Spatial memories that can last a lifetime are thought to be encoded during ‘online’ periods of exploration and subsequently consolidated into stable cognitive maps through their ‘offline’ reactivation. However, the mechanisms and computational principles by which offline reactivation stabilize long-lasting spatial representations remain poorly understood. Here, we employed simultaneous fast calcium imaging and electrophysiology to track hippocampal place cells over 2 weeks of online spatial reward learning behavior and offline resting. We describe that recruitment to persistent network-level offline reactivation of spatial experiences in mice predicts the future representational stability of place cells days in advance of their online reinstatement. Moreover, while representations of reward-adjacent locations are generally more stable across days, offline-reactivation-related stability is, conversely, most prominent for reward-distal locations. Thus, while occurring on the tens of milliseconds timescale, offline reactivation is uniquely associated with the stability of multiday representations that counterbalance the overall reward-adjacency bias, thereby predicting the stabilization of cognitive maps that comprehensively reflect entire underlying spatial contexts. These findings suggest that post-learning offline-related memory consolidation plays a complimentary and computationally distinct role in learning compared to online encoding.
This is a preview of subscription content
Subscribe to Journal
Get full journal access for 1 year
only $4.92 per issue
All prices are NET prices.
VAT will be added later in the checkout.
Tax calculation will be finalised during checkout.
Rent or Buy article
Get time limited or full article access on ReadCube.
All prices are NET prices.
Source data are provided with this paper. All data supporting the findings of this study are available from the corresponding author upon reasonable request.
Custom Matlab code supporting this study will be available at https://github.com/losonczylab upon publication.
Tolman, E. C. Cognitive maps in rats and men. Psychol. Rev. 55, 189–208 (1948).
O’Keefe, J. & Recce, M. L. Phase relationship between hippocampal place units and the EEG theta rhythm. Hippocampus 3, 317–330 (1993).
Dupret, D., O’Neill, J., Pleydell-Bouverie, B. & Csicsvari, J. The reorganization and reactivation of hippocampal maps predict spatial memory performance. Nat. Neurosci. 13, 995–1002 (2010).
Buzsáki, G., Horváth, Z., Urioste, R., Hetke, J. & Wise, K. High-frequency network oscillation in the hippocampus. Science 256, 1025–1027 (1992).
Lee, A. K. & Wilson, M. A. Memory of sequential experience in the hippocampus during slow wave sleep. Neuron 36, 1183–1194 (2002).
Vaz, A. P., Wittig, J. H., Inati, S. K. & Zaghloul, K. A. Replay of cortical spiking sequences during human memory retrieval. Science 367, 1131–1134 (2020).
Buzsáki, G. Two-stage model of memory trace formation: a role for ‘noisy’ brain states. Neuroscience 31, 551–570 (1989).
Girardeau, G., Benchenane, K., Wiener, S. I., Buzsáki, G. & Zugaro, M. B. Selective suppression of hippocampal ripples impairs spatial memory. Nat. Neurosci. 12, 1222–1223 (2009).
Roux, L., Hu, B., Eichler, R., Stark, E. & Buzsáki, G. Sharp wave ripples during learning stabilize the hippocampal spatial map. Nat. Neurosci. 20, 845–853 (2017).
van de Ven, G. M., Trouche, S., McNamara, C. G., Allen, K. & Dupret, D. Hippocampal offline reactivation consolidates recently formed cell assembly patterns during sharp wave-ripples. Neuron 92, 968–974 (2016).
Zaremba, J. D. et al. Impaired hippocampal place cell dynamics in a mouse model of the 22q11.2 deletion. Nat. Neurosci. 20, 1612–1623 (2017).
Ziv, Y. et al. Long-term dynamics of CA1 hippocampal place codes. Nat. Neurosci. 16, 264–266 (2013).
Malvache, A., Reichinnek, S., Villette, V., Haimerl, C. & Cossart, R. Awake hippocampal reactivations project onto orthogonal neuronal assemblies. Science 353, 1280–1283 (2016).
Grosmark, A. D. & Buzsáki, G. Diversity in neural firing dynamics supports both rigid and learned hippocampal sequences. Science 351, 1440–1443 (2016).
Lopes-dos-Santos, V., Ribeiro, S. & Tort, A. B. L. Detecting cell assemblies in large neuronal populations. J. Neurosci. Methods 220, 149–166 (2013).
Peyrache, A., Khamassi, M., Benchenane, K., Wiener, S. I. & Battaglia, F. P. Replay of rule-learning related neural patterns in the prefrontal cortex during sleep. Nat. Neurosci. 12, 919–926 (2009).
Davidson, T. J., Kloosterman, F. & Wilson, M. A. Hippocampal replay of extended experience. Neuron 63, 497–507 (2009).
Foster, D. J. & Wilson, M. A. Reverse replay of behavioural sequences in hippocampal place cells during the awake state. Nature 440, 680–683 (2006).
Wu, X. & Foster, D. J. Hippocampal replay captures the unique topological structure of a novel environment. J. Neurosci. 34, 6459–6469 (2014).
Diba, K. & Buzsáki, G. Forward and reverse hippocampal place-cell sequences during ripples. Nat. Neurosci. 10, 1241–1242 (2007).
Kim, J. J. & Fanselow, M. S. Modality-specific retrograde amnesia of fear. Science 256, 675–677 (1992).
Scoville, W. B. & Milner, B. Loss of recent memory after bilateral hippocampal lesions. J. Neurol. Neurosurg. Psychiatry 20, 11–21 (1957).
Tse, D. et al. Schemas and memory consolidation. Science 316, 76–82 (2007).
Squire, L. R. & Alvarez, P. Retrograde amnesia and memory consolidation: a neurobiological perspective. Curr. Opin. Neurobiol. 5, 169–177 (1995).
Winocur, G. Anterograde and retrograde amnesia in rats with dorsal hippocampal or dorsomedial thalamic lesions. Behav. Brain Res. 38, 145–154 (1990).
Mankin, E. A. et al. Neuronal code for extended time in the hippocampus. Proc. Natl Acad. Sci. USA 109, 19462–19467 (2012).
Cai, D. J. et al. A shared neural ensemble links distinct contextual memories encoded close in time. Nature 534, 115–118 (2016).
Josselyn, S. A. & Frankland, P. W. Memory allocation: mechanisms and function. Annu. Rev. Neurosci. 41, 389–413 (2018).
Tononi, G. & Cirelli, C. Sleep and synaptic down-selection. Eur. J. Neurosci. 51, 413–421 (2020).
Bittner, K. C., Milstein, A. D., Grienberger, C., Romani, S. & Magee, J. C. Behavioral time scale synaptic plasticity underlies CA1 place fields. Science 357, 1033–1036 (2017).
Grosmark, A. D., Mizuseki, K., Pastalkova, E., Diba, K. & Buzsáki, G. REM sleep reorganizes hippocampal excitability. Neuron 75, 1001–1007 (2012).
Cheng, S. & Frank, L. M. New experiences enhance coordinated neural activity in the hippocampus. Neuron 57, 303–313 (2008).
Morris, R. G. M., Garrud, P., Rawlins, J. N. P. & O’Keefe, J. Place navigation impaired in rats with hippocampal lesions. Nature 297, 681–683 (1982).
Atherton, L. A., Dupret, D. & Mellor, J. R. Memory trace replay: the shaping of memory consolidation by neuromodulation. Trends Neurosci. 38, 560–570 (2015).
Kaufman, A. M., Geiller, T. & Losonczy, A. A role for the locus coeruleus in hippocampal CA1 place cell reorganization during spatial reward learning. Neuron 105, 1018–1026.e4 (2020).
English, D. F. et al. Excitation and inhibition compete to control spiking during hippocampal ripples: intracellular study in behaving mice. J. Neurosci. 34, 16509–16517 (2014).
Zhou, H. et al. Cholinergic modulation of hippocampal calcium activity across the sleep-wake cycle. eLife 8, e39777 (2019).
Dragoi, G. & Tonegawa, S. Distinct preplay of multiple novel spatial experiences in the rat. Proc. Natl Acad. Sci. USA 110, 9100–9105 (2013).
Mizuseki, K. & Buzsáki, G. Preconfigured, skewed distribution of firing rates in the hippocampus and entorhinal cortex. Cell Rep. 4, 1010–1021 (2013).
Peyrache, A., Benchenane, K., Khamassi, M., Wiener, S. I. & Battaglia, F. P. Sequential reinstatement of neocortical activity during slow oscillations depends on cells’ global activity. Front. Syst. Neurosci. 3, 18 (2010).
Kudrimoti, H. S., Barnes, C. A. & McNaughton, B. L. Reactivation of hippocampal cell assemblies: effects of behavioral state, experience, and EEG dynamics. J. Neurosci. 19, 4090–4101 (1999).
Sanhueza, M. & Lisman, J. The CaMKII/NMDAR complex as a molecular memory. Mol. Brain 6, 10 (2013).
Giri, B., Miyawaki, H., Mizuseki, K., Cheng, S. & Diba, K. Hippocampal reactivation extends for several hours following novel experience. J. Neurosci. 39, 866–875 (2019).
Foster, D. J. & Knierim, J. J. Sequence learning and the role of the hippocampus in rodent navigation. Curr. Opin. Neurobiol. 22, 294–300 (2012).
Dragoi, G. & Tonegawa, S. Preplay of future place cell sequences by hippocampal cellular assemblies. Nature 469, 397–401 (2011).
Ólafsdóttir, H. F., Barry, C., Saleem, A. B. & Spiers, H. J. Hippocampal place cells construct reward related sequences through unexplored space. eLife 4, e06063 (2015).
Silva, D., Feng, T. & Foster, D. J. Trajectory events across hippocampal place cells require previous experience. Nat. Neurosci. https://doi.org/10.1038/nn.4151 (2015).
Ambrose, R. E., Pfeiffer, B. E. & Foster, D. J. Reverse replay of hippocampal place cells is uniquely modulated by changing reward. Neuron 91, 1124–1136 (2016).
Carey, A. A., Tanaka, Y. & van der Meer, M. A. A. Reward revaluation biases hippocampal replay content away from the preferred outcome. Nat. Neurosci. 22, 1450–1459 (2019).
Stella, F., Baracskay, P., O’Neill, J. & Csicsvari, J. Hippocampal reactivation of random trajectories resembling brownian diffusion. Neuron 102, 450–461.e7 (2019).
Gupta, A. S., van der Meer, M. A. A., Touretzky, D. S. & Redish, A. D. Hippocampal replay is not a simple function of experience. Neuron 65, 695–705 (2010).
Michon, F., Sun, J.-J., Kim, C. Y., Ciliberti, D. & Kloosterman, F. Post-learning hippocampal replay selectively reinforces spatial memory for highly rewarded locations. Curr. Biol. 29, 1436–1444.e5 (2019).
Chang, H. et al. Coordinated activities of retrosplenial ensembles during resting-state encode spatial landmarks. Philos. Trans. R. Soc. Lond. B Biol. Sci. 375, 20190228 (2020).
Norimoto, H. et al. Hippocampal ripples down-regulate synapses. Science 359, 1524–1527 (2018).
Kovács, K. A. et al. Optogenetically blocking sharp wave ripple events in sleep does not interfere with the formation of stable spatial representation in the CA1 area of the hippocampus. PLoS ONE 11, e0164675 (2016).
Gridchyn, I., Schoenenberger, P., O’Neill, J. & Csicsvari, J. Assembly-specific disruption of hippocampal replay leads to selective memory deficit. Neuron 106, 291–300.e6 (2020).
Buzsáki, G., Leung, L. W. & Vanderwolf, C. H. Cellular bases of hippocampal EEG in the behaving rat. Brain Res. 287, 139–171 (1983).
Ólafsdóttir, H. F., Carpenter, F. & Barry, C. Task demands predict a dynamic switch in the content of awake hippocampal replay. Neuron 96, 925–935.e6 (2017).
Stickgold, R. Parsing the role of sleep in memory processing. Curr. Opin. Neurobiol. 23, 847–853 (2013).
Hull, C. L. The goal-gradient hypothesis and maze learning. Psychol. Rev. 39, 25–43 (1932).
Dragatsis, I. & Zeitlin, S. CaMKIIα-Cre transgene expression and recombination patterns in the mouse brain. Genesis 26, 133–135 (2000).
Kaifosh, P., Lovett-Barron, M., Turi, G. F., Reardon, T. R. & Losonczy, A. Septo-hippocampal GABAergic signaling across multiple modalities in awake mice. Nat. Neurosci. 16, 1182–1184 (2013).
Lovett-Barron, M. et al. Dendritic inhibition in the hippocampus supports fear learning. Science 343, 857–863 (2014).
Danielson, N. B. et al. Sublayer-specific coding dynamics during spatial navigation and learning in hippocampal area CA1. Neuron 91, 652–665 (2016).
Pachitariu, M. et al. Suite2p: beyond 10,000 neurons with standard two-photon microscopy. Preprint at bioRxiv https://doi.org/10.1101/061507 (2016).
Sheintuch, L. et al. Tracking the same neurons across multiple days in Ca2+ imaging data. Cell Rep. 21, 1102–1115 (2017).
Friedrich, J., Zhou, P. & Paninski, L. Fast online deconvolution of calcium imaging data. PLoS Comput. Biol. 13, e1005423 (2017).
Skaggs, W. E., McNaughton, B. L., Wilson, M. A. & Barnes, C. A. Theta phase precession in hippocampal neuronal populations and the compression of temporal sequences. Hippocampus 6, 149–172 (1996).
Sheffield, M. E. J., Adoff, M. D. & Dombeck, D. A. Increased prevalence of calcium transients across the dendritic arbor during place field formation. Neuron 96, 490–504.e5 (2017).
Hyvärinen, A. & Oja, E. A fast fixed-point algorithm for independent component analysis. Neural Comput. 9, 1483–1492 (1997).
van der Meer, M. A. A., Kemere, C. & Diba, K. Progress and issues in second-order analysis of hippocampal replay. Philos. Trans. R. Soc. B Biol. Sci. 375, 20190238 (2020).
We thank G. Buzsáki, J. Magee, D. Levenstein, J. Priestley, K. Kay and S. Tuncdemir for invaluable comments and discussions on the analysis and the manuscript. We thank J. Bowler for the behavioral software used in these experiments. This work was supported by the Brain Initiative Grant NIH U19NS104590 (to A.L.) and by the Charles H. Revson Senior Fellowship in Biomedical Science (to A.D.G.).
The authors declare no competing interests.
Peer review information Nature Neuroscience thanks Kamran Diba and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Extended Data Fig. 1 Cross-day registration and hippocampal activity variance across behavioral states.
(a) Eight examples of same-cell registered ROI’s imaged over 9, 8, 7, and 6 days (top 2 rows, second 2 rows, third 2 rows, fourth 2 rows, respectively), note that not all cells were registered on every imaging day. (b) The cross-day registration and recurrence probability is shown for all day N cells (left 2 pie charts) or for day N place cells (right 2 pie charts) either for the next session the animal was run on the same belt (top row) or on all future sessions the animals were run on the same belt (bottom row). Of the place cells registered on day N 59% were also registered on the next same-belt exposure (top right pie chart, yellow + red), of these 64% were also place cells on the next same-belt exposure, such that in total 38% of the place cells registered on day N were also registered place cells on the following same-belt exposure (red). If all future same-belt exposures are considered 58% of day N place cells were also registered on same-belt future days N + B (bottom right pie chart, yellow + red), of which 57% were found to be place cells on those future days, such that 33% of day N place cells were also place cells on future days (red). (c) Cells were registered using an adapted version of the algorithm introduced in66. Briefly, a beta and a gamma distribution were jointly fit to the Jaccard similarity scores (number of overlapping pixels, divided by the total overlapping and non-overlapping pixels) of cross-day ROI pairs located ≤12 µm away from each other (as measured from their nearest pixels). The panel shows the beta and gamma distribution fits for one animal, showing that the joint distribution approach well matches the bi-modality of observed in the scores. The posterior probability of belonging to either of these distributions was subsequently used to cluster ROIs into cross-day groups putatively belonging the same cell imaged over multiple days. (d) Group data for pair-wise Jaccard similarity scores for pairs of ROI’s belonging to different cell-clusters (blue, mean: 0.09, median: 0.06, n = 3×105 ROI-pairs, horizontal bars show mean ± SEM) or to the same cell-clusters (green, mean: 0.73, median: 0.76, n = 101,222 ROI pairs). Note the clear separation between same-cell clustered and non-clustered Jaccard similarity scores. (e) The histogram shows the distribution of the proportion of imaging days each ROI-cluster was imaged for (horizontal overlay shows the box-and whisker plot, red-line shows median, blue box shows 25th and 75th percentiles, dashed line shows the 90th percentile). While due to the conservative approach to clustering (panels B and C) 41.8% of putative pyramidal cells were observed in only one imaging day (that is were not clustered with other ROI’s), 32.3% of putative pyramidal cells were observed across at least 1/3rd of imaging days, and 22.2% putative pyramidal cells were observed in at least half of the imaging days providing a substantial sample in which to examine cross-day pyramidal cell dynamics. In order to conduct a comparison of calcium neural responses in different behavioral states, preliminary transients were detected using a 5 m.a.d. (median absolute deviation) threshold of 5. (f) Marked differences were observed in the median transients (shown ± boot-strapped 95% confidence interval) between different states (n = 3.05×105 total transients, the behavioral states of transients were determined by the time-point of the transient peak). (g) Notably, within place field (PF) transients during the running were found to have the largest amplitude, followed by immobility transients occurring within LFP-detected SWR-events, with outside of PF transients occurring during running, and non-SWR-related transients occurring during immobility displaying significantly lower amplitudes (n = 3.05×105 total transients, whiskers show 1st and 99th percentiles, box shows 25th and 75th percentiles, solid lines show median, dotted lines show means, Kruskal-Wallis test p ~ 0, followed by post-hock Tukey-Kramer tests). (h) Likewise, within PF running transients showed the longest duration, however, SWR-related transients were of the shortest duration when compared to other states, likely reflecting the short-duration of SWR-related excitability burst as well as the relative sparsity of pyramidal cell bursting during SWR’s36. Due to the observed differences in transient amplitude and duration between states, as well as the decreased potential for motion related contamination during immobility epochs, a lower threshold for activity was used for offline (immobility) as compared to online (running) related epochs (plotted as in G, n = 3.05×105 total transients, Kruskal-Wallis test p ~ 0). (i) Example of activity estimation for a place cell during an online RUN epoch. (I, i) The top panel shows the position of the animal on the belt during the RUN session (green lines), with epochs during which the animal was running within the cell’s place field shaded in blue. The bottom of panel I-i shows the m.a.d.-normalized fluorescence trace of the cell (black) as well as the wavelet denoised trace used for activity estimation. Significant online transients (m.a.d. ≥14) are highlighted in red. Panels in I-ii show a zoomed in portion of the panels shown in (i). The top panel I-ii shows the within-frequency z-scored wavelet power of the simultaneously recorded LFP, showing a prominent theta (5-10 Hz) oscillation during running bouts. The second and third panel in I-ii are a zoomed in portion of the panels shown in I-i. The fourth panel in (ii) shows the deconvolved spiking probability (orange), with supra-threshold spike probability events (m.a.d. ≥1.5) shown in red. The bottom panel in (ii) shows the binarized (top-row) spike rasters, as well as the final sparsified binarized raster which was used as the estimate of cellular activity in the current study, unless otherwise specified. (j) Example of activity estimation for a cell during an offline POST epoch. Panel (J, i) shows the whole POST session fluorescence trace (black), the wavelet denoised trace (blue) as well as the detected offline transients (m.a.d. ≥6) – SWR events are shaded in yellow. Panel (J, ii) shows a zoomed in portion of the POST session. The top panel in in J-i shows the per frequency z-scored wavelet power, showing prominent SWR (125-250 Hz) oscillatory events during immobility. Note that transient activity is enriched within, but is not limited to, LFP-detected SWR events. A threshold of 1.25 m.a.d. was used to detect deconvolved spiking events during offline epochs, and these binarized events were subsequently sparsified as described in the Online Methods.
Extended Data Fig. 2 Within and across-day online state oscillatory, excitatory, and spatial coding properties.
(a) Distribution of place cell preferred normalized theta phases during online running (see Fig. 2b), each line shows an individual animal (n = 10 mice). (b) Per animal p-values of Rayleigh tests for uniformity of a circular distribution on the per cell theta phase preferences (group p-value ~ 0 is the one-sided boot-strapped test of the difference of the mean of the per animal p-values from 0.05, n = 10 mice). (c) Distribution of theta phase modulation amplitudes (resultant circular vector lengths, mean: 0.20, median: 0.17, horizontal bars show mean ± SEM). (d) For each cell the observed theta modulation amplitudes were compared to a null distribution of 1,000 theta modulations amplitudes obtained from a circular shuffling of the included epochs. For each cell the observed modulation values were z-scored relative to their null distributions such that the expected normalized modulation due to chance was zero (observed mean: 0.19, median: 0.07, Signed-Rank test p~0). (e) The mean normalized modulation score per animal confirm that theta modulation of the calcium was consistent throughout our dataset (group p-value ~ 0 is the one-sided boot-strapped difference of the per animal means from the expected null of 0, n = 10 mice). (f, left panel) The distribution of place field (PF) peak coding locations is plotted as the percentage of place cells with PF peaks within a circularly sliding ±20 cm window (percentage ± boot-strapped 95% confidence interval). Note the enrichment of PF’s near the rewarded location (dashed line shows the expected percentages based on a uniform distribution of PF’s along the entire belt). (F, middle panel) This enrichment was confirmed by plotting the percentage of place fields found at varying distances from the reward location (percentage within a 20 cm sliding window ± boot-strapped 95% confidence interval, p-value is the Signed-Rank test of the observed PF distances to reward from the 50 cm distance observed under uniform PF tiling of the belt, n = 13,341 place cells, 10 mice for plots F-K). (F, right panel) For each animal the mean PF distance from reward is plotted, confirming the enrichment of PF’s near the reward (p-value ~ 0, is the one-sided boot-strapped difference of the per animal PF reward distance compared to the 50 cm distance expected under a uniform distribution, n = 10 mice). (g, first panel) The distribution of place cell firing rates during running bouts during the RUN epoch (mean: 0.07 Hz, median: 0.039 Hz, n = 13,341 place cells). (G, second panel) The mean online firing rate of place cells during running epochs is plotted within a circularly sliding ±20 cm window (mean ± SEM). (G, third panel) The mean online firing rate is plotted at different distances to the reward (20 cm sliding window ± SEM, r-value from Pearson’s correlation between online firing rate and PF distance to reward across cells, n = 13,341 place cells). (h) The relationship between PF peak firing rate, position and distance to reward is plotted as in G, mean peak firing rate: 0.64 Hz, median: 0.49 Hz). Both overall and peak firing rates showed significant up-regulation farther from the reward zone. (i) The shuffle-normalized spatial information per estimated spike of place cells is shown as a function of position and distance to reward (plotted as in panel G, mean normalized information per spike: 6.6, median: 5.1, 0 is the value expected by chance). (j) The shuffle normalized spatial information per second of place cells is plotted as a function of position and distance to reward as in panel G (mean: 6.9, median: 5.5). Notably, the spatial information measures shown in panels I and J display the opposite relationship to reward distance as do the firing rate measures shown in G and H. (k) Finally, the normalized within session stability (first half of laps compared to second half of laps) is shown as a function of position and PF distance to reward as in panel G (mean: 1.02 a.u., median: 1.35 a.u.) showing a slight but significant positive relationship with distance to reward. (l) The firing rate by position vector of place cells is shown across RUN belt exposure numbers (number of days/sessions on a given RUN belt). The firing-by position vectors observed on the belt exposure number shown on the top of the panel were sorted according to their place field peak firing position on the belt exposure number shown on the left of the panel. Only place cells with significant place coding on at least two RUN sessions on the same belt were included and the data was pooled across both RUN belts. Cells that were not imaged or did not show significant place coding on a given day are colored in black. As the difference between the reference belt exposure number (top labels of panel) and the comparison belt exposure number (left labels of panel) increases, the place fields become less stable (see Fig. 2e). (m) The same data in panel (L) is shown after applying a ± 10 cm rolling average across reference place field peaks (panel L, subpanel y-ordinate). Applying this smoothing reveals that while the spatial code partially decays across time, a substantial amount of stable place coding persists across days. (n, left panel) The cross-day instability of place cells on the same belt (Fig. 2f) is shown as a function of place field peak distance to reward on the first day of the comparison, shown within a 20 cm sliding window (±boot-strapped 95% confidence, peak position determined by the earlier day of the pair). A significant correlation was observed such that place cells with peaks farther from the reward tended to show decreased cross-day spatial stability (Pearson’s correlation between P.F. peak reward distance and cross-day instability (r): 0.11, p < 3.93×10-17, n = 9,721 cross-day pairs). (N, right panel) This relationship was found to be present across animals, each point shows the correlation coefficient calculated in N, left panel, for each of the 10 mice. (o) We calculated the place field onset lap (see Online Methods) for novel place cells, that is for days in which a given place cell displayed its first place field for a given RUN belt. In order to examine de novo PF formations events on the belt, only novel place cells showing PF formation events at or after the fifth lap of the RUN were included for further analysis (mean PF formation lap: 8.58, median: 5, n = 3,316 novel place cells). (p) Place cell firing rate by position vectors are shown relative to the peak firing position on the first lap (graphs show mean ± SEM). It was found that place cell coding location tended to shift backwards (negative values) on laps subsequent to the formation lap (mean shift: -3.3 cm, median shift: -4 cm, Signed-Ranked test p ~ 0, n = 1,672 de novo place cells). (q) The backward shift from the formation lap to subsequent laps was found to be consistent across animals (one-way boot-strap test of the mean, p < 8×10-5). (r) We further calculated place cell’s firing by position vectors either on the first day that the place cell showed a place field on a given RUN belt or for the mean of subsequent days in which the same place cell showed place fields on the same belt relative to the PF peak firing location on the first PF day. Notably, a similar backward shift was observed (mean shift: -2.52 cm, median shift: -2.62 cm, p < 4.91×10-5, n = 2,916 novel place cells). (s) This cross-day backward shifting effect was found to be consistent across animals (plots show mean shift per animal, n = 10, one-way boot-strap test of the mean, p < 0.00112).
Extended Data Fig. 3 Bayesian population decoding accuracy varies by reward distance, and the magnitude of cross-day decoding error is not accounted for by cross-day changes in PC recruitment.
(a) A heat map showing the median decoding error as a function of the animal’s physical position on the RUN belt within (middle row) or across days (other rows, same-belt cross-day comparisons only, medians taken within a ± 20 cm sliding window across the belt (x-axis)). Note that decoding error tends to be largest when the animal is far from the reward. (b) Median decoder ‘shift’ (the signed difference between the decoded and observed positions) is plotted as a function of the animals’ physical location on the RUN belt as in (A). Note that, consistent with the observed backward shift of place fields over days (Supplementary Fig 2o-s) when past day’s place fields are used as a template to decode the animal’s position on future day’s (last three-rows, positive values of train – test day) the decoder tends to predict positions ahead of the animal’s physical location (red heat-map values). (c) The observed per bin distribution in cross-validated within-session error is shown (performed with 5-fold cross-validation across laps, n = 133,276 third of a second RUN bins, horizontal bars show the median and 25th to 75th percentiles). (d) Within-session reconstruction errors are plotted as a function of observed (physical) position along the RUN belt (plots show medians ± boot-strapped 95% confidence interval across bins in a circularly sliding ±20 cm window). (e) Within-session reconstruction errors are shown as a function of each bin’s observed distance to reward (plots show cross-bin in medians ± boot-strapped 95% confidence interval in a sliding 20 cm window, inset shows the correlation coefficient between bins reward distance and reconstruction error, p ~ 0). (f) The within-mouse correlation between bin reward distance and reconstruction error is shown (p-value from one-sided boot-strap test of the mean correlation coefficient, p ~ 0, n = 10 mice). (g) The distribution of same RUN belt, cross-day Bayesian reconstruction errors is shown as in (C, n = 508,779 cross-day decoded RUN bins). (h) Cross-day reconstruction errors are plotted as a function of observed position (plotted as in D). (i) Cross-day reconstruction errors are plotted as a function of observed distance to reward as in (E, p ~ 0). (j) The within-mouse correlation between distance to reward and reconstruction error is plotted as in F (n = 10, one-sided boot-strap test of the mean, p < 6×10-5) revealing that the overall higher population decoding accuracy observed near the reward locations is consistent across animals. (k) The fact that not all day N place cells are registered place cells on future days, raises the possibility that the degradation in cross-day as compared to within-day spatial coding may be attributable not to place cells changing their spatial selectivity, but to different subsets of spatially stable place cells being active on different days. In order to address this alternative hypothesis, we performed Bayesian cross day decoding either across-days (red) or within-day (blue, for this analysis within-day decoding was 2-fold cross-validated between the first and second half of laps for each session, white circles show the median, black bar shows the 25th to 75th percentiles). As expected though well below chance levels (expected error ~100 cm) cross-day Bayesian position reconstruction error (red, median absolute error across session pairs: 29.4 cm, n = 370 session pairs) was higher than within-day cross-validated Bayesian position reconstruction error (blue, median error: 10.8 cm, n = 101 sessions). However, consistent with the recurrence probability of PCs (Supplementary Fig. 1, panel B, bottom right, red) on average only 50.7 (median: 47) double-conditional day N and day N + B PCs per session pair were available to be used for cross-day decoding while a greater number of single-condition day N PCs (mean: 141.58, median: 135 PCs) were used for the within-day decoding. This is consistent with the possibility that the cross-day degradation is due to a smaller place cell sample size. In order to examine this possibility, the for each day N to day N + B session pair, day N within-day decoding was performed using only the subset of cells which were also registered place cells on day N + B, therefore matching the place cell number to the cross-day decoding case (green, mean number of PCs per session pair: 50.7). Notably, the median decoding error using this restricted within day decoding was 16.1 cm, which, while slightly worse than the error when using all day N place cells (blue, median error: 10.8 cm) was markedly better than the observed cross-day reconstruction error (median: 29.4 cm, Kruskal-Wallis p ~ 0, all between-group differences are significant). These results confirm that the incomplete recurrence of neurons to the place-coding network across days does not adequately explain the observed partial cross-day instability in place coding, and that rather cross-day spatial instability within individual cells, as assessed by our primary stability analysis throughout the manuscript, is a major driver of cross-day coding changes.
Extended Data Fig. 4 Evolution of behavior and hippocampal spatial coding over sequential exposures to the RUN belts.
The mice were exposed to each of the two RUN belts, B and C, for a total of 5 days each. The behavioral task required the animals to lick within a hidden rewarded zone to receive water reward. When the animals licked inside the reward zone (either due to learning or by chance) they received water which consistently elicited a ‘cued’, therefore putatively non-learning-related, licking bout. Therefore, in order to analyze the animal’s learning-related behavior we excluded ‘cued’ licks occurring up to 1 second during or after water delivery (‘non-cued licks’). (a) The percentage of non-cued licks by bin (100 bins total) are shown for each of the 5 days one mouse ran on Belt B, non-cued licks on the reward zone are colored in red. Note that the animal suppresses outside of reward licking as it gains experience over days on the belt. (b, left panel) The percentage of within-reward zone non-cued licks is shown as a function of RUN belt (blue and red for belts B and C, respectively) and belt exposure number (the number of sessions the animal had been run on the belt, solid circles/lines shown by belt means, shaded regions show ± SEM). The animals showed increasingly selective within-reward zone licking as they became familiarized with the belt. (B, right panel) the Spearman’s correlation coefficient between belt exposure number and percentage of non-cued licks in the reward zone is shown per animal per belt, confirming this effect is consistent across animals (one sided boot-strap tests of the mean: belt B: p < 0.007, belt C: p ~ 0). (c) The mean PF distance to reward is plotted per animal per belt as in (B, dashed horizontal line shows expected PF reward distance of uniform spatial tiling). While the reward zone was generally over-represented (sessions below the dashed line), consistent with previous reports11,35 reward-related enrichment occurred more robustly on the second belt the animals were exposed to (belt C) as compared to the first (belt B, per animal per belt correlations for belt B: p < 0.14, belt: C p ~ 0). (d) The overall firing rates of place cells during online running bouts during the RUN epoch as well (e) their peak within PF firing rates increased significantly over repeated belt exposures. However, neither the mean shuffle-normalized spatial information per second (f), or per spike (g) showed a clear correlation with repeated RUN belt exposures. Conversely, the within session stability (first to last half of laps spatial coding stability) increased significantly over repeated belt sessions (per animal per belt correlations for belt B: p < 0.026, belt C: p < 0.01). Therefore, while place cells do not become significantly more spatially informative over time, their coding does become more reliably stabilized to a location. (i) As a consequence, Bayesian decoding of position generally becomes more reliable with experience (plots show median absolute Bayesian reconstruction error across sessions (per animal per belt correlations for belt B: p < 0.04, belt C: p < 2.6×10-4)).
Extended Data Fig. 5 Offline activity and synchrony increase from the PRE to the POST epoch and this increase varies by PF distance to reward.
(a) Peri-event time histograms (PETHs) were made around SWR peaks for each place cell on each imaging day (rows), either for the PRE epoch (left panel), the POST (right panel), or for the entire day’s offline activity combined. Cells are arranged from lowest to highest within-SWR firing rates in each epoch. Note that most place cells show a sharp positive modulation by SWR’s during the offline state (10,194/13,341 place cells (76.4%) fired in at least one SWR event throughout the day, and 9,707 place cells (72.7%) showed greater within SWR than overall immobility firing rates). (b) Mean peri-SWR PETH’s for the POST and the PRE epoch showing a small but significant increase in the within-SWR firing rates from PRE to POST (PRE mean: 0.189 Hz, median: 0.077 Hz, POST mean: 0.204 Hz, median: 0.118 Hz, Signed-Rank p ~ 0, n = 13,341 place cells). (c) Overall place cell offline (immobility) firing rates increased significantly from the PRE to POST epochs (left panel, group PRE mean: 0.043 Hz, median: 0.025 Hz, POST mean: 0.048 Hz, median: 0.032 Hz, Signed-Rank test p ~ 0, n = 13,341 place cells). This effect was consistent across animals (each black line represents one mouse, red lines, vertical red bars show the group mean ± SEM across n = 10 mice). (d) Within SWR firing rates also increased significantly from the PRE to the POST epoch (group PRE mean: 0.189 Hz, median: 0.077 Hz, POST mean: 0.205 Hz, median: 0.118 Hz, Signed-Rank test p ~ 0, n = 13,341 place cells). (e) The median SWR-firing rate gain (within SWR firing rate divided by overall offline firing rate) increased significantly across cells grouped across animals (group PRE mean: 5.96, median: 1.71, POST mean: 5.66, median: 3.42, Signed-Rank test p < 0.0034, n = 13,341 place cells), however this effect varied across mice (right panel, black lines). (f) Furthermore, place cells tended to participate (that is have at least one calcium estimated spike occurring within) a greater percentage of POST as compared to PRE SWR events (group PRE mean: 1.57%, median: 0.67%, POST mean: 1.78%, median: 1.03%, Signed-Rank test p ~ 0, n = 13,341 place cells). (g) Together these PRE to POST activity changes led to a robust increase in per cell population offline synchrony calculated as the mean of the Pearson’s correlation coefficient of each place cell’s offline firing rate vector (smoothed with 125 ms Gaussian kernel) with that of every other simultaneously imaged place cell (group PRE mean: 0.007, median: 0.005, POST mean: 0.012, median: 0.007, Signed-Rank test p ~ 0, n = 13,341 place cells). Similar, though slightly smaller magnitude, effects were observed when all (including non-place cell) ROI’s were considered (data not shown). (h, left panel) Place cell offline (immobility) firing rates are plotted for either the PRE (blue) or POST (red) within a circularly sliding ±20 cm window of their PF peak position during the intervening RUN (left-ordinate, mean ± SEM). Overlaid are the by-position POST-PRE differences (green, right ordinate). (H, middle-panel) PRE, POST (left ordinate) and POST – PRE (right-ordinate) offline firing rates are plotted as a function of PF peak distance to reward in a ± 20 cm window. Inset numerical values show the Pearsons correlation coefficients between the group offline firing rates and RUN PF peak distance to reward revealing that while offline firing rates tend to be larger for near-reward coding cells both during PRE and POST (significant negative correlations, Fisher’s z-test, *p < 0.05, **p < 0.005, ***p < 0.0005, n = 13,341 place cells), increases in PRE to POST offline firing rate are largest for cells coding far from the reward (green inset, small but significant positive correlation). (H, right panel) These patterns were confirmed by a calculating the per mouse Pearson’s correlation coefficients between PRE, POST or POST – PRE offline firing rates to RUN PF peak reward distance (y-axis, p-values from one-sided boot-strap test of the mean, solid circles and vertical lines show mean ± SEM across mice, crosses show across mouse median correlation coefficient values). (i) Within-SWR firing rate gains are plotted as in panel H, revealing that while a small correlation with reward distance was observed for within-SWR firing rate gains during the PRE epoch, both the POST epoch and POST – PRE difference in within-SWR firing rate gains are more strongly correlated with RUN PF distance to reward within and across animals. (j) Consistent with this finding the POST epoch and POST-PRE differences in both the percentage of SWR’s each cell participated in, were positively correlated with RUN PF distance to reward across cells within and across animals. (k) Together these effects led to a larger increase in by cell population synchrony for cells father from the reward than cells close to the reward.
Extended Data Fig. 6 Calcium-based ensemble reactivation and sequence replay are statistically robust.
(a) In addition to the within-SWR reactivation shown in Fig. 2k–l, PRE to POST increases in RUN ensemble reactivation were also observed for overall offline activity, and these effects were confirmed across mice (black lines: per animal means, p-value from one-sided boot-strap test of the mean difference, n = 10 mice; red lines: group data, red horizontal lines shows group means, the vertical extent of the red bars show ± SEM, p-value from Signed-Rank test of per RUN ensemble POST – PRE reactivation differences, n = 3,844 RUN ensembles). (b) In order to account for none-specific changes in excitability and synchrony from the PRE to the POST epoch 1,000 null reactivation scores were computed for each ensemble by randomly permuting the template cell id’s – each cell’s observed reactivation value was normalized (z-scored by) its null distribution such that its expected null value was 0. The robust increase in offline POST as compared to offline PRE shuffle-normalized RUN-ensemble reactivation suggests that the reactivation effect was not attributable to non-specific excitability or synchrony changes. (c) The peri-SWR PETH’s of shuffle normalized RUN-ensemble reactivation (±SEM) were constructed, confirming that RUN-specific content increases robustly around the time of SWR’s, (d) an effect which was confirmed at the per-animal level. (e) In order to examine the duration of RUN-ensemble reactivation day N RUN-ensembles were used to measure either to Day N PRE (~20 minutes earlier) or Day N + 1 PRE (~24 hours after) the RUN. Note that for all multi-day comparisons ensembles were recomputed using only the subset of cells which were place cells on the template day and which were registered on both days of the comparison – consequently the templates applied were identical on both days of the comparison (n = 1,750 ensembles). Notably, overall offline reactivations (in addition to the within-SWR effects shown in Fig. 4m,n) of RUN day N ensembles were larger on PRE Day N + 1 compared to PRE day N, an effect observed both at the group level and across mice. Extending the cross-day analysis to all pairs of Day N, versus day B PRE epochs revealed that the reactivation effect is specific to the Day N + 1 time point both for overall offline activity (f) and for within-SWR activity (g), suggesting that though occurring over much longer time-courses (~1 day) than the previously reported minutes-long decay41 though see14,43) reactivation nonetheless decays over time (inset red text shows number of ensembles (templates) per group). While the combinatorial explosion associated with the all-day-to-all-day comparison shown in panels F and G make calculating the shuffled reactivation scores computational unfeasible, we specifically applied this analysis to the PRE day N to PRE day N + 1 comparisons. (h) The shuffle-normalized PRE day N vs. N + 1 effect was enriched around the time of SWR’s, and (i) was significant at the groups and animal levels when restricting the analysis to within-SWR epochs. (k) In order to assess the specificity of ICA ensemble reactivations to given belts, day N RUN ensembles were used to assess POST – PRE changes in reactivation across days either when the animal was run on the same RUN belt (belt B to belt B or belt C to belt C, comparisons) or the other RUN belt (belt B to belt C or belt C to belt B). It was found that both at the group and animal levels PRE to POST increases in reactivation were larger when the two days being compared were run on the same RUN belt. (l) This effect was largest around the time of SWR’s and (m) was significant at the group and animal level when considering only PRE to POST changes in within-SWR activity. Within-PSE sequence content was estimated from the within PSE-event Bayesian posterior probabilities using a version of the weighted-correlation method19 adapted for circular trajectories. This method was validated by visual examination as well as by comparison to the older ‘line-casting’ (or ‘Radon’ transformation) technique which led to similar results. The principal shuffling method employed in the study is the temporal bin shuffle in which the order of the bins within each event is permuted (resampled without replacement) 2,000 times and the identical sequence replay analysis is conducted of each of these shuffled events. Because differences in event activity and duration can lead large differences in each event’s expected null distribution14 each event’s observed sequence score (r) was normalized (z-scored by) their respective shuffled distribution (rZ score). Given that each type of shuffle makes subtly different assumptions14,71 we also tested out dataset using an alternative shuffling procedure in which the cell ID’s of cells active (firing at least one estimated spike) in each event were shuffled within-event, preserving each event’s mean firing rate structure as well as the number of events each cell participated in, though not each cell’s firing rate or firing-structure in each event. (n, left panel) The cumulative distribution of shuffle normalized sequence scores during offline PSE’s occurring during the PRE, RUN or POST epochs is shown (vertical lines show distribution medians). (N, right panel) box-and-whisker plots of rZ scores across epochs (n = 2978 PRE PSEs, 903 RUN PSEs, 3438 POST PSEs, whiskers show 1st and 99th percentiles, box shows 25th and 75th percentiles, solid lines show median, dotted lines show means; Kruskal-Wallis p < 3.4×10-7). (o) The cumulative distribution of per PSE p-values (at p < 0.05, PRE: 4.7%, RUN: 8.0%, POST: 6.2%; error-bars show boot-strapped 95% confidence interval). (p) Following47 the observed sequence content was further verified by jointly comparing the observed sequence score and maximum jump distance (the maximum absolute circular distance between the reconstructed Posterior probability peaks on the successive bins of the event) to those observed in the shuffle distribution. The panels show the empirical p-value (the proportion of shuffled datasets in which at least the same number or more events met the inclusion criteria as the number of events meeting criteria observed in the experimental dataset) for the sequence score and maximum jump distance thresholds indicated on the y and x axes respectively (heat-map show p-value on a log-scale, red values indicate that the experimental data sets show significantly more events passing criteria than would be expected based on the shuffled datasets, dashed lines show the ‘virtual trajectory’ criteria boundary suggested by Silva et. al. 2015). (q) PRE to POST increases in shuffled normalized sequence scores (rZ) were consistent across mice (plot shows mean ± SEM, one-sided boot-strap test of mean difference p ~ 0). Finally, generating synthetic datasets from Poisson distributions matching each cell’s per epoch within PSE firing rate has been proposed as a surprisingly stringent null hypothesis model to test sequence content47. Though there are some concerns whether this is indeed a valid null hypothesis set (see14) out of an abundance of caution we tested our data against this null. Panels (r, s, t and u) are plotted as in N, O, P and Q (all n values are statistical tests are same ass in N-Q, see Supplementary Table 3 for exact p-values), respectively. These results confirm the accessibility of canonical sequential virtual hippocampal trajectories to fast calcium imaging methods.
Extended Data Fig. 7 PRE to POST change in SWR-specific recruitment reliably predict elevated Path zone cell spatial stability.
In order to test whether the observed relationship between PRE to POST changes in SWR-gain and future place field stability, and its difference between the Goal and Path zone cell groups (Fig. 4a–g), were robust across conditions, parallel analyses were carried out across a variety of methods for measuring cross-day stability, different cell inclusion criteria and grouping metrics. (a) PRE to POST increases in overall offline firing rates of novel place cells was not predictive of their long-term spatial coding stability at either the group (panel i-ii are the same as Fig. 6d, n = 1,212 novel place cells per group, *p < 0.05, **p < 0.005, ***p < 0.0005 for all panels) or by animal (panel iii) levels of analysis. Moreover, overall firing rate changes did not predict future place field stability within either the Goal or Path zones (panels iv-vii, plotted as in Fig. 6 panels G-H). (b) Grouping cells by PRE to POST changes in within-SWR firing rates (as opposed to within-SWR firing rate gain as shown in Fig. 6a–i) revealed a similar pattern of predicted stability as the SWR-gain-based analysis, with high ΔSWR-rate coding more stably onto future sessions on the same belt specifically within the Path zone, and this effect being robust across animals. (c) For each place cell in each epoch offline population synchrony was defined as mean correlation of that cell’s per-frame firing rate vector (smoothed with a 125 ms Gaussian kernel) with all other place cell’s firing rate vectors. Consistent with the role of SWR in driving offline synchrony, grouping cells by their PRE to POST changes in offline population synchrony resulted in a similar Path-zone-specific pattern of predicted long-term stability (panels i, ii, iv and v), however, these effects were less robust than the SWR-based grouping at the by-animal level of analysis (panels ii, vi and vii). (d) Consistent results were obtained when using pair-wise Pearson’s correlation coefficents between RUN firing-rate by position vectors as a measure of stability instead of the PF peak distances. (e) The fact that, on average, more cells are recruited near the reward zone (Extended Data Fig. 2f) could potentially lead to a biased estimate of stability since, all else being equal, place cells are more likely to code near the reward zone twice even if these place fields emerge independently. Therefore here we perform a normalization in which each place cell’s cross-day PF peak distance was z-score normalized by (mean subtracted and divided by the S.D. of) the PF peak distances between ith cell’s PF peak distance with all other place cells imaged on both days N and days N + B. Finally, the sign of the resulting normalized stability score was flipped such that larger values corresponded to more highly preserved spatial coding, resulting in the stability units of ‘z-scored –Δcm’. Therefore this normalization makes described stability relative to what would be expected from randomization of cross-day cell id’s. The same pattern of Path-zone specific predicted long-term stability was observed using this normalization as with the standard stability metric used for the rest of the work. (f) Next, we repeated the analysis performed in Fig. 6a–h but restricted the analyses to only those novel place fields occurring during a novel RUN session (that is the first RUN session that the animal had run on that RUN belt, n = 663 novel place fields per Low and High group), further confirming the Path specificity of stability prediction effect. (g) Finally, in order to control for the possibility that the stability predictive effect is not attributable to the differential SWR-recruitment of weakly coding place cells three additional criteria based on day N RUN spatial coding properties were imposed on first-time place cells for inclusion in the analysis: 1) the PCs were active in at least 25% of laps during RUN, 2) the PC’s observed spatial information per spike content had a p-value of < 0.05 when compared to the cell’s 2,000 spatial shuffles (see Online Methods), 3) the PF peaks during the first half and second half of laps of the day N RUN were less than 20 cm apart. These additional criteria reduced the pool of candidate first time place cells observed on multiple days on the same belt from 2,458 entire belt, 1,079 Goal zone, and 1,074 Path zone PCs to 1,359 entire belt, 607 Goal zone and 579 Path zone cells. While the stability predictive effect of ΔSWR-gain was not significant at group level when assessed across the entire belt (panel i, ii), this effect was significant at the per-mouse level of analysis (panel iii) and was further confirmed to be specific to the Path zone (panels iv–vii).
Extended Data Fig. 8 PRE to POST changes in SWR recruitment robustly predicts the stability of place fields onto future days across behavioral conditions but does not predict their future recurrence or spatial information content.
In a complementary analysis to the ‘Per cell’ analysis carried out in Fig. 6a–h and Supplementary Fig. 7, a more inclusive ‘Per same-cell pair analysis’ was conducted in which each pair of times a given cell was a place cell on the same belt was considered separately (see Online Methods). (A) Same cell pairs were divided by their PRE to POST change in within-SWR gain on the earlier day of the pair further confirming that these changes predict the continued stability of place fields onto future days specifically in the Path zone (graphs plotted as in Fig. 6b–h, n = 3,941 all belt, 1,713 Goal zone and 1,678 Path zone same-cell pairs per group low/high group). (B) Due to the cross-day structure of RUN belt exposures some cross-day same cell place cell pairs occur across time points in which the animal did not experience a different RUN belt (for instance, the first and last days the consecutive day triplet: B-B-B) while other pairs do contain such different belt exposures (for instance the first and last days of the consecutive day triplet: B-C-B). In order to examine this potential confound we divided same place cell cross-day pairs by whether or not a different belt ‘distractor’ sessions occurred within their cross-day time-frame. Same place cell cross-day pairs not containing different belt ‘distractor’ sessions recapitulated the Path zone specificity of the ΔSWR-gain stability prediction effect (n = 1,308 all belt, 554 Goal zone and 578 Path zone same-cell pairs per low/high group, the complementary analysis can be found in Supplementary Fig. 4). (C) In a parallel analysis, we examined same place cell pairs occurring over time intervals during which the animal was not run on the same belt between the time of the two sessions being examined (for instance, the first and last days of the day triplet B-C-B) and found the predictive effect of ΔSWR-gain was preserved both for this subset of pairs (n = 1,682 all belt, 738 Goal zone and 706 Path zone same-cell pairs low/high group). (D) In order to better characterize the duration of the ΔSWR-gain predictive effect of future spatial coding stability and its possible dependence on re-exposure to the same RUN spatial context, we specifically looked at the pair of sessions for each animal occurring 5 days apart during which the animal was not re-exposed to the same spatial context. All plots plotted as in Fig. 6b,c,g,h. While no significant predictive effect was found for the population level analysis (n = 197 same cell pairs per group), this effect was present at the per-animal level of analysis (panel iii). Consistent with the unrestricted results no significant predictive effect was found for Goal zone coding place cells at the group level or at the per mouse level (panel vi, n = 86 same cell pairs per group). Conversely, day N ΔSWR-gain predicted future place coding instability for Path zone coding cells both at the group level (n = 81 same cell pairs per group, p < 0.029), and at the per-mouse level of analysis (panel viii). These results suggest that for intervals of at least five-days, as are shown here, belt re-exposure is not a necessary mediator of the day N SWR-predictive effect on future stability. (E) The main stability measurement throughout this study relates to the similarity of place coding for cells displaying significant place fields on different days on the same RUN belt. In order to examine whether ΔSWR-gain also predicts future place cell recurrence day N PCs were divided by whether or not they were registered on the next same-belt exposure (Supplementary Fig 1, panel B, top right, blue vs. red + yellow comparison). Comparing their (previous) day N per-session z-scored ∆SWR-gain did not reveal significant differences between these groups suggesting that ∆SWR-gain does not predict whether a given place cell will be registered on subsequent days (graph shows the cumulative distribution ±95% boot-strapped confidence interval for each group, dashed lines show group medians, p-value from Ranked-Sum test). (F) In a parallel analysis, only cells registered on day N and next same-belt exposure day N + B were considered and cells were instead divided by whether or not they displayed a significant place field (and therefore were considered place cells, see Methods, Extended Data Fig. 1b, top right, red vs. yellow comparison). No significant in previous within-session normalized ∆SWR-gain were observed between these groups. (G) In order to further examine these effects same place cell pairs were split by their day N ΔSWR-gain changes (same grouping as in panel A) and their future day N + B overall RUN (velocity > 5 cm/s) firing rates were examined. (Panel i) No differences in future run-related firing rates were found between groups either across the entire belt or within the Goal or Path zones (p-values from Ranked-Sum test). (Panel ii) Likewise no significant correlations were observed between future day N + B run-related firing rates and previous (day N) SWR-gain PRE to POST changes either for the entire belt or across conditions (graphs show the mean ± SEM for each of the five quintiles of day N ΔSWR-gain, insets show Pearson’s correlation coefficient values, p-values from Fisher’s Z-tests; a similar lack of effect was observed for PF peak firing rates, data not shown). (H) Further, day N ΔSWR-gain did not predict day N + B place cell shuffle-normalized information per second (n = 8052 All Belt place cells, n = 3551 Goal Zone place cells, n = 3465 Path Zone place cells, for panels H and I see Supplementary Table 2 for exact p-values) or (I) future day N + B within-session stability (assessed as the distance in cm between the PF peaks in the first versus second half of laps in a given RUN session). Therefore, the predictive effect of SWR-gain on the continued stability of place coding of place cells is not readily attributable to potentially confounding, non-specific, co-variates either on day N (see Extended Data Fig. 9 and Supplementary Fig. 10) or on day N + B, but is specific to the preservation of spatial coding from day N onto day N + B.
Extended Data Fig. 9 ∆SWR-gain-based future stability prediction is not controlled for by various RUN and cross-day co-variates.
(a) ‘Per cell’ day N PRE epoch within-SWR gain was not found to be significantly correlated with future place coding instability across either all (black), Goal zone (yellow) or Path zone (red) coding place cells (all plots show the mean ± SEM across the 5 quintiles of the variable plotted on the x-axis, insets show the Pearson’s correlation coefficients, p-values from Fisher’s Z-tests). (b) However, day N POST epoch within-SWR gain was found to be negatively correlated to future place coding instability in both the all place cell and Path zone place cell groups, but not in the Goal zone place cell group. (c) Day N PRE to POST epoch ∆SWR-gain was found to be negatively correlated with future place coding instability in both the all place cell and Path zone place cell groups. (d) In order to better quantify and control for non-SWR predictors of future place field stability a 10-fold cross validated, ridge regularized linear regression was performed to predict future (day N + B) place field instability based on an extensive set of eight day N RUN-epoch excitability and place coding metrics. These metrics were: 1) the place cells’ overall run-related firing rates in hertz, 2) their firing rate at their PF firing peaks in hertz, 3) the percentage of laps in which the cells were active inside of their PF, 4) the temporal c.o.m. of their within PF activity which encodes whether they were active earlier or later within the RUN session (see additional Fig. 2), 5) the width in centimeters of their highest firing rate place field, 6) their normalized spatial information content per second (calculated as a z-score relative to 2,000 cell-specific shuffles), 7) their normalized information per spike, and 8) their day N within-session instability as calculated by the distance in cm between their place field peaks for the first and second half of laps during the RUN session. Plots show the observed normalized cross-day instability of cells as a function of their predicted normalized instability based on the above cross-validated 8-parameter ridge-regression model, either for all cells (black), for Goal zone cells (yellow) or for Path zone cells (red, inset show the correlation between predicted and observed instability within each of these groups). (e) Next a simple 10-fold cross-validated linear regression was performed between the single variable, day N POST – PRE ∆SWR-gain, and future place field instability – the results of this model are plotted as in panel D. Note that the ∆SWR-gain-only model predicts future place field instability specifically for Path zone coding place cells. (f) Finally, in order to account for the effect of RUN day N spatial and excitability properties on the relationship between ∆SWR-gain and future spatial instability, the 8 RUN parameters listed above were used to predict either future spatial instability (as in panel E) or ∆SWR-gain. The predicted values of these models was then subtracted from their corresponding observed values leading to the residual ∆SWR-gain (x-axis) or future place field instability (y-axis) not accounted for by the 8-RUN parameter model (note that this is in effect, the partialized linear relationship between ∆SWR-gain and future place field instability controlling for the co-variance accounted for by the 8-RUN parameter model). Residual ∆SWR-gain was found to significantly negatively correlate with residual cross-day instability for Path zone cells, though not quite significantly negatively correlate across all place cells (p < 0.055), while no negative relationship was observed for Path zone cells (yellow, p < 0.26). This additional control confirms that the predictive effect of day N PRE to POST changes in ∆SWR-recruitment is not accounted for by RUN-epoch excitability or spatial coding co-variates. (g) Having examined potentially confounding day N RUN co-variates as a group, we next examined them individually. (First panel) Overall day N RUN-epoch run-related firing rates did not differ between low and high ∆SWR-gain groups (groups are the same as in Fig. 6a–h, whiskers show 1st and 99th percentile, colored box shows 25th, median and 75th percentile, dashed line shows the mean, p-values from Ranked-sum tests). (Second panel) ∆SWR-gain was found to significantly correlate with RUN firing rates across all place cells, and for Goal zone coding place cells, but not Path zone coding place cells. (Third panel) Day N RUN firing rates were significantly predictive of future place field instability specifically for Goal zone coding place cells. (Fourth panel) In order to account for co-variance with RUN related a simple-linear regression was calculated between RUN firing rates and either future PF instability or same day ∆SWR-gain. The predictions from this regression were subtracted from the observed values resulting in the residual future PF instability and residual ∆SWR-gain plotted in the panel (note that this equivalent to taking the partial correlation between day N ∆SWR-gain and future PF instability controlling for day N RUN firing rates, insets shows partial correlation coefficients and Fisher Z-test p-values). Notably, a significant negative correlation between day N ∆SWR-gain and future PF instability was found specifically for day N Path zone coding place cells. Similar analyses were conducted for day N (h) RUN place field peak firing rates, (i) place field width (in cm), (j) shuffled normalized information per spikes, (k) shuffled normalized information per second, (l) within session place field instability (see Methods) and (m) the percentage of laps place cells had higher within than outside-of place field firing rates. Since cells could theoretically migrate into the field of view during a session and therefore skew our results, a measure of within session temporal center of mass was calculated. Briefly RUN temporal center of mass (c.o.m.) was calculated as the mean of the lap-numbers normalized between 0 and 1 weighted by the cell’s running-bout firing rate within its place field (PF) during each of those laps, such that two hypothetical cells that only fired within their PFs in the first or last laps of a given RUN session would have temporal c.o.m.’s of 0 and 1 respectively. No significant differences were found in RUN temporal c.o.m. between low and high ∆SWR-gain cells either for the entire belts or within the Goal or Path zone coding place cell groups. Moreover, temporal c.o.m. did not control for the ∆SWR-gain to future place field instability correlation for either the all place cell or Path zone place cell groups. (o) Since the ‘per cell’ analysis used here averages across future (day N + B) days the animal was RUN on the same belt, different groups could potentially measure differences across different number of days. However, no significant differences were found between the per-cell mean imaging days apart (the average of variable ‘B’ in ‘day N + B’) within either all place cells or Goal or Path zone coding cells. Furthermore, while this variable strongly correlated with future place field instability (third panel, see also Fig. 2e), this correlation did not control for the predictive effect of ∆SWR-gain on future place coding instability (fourth panel). (p) Finally, since ROI registration errors could potentially lead to spurious cross-day place field instability the previous analysis was carried out on the average Jaccard similarity score (# of overlapping pixels in both ROIs divided by # of pixels in either ROI). Importantly, no ROI similarity differences were observed between low and high ∆SWR-gain groups, and this variable did not control for the ∆SWR-gain predictive effect.
Extended Data Fig. 10 Generalized linear model analysis of factors predictive of future instability.
In order to further test which factors predict future place field instability a 10-fold cross-validated ridge regression-based generalized linear model (GLM) was constructed. The model predicted future per-cell instability for novel place cells based on five unique factors normalized by session: 1) the cell’s spatial information per second on the novel place field day, 2) the mean days apart between the novel place field day and the future days on the same RUN belt with which instability was compared, 3) the cell’s normalized within session instability (Δcm between PF peaks on the first and second half of laps on the novel place field day), 4) the cell’s PF peak distance to reward on the novel place field day, and 5) the PRE to POST change in within-SWR gain observed on the novel place field day. In addition, the interactions between the 5 factors were also used as regressors, for a total of 15 regressors. (a, panel i) The GLM was found to be significantly predictive of future place field instability (plot shows the mean ± SEM of the observed normalized cross-day instability broken up into 5 equal parts (‘quintiles’) by predicted cross-day instability (x-axis), Pearson’s correlation between observed and predicted instability scores (r): 0.27, Fisher’s z-test p < 1.1×10-42, n = 2,424 novel place cells). A subtractive analysis was performed by comparing the full-model GLM to reduced GLM’s in which each of the 15 regressors were held out one at a time (all models 10 fold cross-validated). The table in panel ii shows the results of this subtractive analysis for the 15 regressors (left column), arranged from highest to lowest absolute regressor weight in the full model (second column). Reduced models were tested against the full model via two complementary methods: 1) a p-value was computed from the F-statistic of the difference in the mean squared error between the full and reduced models, 2) the difference in median absolute errors between the full vs reduced models was computed within each fold (fourth column, n = 10 folds, red markers show the cross-fold medians), and the significance of the difference in median errors was tested across folds (last column, p – value from one-sided boot-strap test of the mean, n = 10 folds). As expected, the quality of the place fields on the novel place field day as measured by the normalized spatial information content (first row) or within-session instability (third row), as well as the length of the temporal interval across which instability was assessed (Δdays apart, second row) were found to strongly predict future place field instability. Moreover, distance to reward predicted future instability (PF reward distance, fourth row). That is, place cell’s which were found to be more strongly place coding or nearer the reward on their novel place field day tended to retain their place coding in the future while overall instability increased over time (Fig. 2e). However, after accounting for these expected effects via the subtractive GLM analysis and consistent with our other results (Fig. 6a–h) the interaction between novel place field day PRE to POST change in SWR-gain and P.F. reward distance was found to be significantly predictive of future place field instability (sixth row, dashed red box). (b) The GLM analyses were carried out after restricting the data to Goal-zone cells (plotted as in panel A, n = 1,026 novel place cells). As expected, for this sub-group neither PRE to POST changes in SWR-gain or their interaction with P.F. reward distance were found to be significantly predictive of future place field instability. (c) By contrast, in Path-zone cells (plotted as in panel A, n = 1,018 novel place cells) the PRE to POST change in SWR-gain regressor was found to significantly predict future place field instability (fourth row, dashed red box). These results confirm that after controlling for other factors expected to influence cross-day P.F. instability, and consistent with POST epoch-dependent memory consolidation PRE to POST changes in SWR-gain, and their interaction with P.F. reward distance, are significantly predictive of future place field instability.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
Statistical source data.
About this article
Cite this article
Grosmark, A.D., Sparks, F.T., Davis, M.J. et al. Reactivation predicts the consolidation of unbiased long-term cognitive maps. Nat Neurosci 24, 1574–1585 (2021). https://doi.org/10.1038/s41593-021-00920-7