Data and materials availability: Data of this study are available at

Stay Connected to Science

Navigate This Article