Report

Awake hippocampal reactivations project onto orthogonal neuronal assemblies

See allHide authors and affiliations

Science  16 Sep 2016:
Vol. 353, Issue 6305, pp. 1280-1283
DOI: 10.1126/science.aaf3319

Abstract

The chained activation of neuronal assemblies is thought to support major cognitive processes, including memory. In the hippocampus, this is observed during population bursts often associated with sharp-wave ripples, in the form of an ordered reactivation of neurons. However, the organization and lifetime of these assemblies remain unknown. We used calcium imaging to map patterns of synchronous neuronal activation in the CA1 region of awake mice during runs on a treadmill. The patterns were composed of the recurring activation of anatomically intermingled, but functionally orthogonal, assemblies. These assemblies reactivated discrete temporal segments of neuronal sequences observed during runs and could be stable across consecutive days. A binding of these assemblies into longer chains revealed temporally ordered replay. These modules may represent the default building blocks for encoding or retrieving experience.

The concept of “cell assembly” refers to a group of neurons that are coactivated repeatedly for a given brain operation (1). Cell assemblies thus represent a distinct cognitive entity embedded within neuronal networks (2). However, both their basic structural and functional organization, when outside world influences are minimal, as well as their long-term dynamics, remain unknown owing to the experimental difficulty of circumscribing them. In principle, the chained coordinated activation of such neuronal assemblies combines into sequences of neuronal activation supporting complex cognitive processes (3). Therefore, sequences of neuronal activation can represent a remarkable motif for revealing the activation of underlying neuronal assemblies. In the hippocampus, sequences occur at multiple time scales in the CA1 region—e.g., at the time frame of behavior—or compressed within the period of fast network oscillations (2, 4). They can integrate time and/or distance, as well as any contextual information. Of particular interest are the coordinated patterns of neuronal activation that occur during awake immobility and that are related to sharp wave–associated ripples (SWRs), because these are produced when bodily or environmental control over hippocampal dynamics is minimal. Even though these coherent population events include sequential place cell reactivation representing past or future spatial experience, they are indeed also critically shaped by the internal functional organization of local circuits (57). Sequential neuronal reactivation can be split into separate chunks of current or remote experience (811), but their spatiotemporal organization into different cell assemblies remains unknown. So far, the dissection of hippocampal sequences into discrete reactivation patterns has been achieved by mapping them onto an external spatiotemporal template, such as an experienced behavior (812). It is important to minimize external sensory inputs to reveal the default organization of hippocampal dynamics into cell assemblies because local inputs are known to bias the content of both local and remote replay (10). We recently described a paradigm for revealing internally driven spatiotemporal sequences that occur during run behavior, which is particularly well suited to address this issue (13). However, monitoring large-scale multineuronal activity at high cellular density to identify cell assemblies represents a major technical challenge. This is particularly critical in the case of hippocampal population bursts, as they involve local microcircuits within the densely packed pyramidal layer (7, 14). In vivo imaging of hippocampal dynamics is ideally suited to circumvent this limitation.

We used chronic two-photon calcium imaging of awake head-restrained mice allowed to self-regulate their motion in the dark on a nonmotorized treadmill (13). To map neuronal activity across consecutive days, we used a viral vector (AAV2/1.syn-GCaMP5G, -6m, and -6f; see table S1) (15) that allows for the detection of sparse firing through a glass window on the hippocampus (16), as described in (17). Additionally, mice were chronically implanted with an extracellular field electrode placed in the CA1 stratum pyramidale on the contralateral side to monitor the occurrence of network oscillations in the contralateral local field potential (LFP) during awake immobility periods. Particular attention was given to fast frequency domains (100 to 200 Hz) because most, but not all, CA1 population bursts that occur bilaterally during immobility are associated with SWRs. Because a fraction of awake SWRs is coherently observed in both hippocampi, supported by their anatomical interconnections (18), the contralateral LFP can be used as a reference to identify whether specific calcium transients are associated with the occurrence of SWRs. During each daily imaging session, mice spontaneously alternated between run and immobility epochs (Fig. 1A). Sequences of neuronal activation integrating spatiotemporal information (13) recurred during spontaneous run periods (RUN sequences) (Fig. 1A) (n = 3 out of 4 mice, table S1). During immobility periods, significant peaks of synchronous neuronal activity were observed (Fig. 1A) (P < 0.01, supplementary methods) that appeared as transient “flashes” in the fluorescence movies (movie S1). These synchronous calcium events (SCEs) recruited a sparse population of neurons (Fig. 1B). Cell participation was log-normally distributed (fig. S1).

Fig. 1 Imaging hippocampal dynamics in the CA1 region of awake mice alternating between running and rest.

(A) Recurring sequences of neuronal activation during mouse runs. Mouse speed (top row), heat map raster plot of single cell fluorescence signal (middle), and summed activity of cells (over a 200-ms window). Red dashed line indicates significance threshold for synchronous activity detection (five cells in this example); ripple band power (120 to 220 Hz, bottom row) is displayed over time. (B) Contour map of the neurons imaged in (A) during immobility. Cells shown in red were active during one synchronous calcium event [SCE, gray box from (A)]; all other active cells are shown in gray. Scale bar: 100 μm. (C) Sharp wave–associated ripple (SWR) oscillation co-occurring with the event outlined in (A) (black: raw LFP trace, power spectrogram). (D) Co-occurrence rate of contralateral ripples with SCE and vice versa measured during immobility. (E) Raster plot (left) of cell activation during 21 successive SCEs. Neurons were ordered according to their activation onset during the RUN sequence. Identified replay events (ordered reactivation of neurons) are indicated within a 500-ms window (red: forward replay; blue: backward replay). Right: Fraction of SCEs identified as forward or backward replay across all imaging sessions.

Given that SWRs are known as local synchronous events in the CA1 pyramidal layer (19), we next asked whether these synchronous calcium activities could be associated with SWRs. Electrophysiological recordings revealed that SWRs occurred during immobility (Fig. 1C) (see methods), at a rate of 0.12 Hz (n = 3 mice, fig. S1), which is comparable to previous reports (5, 20, 21). The rate of synchronous calcium events was not significantly different from that of SWRs (0.1 Hz, n = 4 mice, fig. S1, P > 0.1). In 15 out of 21 sessions, we found that only a frequency band of the contralateral LFP between 122 and 204 Hz (1/e width) was significantly time-locked to the SCEs (fig. S1). In the remaining six sessions, no frequency band was prominent. This supports the association of SCEs with ripple events rather than other oscillatory events. Accordingly, using two arbitrary detection thresholds for SWRs and SCEs (see methods), we found that 23% of SWRs were time-locked to the synchronous calcium events imaged in the ipsilateral hemisphere (Fig. 1D, n = 3 mice, table S1, movie S1). Conversely, 20% of the synchronous activities detected with optical approaches during immobility co-occurred with an electrophysiological SWR (Fig. 1D). This rate of co-occurrence can reach values (up to 58%) similar to those obtained with the previously reported fraction of bilateral SWRs (18) and was comparable to the co-occurrence rates of SWRs measured with bilateral electrophysiological recordings in our particular conditions (fig. S10). Overall, the fraction of active neurons participating in a SCE was significantly correlated with the ripple power (P < 0.001, Spearman’s r = 0.16; fig. S1, top row, middle panel); bilaterally coherent SCEs and SWRs recruited a slightly higher fraction of active neurons (fig. S1). We conclude that the SCEs detected in this study are often associated with SWRs.

We next investigated the temporal order of cell activation within SCEs. Indeed, a fraction of CA1 population bursts are expected to reactivate, in forward or reverse order, previously experienced behavioral trajectories (2224). We tested for temporally ordered RUN sequence replay in those mice for which recurring sequences reached a significant probability of detection [n = 3 mice; see (13)]. We computed the average activation time for each neuron within a peak of synchronous calcium activity and correlated this value with its average activation order within the RUN sequences (supplementary methods). In this way, we estimated that 14% of the SCEs were replaying RUN sequences (n = 3 mice, P < 0.05, supplementary methods, Fig. 1E). Of these, 24% were identified as backward replay and 76% as forward replay (Fig. 1E). These results are similar to the proportion of SWRs identified as significant replays in previous reports (8, 10, 2225). In addition, they show that internally generated spatiotemporal sequences can be reactivated in a compressed form (6). Sequential replay could be detected independently from the occurrence of a contralateral electrophysiological ripple. The proportion of sequential replay that co-occurred with a ripple (21.3%) was not different from nonreplay (21.8%), further indicating that SCEs are not different, whether they are detected with a contralateral ripple or not. Overall, these results indicate that SCEs occurring during awake immobility can occur with ripples. We next further analyzed the microstructure and dynamics of SCEs.

One SCE recruited around 5 to 10% of active cells (interquartile range, log-normal distribution, fig. S1) (5, 6, 19). The small minority of neurons contributing to most SCEs (90th percentile) were more likely to remain active across consecutive days (fig. S2), as estimated previously (19). Given that SCEs seemed to be variably synchronizing active neurons, we asked whether single SCEs could be preferentially recruiting discrete neuronal assemblies. We used a clustering algorithm that allowed different types of SCEs to be sorted and single cells to be assigned to putative assemblies (see supplementary methods). In 21 out of 26 sessions (table S1), this analysis revealed several functional cell assemblies that were mostly independent from each other. Indeed, the activity of 88% of cells in a given assembly was significantly correlated with that assembly only (Fig. 2A and supplementary methods). To determine how cell assemblies related to reactivation events, we discarded the mouse without RUN sequences from the subsequent analysis (table S1). SCEs comprised three main categories of events (Fig. 2A): The majority of SCEs (63%) repeatedly recruited a single neuronal assembly (“single-assembly SCE”), 3.4% of them combined two distinct neuronal assemblies (“multiple-assembly SCE”), and 33% did not significantly involve any defined set of neurons (“no-assembly SCE”). This segregation of SCEs into three main categories was also observed when only SCEs co-occurring with electrophysiological ripples were included in the analysis (fig. S3). Half of the “multiple-assembly SCE” (50%) were associated with temporally ordered replay. Therefore, the majority of SCEs can be mathematically best described by an orthogonal basis of cell assemblies (fig. S4). These assemblies are bound together into “multiple-assembly” events, often associated with temporally ordered replay. Cell assemblies displayed notable short-term dynamics as they were less likely to repeat within the subsequent 4 s (Fig. 2E).

Fig. 2 SCEs segregated into functional cell assemblies.

(A) Raster plot of all SCEs, within one representative imaging session, sorted by cell assemblies. SCEs displaying ordered sequential replay are indicated in red. (B) Contour map of the imaged cells; scale bar: 100 μm. Filled colored contours indicate different cell assemblies; filled gray contours indicate other active neurons. (C) Number of cell assemblies recruited during each SCE, pooled across all sessions where RUN sequences could be observed. (D) Top: Representative contour maps depicting two assemblies (filled in red and blue, respectively) from one day to the next. Bottom left: Contour map of the cells that remained in the same assembly across consecutive days days; scale bar: 100 μm. Bottom right: Box plot depicting the probability that a given cell pair is active on two consecutive days for cell pairs that do or do not belong to the same assembly on the first day. (E) Temporal dynamics of cell assemblies. Shortly after the activation of an assembly (t = 0), the probability that it was reactivated within a 4-s time interval (black) was lower than for a different assembly (gray); occurrences are normalized; the expected value for a random activation is 1.

We also examined the topological distribution of cell assemblies and found that these were spatially intermingled (73 out of 85 assemblies; Fig. 2B and supplementary methods). Because the same network could be imaged over consecutive days, we asked whether these assemblies were transient or rather displayed some stability. Assembly membership was analyzed for cell pairs across consecutive days (Fig. 2D). The probability that a given cell pair was part of the same assembly on one day was significantly higher (n = 3 mice, supplementary methods) if the two cells already belonged to the same assembly on the previous day (32%) than if they did not (7%). We conclude that most SCEs either reliably reactivated a fixed preconfigured neuronal assembly (“single-assembly SCE”) or recurrently linked together these assemblies (“multiple-assembly SCE”).

In a final step, we asked how the different cell assemblies were recruited during RUN sequences. In the large majority of the imaging sessions (80%), RUN sequences activated successively two to four neuronal assemblies (Fig. 3), whereas only 50% of the cell assemblies were active during RUN sequences (median value, fig. S5). Thus, half of the “single assembly SCE” reactivated non-overlapping segments of RUN sequences (Fig. 3C, colored cells). The remaining events were not reactivating any part of the RUN sequence (Fig. 3C, white), in agreement with previous reports of “nonreplay” SWRs (9, 10, 24). Multiple-assembly SCEs were twice as likely to recruit multiple RUN sequence segments than expected by chance (49 versus 23%, Fig. 3D), indicating that these events preferentially linked together assemblies reiterating ongoing experience (i.e., RUN sequences). Conversely, the coactivation of a sequence segment with an assembly that was not part of the RUN sequences was less likely than expected by chance (32 versus 55%). In addition, even though cell assemblies were successively recruited within RUN sequences, these assemblies could not be inferred by RUN sequence activity alone (supplementary methods and fig. S9).

Fig. 3 Discrete segments of RUN sequences were reactivated during SCEs.

(A) Raster plot displaying neuronal activation over time for a representative imaging session. Onsets of calcium transients were color-coded according to their cell assembly affiliation (labeled A to D). SCEs significantly recruiting one or two assemblies were labeled with the respective letter. RUN sequences were labeled ABCD. (B) Probability distribution pooled across sessions of the number of RUN sequence segments reactivated within SCEs. (C) Raster plot of all SCEs sorted by cell assemblies. Cells were color-coded according to their activation order within RUN sequences. (D) Probability distribution for “multiple-assembly SCE”: (i) to reactivate RUN sequence segments (multiple segments); (ii) to activate “nonreplay” assemblies (no segment); or (iii) a combination of a RUN sequence segment and a “nonreplay” assembly (single segment). Experimental probabilities were tested against random data.

This study sheds light on the basic organization of awake CA1 synchronous population bursts that include ripples. These events fall into three distinct categories. Most of them recruit neurons sampled among a finite repertoire of preconfigured cell assemblies; these are mutually exclusive and therefore mathematically “orthogonal.” Multiple-assembly SCEs preferentially bind these assemblies into the replay of ongoing internal dynamics on a compressed time scale, as reported for the extended replay phenomenon (26). This finding was enabled by the use of large-scale chronic calcium imaging combined with the mapping of ripples onto default recurring spatiotemporal sequences (13). Unfortunately, the temporal resolution of calcium imaging does not allow an exact determination of the time interval within which coactivation occurs for cells in the same assembly or across different assemblies. Multiple-assembly SCEs may indeed still bind together orthogonal assemblies, spreading across different gamma cycles (11). We hypothesize that orthogonal assemblies could be the calcium counterpart of the recently described attractors within trajectory replay, and that their combination would form the discontinuous representation of space in single awake ripples (11). The awake replay of memory-related information during SWRs has been shown to support encoding, consolidation, and retrieval of event memories and to guide memory-guided decision-making (5). The neuronal assemblies revealed here might indeed constitute separate attractor states, forming the building blocks onto which experience is encoded.

SUPPLEMENTARY MATERIALS

www.sciencemag.org/content/353/6305/1280/suppl/DC1

Materials and Methods

Figs. S1 to S10

Table S1

References (2730)

Movie S1

REFERENCES AND NOTES

  1. Acknowledgments: Work was supported by the INSERM and CNRS (R.C.). This project received funding from the European Research Council under the European Union’s FP7 and Horizon 2020 research and innovation program (grant agreement no. 242842 and 646925), from the DFG (Deutsche Forschungsgemeinschaft Project no. RE 3657/1-1), and the William Harvey International Translational Research Academy Cofund (PCOFUND-GA-2013-608765). We gratefully acknowledge the GENIE Program and the Janelia Farm Research Campus specifically for GCaMP5G, 6m, and 6f: L. L. Looger, J. Akerboom, D. S. Kim, from the GENIE Project, Janelia Farm Research Campus, Howard Hughes Medical Institute. We thank S. Feldt-Muldoon for insight on the clustering methods and K. Diba for sharing data analysis. We are indebted to J. Epsztein, R. Khazipov, P. P. Lenck-Santini, D. Robbe, and the team members for helpful comments on the manuscript. All of the data are archived in the Institute of Mediterranean Neurobiology at Aix-Marseille University. A.M., S.R., V.V., and R.C. designed the research. S.R. and V.V. collected the data. A.M. and C.H. designed the analysis. A.M., C.H., and S.R. analyzed the data. R.C., A.M., S.R., and V.V. wrote the paper.
View Abstract

Navigate This Article