Memories of individuals’ experiences are represented across assemblies of neurons in hippocampal and cortical circuits. Several mechanisms of formation and maintenance of these assemblies have been proposed. The most prominent such mechanism is stimulus-induced long-term potentiation of synaptic connectivity7, an energy-demanding process that involves extensive biochemical and morphological adaptations at all levels of neuronal function8,9. There is also evidence for contributions of pre-existing developmental and other intrinsic programmes of individual neurons10, including the baseline expression of the transcriptional factor CREB11 and lineage of developmental origin12,13. Recent focus has also been on the role of the interneuronal perineuronal nets (PNNs) in the stabilization of memory circuits through tightened control of inhibitory inputs to dedicated neuronal assemblies14. Here we explored whether an overarching process could integrate stimulus-dependent and pre-existing mechanisms that underlie the commitment of neurons to memory-specific assemblies.

CFC increases Tlr9 expression

We first performed an analysis of transcriptional profiles of dorso-hippocampal neurons beyond the initial, well-established 24−48 h time window when protein signalling, immediate early gene (IEG) expression and delayed gene expression (for example, growth factors) take place15,16. To this end, we performed bulk RNA-sequencing (RNA-seq) of total RNA isolated from individual mouse hippocampi obtained either 96 h or 21 days after contextual fear conditioning (CFC) (one-trial 3 min exposure to a context followed by a 2 s, 0.7 mA shock, constant current), and noted a robust difference in the gene expression profiles for recent relative to remote memory with little within-group variability (a total of 847, with 440 up-regulated) (Fig. 1a). We previously reported that the 21-day gene expression repertoire revolved around cilium and extracellular matrix genes needed for PNN formation17, but the 96 h gene expression profiles associated with the shaping of recent memory representations remained unexplored. We found that 96 h after CFC the majority of differentially expressed genes were immune response genes involved in nucleic acid sensing and cytokine release (71 up-regulated and 11 down-regulated) (Fig. 1a and Extended Data Fig. 1). Within immune response genes, the applied analyses identified TLR9 and its downstream NF-κB signalling pathway18 as the most prominent functional gene clusters (Extended Data Fig. 1a,b), which is consistent with the reported activation of the NLRP3 inflammasome several days after CFC19. We replicated the RNA-seq data with new sets of individual hippocampal samples using quantitative PCR, confirming the robustness of the observed effect (Extended Data Fig. 1c,d). The up-regulation of Tlr9 gene expression was accompanied by increased TLR9 protein, as well as increased co-localization of TLR9 with LAMP2, a marker of late endosomes and lysosomes (Fig. 1b), an effect that was neuron-specific (Fig. 1c). After CFC, TLR9 co-labelling with LAMP2 was significantly stronger compared with markers of early (EEA1 and RAB7) and recycling (Rab11) endosomes (Fig. 1d), suggesting enhanced TLR9 trafficking to endosomes, enabling DNA recognition and NF-κB activation20.

Fig. 1: Nucleic acid-sensing activity after CFC.
figure 1

a, Bulk RNA-seq showed increased expression of 441 genes in hippocampi obtained 96 h after CFC (recent, n = 7 mice) compared with those collected 21 days (remote, n = 5 mice) after CFC. Volcano plots demonstrate significant increases in expression of genes related to inflammation and TLR signalling. P adj, adjusted P value. b, TLR9 protein levels and co-localization of TLR9 with the mature vesicle marker LAMP2 at different times after CFC. LAMP2 levels did not fluctuate, TLR9 levels and its co-localization with LAMP2 increased 6 h after CFC, peaking 96 h later (n = 6 mice, 360 neurons per time point; one-way ANOVA; LAMP2: P = 0.3104, F(3,19) = 1.278; TLR9: P = 0.0005, F(3,20) = 9.363; co-localization: P < 0.0001, F(3,20) = 21.27). c, TLR9 and LAMP2 signals in glial cells (revealed by nuclear size), show no significant co-localization (n = 6 mice, 12 glial cells per time point; one-way ANOVA; P = 0.8186, F(3,20) = 0.3090). Green arrow, LAMP2; orange and white arrows, LAMP2–TLR9 co-localization; red arrow, TLR9. Scale bars: left, 25 μm; right, 40 μm. d, TLR9–vesicle pool co-localization 96 h after CFC (early endosome: EEA1 and RAB7; recycling endosome: RAB11; late endosome: LAMP2) reveals that the highest overlap is with LAMP2 (orange arrows; n = 6 mice, 30 neurons per time point; one-way ANOVA; P < 0.0001, F(3,20) = 31.53). Note the lack of TLR9 and LAMP2 signals in a glial cell (cyan arrow). Data are mean ± s.e.m. Scale bar, 20 μm. e, Hippocampal cytosolic dsDNA (naive, 24 h or 96 h after CFC) shows no contamination with nuclear DNA, as revealed by lack of ubiquitous amplification of Slc17a7 (which encodes vGlut1) (left). Cloning and sequencing identified genomic dsDNA fragments enriched with non-coding gene GC sequences 24 h and 96 h after CFC (left graph), sized 50–300 bp (right graph). Ctrl, control; miRNA, mitochondrial RNA; ncRNA, non-coding RNA; snoRNA, small nucleolar RNA; snRNA, small nuclear RNA. f, In vitro imaging of primary hippocampal neurons using fluorescent dyes, revealing mobile extranuclear DNA distinct from mitochondrial DNA (Supplementary Video 1). Scale bar, 10 μm. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001; NS, not significant; WT, wild type.

Source Data

TLR9, together with the cyclic GMP–AMP synthase (cGAS)–STING, is the main sensor of extranuclear DNA. In the absence of infection or apoptotic DNA damage, these pathways are typically silent, but under neuronal stress, TLR9 responds to the release of mitochondrial DNA fragments21. To determine whether extranuclear dsDNA fragments with TLR9-activating potential can be found in hippocampal neurons, we collected hippocampi of naive mice or mice trained in CFC 24 or 96 h earlier and isolated their extranuclear DNA. After confirming that the fraction was not contaminated with genomic DNA, we proceeded with cloning and sequencing of the isolated dsDNA fragments (Fig. 1e). Contrary to our expectation that, if any, we would identify mitochondrial dsDNA, all cloned dsDNA fragments were of genomic origin, predominantly stemming from non-coding DNA. We cloned a total of 53 dsDNA fragments belonging to 25 unique genomic sequences, none of which corresponded to mitochondrial genes. After CFC, the frequency of CG-containing sequences, which are putative activators of TLR9, transiently increased from 33% to 77%. We also performed live-cell imaging of primary hippocampal neuronal cultures using dyes that detect dsDNA or mitochondrial DNA to distinguish non-mitochondrial from mitochondrial extranuclear DNA. Non-mitochondrial mobile DNA signals were readily detected in the cytosol (Fig. 1f and Supplementary Video 1).

CFC triggers dsDNA breaks and DDR

Neuronal activity is known to induce transient DNA breaks3, which are required for the learning-related induction of IEGs. However, such breaks occur in enhancers of coding genes and are repaired within minutes without affecting neuronal homeostasis5. To examine the origin of extranuclear dsDNA fragments generated by CFC, we hypothesized that, in discrete neuronal populations, activity-induced DNA damage might be more substantial and sustained. We performed a time-course study up to 96 h after CFC, focusing on dsDNA breaks in individual cells of the hippocampal CA1 region because of its well-established role in the formation and retention of context memories22. To detect DNA damage, we performed immunofluorescent labelling with antibodies specific for the dsDNA break-binding phospho-histone γH2AX3 and quantified the signals using plot profile or particle analyses. To minimize the interference of background γH2AX activity unrelated to dsDNA breaks, foci were selected if their fluorescence intensity was at least two standard deviations greater than the average intensity of the total nuclear signal for each CA1 region of interest. One and three hours after CFC, we identified discrete, patchy clusters of CA1 nuclei exhibiting a significant increase of the total number of neurons with clear γH2AX foci (Fig. 2a, left and Supplementary Video 2). Whereas the number of nuclear foci subsequently decreased, much larger γH2AX-labelled foci emerged in individual neurons, persisting from 6 to 96 h (Fig. 2a, middle). All γH2AX signals were neuron-specific and were not found in astrocytes or microglia (Fig. 2a, right). Co-labelling with the nuclear envelope protein lamin B1 revealed that at the time of maximal dsDNA break detection, a subset of nuclei exhibited envelope ruptures (discontinuation of lamin B1 labelling), resulting in perinuclear release of γH2AX in RNA-rich areas typical of the endoplasmic reticulum (Fig. 2b and Extended Data Fig. 2a), a primary localization site of inactive TLR9. The number of ruptures increased significantly 1 h after CFC and remained detectable in a smaller number of nuclei throughout the 96 h period. The perinuclear γH2AX signals co-labelled with TLR9 (Fig. 2c) and to a lesser extent with the DNA dye Hoechst and with antibodies recognizing dsDNA (Extended Data Fig. 2b). These findings indicated that in some neurons, γH2AX and dsDNA, alone or in complexes, were released from the nucleus in TLR9-containing perinuclear sites. The persistent γH2AX signals found at later time points (6–96 h) had a diameter of 4 µm or larger, and co-localized with the centrosome markers centrin and γ-tubulin, demonstrating pericentrosomal localization rather than localization at sites of dsDNA breaks (Fig. 2d, left), as suggested previously4. Contrary to the small, sharp and focused nuclear signals, all pericentrosomal signals were large and fuzzy. These signals were strongly and consistently co-labelled with 53BP1, the main mediator of DNA repair by nonhomologous end joining23,24 (Fig. 2d, right). Neurons undergoing dsDNA breaks and DNA damage repair (DDR) did not show any nuclear morphological indices of apoptosis. Both nuclear and pericentrosomal γH2AX signals overlapped with cleaved caspase-3 foci, but the patterns of their localization were consistent with the non-apoptotic, probably memory-related roles of cleaved caspase-325 (data not shown). Although consistent with evidence of centrosomal localization of 53BP126 and with increasing recognition of centrosomal control of DDR27, the extent to which dsDNA breaks and centrosomes shared molecular components, including γH2AX, dsDNA fragments and DNA repair enzymes was unexpected and suggested that the proposed role of centrosomes in the maintenance of gene integrity in dividing cells28 also applies to adult neurons undergoing memory-related activity.

Fig. 2: DNA damage and DDR after CFC.
figure 2

a, Left, number of neurons showing γH2AX puncta (n = 360 neurons per group; one-way ANOVA, P < 0.0001, F(5,30) = 65.09) and size of γH2AX foci (n = 360 neurons per group; one-way ANOVA; P < 0.0001, F(5,30) = 38.17) after CFC. Right, localization of γH2AX (green arrows) in neurons (marked by NeuN; cyan arrows) relative to astrocytes (marked by GFAP; purple arrows) and microglia (marked by IBA1; orange arrows) (two-way ANOVA; factor: cell type, P < 0.0001, F(2,90) = 673.8; factor: time, P < 0.0001, F(5,90) = 77.73; cell type × time, P < 0.0001, F(10,90) = 77.95). Bottom right, number of γH2AX foci. b, Nuclear envelope ruptures coinciding with detection of extranuclear γH2AX (red arrows) and DNA (green arrows) (n = 6 mice; one-way ANOVA; P = 0.0038, F(5,30) = 4.445). Scale bars: top row, 25 μm; bottom row, 20 μm. c, Extranuclear γH2AX overlapping with TLR9 (orange arrows) (n = 120 total, 75% overlap). Scale bar, 20 μm. d, Pericentrosomal accumulation of γH2AX shown by co-localization with centrin 2 and γ-tubulin (arrowheads). Additional co-recruitment of 53BP1, revealing centrosomal DDR (n = 30–131 neurons; two-tailed Chi-square test; \({\chi }_{(4)}^{2}=22.98\), P < 0.0001; post hoc analysis using Bonferroni-corrected α = 0.05, 53BP1: 3 h versus 1 h NSP = 0.116, 6 h versus 1 h ****P < 0.0001, 24 h versus 1 h ****P < 0.0001, 96 h versus 1 h ****P < 0.0001; centrin 2: 3 h versus 1 h NSP = 0.5061, 6 h versus 1 h NSP = 0.2061, 24 h versus 1 h ####P < 0.0001, 96 h versus 1 h ####P < 0.0001; adjusted α P < 0.001). e, Co-labelling of γH2AX+ (purple arrows) and Fos+ (green arrow) neurons (20%). f, Significantly lower number of γH2AX+ neurons (orange arrows) relative to PRAM+ neurons (green arrows) show memory reactivation (co-labelling with Fos; purple arrows) (n = 216 neurons; two-tailed Chi-square test; \({\chi }_{(3)}^{2}=6.518\), P = 0.0384; post hoc analysis using Bonferroni-corrected α = 0.05, γH2AX+ versus PRAM+ *P = 0.0215, γH2AX+PRAM+ versus PRAM+ NSP = 0.1007; adjusted α P < 0.025). Data are mean ± s.e.m. Dox, doxycycline; T, total neurons; R, reactivated neurons. Scale bar, 20 μm.

Source Data

We next studied whether the detection of γH2AX signals was related to the CFC-induced up-regulation of IEGs. We collected brain sections 1 h after CFC because this time point is optimal for the detection of Fos in CA129, as well as EGR1, which shows both sizeable baseline expression as well as up-regulation after learning30. We also performed immunostaining for CREB, whose baseline expression has been directly implicated in memory11. Many neurons with the typical, homogenous nuclear labelling for Fos, CREB and EGR1 were devoid of γH2AX signals and showed overall insignificant correlation (Fig. 2e and Extended Data Fig. 3). Even in instances of co-detection, IEG labelling in neurons positive for γH2AX was mainly punctate and scattered. Only 20% of all γH2AX-positive nuclei were Fos-positive (52% of Fos+ nuclei were co-labelled with γH2AX). Instead, γH2AX signals were associated with inflammatory signalling, as shown by a high degree of co-localization with RELA, the most abundant NF-κB family member, which translocates to the nucleus upon activation (Extended Data Fig. 3).

We also examined to what extent IEG or DDR (centrosomal γH2AX signals)-responsive neurons up-regulate Fos during subsequent memory reactivation (Fig. 2f and Extended Data Fig. 3e,f). Because IEG responses peak 1 h after CFC and DDR peaks at 96 h, we permanently labelled the CFC-activated IEG population with GFP using the robust activity marker PRAM–GFP driven by the Fos promoter31. In this way we could examine Fos reactivation 96 h after CFC in both populations by comparing the overlap of γH2AX, GFP and Fos. As γH2AX+Fos+ neurons were scarce, we compared 30–34 neurons per population. Relative to PRAM-labelled neurons, γH2AX+ (DDR population) neurons showed significantly smaller Fos responses after memory reactivation, demonstrating that both after CFC and after memory reactivation, IEG and inflammatory signalling occurred predominantly in non-overlapping neuronal populations.

Tlr9 in CA1 neurons is required for context memory

We next investigated whether the inflammatory response is a side effect of learning-induced DNA damage or whether it contributes to memory formation. We induced a neuron-specific knockout of TLR9 in CA1 dorso-hippocampal neurons of Tlr9fl/fl mice by locally injecting adeno-associated virus (AAV9) expressing Cre recombinase–GFP or GFP under control of the human synapsin promoter (Syn) (Fig. 3a, left). The Syn-cre-injected mice showed impaired context memory, as revealed by significant reduction of freezing behaviour over multiple tests (Fig. 3a, middle). The neuron-specific virus expression was shown by differential staining for astrocytes and microglia (Fig. 3a, right and Extended Data Fig. 4a). Knockdown was validated by quantitative analysis of TLR9 and RELA (Fig. 3b). The memory deficit was replicated by a different approach, using injection of AAV9 expressing Syn-driven Tlr9-targeting short hairpin RNA (shRNA) in the hippocampi of wild-type mice (Fig. 3c). The hippocampal Tlr9 knockdown (Tlr9-KO) also impaired trace fear conditioning (TFC), as shown by reduced freezing during context and tone tests, but did not affect delay fear conditioning (DFC), a form of hippocampus-independent learning (Fig. 3d). Because viral injections are known to cause inflammatory responses of astrocytes and microglia (as confirmed in our post-mortem analysis; Extended Data Fig. 4b), we also determined the contribution of these cell populations to the observed memory deficits. Astrocytic knockdown using GFAP-cre did not affect CFC after injection in Tlr9fl/fl mice (Extended Data Fig. 4c). Consistent with previous findings32, microglial depletion was similarly ineffective, as determined in both wild-type and Tlr9fl/fl mice receiving a CSF1R inhibitor in their diet or regular diet after injection of Syn-cre (Extended Data Fig. 4d,e). Introduction of Syn-cre in the hippocampi of WT mice did not affect CFC relative to Syn-GFP-injected wild-type controls (Extended Data Fig. 4e, bottom right).

Fig. 3: Impaired context memory after neuron-specific deletion of hippocampal Tlr9.
figure 3

a, Left, experimental schematic. Middle, persistent reduction of freezing during context tests of Tlr9fl/fl mice injected intrahippocampally with Syn-cre (n = 11 mice) compared with the control group injected with Syn-GFP (n = 9 mice; two-way ANOVA with repeated measures; factor: virus, P = 0.0007, F(1,18) = 16.54; factor: test, P = 0.0007, F(1.936,34.84) = 9.30, virus × test, P = 0.4358, F(2,36) = 0.85). Right, lack of co-localization of Syn-cre with astrocytic and microglial markers. Scale bar, 40 μm. b, Left, reduction of TLR9 levels (mean optical density per neuron, 60 neurons per mouse, 7 mice per group; two-tailed unpaired t-test; t12 = 17.4700, P < 0.0001) and RELA nuclear signal (60 neurons per mouse, 7 mice per group; two-tailed unpaired t-test; t12 = 3.5679, P = 0.0039) after neuron-specific deletion of TLR9. Right, representative micrographs. Orange arrow indicates TLR9, red arrow indicates RELA. Scale bar, 20 μm. c, Persistent reduction of freezing during context tests of wild-type mice injected intrahippocampally with neuron-specific Tlr9 shRNA (n = 10 mice) compared with scrambled RNA (scrRNA) (n = 8 mice). Two-way ANOVA with repeated measures; factor: virus, P < 0.0001, F(1,16) = 35.50; factor: test, P = 0.2347, F(2,32) = 1.517; virus × test, P = 0.0027, F(2,32) = 7.168. d, After TFC, Tlr9-KO resulted in impaired freezing during context (two-tailed unpaired t-test; t15 = 4.362, P = 0.0006) and tone tests after TFC (two-tailed unpaired t-test; t15 = 3.899, P = 0.0014) (GFP, n = 9 mice; Syn-cre, n = 7 mice), but intact freezing during the tone test after DFC (two-tailed unpaired t-test; t14 = 1.214, P = 0.2448). Data are mean ± s.e.m.

Source Data

We next examined the involvement of TLR9 relative to cGAS–STING signalling in memory using pharmacological manipulations of these pathways. The TLR9 antagonist oligonucleotide ODN2088 significantly impaired CFC, whereas the small drug cGAS–STING inhibitors RU-521 and H-151 were ineffective relative to vehicle (Extended Data Fig. 5a). Consistent with these observations, CFC was intact in mice lacking the Sting1 gene relative to wild-type controls (Extended Data Fig. 5b). Finally, to test the role of endogenous processing of extranuclear dsDNA in memory formation, we also manipulated the levels of two DNases controlling cellular DNA sensing. DNase2 digests dsDNA in shorter dsDNA fragments required for TLR9 binding and activation33. To prevent the generation of TLR9-activating DNA fragments we virally overexpressed Dnase2 shRNA or scrambled shRNA in mouse dorsal hippocampi. Dnase2 shRNA treatment significantly impaired memory formation (Extended Data Fig. 5c). TREX1 is a cytosolic DNase that predominantly restricts the activity of the cGAS–STING pathway33. We disrupted cGAS–STING-mediated DNA sensing by overexpressing TREX1–GFP in the dorsal hippocampus, using GFP overexpression as a control. In these mice, context memory formation was not affected (Extended Data Fig. 5d). In sum, these experiments provided converging evidence for a role of neuron-specific TLR9-mediated but not cGAS–STING-mediated dsDNA sensing in the formation and persistence of context memories.

Tlr9 knockdown disrupts CFC-induced gene expression

To better characterize the effects of CFC and neuron-specific Tlr9-KO on gene expression in individual neuronal and non-neuronal hippocampal populations, we performed single-nucleus RNA-sequencing (snRNA-seq). For this experiment, we expressed the viral vectors in the entire dorsal hippocampus to analyse their effects across all subfields. We collected nuclei and prepared libraries (Lib) from 4 groups of mice (pooled from 5 mice per group): Tlr9fl/fl mice injected with Syn-GFP and euthanized 96 h after CFC (Lib 1); Tlr9fl/fl mice injected with Syn-cre and euthanized 96 h after CFC (Lib 2); naive Tlr9fl/fl mice injected with Syn-GFP (Lib 3); and naive Tlr9fl/fl mice injected with Syn-cre (Lib 4). Lib 4 was sorted to obtain GFP-positive and GFP-negative nuclei and used to determine the cell specificity of cre expression by snRNA-seq (Extended Data Fig. 6a), whereas the other samples were analysed for CFC-induced (Lib 1 versus Lib 3) and Tlr9-KO (Lib 2 versus Lib 1) effects (Fig. 4). An unsupervised algorithm identified 29 clusters, with the highest diversity seen among excitatory CA1 neurons (12 clusters) relative to dentate gyrus granule cells (DGGC) (4 clusters), interneurons (4 clusters) and non-neuronal cells (typically 1 cluster each) (Extended Data Fig. 6b). Analyses of robust gene expression changes (more than 1.5-fold) revealed both increases and decreases of gene expression in the CFC and Tlr9-KO groups (Fig. 4a and Supplementary Tables 1 and 2). Tlr9-KO abolished the CFC-induced gene expression but did not affect the expression of these genes relative to the naive control (Lib 2 versus Lib 3; Supplementary Table 3). An exception was cluster 26, which showed a paradoxical up-regulation of genes involved in axon guidance and adhesion in Tlr9-KO (Extended Data Fig. 7). CFC induced a set of highly conserved genes across neuronal clusters and occasionally in non-neuronal cells, with most genes associated with the endomembrane system and acting as endoplasmic reticulum chaperones and regulators of vesicle trafficking and function, and with interleukin-6 production (Fig. 4b and Supplementary Tables 4 and 5). Some of the proteins encoded by these genes mediate TLR9 folding (HSP90B1) and induce vesicle acidification (ATP6V0C), which are critical for TLR9 activation (Fig. 4c and Extended Data Fig. 6c), whereas others, such as those associated with interleukin-6 signalling, require TLR9 activation. In line with recent reports, we detected many doublecortin-positive (DCX+) neurons among CA1 neurons, DGGC and interneurons34 (Fig. 4d and Extended Data Fig. 8), and found that the conserved gene responses were more pronounced in the DCX+ cluster (Fig. 4e). The expression of Dcx in CA1 and up-regulation of Hsp90b1 after CFC were also confirmed by RNAscope analysis (Extended Data Fig. 9). A reactome analysis of broader gene expression profiles (including all genes showing a significant increase in expression (P < 0.05) in addition to those showing conserved changes) revealed CFC-induced gene expression across pathways involved in RNA and protein metabolism, vesicle trafficking, immunity, cell cycle, DNA repair and cilium assembly, however in this case, DCX excitatory CA1 and DGGC neurons were the most affected population (Fig. 4f). Although we did not detect low abundance transcripts of immune mediators (such as TLR9 or RELA), the findings confirmed that 96 h after CFC, neurons acquire inflammatory phenotypes associated with TLR signalling, DNA repair, ciliogenesis and vesicle trafficking. In addition to the described effects, CFC and Tlr9-KO induced some unexpected phenotypic changes of excitatory neurons, such as fluctuations of vGlut2 and DCX (Extended Data Fig. 8).

Fig. 4: Single-cell changes of gene expression after neuron-specific deletion of hippocampal Tlr9.
figure 4

a, Tlr9fl/fl mice injected with Syn-cre or Syn-GFP were trained in CFC or left untrained (naive), and 96 h later, dorsal hippocampal nuclei were isolated and processed for snRNA-seq. Robust (more than 1.5-fold) changes of gene expression were found after CFC in mice injected with control virus (Syn-GFP CFC versus Syn-GFP naive) or Cre virus (Syn-cre CFC versus Syn-GFP CFC). b, Gene Ontology (GO) analysis reveals that most genes regulated by CFC and Tlr9-KO involve endoplasmic reticulum (ER), mitochondrial function, IL-6 production and inflammation. CFC induced up-regulation of these genes, whereas Tlr9-KO blocked this effect. c, The most conserved genes up-regulated by CFC and down-regulated by Tlr9-KO across cell clusters include Atp6v0c and Hsp90b1, key regulators of TLR9 function (Extended Data Fig. 6). d, Dcx expression superimposed on uniform manifold approximation and projection (UMAP) analysis of snRNA-seq data from dorsal hippocampal cells. The expression of Dcx in the main neuronal clusters is outlined in orange (excitatory CA1), red (inhibitory) and green (DGGC). e, Cell-specific changes in gene expression demonstrates dominant effects of CFC and Tlr9-KO on gene expression in neurons relative to other cell populations with particularly strong effects of Tlr9-KO in DCX+ CA1 neurons. DG, dentate gyrus; oligo, oligodendrocyte; OPC, oligodendrocyte precursor cell. f, Reactome analysis reveals the major functional gene networks affected by CFC and Tlr9-KO in DCX CA1 and DGGC neurons. Circles are scaled to the percentage effect of Tlr9-KO on the gene expression or pathway relative to Syn-GFP with CFC. TLR cascades, DDR and cilium assembly are enriched among the pathways that are most up-regulated by CFC and down-regulated by Tlr9-KO. RNA Pol II, RNA polymerase II.

Source Data

Given that immune cells express high levels of TLR9 and are also a source of circulating, cell-free DNA35, we also examined whether bloodborne infiltrating cells and DNA from extracellular sources might contribute to the observed up-regulation of TLR9 signalling and memory. Such contribution was unlikely, however, given the absence of lymphocytic and myeloid cell markers in our samples (Extended Data Fig. 10a). Similarly, systemic or intrahippocampal infusions of DNase1 (which efficiently degrades cell-free DNA) before and shortly after CFC were ineffective (Extended Data Fig. 10b), as was intrahippocampal injection of DNase1 and S1 nuclease, which degrades single-stranded DNA, before the context test.

In summary, the snRNA-seq approach identified ongoing CFC-induced cell-specific (mainly neuron-specific) gene expression responses and discrete phenotypic changes 96 h after CFC. Tlr9-KO blunted the induction of most of the up-regulated genes without affecting CFC down-regulated genes and induced additional, CFC-unrelated changes, especially in excitatory cluster 26.

TLR9 controls DDR, ciliogenesis and PNN build-up

To determine the cellular consequences of Tlr9-KO, we collected the hippocampi of WT and Tlr9fl/fl mice 24 h after their last memory test following each experiment, and first examined DNA damage and DDR response. In addition, we collected hippocampi of Relafl/fl and interferon receptor 1-floxed (Ifnar1fl/fl) mice undergoing the same genetic and behavioural manipulations to examine the potential contributions of these pathways downstream of TLR9. A control group that was not exposed to CFC was also included. We found a significant increase of the number of neurons with dsDNA breaks in all mice, including wild-type mice, injected with Syn-cre. However, this effect was more pronounced in the genetically modified lines (Fig. 5a and Extended Data Fig. 11) and was also found in mice that were not exposed to CFC. There were large multifocal nuclear accumulations of γH2AX and 53BP1 in wild-type and Ifnar1fl/fl mice injected with Syn-cre, as previously found in nuclear bodies formed around DNA lesions in cell lines undergoing replication stress28, with an average size of 2 µm corresponding to nuclear stress bodies36. Notably, in Tlr9fl/fl and Relafl/fl mice, Syn-cre completely disrupted the recruitment of 53BP1 to sites of DNA damage and to centrosomal DDR sites (Fig. 5a,b). To examine whether this affected other memory-related centrosomal functions, we also examined the consequences of individual gene knockdown on ciliogenesis and cilium-dependent PNN formation17. Whereas wild-type and IFNAR1fl/fl mice injected with Syn-Cre showed intact ciliogenesis in CA1 neurons, as shown by filamentous staining for adenyl cyclase III (ACIII), neurons of Tlr9fl/fl and Relafl/fl mice showed punctate and disorganized ACIII labelling accompanied by disappearance of PNNs, with the scarce remaining PNNs showing markedly reduced complexity (Fig. 5c and Extended Data Fig. 12). These findings demonstrated that lack of TLR9 disrupts the nuclear and centrosomal DDR machinery so that at the time of CFC, CA1 neurons could not recruit DDR complexes or form cilia and PNNs. Whereas both Rela-KO and Ifnar1-KO could contribute to Tlr9-KO-induced genomic instability, the effects on 53BP1, ciliogenesis and PNN formation were most likely to be mediated by the RELA downstream pathway.

Fig. 5: Impaired DDR, ciliogenesis and PNN formation by neuron-specific deletion of hippocampal Tlr9.
figure 5

a, Increased number of neurons showing γH2AX signals in mice hippocampally injected with Syn-cre relative to Syn-GFP (n = 5 mice (150 neurons) per group; two-tailed Chi-square test; \({\chi }_{(4)}^{2}=54.86\), P < 0.0001; post hoc analysis using Bonferroni correction, α = 0.05; versus GFP: WT cre ****P < 0.0001, Tlr9fl/fl ****P < 0.0001, Relafl/fl ****P < 0.0001 and Ifnar1fl/fl ****P < 0.0001). This effect was further potentiated by injection of Syn-cre in hippocampi of Tlr9fl/fl, Relafl/fl and Ifnar1fl/fl mice (versus WT cre: Tlr9fl/fl ##P < 0.002, Relafl/fl ##P < 0.003 and Ifnar1fl/fl ##P < 0.006; adjusted α P < 0.001). The observed genomic instability was accompanied by centrosomal DDR in wild-type and Ifnar1fl/fl mice, but blunted in Tlr9fl/fl and Relafl/fl mice injected with Syn-cre (n = 5 mice (150 neurons) per group; two-tailed Chi-square test; \({\chi }_{(4)}^{2}=124.1\), P < 0.0001; post hoc analysis using Bonferroni correction α = 0.05; versus WT cre: Tlr9fl/fl ####P < 0.0001, Relafl/fl ####P < 0.0001 and Ifnar1fl/fl NSP = 0.9681). Right, up-regulated γH2AX signals (white arrows) were seen in CA1 neurons but not in adjacent astrocytes or microglia. Scale bar, 20 μm. b, Illustration of the findings presented in a. To facilitate signal detection of γH2AX–53BP1 overlap, 53BP1 images were pseudocoloured with cyan. Scale bar, 20 μm. c, Top, whereas Syn-cre injection in wild-type mice and Ifnar1 deletion did not affect the number of cilia (n = 6 mice per group, 30 neurons per mouse) and PNNs (n = 6 mice per group, one slice of dorsal CA1 per mouse), Tlr9 and Rela knockout impaired ciliogenesis and PNN formation (top; one-way ANOVA; ACIII: P < 0.0001, F(4,25) = 97.47; PNNs: P < 0.0001, F(4,25) = 23.31). Representative micrographs depicting ACIII signals (middle, red arrows; scale bar, 20 μm) and PNNs (bottom, blue arrows; scale bar, 100 μm). Wisteria floribunda lectin (WFN). Data are mean ± s.e.m.

Source Data


The recruitment of individual neurons to assemblies is essential not only for encoding individual memories, but also for protecting them from streams of incoming information over time, ensuring stability and persistence of memory representations. On the basis of the evidence presented here, we suggest that in distinct populations of hippocampal CA1 excitatory neurons, this is achieved through learning-induced TLR9 signalling linking DNA damage to DDR. Over several days, such neurons acquired an inflammatory phenotype involving activation of the TLR9 DNA-sensing pathway and pericentrosomal accumulation of DDR complexes. TLR9 activation was most probably triggered by the release of γH2AX and dsDNA fragments, stemming predominantly from non-coding DNA, into the endomembrane system. Some of these neurons underwent more profound phenotypic fluctuations consistent with changes of their differentiation state. The identified genetic, molecular and cellular phenotypes were blocked by neuron-specific knockdown of Tlr9, and some, including disruption of pericentrosomal DDR and ciliogenesis, were replicated by Rela knockdown but not by Ifnar1 knockdown, identifying NF-κB rather than interferon pathways as downstream mediators. The involvement of DNA sensing by TLR9, but not the evolutionary older cGAS–STING pathway in CFC, suggests that neurons have adopted an immune-based memory mechanism involving trafficking of dsDNA and histone to endolysosomes, rather than their release in the cytosol37. This is consistent with the gene expression profiles observed 96 h after CFC, which were primarily related to endoplasmic reticulum function and vesicle trafficking.

Although we could not directly demonstrate activation of TLR9 by specific DNA fragments at the single-neuron level, we found evidence that several TLR9-activating mechanisms are induced by CFC. TLR9 signalling is induced mainly by unmethylated CpG DNA sequences, which are predominantly of bacterial origin38, but such sequences can also be generated through demethylation of mammalian DNA during learning and long-term potentiation15. Moreover, there is increasing evidence of TLR9 activation by mammalian self-DNA39,40 and by histones41, both of which we found to be released from the nucleus following nuclear envelope ruptures.

Given the association of dsDNA damage with neurodegeneration42, neurons undergoing learning-induced dsDNA breaks might be expected to be excluded from memory assemblies. This was also suggested by lack of significant association between the observed inflammatory responses with other mechanisms implicated in learning, including IEG responses. However, we found that these neurons were required for memory formation, but only if their TLR9-mediated DDR machinery was intact. Although many factors could predispose individual neurons to respond with dsDNA damage rather than IEG, it needs to be considered that diverse responses to synaptic input43 could be based on the genetic diversity of neurons generated throughout life44. The occurrence of sustained dsDNA breaks might reflect a particularly intense CA1 response of discrete CA1 neurons best tuned to the context information carried by excitatory input5, and resulting in their lasting representation of that context. A shared history of cycles of DNA damage and repair could similarly organize neuronal assemblies by collective experience as a shared history of IEG response or birthdate, providing yet another dimension to memory organization. An advantage of recruiting distinct neuronal populations to memory representations is that they can uniquely contribute to their stability, specificity, flexibility and other properties31,45. Whereas populations displaying IEG responses during memory formation and reactivation might be better suited for retrieval-mediated memory updates and modifications46, the population displaying an inflammatory DDR phenotype seems better suited to maintain stable representations of context memories. By coordinating centrosome-mediated DDR, TLR9 could integrate genomic responses to synaptic inputs (dsDNA breaks), which subserve memory specificity, with delayed, extrasynaptic memory mechanisms (ciliogenesis and PNN formation), which promote memory persistence14,17,40, at least when memories are hippocampally dependent47.

Our findings support recent data revealing more widespread expression of DCX in hippocampal and cortical neurons than originally believed34,48. Fluctuations of DCX in response to hyperexcitation and inflammation are thought to reflect changes in neuronal maturation states49. We found similar changes within small, DCX+ CA1 clusters (clusters 11, 21 and 24) that lost their DCX+ phenotype in response to CFC, indicating that some neurons, at least transiently, shifted to a more mature state. Given the essential role of DCX in nucleus-centrosome coupling50, it is conceivable that DCX could also contribute to the recruitment of centrosomal DDR complexes in some of the identified neurons.

In addition to its relevance for memory, TLR9 activity was essential for the maintenance of neuronal genomic integrity. Knockdown of TLR9 inflammatory signalling showed that even mild stimuli, such as spontaneous or memory-related neuronal activity (and possibly other endogenous factors such as commensal microbes51) can trigger substantial genomic instability in the absence of TLR9-controlled DNA sensing and DDR. This suggests that the consequences of TLR9-based anti-inflammatory treatment strategies52, although beneficial for restricting astrocytic and microglial activation, may prove detrimental for neuronal health if they compromise the functioning of the neuronal TLR9 pathway. Given that genomic instability is considered to be a gateway to accelerated senescence and psychiatric and neurodegenerative disorders42,53,54, neuron-specific TLR9 and downstream RELA signalling are likely to emerge as promising preventive and therapeutic targets for preserving neurocognitive health.


Animal husbandry

Eight-week-old male and female C57BL/6 N mice were obtained from Envigo. Tlr9fl/fl (C57BL/6J-Tlr9em1Ldm/J) mice with loxP sites flanking exon 1 of the Tlr9, Ifnar1fl/fl (B6(Cg)-Ifnar1tm1.1Ees/J) mice with loxP sites flanking exon 3 of the type-1 interferon α/β receptor gene, Relafl/fl (B6.129S1-Relatm1Ukl/J) mice with loxP sites flanking exon 1 of the Rela gene, and Sting1-knockout mice C57BL/6J-Sting1gt/J; Goldenticket (Tmem173gt) mice were obtained from Jackson Laboratory and bred in institutional facilities. All mice were eight weeks of age at the beginning of experiments, unless otherwise specified. All mice were group housed (12 h:12 h light:dark cycle with lights on at 07:00, temperature 20–22 °C, humidity 30–60%) with ad libitum access to food and water (mice were switched to single housing one week before experiments). All experimental groups were mixed sex, consisting of approximately equal numbers of males and females. All procedures were approved by Northwestern University’s Animal Care and Use Committee (protocols IS00002463 and IS00003359) and Albert Einstein’s Animal Care and Use Committee (protocols 00001289 and 00001268) in compliance with US National Institutes of Health standards, and at Aarhus University in compliance with the Danish National Animal Experiment Committee (protocol 2021-15-0201-00801).

Tissue collection

For all analyses, mice were euthanized by cervical dislocation, dorsal hippocampi immediately dissected and frozen in liquid nitrogen. Frozen tissue was stored at −80 °C until protein or RNA extractions were performed.

Bulk RNA-seq

Read quality was assessed using FastQC46 (v0.10.1) to identify sequencing cycles with low average quality, adapter contamination, or repetitive sequences from PCR amplification. Alignment quality was analysed using SAMtools flagstat with default parameters. Data quality was visually inspected as described previously55. Furthermore, we assessed whether samples were sequenced deep enough by analysing the average per base coverage and the saturation correlation for all samples using the MEDIPS R package56. The saturation function splits each library in fractions of the initial number of reads (ten subsets of equal size) and plots the convergence. The correlation between biological replicates was evaluated using Pearson correlation (function MEDIPS.correlation). Only data passing all quality standards were used for further analyses. Data were aligned to the genome using gapped alignment as RNA transcripts are subject to splicing and reads might therefore span two distant exons. Reads were aligned to the whole Mus musculus mm10 genome using STAR aligner58 (2.3.0e_r291) with default options, generating mapping files (BAM format). Reads were aligned to mouse genome M. musculus mm10 and counted using FeaturesCount as described previously55. Genes with significantly different fold change (false discovery rate-corrected P < 0.05) were classified as up-regulated or down-regulated.

Isolation of nuclei and FACS

Fresh dorsal hippocampal tissue was collected from mice 96 h after CFC. Nuclei were isolated following an established protocol57. RNase inhibitor was added to the 6× homogenization buffer stable master mix at a final concentration of 1.2 U μl−1. Tissue was homogenized using a Polytron homogenizer for 1.5 min. Samples were resuspended with the 50% iodixanol solution (50% iodixanol in 1× homogenization buffer) via gentle pipetting to make a final concentration of 25% iodixanol. Three millilitres of a 35% iodixanol solution (35% iodixanol in 1× homogenization containing 480 mM sucrose) was added to a new 15 ml Falcon tube. Three millilitres of a 29% iodixanol solution (29% iodixanol in 1× homogenization buffer containing 480 mM sucrose) was layered above the 35% iodixanol mixture. A 4 ml 25% iodixanol solution was layered on the 29% solution. In a swinging-bucket centrifuge, nuclei were centrifuged for 20 min at 3,000g. After centrifugation, the nuclei were present at the interface of the 29% and 35% iodixanol solutions. This band with the nuclei was collected in a 300 μl volume and transferred to a pre-chilled tube. Following isolation, nuclei were counted and resuspended in fluorescence-activated cell sorting (FACS) buffer (1% BSA, 1 mM EDTA, and 0.2 U μl−1 RNase inhibitor in PBS) at 2 × 106 nuclei per ml. A subset of nuclei was sorted by GFP expression on a high-speed cell sorter flow cytometer (Beckman Coulter MoFlowXDP cell sorter). See Supplementary Fig. 1 for the detailed gating strategy.


snRNA-seq libraries were generated from up to 10,000 individual cells captured in an oil emulsion on a Chromium Controller (10x Genomics). cDNA was generated in the individual cell–gel bead emulsion micro-reactors while adding barcodes at the cellular and molecular level using the Chromium Next GEM Single Cell 3′ Kit v3.1(10x Genomics kit 1000268). The barcoded cDNAs from the individual cells were combined for the remaining library process. The unique molecular barcodes (UMIs) prevented amplification artefacts from skewing the analysis. The libraries were analysed on a Fragment Analyzer 5200 (Agilent Technologies) to ensure a normal size distribution with an average size of 450 bp and sequenced on an llumina Sequencer with the following read lengths: 28 bp for read 1, 10 bp for i7 index, 10 bp for i5 index and 90 bp for read 2 at a read depth of 20,000 reads per cell. Five libraries were generated and analysed (Supplementary Table 6).

snRNA-seq analysis

The sequencing files in FASTQ format of each sample were aligned against mouse mm10 genome v4.0.0 and converted to gene count matrices using Cellranger software v7.0.1. Quality control and downstream analysis was performed using Seurat R package v4.3.0. Doublets were detected and removed using R package scDblFinder v1.13.13. Ambient RNA was detected and corrected using R package SoupX v1.6.2. Cells with less than 1000 detected genes, more than 4,000 detected genes, or more than 5% mitochondrial genes were excluded from further analysis. Samples were normalized using Seurat SCTransform function and then integrated and clustered using Seurat functions. Differential gene expression analysis was performed to compare cells from different samples in each cluster. Genes with more than 1.5-fold change and an adjusted P value of less than 0.01 were defined as significant. Co-expression of two genes was assessed by comparing the cosine similarity of the target two genes and random chosen pair of genes. Gene set enrichment analysis was run on each cluster comparing samples with different condition against the Gene Ontology database using R package fgsea v1.20.0. Pathways with less than 0.05 adjusted P value are considered significantly enriched. The pathway analysis was performed using the Reactome Patway database (version 86)58. Well-established markers for brain and blood-derived cell populations were used to define individual clusters59,60.

Gene ontology and interaction network analyses of bulk RNA-seq data

Functional protein association network analysis was performed using Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) database (version 11.0)61 for up-regulated genes. The network was processed applying kmeans clustering, setting cluster number to 4. Inflammatory response-associated genes were identified using Mouse Genome Database (MGI) Batch Query Genome Analysis Tool (v6.23). Selection criteria was classification within the GO term ‘Inflammatory response’ (ID: GO:0006954) using IMP (Inferred from mutant phenotype), IEA (Inferred from electronic annotation) and IBA (Inferred from biological aspect of ancestor) evidence codes. Functional classification of up-regulated genes was performed using PANTHER Classification System engine, version 17.0. Genes were analysed through PANTHER Overrepresentation (test type: Fisher; correction: FDR; functional classification: PANTHER GO-Slim Biological Process, Mus musculus - REFLIST (21997)).

RT2 array

Differential expression of inflammatory response-associated genes between recent and remote memory was analysed using The Mouse Innate and Adaptive Immune Responses RT² Profiler PCR Array (Qiagen, 330231 PAMM-052ZA), according to manufacturer’s instructions. Total RNA was extracted from dorsal hippocampi using PureLink RNA Mini Kit (ThermoFisher, 12183018 A). Tissue was dissected either 96 h (n = 5, pooled) or 21 days (n = 5, pooled) after CFC, resuspended in lysis buffer, flash frozen in liquid nitrogen and stored at −80 °C. RNA was extracted following manufacturer’s instructions. cDNA was synthetized using the RT2 First Strand Kit (Qiagen, 330401). For each experimental group, 500 µg of total RNA was used for the cDNA synthesis. Genomic DNA elimination reaction was prepared by mixing RNA with 2 µl of buffer GE in 10 µl total reaction volume. Real-time PCR reaction was performed using RT2 SYBR Green qPCR Mastermix in duplicates for each sample (recent, remote). The PCR reaction was performed in an Applied Biosystems 7300 Real-Time PCR System with the following cycling conditions: 95 °C, 10 min; then 40 cycles of 95 °C, 15 s; 60 °C, 1 min. ΔCt values for each gene were calculated by subtracting the average Ct value of 6 housekeeping genes from the Ct value of each target gene. Differential expression between recent and remote memory was calculated from ΔΔCt value.

Quantitative PCR analysis

Total RNA was extracted using RNeasy Plus Mini Kit (Qiagen, 74136). Reverse transcription was performed on 100 ng of total RNA PrimeScript RT Reagent Kit (Takara RR037A). Real-time PCR analysis was performed on a QuantStudio 6 Flex instrument (ThermoFisher) using SYBR green detection system (Applied Biosystems, 4367659) and primers specific for Tlr9 (330001/PPM04221A-200), Tlr7 (330001/PPM04208A-200) and Tlr13 (330001/PPM41490A-200) (all from Qiagen). Housekeeping genes Gapdh and Hprt were used for normalization. ΔCt for target genes was calculated using average Ct of the two housekeeping genes. mRNA amount for each experimental group was expressed relative to naive group and calculated as \({2}^{\Delta {C}_{{\rm{t}}}-\Delta {C}_{{{\rm{t}}}_{naive}}}\).

Cytosolic DNA extraction and analysis

Cytosolic DNA extraction was performed as previously described62. DNA was extracted using Mitochondria/Cytosol Fractionation Kit (Abcam, ab65320) from fresh tissue (dorsal hippocampus) and cleared extracts were treated with 1 mg ml−1 proteinase K for 1 h at 55 °C, extracted with phenol:chloroform, treated with RNase A (1 mg ml−1) for 1.5 h at 37 °C, and sequentially extracted with phenol:chloroform and chloroform. Purified cytosolic DNA was tested for nuclear DNA contamination by performing a 40-cycle PCR reaction using primers for vGlut1 (also known as Slc17a7) (forward: GTGGAAGTCCTGGAAACTGC, reverse: ATGAGCGAGGAGAATGTGG). For cloning of dsDNA, samples were treated with DNA polymerase I, large (Klenow) fragment (1 U µg−1 DNA; NEB), supplemented with 33 µM of each dNTP, for 15 min. DNA was precipitated with sodium acetate/ethanol as described above, and dissolved in water. DNA samples were treated with Taq polymerase (NEB) for 20 min, and immediately cloned into pCR4-TOPO vector (Invitrogen), according to manufacturer’s instructions. One Shot competent cells were transformed by adding 2 μl of the TOPO Cloning reaction into a vial of One Shot chemically competent Escherichia coli. Cells were incubated on ice for 30 min, heat-shocked for 30 s at 42 °C without shaking, and immediately transferred to ice. After 5 min, 250 μl of room temperature SOC medium was added, and tubes were placed horizontally in a shaker (200 rpm) at 37 °C for 1 h. Cells were pelleted at 6,000 rpm for 10 min, resuspended in 50 µl of SOC medium, and spread on a pre-warmed selective plate. Plates were incubated at 37 °C overnight. All colonies from each plate were picked and placed into individual wells of a 96-well plate containing 50 µl of PBS. The sequence of cloned DNA fragments was determined by direct colony sequencing (ACGT).

Primary cultures, treatment and live imaging

The hippocampi from post-natal day 0 (P1) C57BL/6 N male and female mice were isolated, and dissociated, as described previously63. Cells were plated in a 14-mm-diameter glass dish (MatTek, P35G-1.5-14-C) coated with poly-d-lysine (Sigma-Aldrich) at a density of 50,000 cells per cm2 and grown in neuronal medium (Neurobasal Medium containing 1 mM GlutaMax, and 2% B27, all from ThermoFisher). Neurons were cultured for 14 days in vitro before treatments. Cells were treated with 25 μM N-methyl-d-aspartate (NMDA) for 10 min in fresh neuronal medium, washed twice, and incubated for 1 h with PicoGreen (dsDNA dye, 1:20,000), CellMAsk (cell membrane dye, 1:1,000), and MitoTracker (mitochondria dye, 1:20,000) diluted in neuronal medium. Cells were imaged on Nikon W1 spinning disc confocal microscope (Center for Advanced Microscopy, Northwestern University), 1 frame per 10 s at 100× magnification.


Mice were anaesthetized with an intraperitoneal injection of 240 mg kg−1 Avertin and transcardially perfused with ice-cold 4% paraformaldehyde in phosphate buffer (pH 7.4, 150 ml per mouse). Brains were removed and post-fixed for 24 h in the same fixative and then immersed for 24 h each in 20% and 30% sucrose in phosphate buffer. Brains were frozen and 50-μm sections were cut for use in free-floating immunohistochemistry17 with the primary and secondary antibodies listed in Supplementary Tables 7 and  8, respectively. In addition to manufacturers’ validation, all primary antibodies were validated by comparison to no primary control samples.

PNN imaging

PNNs were visualized using Wisteria floribunda lectin (WFA) staining, a widely used approach for PNN visualization40. WFA staining was performed according to the manufacturer’s instructions. In brief, endogenous peroxidase was inactivated with hydrogen peroxide. Following streptavidin/biotin and Carbo-Free blocking, sections were incubated with biotinylated WFA (Vector Biolaboratories), Vectastain ABC system, and fluorescein, coumarin or rhodamine isothiocyanate (Akoya Biosciences). Sections were mounted using FluorSave (Millipore-Sigma).

Microscopy, image analysis and quantification

Low-magnification images (up to 60× magnification) were captured with a Leica microscope with a Leica DFC450 C digital camera whereas high-magnification images and z-stacks (60–100× magnification) were captured using a confocal laser-scanning microscope (Olympus Fluoview FV10i). All quantifications were performed with ImageJ. Clusters of γH2AX-positive neurons were first identified, and analyses were performed in the surrounding (~100 µm × 100 µm) area. Thus, the numbers are representative of regions of interest rather than average of the CA1 subfield. For time-course analyses, we counted γH2AX neurons in 180 neurons per mouse (3 consecutive slices of 60 neurons). With 6 mice per group, this amounted to 1,000 neurons per time point. In most of the other molecular targets we counted 60 neurons per mouse using a 60–100× objective. All images were converted to binary format, and for each cell showing γH2AX we determined the background and applied a threshold twice above the background signal. We thus obtained similar results across different antibodies and conditions. All analyses were performed with Fiji/ImageJ. The JACoP Plugin for object-based co-localization was used to determine co-localization by comparing the position of the centroids of the nuclei of the colour channels. Their respective coordinates were then used to define structures separated by distances equal to or below the optical resolution45. Volume and 3D viewer Plugins were used for 3D reconstruction of z-stacks, whereas plot profile and surface plot functions were used for analyses of clusters at lower (40×) magnification. Coloc2 was used to determine correlation of expression levels of different fluorophore signals. The analyse particles, plot profile and measure functions were used to determine the number, size, distribution and distance between indicated signals.

Fluorescent multiplex v2 RNAscope

Naive mice (n = 4) or mice subjected to CFC (n = 4) were perfused 96 h later with ice-cold 0.1 M PBS and 4% paraformaldehyde in PBS and processed as described above. RNAscope was performed according to the manufacturer instructions. To visualize Hsp90b1 and Dcx mRNA and NeuN protein RNAscope Multiplex Fluorescent Reagent Kit v2 (ACD Biotechne, 323100) and RNA–Protein Co-Detection Ancillary kit (ACD Biotechne, 323180) were used. In brief, slides were dried at 60 °C in an oven for 30 min, dehydrated in ethanol, treated with hydrogen peroxide for 10 min at room temperature, washed in water and boiled in co-detection target retrieval reagent (around 98 °C) for 5 min. The sections were incubated with anti-Neun antibody (1:500, Sigma, ABN78), fixed with 10% Neutral Buffered Formalin (VWR, GEN0786-1056), protease plus, and rinsed with sterile water. The hybridization step was performed by incubating the sections with the following probes: Mm-Hsp90b1-C1 (ACD Biotechne, 556051), Mm-DCX-C2 (ACD Biotechne, 478671-C2), Mm-PPIB (positive control probe, ACD Biotechne, 320881) and dabB (negative control probe, ACD Biotechne, 320871) for 2 h at 40 °C and stored overnight in 5× saline sodium citrate. The hybridization was amplified with AMP 1 and AMP 2. All amplification and development were performed at 40 °C, and 2 × 2 min of washes in ACD wash buffer was performed after each step. For C1 probe TSA Vivid Fluorophore Kit 520 (Biotechne, 7523) was used, for C2 and C3, positive and negative probes TSA Vivid Fluorophore Kit 650 (7527) was used. Sections were incubated with goat anti-rabbit Alexa-568 secondary antibody (1:300, Invitrogen, A11036) and DAPI solution (ACD bio), and mounted with Prolong Gold Antifade Mountant (ThermoFisher Scientific, 33342). Images were acquired with an Andor BC43 spinning disk confocal microscope (Oxford Instruments) controlled by the Fusion software from Andor, using a 10× air 0.45 NA objective, a 60× oil immersion 1.42 NA objective, a CMOS camera (6.5 μm pixel; 2,048 × 2,000 pixels generating 16-bit, monochrome images) with no binning. Samples were illuminated with 4 fixed wavelengths of 405 nm, 488 nm, 561 nm and 638 nm. Dorsal hippocampus overview was obtained with 3 × 3 stitching (10% overlap) using the 10× objective and the region of interest (CA1 pyramidal layer, closest to the midline) was afterward imaged at higher magnification using the 60× objective. Image analysis was performed after thresholding using the analyse particle function in ImageJ using four slices per mouse. The number of particles was analysed per 60 neurons per slice (240 neurons per mouse).

Fear conditioning

CFC was performed in an automated system (TSE Systems) as previously described29. In brief, mice were exposed for 3 min to a novel context, followed by a foot shock (2 s, 0.7 mA, constant current). TFC was performed by exposing the mice to for 3 min to a novel context, followed by a 30 s tone (75 dB SPL, 10 kHz, 200 ms pulse), a 15 s trace, and a foot shock (2 s, 0.7 mA, constant current)63. DFC was performed as described for TFC, except that trace was omitted so that shock immediately followed the end of the tone. At indicated time points, mice were tested for memory retrieval by re-exposing them to the same context (context test), or to a tone presented over 30 s in a different context (tone test after TFC or DFC). Testing consisted of 3 min in the conditioning context, during which freezing was measured every 10 s. Freezing was expressed as a percentage of the total number of observations during which the mice were motionless. Activity was recorded automatically by an infrared beam system and expressed in cm s−1. The individual experiments with wild-type mice were not performed on littermates, so we did not apply randomization procedures, but with all genetic lines bred in our facility, littermates were randomly assigned to different experimental groups to minimize litter effects. The behavioural tests were performed blindly, either by experimentalists who were unaware of the treatments because the solutions were coded or by experimenters unaware of the experimental design. The experimenter performing the tests was not aware of the numbering code.

Surgery and cannulation

Double-guided cannulas (Plastic One) were implanted in the dorsal hippocampus as described previously17. Mice were anesthetized with 1.2% tribromoethanol (vol/vol, Avertin) and implanted with bilateral 26-gauge cannulas using a stereotaxic apparatus (Kopf, model 1900). Stereotaxic coordinates for the dorsal hippocampus were 1.8 mm posterior, ±1.0 mm lateral and 2.0 mm ventral to bregma.

Pharmacological treatments

All oligonucleotides and drugs were injected into the dorsal hippocampus at a volume of 0.25 μl per side, at a rate of 0.15 μl min−1 using microinfusion pumps (Model UMP3T-1 UltraMicroPump 3 with SMARTouch Controller). ODN2088 was injected at doses of 125 and 300 ng per mouse corresponding to 4 and 8 nmol, respectively. Based on pilot experiments, the cGAS and STING antagonists were injected at doses of 10 and 50 ng in 250 nl per mouse.

DNase I (Sigma-Aldrich D4513, lot SLCQ3662, diluted in water) was injected intraperitoneally with 50 U DNase I in 200 µl saline, 24 h and 12 h before and 1 h after CFC, or intrahippocampally with 50 U DNase I per mouse at volume of 0.5 μl per side, at a rate of 1 μl min−1 24 h before and 1 h after test. DNase I and S1 nuclease (Thermo Scientific, EN0321) treatment were injected into the dorsal hippocampus at dosage 5 U DNase I + 10U S1 in 20% glycerol/artificial cerebrospinal fluid, vehicle: 20% glycerol/artificial cerebrospinal fluid 4 h before the memory test. Treatment schedules and doses were designed based on published data64 and including both high and low doses as well as pre-CFC and pre-test injections.

For depletion of microglia, Pexidartinib (PLX-3397, HY-16749 lot: 212013 MedChemExpress) was formulated into 5053 PicoLab Rodent Diet 20 at the concentration of 290 ppm (W.F. Fisher and Son). The mice were fed for four weeks before the test. LabDiet 5053 served as a control diet. At the end of experiments, all brains were collected for histological determination of cannula placements or immunohistochemistry.

Virus injections

All viruses (listed in Supplementary Table 9) were injected at a volume of 0.5 μl per side, 1.8 mm posterior, ±1.0 mm lateral and 2.0 mm ventral to bregma, at a rate of 0.15 μl min−1. At the time of virus infusion, mice were eight weeks old. CFC and memory tests were performed from weeks 13 to 17. One day after the completion of behavioural testing, mice were intracardially perfused, and all brains were collected, and virus spread was confirmed by immunohistochemical analysis of GFP or mCherry.

Labelling CFC-activated CA1 neurons using robust activity marking

We labelled the CA1 neurons in the dorsal hippocampus activated during CFC using doxycycline-off (off-Dox) activity-dependent cell tagging with the Robust Activity RAM system31 coupled to the human Fos minimal promoter with four tandem repeats of an enhancer module (PRAM) as recently described in detail11. In brief, we put two groups of wild-type mice on a doxycycline diet one day before injecting AAV2/9-PRAM:d2tTA-TRE:NLS-mKate2 and took them off the diet after 9 days, followed by CFC the day after. Mice were left undisturbed in our testing facility without doxycycline for additional two days. Following two days, mice were put back on the doxycycline diet and 96 h after CFC, one group was subjected to memory reactivation (context test) and euthanized 1 h later. An additional, non-reactivated group served as control. This later group was only used internally to ensure that there was no Fos response without memory reactivation.

Statistics and reproducibility

Statistical power to detect anticipated effect sizes was determined using power analysis (calculator at conducted on representative samples of previous work and pilot experiments. For all proposed experiments, minimum power is set at 0.90 to detect an α = 0.05 (two-sided test) for a difference in means from 20% to 40%, with a 15% common standard deviation. To prevent litter effects, mice from the same litter were assigned to different experimental groups. Viruses were injected by experimenters aware of the construct but the mice were then assigned coded numbers by the laboratory technician. The code was available after quantification and before analyses. Statistical analyses were performed using GraphPad Prism. Mice with misplaced virus infusion or cannulas, and mice with less than 70% viral expression in the CA1 were excluded. One-way ANOVA followed by Tukey’s test was used for post hoc comparisons of three or more experimental groups (only when ANOVA was significant) whereas Student’s t-test was used for comparison of two experimental groups. Homogeneity of variance was confirmed with Levene’s test for equality of variances. On indicated data, we performed correlation analyses and report Pearson’s r coefficients. Significant changes of co-localization or activation (%) were determined using the Chi-square test. A priori determined post hoc analyses following Chi-square tests were performed using Bonferroni-corrected alpha levels as the original overall alpha level (α = 0.05) was divided by the number of tests being conducted. These adjusted alpha levels were used as the new significance threshold for each individual test. All comparisons were conducted using two-tailed tests and the P value for all cases was set to <0.05 for significant differences. Data are expressed as mean ± s.e.m. Statistically significant differences are indicated as *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 and *****P < 0.00001. All cellular and molecular effects were shown in a minimum of six biological replicates (Figs. 1e and 2e and Extended Data Figs. 3f and 9). Two experimental replicates were performed for all time-course, Tlr9-KO, Rela-KO, Ifnar1-KO and WT–cre experiments (Figs. 1c,d, 2a–d and 5a,c and Extended Data Figs. 24, 11 and 12). All significant behavioural effects were replicated at least three times using wild-type, virus-specific and cell-specific control groups. All main gene expression effects were replicated with four different approaches (bulk RNA-seq, quantitative PCR arrays, quantitative PCR and snRNA-seq). Replicates produced similar results relative to initial or representative experiments.


Figures were created and edited using Adobe Illustrator CS6 (Adobe, v27.5, RRID: SCR_010279). Figs. 1a, 1b, 2a, 2e, 3a, 3c, 3d, 4a, and 5a, as well as Extended Data Figs. 4c, 4d, 5a, 5b, 5c, 5d and Supplementary Fig. 1 were created using

Ethics declaration

All animal procedures used in this study were approved by the Northwestern University IACUC, Albert Einstein Medical College IACUC and Danish National Animal Experiment Committee, and complied with federal regulations set forth by the National Institutes of Health.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.