Excitability mediates allocation of pre-configured ensembles to a hippocampal engram supporting contextual conditioned threat in mice

SUMMARY

Little is understood about how engrams, sparse groups of neurons that store memories, are formed endogenously. Here, we combined calcium imaging, activity tagging and optogenetics to examine the role of neuronal excitability and pre-existing functional connectivity on allocation of mouse CA1 hippocampal neurons to an engram ensemble supporting a contextual threat memory. Engram neurons (high activity during recall or TRAP2-tagged during training) were more active than Non-Engram neurons 3h (but not 24h-5d) before training. Consistent with this, optogenetically inhibiting scFLARE2-tagged neurons active in homecage 3h, but not 24h, before conditioning disrupted memory retrieval, indicating neurons with higher pre-training excitability were allocated to the engram. We also observed stable pre-configured functionally connected sub-ensembles of neurons whose activity cycled over days. Sub-ensembles more active before training were allocated to the engram and their functional connectivity increased at training. Therefore, both neuronal excitability and pre-configured functional connectivity mediate allocation to an engram ensemble.

Graphical Abstract

eTOC Blurb

Neurons are allocated to an engram ensemble during an event. Mocle et al. found engram neurons showed endogenously elevated activity hours before an event, but pre-existing functional connectivity days before an event, suggesting that small sub-ensembles of active neurons are allocated as groups and their functional connectivity modified by learning.

INTRODUCTION

Memory of an experience is thought to be stored by a sparse group of neurons referred to as an engram ensemble, and subsequent re-activation of these “engram neurons” initiates memory retrieval^1–7. Consistent with this, ablating or temporarily silencing engram ensemble neurons specifically disrupts retrieval of that particular memory^8,9. Moreover, artificially activating engram neurons is sufficient to induce behavior consistent with memory recall in the absence of external sensory retrieval cues^10–12. Despite the range of studies showing the critical importance of engram ensembles to memory, understanding how engram ensembles are formed endogenously, particularly how a sparse subset of neurons is allocated to an engram ensemble, remains elusive.

Previous research shows artificially exciting a small random subset of individual neurons in a relevant brain region before an experience biases their allocation to the engram ensemble supporting the memory of that experience^9,13–16. Similar to endogenously-allocated engram neurons, artificially-allocated engram neurons express activity-dependent immediate early genes (IEGs) following a memory test and the intact function of artificially-allocated neurons is necessary for memory retrieval^9,13–15. In contrast, artificially decreasing the excitability of a similar number of random neurons before an experience decreases the likelihood of their allocation to an engram ensemble and the function of these neurons is not necessary for intact memory retrieval^13,17,18. Importantly, the overall size of the engram in any one relevant brain region (e.g., the number of neurons expressing IEGs after memory encoding or retrieval) remains relatively constant despite the use of multiple methods to manipulate neuronal excitability, different types of memory tested and different strengths of memory^19–21. Together, these findings led to the hypothesis that eligible individual neurons compete for allocation to an engram ensemble; the outcome of this competition is based on relative neuronal excitability at the time of an experience, with more excitable neurons “winning” the competition to be allocated to an engram. In these studies, though, the excitability of random individual neurons was artificially manipulated before a training experience. Whether a similar process is engaged endogenously, and whether other processes modulate allocation to an engram supporting a specific memory is unknown. Here we examined the factors mediating endogenous allocation to an engram (without manipulating neuronal excitability).

RESULTS

Experimentally increasing the excitability of a small portion of random dCA1 pyramidal neurons biases their allocation to an engram ensemble supporting a contextual threat memory

Previously, we and others, showed that a sparse subset of principal neurons in the lateral amygdala (LA) experimentally manipulated to be more excitable before a training episode are biased to become key components of an engram ensemble supporting the memory for that episode^13,17,18,22. We examined whether increasing the excitability of a similar sparse subset of CA1 pyramidal neurons in the dorsal hippocampus (dCA1) also biases their allocation to an engram ensemble. We chose to examine contextual threat memory, in which a particular environmental context is paired with an aversive footshock, because this task critically engages the dCA1 region^8,23–25. Memory is tested typically 24h after training, by assessing the time mice spend in a defensive crouched motionless posture (freezing) when replaced in the conditioning context. Freezing is an active (rather than a passive) species-specific behavioral response to threatening stimuli^26,27.

We applied an allocate-and-silence approach using a viral vector that transfects a sparse population of pyramidal neurons and expresses an NpACY construct, with both channelrhodopsin-2(H134R) [ChR2, blue light− (BL) sensitive excitatory opsin] and halorhodopsin [eNpHR3.0, red light− (RL) sensitive inhibitory opsin]^15,28,29. Expression of these two opsins allows the excitability of the same small population of dCA1 pyramidal neurons to be bidirectionally manipulated depending on the wavelength of light applied (increasing with BL and decreasing with RL) (Figure 1A, Figure S1A). Mice in the experimental (Allocated) group received BL photostimulation (30 s, 473 nm, 4 Hz, to increase the excitability of NpACY+ neurons) immediately before contextual threat conditioning (3 footshocks, 1 min inter-shock interval) in an attempt to bias the allocation of these neurons into the engram ensemble. Control mice did not receive BL photostimulation (BL−), such that NpACY+ neurons were not experimentally excited or preferentially biased for allocation. Memory was probed 24h later by returning mice to the conditioned context and assessing freezing behavior under basal conditions (no red light, RL−) and while inhibiting NpACY+ neurons with RL+ photostimulation (660 nm) (Figure 1A). Allocated mice froze at high levels in the RL− condition, but at lower levels when NpACY+ neurons were inhibited with RL (Figure 1B). In contrast, Control mice froze equivalently high in the presence and absence of RL, indicating that silencing a small, random (not experimentally-allocated) population of dCA1 neurons does not impair memory recall. These data verify it is possible to bias the allocation of neurons in the dCA1 to an engram ensemble by artificially increasing their excitability immediately before training.

Figure 1. Optogenetically increasing the excitability of a sparse population of dCA1 pyramidal neurons biases their allocation to an engram supporting contextual threat memory.

(A) Schematic of allocate-and-silence strategy. Sparse subset of dCA1 pyramidal neurons express NpACY construct (red) containing both a blue light (BL)-responsive excitatory opsin (ChR2(H134R)) and a red light (RL)-responsive inhibitory opsin (eNpHR3.0). In Allocated group (Alloc), NpACY+ neurons excited with BL immediately before contextual threat conditioning, biasing NpACY+ neurons to become Engram neurons. In Control group, activity of NpACY+ neurons not experimentally manipulated before conditioning. Memory recall (freezing) assessed with No Light and when NpACY+ cells are inhibited with RL.

(B) Inhibiting NpACY+ neurons disrupted freezing only in Allocated (BL before Training), not Control, mice. Two-way ANOVA, BL before Training × RL during Test, F(3, 24) = 6.74, p = 0.0019, N = 5-7 per group. **: p < 0.01, by Tukey HSD post-hoc. Data are mean ± SEM unless otherwise specified. Dots represent individual mice.

(C) Schematic of scFLARE2 tag-and-silence strategy. Mice expressing scFLARE2 and TRE-eNpHR3.0 such that in the presence of BL active neurons are tagged and express eNpHR3.0. BL delivered either during Training (Training) or in Homecage 10min (10m), 3h or 24h before Training. Control mice did not receive BL (BL−). Memory recall assessed with No Light and when tagged neurons optogenetically inhibited with RL.

(D) Inhibiting highly active neurons tagged by scFLARE2 during Training or in the minutes (10m and 3h), but not 24h, before Training disrupts memory recall. RL does not disrupt freezing in Control (BL−) mice.

(E) Quantification of (D). Difference in percent freezing between No Light and RL Inhibition conditions for each experimental group (24h, 3h, 10m, Training, -BL). Each dot represents change in freezing per mouse. One-way ANOVA, F(4, 15) = 6.44, p = 0.0032, N = 7 per group. *: p < 0.05, one-sample t-test, Holm-Sidak correction for multiple comparisons. Red line indicates no change in freezing with RL.

(F) Schematic of allocate-and-record strategy. Sparse subset of dCA1 pyramidal neurons express both RL-responsive excitatory opsin (ChRmine) and calcium indicator (GCaMP6m). In Allocated mice, GCaMP+ neurons are excited with RL immediately before Training, biasing them to become Engram neurons. In Control mice, GCaMP+ neurons are not experimentally manipulated before training. During memory recall test, freezing behavior measured and GCaMP fluorescence recorded from Allocated or Control neurons.

(G) Example GCaMP traces relative to freezing bout onset (dashed line) during memory test in Allocated (RL+) and Control (RL−) mice.

(H) Time-course of normalized GCaMP6 fluorescence in Allocated (Alloc) and Control neurons relative to mouse freezing bout onset. Number of freezing bouts = 142 and 144 in Allocated and Control mice, respectively.

(I) Fluorescence (arbitrary units, a. u.) higher in Allocated than Control neurons in 2 sec before freezing bout onset, but not during freezing bouts, p < 0.05, Mann-Whitney U test.

Silencing tagged neurons that were endogenously active in the homecage 3h, but not 24h, before contextual threat conditioning impairs subsequent memory retrieval

Next, we used a tag-and-silence strategy to ask whether dCA1 neurons with high endogenous activity before training are similarly biased for allocation to an engram ensemble such that silencing their activity would disrupt subsequent memory retrieval. We tagged neurons with high activity over a brief, temporally precise time window, using the scFLARE2 (Single-chain FLARE, Fast Light- and Activity-Regulated Expression 2)^30–32 tagging system. scFLARE2 tags neurons with high activity (those with increased intracellular calcium) in the presence of a BL inducer that enables expression of the same RL-sensitive inhibitory opsin, eNpHR3.0, as used above. First, we first validated the tag-and-silence approach using scFLARE2 by tagging neurons (applying BL, 20 Hz) during the 5-min contextual threat training period (Training). To probe the necessity of scFLARE2-tagged neurons active during training to subsequent memory recall, we tested mice both in the absence and presence of RL to silence tagged neurons. A Control group did not receive BL during training (BL−) but was tested similarly. RL silencing of neurons active during training disrupted recall of a contextual threat memory whereas RL had no effect in the Control group (Figure 1C–E).

Next, we used scFLARE2 to examine the importance to subsequent memory recall of neurons with endogenously high activity in the homecage at different times before training. Different groups of mice received BL (5-min x 2) to tag neurons with high activity 24h (24h-Pre), 3h (3h-Pre), or 10 min (10m-Pre) before training (Figure 1C–E, Figure S1C–D). Similar to the Training group, silencing scFLARE2-tagged neurons disrupted memory recall in the 3h-Pre and 1m-Pre groups. In contrast, inhibiting cells tagged in the homecage 24h before training did not decrease freezing levels. Therefore, neurons that are endogenously more active in the three hours or less, but not 24 hours, before training are selectively allocated to the engram ensemble supporting the memory of that event.

Artificially allocated Engram neurons show ramping activity in the seconds before memory retrieval

We next used an allocate-and-record strategy to examine the activity dynamics of engram ensemble neurons during a memory test. We used a viral vector expressing the RL-sensitive excitatory opsin ChRmine³³ (to allocate neurons to the engram ensemble) co-expressed with GCaMP6m (to assess their activity during a memory test) (Figure 1F, Figure S1B). We used ChRmine rather than ChR2 to avoid opsin-mediated activation of GCaMP6m. In the experimental Allocated group, we biased the allocation of GCaMP-ChRmine+ neurons with RL+ before contextual threat conditioning, whereas in the Control group, no RL was given (RL−). During the memory test, fiber photometry recorded the bulk GCaMP activity from GCaMP-ChRmine+ neurons that were excited by RL+ before training (Allocated) or not (RL−, a similar number of random neurons, Control)(Figure 1F, Figure S1E). In Allocated, but not Control, neurons we observed a ramping of fluorescence signal in the seconds before freezing bout onset, but not during the freezing bout itself (Figure 1G–I). Importantly, both Allocated and Control mice showed similar freezing levels, indicating differences in neuronal activity before freezing bout onset were not due to differences in threat memory expression (Figure S1F,G). These data are consistent with the hypothesis that contextual threat memory recall is supported by the emergence of neuronal activity in allocated engram ensemble neurons on a seconds-long timescale.

2-Photon imaging shows neurons with increased activity on the day of conditioning, and not in the days before conditioning, are selectivity allocated to an engram ensemble

The results from the above population level experiments indicate that experimentally increasing the excitability of neurons before a training episode biases their allocation to an engram ensemble. We next asked whether endogenous neuronal excitability similarly modulates allocation at the level of single neurons. Two photon (2P) microscopy was used to image dCA1 pyramidal neurons expressing GCaMP7f at different times before training and engram ensemble neurons active during contextual threat training were identified with the TRAP2 system. TRAP2 tags active (via a cFos-based promoter) neurons with a Cre-dependent transgene in the presence of the inducer, 4-Hydroxytamoxifen (4-OHT)³⁴. TRAP2 transgenic mice were microinjected with two viruses; the first expressed CaMKII-GCaMP7f and the second expressed a Cre-dependent mCherry to tag Engram neurons (Figure 2A, Figure S2A). GCaMP+ neurons were imaged on the 2P microscope for 20 min twice a day for 5 days in head-fixed mice. Immediately after the final baseline imaging session, freely-behaving mice were trained in contextual threat conditioning as above and administered 4-OHT to tag active (Engram ensemble) neurons with mCherry. Consistent with previous data, contextually threat conditioned mice showed high freezing specific to the conditioning context (Figure S2B,C).

Figure 2. The activity of stably functionally connected neurons changes over days and those with highest activity on the day of conditioning are allocated to become Engram neurons as assessed by 2P imaging.

(A) Schematic of strategy to image activity of TRAPed (tagged) Engram neurons before contextual threat Training using 2P imaging in head-fixed mice. TRAP2 transgenic mice express GCaMP7f in dCA1 neurons and mCherry in neurons active during Training (via 4-OHT injection). Neuronal activity of GCaMP+ neurons [either Engram (mCherry+) or Non-Engram (mCherry−)] imaged for 5 days before Training.

(B) Examples of tagged (mCherry+) neurons. Top row, raw mCherry signal after 4-OHT treatment (left), background subtracted mCherry signal (middle), and spatial footprints of tagged Engram neurons. Middle and bottom rows, examples of spatial footprints of neurons classified as tagged with background-subtracted mCherry signal shown below.

(C) Fluorescence AUC (area under curve) of Engram (blue) and Non-Engram (gray) neurons across baseline days, normalized to each neuron’s average AUC across all baseline sessions. Linear mixed effects model, Engram × Day, p = 0.015, N = 4 mice. ***: p < 0.001, linear mixed effects model between Engram and Non-Engram neurons, Holm-Sidak correction.

(D) Relative functional correlation strength between Engram and Non-Engram neurons across baseline days. Linear mixed effects model, Engram × Day, p = 0.97, Engram main effect, p < 0.001. N = 56-71 Engram, 752-1003 Non-Engram neurons per session, N = 4 mice. *: p < 0.05, **: p < 0.01, linear mixed effects model between Engram and Non-Engram neurons, Holm-Sidak correction.

The mCherry+ (Engram neurons, roughly 5% of all neurons imaged) and mCherry− (Non-Engram neurons) were traced back to the baseline imaging days (Figure 2B, S2D, E). Using CellReg, we registered cell footprints across imaging sessions (Figure S2F) and found no difference in average fraction of detected neurons (Figure S3A), or fraction of detected Engram neurons (Figure S3B) across baseline imaging days, but, as expected, noted a small decrease in neuronal overlap with the initial imaging session across baseline days (Figure S3C). There was no difference in the overall average transient rate of Engram and Non-Engram neurons across sessions (Figure S3D, E). Although locomotion may impact hippocampal neuronal activity^35,36 Engram and Non-Engram cells showed similar correlation to locomotion across baseline days (Figure S3F).

Pre-configured ensembles of neurons cycle over time; ensembles with relatively high excitability in the hours before an episode are preferentially allocated to an engram ensemble

To examine the relative activity of Engram and Non-Engram neurons across baseline sessions, we integrated GCaMP fluorescence for each neuron over each baseline session (area under the curve, AUC), and subtracted the average AUC for that neuron across all baseline sessions in which the neuron was detected to obtain a normalized AUC. This measures the extent to which a neuron is active relative to its own average activity over the entire six-day baseline period. Engram neurons showed greater relative activity than Non-Engram neurons only on the day of conditioning (Figure 2C), consistent with our findings from the scFLARE2 experiment (Figure 1D) in which we found that neurons endogenously more active in the homecage in the hours before contextual threat training were critical for subsequent memory expression. The activity of the small population of Engram neurons (5%) was not constant across baseline days but showed a cyclic pattern. This cyclic pattern was not evident in the much larger population (95%) of Non-Engram neurons (Figure 2C), which likely correspond to many ensembles cycling out of phase with each other, resulting in stable average activity.

We next examined the functional connectivity among Engram and Non-Engram neurons over baseline days, by calculating pairwise correlations between neurons and z-scoring this value to yield a normalized pairwise correlation value for each neuron pair in each baseline session. We then calculated the relative functional connectivity strength among Engram and Non-Engram populations by averaging the normalized pairwise correlation among Engram and Non-Engram populations in each baseline session. Engram neurons showed higher functional connectivity over days than Non-Engram neurons (Figure 2D), suggesting that even before training, Engram neurons constituted a fairly stable, functional neuronal ensemble, at least at the pairwise level. This finding is consistent with the hypothesis that there are stable functional ensembles of neurons whose activity cycles over days and that the ensemble with increased activity on the day of training is allocated to the Engram ensemble. Similar to the single neuron activity data (Figure 2C), the apparent lack of functional connectivity in the Non-Engram population likely reflects the contribution of many small ensembles that internally have high functional connectivity but do not show high connectivity when averaged across all ensembles.

Patterns of activity among engram ensembles during learning are preferentially reinstated during memory recall

Together, these findings suggest that neurons endogenously allocated to an engram supporting a contextual threat memory show increased activity in the hours, and significantly correlated activity in the days, before conditioning. We next asked whether the activity patterns among Engram neurons changes with learning in freely-behaving mice. Using a custom-built, open-source miniature fluorescence endoscope (CHEndoscope)³⁷ we imaged dCA1 pyramidal neurons expressing GCaMP6f (Thy1-GCaMP6f transgenic mice) before training in the homecage [24h-Pre, 3h-Pre and immediately (Imm-Pre)], during training [5-min in the conditioning context (Pre-shock), then 3 footshocks, as above (Training)] and during test sessions in the conditioning context (Context A, Ctx A) and in a novel context (Context B, Ctx B) (Figure 3A). Each imaging session was 5 min, except for the 8-min Training session. As expected, mice showed low freezing in the Pre-Shock Training period and higher freezing after footshock delivery (Post-Shock Training). Mice also froze at high levels in the memory Test in the conditioning (Ctx A), but not in the novel (Ctx B), context (Figure 3B, see also Figure S4A).

Figure 3. Neurons show above-chance levels of functional connectivity and those with high activity in the homecage 3h, but not 24h, before training are allocated to Engram ensemble as assessed by 1P imaging.

(A) Schematic of strategy to image activity of Engram and Non-Engram neurons in Homecage before and during Training and in Memory Tests using 1P imaging in freely-behaving mice. Top panel. Transgenic mice expressing GCaMP6f in dCA1 hippocampal neurons imaged in Homecage 24h (24h-Pre), 3h (3h-Pre), and Immediately (Imm-Pre) before contextual threat training (5 min context exposure followed by 3 footshocks spaced 1-min apart). Memory assessed in Training context (Ctx A Test) and novel context (Ctx B). Bottom panel. Examples of calcium activity traces.

(B) Mice showed contextual threat memory, freezing at higher levels when replaced in the training context during Test than during Training [both before (Pre-S. Training) or after (Post-S. Training) shock delivery] or when tested in novel Ctx B. One-way repeated measures ANOVA, F(3, 21) = 29.4, p < 0.001. *: p < 0.05, **: p < 0.01, Tukey HSD post-hoc test. N = 8 mice. Data are mean ± SEM unless otherwise specified. Dots represent individual mice.

(C) Classification of neurons detected in the Test as Engram vs. Non-Engram neurons based on average transient rate during Test. Classified as Engram neurons if average transient rate in Test exceeds 50th percentile (z-score > 0), and Non-Engram if not.

(D) Proportion of all imaged neurons Detected and Not detected in Test. Of neurons Detected in Test, 46% categorized as Engram neurons. Neurons with lower activity in Test (z-score < 0) and neurons Not detected in Test categorized as Non-Engram neurons. Engram neurons constitute 9% of total neurons detected throughout experiment.

(E) Average time-course of transient probability in Engram (blue) and Non-Engram (gray) neurons relative to freezing bout onset during Ctx A memory Test (all freezing bouts aligned to Time 0). n = 135 Engram neurons, 149 Non-Engram neurons, N = 6 mice. Error bars represent +/− SEM.

(F) During training, Engram neurons show greater shock information than Non-Engram neurons. Left, example firing fields of shock-coding neurons. Right, average normalized shock information during Training of Engram and Non-Engram neurons. Linear mixed effects model, p = 0.06, n = 135 Engram neurons, 149 Non-Engram neurons, N = 6 mice.

(G) During training, Engram neurons show greater spatial information than Non-Engram neurons. Left, example firing fields of place-coding neurons. Right, Engram neurons show higher average normalized spatial information than Non-Engram neurons. Linear mixed effects model, p = 0.02, n = 135 Engram neurons, 149 Non-Engram neurons, N = 6 mice, data are mean ± SEM.

(H) Engram neurons show higher relative transient rates than Non-Engram neurons in Training, immediately and 3h before Training (Imm-Pre, 3h-Pre), but not 24h before Training (24h-Pre) in homecage. Linear mixed effects model, Engram × Session, p = 0.011, n = 81-149 Engram neurons, 510-983 Non-Engram neurons per session, N = 6 mice. *: p < 0.05, **: p < 0.01, ***: p < 0.001, by pairwise linear mixed effects model with Holm-Sidak post-hoc correction.

(I) Test patterns of pairwise functional connectivity in both Engram and Non-Engram neurons greater than chance (gray line, circular shuffle) across sessions. For all neurons detected in Test, correlation computed between pattern of pairwise functional connectivity for that neuron in Test and session of interest. In all sessions, both Engram and Non-Engram neuronal functional connectivity showed higher than chance similarity to Test session, ***: p < 0.001 different than circular shuffle in a linear mixed effects model, n = 81-149 Engram neurons, 510-983 Non-Engram neurons per session, N = 6 mice. No difference between Engram and Non-Engram neurons for any session.

To identify Engram neurons in this experiment, we chose a widely-used definition of an Engram neuron as one showing high activity during memory recall, and therefore, post-recall expression of activity-dependent IEGs such as cFos or Arc^8,38,39. Accordingly, Engram neurons were defined as those with high activity (relative GCaMP transient rates, z-score > 0) during the Ctx A memory Test (Figure 3C). Of all the GCaMP+ neurons imaged throughout this experiment, 18% were detected in the Ctx A Test, and, of these, 46% were categorized as Engram neurons (Figure 3D). Therefore, roughly 9% of the total imaged population were categorized as Engram neurons, and the remaining as Non-Engram neurons for this contextual threat memory, consistent with our findings above using the TRAP2 system to tag Engram neurons active during training and previous reports^20,40.

In this freely-moving experiment, we also imaged neurons during training and memory recall, allowing general features of endogenously allocated Engram neurons to be examined. Endogenously allocated Engram neurons show increased GCaMP-transient probability (Figure 3E) and transient magnitude (Figure S4C) in the seconds before freezing bout onset in the Ctx A Test than Non-Engram neurons. Even when using a more stringent threshold to categorize Engram neurons (defined as mean transient rate during the Ctx A Test z-score > 1, categorizing 3% of neurons as Engram neurons), Engram neurons showed increased activity in the seconds before freezing bout onset in the Ctx A Test while Non-Engram neurons did not (Figure 3E, Figure 1H, Figure S4D). Conversely, in the Ctx B test, we did not observe pre-freezing activity increases in Engram neurons (Figure S4E), suggesting the activity of Ctx A Engram neurons form a specific memory representation. The time-course of increased activity of Engram neurons before freezing bout onset during the Ctx A Test in this experiment paralleled that observed in experimentally- (optogenetically-) allocated Engram neurons (Figure 1F–I). An increase in the activity of a subpopulation of hippocampal neurons in the seconds before memory retrieval has also been reported in rodents⁴¹ and humans^42,43. Therefore, although there may be differences between experimentally-allocated and endogenously-allocated neurons, these data converge to suggest the conservation of this ramping “neural activity signature” preceding memory retrieval across tasks and species.

Engram neurons identified during the Test showed higher shock (Figure 3F), and spatial (Figure 3G) information than Non-Engram neurons during Training. Considering spatial and shock information exceeding two standard deviations from the shuffled distribution as significant, we found ~15% of Engram neurons and ~8% of Non-Engram neurons are significantly shock-responsive, and ~30% of Engram and ~15% of Non-Engram neurons are significantly spatially modulated. As in the head-fixed 2P experiment above, mouse movement was not differentially correlated with the activity of Engram versus Non-Engram neurons during Training or Test sessions (Figure S4I), indicating the observed differences in activity are not simply due to differing levels of motion correlation.

In Figure 2C, we imaged GCaMP+ neurons in head-fixed mice on the 2P microscope and found that neurons with relatively higher activity on the day of Training went on to become Engram neurons (defined using TRAP2 tagging during training). Here we imaged GCaMP+ neurons in freely-behaving mice at different times in the homecage before Training to ask whether excitability-dependent allocation to an Engram similarly occurs. As above, neurons were registered across all imaging sessions using the CellReg algorithm⁴⁴ (Figure S4F–H). We observed a high correlation between the GCaMP transient rates measured in the 5 min homecage 1P imaging to the first 5 min and last 15 min of our 2P imaging data (Figure S4J–K), lending validity to our 5 min 1P imaging data.

First, we compared the average transient rates of Engram (Test activity z-score > 0) and Non-Engram neurons (Test activity z-score < 0 + neurons not detected in Test) in the homecage, Training and Ctx B imaging sessions. Relative to Non-Engram neurons, Engram neurons showed higher activity during Training, as well as in homecage imaging sessions immediately and 3h before Training (Figure 3H). In contrast, Engram and Non-Engram neurons were equally active in the homecage session 24h before Training, indicating that Engram neurons are not always more active than their Non-Engram counterparts. This result is robust to the specific choice of activity threshold to define an Engram neuron (Figure S4L), or reference session (Training or Test, Figure S4M), and is consistent with findings in Figure 2C in which we tagged Engram neurons active during training using the TRAP2 system.

Results from our previous 2P experiment (Figure 2D) indicated Engram neurons showed above-chance levels of pairwise functional connectivity strength with each other even five days before training. However, in this previous 2P experiment we did not image neuronal activity during training or memory recall, so were unable to assess whether the specific pattern of functional connectivity observed during Test was present during baseline sessions. Therefore, to probe the degree to which functional connectivity patterns observed during Test were also observed in prior sessions in the present 1P experiment, we calculated the similarity between the pattern of pairwise functional correlations for each pair of neurons observed in the memory recall Test session in Ctx A to the pairwise functional correlations observed in all other sessions (24h-Pre, 3h-Pre, Training, Ctx B test). Patterns of pairwise functional correlations among Engram and Non-Engram neurons were preserved at equal and above chance levels (assessed by a circular shuffle) in all other imaging sessions (Figure 3I). These findings are consistent with the interpretation that even before conditioning, there are pre-existing ensembles that change in average excitability but not functional connectivity over days, and those ensembles with high relative excitability in the hours before conditioning are preferentially recruited to the Engram.

Higher-order engram ensemble activity evoked during training is reinstated during memory recall

To further explore how activity patterns among the Engram population are impacted by training, we moved beyond simple pairwise activity correlation and used non-negative matrix factorization (NMF) to decompose neurons observed during Test into sub-ensembles of two or more neurons with correlated activity⁴⁵ (Figure 4A–B, S5A). Similar to our categorization of Engram neurons in Figure 3, we categorized each sub-ensemble as either Engram or Non-Engram by first calculating the average transient rate of all neurons active during Test, weighted by their value in the NMF pattern vector and then normalizing to the mean and standard deviation of all Test sub-ensemble transient rates. Sub-ensembles with high Test activity (z-score > 1) were classified as Engram sub-ensembles, with all others classified as Non-Engram sub-ensembles. The number of member neurons in Engram and Non-Engram sub-ensembles was similar (Figure S5B), but, as expected, Engram sub-ensembles contained a higher proportion of individual Engram neurons (roughly 80%) than Non-Engram sub-ensembles (Figure S5C). Accordingly, Engram sub-ensembles and not Non-Engram sub-ensembles also showed high activity in the seconds before freezing bout onset (Figure S5D).

Figure 4. Selective stabilization of Engram ensembles by training.

(A) Schematic of strategy to examine functional neuronal connectivity in high dimensions. Activity of neuronal population imaged during Test decomposed into sub-ensembles of functionally-connected neurons (sub-ensemble vector) and their corresponding time courses of activation (sub-ensemble activation) using non-negative matrix factorization (NMF). Sub-ensembles classified as either Engram or Non-Engram based on average transient rate during Test. Activity pattern of Engram and Non-Engram sub-ensembles observed during Test compared to other imaging sessions before (24h-Pre, 3h-Pre, Imm-Pre) Training, during Training and novel Ctx B Test.

(B) Test neuronal population activity for one mouse decomposed into sub-ensembles as described above. For each sub-ensemble, activity of each significant member neuron is plotted. Each row separated by dashed lines represents a single sub-ensemble. Time bins with significant sub-ensemble activation are colored.

(C) Engram sub-ensembles identified in Test (blue) showed greater activation during Training, but not in Homecage or Ctx B, than Non-Engram sub-ensembles (gray), suggesting structure of Engram sub-ensembles specifically stabilized during Training. n = 12-22 Engram sub-ensembles, 69-102 Non-Engram sub-ensembles per session, N = 6 mice. **: p < 0.01, by pairwise linear mixed effects model with Holm-Sidak post-hoc correction.

(D) Time-course of Engram and Non-Engram sub-ensemble activation during Training in 1-min time bins. Highest activation of Engram sub-ensembles identified during Test occurred after footshock delivery. n = 22 Engram sub-ensembles, 102 Non-Engram sub-ensembles per session, N = 6 mice. *: p < 0.05, ***: p < 0.001, by pairwise linear mixed effects model with Holm-Sidak post-hoc correction.

(E) Model of sub-ensemble allocation during contextual threat memory acquisition and recall. Average excitability of pre-existing neuronal sub-ensembles changes over days. During training, sub-ensembles that happen to be highly excitable, either by chance or previous memory recall, are allocated to the Engram ensemble supporting that memory. Functional connectivity of neurons in an Engram ensemble observed during Training is selectively stabilized by synaptic plasticity, such that the pattern of activity in Engram ensembles is reinstated by retrieval cues during memory recall Test.

Similar to our analysis with individual neurons in Figure 3, we retrospectively traced the template activation pattern of each sub-ensemble (Engram and Non-Engram) observed in the Test to the other imaging sessions. Engram sub-ensembles showed higher activation rates (higher similarity with the Test activity template) than Non-Engram sub-ensembles in Training (Figure 4C), especially after footshock delivery (Figure 4D), but not in other sessions. Furthermore, unlike pairwise correlation similarity (Figure 3I), sub-ensemble activation rates were not greater than chance levels in the Pre-training homecage imaging sessions (Figure S5E). This effect was not sensitive to the activity threshold used to categorize Engram and Non-Engram sub-ensembles (Figure S5F). These findings indicate that higher-order ensemble activity shows greater context/memory-specificity than simple pairwise metrics of activity patterns. Moreover, the finding that the activity pattern of Engram sub-ensembles evoked by memory retrieval is highly similar to the Training structure suggests the structure of Engram sub-ensembles (not simply pairs of Engram neurons) is selectively stabilized during Training and subsequently reactivated before memory retrieval (Figure 4E).

DISCUSSION

Here we examined neuronal activity dynamics before, during, and after neuronal allocation to an engram ensemble using different techniques. First, we showed that optogenetically exciting a small random population of dCA1 pyramidal neurons immediately before training biases their allocation to an engram supporting a contextual threat memory as disrupting their activity impaired subsequent memory retrieval. Using fiber photometry, we found that optogenetically-allocated engram neurons show ramping activity in the seconds before memory retrieval. Using scFLARE2, an activity tagging system with a precise temporal tagging window, we showed that neurons endogenously active in the homecage in the 1m or 3h, but not 24h, before contextual threat conditioning were critical for subsequent memory retrieval. Using 2P imaging, we found that neurons active during training (and tagged via the TRAP2 system as engram neurons) also showed increased activity when imaged in the hours, but not days, before conditioning. Using 1P imaging in freely-moving mice, we identified ensembles of functionally connected neurons whose structure is fairly stable in the homecage before training. Those ensembles with relatively higher activity at the time of training became allocated to the engram. Learning “imprinted” the higher-order ensemble structure evoked during training in this neural population, and this pattern of activity was reinstated during memory recall.

Together, our results indicate at least two processes modulate endogenous neuronal allocation to an engram, neuronal excitability and pre-configured ensemble structure. In terms of neuronal excitability, we observed the excitability of neurons varies over time as if maintaining a predetermined internal neuronal activity dynamic. Neurons with relatively higher activity in the hours, but not 24 hours or longer, before an experience are preferentially allocated to an engram supporting the memory of that experience. This time course is consistent with memory linking studies in which the engram ensembles supporting two similar events occurring within several hours, but not 24 hours, of each other are linked via co-allocation to overlapping engrams^15,46,47.

The present findings that neurons with relatively higher activity in the hours before an experience are preferentially allocated to a sparse engram are in agreement with previous experiments examining intrinsic neuronal features that mediate the formation of dCA1 place cells (cells that fire in a specific location in an environment)⁴⁸. In any environment, only a sparse population of dCA1 pyramidal neurons develop place fields, with the majority of neurons remaining silent (non-place cells for that environment)⁴⁹. Similar to future engram neurons, future place cells tend to have higher activity (higher burst rates and lower spiking thresholds) than future silent (non-place cells) before exposure to a novel environment⁵⁰. Moreover, small ensembles of neurons with higher firing rates before spatial navigation on multiple linear mazes showed higher probabilities for being recruited as place cells during navigation⁵¹. Similar to our optogenetically-induced allocation finding, several groups showed that experimentally increasing the excitability of a dCA1 neuron was sufficient to induce a place field in an otherwise silent cell^52–55. Therefore, although there may be similarities and differences between engram neurons and place cells⁵⁶, our findings of endogenous and optogenetically-mediated neuronal allocation to an engram fit well with research examining the formation of place cells.

We also found that neuronal excitability-modulated allocation to an engram occurs in a context of pre-existing functional connectivity. Similar to individual neurons, the overall excitability of pre-configured sub-ensembles of functionally connected neurons cycles over time. Sub-ensembles that are relatively more excitable in the hours before a training experience are allocated to the larger engram ensemble and their internal pattern of functional connectivity during learning is stabilized by the training experience. The recapitulation of training activity patterns of Engram sub-ensembles precedes context-specific memory retrieval.

The finding that pre-existing functional connectivity is a factor mediating allocation to an engram ensemble supporting a specific memory is consistent with the general notion that dCA1 pyramidal neurons belong to pre-existing functionally connected sub-ensembles^57–59. Moreover, the present findings agree with reports of preplay in which there is a rapid recruitment of small, pre-configured, internally-generated neuronal sequences to represent spatial trajectories in rodents^51,60–63 and even memory in humans⁶⁴. Although our methods did not permit the examination of the sequence of activity in engram ensembles, our data suggests experience writes and stabilizes new patterns of functional connectivity on top of pre-existing patterns to encode a memory. That is, although Engram ensembles showed pre-existing functional connectivity that was stable over days before training in terms of above chance levels of pairwise functional connectivity, we found that the higher-order activity patterns at the sub-ensemble level observed at Test showed greater context-specificity. These findings suggest that learning may link small, stable sub-ensembles to create higher-order activity patterns that represent the learned associations during a new experience, consistent with previous findings showing patterns in neuronal ensembles observed during training are reinstated during a test and stabilized post-training^63,65.

Learning-induced engram ensemble stabilization may be supported by enhanced synaptic connectivity and neurotransmission with upstream engram neurons, perhaps in CA3 of the hippocampus^20,66 and may occur during offline periods when CA1 engram neurons actively participate in sharp-wave ripple events^67–69. During a subsequent memory test, we observed Engram ensembles undergo reactivation by coordinating and ramping their activity in the seconds preceding memory-guided freezing behavior, similar to reports of enhanced hippocampal coordination in the seconds preceding freezing during remote fear recall in mice⁴¹ and verbal report of memory recall in humans^42,43. The present findings add to the growing literature that pre-existing functionally connected sub-ensembles are a fundamental unit of information in the hippocampus, and further elucidates how the hippocampus transforms patterns of activity during learning into representations that support memory recall.

STAR METHODS

RESOURCE AVAILABILITY

Lead Contact

Further information and requests for resources should be directed to Lead Contact, Sheena A. Josselyn (sheena.josselyn@sickkids.ca).

Materials availability

All unique/stable reagents generated in this study are available from the lead contact upon request with a completed Materials Transfer Agreement.

Data and code availability

• Requests for behavioral and/or calcium imaging data reported in this paper will be shared by the lead author upon reasonable request.

• All code used for analysis of this data has been deposited at Zenodo and is publicly listed and is available as of the date of publication. The DOI is listed in the key resources table.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCES	IDENTIFIERS
Experimental models: Organisms/strains
129SvEv mice crossed with C57BL6n mice	Taconic & Taconic	129SVE & B6
129SvEv mice crossed with GP5.17 mice	Taconic & Jackson Laboratories	129SVE & 025393
Software and algorithms
CNMF-E	Zhou et al.78	N/A
CellReg	Sheintuch et al.44	N/A
Custom analysis software written in Python	Zenodo	https://doi.org/10.5281/zenodo.10573125
Other
CHENdoscope	Jacob et al.37	RRID:SCR_021587

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

All procedures were conducted in accordance with policies of the Hospital for Sick Children Animal Care and Use Committee and conformed to both Canadian Council on Animal Care (CCAC) and National Institutes of Health (NIH) Guidelines on Care and Use of Laboratory Animals. Three mouse lines were used in this study. All mice (male and female) were approximately 8 weeks old at the time of surgery. Wild-type (WT) mice were used in the optogenetics and fiber photometry experiments. These mice were the F1 generation of a cross between C57BL/6N and 129S6/SvEv mice (Taconic Farms). Mice used in the calcium imaging experiment were from the Thy1-GCaMP6f line. We obtained mice hemizygous for the Thy1- GCaMP6f transgene on a BL6J background (GP5.17 line, Jackson Laboratories)⁷⁰ and crossed these with WT 129SvEv mice. The resulting offspring were on a BL6J/129SvEv hybrid background, similar to the WT mice used in the optogenetics and fiber photometry experiments. Only offspring hemizygous for the GCaMP6f transgene were used for the GCaMP imaging experiment. Mice used in the 2p imaging experiment were the F1 generations of C57BL/6N-TRAP2 mice³⁴ (Jackson Laboratories) and 129S6/SvEv mice (Taconic Farms). All mice were bred at the Hospital for Sick Children. Mice were co-housed in groups of 3-5 and had ad libitum access to food and water throughout the duration of the experiments. Rooms were maintained on a 12h light-dark cycle, and all experiments were conducted during the light phase. For the calcium imaging study, 8 animals were included in the dataset, however extracted cells from two animals were not reliably registered across all sessions. For analyses that required registration across all session we included six animals.

METHOD DETAILS

Viruses

Herpes simplex virus (HSV) co-expressing channelrhodopsin2 (ChR2-H134R) fused to enhanced yellow fluorescent protein (eYFP) (ChR2-eYFP) and halorhodopsin 3.0 (eNpHR3.0) (HSV-NpACY)^28,29 was packaged in-house as previously described^15,71. The ChR2-eYFP and eNpHR3.0 genes were connected by a p2A self-cleavage linker and were both expressed under the endogenous HSV promoter IE4/5. Using this construct, the excitability of the same population of neurons expressing NpACY could be increased with blue light (BL) (before training) and decreased with red light (RL) (during testing). Previous whole-cell current-clamp experiments of hippocampal neurons verified the activation spectra for ChR2 and eNpHR3.0 are separable by ~100nm, allowing distinct activation of each opsin²⁹. Moreover, BL (473 nm) increases, while RL (594 nm, 639 nm) decreases, the activity of cells expressing this construct, with minimal cross-talk^15,28. Transgene expression using this HSV systems peaks 3-4 days following surgery^17,18,72. When microinjected into the dorsal hippocampus, this virus preferentially infects pyramidal neurons^73,74.

The AAV plasmid containing ChRmine (the RL-sensitive excitatory opsin) and GCaMP6m was a gift from Dr. Karl Deisseroth³³. This plasmid was packaged in-house into an AAV(DJ) viral vector [AAV(DJ)-αCaMKII-GCaMP6m-p2a-ChRmine-TS-Kv2.1-HA]. We diluted the AAV immediately prior to surgery in sterile phosphate buffered saline (1:10 dilution) concentration to achieve ~10-15% infection of cell layer neurons in CA1 (estimated titer 4.0 x 10⁶ infectious units/mL).

For the scFLARE2 experiment, we packaged the scFLARE2 plasmids (scFLARE2; NRX-TM-CaTEV(uTEV1Δ)-hLOV-TEVcs-tTA-VP16, Addgene plasmid # 158700 AAV-TRE-mCherry-P2A-NpHR3.0-TS, Addgene plasmid # 158703) into AAV(DJ) viral vectors.

Surgical procedures

Anesthesia.

Mice were pretreated with atropine sulfate (0.1mg/kg, i.p.), anesthetized with isofluorane-oxygen mix (3% isofluorane for initial induction and 1-2.5% through nose cone thereafter), and then placed into stereotaxic frames. After surgery, mice were treated with analgesic (Metacam, 4 mg/kg, s.c.) and 1 mL of 0.9% saline (subcutaneously) to prevent dehydration.

Viral microinjection and fiber implantation.

For the optogenetic experiment, HSV-NpACY vector was bilaterally microinjected (1.0 μl/side, 0.1 μl/min) into CA1 (AP −1.8, ML ±1.5, DV −1.5 mm) by pulled glass micropipettes connected to a microsyringe (Hamilton) via polyethylene tubing. The pipette was left in the brain for an additional 10 min following injection to prevent reflux of the virus. Optical fibers were constructed by inserting and securing a 10 mm length of optical fiber (200 μm diameter, 0.37 NA) into a 1.25 mm zirconia ferrule such that the fiber extended ~3 mm beyond the ferrule. These fibers were implanted 0.1 mm dorsal to the CA1 injection site and secured to the skull with dental cement and 2-3 jeweler screws. Mice were allowed to recover for 3 days before the experiment.

For the fiber photometry experiment, 1.0 μl of diluted AAV-ChRmine-GCaMP6m vector was unilaterally microinjected into the right CA1 as described above. A single optical fiber was implanted 0.1 mm dorsal to the CA1 injection site and secured with dental cement and jeweler screws. Mice were allowed to recover for 4 weeks before the start of the photometry experiment. For the scFLARE2 experiment, viruses AAV-scFLARE2 and AAV-TRE-mCherry-P2A-NpHR3.0-TS were co-microinjected into the dCA1 hippocampal region of mice and optogenetic ferrules inserted just above the dCA1.

GRIN lens implantation.

Lenses for calcium imaging were implanted as previously described^37,75. In addition, mice were pretreated with dexamethasone (5 mg/kg, i.p.) to reduce brain inflammation and swelling during surgery. A craniotomy was performed above the right CA1 (AP −2.0, ML +1.5), the dura was carefully removed, and cortical tissue was gently aspirated while continuously applying chilled artificial cerebrospinal fluid. Finally, a 2mm-diameter gradient index (GRIN) lens (ILW-200-P0250-055-NC, GoFoton) was implanted at a depth of −1.5 mm relative to the skull surface and fixed to the skull using dental cement and 3 jeweler screws. For the two-photon experiment, 1 mm relay lenses (130-000143, Grintech) were similarly implanted with a custom baseplate attached to a metal headbar for head-fixation. Mice were allowed to recover for 4-6 weeks before the start of the imaging experiment and were single-housed 1 week before the start of the experiment.

Recording and preprocessing

Fiber photometry.

Fiber photometry recordings were conducted using a custom setup, similar to previous studies⁷⁶. Briefly, 473 nm (signal channel) and 405 nm (isosbestic channel) were modulated sinusoidally at 531 and 211 Hz, respectively. Filtered emission light was collected by a photoreceiver (Doric). The resulting AC signal was passed to a dual channel lockin amplifier (Signal Recovery) which was used to demodulate the signal and isosbestic channels. Both channels were digitized and acquired at a sampling rate of 20 Hz using a U6 DAQ (LabJack). Both channels were bandpass filtered with corner frequencies of 2 and 0.001 Hz. Finally, dF/F was calculated at (signal – isosbestic)/isosbestic. Behavior videos for freezing quantification were acquired at 10 Hz.

2-photon (2P) calcium imaging.

All 2P imaging was conducted on a dual-channel two-photon microscope (Neurolabware). 920nm and 1040nm illumination were supplied by a Coherent Chameleon NX femtosecond laser to excite GCaMP and mCherry, respectively. GCaMP7f recordings were acquired at 20 Hz. On the first imaging session, the PMT gain and laser illumination intensity was set for each mouse and these values were used for all subsequent recording sessions. Images of mCherry expression were acquired by averaging 3000 scans of mCherry fluorescence.

Extraction of fluorescence traces from raw video data was carried out using suite2p⁷⁷ on data from each recording session. The resulting traces were denoised using OASIS as with the CHEndoscope data. We found that the OASIS spike deconvolution on this data resulted in a large number of spuriously detected transients. So to extract significant transients we conducted peak detection on the denoised traces that were filtered with a gaussian kernel with a sigma of 1 second. Peaks were considered significant transients if their height exceeded one standard deviation in the pooled fluorescence values from all cells and timepoints in a given recording session, see examples in Figure S3D. To register cells across sessions, we used the spatial footprints output from suite2p for each session in CellReg, as with the CHEndoscope data.

1-photon imaging.

The design, manufacturing, and applications of the CHEndoscope system used in the current study is described in full detail in Jacob et al., 2018. Here, we used the CHEndoscope system to study neuronal dynamics associated with contextual fear memory in freely behaving mice. Raw images were acquired at 20 frames per second and downsampled to 5 frames per second by frame averaging. Calcium traces and spatial footprints were extracted from the raw video data using CNMF-E using the previously described parameters^37,78, and subsequently manually reviewed for quality using custom Python scripts⁷⁹. Footprints were matched across sessions using the MATLAB package CellReg⁴⁴, which uses a well-validated probabilistic approach to register cells across sessions. For all analyses except for those quantifying mean transient rates, denoised traces from CNMF-E were scaled such that they had unit variance across time. To quantify transient rates, the OASIS spike deconvolution algorithm⁸⁰ was used to recover transient times and to subsequently estimate transient rates.

Behavioral procedures

Contextual threat conditioning.

Unless otherwise specified, training consisted of placing mice in the contextual threat chamber (Ctx A, small box with metal bar flooring, 31×24×21 cm; MED Associates) and then delivering three 0.5 mA, 2 s footshocks spaced one minute apart. 24h after training, mice were replaced in the conditioned context for a 5-min memory test during which the amount of time spent freezing was assessed. 48h after training, mice were placed in a novel context (Ctx B) for 5 min and freezing assessed.

Quantification of behavior.

Our primary measure of conditioned fear memory was the amount of time mice spent freezing in the conditioning context. Freezing is an active defensive response²⁶ and defined as an immobilized, crouched position, with an absence of any movement except respiration⁸¹. The onset and offset of freezing bouts were analyzed manually from the behavior video by two trained scorers unaware of the treatment conditions of the mouse. Freezing bouts lasting less than 1s were excluded and freezing bouts separated by less than 1s were joined. To verify and supplement our hand scoring, we also used DeepLabCut⁸² to track three points (nose, head, and base of tail) along the mouse body. Instantaneous motion was gauged by calculating the difference between body part coordinates at each point in time and an average of the previous five timebins.

Optogenetic experiment.

One day before training, mice were habituated to the optic patch cables for 2 × 5 min. Immediately before training, mice were placed in a clean homecage. Mice in the Allocated (BL⁺) group received BL (473 nm, 30 s, 1 mW, 4 Hz, 15 ms pulse width, to activate ChR2 and excite infected neurons) whereas mice in the Control group received no light stimulation. Contextual fear conditioning took place as above. 24h later, mice were replaced in the conditioned context for a 6-min memory test, the last 3-min of which all mice received continuous red light (RL⁺, 655 nm, 7 mW, to activate eNpHR3.0 and silence infected neurons). The amount of time mice spent freezing was recorded.

scFLARE2 experiment.

To tag neurons active at different times during a contextual threat experiment, we tagged dCA1 neurons with BL illumination (473 nm, 20 Hz, 1 mW). First, we first validated scFLARE2 in this experiment by tagging neurons (applying BL) during the 5-min contextual threat training period (Training). scFLARE2+ neurons active during the BL expressed the red-light (RL) inhibitory opsin NpHR3.0. To probe the necessity of scFLARE2-tagged neurons active during training to subsequent memory recall, we tested mice both in the absence and presence of RL to silence tagged neurons. A Control group did not receive BL during training (BL−) but was tested similarly. Next, we applied BL for 2 x 5-min periods with a 1 min interval between periods in an empty cage 24h, 3h or immediately before mice were threat trained fear as before. 24h later, mice were tested in the conditioning context with optogenetic cables attached. For the first min of memory test, no light was given. During the second minute of the test, RL illumination (660 nm, square pulse, ~7 mW output, 1 min) was used to inhibit NpHR3.0+ tagged neurons. After the memory test, mice were perfused and mCherry expression analyzed in brain slices. mCherry+ neurons in CA1 were quantified and expression was calculated as a percentage of total CA1 neurons as measure by DAPI expression. For each mouse, at least 3 sections were analyzed.

Fiber photometry experiment.

One day before training, mice were habituated to the optic patch cables 2 × 15 min. Immediately before training, mice were placed in a clean cage. Mice in the Allocated (RL⁺) group received RL (655 nm, 30 s continuous stimulation at 10 mW, to activate ChRmine and excite infected neurons) whereas mice in the Control group received no light stimulation. Mice received contextual fear conditioning as above. 24h later, mice were replaced in the conditioned context for a 5-min memory test.

2P experiment.

To ensure that mice were highly habituated to the head-fixed imaging rig, we first habituated each mouse to head fixation for four days, starting at 5 min of head-fixation per day and ending with 20 min of head-fixation. Following these habituation sessions, we habituated the mice to the 2P head fixed rig for 30 min per day for two days. Imaging and habituation sessions in the 2P rig were conducted in darkness on a running wheel. Baseline recording sessions consisted of one 20 min recording followed by a 20 min recording 3 h later. This was repeated for 5 days. On the 5th day, after the second recording, mice underwent contextual threat training as before. Immediately after training, TRAP2 x mCherry mice were injected with 50 mg/kg 4-OHT (i.p.). Three days later, a 20 min recording was conducted, and mCherry fluorescence imaged. This was repeated for two additional days. After the first posttagging imaging session, mice were tested in the conditioned context as before. After the second post-imaging session, mice were tested in a novel context as before. One month after tagging, the mCherry signal was imaged again, where in some cases improved mCherry expression was observed. The mCherry images were used to classify neurons as tagged or not tagged.

1P experiment.

Mice were habituated to the CHEndoscope for 15 min/d for 3 d before training. 24h, 3h and immediately before training, mice were imaged in the homecage for 5 min. Training consisted of placing mice in the contextual threat chamber and, 5 min later, delivering three 0.5 mA, 2 s footshocks spaced one minute apart. 24h after training, mice were replaced in the conditioned context for a 5-min memory test during which the amount of time spent freezing was assessed. 48h after training, mice were placed in a novel context (Ctx B) for 5 min and freezing assessed.

QUANTIFICATION AND STATISTICAL ANALYSIS

Fiber photometry data analyses

After pre-processing the photometry calcium traces, we smoothed the dF/F values with a Gaussian kernel with a sigma of 1s. We then plotted the fluorescence intensity as a difference from baseline (mean fluorescence from 7 to 5 sec before freezing onset) across all freezing bouts. With respect to any particular freezing bout, any data corresponding to a previous freezing bout or a subsequent freezing bout was excluded. To quantify fluorescence across freezing bout onset, within each freezing bout we averaged the dF/F values in the 2 s before freezing bout onset and 4 s after freezing onset.

2P experiment

Identification of tagged neurons.

To correct for the small chromatic aberration introduced by the GRIN lens, we first registered the mCherry image to an average projection GCaMP image from the same session using basin hopping to maximize the correlation between the two images by rotations and translations. Once the mCherry image was registered to GCaMP images, we corrected for local differences in illumination and viral expression by local background subtraction. To estimate local background, the mCherry image was blurred by convolution with a gaussian kernel with a sigma value of 10 pixels. The normalized red intensity was calculated as (red − bg)/bg, where bg is the estimated local background. We then calculated both the pixel correlation between each cell footprint (from suite2p) and the mCherry image, as well as the average normalized red signal within each cell footprint (Fig. 2B). Cells that showed a correlation value greater than 0.4 and a normalized intensity value greater than 0.3 were considered Engram neurons. To quantify the quality of tagged cell detection, we calculated the fraction of cells considered tagged using this procedure under small random changes to the mCherry channel position (between ± 20 pixels translation, −5 to 5 degrees rotation). In all four mice, the actual number of detected tagged cells far exceeded the chance-level distribution.

Quantification of activity across baseline sessions.

Inhomogeneities in viral expression may simultaneously influence the detection of both mCherry fluorescence and GCaMP7f transients, and therefore may introduce an artificial association between transient rate and Engram identity. Furthermore, it is possible that some neurons are directly activated by the head-fixed imaging context, and may show relatively higher activity than other cells across all baseline sessions. To control for these effects, we integrated fluorescence for each neuron in each baseline session (area under the curve, AUC), and subtracted the average AUC for that neuron across all baseline sessions in which the neuron was detected to obtain a normalized AUC. The normalized AUC was plotted in Figure 2C.

Quantification of pairwise functional connectivity strength across baseline sessions.

To quantify the strength of correlations among Engram and Non-Engram populations we first calculated the Pearson correlation between all pairs of neurons. To control for various effects such as total activity or transient shape, we then circularly shuffled the traces 1000 times and normalized each pairwise correlation to the mean and standard deviation of its corresponding shuffled distribution. Within each session we then z-scored all normalized pairwise correlation values to enable comparison across mice and sessions. The functional connectivity strength for an Engram neuron is then simply the average z-scored normalized pairwise correlation between the given neuron and all other Engram neurons active in the particular session. Similarly, for Non-Engram neurons we averaged the z-scored normalized pairwise correlations between all active Non-Engram neurons.

1P experiment

Identification of putative engram and non-engram neurons.

Neurons were classified as Engram or Non-Engram based on their relative average transient rate in the Test session (Cxt A, conditioning context). The average transient rate of all neurons active in the Test session was calculated and then normalized to the mean and standard deviation of the distribution of transient rates (z-score). Neurons were classified as Engram if their normalized transient rate exceeded 0, but the sensitivity of certain analyses to this parameter was examined in Figure S4D, K, and L.

Identification of place and shock cells.

We identified neurons that responded to shock and spatial cells using a mutual information metric⁸³ between the calcium trace of a particular cell with a given stimulus:

$$I=\sum _i{p}_i{\lambda }_i{log}_2\frac{{\lambda }_i}{\lambda }$$

For spatially tuned neurons, $p_i$ is the probability of the mouse being in the i’th spatial bin, ${\lambda }_i$ is the mean fluorescence in the i’th spatial bin, and $\lambda$ is the total mean fluorescence of the cell (training arena was binned in a 13 × 13 grid). Similarly, for shock-tuned neurons, $p_i$ is the fraction of the total training time in a 10 s window following shock delivery, and ${\lambda }_i$ is the mean fluorescence either outside or inside the 10 s post-shock window (i=0 or 1, respectively). To compensate for sampling bias in low transient rate neurons, we circularly shuffled calcium traces 1000 times relative to position or shock delivery times and computed the mutual information metric. The information score was calculated as the actual calculated mutual information that was z-scored using the mean and standard deviation from the shuffled distribution.

Pairwise functional connectivity.

To quantify functional connectivity between pairs of neurons, we calculated a Pearson’s correlation coefficient using the neurons’ corresponding calcium traces. To account for differences in calcium activity rates between neurons that can potentially inflate these correlations, we circularly shuffled each neuron’s calcium traces to compute the chance-level correlation 1000 times and used this to normalize the correlation coefficient of each neuron pair. Neurons were considered significantly functionally connected if their actual correlation exceeded two standard deviations from the mean of the shuffled distribution.

Sub-ensemble decomposition using NMF.

We used non-negative matrix factorization (NMF) to decompose population activity in the test session into ensembles and their corresponding time courses of activation, similar to a previous approach⁴⁵ with some modifications. We used the scikit-learn implementation of NMF⁸⁴, initialized with singular-value decomposition and optimized with using multiplicative updates rules⁸⁵ and with a L1 constraint to encourage sparse ensemble vectors. To choose the number of components in the model we used a permutation-based approach. For each component number, we decomposed the activity matrix, and calculated the reconstruction error:

$$\text{ε}={\left|\left|A-WH\right|\right|}^2$$

Where $A$ is the original activity matrix with shape time × cell number, but downsampled to 1s time bins. $W$ is the time × ensemble number timecourse of ensemble activity, and $H$ is the ensemble number × cell number ensemble matrix, containing the information about ensemble membership for each neuron.

To choose the number of ensembles into which we decompose the population activity in an unbiased manner, we used a permutation-based approach. We first calculated the decay in reconstruction error when decomposing the actual activity matrix with increasing numbers of ensembles. Adding ensembles will always decrease the reconstruction error, however we only wish to add ensembles if it decreases the reconstruction error to a larger degree than chance levels. To measure this chance level of reconstruction error reduction in each session, we decompose a circularly-shuffled activity matrix with increasing number of ensembles. We average 10 such shuffles to get a chance-level reconstruction error decay curve. For the actual and shuffled curve, we calculate the rate of change in reconstruction error for each added ensemble. The final ensemble number for a particular session is the ensemble number where the actual rate of change is equal to the shuffled rate of change. For the analyses involving classifying whether neurons were members of an ensemble, we thresholded the ensemble vector at a value of 0.5.

To quantify the degree to which to a particular sub-ensemble is active within a single time-bin, we calculated the ensemble similarity as the dot product similarity:

$${\text{s}}_{\text{i,t}}=a_t\cdot e_i$$

Where $e_i$ is the i’th ensemble vector (the i’th row of $H$) and $a_t$ is the population vector at time bin t in the session of interest. To compensate for inflated similarities during time bins with high activity, we normalized the similarity in each time bin to the mean and standard deviation of 1000 cell ID shuffles of the ensemble vector. We refer to this as the sub-ensemble activation score. An ensemble was considered significantly active at a particular timebin if its activation score exceeded 2 standard deviations.

To estimate the absolute level of activity in each ensemble during the Test session, we calculated the ensemble activity as an average of the transient rates of the individual neurons in the ensemble, weighted by their corresponding value in the ensemble vector.

Identification of Engram and Non-Engram sub-ensembles.

We categorized NMF-identified neuronal sub-ensembles as Engram or Non-engram sub-ensembles based on their relative average transient rate in the Test session. We calculated the sub-ensemble average transient rate for a given ensemble by taking an average of all transient rates for each neuron, weighted by each neuron’s value in the ensemble vector. The average transient rate of all sub-ensembles in the Test session was calculated and then normalized to the mean and standard deviation of the distribution of ensemble transient rates (z-score). Sub-ensembles were classified as Engram if their normalized transient rate exceeded 1. This threshold resulted in a similar proportion (41% of neurons active in Test, 7% of all recorded neurons) of neurons that are members of Engram sub-ensembles as Engram neurons (using the 0 std. dev. threshold).

Statistics

All statistical tests were conducted using custom scripts in Python. Statistical significance was assessed using paired and unpaired t-tests where normality assumptions held, and Mann-Whitney U tests if normality assumptions were violated using a Shapiro-Wilk test. Multiple comparison tests were carried out using one-way and two-way ANOVAs with repeated measures, where appropriate. All multiple comparison tests were done using the statsmodels Python library, and post-hoc comparisons were carried out using the Tukey HSD procedure.

To account for inflated type I error rates in the nested designs⁸⁶ (neurons within animals) of the calcium imaging experiment, we fit hierarchical linear mixed models (random intercepts and fixed slope models) to assess the taking into account the non-independence of neurons recorded in different animals where appropriate. Post-hoc comparisons were conducted by calculating the p-value of a pairwise linear mixed effects models, adjusted using a Holm-Sidak correction. Sample sizes were chosen to match previously published similar work. Details regarding the statistical test used, the value and definition of n, and dispersion measures can be found in the relevant figure captions.

Highlights.

• Active neurons in hours before experience preferentially allocated to engram ensemble

• Active neurons in hours before experience necessary for subsequent memory retrieval

• Functionally connected neuronal ensembles detected days before conditioning event

• Learning modifies functional connectivity between neurons in engram ensembles

Inclusion and diversity

We support inclusive, diverse, and equitable conduct of research.