A weakened recurrent circuit in the hippocampus of Rett syndrome mice disrupts long-term memory representations

Summary

Successful recall of a contextual memory requires reactivating ensembles of hippocampal cells that were allocated during memory formation. Altering the ratio of excitation-to-inhibition (E/I) during memory retrieval can bias cell participation in an ensemble and hinder memory recall. In the case of Rett syndrome (RTT), a neurological disorder with severe learning and memory deficits, the E/I balance is altered but the source of this imbalance is unknown. Using in vivo imaging during an associative memory task, we show that during long-term memory retrieval, RTT CA1 cells poorly distinguish mnemonic context and form larger ensembles than wild-type mice. Simultaneous multiple whole-cell recordings revealed that mutant somatostatin expressing interneurons (SOM) are poorly recruited by CA1 pyramidal cells and are less active during long-term memory retrieval in vivo. Chemogenetic manipulation revealed that reduced SOM activity underlies poor long-term memory recall. Our findings reveal a disrupted recurrent CA1 circuit contributing to RTT memory impairment.

Graphical Abstract

Introduction

Rett syndrome is a neurodevelopmental disorder primarily caused by de novo mutations in the X-linked methyl-CpG-binding protein 2 gene (MECP2) (Amir et al., 1999). In females, X-chromosome inactivation during development leads to a mosaic expression of the dysfunctional gene in 50% of neurons. Girls with the syndrome appear to develop normally for the first 12–18 months but then regress, losing most of their acquired language, cognitive, social, and motor skills (Hagberg et al., 1983). This regression is often accompanied by ataxia, respiratory problems, seizures, and other neurological signs (Ip et al., 2018). Heterozygous female mice (Mecp2^+/−, hereafter referred to as RTT mice) (Guy et al., 2007) develop a similar set of symptoms and exhibit profound learning and memory impairments (Samaco et al., 2013). Studies of neuronal activity in mouse models that under-express MeCP2 suggest that these learning and memory impairments result from disruptions in the balance of excitation to inhibition (E/I) (Calfa et al., 2011; Kee et al., 2018; Lu et al., 2016; Zhang et al., 2008). Indeed, RTT mice have neurological defects that could disrupt the E/I balance, including reduced synaptic efficacy (Chao et al., 2007), less elaborate dendritic morphologies (Belichenko et al., 2009), and reduced long-term potentiation (Guy et al., 2007; Na et al., 2013). While conceptually helpful, the premise of a perturbed E/I balance does not identify specific cellular elements that can be causally linked with learning and memory deficits in Rett syndrome. To form this link, it is essential to both record and manipulate the activities of specific hippocampal cell types during a learning and memory task.

The hippocampus is necessary to process contextual memories, memories that encode an event and the circumstances in which the event was experienced (Kim and Fanselow, 1992). Two excitatory pathways in CA1 are thought to independently support the encoding and decoding of contextual memories. The temporoammonic pathway carries on-going sensory information from the entorhinal cortex and synaptically targets the distal dendrites of pyramidal cells (Amaral and Lavenex, 2007; Maccaferri and McBain, 1995) while the Schaffer collateral pathway carries internal mnemonic representations from CA3 and targets the proximal dendrites (Amaral and Lavenex, 2007; Cutsuridis et al., 2009). Complementing these spatially distinct excitatory inputs, local inhibitory neurons provide compartment-specific inhibition. In particular, somatostatin expressing interneurons (SOM) target the distal dendrites of pyramidal cells. Previous studies have revealed that SOM cells regulate bursting (Lovett-Barron et al., 2012), gate sensory information flow by reducing the influence of temporoammoammonic inputs (Leao et al., 2012), and are necessary for forming and recalling contextual memories (Lovett-Barron et al., 2014).

Given the memory deficits and altered E/I balance in RTT mice, we hypothesized that RTT CA1 pyramidal cells poorly differentiate between distinct mnemonic contexts and that this memory impairment is a consequence of reduced inhibitory synaptic input during the recall of a contextual memory. To test this hypothesis, we leveraged large-scale in vivo calcium imaging in freely-behaving animals and ex vivo slice recordings from identified cell types in the CA1.

Results

Long-term contextual fear memory recall is impaired in RTT mice.

To study neural population activity during contextual memory recall, we used a classic associative memory paradigm, the contextual fear conditioning (CFC) task, in conjunction with microendoscopic one-photon calcium imaging (Figure 1A). Female RTT mice and their WT littermates were exposed to two mild foot-shocks in a fear-conditioning context and subsequently re-exposed to this context 1 hr. and 1 day later so we could observe activities from CA1 neurons during short-term and long-term recall, respectively. As a control, mice were exposed to a novel neutral context 2 hrs. after training (Figure 1A). To behaviorally assess recall, video recordings were made from the top of each context chamber and the percentage of time each mouse spent freezing was measured. Four-month-old symptomatic RTT mice froze 57% less than WT mice during the long-term (1 day, Mann-Whitney, p < 0.01) but not short-term (1 hr., Mann-Whitney, p > 0.05) fear recall context (Figure 1B; Table S1). Importantly, no difference in freezing percentage was observed in the neutral context or conditioning context (Figure 1B inset) verifying that freezing differences between RTT and WT mice resulted from long-term memory impairment and not motor deficits or differences in pain thresholds (Figure S1A–B).

Figure 1.

Hippocampal imaging and event inference during the CFC task

(A) Depiction of experimental preparation timeline and task. Bottom: calcium signals were recorded while mice were exposed to environmental contexts (colored boxes) for 5-minutes. After exposure to a conditioning environment with foot-shocks, mice were re-exposed to the conditioning environment 1 h (short-term) and 1 day later (long-term) (green boxes). To assess baseline activity, mice were also exposed to a novel neutral environment (blue box).

(B) Assessment of memory recall by percentage of total time WT (n = 13) and RTT (n = 11) mice froze in recall and neutral contexts. No difference in freezing percentages were observed in the training or neutral context (inset).

(C) Left: Representative raw fluorescent image of hippocampal neurons labeled with GCaMP6f. Right: Maximum intensity projection of independent component images returned from the SVD-ICA segmentation model (left and right: scale bar= 125 μm).

(D) ΔF/F traces for a subset of 9 neurons displayed in C.

(E) Example of change-point event detection algorithm for section of blue trace shown in D. Blue dots are recorded fluorescent values, gray line is model’s predicted fluorescence and black ticks denote inferred events.

Boxplots show median and interquartile range (IQR) with whiskers extending to 1.5*IQR. Mann-Whitney used for independent comparisons, **p < 0.01; NS indicates nonsignificance p > 0.05.

To record the activity of the dorsal CA1 pyramidal cell population, we expressed the calcium indicator GCaMP6f under the CAMKIIα promoter in RTT and WT mice (Chen et al., 2013) (Figure S1C) and recorded fluorescent responses using a chronically implanted gradient index (GRIN) lens (Figure S1D). This system allowed us to monitor an average of 186 ± 24 neurons per animal while they performed the CFC task (Figure 1C). Since the large excitation focal volume of single-photon imaging introduces significant out-of-focus background contamination, we developed an imaging pipeline (Figure S2A–G) that combines Singular Value Decomposition (SVD) and Independent Component Analysis (ICA) (Mukamel et al., 2009) with annular background subtraction methods (Kerlin et al., 2010) to yield relative fluorescence (ΔF/F) signals (Figure 1D). Spikes were inferred from these signals using a recently developed spike inference algorithm (Jewell and Witten, 2018) (Figure 1E) that we tuned and validated on our data using a 2-fold cross-validation procedure (Figure S2K–M). Given our slow rate of image acquisition and the slow decay dynamics of our calcium reporter, we refer to these inferred spikes as events (Figure 1E black ticks).

Event rates poorly discriminate contextually distinct memories in RTT mice.

A shift in the balance of excitation and inhibition has been associated with alterations in the spontaneous firing rates of hippocampal and cortical neurons in mouse models of Rett syndrome (Dani et al., 2005; Zhang et al., 2008). To determine whether event rates are altered during real-time hippocampal processing of memories, we compared the population-averaged event rates in WT and RTT mice across the CFC task. Since each experiment spanned multiple contexts over two days, we confirmed that we could track individual cells within and across contexts of the task (Figure S2B–D, H–J). Analysis of cell responses during the training context was not performed due to large motion artifacts elicited by the conditioning foot-shocks (Figure S2B, D). WT mice exposed to the fear-conditioning context 1 hr. and 1 day later showed a 41% decrease in event rate compared to the neutral context (Wilcoxon signed-rank, p < 0.001) (Figure 2A; Table S1). This decrease in event rate accords with theoretical models that predict that memories are encoded by sparsely active populations of neurons (Kanerva, 1988; Rolls and Treves, 2011). Furthermore, we verified that this pattern of decreased event rates in the long-term (i.e. familiar) recall context was not due to the novelty of the neutral context (Figure S3A–C). In contrast, RTT mice displayed a lower event rate 1 hr. (Wilcoxon signed-rank, p < 0.05) but not 1 day (Wilcoxon signed-rank, p > 0.05) after fear conditioning relative to the neutral context. Thus, the average event rate of the CA1 pyramidal population in RTT mice, unlike WT mice, does not discriminate between a distantly experienced fearful and neutral environment.

Figure 2.

Decreased long-term context discrimination in RTT mice

(A) Average event rates in the neutral, short-term, and long-term contexts for WT (n = 13) and RTT (n = 11) mice. Each data point (not shown) of the boxplot is the cell-averaged event rate for a mouse.

(B) Percentage change in event rates during short and long-term memory recall relative to the neutral context in WT (n = 13) and RTT (n = 11) mice. Each data point (not shown) of the boxplot is the cell-averaged percent change in event rate. RTT mice have a reduction index closer to 0 during the long-term memory recall indicating less discrimination of a long-term context from a neutral context than WT mice.

(C) Empirical cumulative distribution function of percent change in rates across all recorded cells in WT (n = 2393) and RTT (n = 2043).

Boxplots show median and interquartile range (IQR) with whiskers extending to 1.5*IQR. Mann-Whitney used for independent comparisons and Wilcoxon-signed rank used for paired comparisons, *p < 0.05, ***p < 0.001; NS indicates nonsignificance p > 0.05.

To quantify the strength of the event rate discrimination between different contexts in WT and RTT mice, we calculated a reduction index that compares the event rate in each recall context relative to the neutral context (Figure 2B; Table S1). The more negative the value, the greater the difference in event rates between the recall and neutral context and therefore greater context discrimination. The rate reduction index of WT mice was 127% more negative than that of their RTT littermates during the long-term (Mann-Whitney, p < 0.05) but not short-term (Mann-Whitney, p > 0.05) fear memory recall. The poor discrimination in RTT mice was not due to non-mnemonic changes in event rates as the neutral context evoked similar event rates in both WT and RTT mice (Figure 2A). We should note that none of the RTT mice at this age showed indications of epileptic discharges that would have confounded these event rate measures. Consistent with these cell averaged reduction indices (Figure 2B), the distribution of rate reduction indices recorded across all cells (WT: n = 2393 cells, RTT: n = 2043 cells) had the greatest rightward shift for RTT mice during the long-term contextual memory recall (Figure 2C).

These findings indicate that poor memory recall in RTT mice correlates with a loss of event rate discrimination between a neutral and distantly experienced fear-inducing environment. We next sought to understand how the temporal pattern of activity within individual neurons relates to the recall of a contextual memory.

Long-term contextual memories activate larger and more functionally connected ensembles in RTT mice.

Synchronous activation of CA1 pyramidal cells and their associated local field potential oscillations (Buzsáki, 2015) is necessary for the consolidation of recently experienced events (Ego-Stengel and Wilson, 2009; Girardeau et al., 2009) and is thought to promote the transfer of memories to the neocortex for long-term storage (Buzsáki, 1989; Káli and Dayan, 2004). Consistent with prior studies of CA1 activity in freely moving rats (Buzsaki et al., 1992), we observed transient co-activation of multiple CA1 pyramidal cells in both WT and RTT mice during each context of the CFC task (Figure 3A–B). To quantify the maximum coordinated network activity, we identified the largest of these transient ensembles by convolving each inferred raster with a Gaussian window and averaging across cells (Figure 3A–B, Figure S3D). Comparing the percentages of cells active during the largest peak that exceeded a chance correlation threshold (Figure 3A–B, arrowheads and green threshold; see STAR Methods) revealed that RTT mice have a greater percentage of transiently co-activated neurons than WT mice during long-term memory recall (Mann-Whitney, p < 0.05) (Figure 3C; Table S1). Importantly, this difference in the number of co-activated cells was not due to an increase in the event rates of coactive cells in RTT mice as their average event rates were indistinguishable from the average event rates of coactive cells in WT mice and the non-coactive cells of RTT mice (Figure 3D, Kruskal-Wallis, p > 0.05).

Figure 3.

Context-dependent increases in ensemble sizes and functional connectivity in RTT mice

(A-B) Averaged activity traces from 200 WT and 186 RTT cells (top) and 25 randomly selected event rasters (bottom) during long-term memory recall. Arrowheads mark the time-point with the largest co-activation of cells. Gray rectangles mark periods of freezing and green horizontal line is the threshold for significant peaks using time shifted surrogate data (STAR Methods).

(C) Boxplots of percentage of coactive cells during largest activity peak in neutral and recall contexts for WT and RTT mice (WT; Neutral: n = 11, 1 hr.: n = 12, 1-day: n = 12; RTT; Neutral: n=9, 1 hr.: n = 9, 1-day: n = 10).

(D) Boxplot of average event rates from all WT cells (n = 13 mice), WT coactive cells (n = 12 mice), WT non-coactive cells (n = 13 mice), all RTT cells (n = 11 mice), RTT coactive cells (n = 10 mice) and RTT non-coactive cells (n = 10 mice) reveal that event rate differences between RTT coactive and WT coactive cells are not driving the increases in coactivity in RTT mice (Figure 3C).

Each data point (not shown) of the boxplot is the cell-averaged event rate for each mouse and group.

(E) Activity traces (top) of two sample cells with a high correlation value (> 0.3). Histograms of number of correlated partners for a WT and RTT mouse in the neutral (middle) and long-term fearful context (bottom). Blue and orange dashed lines represent the boundary between low and high-degree cells for the WT and RTT mouse respectively (STAR Methods).

(F) XY positions of the high-degree cells from 3 WT (top) and 3 RTT (bottom) mice in the neutral and long-term fear contexts.

(G) Number of neurons comprising each high-degree ensemble for WT (n = 13) and RTT mice (n = 11) separated by context.

(H) Average number of correlated partners within each high-degree ensemble for WT (n = 13) and RTT (n = 11) mice separated by context.

(I) Boxplots of event rates for RTT high-degree ensemble cells (HD) in the neutral context (n=171 cells) and all RTT cells (n = 2043), RTT HD cells (n = 494) and RTT non-high-degree (non-HD) (n = 1549) cells in the long-term recall context reveal that larger ensemble sizes in RTT are not driven by increased firing rates in RTT HD cells during the long-term recall. Each data point (not shown) of the boxplot is an individual RTT cell.

Boxplots show median and interquartile range (IQR) with whiskers extending to 1.5*IQR. Mann-Whitney used for independent comparisons. Kruskal-Wallis used for unpaired omnibus test. Wilcoxon-signed rank used for paired comparisons. Bootstrapped confidence intervals of median differences (10000 resamples) used for comparisons of cell event rates. *p < 0.05, **p < 0.01; NS indicates nonsignificance p > 0.05.

In addition to transient ensembles, we observed cells whose activities were persistently correlated (Pearson’s > 0.3) (Cossell et al., 2015) for the duration of the context exposure (Figure 3E top). While the activities of many CA1 cells in WT and RTT mice were correlated with a few other cells, a sub-population, termed high-degree cells, had correlated activities with many other neurons (Figure 3E middle and bottom) forming a high-degree ensemble. Given the large number of recorded cells and the likelihood of chance correlations in their activity patterns, we identified cells belonging to this ensemble with a graph theory approach. Specifically, we identified neurons within each context that possessed a statistically higher number of correlated partners than expected by chance (cells to the right of blue (WT) and orange (RTT) dashed lines in Figure 3E middle and bottom; see STAR Methods). Cells belonging to these high-degree ensembles were evenly distributed throughout the field of view in both WT and RTT mice (Figure 3F, top: 3 WT mice, bottom: 3 RTT mice). To analyze the composition of these ensembles, we looked at both the number of neurons comprising each ensemble (i.e. dots in Figure 3F) (Figure 3G) and the number of correlated partners per neuron (i.e., number of lines per dot in Figure 3F) (Figure 3H) across contexts of the CFC task. In WT mice, the sizes of the high-degree ensembles were stable across contexts. In RTT mice, however, the long-term fear memory recruited 236% more high-degree neurons than the neutral context (Wilcoxon signed-rank, p < 0.01) (Figure 3G; Table S1). In addition, RTT mice also showed a 409% increase in the number of correlations per neuron during the long-term memory recall relative to the neutral context (Wilcoxon signed-rank, p < 0.01) (Figure 3H; Table S1). To ensure that high-degree ensemble neurons in RTT mice did not simply have higher event rates during the long-term recall that artificially inflated the ensemble size, we compared the high-degree cell event rates during the neutral context (n=171 cells) to the long-term recall context (n=494 cells) in RTT mice (Figure 3I) and found no statistical difference in the event rates of the high-degree cells (Bootstrapped 95% confidence interval of the difference of the medians using 10000 resamples) (Figure 3I; Table S1).

These data demonstrate that RTT mice have larger and more correlated ensembles of CA1 pyramidal cells than WT mice during the long-term recall of a fearful memory. Given that artificial inflation of ensemble sizes by optogenetic inactivation of inhibitory cell types impairs memory recall in the CFC task (Stefanelli et al., 2016), we next explored if reduced inhibition in the CA1 circuit might support the formation of larger ensembles during long term fear memory recall in RTT mice.

SOM cells lacking MeCP2 are weakly recruited by CA1 pyramidal cells.

GABAergic interneurons are the primary regulators of neuronal excitability, a critical mechanism that controls ensemble size (Rao-Ruiz et al., 2019). To identify specific interneuron types that contribute to increased event rates and ensemble sizes in RTT mice, we used whole-cell recordings in conjunction with cre-dependent viral expression of fluorescent proteins and biocytin filling to identify specific cell types on the basis of their location and morphology in CA1. Furthermore, since random X chromosome inactivation leads to MeCP2 expression in half of all neurons in female RTT mice, we used post hoc immunostaining to differentiate and compare cells with and without the MeCP2 protein. To assess alterations in inhibition across the pyramidal cell population, we measured spontaneous inhibitory post-synaptic currents (sIPSCs) at their somas in the presence of CNQX and APV to block AMPA and NMDA receptors respectively while holding the cell at −70 mV (Figure 4A–B). We found that both the frequency and amplitude of sIPSCs were similar across WT and RTT mice, irrespective of the presence of MeCP2 (Kruskal-Wallis, p > 0.05) (Figure 4C; Table S1). Post hoc biocytin and immunostaining was used to confirm the cell type (i.e., pyramidal cells) and segregate cells on the basis of MeCP2 expression (Figure 4D). The absence of any measurable difference in sIPSCs at the somatic recording site coupled with a lack of distance-dependent scaling of inhibitory post-synaptic currents (IPSCs) (Andrásfalvy and Mody, 2006) suggested that any changes in inhibition might arise from interneurons that synaptically target more distal dendrites.

Figure 4.

Reduced excitatory synaptic responses in RTT-OLM cells lacking MeCP2

(A) Recording schematic for pyramidal cells (left) and reconstruction (right).

(B) sIPSCs in WT (blue), RTT MeCP2+ (orange) and RTT MeCP2− (teal) pyramidal cells.

(C) Pyramidal cell sIPSC frequencies (left) and amplitudes (right) (WT, n = 16, RTT MeCP2+, n = 9, RTT MeCP2−, n = 15).

(D) Confocal images of an MeCP2− pyramidal cell (arrowhead); biocytin (left, scalebar = 80 μm), DAPI (middle-left), MeCP2 antibody (middle-right) and merge (right). Box in biocytin image is zoomed region for remaining images (scale bar = 50 μm).

(E) Recording schematic for OLM cells (top), OLM cell current injection responses (bottom-left) and biocytin reconstruction (bottom-right).

(F) sEPSCs in SOM-cre, SOM-cre/RTT MeCP2+ and SOM-cre/RTT MeCP2− OLM cells (same colors as B).

(G) OLM cell sEPSC frequencies (left) and amplitudes (right) (OLM, n = 43, RTT-OLM MeCP2+, n = 26, RTT-OLM MeCP2−, n = 27).

(H) Images of an MeCP2+ OLM cell (arrowhead); staining (left, scale bar = 60 μm), confocal tdTomato and Biocytin image (middle-left), MeCP2 antibody (middle-right) and merge (right). Box in DAB image is zoomed region for remaining images (scale bar = 50 μm).

(I) Triple recording schematic from OLM cell (red) and two pyramidal cells (green and black) (top-left) with biocytin reconstructions (top-right) and OLM responses to pyramidal cell current injection (bottom).

(J) Pyramidal to OLM cell connectivity percentages.

(K) OLM cell unitary EPSP amplitudes for cell pairs in J.

(L) OLM to pyramidal cell connectivity percentages.

(M) Pyramidal cell unitary EPSP amplitudes for cells pairs in L.

Barplots show mean and standard deviation. Kruskal-Wallis used for omnibus test. Mann-Whitney used for independent comparisons. Binomial used for connectivity comparisons, *p < 0.05, **p < 0.01; NS indicates nonsignificance p > 0.05.

A subset of SOM interneurons, the Oriens-lacunosum moleculare (OLM) cells, have cell bodies localized in the stratum oriens, receive glutamatergic inputs from CA1 pyramidal cells and recurrently target the distal dendrites of pyramidal cells (Müller and Remy, 2014). Interneurons located in the oriens layer receive less excitatory input in RTT mice (Lu et al., 2016), but it is unknown whether OLM cells are specifically poorly recruited in the CA1 circuit. We therefore crossed RTT mice with Som-Cre mice (RTT/Som-Cre) and injected cre-dependent AAVs (AAV-hSyn-LSL-tdTomato) expressing tdTomato and performed whole-cell patch recordings from SOM neurons in the oriens layer (Figure 4E top). The electrophysiology of the targeted SOM cells was consistent with OLM cells (low-threshold spiking, sustained outputs, and a “sag” in response to hyperpolarizing current injections) (Maccaferri and McBain, 1996) (Figure 4E bottom left). Additionally, post hoc biocytin staining verified that the recorded cell’s morphology matched that of OLM cells (Figure 4E bottom right). We compared the spontaneous excitatory post-synaptic currents (sEPSCs) in OLM cells at a holding potential of −70 mV from WT and RTT mice (Figure 4F–G). As before, post hoc immunostaining was used to segregate OLM cells in RTT mice based on MeCP2 expression (Figure 4H). RTT-OLM MeCP2− cells had a 62% lower sEPSC rate than WT-OLM and 59% lower rate than RTT-OLM MeCP2+ cells (Mann-Whitney, p = 0.001 and p = 0.003 respectively) (Figure 4G left; Table S1). Similarly, RTT-OLM MeCP2− cells had a 19% lower sEPSC amplitude than WT-OLM cells (Mann-Whitney, p = 0.012). (Figure 4G right; Table S1). This decrease in excitatory input was not accompanied by a commensurate decrease in inhibitory input (Figure S4A–C) providing further evidence for decreased recruitment as opposed to basal activity deficits in MeCP2− RTT-OLM cells. Furthermore, this result highlights the E/I imbalance in OLM cells lacking MeCP2.

To understand whether this decrease in sEPSC rate and amplitude results from a reduction in connectivity between CA1 pyramidal and OLM cells, we performed simultaneous multiple whole-cell patch recordings (Figure 4I). A cell pair was defined as being connected if a unitary postsynaptic potential was observed in a neuron in response to an action potential in its putative presynaptic partner (Figure 4I bottom). Given the low rate of finding connected pairs, we did not separate MeCP2+ and MeCP2− expressing cells in RTT mice. Nevertheless, we observed a significant reduction of connectivity between pyramidal to OLM cells in RTT mice with only 2 of 68 tested pairs yielding connected cells in RTT mice compared with 10 of 81 tested pairs in WT mice (Binomial, p = 0.015) (Figure 4J). Although the connectivity in RTT mice was low, we found no evidence for a difference in the unitary excitatory post-synaptic potential (EPSP) amplitude between WT and RTT mice (Figure 4K). To understand if this synaptic deficiency was limited to the pyramidal to OLM synapse, we measured the connectivity rate from OLM to pyramidal cells and found no difference between WT and RTT mice (Figure 4L, M). Importantly, we confirmed that the poor recruitment of MeCP2− OLM cells in RTT mice was not due to poor recruitment of MeCP2− pyramidal cells as WT, RTT-MeCP2+ and RTT-MeCP2− pyramidal cells had similar sEPSC rates and amplitudes (Figure S4D–F).

These data demonstrate that MeCP2− OLM cells are poorly recruited in RTT mice, resulting in a weakened dendritic inhibition of CA1 pyramidal cells. Given that dendritic inhibition plays a critical role in regulating the input-output relationship of CA1 pyramidal cells (Lovett-Barron et al., 2012) and controls ensemble sizes (Stefanelli et al., 2016), the impaired synaptic connectivity from CA1 pyramidal cells to OLM cells in RTT mice could explain the higher event rates, larger ensembles and, ultimately, the poor memory recall observed in RTT mice. We next tested whether enhancing SOM activity during the CFC task could improve memory recall in RTT mice.

Enhancing SOM activity in RTT mice restores long-term contextual memory recall.

Contextual fear memory recall requires the recruitment of SOM cells (Lovett-Barron et al., 2014). Thus, the impaired recruitment of OLM cells in RTT mice may be one of the reasons for their poor performance in the CFC task. If so, we would expect to see a reduction of SOM activity during the long-term recall component of the CFC task in RTT mice compared to WT mice. To test this hypothesis, we injected AAV1-Flex-GCaMP6f in Som-Cre and RTT/Som-Cre mice to express GCaMP6f in SOM cells and performed microendoscopic imaging (Figure 5A). Consistent with the high frequency of sEPSCs in SOM cells observed in vitro (Figure 4G), we found identifying a quiescent period of fluorescence for estimating relative changes in fluorescence (ΔF/F) difficult. Instead, we used a z-score of the fluorescence values to estimate the calcium signal deflections around the mean (Figure 5B). Taking the area below this z-score trace as a measure of the response, we found that the average SOM cell response in RTT/Som-Cre was smaller than SOM cells in Som-Cre mice during the long-term recall context (Mann-Whitney, p < 0.05) (Figure 5C; Table S1). This modest reduction is consistent with a subset of SOM cells, the MeCP2− OLM cells, being poorly recruited by pyramidal cells (Figure 4G, J). To determine whether reduced SOM activity (Figure 5C) diminishes long-term recall of a contextual fear memory in RTT mice, we conditionally expressed the Designer Receptors Exclusively Activated by Designer Drugs: AAV-FLEX-hM3D(Gq)-mcherry, AAV-FLEX-hM4D(Gi)-mcherry in SOM cells (Figure 5D and Figure S5A). These receptors enhance (hM3D (Alexander et al., 2009)) or suppress (hM4D (Zhu et al., 2014)) neural activity when exposed to the chemical activator clozapine-N-oxide. Activation of SOM cells in Som-Cre mice enhanced fear learning (Figure S5B), while suppression of SOM cells in Som-Cre mice reduced freezing both 1 hr. (Mann-Whitney, p = 2.6 × 10⁻⁴) and one day (Mann-Whitney, p = 1.30 × 10⁻⁵) after fear conditioning (Figure 5E; Table S1). In contrast, enhancing SOM cell activity in RTT/Som-Cre mice increased freezing during both the short (Mann-Whitney, p = 0.009) and long-term (Mann-Whitney, p = 4 × 10⁻⁴) recalls with the largest effect size of 0.81 occurring for the long-term memory recall. We verified that mCherry had no impact on the poor recall in RTT mice (SOM-Cre-mCherry 1 day to RTT/SOM-Cre-mCherry 1 day, Mann-Whitney, p = 0.0006 Table S1) and that changes in freezing were task related as evidenced by similar responses in the neutral context (Kruskal-Wallis, p > 0.05) (Figure 5E; Table S1).

Figure 5.

Enhancing SOM cell activity in RTT mice during the CFC task improves long-term memory recall

(A) Representative fluorescent image of CA1 SOM neurons expressing GCaMP6f (top) and maximum intensity projection of independent component images returned from SVD-ICA segmentation (bottom, scale bar = 125 μm).

(B) Selection of z-scored fluorescence traces from cells shown in A. Gray box is zoomed region for bottom trace, black line indicates…...

(C) SOM responses as measured by area under z-scored activity traces for WT (n = 13) and RTT (n = 11) mice. Boxplot data are the average area of all SOM cells recorded in a mouse for a given context.

(D) Confocal images of CA1 SOM cells stained with DAPI (top-left), Somatostatin antibody (top-right), hM3D(Gq)-mcherry (bottom-left) and merge (bottom-right, scale bar = 50 μm).

(E) Freezing during CFC task of SOM-Cre mice conditionally expressing mcherry (blue, n = 19 mice), hM4D (light blue, n = 21 mice) and RTT/SOM-Cre mice conditionally expressing mcherry (orange, n = 20 mice) and hM3D (yellow, n = 27 mice). Boxplots show median and interquartile range (IQR) with whiskers extending to 1.5*IQR. Kruskal-Wallis used for omnibus test. Mann-Whitney used for independent comparisons, *p < 0.05, **p < 0.01, ***p < 0.001; NS indicates nonsignificance p > 0.05.

Discussion

While it is increasingly clear that imbalance in excitation and inhibition contributes to the pathophysiology in Rett syndrome (Dani et al., 2005; Marín, 2012; Zhang et al., 2008), identifying the source of this imbalance and its impact during a specific cognitive task remains unresolved. Here, we imaged the CA1 pyramidal cell population of RTT and WT mice during the contextual fear conditioning task. We found that event rates of pyramidal cells in RTT mice, unlike those in WT mice, poorly discriminate a distantly experienced fear-inducing environment from a neutral environment. We also discovered that RTT mice have larger and more correlated cell assemblies during the long-term memory recall component of the task than WT mice. Furthermore, the increase in pyramidal cell co-activity suggests that seizures that develop later in the progression of Rett syndrome may have their origins in the enhanced co-activation of excitatory cells within learning and memory circuits.

Several lines of evidence support that OLM cells, a subset of SOM cells (Pelkey et al., 2017), are critical for learning and memory. These cells modulate synaptic plasticity in CA1 (Nakauchi et al., 2007), increase their spine density during learning (Schmid et al., 2016), and are necessary for contextual learning and memory retrieval (Lovett-Barron et al., 2014). In this study, we found that OLM cells which lack MeCP2 in RTT mice receive less excitation from the surrounding pyramidal cell population. This deficiency was local, as optogenetic stimulation of the medial septum, a brain region that provides direct cholinergic input to OLM cells, evoked similarly-sized EPSCs in OLM cells irrespective of the presence of MeCP2 (Figure S4G–I). Based on the literature and the current study, we propose that the poor recruitment of OLM cells is causally related to the increased event rates and ensemble sizes measured in RTT mice. Indeed, OLM cells have been shown to strongly regulate CA1 pyramidal cell excitability, a primary factor in determining which cells participate in a memory encoding ensemble (Park et al., 2016). One interesting and unexplored facet of this study is that despite the poor recruitment of OLM neurons (Figure 4G, J), only the long-term memory recall appears impacted (Figure 1B). Indeed, the event rates (Figure 2A), coactive cell percentages (Figure 3C), ensemble sizes (Figure 3G–H), and SOM in-vivo activity (Figure 5C) all show alterations during the long-term but not short-term memory recall in RTT mice. While speculative, this finding suggests that short and long-term recalls of a fearful memory partially rely on independent circuit mechanisms.

We established a causal relationship between impaired OLM cell recruitment and poor memory recall by chemogenetically enhancing SOM cell activity during the CFC task. This improved behavioral recall in both the short- and long-term fearful contexts without affecting behavior in the neutral context. A parsimonious explanation for why both recall contexts were affected despite only long-term recall being altered in vivo, is that our manipulation enhanced the activity of both OLM and non-OLM SOM cells. Future studies that use more specific Cre-lines such as the Chrna2-Cre mouse line (Leao et al., 2012) that specifically labels OLM cells can refine this analysis.

Given the marked improvement in performance upon activation of SOM cells in RTT mice, it is natural to ask what possible therapeutic strategies might be employed to relieve the cognitive symptoms of the disease in human patients. Deep brain stimulation (DBS) of the fornix has shown remarkable efficacy in restoring hippocampal-dependent learning and memory in mice (Hao et al., 2015). DBS increases the frequency of sEPSCs in interneurons of the oriens layer where OLM cells reside (Lu et al., 2016). While the mechanisms of DBS remain poorly understood, these previous results paired with the current study suggest that recruitment of OLM cells helps to restore learning and memory deficits. Importantly, the medial septum provides direct cholinergic input to OLM cells leading to sustained depolarization, and is required for the CFC task (Calandreau et al., 2007). Thus, the medial septum could serve as a target for DBS to treat cognitive disabilities in RTT. A potential advantage of using DBS in the medial septum is that it could treat seizures that develop in patients in the late stages of the disease. Indeed, closed-loop DBS stimulation of the medial septum in rats can arrest complex partial seizures, a common seizure type in RTT (Operto et al., 2019) in real-time (Takeuchi et al., 2021). This dual role of cognitive improvement and seizure prevention makes the medial septum an attractive target for DBS intervention.

While this study has focused on Rett syndrome, many neurological disorders of disparate molecular etiologies converge to alter the development and function of inhibitory circuits (Marín, 2012). Identifying the relevant interneurons and the resulting circuit disruptions in the context of a specific cognitive task will provide an anatomical basis for the illness, lead to insights into pathogenesis, and perhaps to more targeted therapies (Akil et al., 2010).

STAR METHODS

Detailed methods include the following:

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Huda Y. Zoghbi (hzoghbi@bcm.edu)

Materials availability

No new materials were generated in these studies.

Data and code availability

The calcium signals, behavior data, and all associated data needed to generate the figures in this study are stored at Zenodo: https://doi.org/10.5281/zenodo.5999292. The code to generate the figures in this manuscript can be found at Github: https://github.com/mscaudill/rett_memory. The original calcium imaging movies were not deposited because of their large size but will be made available upon request to the lead contact.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Male C57BL/6 mice were mated with female heterozygous Mecp2^+/− mice on 129 background to obtain F1 hybrid offspring. We housed the adult female Mecp2^+/− mice (15–16 weeks of age) with standard food and water ad libitum in an AAALAS-accredited Level 3 facility on a 14-hr. light cycle (light on at 6am). Mice were housed up to five per cage before surgery and singly housed after surgery. Male SOM-IRES-cre (sst^{tm2.1(cre)Zjh/J}, Jackson Laboratory stock number 13044) allele were crossed with Mecp2^+/− mice under 129 background to label SOM positive cells. To activate cholinergic fibers in CA1, male ChAT-ChR2-EYFP (B6.Cg-Tg(Chat-COP4*H134R/EYFP, Slc18a3)6Gfng/J, Jackson Laboratory stock number 14546) mice were bred to Mecp2^+/− mice under 129 background. Institutional Animal Care and Use Committee (IACUC) of Baylor College of Medicine reviewed and approved all husbandry and research protocols.

METHOD DETAILS

Stereotactic surgeries

All surgical procedures were performed in accordance with Baylor College Medicine’s guidelines. The mice were given slow-release buprenorphine (1 mg/kg) 1 hr. and meloxicam (5 mg/kg) 0.5 hr. prior to anesthesia. Mice were weighed and anesthetized preoperatively with intraperitoneal injection of Rodent Combo III (1.5 ml/kg, ketamine 37.5 mg, xylazine 1.9 mg and acepromazine 0.37 mg) and placed on a stereotaxic rig (David Kopf Instrument). Constant isoflurane (1–2% in O₂) was also administrated via vaporizer (Veterinary Anesthesia Systems, Inc) until mice no longer had a toe-pinch response. Ophthalmic ointment was then dabbed on the eyes, the scalp was shaved, and the skin was cleansed with betadine and 70% alcohol (alternately, three times). All surgical tools were autoclaved before the surgeries and sterilized with bead sterilizer between animals. For all the survival surgeries, mice were monitored and injected with buprenorphine and meloxicam every 24 hours for 3 days post-operatively.

For virus injection, a craniotomy was created on the skull and mice were injected with AAVs expressing GCaMP6f (pENN.AAV.CamKII.GCaMP6f.WPRE.SV40, Addgene: 100834-AAV1) into the dorsal hippocampal CA1 (anteroposterior: −1.9 mm from bregma; mediolateral: 1.5 mm from midline; −1.4 mm from skull surface). We unilaterally microinjected 500 nl of virus at 70 nl/min with a 10 uL syringe (Hamilton) and a pump controller (Micro 4 MicroSyringe Pump Controller). For DREADD expression in the CA1, we bilaterally injection 500 nl of virus (AAV-hSyn-DIO-hM3D(Gq)-mCherry, Addgene: 44361-AAV2; pAAV-hSyn-DIO-hM4D(Gi)-mCherry, Addgene: 44362-AAV2; pAAV-hSyn-DIO-mCherry, Addgene: 50459-AAV2) at three coordinates above the dorsal CA1. We waited for 10 min for virus infusion after the injection was done. The scalp was sutured with 5–0 vicryl suture (MCKesson) and the animal was placed in the recovery cage on the heatpad until mobility was observed.

Three to four weeks after injecting AAVs expressing GCaMP6f, we performed a surgery to insert a cannula for the endoscope. A 1.8 mm circular craniotomy centered 1 mm to the injection site with a 1.8 mm trephine was created (Fine Science Tools). The dura was removed with blended forceps and the cortical tissue below the dura was aspirated with a 27-gauge blunt needle (VWR) down to 1.0 mm. A 30-gauge blunt needle (VWR) was then used to aspirate the cortex down to the corpus callosum. Saline was continuously perfused to stop bleeding from the brain tissue. The cannula (1.8 mm, Inscopix) was implanted into the corpus callosum and sealed with VetBond Tissue Adhesive (3M). To head-fix the mouse for subsequent procedures, we attached an aluminum head bar to the skull and applied Metabond (Parkell) to the skull surface. Three to four weeks after cannula implantation, a gradient refractive index (GRIN) lens (4 mm length, 1 mm diameter, Inscopix Inc.) was inserted into the cannula. To help place the GRIN lens, a miniature microscope (nVistaHD, Inscopix Inc.) was lowered towards the GRIN lens until the GCAMP6f fluorescence was clearly observed. The GRIN lens was fixed with VetBond Tissue Adhesive (3M) along with a baseplate fixed using Metabond (Parkell) above the lens for later attachment of the miniature microscope.

Drugs and AAVs

Clozapine N-oxide (CNO, Tocris) was prepared in dimethyl sulfoxide (DMSO) at concentration of 20 mg/ml and intraperitoneally injected at 5 mg/kg in mice 1 hr. before fear conditioning test. pENN.AAV.CamKII.GCaMP6f.WPRE.SV40 (Addgene 100834-AAV1) was injected to CA1 area 3–4 weeks before lens implantation. The DREADD viruses pAAV-hSyn-DIO-hM3D(Gq)-mCherry (Addgene: 44361-AAV1), pAAV-hSyn-DIO-hM4D(Gi)-mCherry (Addgene: 44362-AAV1), or control virus pAAV-hSyn-DIO-mCherry (50459-AAV1) were injected in the dorsal CA1 to express hM3D, hM4D, or mCherry only, respectively. AAV1.CAG.LSL.tdTomato.bGH (Penn Vector Core: AV-1-ALL856) conditionally expressing tdTomato was injected to CA1 to label SOM cells for slice recording.

Behavioral procedures and Ca²⁺ imaging

All behavioral tests were conducted during the light cycle. Before each test, the miniature microscope was attached to the baseplate of a mouse under isoflurane anesthetization (1–2% in O₂) and the focus ring was adjusted until the fluorescence from the GCaPM6f was clearly observed. We secured the microscope on to the baseplate with a screw and Kwik-cast silicone sealant (World Precision Instruments). The focus was maintained at the same position across the fear conditioning and memory recall tests. Each mouse was habituated in the testing room for at least 1 hr before any behavioral tests were performed. The environment for habituation and testing included white noise at 60 dB and the light intensity at 100 lx. Animal behaviors and calcium imaging movies were recorded simultaneously with the TTL-trigger interface. Calcium signals were acquired at 20 Hz using the nVista software (Inscopix Inc.) with the LED power 60–150 uW. Mice were excluded if acquired data sets were incomplete or AAV injection was not successfully targeted to CA1.

Fear conditioning

We used a modified contextual fear conditioning protocol (Hao et al., 2015) to evaluate hippocampal-dependent memory recall. The conditioning chamber was an opaque square box with a stainless-steel grid on the floor for delivering electric current (Med Associates, Inc.). A digital camera was placed above the chamber to record animal position. A mouse was allowed to freely explore for 2.5 mins before the delivery of a time scrambled foot-shock (2 sec duration at 0.7 mA). This initial foot-shock was followed by a second shock of the same intensity 1.5 mins later. The mouse was removed from the chamber 1 min after the second foot shock and returned to its home cage. The conditioning chamber was cleaned with 70% alcohol between mice. Short- and long-term memory were assessed 1 hr. and 1 day later in the same chamber for 5 min. As a control, mice were exposed to a novel neutral context consisting of a round plexiglass cage with corncob bedding. This chamber was cleaned between mice with 1% acetic acid. As with the conditioning chamber, behavior was recorded with a video camera placed above the chamber. Freezing was defined as the cessation of all non-respiratory movement for at least 1 s and was automatically detected with the ANY-maze (Stoelting) software system.

Open field test

Open field locomotion was tested in a clear Plexiglas box (40 × 40 × 30 cm, Stoelting) chamber with an overhead camera to record the animals’ behaviors. Total and center distances were automatically quantified by ANY-maze (Stoelting).

Pain threshold

Pain thresholds for flinching, vocalization and jumping were measured in the conditioning chamber. The electric current thresholds were tested by increasing the current delivered during a 2-s foot shock with 0.05 mA increment applied every 30 s.

Slice preparation

Acute hippocampal slices were prepared from 4-month-old female WT, RTT, SOM-cre, RTT/SOM-cre, ChAT-ChR2 or RTT/ChAT-ChR2 mice according to previously described methods (He et al., 2014). Briefly, the mice were deeply anesthetized with 2% isoflurane in the induction chamber and decapitated. The brain was dissected and placed into ice-cold oxygenated cutting solution (93 mM NMDG, 93 mM HCl, 2.5 mM KCl, 1.2 mM NaH₂PO₄, 30 mM NaHCO₃, 20 mM HEPES, 25 mM glucose, 5 mM sodium ascorbate, 2 mM Thiourea, 3 mM sodium pyruvate, 10mM MgSO₄ and 0.5 mM CaCl₂, pH 7.35, bubbled with 5% CO₂ and 95% O₂). Coronal hippocampal slices were prepared from brain tissue with Vibratome (Leica Biosystems) at 400 uM thickness. Slices were incubated with physiological aCSF (125 mM NaCl, 2.5 mM KCl, 1.25 mM NaH₂PO₄, 25 mM NaHCO₃, 25 mM glucose, 1 mM MgCl₂ and 2 mM CaCl₂, pH 7.4, bubbled with 5% CO₂ and 95% O₂) at 34°C for 1 hour and then transferred to room temperature for > 30 min before recording. The physiological aCSF was perfusing the recording chamber at ~2mL/min at room temperature.

Electrophysiological recordings

Single whole-cell recordings were carried out from CA1 pyramidal neurons or SOM cells in the oriens layer under infrared-differential interference contrast (DIC) visualization and a charge-coupled device camera (CCD). Current and voltage clamping recordings were performed using an Axon 700B amplifier (Molecular Devices). Electric signals were low-pass filtered at 4 kHz and digitized (Digidata 1322A, Molecular Devices) at 10 kHz. The data were acquired with Axon Clampex 10.3 (Molecular Devices) and subsequently analyzed with Minianalysis 6.0.3 (Synaptosoft Inc.).

Spontaneous IPSCs from pyramidal cells or SOM cells in the oriens layer were recorded at −70 mV in voltage-clamp mode with bath application of 50 uM CNQX and 25 uM APV. Cs-containing intracellular solution contained (in mM) 132 CsCl, 2 MgCl₂, 0.16 CaCl₂, 10 HEPES, 5 QX-314 (not included for recording sIPSCs in SOM cells), 0.5 EGTA, 2 MgATP, 0.4 Na₃GTP, 10 Sodium phosphocreatine and 0.5% biocytin (pH 7.3).

For spontaneous EPSCs recordings, SOM cells in the oriens layer or pyramidal cells were held at −70 mV in voltage-clamp mode in the presence of 100 uM Picrotoxin. The recording pipettes (4–6 M) were filled with internal solution containing (in mM) 120 potassium gluconate, 10 HEPES, 4 KCl, 4 MgATP, 0.3 Na₃GTP, 10 sodium phosphocreatine and 0.5% biocytin (pH 7.25).

In the experiments of light evoked IPSCs, to active ChR2+ fibers in the oriens layer, blue light was emitted from a LED (470 nm, Thorlabs M470L3) through a 60× water objective lens above the CA1 area. Evoked EPSCs were recorded from OLM cells in the presence of 50 uM CNQX, 25 uM APV, and 100 uM Picrotoxin. EPSC amplitudes were calculated as the peak amplitudes of pre-defined window after the light stimulation subtracting baseline amplitudes. The failure rate was counted as the percentage of responsive episodes out of all recorded episodes.

Simultaneous multiple whole-cell patch recordings were performed as previously described (Jiang et al., 2015). Briefly, two Quadro EPC10 amplifiers and a built-in LIH 8+8 interface board were used to collect electric signal and simultaneously perform A/D and D/A conversion. The glass pipette contained (in mM) 120 potassium gluconate, 10 HEPES, 4 KCl, 4 MgATP, 0.3 Na₃GTP, 10 sodium phosphocreatine and 0.5% biocytin (pH 7.25).

Current steps were injected into the recorded cells to determine the intrinsic properties. We calculated the amplitudes of unitary IPSCs or EPSCs by averaging 20–50 stimulation episodes.

For all recordings, the electrodes were slowly pulled from the recorded cells to keep the cell membrane intact. High concentration of biocytin (0.5%) was added in the internal solution for reconstructing morphology and further separating cells based on MeCP2 expression. Cells were discarded when the change in the series resistance was >20% during the recording.

Immunohistochemistry

After the slice recording, the slices were fixed in a freshly prepared 4% paraformaldehyde with 2.5% glutaraldehyde in 0.1 M PBS at 4° for 24 hrs. The slices were rinsed with PBS and subsequently incubated with rat polyclonal antibody against somatostatin (1:250, Millipore Cat# MAB354), rabbit polyclonal antibody against MeCP2 (1:1000, Cell Signaling Technology Cat# 3456S) overnight at 4°. Secondary antibody Alexa Fluor 488 antibody against rabbit IgG (1:1000, Thermo Fisher Scientific Cat# A-11008 or Alexa Fluor 647 antibody against rat IgG (1:500, Thermo Fisher Scientific Cat# A-21247) incubated the slices for 2 hrs at room temperature. Alexa Fluor 568 conjugated streptavidin (1:1000, Thermo Fisher Scientific Cat# S11226) was used for biocytin visualization. Nuclei were stained with DAPI (4′,6-diamidino-2-phenylindole). Fluorescence images of hippocampal sections were acquired using a Leica SP8X confocal microscope with × 10 or × 25 water emersion objective lens, followed by image processing in Fiji (National Institutes of Health).

Morphological reconstructions

We performed avidin-biotin-peroxidase staining as previously described (Jiang et al., 2015). After acquiring fluorescence images, the slices were then rinsed with PBS for 3 times followed with incubation in 3% H₂O₂ solution for 30 min at room temperature. To better recover the axonal arbors of OLM cells, we incubated the slices in avidin-biotin complex and high concentration detergents (Triton-X 100, 5%) overnight prior to DAB staining (0.5 mg/ml). The morphology was examined and reconstructed under a camera lucida system with a × 100 oil emersion objective lens (numerical aperture 1.25).

QUANTIFICATION AND STATISTICAL ANALYSIS

Calcium imaging analysis

Raw fluorescent images collected with the endoscope were down-sampled using a 2 × 2 binning to a resolution of 540 × 720 pixels and motion corrected using the Mosaic (Inscopix) software. All subsequent processing and analysis were carried out with a custom imaging pipeline developed in Python (Figure S2A–G). This open-source pipeline leveraged the following packages: Numpy (Harris et al., 2020), Scikit-Image (van der Walt et al., 2014), and Matplotlib (Hunter, 2007).

Neuronal segmentation.

A variation of the Singular Value Decomposition (SVD) and Independent Component Analysis (ICA) algorithm (Mukamel et al., 2009) was used to segment neurons from the calcium movies. First, each image was centered by subtracting the mean image and reshaped into a column vector. A matrix of these image columns, M, was then decomposed into a set of basis images U, singular values Σ, and standardized activity traces V according to

$$\text{M}\text{=}\text{U}\sum {\text{V}}^{\text{T}}.$$

The SVD basis images, U, are a set of orthonormal images and their corresponding singular values describe their contribution to the summed variance in the movie. We kept only the basis images of the largest 220 singular values; equating to 20% more images than the expected number of cells per recording session. Following Mukamel et al. 2009 (Mukamel et al., 2009), we discarded basis images with large singular values as these images contained uncorrected motion artifacts that inflated their contribution to the movie’s variance. The specific number of dropped basis images was determined per experiment by examining a skew plot of the singular values and ranged from 5–20 images. The remaining basis images and their associated singular values were ordered by singular value magnitude. Images with larger singular values contained cells of interest (Figure S2E) and images with small singular values represented noise. To segregate these, we computed the largest expected singular value of a random matrix of size M using the Marchenko-Pastur distribution. Basis images with singular values which failed to exceed 1.5 times this critical singular value were discarded.

The orthonormality of the curated basis images ensures strict spatial segregation of cells but also implies that some image components will be negative. To isolate positive cellular structures from the basis images, we applied the Fast ICA algorithm (Hyvärinen and Oja, 2000). This algorithm determines a demixing matrix W that when applied to the basis images, U, returns a set of independent components S according to

$$\text{S}=\text{WU}$$

where the skew of the pixel values in each column (image) of S is maximal. These independent components revealed clear spatial footprints for each cell (Figure S2F). Lastly, a multi-level thresholding (Yen et al., 1995) was applied to extract each cell’s boundary as a polygon (Figure S2G orange and green boundaries).

The SVD/ICA algorithm detects structures that contribute to the overall variance in the calcium movie including small structures such as dendrites and large structures such as out-of-focus neurons. The following geometrical constraints were employed to ensure the detected regions of interests (ROIs) reflected the expected size and shape of neurons. First, ROI’s that had a main axis diameter of less than 8 μm or greater than 32 μm were discarded. Second, we required that the eccentricity of the polygon regions, approximated from an ellipse with the same second moments as the region, not exceed 0.9. Third, the polygon regions were shrunk to include only the 70% brightest pixels for each detected region to compensate for smearing due to movement. Lastly, the ROI boundaries were drawn onto the source images, S, and were manually curated to ensure only clearly identifiable cells were used.

Signal Extraction.

The calcium traces were extracted using an Annular Background Correction (ABC) procedure for each ROI (neuron). First, an annulus polygon was automatically constructed around each ROI with an inner and outer diameter of 1.3 and 1.6 times the ROI’s diameter at each vertex (Figure S2G top; blue region). Since the annulus may overlap with nearby cells, we excluded these overlapping pixels from each annulus (Figure S2G top). Next, we computed the average pixel value within the ROI (<F>_ROI) and the average pixel value within the annulus surrounding the ROI (<F>_annulus) for each frame. The <F>_ROI contains a contribution from out-of-focus neurons and the neuropil’s response. A common method to correct this contamination is to subtract a fraction (η) of the annulus’ response from the ROI’s response; F_corrected (t) = <F>_ROI (t) − η*<F>_annulus (t). We chose η=0.7 from previous studies (Chen et al., 2013; Kerlin et al., 2010) and visually confirmed that the corrected signal no longer included annular contributions (Figure S2G bottom).

Cell alignment.

To validate the Inscopix motion correction algorithm and ensure that individual cells could be tracked within and across contexts of the CFC task, we applied a second “validating” motion correction algorithm (Guizar-Sicairos et al., 2008) to the motion corrected movies returned from Inscopix. This validation analysis revealed that the training context of the CFC task had large uncorrected motion artifacts (Figure S2B, D). For this reason, responses from the training context were not considered in this study. This motion analysis also revealed that greater than 98.3% of the recorded images across all mice used in this study have a displacement below 10 μm (Figure S2D insets). To specifically confirm cell tracking across days of the task, we computed the maximum intensity projection of source images independently measured on each day of the task (Figure S2H) and used our validating motion correction algorithm (Guizar-Sicairos et al., 2008) to measure the inter-day displacements for all mice (Figure S2J) and found no displacement exceeding 4 μm. As a final verification of the quality of the motion correction, we visually examined the positions of cells relative to the ROIs across contexts (Figure S2H–I).

Spike inference

Spike events were estimated using a first-order auto-regressive (AR1) model

$$y_t=c_t+{\epsilon }_t,{\epsilon }_t~N\left(0,{\sigma }^2\right),t=1\dots T$$

$$c_t=\phi c_{t-i}+s_t,t=p+1\dots T$$

where y_t is the observed fluorescence at time t, c_t is the unobserved calcium concentration, s_t is the number of spikes at time t and φ is the calcium concentration decay rate. By reinterpreting s_t as an indicator variable for the presence of a spike at time t, this leads to the following constrained objective function:

$$\underset{c_1\dots c_T,s_2-s_T}{\mathrm{minimize}}\left\{\frac{1}{2}\sum _{t=1}^T{\left(y_t-c_t\right)}^2+\lambda \sum _{t=2}^T{1}_{s_t\ne 1}\right\}\text{subject}\text{to}{s}_t=c_t-\phi c_{t-1}⩾0$$

where 1 is an indicator variable that equals 1 if at least one spike occurred at time t and 0 otherwise. Here, the tuning parameter λ controls the trade-off between the number of spikes and the calcium concentration fit. Importantly, this objective function can be exactly solved for the global optimum by recasting it as a changepoint detection problem (Jewell and Witten, 2018). The changepoints that solve the above objective function directly correspond with the spike event times. Thus, the model has two parameters that must be determined, namely the decay rate φ and the penalty λ. The decay parameter was chosen by fitting manually selected sections of the calcium traces that appear to have a single exponential decay. To determine the penalty parameter, we randomly chose 100 cells from across all mice used in this study and performed the two-fold cross validation procedure outlined in Jewell et al 2018 (Jewell and Witten, 2018). This optimal penalty was found to be cell-dependent (Figure S2K blue). Since this study is only concerned with relative spike event rates, we took a conservative approach to setting this penalty parameter. For the 100 randomly chosen cells, we measured the mean and variance of the optimal penalties across cells and applied a 3 standard error rule resulting in a global penalty of 0.01 (Figure S2K green). This conservative penalty was applied to all cells used in this study. To ensure this penalty does not over-count spike events, we constructed a histogram of event rates for all cells and confirmed that the vast majority of cells have event rates below previous electrophysiology studies (Hirase et al., 2001) (Figure S2L). We further confirmed this by visually inspecting traces with the optimal and global penalty and found many putative single spike events were detected with the per cell optimal penalty but missed with the global penalty (Figure S2M).

Transiently co-active cell ensembles

Transient ensembles were defined as a group of cells that temporarily exhibited simultaneous activity during a short (< 1 second) window of time. Since we recorded a large number of cells simultaneously (186 ±24 neurons), the pairwise correlation coefficient is a biased estimator that fails to account for chance correlations due to the large sample size (Quaglio et al., 2018). Thus, we used a time-shifted surrogate data approach (Grün et al., 2010). First, we convolved each event train with a Gaussian window of width 250 ms yielding an activity trace for each cell. Next, we averaged these activity traces to estimate the coordinated “Network” activity (Figure 3 A–B, Figure S3D, blue and orange traces). Given the large number of cells recorded some of the peaks in this network activity trace represent chance alignments of spike events. To determine if a peak in the network activity trace resulted from a significant number of correlated neurons, we randomly dithered each spike train up to 75 seconds and computed a surrogate network activity trace. Repeating this process 10 times and recording the maximum surrogate peak value for each trial provided a distribution of the largest peaks expected by chance. Applying a 1.5 standard error rule to this distribution yielded a threshold (Figure 3A–B, green line) for significant network peaks. In this study, we were interested in the largest amount of co-activation in the WT and RTT genotypes. We therefore took a 1 second window around the largest significant peak from each network activity trace and counted the number of cells active during this window as a fraction of the total recorded cells. This value is our estimate of the co-active cell percentages (Figure 3C). During the long-term recall, all but one of the WT mice and one of the RTT mice exhibited significant network peaks. These two mice could not be included in the coactive cell percentage (Figure 3C, Table S1).

High-degree cells

Persistently correlated cell pairs were defined as two cells whose activity traces had a Pearson’s correlation coefficient exceeding 0.3 (Cossell et al., 2015). Histograms of the number of correlated partners (Figure 3E) revealed that the majority of cells in each context had a low number of partners. As with the co-active cells, the pairwise correlation coefficient fails to account for chance correlations in our large sample of cells (186 ± 24) per experiment. To determine a threshold separating the cells with a low number of partners (presumably chance correlations) from cells with a significantly large number of partners (i.e. high-degree cells; right of dashed lines Figure 3E), we took a graph theory approach. First, for each context we took all the correlated pairs and constructed a graph where each node represents a cell and each edge a correlation exceeding 0.3. We then recorded the degree (i.e. number of partners) of each node in this graph. Second, we constructed a surrogate Erdős-Rényi random graph with the same number of nodes (cells) and edges (correlations) and recorded the largest degree (number of partners) for any single node in this random graph. This value represents the number of partners that would be expected by chance for this number of cells (nodes) and correlations (edges). Repeating this process 100 times and applying a 1.5 standard error rule to the recorded degrees provided a threshold for the degrees in the original graph. This procedure yields cells with higher numbers of partners than expected by chance for each experiment and context. These cells were termed “high-degree cells” and constituted the high-degree ensemble (Figure 3E, F–H) specific to each experiment and context.

Data analysis and statistics

All data analysis was carried out in Python. No a priori power analysis of the number of experiments needed to reach a given confidence level was conducted. Normality test revealed that much of the data in this study is not normally distributed. Therefore, we used only non-parameteric two-tailed statistical tests. For unpaired 2-sample tests we used the Mann-Whitney U test. For paired 2-sample tests we used the Wilcoxon-signed rank tests. For multiple comparisons we used the Kruskal-Wallis omnibus test followed by post-hoc comparisons between sample groups. For categorical connectivity (Figure 4J, L) we used a Binomial test. When the number of samples was large (Figure 3I) we performed hypothesis testing with bootstrapped confidence intervals of the median differences using 10,000 resamples of the data to avoid spuriously low p-values due to large sample sizes. The largest number of independent simultaneous test conducted was four and no correction for multiple comparisons was applied. Tables were consulted for p-values when the number of samples was less than 20 and otherwise were assumed asymptotically normal. Data are presented with the median and interquartile range (IQR) on a boxplot with whiskers extending to 1.5 * IQR unless otherwise noted in the text. Full statistical tests and values of main figures are presented in Table S1.

KEY RESOURCES TABLE.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
somatostatin	Millipore	Cat# MAB354, RRID:AB_2255365
MeCP2	Cell Signaling Technology	Cat# 3456, RRID:AB_2143849
CamKII	Abcam	Cat# ab22609, RRID:AB_447192
Alexa Fluor 488 antibody against rabbit IgG	Thermo Fisher Scientific	Cat# A-11008, RRID:AB_143165
Alexa Fluor 647 antibody against rat IgG	Thermo Fisher Scientific	Cat# A-21247, RRID:AB_141778
Alexa Fluor 568 conjugated streptavidin	Thermo Fisher Scientific	Cat# S-11226, RRID:AB_2315774
Drugs and Virus
Clozapine N-oxide	Tocris	4936
Picrotoxin	Tocris	1128
CNQX	Tocris	0190
D-AP5	Tocris	366877-50mg
pENN.AAV.CamKII.GCaMP6f.WPR E.SV40	Addgene	100834-AAV1, RRID:Addgene_100834
pAAV-hSyn-DIO-hM3D(Gq)-mCherry	Addgene	44361-AAV1, RRID:Addgene_44361
pAAV-hSyn-DIO-hM4D(Gi)-mCherry	Addgene	44362-AAV1, RRID:Addgene_44362
pAAV-hSyn-DIO-mCherry	Addgene	50459-AAV1, RRID:Addgene_50459
AAV1.CAG.LSL.tdTomato.bGH	Penn Vector Core	AV-1-ALL856
Experimental Models
SOM-IRES-cre	Jackson Laboratory	Jax stock: 13044
ChAT-ChR2-EYFP	Jackson Laboratory	Jax stock: 14546
Software and algorithms
Python 3.7	Python	https://www.python.org/, RRID:SCR_008394
Numpy	Harris et al., 2020	https://numpy.org/, RRID:SCR_008633
Scikit-Image	Van der Walt et al., 2014	https://scikit-image.org/, RRID:SCR_021142
Matplotlib	Hunter, 2007	https://matplotlib.org/, RRID:SCR_008624
FIJI (ImageJ)	NIH	https://fiji.sc, RRID:SCR_002285
Adobe Illustrator	Adobe	http://www.adobe.com/products/illustrator.html, RRID:SCR_010279
Minianalysis 6.0.3	Synaptos oft Inc.	http://www.synaptosoft.com/MiniAnalysis/, RRID:SCR_002184
Mosaic	Inscopix	https://www.inscopix.com/
Deposited Data
Zenodo database		Zenodo: https://doi.org/10.5281/zenodo.5999292, RRID:SCR_004129
Github repository		Github: https://github.com/mscaudill/rett_memory, RRID:SCR_002630