Abstract
The physical manifestations of memory formation and recall are fundamental questions that remain unresolved1. At the cellular level, ensembles of neurons called engrams are activated by learning events and control memory recall1-5. Astrocytes are in close proximity to neurons and engage in a range of activities that support neurotransmission and circuit plasticity6-10. Moreover, astrocytes exhibit experience-dependent plasticity11-13; however whether specific ensembles of astrocytes participate in memory recall remains obscure. Here we show that learning events induce c-Fos expression in a subset of hippocampal astrocytes, which subsequently regulates hippocampal circuit function. Intersectional, c-Fos based labeling of astrocyte ensembles after learning events reveals that they are closely affiliated with engram neurons, while re-activation of these astrocyte ensembles stimulates memory recall. At the molecular level, learning-associated astrocyte ensembles exhibit elevated expression of NFIA and its selective deletion from this population suppresses memory recall. Together, our studies identify learning-associated astrocyte ensembles as a new form of plasticity that is sufficient to provoke memory recall, while implicating astrocytes as an active component of the engram.
Ensembles of neurons activated during a learning event, whose activity is necessary and sufficient for subsequent memory recall, form the cellular basis of memory, and are referred to as engrams2-5. A hallmark feature of engram neurons is the induction of immediate early genes (IEGs), which mediate synaptic plasticity, a critical aspect of memory formation14,15. Accordingly, IEG-based tools have been used to tag and manipulate engram neurons – and the memories they represent – across a host of brain regions2-5,16. Astrocytes are closely affiliated with neurons and, as part of the tripartite synapse, play essential roles in regulating circuit function, including those circuits associated with learning and memory9,17-20. Moreover, astrocytes exhibit experience-dependent plasticity, where their activation states, transcriptional responses, and functional properties are tuned to environmental stimuli and internal states11,12,21,22. Astrocytes are in close proximity to neurons, respond to a range of experiential modalities, and regulate learning behaviors, raising the possibility that they actively participate in memory formation and retrieval in coordination with neuronal ensembles. Nevertheless, whether subsets of astrocytes respond to memory-inducing stimuli and form durable ensembles that are necessary and sufficient for memory storage and recall is unknown.
Learning-dependent astrocyte activation
The function of hippocampal circuits and the memory-related behaviors they subserve depends on interactions with astrocytes23-25. We previously found that olfactory bulb astrocytes undergo transcriptional and functional changes in response to neuronal activity in order to facilitate the processing of odor11. To examine how hippocampal astrocytes respond to neuronal activity we chemogenetically activated dentate gyrus (DG) neurons by microinjecting AAV-hSyn-hM3D-mCherry into the DG and, several weeks later, injecting mice with the hM3D agonist clozapine N-oxide (CNO, 0.3 mg/kg). This stimulation induced expression of the IEG c-Fos in a subset of astrocytes in both DG and CA1 (Extended Data Fig. 1a-d). We also found that selectively activating excitatory dentate gyrus neurons using AAV-CaMKII-hM3D-mCherry induced c-Fos expression in astrocytes (Extended Data Fig. 1e-g). Induction of astrocytic c-Fos upon neuronal activation was preserved when inputs to the hippocampus were silenced using retrograde AAV-hSyn-hM4D-mCherry, indicating that the activity of local, excitatory neurons is sufficient to drive c-Fos expression in astrocytes (Extended Data Fig. 1e-f). Next, we used a contextual fear conditioning paradigm to elicit natural activity within the hippocampal circuit and similarly found an activity-dependent increase in c-Fos-expressing astrocytes in CA1 and DG that peaked 1.5 hours after the learning event (Fig. 1a-d). These findings indicate that hippocampal neuronal activity activates a subset of astrocytes.
Figure 1. Experience-dependent induction of astrocytic c-Fos regulates memory.
a. Experimental timeline and representative immunostaining of c-Fos expression in a subset of hippocampal astrocytes after fear learning. b. Quantification of c-Fos+ astrocytes following fear conditioning (n = 4 mice per group, one-way ANOVA and Dunnett’s post-hoc tests, *P = 0.023, ****P < 0.0001). c. Schematics of AAV constructs, timeline, and representative images for stimulating Gq-coupled signaling in astrocytes in control (hM3D) and Fos cKO (hM3D+Cre) mice. d. Quantification of c-Fos+ astrocytes following CNO injection in hM3D vs. hM3D+Cre (Fos cKO) mice (n = 3 mice per group, two-tailed t test, ****P < 0.0001). e. Schematic and LTP traces in control (GFAP-mCherry) and Fos cKO (GFAP-Cre-4x6T) mice. Plot in upper right shows mean fEPSPs from the last five minutes of recordings (n = 8 control, n = 7 Fos cKO, two-tailed t test, **P = 0.005). f. Behavioral experiment timeline comparing control and Fos cKO mice. g. Quantification of c-Fos expression in hippocampal astrocytes from control (GFAP-mCherry) and Fos cKO mice (n = 4 control, n = 3 Fos cKO, Welch’s t test, **P = 0.002). h. Left panel: freezing behavior in Context A during the two minutes before foot shocks (n = 8 mice per group, two-tailed t test, P = 0.503). Middle panel: freezing behavior during the recall test in Context A (n = 8 mice per group, two-tailed t test, **P = 0.006). Right panel: Place discrimination in the novel place recognition task (n = 8 mice per group, two-tailed t test, *P = 0.046; horizontal line indicates median, box indicates quartiles, whiskers indicate minimum and maximum values). Data are mean±SEM except where otherwise noted. Panel c, e, f were created using Biorender.com
Neuronal expression of c-Fos is activity-dependent and is critical for synaptic plasticity and memory-dependent behaviors14,15,26, however the role of astrocytic c-Fos has not been studied in this context. To examine how experience-dependent c-Fos induction in astrocytes influences circuit plasticity in the hippocampus, we selectively knocked out Fos in hippocampal astrocytes using Fos-flox mice microinjected with GFAP-Cre-4x6T AAVs, which contain a microRNA targeting cassette to eliminate off-target transgene expression in neurons (Extended Data Fig. 2)27. We confirmed the efficiency of this knockout strategy by artificially inducing astrocytic c-Fos expression through chemogenetic activation of the Gq pathway in astrocytes24 (Fig. 1c-d), and confirmed Cre expression in all hippocampal subfields (Extended Data Fig. 3). Astrocyte-specific Fos deletion (Fos cKO) diminished long-term potentiation (LTP) in hippocampal slices, suggesting that astrocyte activation is critical for synaptic plasticity and learning-dependent behaviors (Fig. 1e). Accordingly, Fos cKO mice also exhibited impaired performance in contextual fear conditioning and novel place recognition tasks (Fig. 1f-h, Extended Data Fig. 3). These behavioral deficits were coupled with mild reductions in astrocytic microdomain calcium activity and spontaneous neuronal activity in Fos cKO mice (Extended Data Fig. 4). Together, these findings demonstrate that activation of a subset of hippocampal astrocytes during learning is critical for circuit plasticity and memory recall.
Learning-associated astrocyte ensembles
Our findings indicate that only a subset of astrocytes express c-Fos in response to hippocampal neuronal activity. These observations led us to hypothesize that an ensemble of astrocytes may selectivity participate in the consolidation, storage, and retrieval of the memory associated with the learning event in which they were activated. To study these ensembles we developed a strategy that enables genetic access to learning-associated astrocytes. We generated an Aldh1l1-CreER; Rosa-CAG-FSF-tdTomato reporter mouse line (FSF = FRT-stop-FRT), which allows for astrocyte-specific and tamoxifen-inducible Cre-mediated recombination in conjunction with Flp-dependent expression of tdTomato (Fig. 2a). Next, we created a Fos-Flex-Flp AAV, which drives Flp expression in Fos-expressing (activated) cells and when used with the reporter line labels Fos-expressing astrocytes with tdTomato (Fig. 2a). Using this strategy, we injected the hippocampus of the reporter line with Fos-Flex-FLP AAV, treated with tamoxifen for three consecutive days, after which we subjected these mice to contextual fear conditioning or contextual exposure only, and homecage controls (Fig. 2a). We observed a significant increase in the number of tdTomato-labeled cells that express GFAP throughout the hippocampus in all context exposure groups relative to homecage controls, indicating successful labeling of astrocytes in an experience-dependent manner (Fig. 2b-c) (Extended Data Fig. 5). We validated that all tdTomato+ cells examined express GFAP (Extended Data Fig. 5), while confirming that tamoxifen administration itself did not affect astrocytic expression of c-Fos (Extended Data Fig. 6). It is important to note that using Fos or any other single genetic entry point may underestimate the population of astrocytes engaged by learning28. Another caveat of this labeling system is that the amount of labeling is dependent on the interval between tamoxifen injection and perfusion.
Figure 2. Genetic access to learning-associated astrocyte ensembles.
a. Schematic of genetic system and timeline for labeling learning-associated astrocytes using AAV-Fos-Flex-Flp in Aldh1l1-CreER; Rosa-CAG-FSF-tdTomato mice. b. Representative images of tdTomato+ hippocampal astrocytes in homecage and fear conditioned mice. c. Quantification of labeled astrocytes (n = 5 mice per group, Welch’s corrected t test, *P = 0.023, two-tailed t test, ***P = 0.0002). d. Schematic of genetic system and timeline for examining Ca2+ dynamics. e. Representative two photon images and GCaMP signal traces from fear conditioning tagged and home cage tagged astrocytes. Each row in the heatmaps represents a single astrocyte. f. Quantification of Ca2+ signaling parameters (n = 14 cells per group, n = 3 mice per group, two-tailed t tests, **P = 0.004 (amplitude), 0.002 (AUC), 0.008 (spike width). Data are mean±SEM except where otherwise noted. Panel a, d were created using Biorender.com
Memory consolidation involves changes to ensembles of cells associated with a memory trace1,29. Astrocytes display complex intracellular calcium signals that reflect interactions with neurons and can influence associated circuit function23,30,31. To examine whether learning-associated astrocytes, herein termed LAAs, display alterations in Ca2+ activity, we co-injected Fos-Flex-Flp and GFAP-FSF-GCaMP6 (FSF = FRT-stop-FRT) AAVs into the hippocampus of Aldh1l1-CreER mice (Fig. 2d). We injected mice with tamoxifen, after which they were left in the homecage or fear conditioned. After allowing 3 days for sufficient expression of GCaMP6, we recorded spontaneous Ca2+ activity in labeled astrocytes with 2-photon imaging of acute slices (Fig. 2d-f). We observed that LAAs from fear conditioned mice displayed an overall increase in Ca2+ activity within the soma and main branches, including increased amplitude and duration of events. Ca2+ activity was also increased within microdomains of LAAs (Extended Data Fig. 5c). Notably, calcium activity within Fos cKO astrocytes was modestly reduced (Extended Data Fig. 4). Together, these findings reveal elevated Ca2+ activity as a core feature of astrocyte ensembles activated by learning and may reflect changes in neuron-astrocyte interactions during memory consolidation.
LAAs and engram neurons interact
We next sought to examine the interactions between learning-associated astrocytes and engram neurons associated with a single learning event. To achieve this, we developed a system that employs a tamoxifen-inducible Flp whose expression is driven by the Fos promoter (Fos-FlpER; Fig. 3a, Extended Data Fig. 7). Co-injecting Fos-FlpER with GFAP-fDIO-tdTomato-4x6T (fDIO = FRT double-flanked inverted open reading frame) and hSyn-fDIO-EYFP allows us to simultaneously label and map engram neurons and LAAs in wildtype mice after fear conditioning by injection of 4-hydroxytamoxifen (Fig. 3b). Importantly, this tagging system labels a proportion of astrocytes that is in line with the number that express c-Fos after fear conditioning (Extended Data Fig. 8). Moreover, astrocytes labeled during fear conditioning re-express c-Fos at rates over 25-fold greater than unlabeled astrocytes after a recall test, which indicates that our labeling of LAAs is highly specific to fear memory (Extended Data Fig. 8).
Figure 3. Learning-associated astrocytes interact with engram neurons.
a. Schematic for labeling LAAs and engram neurons. b. Image of EYFP+ engram neuron and tdTomato+ LAA. c. Immunostaining of engram neuron within territories of LAA and non-LAA. d. Quantification of engram neuron coverage of astrocyte territories (upper panel) and astrocyte territory size (lower panel). n = 186 tdTomato− astrocytes, n = 38 tdTomato+ astrocytes, n = 8 mice, two-sided nested t tests, ****P < 0.0001 for upper panel, P = 0.657 for lower panel. Middle line indicates median, upper and lower lines indicate quartiles. e. Schematic for labeling LAAs and CA3-CA1 engram-engram synapses. f. Image of eGRASP engram synapses within territories of LAA and non-LAA. g. Quantification of eGRASP puncta within astrocyte territories (upper panel) and astrocyte territory size (lower panel). n = 276 tdTomato− astrocytes, n = 25 tdTomato+ astrocytes, n = 4 mice, two-sided nested t tests, ****P < 0.0001 for upper panel, P = 0.295 for lower panel. Middle line indicates median, upper and lower lines indicate quartiles. h. Schematic for expressing hM3D-mCherry in LAAs and EYFP in engram neurons. i. Diagram of recording from engram and non-engram neurons; representative traces. j. Quantification of EPSCs before and after CNO treatment in engram and non-engram neurons. n = 8 engram neurons from 8 mice, n = 7 non-engram neurons from 7 mice. Tukey test, **P = 0.009 for upper panel. One-way ANOVA, P = 0.622 for lower panel. k. Timeline for examining engram neuron c-Fos expression. l. Immunostaining of c-Fos in EYFP+ engram neurons. Arrow denotes c-Fos+ EYFP+ cell. Arrowhead denotes c-Fos− EYFP+ cell. m. Quantification of engram neuron activity (c-Fos+/EYFP+). n = 10 saline, n = 7 CNO. Two-tailed t test, ****P < 0.0001. Data are mean±SEM except where otherwise noted. Panel e, i were created using Biorender.com
We used the Fos-FlpER system to label engram neurons and LAAs active during fear conditioning. We examined the proximity of engram neurons with LAAs and non-LAAs, using GFAP to estimate the territory of each astrocyte. Quantification of the volume of EYFP+ processes of engram neurons within individual astrocyte territories revealed that the processes of engram neurons were preferentially localized within the domains of tdTomato+ LAAs (Fig. 3c-d), with every LAA territory containing EYFP+ processes. We observed no differences in the territory size between LAAs and non-LAAs. These observations suggest a close association between engram neurons and the ensemble of LAAs.
Prior studies showed that fear learning increases the number of synapses between engram neurons without changing the number of synapses with non-engram neurons32. This suggests that engram-engram synapses are a structural substrate of memory; in other words, a “synaptic engram.” To investigate whether LAAs preferentially interact with synaptic engrams, we selectively labeled engram-engram synapses using Cre-dependent eGRASP (enhanced GFP reconstitution across synaptic partners)32 constructs. eGRASP uses presynaptic and postsynaptic targeted components of a split GFP to generate fluorescent puncta at synapses where both components are present. We targeted the pre-synaptic eGRASP component to left CA3 engram neurons by co-injection of Fos-CreER and Ef1α-DIO-Pre-eGRASP AAVs (Fig. 3e). In right CA1, we co-injected Fos-CreER and Ef1α-DIO-Post-eGRASP AAVs, to enable selective expression of the postsynaptic eGRASP component in CA1 engram neurons, along with Fos-FlpER and GFAP-fDIO-tdTomato-4x6T to label LAAs. We then fear conditioned mice and injected them with 4-OHT to label cells activated during learning. We examined right CA1 in which eGRASP+ puncta marking sparse engram-engram synapses were evident. We quantified the number of eGRASP puncta within the territory of individual astrocytes as defined by GFAP. tdTomato+ astrocyte territories contained significantly more eGRASP puncta, indicating that the synaptic engram is enriched within the territories of LAAs (Fig. 3f-g). These observations suggest that LAAs are well positioned to influence the activity of engram neurons. Furthermore, these findings imply that memory recall may involve the coordinated interactions between unique ensembles of astrocytes and engram neurons.
LAAs regulate neuronal engram activity
The identification of LAA ensembles that are juxtaposed with engram neurons prompted us to investigate the role of these astrocyte ensembles in circuit plasticity. Previous studies showed that activating CA1 astrocytes via the Gq pathway is sufficient to induce long-term potentiation (LTP) in the CA3-CA1 circuit of the hippocampus24. To test whether activation of the relatively small ensemble of LAAs was sufficient to induce LTP, we expressed hM3D in hippocampal astrocytes tagged during fear conditioning using Aldh1l1-CreER mice injected with Fos-Flex-Flp and GFAP-FSF-hM3D-mCherry (Extended Data Fig. 9a-b). Using slice recordings and a subthreshold stimulus (10 stimuli at 40 Hz)33, we found that homecage controls and LAA groups treated with saline did not induce LTP. However, when LAA groups were treated with CNO to activate hM3D we observed a robust induction of LTP that was comparable to that observed with pan-astrocyte hM3D activation (i.e., CNO treated, non-fear conditioned, hM3D expressed in all astrocytes) (Extended Data Fig. 9c-d). Moreover, activating hM3D signaling in LAAs in the absence of Schaffer collateral stimulation induced a mild potentiation (Extended Data Fig. 9e). Together, this data suggests that the ensemble of astrocytes activated by a single learning event are capable of mediating plasticity within hippocampal circuits.
We next sought to examine whether LAA ensembles could affect the function of engram neurons. Using our activity-dependent labeling approach, we directed expression of hM3D to LAAs and EYFP to engram neurons active during fear conditioning (Fig. 3h). One week later, we recorded excitatory postsynaptic currents (EPSCs) from engram (EYFP+) and non-engram (EYFP−) neurons in acute slices. We obtained recordings before and after application of CNO to activate hM3D in LAAs. Activating LAAs selectively increased the frequency of EPSCs in engram neurons without affecting EPSC amplitude (Fig. 3i-j). This observation demonstrates that LAAs can regulate the synaptic activity of engram neurons.
Next, we tested whether activating LAAs increases fear engram activation at the molecular level. We expressed EYFP in engram neurons and hM3D in LAAs active during fear conditioning in a novel context (i.e. Context A). One week later, we injected mice with either saline or CNO (3 mg/kg, 30 minutes prior to testing), to activate hM3D-expressing astrocytes, and placed them into a distinct and innocuous context (i.e. Context B). 90 minutes later, we collected brains and stained for c-Fos to determine the activation of engram neurons within Context B (Fig 3k). CNO-injected mice showed an increase in the reactivation rate of fear-tagged engram neurons (c-Fos+/EYFP+) (Fig. 3l-m, Extended Data Fig. 10a-c). Collectively, these observations suggest that LAA ensembles can regulate hippocampal circuit function by modulating the activity of engram neurons.
LAA reactivation elicits recall
Our findings that LAA ensembles regulate hippocampal circuit plasticity and engram neuron activity raise the possibility that they may also control the recall of memories specific to the learning event in which they were activated. Therefore, we examined whether selectively reactivating LAAs tagged during fear conditioning could cause expression of adaptive behaviors (i.e., freezing) that mice engage in when experiencing fear. To achieve this, we labeled astrocyte ensembles with hM3D during fear conditioning in a novel context (i.e., Context A) using Aldh1l1-CreER mice injected with Fos-Flex-Flp and GFAP-FSF-hM3D-mCherry (Fig. 4a). This configuration allowed us to introduce hM3D into LAAs, enabling their subsequent reactivation. We found that reactivating LAAs by injection of CNO (3 mg/kg, 30 minutes prior to testing) 3 or 7 days after fear conditioning resulted in a significant increase in freezing behavior after mice were placed in a distinct context (i.e., Context B) (Fig. 4b-d). In addition, using the AAV-Fos-FlpER system, we observed a similar induction of freezing behavior by reactivating LAAs 3 weeks after fear conditioning, which demonstrates that the phenomenon of astrocyte ensemble-mediated memory recall is durable (Extended Data Fig. 11). We confirmed that the expression of hM3D was specific to astrocytes so as to exclude the possibility that increased freezing behavior could be attributed to off-target expression of hM3D in neurons (Fig. 4g-h). Moreover, to confirm that hippocampal astrocyte activation does not non-specifically cause freezing behavior, we bulk labeled astrocytes with AAV-GFAP-hM3D-mCherry (or GFAP-mCherry control; Fig. 4e). Three days after fear conditioning in Context A, we placed mice into Context B following injections of saline or CNO (3 mg/kg). We found no difference in freezing across groups (Fig. 4f), which demonstrates that stochastic astrocyte activation does not cause freezing. Together, these results demonstrate that selective reactivation of hippocampal astrocyte ensembles that were previously activated during fear learning is sufficient to elicit memory recall.
Figure 4. Reactivation of learning-associated astrocytes elicits memory recall.
a. Schematic of genetic system and timeline for tagging astrocyte ensembles active during fear learning with hM3D and subsequently reactivating them in a distinct context. b. Freezing behavior in Context B after reactivation of learning-associated astrocytes at 3 days after fear conditioning (n = 9 mCherry+saline, 5 mCherry+CNO, 7 hM3D+saline, 7 hM3D+CNO mice, one-way ANOVA and Dunnett’s post-hoc tests, **P ≤ 0.001). c. Freezing behavior in Context B after reactivation of learning-associated astrocytes at 7 days after fear conditioning (n = 8 mice injected with Fos-Flex-Flp and GFAP-FSF-hM3D-mCherry per group, two-tailed t test, **P = 0.002). d. Quantification (top) and representative images (bottom) of c-Fos expression in Sox9+ astrocytes after CNO injection in Aldh1l1-CreER+ and Aldh1l1-CreER− mice (n = 4 mice per group, two-sided Welch’s corrected t tests, *P = 0.033, **P = 0.003). e. Timeline and schematic of genetic system for activation of random astrocytes during recall. f. Freezing behavior in Context B after random astrocyte activation (n = 4 mCherry+saline, 5 mCherry+CNO, 4 hM3D+saline, 6 hM3D+CNO mice, one-way ANOVA, P = 0.974). g. Quantification (top, n = 4 mice) and representative images (bottom) of hM3D-mCherry expression in GFAP+ astrocytes. h. Quantification (top, n = 4 mice) and representative images (bottom) showing absence of hM3D-mCherry expression in neurons. Data are mean±SEM. Panel a, b, e, f were created using Biorender.com
LAA NFIA is essential for recall
The foregoing observations suggest that LAAs exhibit distinct functional properties during experience-dependent plasticity after learning events. To examine the molecular properties of LAAs, we employed the Fos-Flex-Flp AAV system and a dual reporter mouse line (Aldh1l1-CreER; Aldh1l1-GFP; Rosa-CAG-FSF-tdTomato), in which all astrocytes express GFP, and Flp-expressing astrocytes additionally express tdTomato (Fig. 5a). With this system, we were able to label LAAs (GFP+, tdTomato+) and non-LAAs (GFP+, tdTomato−). After tamoxifen treatment and fear conditioning, we used FACS to purify the respective populations and performed transcriptomic RNA-sequencing (Fig. 5a-d, Extended Data Fig. 12). Analysis of the differentially expressed genes (DEGs) revealed gene ontologies associated with synapse-related genes in LAAs (Fig. 5b-d; Supplementary Table 1). Subsequent HOMER transcription factor motif analysis of the DEGs revealed that the NFIA DNA binding motif was enriched among genes upregulated in LAAs (Fig. 5c). We further validated that NFIA mRNA and protein expression were elevated in LAAs (Extended Data Fig. 13).
Figure 5. Ensemble-specific NFIA is necessary for context-specific memory.
a. Schematic for labeling and sorting GFP+ tdTomato+ LAAs and GFP+ tdTomato− non-LAAs for RNA-seq. b. Volcano plot depicting DEGs in LAAs vs non-LAAs. Uncorrected P values from DEseq2. N = 3 biological replicates per group (hippocampi from 3 mice pooled per replicate). c. HOMER motif analysis of transcription factor motifs enriched among upregulated (left) and downregulated (right) genes. Cumulative binomial distribution test from HOMER. d. Gene Ontology (GO) analysis of upregulated (left) and downregulated (right) genes. Two-sided Fisher exact test. e-f. Timeline and schematic for knockout of NFIA in LAAs. g. Representative immunostaining (left) and quantification (right) of NFIA expression in tdTomato-labeled (control) and Cre-HA labeled (NFIA cKO) astrocytes (n = 4 mice per group, two-tailed t test, ****P < 0.0001). h. Freezing during recall test in Context A (n = 10 mice per group, two-tailed t test, **P = 0.009). i. Quantification of novel place discrimination (n = 9 tdTomato, n = 10 NFIA cKO, two-tailed t test, P = 0.960). j-k. Timeline and schematic for knockout of NFIA in LAAs and hM3D expression in engram neurons. l. Freezing prior to shocks in Context A (n = 10 mice per group, one-way ANOVA, P = 0.152). m. Freezing in Context B 30 minutes after CNO injections (n = 10 mice per group, one-way ANOVA (P = 0.002) and Fisher’s post hoc tests, **P ≤ 0.005). n. Representative immunostaining of c-Fos expression in hM3D-mCherry+ engram neurons. o. Quantification of c-Fos expression in CNO activated hM3D-mCherry+ engram neurons (n = 4 mice per group, two-tailed t test, ****P < 0.0001). p. Quantification of mCherry co-labeling with NeuN or GFAP (n = 4 mice). Data are mean±SEM. Panel a, e, j were created using Biorender.com
Prior studies have shown that astrocytic NFIA regulates hippocampal circuit function23, however, it is unknown whether it participates in the process of memory consolidation and recall. Moreover, whether NFIA facilitates circuit function through ensemble-specific actions is unknown. To examine the role of NFIA in astrocyte ensembles after fear learning, we developed a system that enables selective deletion of NFIA in LAAs during fear conditioning via injection of 4-hydroxytamoxifen (Fig. 5e-f). By co-injecting AAV-Fos-FlpER with GFAP-fDIO-Cre-HA-4x6T into NFIA-flox mice, we were able to induce deletion of NFIA in LAAs in the hippocampus after fear conditioning (Fig. 5f-g). To assess whether loss of NFIA from LAAs affected memory recall, we tested fear recall in the conditioned context (i.e., Context A) and found reduced freezing behavior in NFIA cKO mice compared to controls (Fig. 5h). Similarly, selectively deleting c-Fos in LAAs impaired fear recall (Extended Data Fig. 14). We next evaluated whether selective deletion of NFIA from the LAA ensemble specifically impaired fear memory, or if other hippocampus-dependent behaviors were also affected. We tested the same NFIA cKO mice on the novel place recognition task and found that they performed similarly to control mice (Fig. 5i). These findings demonstrate that memory impairment due to deletion of NFIA in LAAs is specific to memory recall associated with the original fear learning event, whereas memory for other events remained intact. Therefore, these data suggest that a distinct ensemble of astrocytes selectively coordinates the consolidation and recall of a distinct memory.
Next, we sought to determine whether fear memory recall could be rescued in NFIA cKO mice by artificially reactivating the neuronal engram. To achieve this, we co-injected AAV-Fos-FlpER and hSyn-fDIO-hM3D-mCherry with GFAP-fDIO-Cre-HA-4x6T (or GFAP-fDIO-TdTomato-4x6T) into NFIA-Flox mice (Fig. 5j-k). Four weeks after AAV injection mice were subjected to fear conditioning and 4-hydroxytamoxifen treatment to label engram neurons with hM3D and delete NFIA from LAAs. We observed no differences in pre-shock freezing between groups (Fig. 5l). After an additional three weeks, we chemogenetically reactivated engram neurons by treating mice with CNO (3 mg/kg) 30 minutes prior to exposure to an innocuous context (i.e. Context B). Critically, hM3D-mediated reactivation of engram neurons increased freezing behavior regardless of selective NFIA deletion in LAAs (Fig. 5m). We validated that CNO treatment selectively activated hM3D+ engram neurons and that hM3D expression was restricted to neurons (Fig. 5n-p). These findings indicate that activation of engram neurons can restore memory recall in mice with NFIA deficient LAAs and suggest that recall initiated by LAA reactivation is reliant upon neuronal engram function.
Discussion
Ensembles of neurons activated by learning events control memory recall, which led us to examine whether other brain cell types also participate in memory formation, storage, and recall. We found that learning events activate subsets of astrocytes in the hippocampus and that these learning-associated astrocyte ensembles regulate memory recall, suggesting that astrocytes themselves participate in the physical manifestation and expression of memories. These findings expand upon the notion that astrocytes exhibit experience-dependent plasticity, where astrocyte function is tuned to sensory or social experiences11,12,34, by illustrating new roles in memory consolidation and recall. Recent studies highlighted functionally diverse subpopulations of astrocytes17,19,35-37, which raises the possibility that LAAs are derived from specialized subsets of astrocytes or, alternatively, represent a generalized adaptation reflecting proximity to neuronal engrams. Dissecting the ontogeny of LAAs and their interrelationships with neuronal engram ensembles will shed important light on this new mechanism underlying memory recall. In much the same way that distinct ensembles of neurons in a variety of brain regions are activated in response to a host of experiences4,38, it is likely that complementary ensembles of astrocytes are similarly activated after experience and contribute to the associated behavioral adaptations. Accordingly, the tools and methods developed in our study can be widely applied to astrocytes across a host of experiences and brain regions.
IEGs serve as the molecular entry point for labeling neural substrates responsible for memory recall2-5, including LAAs. Leveraging these findings, we discovered a new role for astrocytic Fos in the regulation of hippocampal circuits and learning behaviors. These observations, coupled with prior findings describing experience dependent Sox9 transcriptional plasticity in olfactory bulb astrocytes11,39, illustrate adaptive transcriptional responses in astrocytes that parallel IEG activation during neuronal plasticity40. At the molecular level, LAAs exhibit unique transcriptomes, highlighted by the upregulation of NFIA and enrichment of its prospective target genes. Astrocytic NFIA plays a critical role in hippocampal circuit function23 and consistent with this, its deletion from LAAs inhibited memory recall in a context-specific manner. Astrocytes exhibit region-specific transcriptional dependencies21,36,41,42 and our observations in the hippocampus with NFIA, along with Sox9 in the olfactory system11,23,39, suggest that transcriptional plasticity in astrocytes is also mediated by these regional dependencies. Critically, future studies will be aimed at relating these transcriptional responses to the biochemical quanta used by astrocytes as memory substrates. Collectively, our studies identify LAAs as key components of the adaptive response to learning experiences, which regulate the flow of information during circuit plasticity and memory recall.
Methods
Animals
Animals were used in compliance with protocols approved by the Institutional Animal Care and Use Committee at Baylor College of Medicine. Mice were housed with food and water available ad libitum under a 12h-12 h light–dark cycle in a 20–22°C and 40–60% humidity environment. The following transgenic strains were used: Fos-flox (037115-JAX), Rosa-CAG-LSL-tdTomato (Ai14; Jackson Laboratory, 007914), Rosa-CAG-FSF-tdTomato (Ai65F; Jackson Laboratory, 032864), Aldh1l1-CreER (Jackson Laboratory, 029655), Aldh1l1-GFP (RRID:MMRRC_011015-UCD), and NFIA-flox43. All mice were on a C57/b6 background. Transgenic strains were crossed as described in the Results. We used 2-6 month old mice of both sexes and randomly allocated littermates to groups. Experiments and analyses were conducted blinded to group allocation.
To induce recombination with Fos-Flex-Flp constructs, mice were injected i.p. for three consecutive days with 100 mg/kg Tamoxifen (Sigma, T5648) in corn oil. To induce recombination with Fos-FlpER constructs, mice were injected i.p. with 50 mg/kg 4-hydroxytamoxifen (4-OHT; Sigma, H6278) dissolved in 9:1 corn oil:ethanol. Two 4-OHT i.p. injections were given 4 hours apart, with the second dose delivered immediately following fear conditioning for mice subjected to the behavioral paradigm. For stimulation of Gq-coupled signaling via hM3D, mice were injected i.p. with Clozapine N-oxide (CNO; Abcam, ab141704) dissolved in saline (doses are indicated for each experiment in the text). CNO was administered 30 minutes prior to behavioral testing and 90 minutes prior to perfusion.
Plasmid and virus production
The following plasmids were used for AAV packaging or creation of new constructs: GFAP-Cre-4x6T (a gift from S. Thomas Carmichael; Addgene plasmid #196410), GFAP-hM3D-mCherry (a gift from Bryan Roth; Addgene plasmid #50478), hSyn-hM3D(Gq)-mCherry (a gift from Bryan Roth; Addgene plasmid #50474), hSyn-fDIO-EYFP (a gift from Ulrik Gether; Addgene plasmid #154870), gfaABC1D-cyto-GCaMP6f (a gift from Baljit Khakh; Addgene plasmid #52925), Villinpromoter-blue-FlpOERT2 (a gift from Jannik Elverløv-Jakobsen; Addgene plasmid #67278), pOTTC475-pAAV-c-fos-iRFP (a gift from Brandon Harvey; Addgene plasmid #47906), EF1a-fDIO-Cre-HA (a gift from Esteban Engel & Alexander Nectow; Addgene plasmid #121675), Fos-CreER (a gift from Jun-Hyeong Cho, Addgene plasmid #194643) pAAV-EWB-DIO-cyan pre-eGRASP(p30) (a gift from Bong-Kiun Kaang, Addgene plasmid #111583), pAAV-EWB-DIO-myriRFP670V5-P2A-post-eGRASP (a gift from Bong-Kiun Kaang, Addgene plasmid #111585), pAAV-hSyn-Cre-P2A-dTomato (a gift from Rylan Larsen, Addgene plasmid #107738), pAAV-GFAP-HA-rM3D(Gs)-IRES-mCitrine (a gift from Bryan Roth; Addgene plasmid #50472), Ef1a-fDIO-tdTomato (a gift from Patricia Jensen; Addgene plasmid #128434), pAAV-hSyn-fDIO-hM4D(Gi)-mCherry-WPREpA (a gift from Ulrik Gether; Addgene plasmid #154867), and hSyn-fDIO-hM3D(Gq)-mCherry (a gift from Ulrik Gether; Addgene plasmid #154868).
The following constructs were made by Gibson assembly or restriction cloning: Fos-Flex-Flp, GFAP-FSF-GCaMP6, GFAP-FSF-hM3D-mCherry, Fos-FlpER, GFAP-fDIO-tdTomato-4x6T, GFAP-fDIO-Cre-HA-4x6T, GFAP-FSF-hM4D-mCherry, GFAP-fDIO-rM3D-mCitrine, and GFAP-fDIO-Cre-P2A-dTomato-4x6T.
All new constructs were verified by whole plasmid sequencing before AAV packaging. AAVs were packaged by the Optogenetics and Viral Vectors Core at the Jan and Dan Duncan Neurological Research Institute. AAV2/9 was produced at the following titers (in GC/mL): GFAP-Cre-4x6T: 7.3×1012, GFAP-mCherry: 3.5×1012, GFAP-hM3D-mCherry: 1.2×1012, hSyn-hM3D-mCherry: 3.5×1012, Fos-Flex-Flp: 9.3×1012, GFAP-FSF-GCaMP6: 9.3×1011, Fos-FlpER: 2.6×1012, GFAP-fDIO-tdTomato-4x6T: 2.8×1012, hSyn-fDIO-EYFP: 1.8×1012, GFAP-FSF-hM3D-mCherry: 2.3×1012, GFAP-fDIO-Cre-HA-4x6T: 3.7×1012, hSyn-fDIO-hM3D-mCherry 2.3×1012, CaMKIIa-hM3D-mCherry 3.9×1012, Fos-CreER 6.7×1012,Ef1α-DIO-Pre-eGRASP 4.0×1012, Ef1α-DIO-Post-eGRASP 2.1×1012, GFAP-FSF-hM4D-mCherry 3.9×1012, GFAP-fDIO-rM3D-mCitrine 4.2×1011, and GFAP-fDIO-Cre-P2A-dTomato-4x6T 4.9×1011. Retrograde AAV hSyn-hM4D mCherry (titer 2.3×1013 GC/mL) was purchased from Addgene (50475-AAVrg).
AAV injection surgery
Mice were anesthetized with isoflurane (3% induction, 2% maintenance) in oxygen and placed in a stereotaxic frame. The scalp was shaved and cleaned with alternating applications of 70% ethanol and Betadine. After a midline scalp incision, burr holes were drilled in the skull and 250-500 μL of AAV was injected in each target site using a pulled pipette attached to a Drummond Nanoject III. Injections were targeted bilaterally to CA1 (2.0 mm posterior, 1.4 mm lateral and 1.2 mm ventral relative to Bregma) and DG (2.0 mm posterior, 1.4 mm lateral and 1.9 mm ventral relative to Bregma). Combinations of AAVs for each experiment are listed in the Results and Figures. After injections, the scalp was sutured closed. 0.6 mg/kg of buprenorphine was given s.c. 30 minutes prior to surgery and every 12 hours afterwards for three days.
Histology
Mice were perfused with phosphate buffered saline (PBS) followed by 4% paraformaldehyde in PBS. Brains were dissected and postfixed overnight in 4% paraformaldehyde at 4°C and then cryoprotected with 20% sucrose overnight. Brains were sectioned at 35 μm with a cryostat.
For immunostaining, sections were washed three times in PBS and then incubated for an hour in PBS with 0.25% Triton and 10% donkey serum (blocking buffer), then incubated overnight at room temperature with primary antibodies. The following primary antibodies were used: rat anti-c-Fos (1:5000; Synaptic Systems, 226 017), guinea pig anti-Cre (1:500; Synaptic Systems, 257 004), chicken anti-GFAP (1:1000; abcam, ab4674), rabbit anti-GFAP (1:1000; DAKO, Z0334), rabbit anti-HA (1:500; Roche, 11867423001), rabbit anti-NeuN (1:2000, Millipore, MABN140), rabbit anti-NFIA (1:500; Sigma, HPA006111), rabbit anti-RFP (1:500; Rockland, 600-401-379), and rabbit anti-Sox9 (1:500; Millipore, AB5535). Sections were then washed three times in PBS followed by incubation with secondary antibodies (1:500) in blocking buffer for one hour. The following secondary antibodies were used: Alexa Fluor 488-conjugated donkey anti-chicken (Jackson ImmunoResearch, 703-545-155), Alexa Fluor 647-conjugated donkey anti-chicken (Jackson ImmunoResearch, 703-605-155), Alexa Fluor 594-conjugated donkey anti-guinea pig (Jackson ImmunoResearch, 706-585-148), Alexa Fluor 488-conjugated donkey anti-rabbit (Jackson ImmunoResearch, 711-545-152), Alexa Fluor 594-conjugated donkey anti-rabbit (Jackson ImmunoResearch, 711-585-152), Alexa Fluor 647-conjugated donkey anti-rabbit (Jackson ImmunoResearch, 711-605-152), Alexa Fluor 488-conjugated donkey anti-rat (Jackson ImmunoResearch, 712-545-153), Alexa Fluor 594-conjugated donkey anti-rat (Jackson ImmunoResearch, 712-585-153), Alexa Fluor 647-conjugated donkey anti-rat ( Jackson ImmunoResearch, 712-605-153). Sections were washed three additional times in PBS prior to mounting and coverslipping.
Confocal imaging and analysis
Confocal images were acquired with a Zeiss LSM 880 laser scanning confocal microscope with 20× air, 40× oil, or 63× oil objective. Images were taken with 1-2 μm steps in the z plane and 1024 × 1024 or 2048 × 2048 pixel resolution. Images were quantified using ImageJ/FIJI (v1.54f) by experimenters blind to group identity. Cell numbers were quantified manually using the multi-point selection tool. For examination of c-Fos expression in astrocytes, we excluded analysis of the subgranular zone due to expression of the astrocyte marker Sox9 in neural stem cells in this region. For quantification of EYFP+ neuron coverage of GFAP+ astrocyte territories, maximum intensity z-projections were created encompassing the territory of a given astrocyte. The polygon selection tool was used to trace the territory of each astrocyte based on the GFAP channel. Each region of interest selection was applied to the corresponding region in the thresholded EYFP channel and the area fraction of EYFP+ structures was computed. To account for regional differences in neuronal labeling density and differences in laser power needed to image different neurons, area fraction from each astrocyte was normalized to the mean of non-LAAs (tdTomato−) for each image. The area measurement output was used to measure territory size. eGRASP puncta were counted within the 3D territory of each astrocyte based on the GFAP channel. 2D projections were used to measure astrocyte territory based on the GFAP channel. Data were derived as the number of puncta (within 3D space) per astrocyte 2D territory. To account for regional variation in the number and proximity of labeled neurons and differences in fluorescence intensity, the number of puncta was normalized to the mean of non-learning-associated astrocytes (tdTomato−) for each image. A caveat of this analysis is that an astrocyte may not functionally interact with every eGRASP+ synapse within its territory. For fluorescence quantification, the integrated density was measured from manually drawn regions of interest.
Contextual fear conditioning
The fear conditioning chamber (Context A) was a 25 cm × 35 cm arena with a metal grid floor. The alternate context (Context B) was a 25 cm × 45 cm arena with a plastic floor. The arenas had distinct patterns on certain walls as visual cues. Context A was cleaned with 35% isopropyl alcohol and Context B was cleaned with 70% ethanol to give the contexts distinct scents. A fan provided approximately 65 dB of ambient noise for both chambers. Mice were recorded from above during all behavioral sessions.
For fear conditioning, mice were placed into Context A for 5 minutes. Mice were allowed to explore the chamber for the first 120s. Then, 1.5 mA foot shocks lasting 2s were delivered at 120s, 180s, and 240s after mice were initially placed in the chamber. Mice were removed from the chamber 60s after the final foot shock and returned to their home cage. For subsequent tests, mice were placed into Context A or Context B for 5 minutes. Videos were manually scored for freezing behavior by a blinded investigator. Freezing was defined as the absence of movement, with the exception of breathing-related movement, lasting >1s.
Novel place recognition
Mice were placed in a square arena (40 cm × 40 cm) with two identical objects (12 cm tall cylinders) in adjacent quadrants. Mice were trained three consecutive days, 10 minutes per day, to learn the position of the objects. The arena had distinct shapes on certain walls as orientation cues. On the testing day, the location of one object was moved to a different quadrant. Mice were placed in the arena for five minutes and recorded from above. The discrimination index, a measure of spatial memory performance, was calculated as: (time spent investigating object in novel place – time spent investigating object in original place) / total time exploring either object.
Acute slice preparation
Mice were deeply anesthetized with isoflurane and decapitated. The brain was dissected and placed in ice-cold oxygenated (95% O2, 5% CO2) cutting solution that contained: 130 mM NaCl, 24 mM NaHCO3, 1.25 mM NaH2PO4, 3.5 mM KCl, 1.5 mM CaCl2, 1.5 mM MgCl2, and 10 mM D(+)-glucose, pH 7.4. A vibratome was used to collect 300 μm hippocampal sections, which were transferred to oxygenated artificial cerebrospinal fluid solution (ACSF, 125 mM NaCl, 25 mM glucose, 25 mM NaHCO3, 2.5 mM KCl, 2 mM CaCl2, 1.25 mM NaH2PO4 and 1 mM MgCl2, pH 7.3, 310–320 mOsm). Slices were then recovered in oxygenated ACSF for at least one hour prior to recording or imaging.
Slice electrophysiology
Hippocampal slices were prepared as above and then acclimated at room temperature for at least one hour with continuous perfusion of oxygenated ASCF. LTP was induced by theta burst stimulation (TBS, 10 trains of 4 half maximal stimuli at 100 Hz within 200 ms interval) onto the Shaffer collateral pathway. Subthreshold stimulation was 10 pulses at 40 Hz33. Field EPSPs (fEPSPs) were measured in the CA1 stratum radiatum. The recording pipette was filled with 1M NaCl solution. fEPSPs were calculated as the slope of the response (mV/ms) and normalized to the average of the baseline. Final potentiation was calculated by averaging the last 5 minutes of responses. Whole-cell recording was performed with pCLAMP10 and Multi-Clamp 700B amplifier (Axon Instrument, Molecular Devices) from hippocampal CA1 neurons. The holding potential was −60 mV. Pipette resistance was typically 5–8 MΩ. The pipette was filled with an internal solution: 135 mM CsMeSO4, 8 mM NaCl, 10 mM HEPES, 0.25 mM EGTA, 1 mM Mg-ATP, 0.25 mM Na2-GTP, 30 mM QX-314, pH adjusted to 7.2 with CsOH (278–285 mOsmol) for EPSC measurement; 135 mM CsCl, 4 mM NaCl, 0.5 mM CaCl2, 10 mM HEPES, 5 mM EGTA, 2 mM Mg-ATP, 0.5 mM Na2-GTP, 30 mM QX-314, pH adjusted to 7.2 with CsOH (278–285 mOsmol) for IPSC measurement. Spontaneous EPSCs were measured in the presence of GABAAR antagonist, bicuculline (20 μM, Tocris). IPSCs were measured in the presence of ionotropic glutamate receptor antagonists, APV (50 μM, Tocris) and CNQX (20 μM, Tocris). Electrical signals were digitized and sampled at 50 ms intervals with Digidata 1550B and Multiclamp 700B amplifier (Molecular Devices, CA) using pCLAMP (10.7) software. Data were filtered at 2 kHz. The recorded current was analyzed with ClampFit (10.7) software.
Calcium imaging and analysis
Acute slices were prepared as described above. Slices were maintained in oxygenated ACSF with a perfusion system. Calcium traces were recorded using a two-photon microscope (LSM 7MP, Zeiss) equipped with a Coherent Chameleon Ultra (II) Ti-sapphire laser tuned to 900 nm and a 20x, 1.0 NA water-immersion Zeiss objective. GCaMP signals were recorded for 3-5 min per trial at 1,024 × 1,024 pixel resolution and 1 Hz temporal resolution from astrocytes at depths of approximately 30 μm below the surface. All multiphoton imaging experiments were performed within 4 hours of slicing. Images were quantified using GECIquant44, ImageJ (v.54f), and Clampfit (10.7) software. The region of interest detection for soma and main processes or microdomains was performed in a semi-automated manner using the GECIquant algorithm as previously described44. The amplitude, area under the curve, and frequency values of GCaMP signals were computed using Clampfit. Spike width was calculated as the full width at half maximum for each spike.
Tissue dissociation and fluorescence activated cell sorting
Mice were perfused with cold saline and hippocampi were dissected on ice. Hippocampi from three mice were pooled per replicate and dissociated as previously described35. Dissociated GFP+ tdTomato− and GFP+ tdTomato+ astrocytes, and GFP− tdTomato− cells, were sorted using a Becton Dickinson FACSAria II with BD FACSDiva software. Samples were sorted using a 70 micron nozzle with 70 PSI of pressure. After singlet gating with forward and side scatter, singlet events were assessed for GFP and tdTomato fluorescence. Non-fluorescent brain tissue was used as a negative control to set the double negative population gate. We also performed single color controls to establish ideal voltage, potential compensation, and gating. The GPF signal was detected with a 488nm laser, through a 505nm long pass filter, and 530/30 detector. tdTomato signal was detected with 561nm laser, through a 570 long pass filter, and 585/51 detector. 100,000 events of each type were collected into 1.5 ml tubes containing Qiagen lysis buffer. Samples were then frozen at −80°C until RNA extraction.
RNA extraction, library preparation, and sequencing
RNA was extracted from sorted cells using RNeasyMicro Kit (Cat.No. 4004, QIAGEN). RNA integrity (RINR8.0) was assessed using the High Sensitivity RNA Analysis Kit (DNF-472-0500, Agilent) on a 12-Capillary Fragment Analyzer. cDNA synthesis and Illumina sequencing libraries with 8-bp single indices were constructed from 10 ng total RNA using the Revelo RNA-seq High Sensitivity kit (Cat.No. 30184149, TECAN). The resulting libraries were validated using the Standard Sensitivity NGS Fragment Analysis Kit (DNF-473-0500,) on a 12-Capillary Fragment Analyzer and quantified using the Quant-it dsDNA assay kit (Cat. Q33120). Equal concentrations (2 nM) of libraries were pooled and processed for paired-end (R1: 75, R2: 75) sequencing of approximately 40 million reads per sample using the High Output v2 kit (FC-404-2002, Illumina) on an Illumina NextSeq 550 following the manufacturer’s instructions.
RNA-seq analysis
Sequencing files from each flow cell lane were downloaded and the resulting fastq files were merged. Quality control was conducted using fastQC (v.0.10.1) and MultiQC (v.0.9)45. Reads were mapped to the mouse genome mm10 assembly using STAR (v2.5.0a)46. RNA-seq data were analyzed and plotted as previously described23. DESeq2 (v1.20.0) was used for both differential gene expression analysis (Wald test) and read count normalization. We defined DEGs as those with normalized reads per million >5 in at least two of the replicates and expression fold-change >1.5 at p < 0.01. Gene Ontologies associated with DEGs were determined using Enrichr. Motif analysis was performed using Hypergeometric Optimization of Motif Enrichment (HOMER, v4.10) to identify transcription factor motifs enriched near the transcription start site of differentially expressed genes.
RT-qPCR
RT-qPCR was conducted on cDNA libraries using Quantabio Perfecta SYBR Green Fast Mix (Cat. 95072-012) on a Roche Light Cycler 480 instrument. Reactions were set up using 2 ng cDNA, 250 nM primers and 1x SYBR mix. qPCR was carried out at 30 s of 95°C, 40 cycles of 95°C for 5 s and 60°C for 30 s, with subsequent melting curve analysis. Expression of transcripts of target genes was normalized to Gapdh. The primers for RT-qPCR used are as follows: Gapdh (forward 5′- TGGCCTTCCGTGTTCCTAC-3′, reverse 5′- GAGTTGCTGTTGAAGTCGCA-3′), Tuj1 (Forward 5’- TAGACCCCAGCGGCAACTAT-3’. Reverse 5’- GTTCCAGGTTCCAAGTCCACC-3’), Nfia (Forward 5’- GAAGCGCATGTCGAAAGAAGA-3’, Reverse 5’- GGCGGAGGCAGTCAATTCTC-3’), and Sox9 (Forward 5’- AGTACCCGCATCTGCACAAC-3’, Reverse 5’- ACGAAGGGTCTCTTCTCGCT-3’).
Statistical analysis
All sample sizes and statistical tests are noted in Figure Legends. Alpha was set at P < 0.05. Graphpad Prism (v10) was used for all statistical analyses. Independent samples were compared with two-tailed t tests. Variance was assessed with F tests, and Welch’s corrected t tests were used when variance was significantly different. One-way ANOVA and nested t tests were used as noted in the legends. Post-hoc tests were used following significant ANOVA, as noted in the legends.
Extended Data
Extended Data Figure 1. Chemogenetic activation of hippocampal neurons induces c-Fos expression in a subset of astrocytes.
a. Genetic system and timeline for chemogenetic activation of dentate gyrus neurons. b. Representative image of hM3D-mCherry labeling in the hippocampus (representative of 3 mice). c. Representative immunostaining of c-Fos expression in hippocampal astrocytes. d. Quantification of c-Fos+ Sox9+ astrocytes following saline or CNO injection (n = 4 mice per group, upper panel: two-tailed t test, ****P < 0.0001, lower panel: two-sided Welch’s corrected t test, **P = 0.003). e. Genetic system and timeline for chemogenetic activation of excitatory dentate gyrus neurons. Retrograde AAV-hSyn-hM4D-mCherry was used to inhibit inputs. f. Representative immunostaining of c-Fos expression in hippocampal astrocytes. g. Quantification of c-Fos+ Sox9+ astrocytes following CNO injection (n = 5 mice per group, upper panel: one-way ANOVA (P<0.0001) and Tukey tests, ****P < 0.0001, lower panel: one-way ANOVA (P<0.0001) and Tukey tests, ***P = 0.0002, ****P < 0.0001). Data are mean±SEM.
Extended Data Figure 2. Specific targeting of astrocytes with AAV-GFAP-Cre-4x6T.
a. Schematic of genetic system for testing specificity of AAV-GFAP-Cre-4x6T in tdTomato Cre reporter mice. b. Representative immunostaining for Sox9+ astrocytes and NeuN+ neurons. c. Quantification of overlap between tdTomato and Sox9 or NeuN (n = 8 mice). Data are mean±SEM. Panel a was created using Biorender.com
Extended Data Figure 3. Additional characterization of c-Fos expression and Fos cKO.
a. Representative images of c-Fos expression in hippocampal neurons from mice left in homecage or 90 minutes after fear conditioning. b. Quantification of neuronal c-Fos expression (n = 4 mice per group (HC = homecage, FC = fear conditioning), two-tailed t tests, **P = 0.002 (upper), 0.008 (lower). c. Representative images encompassing portions of stratum radiatum, stratum lacunosum, and stratum moleculare of Sox9 and c-Fos. Arrows indicate c-Fos+ Sox9+ cells. Representative of 4 mice/group. d. Representative images showing astrocytic expression of Cre following injection of AAV-GFAP-Cre-4x6T. Images were processed with a dehaze filter. Representative of 4 mice. e. Representative images of astrocytic c-Fos expression from GFAP-mCherry and GFAP-Cre-4x6T injected mice. Arrows indicate c-Fos+ Sox9+ cells. Representative of 8 mice/group. Data are mean±SEM.
Extended Data Figure 4. Calcium imaging and electrophysiology in Fos cKO mice.
a. Calcium signal traces from control and Fos cKO astrocytes. Each row in the heatmaps represents a single astrocyte (n = 20 control, 18 Fos cKO astrocytes, n= 3 mice per group). b-c. Quantification of calcium signaling parameters in soma/major banches (b) and microdomains (c) (n = 20 control, 18 Fos cKO astrocytes, n= 3 mice per group for some/main branches; n = 16 astrocytes, n = 3 mice per group for microdomains). Two-tailed t test, *P = 0.038. d. Representative traces and quantification of sEPSCs in control (n = 8) and Fos cKO (n = 10) mice. Two-tailed t test, *P = 0.026. e. Representative traces and quantification of sIPSCs in control (n = 8) and Fos cKO (n = 9) mice. Data are mean±SEM.
Extended Data Figure 5. Additional characterization of learning-associated astrocyte distribution and microdomain calcium activity.
a. Quantification of learning-associated astrocytes across hippocampal layers (see Fig. 2). Comparisons on graph are P values from Dunnett’s tests. N= 6 homecage, N = 5 context only, N = 5 0.75 mA shock, N = 8 1.5 mA shock mice. b. Representative images of tdTomato+ learning-associated astrocytes in the hippocampus. Str. oriens = stratum oriens, str. pyr. = stratum pyramidale, str. rad. = stratum radiatum, str. l-m. = stratum lacunosum-moleculare, str. mol. = stratum moleculare of the dentate gyrus, str. gr. = stratum granulosum. Representative of 5 mice each. c. Quantification of microdomain calcium activity parameters related to Figure 2d-e. Middle panel: two-tailed Mann Whitney test, *P = 0.030. Right panel: two-tailed t test, *P = 0.036. n = 16 homecage and 17 fear conditioning cells from 3 mice per group. Data are mean±SEM.
Extended Data Figure 6. Tamoxifen does not affect astrocytic c-Fos expression.
a. Timeline for examining astrocytic c-Fos expression after injection of tamoxifen, 4-hydroxytamoxifen, or vehicle. b. Representative images of Sox9 and c-Fos immunostaining in CA1 and DG. c. Quantification of c-Fos expression in Sox9+ cells. One-way ANOVA, P ≥ 0.268. n = 4 mice per group. Data are mean±SEM.
Extended Data Figure 7. Tagging learning-associated astrocyte ensembles with Fos-FlpER.
a. Schematic of genetic system and timeline for evaluating activity-dependent labeling of astrocytes with Fos-FlpER. b. Representative images of hippocampal astrocytes labeled in home cage and fear conditioned mice. c. Quantification of tdTomato+ cells (n = 4 mice per group, two-tailed t test, **P = 0.001). d. Representative immunostaining showing colocalization of tdTomato with GFAP. Representative of 4 mice/group. Data are mean±SEM.
Extended Data Figure 8. Learning-associated astrocytes re-express c-Fos after memory recall.
a. Schematic of genetic system for labeling learning-associated astrocytes with tdTomato and experimental design for labeling active astrocytes during fear conditioning or recall. b. Representative images of tdTomato+ learning associated astrocytes. c. Quantification of density of tdTomato+ astrocytes tagged during fear conditioning (FC) or recall (N = 4 mice per group). Two-tailed t test, P = 0.818. d. Schematic of genetic system for labeling learning-associated astrocytes with tdTomato during fear conditioning and experimental design for examining c-Fos expression after memory recall. e. Representative image of c-Fos expression in a learning-associated astrocyte after memory recall. Arrow indicates a c-Fos+ tdTomato+ astrocyte. f. Quantification of c-Fos expression after recall in Sox9+ non-learning associated astrocytes (tdTomato−) and learning-associated astrocytes following recall (tdTomato+). tdTomato+ astrocytes were ~2.4% of all Sox9+. N = 4 mice, two-tailed Welch’s t test, **P = 0.002. Data are mean±SEM.
Extended Data Figure 9. Learning-associated astrocytes facilitate long-term potentiation.
a. Schematic of genetic system for expressing hM3D-mCherry in learning-associated astrocytes. b. Experimental timeline and schematic of LTP recordings. c. fEPSP traces from slices treated with CNO or saline followed by subthreshold stimulation (n = 6 fear conditioned hM3D + saline, n = 9 homecage hM3d + CNO, n = 7 fear conditioned hM3D + CNO, n = 7 pan-astrocyte hM3D + CNO mice). d. Summary of fEPSP slope from the last 5 minutes of recordings in panel c (n = 6 fear conditioned hM3D + saline, n = 9 homecage hM3d + CNO, n = 7 fear conditioned hM3D + CNO, n = 7 pan-astrocyte hM3D + CNO mice). One way ANOVA, P = 0.004; Dunnett’s tests, *P = 0.047, **P = 0.007. e. fEPSP traces from slices treated with CNO but without electrical stimulation. Two-tailed t tests comparing 5-minute time bins, *P ≤ 0.049. n = 9 LAA hM3D and 8 mCherry mice. Data are mean±SEM. Panel a, b were created using Biorender.com
Extended Data Figure 10. Additional characterization of learning-associated astrocyte-engram neuron interactions and viral labeling.
a. Schematic of AAVs and experimental timeline related to Fig. 3k-m. b. Images showing AAV-mediated labeling and immunostaining for c-Fos in dentate gyrus. Arrow indicates an EYFP+ c-Fos positive engram neurons. Arrowhead indicates an hM3D-mCherry+ astrocyte located within the dendritic arbor of the neuron. The magenta channel fluorescence was subtracted from the red channel in order to offset spectral bleed-through. Representative of 4 mice. c. Re-activation of engram neurons in dentate gyrus was increased by learning-associated astrocyte activation (two-tailed t test, **P = 0.007). n = 10 saline, 7 CNO mice. d. Low magnification images showing AAV-mediated labeling of engram neurons and immunostaining for c-Fos. e. Images of mCherry and EYFP labeled cells in CA1 and DG. Representative of 4 mice. f. Quantification of the ratio of engram neurons (EYFP+) to learning-associated astrocytes (mCherry+) in CA1 and DG. N = 3 mice per region. g. Low magnification image showing viral targeting of the hippocampus with AAV-GFAP-mCherry. Representative of 8 mice. h. Higher magnification image of viral labeling representative of 6 mice. i. Quantification of viral labeling of astrocytes. n = 6 mice per region. Data are mean±SEM.
Extended Data Figure 11. Remote reactivation of learning-associated astrocyte ensembles elicits recall.
a. Timeline. b. Schematic of genetic system for evaluating recall after remote reactivation of the fear-tagged astrocyte ensemble. c. Quantification of freezing behavior prior to foot shocks (n = 8 saline, n = 7 CNO mice, two-tailed t test, P = 0.368). d. Quantification of freezing behavior in a neutral Context B 30 minutes after injection of 3 mg/kg CNO (n = 8 saline, n = 7 CNO mice, two-tailed t test, *P = 0.012). Data are mean±SEM. Panel a was created using Biorender.com
Extended Data Figure 12. Gating strategy for cell sorting.
Example of the gating strategy used for sorting GFP+ tdTomato+ (learning-associated) and GFP+ tdTomato− (non-learning-associated) astrocytes.
Extended Data Figure 13. Validation of increased NFIA in learning-associated astrocytes.
a. Schematic of genetic system. b. Timeline for labeling learning-associated astrocytes. c. Representative immunostaining for NFIA. Arrow indicates tdTomato+ learning-associated astrocyte, arrowheads indicate tdTomato− astrocytes. d. Quantification of NFIA fluorescence intensity in tdTomato+ and tdTomato− astrocytes (n = 62 CA1 tdTomato− astrocytes, n = 61 CA1 tdTomato+ astrocytes, n = 48 DG tdTomato− astrocytes, n = 52 tdTomato+ DG astrocytes, n = 4 mice, nested t tests, *P = 0.021, ***P = 0.001). e. RT-qPCR quantification of neuronal Tuj1 mRNA in sorted cell populations. Tuj1 was de-enriched in GFP+ samples (n = 3 technical replicates, n = 3 biological replicates per group, one-way ANOVA and Tukey’s post hoc tests, ****P < 0.0001). f. RT-qPCR quantification of astrocytic Sox9 mRNA in sorted cell populations. Sox9 was enriched in GFP+ samples (n = 3 technical replicates, n = 3 biological replicates per group, one-way ANOVA and Tukey’s post hoc tests, **P = 0.002, ****P < 0.0001). g. RT-qPCR quantification of Nfia mRNA in sorted cell populations. Nfia was enriched in GFP+ tdTomato+ samples (n = 3 technical replicates, n = 3 biological replicates per group, one-way ANOVA and Tukey’s post hoc tests, *P = 0.022 (top), 0.031 (bottom). Data are mean±SEM. Panel a was created using Biorender.com
Extended Data Figure 14. Fos deletion in learning-associated astrocytes impairs memory recall.
a. Timeline for Fos knockout in learning-associated astrocytes and examining memory recall. b. Schematic of genetic system. c. Confirmation of lack of Fos expression in dTomato+ learning-associated astrocytes relative to dTomato− astrocytes (N = 5 mice, two-tailed Welch’s t test, ****P < 0.0001). d. Quantification of freezing behavior during recall test in Context A. N = 10 Control (vehicle injected), N = 11 Fos cKO mice; two-tailed t test, *P = 0.014. e. Representative image showing lack of c-Fos expression in dTomato+ astrocytes. Representative of 5 mice. f. Low magnification image of learning-associated astrocyte labeling in CA1 and DG. Note that no dTomato+ astrocytes express c-Fos. Representative of 5 mice. Panel a was created using Biorender.com
Supplementary Material
Acknowledgements
This work was supported by US National Institutes of Health grants R35-NS132230, R21-MH134002, and R01-AG071687 to BD. MRW is supported by AHA-23POST1019413. WK is supported by National Research Foundation of Korea (RS-2024-00405396). We are thankful for support from the David and Eula Wintermann Foundation. We would like to acknowledge the Optogenetics and Viral Vectors Core at the Jan and Dan Duncan Neurological Research Institute. Research reported in this publication was supported by the Eunice Kennedy Shriver National Institute of Child Health & Human Development of the National Institutes of Health under Award Number P50HD103555 for use of the Microscopy Core facilities and the Animal Phenotyping & Preclinical Endpoints Core facilities. Images in schematics were created using Biorender.com.
Footnotes
Competing Interests
The authors declare no competing interests.
Data availability
RNA-sequencing data have been deposited at the NCBI GEO under accession number GSE254016. All other data in this Article are available from the corresponding author on reasonable request. Source data are provided with this paper.
References
- 1.Josselyn SA & Tonegawa S. Memory engrams: Recalling the past and imagining the future. Science 367, eaaw4325 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Liu X. et al. Optogenetic stimulation of a hippocampal engram activates fear memory recall. Nature 484, 381–385 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Ramirez S. et al. Creating a false memory in the hippocampus. Science 341, 387–391 (2013). [DOI] [PubMed] [Google Scholar]
- 4.Roy DS et al. Brain-wide mapping reveals that engrams for a single memory are distributed across multiple brain regions. Nat Commun 13, 1799 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Han J-H et al. Selective Erasure of a Fear Memory. Science 323, 1492–1496 (2009). [DOI] [PubMed] [Google Scholar]
- 6.Dallérac G, Zapata J. & Rouach N. Versatile control of synaptic circuits by astrocytes: where, when and how? Nat Rev Neurosci 19, 729–743 (2018). [DOI] [PubMed] [Google Scholar]
- 7.Kofuji P. & Araque A. Astrocytes and Behavior. Annu Rev Neurosci 44, 49–67 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Allen NJ & Eroglu C. Cell Biology of Astrocyte-Synapse Interactions. Neuron 96, 697–708 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Nagai J. et al. Behaviorally consequential astrocytic regulation of neural circuits. Neuron 109, 576–596 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Soto JS et al. Astrocyte–neuron subproteomes and obsessive–compulsive disorder mechanisms. Nature 616, 764–773 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Sardar D. et al. Induction of astrocytic Slc22a3 regulates sensory processing through histone serotonylation. Science 380, eade0027 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Cheng Y-T et al. Social deprivation induces astrocytic TRPA1-GABA suppression of hippocampal circuits. Neuron 111, 1301–1315.e5 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Lawal O, Ulloa Severino FP & Eroglu C. The role of astrocyte structural plasticity in regulating neural circuit function and behavior. Glia 70, 1467–1483 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Fleischmann A. et al. Impaired Long-Term Memory and NR2A-Type NMDA Receptor-Dependent Synaptic Plasticity in Mice Lacking c-Fos in the CNS. J. Neurosci 23, 9116–9122 (2003). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Katche C. et al. Delayed wave of c-Fos expression in the dorsal hippocampus involved specifically in persistence of long-term memory storage. Proceedings of the National Academy of Sciences 107, 349–354 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Lacagnina AF et al. Distinct hippocampal engrams control extinction and relapse of fear memory. Nature Neuroscience 22, 753–761 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Khakh BS & Sofroniew MV Diversity of astrocyte functions and phenotypes in neural circuits. Nat Neurosci 18, 942–952 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Papouin T, Dunphy J, Tolman M, Foley JC & Haydon PG Astrocytic control of synaptic function. Philosophical Transactions of the Royal Society B: Biological Sciences 372, 20160154 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Endo F. et al. Molecular basis of astrocyte diversity and morphology across the CNS in health and disease. Science 378, eadc9020 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Kol A. et al. Astrocytes contribute to remote memory formation by modulating hippocampal–cortical communication during learning. Nat Neurosci 23, 1229–1239 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Cheng Y-T et al. Inhibitory input directs astrocyte morphogenesis through glial GABABR. Nature 617, 369–376 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Henneberger C, Papouin T, Oliet SHR & Rusakov DA Long-term potentiation depends on release of D-serine from astrocytes. Nature 463, 232–236 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Huang AY-S et al. Region-Specific Transcriptional Control of Astrocyte Function Oversees Local Circuit Activities. Neuron 106, 992–1008 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Adamsky A. et al. Astrocytic Activation Generates De Novo Neuronal Potentiation and Memory Enhancement. Cell 174, 59–71.e14 (2018). [DOI] [PubMed] [Google Scholar]
- 25.Suthard RL et al. Basolateral amygdala astrocytes are engaged by the acquisition and expression of a contextual fear memory. J. Neurosci JN-RM-1775-22 (2023) doi: 10.1523/JNEUROSCI.1775-22.2023. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Yap E-L & Greenberg ME Activity-Regulated Transcription: Bridging the Gap between Neural Activity and Behavior. Neuron 100, 330–348 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Gleichman AJ, Kawaguchi R, Sofroniew MV & Carmichael ST A toolbox of astrocyte-specific, serotype-independent adeno-associated viral vectors using microRNA targeting sequences. Nat Commun 14, 7426 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Sun W. et al. Spatial transcriptomics reveal neuron–astrocyte synergy in long-term memory. Nature 1–8 (2024) doi: 10.1038/s41586-023-07011-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Chen MB, Jiang X, Quake SR & Südhof TC Persistent transcriptional programmes are associated with remote memory. Nature 587, 437–442 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Yu X. et al. Reducing Astrocyte Calcium Signaling In Vivo Alters Striatal Microcircuits and Causes Repetitive Behavior. Neuron 99, 1170–1187.e9 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Ung K, Tepe B, Pekarek B, Arenkiel BR & Deneen B. Parallel astrocyte calcium signaling modulates olfactory bulb responses. J Neurosci Res 98, 1605–1618 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Choi J-H et al. Interregional synaptic maps among engram cells underlie memory formation. Science 360, 430–435 (2018). [DOI] [PubMed] [Google Scholar]
- 33.Park H. et al. Channel-mediated astrocytic glutamate modulates hippocampal synaptic plasticity by activating postsynaptic NMDA receptors. Molecular Brain 8, 7 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Yu X. et al. Context-Specific Striatal Astrocyte Molecular Responses Are Phenotypically Exploitable. Neuron 1–17 (2020) doi: 10.1016/j.neuron.2020.09.021. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Lin C-CJ et al. Identification of diverse astrocyte populations and their malignant analogs. Nature Neuroscience 20, 396–405 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Morel L. et al. Molecular and Functional Properties of Regional Astrocytes in the Adult Brain. Journal of Neuroscience 37, 8706–8717 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.de Ceglia R. et al. Specialized astrocytes mediate glutamatergic gliotransmission in the CNS. Nature 1–10 (2023) doi: 10.1038/s41586-023-06502-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Ghandour K. et al. Orchestrated ensemble activities constitute a hippocampal memory engram. Nat Commun 10, 2637 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Ung K. et al. Olfactory bulb astrocytes mediate sensory circuit processing through Sox9 in the mouse brain. Nat Commun 12, 5230 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Rao-Ruiz P. et al. Engram-specific transcriptome profiling of contextual memory consolidation. Nat Commun 10, 2232 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41.Lozzi B, Huang T-W, Sardar D, Huang AY-S & Deneen B. Regionally Distinct Astrocytes Display Unique Transcription Factor Profiles in the Adult Brain. Frontiers in Neuroscience 14, (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Cheng Y-T, Woo J. & Deneen B. Sculpting Astrocyte Diversity through Circuits and Transcription. The Neuroscientist 107385842210826 (2022) doi: 10.1177/10738584221082620. [DOI] [PMC free article] [PubMed] [Google Scholar]
Methods References
- 43.Scavuzzo MA et al. Pancreatic Cell Fate Determination Relies on Notch Ligand Trafficking by NFIA. Cell Reports 25, 3811–3827.e7 (2018). [DOI] [PubMed] [Google Scholar]
- 44.Srinivasan R. et al. Ca2+ signaling in astrocytes from Ip3r2−/− mice in brain slices and during startle responses in vivo. Nat Neurosci 18, 708–717 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45.Ewels P, Magnusson M, Lundin S. & Käller M. MultiQC: summarize analysis results for multiple tools and samples in a single report. Bioinformatics 32, 3047–3048 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46.Dobin A. et al. STAR: ultrafast universal RNA-seq aligner. Bioinformatics 29, 15–21 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
RNA-sequencing data have been deposited at the NCBI GEO under accession number GSE254016. All other data in this Article are available from the corresponding author on reasonable request. Source data are provided with this paper.