Generalization of negative memories drives the development of psychological distress-related behaviors

Abstract

Individuals with depression and anxiety often exhibit intensified and overgeneralized negative memories. However, the causal contribution of memory features, especially the relative roles of memory strengthening versus generalization in shaping subsequent affective disturbances, remains poorly understood. Here, we developed behavioral paradigms to dissociate memory generalization from strengthening and found that negative experiences leading to memory overgeneralization, but not those enhancing memory strength alone, triggered stress-related behaviors in mice. We identified the projection from medial prefrontal cortex (mPFC) to the bed nucleus of the stria terminalis (BNST) as a key circuit linking generalized memory to stress-related phenotypes. Single-cell calcium imaging revealed a significant overlap between the BNST-projecting mPFC (mPFC^BNST) neuronal ensemble encoding stress-related behaviors and that encoding generalized memory, but not memory strength. Circuit-specific transcriptomic profiling, chromophore-assisted light inactivation, and short hairpin RNA-mediated gene knockdown further demonstrated that actin cytoskeleton remodeling within the mPFC^BNST neurons, a molecular event underlying the formation of memory generalization, is essential for the emergence of stress-related behaviors. Collectively, these findings identify the generalization of negative memories as a principal driver of stress-related behaviors. More broadly, they provide mechanistic insight into how overgeneralized negative memories, a cognitive feature commonly observed in psychological disorders, may contribute to vulnerability to psychological distress, thereby offering a conceptual framework relevant to early identification and intervention in affective disorders.

Graphical abstract

Public summary

• •Novel mouse behavioral paradigms were developed to double-dissociate memory strengthening and generalization.
• •Generalized negative memories selectively triggered psychological distress-related behaviors.
• •Actin remodeling in medial prefrontal cortex to bed nucleus of the stria terminals links memory to distress.
• •Post-encoding modulation of memory generalization altered stress-related phenotypes, minus stress intensity.
• •Rectifying memory overgeneralization may reduce maladaptive affective responses.

Introduction

Depression is a prevalent and debilitating mental disorder.¹ Despite the availability of various antidepressants that can alleviate depressive symptoms to some extent, their efficacy in altering the trajectory of the disorder remains limited, and recurrence rates are high.² Thus, there is an urgent need to elucidate the underlying mechanisms of depression and develop effective strategies capable of slowing or reversing its progression.

A specific cognitive model of depression proposes that negative experiences give rise to maladaptive cognitive biases, which are thought to contribute to vulnerability for depression.³ Accordingly, therapeutic interventions aimed at modifying these cognitive biases hold considerable clinical potential.³ This framework is supported by extensive clinical reports documenting distortions of autobiographical memories among depressed individuals. Specifically, individuals with depression tend to frequently recall vivid and emotionally charged negative memories, which can trigger repetitive negative thinking associated with increased risk for depression.^4,5 Furthermore, they often overgeneralize negative experiences associated with specific contexts into a broader sense of self-referential beliefs, which may disproportionately shape behavior linked to depressive symptoms even in neutral environments.^6,7,8 However, despite these observations, the specific components of negative memories that influence the development of depression and the underlying neurobiological mechanisms remain poorly understood. Indeed, in the field of depression research, memory alterations are often viewed as cognitive consequences of stress rather than as contributing factors to the disorder itself.^9,10,11

Memory, a fundamental cognitive function of the brain, comprises two key components: strength, which refers to the intensity of recalled memories, and generalization, the extension of memories to similar yet distinct situations. Research in the field of memory has predominantly focused on memory strength and its clinical relevance,^{12,13,14,15,16} whereas memory generalization has received comparatively less attention and is often treated as a parametric feature of memory. The impact of negative memory generalization on psychological distress associated with depression and anxiety warrants further investigation. It remains unclear how the strengthening or generalization of negative memories differentially contribute to stress-related behaviors. Identifying the key component that links negative memories to psychological distress is essential for the early detection of at-risk individuals and for determining whether reducing the strength or limiting the generalization of negative memories should be the primary focus of therapies aimed at mitigating the effects of negative cognitive bias.^13,14,15,17

Memory strengthening and generalization are often intertwined, with stronger negative memories typically accompanied by increased generalization.^18,19 However, there is a scarcity of experimental models that can clearly distinguish between these two key components of memory. In this study, we introduced behavioral paradigms that clearly dissociated memory strengthening and generalization. Unexpectedly, our results indicate that the generalization of negative memory, rather than its strengthening alone, drives stress-related behaviors. These findings highlight the importance of rectifying overgeneralization of negative memories and its related mechanisms as potential therapeutic approaches for stress-related neuropsychiatric disorders.

Materials and methods

The detailed methods for mice, behavioral training and testing, stereotaxic surgery, circuit manipulations, immunofluorescence staining and image analysis, single-cell calcium imaging, Retro-TRAP (translating ribosome affinity purification), RNA sequencing, and statistical analyses can be found in method S1.

Results

Generalization of negative memories is associated with stress-related behaviors

To explore whether the strength or generalization of negative memories may contribute to stress-related behavioral alterations, we needed to establish new behavioral paradigms that could distinguish these two aspects of memory. Toward that end, we examined how the number and contextual diversity of negative experiences affected the strength and generalization of memories. We trained mice with single contextual fear conditioning (CFC) training (referred to as “1xTr”), or two CFC training sessions in either the same (“2xTr same”) or different contexts (“2xTr altered”). Each training session consisted of three foot-shocks at 0.75 mA, an intensity within the commonly used range in CFC studies that reliably induces robust memory.^20,21,22 Memory strength was measured in the same context as training, and memory generalization was measured in a modified context that differed substantially from the training context. These memory performance tests were conducted at two time points: 1–2 and 14–15 days after the last training, with a 24 h interval between the two tests at each time point (Figure 1A).

Figure 1.

Stress-related behaviors are associated with the generalization of negative memories

(A) Schematics of contextual fear conditioning (CFC) training (Tr), fear memory strength test (S test) and fear generalization test (G test), novelty-suppressed feeding (NSF) test, splash test (ST), tail suspension (TS) test, social interaction test (SIT), and sucrose preference test (SPT). Naive, control mice without training; 1xTr, single CFC training; 2xTr same, two CFC training trials in the same context; 2xTr altered, two CFC training trials in two different contexts.

(B and C) Quantification of percentage of time spent freezing during the S test and G test (B), and the generalization index (G index) (C). G index: % freezing in the G test divided by % freezing in the S test. n = 9–10 per group.

(D and I) Quantification of latency to eat in the NSF test.

(E and J) Quantification of grooming duration in the ST.

(F and K) Quantification of immobile time in the TS test.

(G and L) Quantification of time exploring the non-social target (NS) and the social mouse (S) in the SIT, and the differential index (DI). DI: time exploring the stranger mouse subtracted by time exploring the toy mouse as a fraction of total exploration time.

(H and M) Quantification of sucrose preference (DI) in the SPT.

(D–H) Tests starting at 1 day after training. (I–M) Tests starting at 2 weeks after training. (D–G) and (I–L) n = 10–12 per group. (H and M) n = 4–8 per group. Data are presented as mean ± SEM and analyzed by one-way ANOVA followed by Turkey’s post hoc test or two-way ANOVA followed the Sidak’s post hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001.

Compared with 1xTr, the 2xTr same paradigm significantly enhanced memory strength in mice 1 day after the final training session, whereas the 2xTr altered paradigm did not further enhance memory strength beyond the 1xTr level (Figure 1B). During the memory generalization test, mice in the 1xTr and 2xTr same groups exhibited similarly low freezing levels, whereas the 2xTr altered paradigm significantly enhanced freezing levels in the memory generalization test (Figure 1B). Since memory strength has an impact on the outcome of memory generalization test, we also calculated the generalization index (G index), which was derived from the ratio of freezing in the generalization test to that in the memory strength test to offer a more robust measure of memory generalization.^19,23 Compared with the 1xTr and 2xTr same groups, the 2xTr altered group also exhibited a significantly increased G index (Figure 1C). Notably, training/testing order and sex had no significant effects on memory performance (Figure S1A). In an independent cohort assessed 2 weeks after training, similar results were obtained (Figure S1B). These results suggest that a second training in the same versus different contexts induced memory strengthening and generalization, respectively, providing a double dissociation between the two components of memory.

We next evaluated the ability of different patterns of CFC training to induce persistent stress-related behavioral alterations. We found that, starting 1 day after training, mice in the 2xTr altered, but not the 1xTr or 2xTr same group, exhibited significant stress-related behaviors across various tests. These included increased latency to eat in the novelty-suppressed feeding test, decreased grooming duration in the splash test, increased immobility in the tail suspension test, decreased social preference in the social interaction test, and decreased sucrose consumption in the sucrose preference test (Figures 1D–1H). Together, these changes suggest altered behavioral responses related to affect and motivation. Notably, these behavioral phenotypes persisted for at least 2 weeks after training, indicating a sustained stress-related phenotype rather than transient responses to conditioning (Figures 1I–1M). Additionally, in a subset of animals that underwent both memory and stress-related behavior assessments, we examined the correlations between memory performances and stress-related behaviors across the 1xTr, 2xTr same, and 2xTr altered groups. Our analysis revealed significant correlations between stress-related behaviors and indicators of memory generalization, including the freezing levels during generalization test and the G index. In contrast, no significant correlation was observed between stress-related behaviors and freezing levels during the memory strength test (Figure S1C).

Taken together, by differentiating between memory strengthening and generalization, our results unexpectedly reveal a selective association between negative memory generalization and stress-related behaviors.

Activation of the mPFC-to-BNST projection promotes the generalization, but not strengthening of negative memories

The unpredictability of negative experiences is recognized as a crucial factor in the onset of depression.²⁴ Yet, the distinctions between memories of unpredictable versus predictable negative events and their impact on stress-related behaviors remain elusive. Given that the 2xTr altered and 2xTr same paradigms selectively induced memory generalization and strengthening, respectively, we hypothesized that memory generalization may serve as a key cognitive process through which unpredictable negative experiences influence stress-related behavioral outcomes.

To identify the neural circuit underpinning memory generalization but not strengthening, we performed a brain-wide c-FOS mapping following distinct CFC training protocols. This approach allowed us to compare the activation patterns of brain networks recruited by training paradigms that induced generalized versus strengthened negative memory (Figure 2A). Since multiple brain regions were activated by CFC trainings, we computed the interregional correlations of the c-FOS⁺ cell density to assess functional connectivity, from which we generated networks of co-regulated brain regions (Figures 2B–2D; Table S1).^25,26 Compared with other conditions, the 2xTr altered condition exhibited a much denser connectivity, indicating increased interregional crosstalk (Figures 2B–2D). By calculating changes in the network global efficiency after isolating each node,^26,27 we identified the bed nucleus of the stria terminalis (BNST) as the top hub for network recruited in the 2xTr altered paradigm that produced generalized CFC memory (Figure 2D). In contrast, the central amygdala (CeA) was identified as the top hub in the networks recruited in both the 1xTr and 2xTr same paradigms (Figures 2B and 2C).

Figure 2.

The mPFC-to-BNST projection is activated and required for negative memory generalization following negative experiences in distinct contexts

(A) Schematic of CFC training and brain collection (sac).

(B–D) Left: network graphs derived from c-FOS mapping, with nodes representing each brain region (see Table S1 for the list) and the edges representing significant correlations (Pearson’s p < 0.05; blue, negative correlation; red, positive correlation). Right: top 5 brain regions contributing to network global efficiency. n = 7–9 per group.

(E) Schematic of the experiment timeline, and cholera toxin B (CTB) injection into the BNST and co-staining with c-FOS. A representative image of CTB infusion into the BNST. Scale bar, 350 μm.

(F) Fold changes in c-FOS expression in CTB⁺ BNST-projecting neurons in upstream regions induced by the G test, normalized to the mean value of the 1xTr group in each region. mPFC, medial prefrontal cortex; ACC, anterior cingulate cortex; LS, lateral septum; vSUB, ventral subiculum.

(G) Representative images of c-FOS staining (gray) in the CTB-labeled BNST-projecting neurons (red) in the mPFC and ACC. Scale bar, 40 μm. n = 9–10 per group.

(H) Schematics of the behavioral paradigm and AAV infusion for circuit-specific chemogenetic manipulation.

(I–K) Left: quantification of percentage of time spent freezing during the memory strength test (S test) and generalization test (G test) with chemogenetic inhibition of different projections to the BNST after the 2xTr altered paradigm. Right: the generalization index (G index), calculated as % freezing during the G test divided by % freezing during the S test. n = 7 (I) (mPFC-to-BNST), 8 (J) (LS-to-BNST), 5 (K) (ACC-to-BNST) per group.

(L) Schematics of the behavioral paradigm and AAV infusion for circuit-specific optogenetic manipulation.

(M) Quantification of the G index during light on/off phases for optogenetic activation of the vSUB-to-BNST projection.

(N) Quantification of percentage of time spent freezing during the S and G tests for optogenetic activation of the vSUB-to-BNST projection. n = 6 per group. Data are presented as mean ± SEM and analyzed by Student’s t tests or two-way ANOVA followed by Sidak’s post hoc test. ∗p < 0.05.

While the amygdala has been the focus in fear research, the BNST is a critical relay station connecting limbic and forebrain structures to the hypothalamic-pituitary-adrenal stress axis.²⁸ We then wondered whether the upstream inputs integrated by BNST prior to regulating negative memory expression were different in the memory strength test versus generalization test after the 2xTr altered paradigm. Combining retrograde tracing with c-FOS staining, our findings revealed that the enhanced memory generalization after the 2xTr altered paradigm was associated with more activated BNST-projecting neurons in the medial prefrontal cortex (mPFC^BNST) and anterior cingulate cortex (ACC^BNST), compared with the 2xTr same paradigm (Figures 2E–2G). Additionally, relative to the 1xTr condition, the 2xTr altered paradigm also produced greater activation of mPFC^BNST, ACC^BNST, and BNST-projecting neurons in the lateral septum (LS^BNST), but reduced activation of BNST-projecting neurons in the ventral subiculum (vSUB^BNST) (Figures S2A–S2D). These patterns of activation were largely specific to BNST-projecting neurons and were not reflected at the level of whole-region c-FOS expression, except in the LS (Figure S2E). In addition, these activation patterns were specific to the generalization test and were not present following the memory strength test (Figures S2F–S2I).

To examine the functional contribution of different upstream inputs to the BNST in memory expression after the 2xTr altered paradigm, we used chemogenetic inhibition to suppress the activity of BNST-projecting neurons in the mPFC, ACC, or LS during memory tests (Figure 2H). Suppression of the mPFC^BNST, but not ACC^BNST or LS^BNST, neuronal activity significantly reduced freezing during the memory generalization test, while leaving performance in the memory strength test unaffected (Figures 2I–2K). Furthermore, consistent with the observation that memory generalization was accompanied by reduced activity in the vSUB^BNST neurons, optogenetic activation of these neurons significantly reduced memory generalization (Figures 2L–2N).

The mPFC is known to play a critical role in both memory and depression.^29,30,31,32 We further examined whether the mPFC-to-BNST projection played a role in generalization of fear memory under other conditions in addition to the 2xTr altered condition. Memory generalization is known to be graded.^33,34 Mice trained with 1xTr showed significantly more freezing in a context similar to the training context (context A′) compared with a context that deviated greatly from training (context C) (Figure 3A). Inhibition of the mPFC^BNST neurons significantly reduced generalization to the similar context (Figures 3B and 3C). In the reverse fashion, following the 1xTr or 2xTr same paradigm in mice, optogenetic activation of the mPFC-to-BNST projection significantly enhanced the expression of fear memory generalization (Figures 3D–3K). Of note, activation of the mPFC-to-BNST projection decreased freezing level during the memory strength test in the training context after 1xTr (Figure 3E), supporting the notion that memory generalization and strength can be oppositely regulated. Furthermore, optogenetic inhibition of mPFC terminals in the BNST during memory tests significantly reduced memory generalization (Figure 3L–3O), reinforcing the critical role of this circuit in mediating generalized memory.

Figure 3.

Activation of the mPFC-to-BNST projection promotes negative memory generalization, but not memory strengthening

(A) Left: schematic of the experimental timeline. Right: quantification of the percentage of time spent freezing during memory tests in the same context as used in training (ctx A), in a context largely deviated from the training context (ctx C), or in a context similar to training (ctx A′). n = 11–13 per group.

(B, D, H, and L) Schematics of AAV infusion, the experimental timeline, and representative images of AAV-mediated fluorescent protein expression in the BNST and mPFC. Scale bar, 300 μm.

(C) Quantification of changes in the percentage of time spent freezing during the G test in context A′ by inhibition of the mPFC-to-BNST projection. n = 7–9 per group.

(E, F, I, J, M, and N) Quantification of percentage of time spent freezing in the S test or G test, during light on/off phases for optogenetic activation of the mPFC-to-BNST projection after 1xTr (E and F) (n = 6–8 per group) and 2xTr same (I and J) (n = 7–9 per group), and for optogenetic inhibition of the mPFC-to-BNST-projection after the 2xTr altered (M and N), (n = 8 per group) paradigm.

(G, K, and O) Quantification of the G index for optogenetic manipulation of the mPFC-to-BNST projection. Data are presented as mean ± SEM and analyzed by one-way ANOVA or two-way ANOVA followed by Sidak’s post hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001.

Altogether, these findings suggest activation of mPFC-to-BNST projection as a key circuit mechanism that preferentially supports the expression of generalized negative memories following negative experiences in distinct contexts.

The mPFC^BNST neuronal ensemble encoding negative memory generalization, but not strength, significantly overlaps with that encoding stress-related behavior

If memory generalization underlies the link between unpredictable negative experiences and stress-related behaviors, we postulated that there might exist overlapping neuronal ensembles encoding memory generalization and stress-related behaviors following unpredictable training. We utilized a miniaturized microscope to visualize Ca²⁺ activity in individual mPFC^BNST neurons in free-moving mice during consecutive behavioral assays, including two CFC trainings in the same or altered contexts, a memory strength test, a memory generalization test, and a tail suspension test to assess immobility, a commonly used measure of stress-related responses in rodents (Figures 4A–4C).

Figure 4.

The mPFC^BNST neuronal ensemble encoding the generalization of negative memory, but not its strength, overlaps with that encoding stress-related behavior following negative experiences in distinct contexts

(A) Schematic of AAV infusion, lens implantation, and the attachment of miniaturized microscope.

(B) Representative images of jGCaMP7s expression and a field of view (FOV) in the mPFC. Scale bars, 300 and 150 μm.

(C) Schematic of two CFC trainings in the same (2xTr same) or different contexts (2xTr altered), memory strength test (S test), generalization test (G test), and tail suspension (TS) test.

(D and E) Distribution (D) and quantification (E) of the overall calcium activity during memory tests after 2xTr same (6 animals; 804 cells for S test, 842 cells for G test) or 2xTr altered (9 animals; 810 cells for S test, 814 cells for G test) training.

(F) Representative correlation graphs during memory tests after 2xTr same (6 animals; 373 cell pairs for S test, 514 cell pairs for G test) or 2xTr altered (9 animals; 616 cell pairs for S test, 749 cell pairs for G test) training.

(G) Quantification of correlated pair ratio and clustering coefficient.

(H) Decoding performance of the random forest classifier trained on real versus shuffled data to predict immobility behavior in the TS test.

(I) Quantification of the importance scores of neuronal activity features derived from the memory generalization versus strength tests for classifier performance.

(J) Classification of mPFC^BNST neurons based on their responses to the onset/offset of freezing during memory tests and immobility during the TS test. Lines indicate the mean value and shades indicate the SEM.

(K and M) The percentages of “strength cells,” “generalization cells,” and “immobility cells” among all recorded cells, and the overlaps between these cell classes. Schematic defining “strength only” (responding to freezing only in the S test), “generalization only” (responding to freezing only in the G test) and “generalization & strength shared” cells (responding to freezing in both memory tests).

(L and N) Histograms showing the null distribution (blue shading) of overlap between “immobility cells” and “strength/generalization cells,” compared with the observed number (red line), after 2xTr same (L), or 2xTr altered (N) training.

Analysis of Ca²⁺ activity during CFC training sessions revealed that a majority of visualized mPFC^BNST neurons were either activated (47%–56%) or inhibited (33%–37%) by foot-shocks (Figure S3). After the 2xTr altered but not 2xTr same paradigm, the mPFC^BNST neurons exhibited significantly higher overall activity during the memory generalization test compared with the memory strength test, suggesting that these neurons developed an increased response to context change following negative experiences occurring in distinct contexts (Figures 4D and 4E). Functional network analysis further revealed significantly increased correlated pair ratio and clustering coefficient during the generalization test when comparing the 2xTr altered group with the 2xTr same group (Figures 4F and 4G), indicating a more synchronized and densely connected ensemble supporting generalized fear expression.

To assess whether and how neuronal activity during memory tests predicts stress-related behavior, we trained random forest classifiers to distinguish between low and high immobility levels in the tail suspension test, using features extracted from Ca²⁺ activity recorded during the memory strength and generalization tests (Figure 4H). Features from the memory generalization test had significantly higher importance scores than those from the strength test (Figure 4I), suggesting that neuronal activity during generalization more strongly predicted stress-related outcomes.

We further aligned the Ca²⁺ activity of mPFC^BNST neurons to the onset or offset of freezing behavior during memory tests and immobility behavior during tail suspension, and then utilized the K-means clustering, an unsupervised machine learning algorithm, to categorize these neurons. We defined the mPFC^BNST neurons that exhibited activity changes in response to freezing during the memory generalization test and strength test as “generalization cells” and “strength cells,” respectively, and those that showed activity changes in response to immobility during tail suspension were referred to as “immobility cells” (Figures 4J and S4). To determine overlaps between these cell populations, we conducted a permutation test and compared the actual number of overlapped cells with the null distribution. The results showed significant overlaps of the “immobility cells” with the “generalization and strength shared cells,” as well as with those identified only as “generalization cells” following the 2xTr altered paradigm, but not the 2xTr same paradigm (Figures 4K–4N). No significant overlap was observed between “immobility cells” and neuron identified solely as “strength cells” (Figures 4K–4N). Instead, there was a trend toward separation between these two neuronal ensembles following the 2xTr same paradigm, which primarily increased memory strength (Figure 4L).

Collectively, these findings suggest that negative experiences in distinct contexts induce an increase in overall mPFC^BNST neuronal activity and connectivity in response to context change. mPFC^BNST neuronal activity during memory generalization is more predictive of stress-related behavior than activity measured during the memory strength test. Critically, there is a significant overlap between the mPFC^BNST neuronal ensemble encoding stress-related behavior and that encoding negative memory generalization.

Connectivity of the mPFC-to-BNST projections

To gain an understanding of the connectivity properties of the mPFC-to-BNST projections, we examined the distribution and cell-type identity of mPFC^BNST neurons. The results showed that, within the mPFC subregions, a substantial proportion of the BNST-projecting neurons were located in the deep layers of the infralimbic (IL) subdivision, and the majority of these neurons were excitatory (Figures 5A–5E). Retrograde trans-monosynaptic tracing with recombinant rabies virus revealed that the mPFC^BNST neurons received monosynaptic inputs primarily from the ACC and the anterodorsal and mediodorsal thalamus, which are involved in information integration (Figures 5F–5H).^35,36

Figure 5.

Connectivity of the mPFC-to-BNST projections

(A and B) Schematic of cholera toxin B (CTB) infusion into the BNST, and representative images of CTB in the BNST and mPFC. Scale bar, 350 μm.

(C) Quantification of CTB-labeled neurons in different subregions as a percentage of total mPFC^BNST neurons. PL: prelimbic cortex, IL: infralimbic cortex.

(D) Representative images of CTB-labeled mPFC^BNST neurons (red) in the mPFC, co-stained with CaMKIIa or GABA (gray). Scale bar, 100 μm.

(E) Quantification of the percentage of CaMKIIa⁺ and GABA⁺ neurons in mPFC^BNST neurons.

(F) Schematics of AAV infusions for labeling inputs to mPFC^BNST neurons, and representative images of mPFC^BNST neurons identified as starter cells co-expressing EGFP and dsRed (yellow). Scale bar, 50 μm.

(G) Representative images of dsRed-labeled input neurons targeting mPFC^BNST neurons. Scale bar, 300 μm.

(H) Quantification of the amount of input cells in each region as a percentage of total input cells. n = 5 per region.

(I and L) Schematics of AAV infusions for labeling ^mPFCBNST neurons.

(J) Representative images of mCherry-labeled ^mPFCBNST neurons from anterior to posterior BNST subregions. Scale bar, 400 μm.

(K) Quantification of the density of mCherry-labeled ^mPFCBNST neurons in different BNST subregions. Ju, juxtacapsular nucleus; fu, fusiform nucleus; al, anterolateral BNST; am, anteromedial BNST; ov, oval nucleus; pr, principle nucleus; if, interfascicular nucleus; tr, transverse nucleus.

(M) Representative images of ^mPFCBNST neurons (red) co-stained with CaMKIIa or GABA (gray). Scale bar, 100 μm.

(N) Quantification of the percentage of CaMKIIa⁺ and GABA⁺ neurons in all mCherry-labeled ^mPFCBNST neurons.

(O) Representative images of mCherry-labeled ^mPFCBNST projections in downstream regions. Scale bar, 300 μm.

(P) Quantification of BNST fiber-covered area as a percentage of the area of each downstream region. n = 4–5 per region. Data are presented as mean ± SEM.

Within the BNST, projections from the mPFC were most enriched in the lateral division, including the juxtacapsular, fusiform, and anterolateral subdivisions, although other BNST subdivisions also received inputs from the mPFC (Figures 5I–5K). The BNST neurons targeted by these mPFC projections (^mPFCBNST) were predominantly inhibitory (Figures 5L–5N). Anterograde trans-monosynaptic tracing with AAV1 showed that the ^mPFCBNST neurons projected to several subcortical regions implicated in motivation and valence processing, including the ventral tegmental area (VTA) and the lateral habenula (LHb) (Figures 5O–5P).

Activation of the mPFC-to-BNST projections promotes stress-related behaviors

We next asked whether manipulations of the mPFC-to-BNST projection, which supported generalization of negative memories but not memory strengthening, after negative experiences could alter stress-related behaviors. We selectively expressed hM4Di in mPFC^BNST neurons and trained mice using the 2xTr altered paradigm (Figure 6A). Suppression of the mPFC^BNST neuronal activity reduced stress-related behaviors, as indicated by a decrease in latency to eat in the novelty-suppressed feeding test, an increase in grooming duration in the splash test, a decrease in immobility in the tail suspension test and an increase in social interaction. These effects were observed both at 1 and 2 weeks after training (Figures 6B–6E and S5A–S5E). Similarly, optogenetic inhibition of mPFC^BNST projection terminals in the BNST also reduced stress-related behaviors induced by the 2xTr altered paradigm at both early and late time points (Figure S6).

Figure 6.

Activation of the mPFC-to-BNST projections promotes stress-related behaviors

(A and F) Schematics of AAV infusion, representative images of mCherry or EYFP expression in the mPFC, and experimental timelines. Scale bar, 500 μm.

(B and G) Quantification of latency to eat in the novelty-suppressed feeding (NSF) test.

(C and H) Quantification of grooming duration in the splash test (ST).

(D and I) Quantification of immobile time in the tail suspension (TS) test.

(E and J) Quantification of exploration time and differential index (DI) in the social interaction test (SIT). (B–E), behavioral changes upon chemogenetic inhibition of the mPFC-to-BNST projections after 2xTr altered. n = 8–11 per group.

(G–J) Behavioral changes upon optogenetic activation of the mPFC-to-BNST projections after 2xTr same. n = 6–9 per group.

(K) Schematic of AAV injection, optogenetic activation, and sample collection (sac).

(L) Fold changes in the density of c-FOS⁺ cells across different downstream regions of the BNST in response to optogenetic activation of the mPFC-to-BNST projection. n = 7–11 per group.

(M) Representative images of c-FOS staining in the periventricular nucleus of hypothalamus (PVN), the lateral habenula (LHb), and the ventral tegmental area (VTA) after optogenetic activation of the mPFC-to-BNST project. Scale bar, 200 μm. Data are presented as mean ± SEM and analyzed by t tests and two-way ANOVA followed by Sidak’s post hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 0.001.

Conversely, we expressed ChR2 in the mPFC^BNST neurons and implanted fibers in the BNST for optogenetic activation of the mPFC-to-BNST projection (Figure 6F). After the 2xTr same paradigm, activation of the mPFC-to-BNST projection enhanced stress-related behaviors both at 1 and 2 weeks after training (Figures 6G–6J and S5F–S5J). Consistent with the behavioral phenotypes, optogenetic activation of the mPFC-to-BNST projection resulted in activation of the LHb and decreased activities in the VTA, patterns consistent with those reported in aversive- or stress-related states in both rodents and humans (Figures 6K–6M).^37,38

Collectively, these findings demonstrate that, at least under the conditions tested here, activity in the mPFC-to-BNST projection promotes both generalization of negative memories and stress-related behaviors, but not memory strengthening.

Activation of a memory generalization-specific consolidation mechanism promotes stress-related behaviors

To further investigate the hypothesis that negative memory generalization can promote stress-related behaviors, we postulated that activation of the consolidation process specific to memory generalization following negative experiences would be sufficient to influence such behavioral outcomes. The consolidation process necessary for the formation of long-term memory involves de novo gene expression.^39,40 To identify how the consolidation mechanisms of memory generalization differ from those supporting memory strengthening, we utilized the Retro-TRAP method to capture transcriptional changes in mPFC^BNST neurons following training.^41,42 We selectively expressed NBL10 (anti-GFP Nanobody-RPL10A fusion) in the mPFC^BNST neurons, and subjected the mice to different patterns of training. One hour after the last training, the mPFC was collected for immunoprecipitation of NBL10-tagged ribosomes, from which the attached RNAs were purified and sequenced (Figure 7A). As expected from prior anatomical tracing data, we observed an enrichment of excitatory neuronal markers and a depletion of inhibitory neuronal and glial markers in the TRAP samples compared with the total inputs, reflecting the selective capture of RNAs from excitatory mPFC^BNST projection neurons by Retro-TRAP. This pattern confirms the cell-type specificity of the Retro-TRAP approach (Figure 7B).^43,44

Figure 7.

Activation of the consolidation mechanism specific to memory generalization promotes stress-related behaviors

(A) Schematic of the Retro-TRAP procedure.

(B) Relative abundance of neuronal and glia marker genes in the translating ribosome affinity-purified (TRAP) RNA compared with input RNA.

(C) Principal-component analysis (PCA) plot of TRAP samples. 1Tr, single CFC training in either context A or B; 2TrS, two CFC trainings in the same context; 2TrA, two CFC trainings in altered contexts.

(D and E) Top 5 enriched Gene Ontology terms for DEGs comparing 2TrA versus 2TrS (D) and 2TrA versus 1Tr (E).

(F) Heatmaps showing the relative expression of DEGs related to the actin cytoskeleton across groups.

(G) Schematic of AAV infusion and optic fiber implantation for inducing actin remodeling in the mPFC^BNST neurons.

(H) Schematic of behavior paradigm and chromophore-assisted light inactivation (CALI) application.

(I and J) Representative images and quantification of phalloidin staining in the experimental group expressing cofilin fused to SuperNova (CFL-SN) and in control mice expressing SuperNova (SN) alone. Scale bar, 50 μm. n = 9 per group.

(K) Quantification of the percentage of time spent freezing during the memory strength test (S test) and the generalization test (G test), and the G index.

(L–O) Quantification of latency to eat in the novelty-suppressed feeding (NSF) test (L), grooming duration in the splash test (ST) (M), immobile time in the tail suspension (TS) test (N), and exploration time and differential index (DI) in the social interaction test (SIT) (O). n = 10–12 per group. Data are presented as mean ± SEM and analyzed by t tests and two-way ANOVA followed by Sidak’s post hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

We first analyzed transcriptional changes in the total mPFC input samples to characterize the broader molecular landscape. Differentially expressed genes (DEGs) comparing 2xTr same versus 1xTr were associated with the consolidation mechanisms linked to memory strengthening, whereas DEGs comparing 2xTr altered versus 1xTr were associated with those linked to memory generalization (Figure S7A–S7C). Both paradigms showed altered expression of immediate-early genes and nuclear receptors relative to 1xTr, consistent with activity-dependent transcriptional changes (Figures S7A–S7D). However, functional enrichment analyses revealed distinct downstream pathways. DEGs in 2xTr altered versus 1xTr were enriched for terms related to postsynaptic membrane and regulation of synapse structure or activity, whereas DEGs in 2xTr same versus 1xTr were enriched for terms related to response to hypoxia and purine nucleoside triphosphate metabolic process (Figures S7E and S7F). Thus, at the whole mPFC level, memory generalization and strengthening recruit distinct transcriptional programs.

To examine molecular changes more precisely within the circuit implicated in memory generalization, we analyzed TRAP-isolated RNAs from the mPFC^BNST neurons. Principal-component analysis revealed that the 2xTr altered group separated clearly from both the 1xTr and 2xTr same groups, whereas the latter two showed partial overlap (Figure 7C). Interestingly, although mPFC^BNST neurons exhibited comparable overall calcium event rates during training across the 2xTr same and 2xTr altered paradigms (Figure S3), transcriptomic profiling revealed markedly different downstream molecular programs. The DEGs identified in the “2xTr same” versus 1xTr comparison showed minimal overlap with those from the “2xTr altered” versus “1xTr” comparison (Figure S7G). Notably, Gene Ontology analysis revealed strong enrichment for terms related to actin cytoskeleton in the 2xTr altered group compared with the other two groups (Figures 7D and 7E). Further analysis of the DEGs related to the actin cytoskeleton showed upregulation of genes involved in actin filament depolymerization and axonal connectivity (Figure 7F).^{45,46,47,48,49,50} In contrast, DEGs in the 2xTr same versus 1xTr comparison were enriched for terms related to extracellular matrix and scavenger receptor activity (Figure S7H). Transcription factor (TF) prediction further supported this distinction: only one TF (ZNF219) was shared between the 2xTr same versus 1xTr and the 2xTr altered versus 1xTr comparisons, and this TF regulated largely divergent gene sets in each comparison (Figures S7I–S7J). Moreover, distinct sets of calcium-responsive TFs linked to largely different downstream targets across the two comparisons (Figure S7K), suggesting that similar calcium activation in mPFC^BNST neurons during training can nonetheless give rise to fundamentally different downstream molecular profiles in a context-dependent manner.

To test whether actin remodeling in mPFC^BNST neurons is sufficient to drive memory generalization, chromophore-assisted light inactivation (CALI) was employed to induce acute actin destabilization in mPFC^BNST neurons specifically during the memory consolidation phase (Figure 7G).^51,52,53 Cofilin fused to the photosensitizer SuperNova was expressed in mPFC^BNST neurons. Illumination at 593 nm was used to activate SuperNova to generate reactive oxygen species, leading to cofilin inactivation and subsequent actin cytoskeleton destabilization.^51,52,53 Mice were trained with the 2xTr same paradigm, which typically results in memory strengthening without generalization. Light stimulation was applied to trigger actin destabilization in the mPFC^BNST neurons 2 min after the second training (Figures 7H–7J). This manipulation altered the paradigm’s outcome, leading to memory generalization while preventing memory strengthening (Figure 7K). Importantly, it also rendered the 2xTr same paradigm capable of inducing stress-related behaviors (Figures 7L–7O). Because CALI was applied only after training, these effects cannot be attributed to altered stress perception during conditioning. These data suggest that perturbing actin cytoskeleton dynamics in mPFC^BNST neurons during consolidation can bias the outcome toward memory generalization and downstream stress-related phenotypes.

Among the DEGs uniquely upregulated in the 2xTr altered group, Gsn (Gelsolin), a potent actin filament severing and capping protein,^54,55 emerged as a top candidate (Figure 8A). Consistent with RNA sequencing data, Gelsolin protein levels were elevated in the somata of mPFC^BNST neurons 1 h after the last training in the 2xTr altered group compared with the 2xTr same group (Figures 8B–8D). To assess its functional relevance, we expressed short hairpin RNA targeting Gsn in mPFC^BNST neurons and trained mice using the 2xTr altered paradigm (Figures 8E and 8F). Projection-specific knockdown of Gelsolin reduced memory generalization and stress-related behaviors, without significantly affecting memory strength (Figures 8G–8P). Mechanistically, Gelsolin knockdown in mPFC^BNST neurons increased the F-actin/G-actin ratio after the 2xTr altered training, confirming reduced actin depolymerization (Figures S8A–S8G). Importantly, CALI-mediated actin destabilization delivered immediately after the second training reversed the behavioral effects caused by Gelsolin knockdown, restoring both memory generalization and stress-related behaviors (Figures S8H–S8N).

Figure 8.

Gelsolin is a key regulator of memory generalization and stress-related behaviors in mPFC^BNST neurons

(A) Volcano plot of differentially expressed genes (DEGs) by the 2xTr altered (2xTrA) versus 2xTr same (2xTrS) paradigms.

(B) Schematics of AAV infusion and the experimental timeline.

(C) Representative images of Gelsolin staining in EYFP-labeled mPFC^BNST neurons. Yellow dashed lines indicate the soma of mPFC^BNST neurons. Scale bar, 50 μm.

(D) Quantification of the fluorescence intensity of Gelsolin staining in EYFP-labeled mPFC^BNST neurons. n = 8–9 per group.

(E) Schematics of AAV infusion and a representative image of AAV-mediated EGFP expression in the mPFC. Scale bar, 500 μm.

(F) Experimental timeline.

(G and L) Quantification of the percentage of time spent freezing during memory tests and the generalization index (G index). G index: % freezing during G test divided by % freezing during S test.

(H and M) Quantification of latency to eat in the novelty-suppressed feeding (NSF) test.

(I and N) Quantification of grooming duration in the splash test (ST).

(J and O) Quantification of immobile time during the tail suspension (TS) test.

(K and P) Quantification of exploration time and differential index (DI) in the social interaction test (SIT). n = 8–11 per group. Data are presented as mean ± SEM and analyzed by t tests and two-way ANOVA followed by Sidak’s post hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

Taken together, these findings demonstrate that actin cytoskeleton remodeling in mPFC^BNST neurons, driven in part by altered expression of Gelsolin, represents an important molecular mechanism underlying the consolidation of negative memory generalization and its downstream behavioral consequences.

Discussion

Negative experiences are well-established external factors in the onset of depression and involve two components: the stress experienced during the events and enduring memories that follow. Diverging from conventional depression research, which has predominantly emphasized the stress component, our study examines the contribution of memory processes to stress-related behavioral outcomes. By introducing behavioral paradigms that dissociate memory strengthening from generalization, our study provides empirical evidence across four levels of analysis (behavioral, neural circuit, neuronal ensemble, and molecular) that collectively demonstrates that, under our experimental conditions, the generalization of negative memories exerts a stronger influence on stress-related behaviors than memory strengthening.

Memories of past experiences are integral to human mental well-being.^12,16 Although the exact causes of depression remain unclear, cognitive theories posit that maladaptive negative memories contribute to persistent negative thought patterns, which influence the onset and maintenance of depressive symptoms. Depressed individuals tend to recall negative personal memories in an overly generalized and highly intense manner.^4,5,8 However, experimental evidence disentangling the relative contributions of memory strengthening versus generalization to psychological distress has been limited, in part because these two memory components are tightly interwined. To address this gap, we developed new mouse behavioral paradigms showing that repeated negative experiences in the same context strengthen memories, whereas negative experiences in distinct contexts enhance generalization. These paradigms establish a framework for probing mechanistic links between memory processes and neuropsychiatric disorders, and provide behavioral evidence that the overgeneralization of negative memories is closely linked to stress-related behaviors.

Our findings align with clinical and animal studies reporting that, as opposed to repeated exposure to the same stressor, exposure to negative experiences occurring in distinct contexts is associated with increased vulnerability to depression.^24,56 Importantly, our study extends these observations by demonstrating that post-training manipulations of a neural circuit promoting memory generalization, without altering the perception of unpredictability or the stress responses during training, were sufficient to modify stress-related behaviors. These results highlight a meaningful contribution of post-encoding memory processes to emotional outcomes, independent of stress intensity during encoding. Specifically, negative experiences encountered across different contexts may bias the consolidation of aversive memories toward overgeneralization, rather than toward intensifying their strength, and this shift in memory quality can promote stress-related behaviors.

This interpretation suggests a potential therapeutic direction: instead of attempting to erase or reframe emotionally charged memories, it may be more effective to reduce overgeneralization and improve the accuracy of past negative memories. Additionally, assessing the precision of negative memories could serve as a cognitive marker for evaluating the risk and severity of depression and for monitoring responses to therapeutic interventions. However, we acknowledge that the behavioral assays of stress-related phenotypes used in this study, while widely employed in preclinical research to assess depression-relevant phenotypes, are limited by their debated translational relevance to human depressive symptoms. Future work should incorporate behavioral assays with clearer clinical correspondence, such as those probing negative bias, reward learning, and other clinically aligned measures, to more directly model depressive processes in animals.^57,58

At a mechanistic level, we provide evidence that memory generalization and strengthening rely on partially dissociable neural circuits. While prior research has made strides in unraveling the regulatory mechanisms underlying memory strength,^40,59 the mechanistic understanding of memory generalization remains limited. Our study, leveraging c-FOS mapping, functional connectivity analysis, circuit tracing and manipulations, uncovers that the BNST, a structure less emphasized in fear memory literature,⁶⁰ serves as a central hub within the brain network engaged by negative experiences occurring across distinct contexts, which promote memory generalization. In contrast, the CeA functions as the core hub for networks activated by experiences that generate precise memories. We further show that activation of the mPFC-to-BNST projection promotes both memory generalization and stress-related behaviors, but not memory strengthening.

Employing single-cell calcium imaging, we show that training in different contexts increases the overall responsiveness and synchronization of mPFC^BNST neurons to contextual changes. Prior work from the hippocampus and amygdala has demonstrated that memory linking or generalization often arises through temporal proximity and excitability-driven co-allocation of neuronal ensembles, whereby events encoded close in time recruit overlapping engrams that promote generalization; such overlap can be further shaped by stress, altered excitability, or time-dependent engram reorganization.^{61,62,63,64,65,66,67} Consistent with these principles, we also observed greater overlap in activated mPFC^BNST neurons when training occurred in different contexts. However, our findings extend this framework in several ways. First, the two training sessions were separated by 24 h, well beyond the classical temporal-linking window, yet generalization still emerged. Second, generalization was not limited to linking the two training episodes, as mice also generalized fear to a third context. Third, mPFC^BNST neurons exhibited increased activity and synchrony during generalization tests only after training in distinct contexts, suggesting that generalization here may rely critically on experience-dependent increases in coordinated ensemble dynamics.

Moreover, mPFC^BNST neuronal activity patterns during the memory generalization test were more predictive of stress-related outcomes than the activity patterns during the memory strength test. Importantly, the mPFC^BNST neuronal ensembles encoding stress-related behaviors showed substantial overlap with those supporting memory generalization, but not with ensembles associated with memory strength. Given the mPFC’s involvement in schema-related memory processing⁶⁸ and the BNST’s role as a relay station regulating hypothalamic-pituitary-adrenal axis activity,²⁸ these findings support a functional model in which the mPFC-to-BNST projection integrates contextual-schema information with downstream stress responses, thereby contributing to maladaptive behaviors. This projection may therefore represent a potential target for neuromodulation therapies aimed at reducing memory overgeneralization and maladaptive stress-related behaviors.^69,70

Our findings, together with other studies indicating complex regulation of memory generalization by BNST, mPFC, and VTA (the latter identified in this study as a downstream target of the mPFC-to-BNST projection), collectively advocate for more comprehensive investigation into the regulation and connectivity of the mPFC-to-BNST circuit across different stages of memory generalization process.^71,72,73,74 It is also important to note that the mPFC comprises subdivisions with functional differences in fear regulation and depression.^32,75 A limitation of our study is the lack of functional differentiation between these subdivisions. Although our AAV injections targeted both the PL (prelimbic cortex) and IL subdivisions of the mPFC, a substantial proportion of BNST-projecting neurons were located within the IL. Further research is needed to elucidate the distinct roles of different mPFC subdivision in linking memory generalization to stress-related behaviors.

Different from traditional depression models using chronic stress exposure,^9,10,76 our study demonstrates that two negative experiences in distinct contexts are sufficient to elicit stress-related behaviors that persist for at least 2 weeks. This observation is consistent with clinical evidence linking acute stressful events to the onset or recurrence of major depression episodes.^77,78 Using Retro-TRAP, we found that memory strengthening and generalization may be governed by distinct molecular pathways. By inducing actin remodeling in mPFC^BNST neurons in a temporally and cell-type-specific manner during memory consolidation, we observed compromised memory quality, reflected by reduced memory strength but increased memory generalization. Notably, this manipulation, which did not alter stress levels during training, was sufficient to induce stress-related behaviors. Among the molecular candidates, Gelsolin, an actin-severing protein, was selectively upregulated after negative experiences in distinct contexts. Projection-specific knockdown of Gelsolin in mPFC^BNST neurons attenuated both memory generalization and stress-related behaviors without affecting memory strength. These findings imply that, under the conditions examined here, the strengthening of negative memories may not be the primary determinant of stress-related behavioral outcomes,^12,14,15,16 whereas the degree of memory generalization emerges as a more prominent contributing factor. They further underscore the importance of memory processes in shaping emotional states and support the hypothesis that enhanced memory generalization can increase vulnerability to psychological distress.

Although this perspective challenges the traditional emphasis on memory strengthening as the primary contributor to stress-related states, out results do not exclude the possibility that memory strengthening may exert behavioral effects under different circuit or behavioral conditions, particularly through the CeA-related pathways. Future studies will be needed to determine under what conditions memory-strengthening circuits contribute to stress-related phenotypes. Finally, our results highlight actin remodeling as a potential molecular target for therapeutic strategies aimed at limiting the overgeneralization of negative memories in neuropsychiatric disorders.

In summary, our findings underscore the need to address the overgeneralization of negative memories in the context of depression interventions. Early efforts in this direction have shown promising outcomes.¹⁷ Future research should prioritize the development of improved psychotherapy protocols, potentially combined with neuromodulation strategies, to achieve effective memory specificity training. By identifying the key underlying neural circuits and molecular mechanisms, our findings lay a foundation for developing novel interventions for the prevention and treatment of depression. Given that the overgeneralization of negative memories is a common cognitive symptom across multiple stress-related neuropsychiatric disorders,^79,80 this study may also offer broader insights into how maladaptive memory processing contributes to psychopathology.

Resource availability

Materials availability

No new materials were generated in this study.

Data and code availability

• •The RNA sequencing data generated for this study have been deposited to the GEO repository under accession number GSE248449.
• •All other data are included in the main text and supplemental materials of this article.
• •Further information of this study and the custom code are available upon reasonable request from the corresponding authors.

Funding and acknowledgments

We thank Dr. Thomas J. Carew at New York University and Dr. Xueyi Shen at University of Edinburgh for helpful comments and editing the manuscript. We thank the Animal Facility and the Core Facility of Medical Science (Guangzhou Campus), Sun Yat-sen University for their help in equipment use and data analysis. We thank all the lab members for discussion and technical assistance during the execution of this project. This work is supported by grants from the STI2030-Major Projects (no. 2021ZD0202000 to X.Y. and Yanni Zeng), National Natural Science Foundation of China (no. 32271068 and 81873797 to X.Y., 81972967 to W.J.-L., and 81971270 to Yanni Zheng), Guangdong Basic and Applied Basic Research Foundation (no. 2024A1515011474 to X.Y. and 2025A1515011369 to W.J.L.), the Science and Technology Planning Project of Guangdong Province (no. 2023B1212060013 and 2020B1212030004 to W.J.L. and 2023B1212060018 to X.Y. and W.J.L.), China Postdoctoral Science Foundation (no. 2025M772577 to X.C.) and Guangdong Project (no. 2019QN01Y202 to X.Y.). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Author contributions

X.Y. and W.J.L. designed the research project. X.C., Y.H., Y. Z., P.Z., Y.C., D.K., S.Z., Y.N. Zheng, Y.W., C.X., and Y.X. carried out the experiments. J.C., X.B., and Yanni Zeng helped analyzed the data. X.Y., W.J.L., Yanni Zeng, X.C., H.Y., Y. Z., and H.Z. wrote the manuscript. All authors contributed to the article and approved the final version.

Declaration of interests

The authors declare no competing interests.

Published Online: February 4, 2026