Behavioral and Molecular Signatures of Social Stress-Escalated Alcohol Drinking
A pose-based resilient/susceptible behavioral phenotype and the single-nucleus transcriptomic signature of the stress-plus-alcohol PVT ensemble
Abstract
Background. Social stress is pervasive in humans and is associated with several neuropsychiatric conditions, including addiction. In both humans and animals repeated exposure to social stress escalates alcohol consumption and increases the risk of relapse to alcohol seeking. However, individuals vary markedly in their susceptibility to these effects, with some developing persistent increases in alcohol intake while others remain resilient. The molecular and behavioral mechanisms that underlie social stress escalated drinking that distinguish susceptible from resilient individuals remain poorly understood.
Methods. Two related studies were carried out using the mouse model of social defeat stress (SDS)-escalated alcohol drinking. In Study 1, C57BL/6J mice (8 control, 8 defeated) were subject to SDS followed by three automated, pose-derived measures: an elevated plus maze (EPM) anxiety score from DeepLabCut pose and a dlc2action head-dip detector, a social investigation score from calibrated three-dimensional dyadic pose estimation from a dyadic social interaction test with an unfamiliar juvenile mouse, and intermittent-access two-bottle-choice alcohol intake. Further, an unsupervised repertoire of dyadic behavioral modules was determined from pose alone for the dyadic social interaction test as an independent test of the resulting phenotype. In Study 2, activity-dependent tagging of stress activated neural ensembles was used to permanently label the SDS-activated ensemble with tdTomato in Fos-2A-Cre:Ai14 mice. Mice were then subject to intermittent access two-bottle choice alcohol consumption. Single-nucleus RNA sequencing (snRNA-seq) of the paraventricular thalamus (PVT) was performed to identify stress and alcohol-induced neuroadaptations in the PVT. We identified stress (tdTomato expression) and alcohol (immediate-early-gene expression) activated neuronal subtypes across the anterior-to-posterior axis of the PVT, and used cell-type-stratified differential expression to identify unique cell-type specific neuroadaptations that occur in cells activated by both SDS and alcohol in the posterior PVT.
Results. In Study 1, SDS led to a trend in increased anxiety-like behaviors and increased social investigation. Further, anxiety and social investigation became tightly coupled only in stressed mice (defeated rho = +0.90). Unbiased k-means clustering of anxiety, social investigation, and alcohol drinking behaviors identified resilient and susceptible subpopulations. Importantly, an independent, unbiased analysis of dyadic social interaction behavior recovered the same resilient and susceptible split, identifying 14 behavioral modules, three of which were significantly altered between the two groups. In Study 2, snRNA-seq resolved five transcriptionally distinct PVT neuronal subtypes (PVT1-5). Differential gene expression (DEG) analysis revealed that the stress and alcohol activated cells in the posterior PVT had the highest number of DEGs thereby identifying this population as being highly vulnerable to stress and alcohol induced neuroadaptations. These DEGs included neuropeptide and secretory genes including Vgf, Scg2, Pcsk1, and Bdnf.
Conclusions. Pose estimation studies revealed unique behavioral signatures of vulnerability and resilience that could be gleaned from EPM and dyadic social interaction behaviors. These signatures could predict escalated alcohol consumption. Our results also demonstrate that stress and alcohol activated overlapping posterior-PVT ensembles that showed a unique transcriptional signature that may drive escalated alcohol intake.
Keywords: paraventricular thalamus, social defeat stress, alcohol use disorder, activity tagging, single-nucleus RNA sequencing, pose estimation, behavioral phenotyping, susceptible, resilient
1 Background and Significance
1.1 Introduction
Stressful events underlie many psychiatric illnesses, and social stress is one of the most common forms (Sinha 2008). Social stress is associated with mood disorders, anxiety, and substance use (Sinha 2001; Keyes et al. 2012). For alcohol in particular, the link between social stress and problem drinking is very well established (Keyes et al. 2012); however, how a history of social stress changes the brain at the level of individual cell types to increase drinking, and why some individuals are vulnerable to the effects of social stress while others are resilient is not well understood. Here we attempt to fill this gap in knowledge by first using high resolution behavioral phenotyping to determine behavioral signatures of resilience and susceptibility to social stress-escalated alcohol consumption. We then examine molecular signatures of stress-escalated alcohol intake in the paraventricular thalamus (PVT).
1.2 Alcohol use disorder, stress, and drinking to cope
AUD is one of the most common and costly substance use disorders in the United States, with a lifetime prevalence near 30% in national survey data (Grant et al. 2015), and a large fraction of cases fit a “drinking-to-cope” profile, in which alcohol is used to manage negative affect (Cooper et al. 1995; Cooper 1994). Further, people who drink to ease anxiety or shame after difficult social experiences are more likely to go on to meet DSM-5 criteria for AUD (American Psychiatric Association 2013) than people who drink for hedonic or social reasons alone, and this risk grows with the severity of the stressor and grows further once coping-motivated drinking has taken hold.
Current AUD pharmacotherapies (naltrexone, acamprosate, disulfiram) target broad reward and aversion circuitry, and they produce only modest effect sizes in unselected AUD populations (Jonas et al. 2014). However, none was designed against the cell-type-specific changes that stress-driven drinking is likely to recruit, and this mismatch is a key reason why studying the molecular machinery in the actual cells engaged by the stress experience is important.
1.3 Psychosocial stress as a risk factor for AUD
Across human studies, stress predicts new drinking problems and relapse in people already in remission (Sinha 2001, 2008; Keyes et al. 2012), and the relationship is bidirectional and dose-dependent, as more severe stress predicts steeper escalation of intake while heavier drinkers report stronger stress reactivity. Sex modulates this trajectory, in that women are more likely than men to report negative affect as a trigger for drinking, show greater HPA reactivity to laboratory stressors, and are over-represented in the drinking-to-cope phenotype (Peltier et al. 2019; Becker and Koob 2016; Becker, McClellan, and Reed 2017). Further, acute laboratory stressors such as the Trier Social Stress Test cause measurable increases in alcohol craving and self-administration in people with AUD (Sinha et al. 2009; Childs, O’Connor, and Wit 2011), which gives a controlled human analog of what longer-term social stress is presumed to do.
Rodent work parallels the human findings, as chronic or repeated social defeat increases voluntary alcohol intake, accelerates transition to compulsive-like drinking, and primes relapse-like behavior after abstinence (Miczek, Yap, and Covington 2008; Hwa et al. 2016; Newman, Albrechet-Souza, et al. 2018; Newman, Leonard, et al. 2018). Both SDS and alcohol consumption cause widespread neuronal activation in the brain (Newman et al. 2019; Vilpoux et al. 2009). SDS activates distinct transcriptional programs in several brain regions (Lorsch et al. 2019). Typically, such transcriptome studies are conducted using RNA extracted from brain homogenates that are comprised of multiple cell types (Lorsch et al. 2019; Bagot et al. 2016). These studies lack cellular resolution as they mask gene expression changes from relevant cell subpopulations that are activated by stress and likely drive neuroadaptive changes in circuit function.
There is also wide individual variability in susceptibility to stress in both human subjects and in animal models (Krishnan et al. 2007; Russo et al. 2012). However, the behavioral mechanisms that underlie the development of resilience and susceptible phenotypes is not well understood in the context of social stress. Here we use machine learning algorithms to perform high resolution behavioral phenotyping to determine behavioral signatures of resilience and susceptibility to social stress-escalated alcohol consumption.
1.4 Modeling social defeat stress (SDS) in mice
SDS is an ethologically relevant form of stress, because the protocol captures the key aspects of a real social threat: territoriality, asymmetric size and aggression, and repeated exposure. The resident-intruder design pairs a larger, aggressive conspecific with a smaller intruder, and it produces stable HPA activation, social avoidance, and altered reward sensitivity that translate well across species (Miczek, Yap, and Covington 2008; Koolhaas et al. 2013; Golden et al. 2011; Newman et al. 2019; Norman et al. 2015).
The episodic protocol used in this thesis was developed in the Miczek lab, and it repeats brief defeat episodes daily for ten days and is well-characterized for producing escalated two-bottle-choice alcohol intake in both sexes, with the resident configuration adapted appropriately for female intruders (Hwa et al. 2016; Norman et al. 2015; Newman, Albrechet-Souza, et al. 2018; Newman et al. 2019). Further, compared to non-social stress models such as restraint or chronic variable stress, SDS more closely mirrors the kind of stressor that, in humans, is most reliably linked to coping-motivated alcohol use. Standard readouts beyond intake include social-interaction tests for the susceptibility/resilience distinction (Krishnan et al. 2007; Russo et al. 2012) and anxiety-like assays such as the elevated plus maze and open field test (Pellow et al. 1985; Walf and Frye 2007).
Previous work performed in the Maiya lab using this model has led to several key discoveries. The Maiya lab has found that repeated social stress leads to robust and persistent escalation in alcohol consumption and reward in both males and females. The lab has also identified a potential role for Dyn/KOR signaling in the dorsal raphe nucleus to bed nucleus of the stria terminalis to basolateral amygdala circuit in regulating stress-escalated alcohol intake (Paliarin et al. 2024). Using a mouse line for activity-dependent genetic labeling of neural ensembles, the lab has also established a functional role for a stress activated neural ensemble in the paraventricular thalamus in regulating social stress-escalated alcohol intake.
1.5 Limits of bulk transcriptomics for stress and alcohol phenotypes
Stress and alcohol each drive transcriptional reprogramming in many brain regions, and much of the existing literature has mapped that reprogramming with bulk RNA-sequencing of dissected tissue. However, the problem with this approach for the present question is direct, because any tissue punch from a region like the PVT contains many cell types: neurons, glia, and endothelial cells, plus several neuronal subtypes with different roles in motivated behavior. Averaging transcripts over that mixture dilutes any signal that lives in a small subpopulation, and by construction the cells activated by both a stressor and alcohol are a small fraction of the total; thus, bulk RNA-seq is biased away from detecting specific adaptations in relevant cell types that most likely drive stress-escalated drinking.
Further, there is growing evidence that drugs of abuse and stressors recruit overlapping programs in shared neuronal populations, the clearest example being the ventral tegmental area, where both repeated social defeat and several drug classes (including alcohol) alter the excitability and reward signaling of the same dopaminergic neurons (Berton et al. 2006; Cao et al. 2010; Chaudhury et al. 2013; Saal et al. 2003; Xiu et al. 2014). For example, both SDS and drugs increase the firing of ventral tegmental area (VTA) dopamine neurons (Saal et al. 2003; Barik et al. 2013). Hence, cells that are stimulated by both stress and alcohol consumption in the same animal are particularly interesting because they may be causally related to escalated drinking. To better understand how SDS leads to escalated alcohol consumption, it is necessary to first understand the cellular and molecular changes that are unique to cells that are activated by both stress and alcohol. In this study, we plan to address this gap in knowledge by profiling the transcriptomes of cells that are activated by both stress and alcohol at single cell resolution in the paraventricular thalamus (PVT).
1.6 The paraventricular thalamus
The PVT sits along the dorsal midline of the thalamus, where it forms a long, thin nucleus that runs across most of the rostro-caudal extent of the thalamic complex, and it is functionally heterogeneous along its anterior-to-posterior axis (Penzo and Gao 2021; Kirouac 2015; Gao et al. 2020). Anterior PVT receives more arousal- and prefrontal-related input, whereas posterior PVT is more closely tied to extended-amygdala and BNST circuitry. The PVT is a hub that collects salience- and arousal-weighted information from cortical, hypothalamic, and brainstem afferents and routes it onward to limbic targets such as the nucleus accumbens, amygdala, BNST, and prefrontal cortex (Li and Kirouac 2008; Penzo et al. 2015; Do-Monte, Quiñones-Laracuente, and Quirk 2015). Those targets govern motivated behavior, and PVT output tracks stimulus salience and balances reward against threat (Choi and McNally 2017; Zhu et al. 2018); thus, the PVT is a well-positioned node for a stress signal to modify reward-related output.
The PVT is almost entirely glutamatergic, and the field has converged on a small set of marker genes that subdivide the population. Drd2, galanin (Gal), and Ppp1r17 each mark partially overlapping subtypes with distinct projection targets. The PVT is awash in neuropeptides. Other neuropeptides expressed in the PVT include Tac2, Nts, Adcyap1, Oprk1/Dyn, and Crh (Curtis, Oakes, and Barson 2021; Gao et al. 2020, 2023). The PVT has also been implicated across stages of alcohol and drug use (James and Dayas 2013; Hartmann and Pleil 2021), peptidergic and orexinergic signaling within the PVT modulates intermittent-access and binge-like consumption (Barson, Ho, and Leibowitz 2015; Matzeu and Martin-Fardon 2020), and a posterior-PVTVGLUT2 projection to the BNST is causally involved in binge alcohol drinking (Levine et al. 2021). Thus, these prior results make PVT an attractive candidate brain region that regulates how stressful experiences are translated into maladaptive behavior (Section 4).
1.7 Activity-dependent genetic access to stress ensembles
The immediate-early gene Fos is the standard marker for neuronal activation (Cruz et al. 2013; Kawashima, Okuno, and Bito 2014), as it is induced within minutes of strong neuronal activation and resolves on a one to two hour timescale at the protein level; however, a transient Fos pulse cannot be followed across days. The Transient Access to Activated Populations (TRAP) mouse line solves this problem by providing permanent genetic access to neurons activated by stimuli (Reijmers et al. 2007; Guenthner et al. 2013; Josselyn and Tonegawa 2020).
In TRAP mice (Guenthner et al. 2013; DeNardo et al. 2019), the Fos locus drives expression of an estrogen-receptor-fused inducible Cre recombinase, and the Cre becomes active only in the nuclei of cells that experience both (i) Fos induction and (ii) 4-hydroxytamoxifen (4-OHT). Active Cre can then be leveraged to drive the permanent expression of fluorescent reporter molecules, chemogenetic, and optogenetic actuators.
Here we used bigenic Fos-2A-Cre × Ai14 (“Fos-2A-Cre:Ai14”) mice, in which the Ai14 floxed-tdTomato reporter is expressed in activated cells in the presence of 4-OHT. The strengths of the system are temporal control of labeling (set by the 4-OHT injection time), permanence (no need to re-tag for downstream experiments), and clean intersectional access for snRNA-seq. However, the known caveats are: (1) sub-saturating capture efficiency, so that tdTomato is a high-specificity but moderate-sensitivity marker of the activated ensemble, and (2) a narrow temporal window that may miss late-recruited cells.
1.8 snRNA-seq analysis of activated populations
Once the stress-activated ensemble is permanently labeled, activated cells can be detected in snRNA-seq datasets by tracking tdTomato expression. Alcohol exposure also leads to immediate early gene expression that can be tracked in snRNA-seq studies. Thus, the bigenic Fos-2A-Cre:Ai14 mice allow us to examine stress- and alcohol-induced neuroadaptations with cell type specificity in behaviorally relevant cell populations within the same animal. Single-cell dissociation of adult brain tissue is known to induce stress-response and immediate-early-gene programs during the dissociation itself, which is an obvious problem when the readout of interest is an IEG-driven population. However, single-nucleus RNA sequencing (snRNA-seq) circumvents this problem by isolating nuclei rather than whole cells, as nuclei tolerate the dissociation steps far better, and yield a transcriptome that more faithfully reflects the in-vivo state (Lacar et al. 2016; Habib et al. 2017).
1.9 Behavioral profiling and individual differences
A consistent finding across SDS-plus-two-bottle-choice cohorts is that the stress effect on drinking is not uniform across controls and stressed mice. Indeed, some stressed mice show large escalations in intake while others don’t. The standard susceptibility/resilience distinction, usually drawn from social-interaction-test data, captures some of this variance but is coarse and assay-specific; however, there is no reason to think the most informative behavioral readouts of a given mouse’s vulnerability are limited to the handful of summary statistics traditionally extracted from each assay (open-arm time, total distance, social-interaction ratio).
Thus, markerless pose estimation gives a high-resolution alternative, as modern tools include DeepLabCut for keypoint tracking (Mathis et al. 2018; Nath et al. 2019), action-detection models built on top of keypoint trajectories, and unsupervised methods such as MoSeq and Keypoint-MoSeq that decompose movement into a data-driven repertoire of recurring behavioral states without hand-defined categories (Wiltschko et al. 2015; Weinreb et al. 2024). Together they produce a frame-by-frame description of locomotion, posture, and ethologically defined actions, and aggregating those features per mouse and relating them to the drinking outcome reframes susceptibility/resilience as a data-driven, multi-dimensional problem rather than a binary split based on a single summary metric (Anderson and Perona 2014; Datta et al. 2019; Pereira, Shaevitz, and Murthy 2020). Thus, the two main questions that are being addressed in Study 1 (Section 3) are: 1) can a resilient/susceptible phenotype be recovered from automated pose-derived behavior, and 2) does an unsupervised behavioral repertoire learned from pose alone independently confirm it?
2 Specific Aims and Overall Project Design
2.1 Project aims
Social defeat stress (SDS) escalates voluntary alcohol consumption, but individual responses to defeat vary widely, and group averages obscure the animals that actually escalate. The goals of this proposal are two-fold: First, we leveraged machine learning strategies to detect behavioral signatures of susceptibility and resilience to social stress. Second, using a tool for activity dependent tagging of neural ensembles, we performed snRNA-seq to determine molecular signatures of social stress escalated alcohol consumption in the PVT.
2.1.1 Study 1: Behavioral phenotyping of stress-escalated drinking
Working hypothesis. The working hypothesis is that social defeat reorganizes affective and social behavior in a way that stratifies animals along a vulnerability axis that is correlated with drinking. We hypothesize that high resolution behavioral phenotyping using tools such as DeepLabCut (DLC) will allow us to identify behavioral signatures of susceptibility and resilience that predicts escalated alcohol intake across our cohort of control and defeated mice.
Aim 1.1. This aim builds an automated, pose-derived behavioral phenotype for each mouse from three measures: an elevated plus maze (EPM) anxiety score (DeepLabCut pose plus a trained dlc2action head-dip detector), a social investigation score derived from calibrated three-dimensional dyadic pose, and intermittent-access two-bottle-choice alcohol intake.
Aim 1.2. This aim uses K-means clustering to determine if clustering mice based on their joint anxiety × social-investigation × drinking score defines a resilient/susceptible split, and characterizes how the two phenotypes differ across independent behavioral readouts.
Aim 1.3. This aim learns an unsupervised repertoire of dyadic behavioral modules from pose alone, and tests whether it recovers the resilient/susceptible phenotype without using anxiety, drinking, or the cluster labels, and whether the modules that mark stress exposure are dissociable from those that mark stress vulnerability.
2.1.2 Study 2: Single-nucleus transcriptomics of the stress-plus-alcohol PVT ensemble
This aim builds on previous work in the Maiya lab demonstrating that both repeated SDS and IA alcohol consumption leads to robust neuronal activation in the PVT. Further, there is a high degree of overlap between stress and alcohol activated cells in the PVT suggesting that stress biases the allocation of the alcohol ensemble in the PVT. Chemogenetic inhibition of stress activated neurons in the PVT significantly reduced alcohol intake suggesting a functional role for stress activated neurons in the PVT in stress-escalated alcohol intake. These findings suggest that repeated SDS recruits a discrete population of PVT neurons that is reactivated during alcohol consumption, and that activity within this ensemble promotes alcohol intake.
Working hypothesis. The advantage of the TRAP model is that it provides permanent genetic access to neurons activated by social stress, allowing us to identify and track these cells in snRNA-seq datasets. Furthermore, because acute alcohol exposure also induces neuronal activation, we can identify alcohol-responsive neurons by quantifying immediate early gene (IEG) expression. The working hypothesis for this project is that cells in the PVT that are activated by both stress and alcohol display a unique cell-type-specific transcriptional signature, distinct from cells activated by stress alone, alcohol alone, or neither, and that this signature underlies social stress-escalated drinking.
Aim 2.1. This aim resolves the cell-type and neuronal-subtype taxonomy of the PVT by snRNA-seq, and places each subtype along the anterior-intermediate-posterior axis.
Aim 2.2. This aim defines the cell-type-specific transcriptional signature of stress- and alcohol-activated PVT neurons by cell-type-stratified differential expression, comparing the stress-plus-alcohol population against stress-only, alcohol-only, and unactivated cells.
2.2 Overall project design
The two studies share the same core paradigm, which was applied to separate cohorts. The behavioral phenotyping experiments used C57BL/6J mice whereas the molecular phenotyping experiments used bigenic Fos2A-Cre:Ai14 mice.
Study 1 used a focused 16-animal cohort (SDS84; 8 control, 8 defeated), in which each mouse contributed one EPM recording, one dyadic social-interaction recording, and a complete drinking record. Here the analysis was deliberately behavioral and individual-level: each animal was summarized with pose-derived scores, animals were clustered in the joint behavioral space, and an unsupervised behavioral repertoire was learned as an independent check on the resulting phenotype.
Study 2 used separate, larger cohorts in which the SDS-activated ensemble was permanently tagged with tdTomato through a single 4-hydroxytamoxifen injection during the final defeat session (TRAP2), and the PVT was profiled by single-nucleus RNA sequencing a few hours into a later alcohol session, so that tdTomato marked the stress ensemble while concurrent immediate-early-gene expression marked the alcohol ensemble (Figure 12). This design allowed the stress-plus-alcohol population to be identified in silico and contrasted against cells activated by stress alone, alcohol alone, or neither.
3 Study 1: Behavioral Phenotyping of Stress-Escalated Drinking
3.1 Introduction
A defining feature of stress-escalated drinking is that the risk is not uniform, because among individuals exposed to comparable social stress only a subset escalate their drinking, while others remain resilient (Section 1.3) (Krishnan et al. 2007; Russo et al. 2012). Our lab is interested in identifying behavioral signatures of stress resilience and susceptibility in our cohort of stressed mice before they are exposed to alcohol. To achieve this, we did high resolution behavioral phenotyping of stressed and control mice as they went through elevated plus maze and dyadic social interaction tests post stress.
These data were then analyzed using markerless pose estimation, which tracks body keypoints from video, and unsupervised methods can then decompose the resulting movement into a repertoire of recurring behavioral states without any pre-defined categories (Weinreb et al. 2024; Datta et al. 2019). In Study 1 this approach was applied to a 16-animal cohort (SDS84), in which each mouse was subject to an elevated plus maze (EPM) anxiety score and, from calibrated three-dimensional pose of a dyadic social-interaction assay with a juvenile intruder, a social investigation score, and both of these were related to that animal’s voluntary alcohol intake. An unsupervised repertoire of dyadic behavioral modules was then estimated directly from the pose, which provides a test of whether a resilient/susceptible phenotype defined on anxiety, social investigation, and drinking can be recovered from behaviors that were never used to define it.
The working hypothesis is that social defeat reorganizes affective and social behavior and stratifies animals along a vulnerability axis that is correlated with subsequent alcohol drinking.
3.2 Materials and Methods
3.2.1 Animals
For all experiments 8-week-old male C57BL/6J purchased from Jackson Labs (Bar Harbor, Maine) mice were used. Mice were singly housed on a 12 h light/dark cycle (lights on at 10 PM and off at 10 AM) with food and water available without restriction, and every procedure was carried out under an approved institutional IACUC protocol. Mice were maintained on Teklad 2019S and switched to LD5053 seven to ten days prior to the start of drinking studies (Zaparte et al. 2024). The behavioral phenotyping cohort reported here comprised 16 mice, 8 non-defeated controls (CTRL) and 8 defeated animals (SDS). Following SDS, mice were subjected to EPM recording, dyadic social interaction with a juvenile male intruder, and a complete two-bottle-choice drinking record.
3.2.2 Social defeat stress paradigm
We performed SDS following a protocol developed by the Miczek lab (Miczek, Yap, and Covington 2008; Norman et al. 2015), in which the singly housed male eight-week-old C57BL/6J mice were randomly assigned to a defeat group or a non-defeat control group at the start of the protocol. Each daily session ran in three phases: a pre-defeat threat period in which the intruder was held behind a perforated, transparent barrier inside the resident’s home cage (sensory exposure without contact), a defeat period of up to five minutes that was capped at thirty attack bites, and a post-defeat threat period of continued sensory-only exposure. One session was run every 24 h for ten consecutive days, and each intruder was paired with a different retired-breeder CFW resident on each session to prevent habituation to a single aggressor. The non-defeat controls went through the same daily handling and enclosure exposure, but with no attacking conspecific present. Following the tenth session, the mice were left undisturbed in their home cages for ten days before two-bottle-choice drinking began. Defeat was confirmed from the attack-count logs, the appearance of social-avoidance behavior in subsequent assays, and the downstream escalation of alcohol intake (Section 3.3.2).
3.2.3 Two-bottle-choice alcohol consumption
Drinking was assessed with a standard intermittent-access (IA) two-bottle-choice schedule, which produces voluntary high-intake drinking in C57BL/6J mice (Simms et al. 2008; Hwa et al. 2011). Mice were given 24 h access to 15% (v/v) ethanol versus water on Monday, Wednesday, and Friday across three weeks, for nine IA sessions total. Bottles were introduced into their home cages three hours into the dark phase, and ethanol was presented at the full 15% concentration from the first session without a graded ramp-up. On alcohol-off days the home cage received two water bottles, and the bottle positions were alternated between sessions to control for side preference. For each session, ethanol intake was computed as grams of ethanol per kilogram of body weight (g/kg), and ethanol preference as the ratio of ethanol-bottle volume to total fluid consumed. Per-mouse drinking was summarized both as the mean intake across the nine sessions and as the Session-1 intake (the first, naive exposure), which is treated separately because it precedes any acquired tolerance or escalation.
3.2.4 Elevated plus maze: apparatus, recording, and anxiety score
Apparatus and recording. Anxiety-like behavior was assessed on a standard elevated plus maze (Pellow et al. 1985; Walf and Frye 2007) (Maze Engineers, IL). The maze was comprised of two open and two closed arms. The arms are 35 cm × 5 cm. The closed arms are surrounded by walls that are 20 cm high. The open arms had a 0.5 cm outer lip. The maze is elevated 60 cm above the floor. Each mouse was recorded for a single five-minute session at 30 frames per second from one overhead camera (2304 × 1728 px), which yielded one top-down video per animal across the 16-mouse cohort.
Pose estimation. Body pose was estimated with DeepLabCut v3 (Mathis et al. 2018; Lauer et al. 2022) (PyTorch backend, DEKR-W32 architecture (Geng et al. 2021)), tracking seven keypoints per mouse: Nose, Head, Neck, BodyCenter, TailBase, LeftEar, and RightEar. The per-frame keypoint likelihoods were retained, and the downstream spatial analyses were gated on a likelihood threshold of 0.6.
Head-dip detection. Head dips, an ethologically relevant EPM risk-assessment behavior (Rodgers and Johnson 1995), were detected from pose with an action-detection model (dlc2action) trained for a single action class. The detected dips were then split by maze zone into Protected (head dip initiated from a closed arm or the center) and Unprotected (head dip from an open arm) subtypes by applying a vectorized point-in-polygon rule to the TailBase and BodyCenter keypoints against five hand-traced regions of interest (two closed arms, center, two open arms), and each dip bout was assigned by a majority vote of its frames, resolving ties to Protected.
Spatial and kinematic summaries. The same pose was also summarized spatially against the maze regions of interest to give per-mouse open-arm time, open-arm entries, total distance, and mean locomotor speed.
EPM anxiety score. Each mouse received a single composite anxiety score built from four EPM metrics that index avoidance and behavioral inhibition. Each component was z-scored across the cohort and sign-aligned so that a higher value always indicates a more anxious animal, and the score was taken as the mean of the four aligned components (Figure 4). The components and their anxiety-increasing direction are: mean locomotor speed (more anxious when lower), time in open arms (lower), open-arm entries (lower), and head-dip duration (lower).
3.2.5 Dyadic social interaction: apparatus, 3-D pose, and social investigation score
Apparatus and recording. Social behavior was assessed in a dyadic social-interaction assay in which the adult experimental (“Focal”) mouse freely interacted with a smaller, novel juvenile social target for 5 minute duration. Each of the 16 sessions was recorded with SqueakShot, a custom three-camera, calibrated and synchronized recording setup (Figure 3), at 56 frames per second for approximately 90 s (about 5,040 frames per session). The same juvenile social targets were reused across multiple sessions; however, because the focal animal is the unit of analysis, the focal contrast is treated as independent in the group statistics, and the per-animal summary tables were structured so that a juvenile-identity random effect can be added in a later mixed-model analysis.
Recording system. Multi-camera video was acquired with SqueakShot, a custom synchronized recording system developed in-house on the Raspberry Pi platform (Doré 2026). Three Raspberry Pi units (one server and two clients), each fitted with a Camera Module 3 (Sony IMX708, wide-field-of-view variant), recorded H.264 video at 56 frames per second (2304 × 1296 sensor mode, 1536 × 864 output, 25 Mbps) with per-frame microsecond timestamps. The units were disciplined to a shared clock over NTP and triggered to open their encoders at the same scheduled instant, and a desktop controller (a Flask web application) drove preview, recording, footage retrieval, MP4 encoding, cross-camera frame matching to within a 20 ms tolerance, and trimming to the interaction window. The three cameras looked down into a custom 3D-printed housing designed to fit the specific mouse cage used for the assay (Allentown PC75JHT; 7.5 × 11.75 × 5 in, 186 × 298 × 128 mm; 484 cm² floor area).
Two-dimensional pose and identity tracking. Multi-animal pose in each camera view was estimated with DeepLabCut (DEKR-W32 architecture) (Mathis et al. 2018; Lauer et al. 2022), tracking the same seven keypoints per animal. Identity swaps between the two animals were resolved with a Viterbi re-tracking pass that re-runs inference on the video and performs a global Focal/Juvenile channel assignment anchored to the filtered production tracks, followed by manual tracklet review where needed.
Three-dimensional reconstruction. The cameras were calibrated with a charuco-board procedure, and the per-frame two-dimensional keypoints were triangulated to three dimensions with anipose-style triangulation (aniposelib) (Karashchuk et al. 2021; Nath et al. 2019), retaining points seen by at least one camera and reconstructing from as few as one view where coverage was sparse. Real-world coordinates are in meters. The residual gaps were gap-filled during a Blender export step so that the analyzed trajectories are continuous across the full session.
Relational kinematics. For every session, dyadic relational kinematics were computed describing how the focal animal behaves relative to the juvenile target, with feature definitions adapted from established three-dimensional social-behavior frameworks (DeepOF (Bordes et al. 2023), the Social Behavior Atlas / SBeA (Han et al. 2024), and DANNCE-based 3-D social pose (Dunn et al. 2021; Klibaite et al. 2025)). These features included the full 7 × 7 inter-keypoint distance matrix and closest body-part approach, nose-to-body investigation (focal nose to the nearest juvenile body part, and the reverse), a focal-nose-to-juvenile-tail-base proximity measure (the “anogenital” investigation metric used in the composite score below), directional orientation and mutual-orientation measures, side-by-side and antiparallel huddle configurations, and behavioral synchrony. The inter-animal distances were normalized to focal body length (the 97.5th-percentile nose-to-tail-base distance) because the adult focal and smaller juvenile differ in size, which would otherwise confound raw distances.
Social investigation score. Each mouse received a single composite social investigation score built from four social metrics, each of which was z-scored across the cohort and sign-aligned so that a higher value always indicates more investigation, and the score was taken as the mean of the four aligned components (Figure 4). The components and their investigation-increasing direction are: body-to-body distance (more investigation when lower, that is, closer), contact time (higher), anogenital (nose-to-tail-base) investigation time (higher), and mean social-zone bout duration (higher).
3.2.6 Data-driven dyadic behavioral modules
To identify recurring patterns of social interaction, we applied an unsupervised machine learning approach to the three-dimensional pose data obtained from the dyadic social interaction sessions. The analysis was performed without using information about experimental group, alcohol drinking, or behavioral scores. For each video frame, the relative positions of the two animals and their movement over time were used to describe their social configuration. Frames from all 16 sessions were pooled and embedded into a two-dimensional behavioral space using UMAP (n_neighbors = 50, min_dist = 0.3) (McInnes, Healy, and Melville 2018), and the embedding was partitioned into discrete behavioral modules by a watershed segmentation of the kernel-density estimate of the embedding (density peaks become module seeds). This procedure yielded 14 dyadic modules (m1 to m14), which were ordered from the closest to the most distant inter-animal geometry and given short interpretive labels from each module’s mean geometry (Table 1). For each session, module occupancy was computed as the percentage of session frames falling in each module; the session is the unit of analysis.
| Module | Short label | Interpretation |
|---|---|---|
| m8 | anogenital | Anogenital investigation (focal to juvenile) |
| m9 | anogen loose | Loose anogenital sniffing (antiparallel) |
| m10 | n2n / sbs | Nose-to-nose plus side-by-side huddle |
| m11 | anogen-anti | Anogenital sniffing (antiparallel) |
| m12 | n2n face-off | Mutual nose-to-nose face-off |
| m13 | mut. anogen | Mutual anogenital (antiparallel) |
| m14 | nose-nose | Nose-to-nose (face-to-face) |
| m4 | following | Following / pursuit |
| m3 | approach | Rapid approach (closing) |
| m1 | turn-away | Disengaging (focal turns away) |
| m2 | leaving | Separating (juvenile leaving) |
| m6 | transitional | Mid-distance reorienting (mixed) |
| m5 | far | Far apart (drifting closer) |
| m7 | apart | Separated / no interaction |
3.2.7 Joint resilient/susceptible phenotype clustering
To define data-integrated phenotypes, unsupervised k-means clustering (k = 2) was applied to each mouse’s three standardized summary axes: EPM anxiety score, social investigation score, and mean ethanol intake across sessions 2–9 (the post-naive sessions, excluding the first-exposure Session 1). This split the cohort into a Susceptible cluster (7 mice) and a Resilient cluster (9 mice). Because the clusters were defined on these three axes, any analyses that contrast the clusters on anxiety, social investigation, or drinking are descriptive; however, the behavioral-module and social-bout contrasts, which use information that does not enter the clustering, are the independent tests of whether the phenotype generalizes.
3.2.8 Statistical analysis
Continuous relationships between the per-mouse axes (anxiety, social investigation, drinking, and individual social metrics) were assessed by Spearman rank correlation (rho). Ethanol intake and preference across sessions were analyzed by two-way repeated-measures (RM) ANOVA with factors stress and session; per-session Control-versus-SDS comparisons were made only where the stress × session interaction was significant, using Sidak-corrected post-hoc tests. All other group contrasts, namely the composite anxiety and social scores (CTRL versus SDS), the resilient/ susceptible phenotype characterization (Resilient versus Susceptible), and per-module occupancy, were assessed by two-sided Mann-Whitney U tests and reported as raw, uncorrected p-values, with significance read at raw p < 0.05. The composite scores were formed by z-scoring and sign-aligning their components and taking the mean. This exploratory threshold is intended to surface candidate behavioral signatures in a 16-animal cohort rather than to provide confirmatory inference.
3.3 Results
3.3.1 Repeated social defeat escalates alcohol intake
Each mouse was summarized by a composite EPM anxiety score and a composite social investigation score (Section 3.2.4, Section 3.2.5; Figure 4). Defeated mice trended toward higher anxiety scores than controls (CTRL −0.25 ± 0.30 versus SDS +0.25 ± 0.22; Mann-Whitney p = 0.328) and higher social investigation scores (CTRL −0.31 ± 0.29 versus SDS +0.31 ± 0.23; p = 0.105). Both scores thus shifted in the expected direction after defeat but, on their own, separated the groups only weakly, which motivates examining how the two measures relate within individuals.
3.3.2 Defeated mice drink more alcohol across sessions
Across the nine intermittent-access sessions, defeated mice drank more ethanol than controls, holding near 10 g/kg while controls stayed near 7 g/kg (Figure 5). A two-way RM ANOVA of alcohol consumption across the nine sessions revealed a strong trend toward a main effect of stress (FStress(1, 14) = 4.55, P = 0.051) but no effect of session or a session × stress interaction. Ethanol preference followed the same pattern, with a comparable group-level trend (FStress(1, 14) = 3.99, P = 0.066). We also found a significant session × stress interaction for alcohol preference (FStress×Session(8, 112) = 2.059, P = 0.0458). Post-hoc Sidak testing revealed that alcohol preference was significantly different between control and stress groups on IA Session 6. However, when we examined drinking data from individual mice, we found that this apparent group-level escalation was carried by a subset of defeated animals rather than a uniform shift (Figure 5), which motivated us to follow up on analyses of individual behaviors. Notably, we also found large individual differences in alcohol consumption in control mice as well. Overall, there was marked individual variability in drinking across control and stressed mice, with stressed mice on average consuming more alcohol and showing increased preference for alcohol than unstressed controls.
3.3.3 The three measures and their pairwise relationships
Elevated plus maze data for each mouse was summarized by an EPM anxiety score. Similarly, dyadic juvenile social interaction data for each mouse was summarized by a social investigation score (Section 3.2.4, Section 3.2.5, Section 3.2.3). The two behavioral scores each relate to drinking, but on different drinking measures. Anxiety was highly correlated with Session-1 intake (Spearman rho = +0.63, p = 0.009), so that more anxious mice drank more on first exposure, whereas social investigation instead tracked mean intake across all nine sessions (rho = +0.57, p = 0.021), so that mice that investigated the juvenile more were the heavier sustained drinkers.
Further, heavier drinkers reorganized their social interaction rather than simply doing more of it, as mean intake correlated negatively with the number of social-zone bouts (rho = −0.66) and positively with mean bout duration (rho = +0.76, p < 0.001). This indicates that heavier drinkers consolidated their social time into fewer but longer bouts of engagement, a signature examined categorically in Section 3.3.6.
3.3.4 Anxiety and social investigation scores are tightly coupled only after defeat
Pooled across all 16 mice, the anxiety and social investigation scores were only weakly correlated (rho = +0.39, not significant; Figure 6). However, splitting by group revealed a strong positive correlation between anxiety and social indices in SDS mice (SDS rho = +0.90, p = 0.002), whereas controls remained flat (CTRL rho = +0.12, not significant; Figure 6). This indicates that social defeat binds the affective (anxiety) and social-investigation dimensions that are independent at baseline, so that the more anxious defeated animals are also the ones investigating the juvenile most.
3.3.5 A joint resilient/susceptible phenotype
Unsupervised k-means (k = 2) applied to the joint anxiety × social-investigation × drinking space split the cohort into a Susceptible cluster (7 mice, mostly defeated) and a Resilient cluster (9 mice; Figure 7). On the three standardized axes, the Susceptible cluster was higher in anxiety, social investigation, and sustained ethanol intake, which indicates a coordinated, multi-measure shift rather than a change in any single behavior (Figure 8).
3.3.6 Characterizing the phenotypes
Across six per-animal readouts spanning social bouts, drinking, and investigation intensity, Susceptible mice had fewer but longer social bouts, drank more, and investigated the juvenile more. Five of the six were significant by two-sided Mann-Whitney U and are shown in Figure 8; the sixth, the percentage of time in the social zone, trended in the same direction (p = 0.071), and all six are tabulated in Table 2. Crucially, the social bout structure and contact metrics are not part of the clustering axes, so their separation is an independent validation of the phenotype rather than a restatement of how it was defined.
| Theme | Metric | Resilient | Susceptible | p |
|---|---|---|---|---|
| Fewer, longer bouts | N social-zone bouts | 10 | 6 | 0.033 |
| Mean bout duration (s) | 6.6 | 13.6 | 0.016 | |
| Drink more | Session-1 ethanol (g/kg) | 6.8 | 9.7 | 0.016 |
| Mean ethanol, S1 to S9 (g/kg) | 7.2 | 10.5 | 0.023 | |
| Investigate more | % time in social zone | 65% | 74% | 0.071 |
| Total contact time (s) | 38.5 | 53.7 | 0.016 |
3.3.7 Behavioral modules independently identify susceptible and resilient animals
To identify recurring patterns of social interaction, we used an unsupervised machine learning approach that analyzed the three-dimensional positions and movements of both animals throughout each social interaction session. Importantly, the analysis was performed without using information about experimental group, alcohol drinking, or any behavioral scores, allowing the behavioral patterns to emerge directly from the pose data (Section 3.2.6). Despite this, differences in module occupancy nevertheless distinguished susceptible and resilient mice across the behavioral embedding (Figure 9). Quantifying occupancy module by module, the clearest discriminator was m4 (following; 9.54% versus 3.06%, p = 0.0003), which Resilient mice occupied far more than Susceptible mice, while Susceptible mice trended toward more m9 (loose anogenital investigation; 6.16% versus 4.41%, p = 0.091; Figure 10). Complete module occupancy data for both group comparisons are provided in Table 3 and Table 4.
We next asked whether the same behavioral modules distinguished mice based simply on stress exposure (Control vs. SDS), rather than stress vulnerability (Resilient vs. Susceptible). In contrast to the vulnerability comparison, Control and SDS mice differed in a distinct set of behavioral modules. SDS mice spent more time in both m13 (mutual anogenital investigation; 4.65% vs. 1.73%, p = 0.038) and m1 (focal turns away / disengaging; 7.90% vs. 3.20%, p = 0.038; Figure 11). Thus, the behavioral modules associated with stress exposure (m13 and m1) were distinct from the module associated with stress vulnerability (m4), which did not itself separate the exposure contrast (CTRL versus SDS p = 0.161). Similar patterns were observed for modules that approached significance (Resilient vs. Susceptible: m9, p = 0.091; Control vs. SDS: m14, nose-nose, p = 0.065; m7, p = 0.093).
| Module | Resilient | Susceptible | p | Sig. |
|---|---|---|---|---|
| m11 anogen-anti | 2.52 | 4.07 | 0.408 | |
| m10 n2n/sbs | 18.80 | 13.82 | 0.211 | |
| m13 mut. anogen | 2.17 | 4.50 | 0.351 | |
| m9 anogen loose | 4.41 | 6.16 | 0.091 | |
| m14 nose-nose | 2.22 | 5.86 | 0.837 | |
| m8 anogenital | 36.02 | 36.49 | 0.918 | |
| m12 n2n face-off | 2.45 | 1.36 | 0.791 | |
| m4 following | 9.54 | 3.06 | 0.0003 | ✓ |
| m1 turn-away | 3.98 | 7.56 | 0.142 | |
| m2 leaving | 1.95 | 2.61 | 0.458 | |
| m6 transitional | 2.88 | 3.78 | 0.408 | |
| m3 approach | 3.48 | 3.42 | 0.758 | |
| m5 far | 2.53 | 2.16 | 0.396 | |
| m7 apart | 7.06 | 5.13 | 0.427 |
| Module | CTRL | SDS | p | Sig. |
|---|---|---|---|---|
| m11 anogen-anti | 4.07 | 2.32 | 0.574 | |
| m10 n2n/sbs | 16.43 | 16.81 | 0.721 | |
| m13 mut. anogen | 1.73 | 4.65 | 0.038 | ✓ |
| m9 anogen loose | 5.00 | 5.35 | 0.721 | |
| m14 nose-nose | 1.24 | 6.38 | 0.065 | |
| m8 anogenital | 38.06 | 34.39 | 0.505 | |
| m12 n2n face-off | 2.80 | 1.15 | 0.248 | |
| m4 following | 7.93 | 5.49 | 0.161 | |
| m1 turn-away | 3.20 | 7.90 | 0.038 | ✓ |
| m2 leaving | 2.03 | 2.45 | 0.713 | |
| m6 transitional | 2.95 | 3.61 | 0.382 | |
| m3 approach | 3.93 | 2.98 | 0.442 | |
| m5 far | 2.14 | 2.60 | 0.792 | |
| m7 apart | 8.50 | 3.92 | 0.093 |
These results demonstrate that stress exposure and stress vulnerability are associated with distinct behavioral modules, indicating that the behaviors identifying mice that experienced social defeat are different from those identifying mice that escalated drinking, even within the same unsupervised behavioral repertoire.
3.3.8 Module phenotype on the embedding
Mapping the phenotype contrast back onto the behavioral embedding revealed the same pattern observed in the per-module analysis. Susceptible mice were enriched in the close-investigation region of the map (m9 and neighboring modules), whereas resilient mice were enriched in the disengagement region (m5 and m7) (Figure 9). For reference, the module that showed the greatest difference between susceptible and resilient mice, the following/pursuit module m4, is shown across the two groups (Figure 10).
3.4 Discussion
Study 1 tested whether resilient and susceptible mice could be identified from automated, pose-derived behavior, and whether this behavioral distinction was separate from simply identifying which animals had experienced social defeat.
First, repeated social defeat escalated alcohol intake and preference, but the escalation was carried by a subset of defeated animals rather than a uniform group shift (Figure 5). This individual variability is the premise of the whole study, because a CTRL-versus-SDS comparison captures the average effect of defeat but not the individual variability in escalation.
Second, anxiety and social investigation are largely independent at baseline but become tightly coupled after defeat (defeated rho = +0.90 versus control rho = +0.12; Figure 6). This suggests that, following social defeat, anxiety and social investigation no longer vary independently but instead become tightly linked, such that the most anxious animals are also the ones that investigate the juvenile the most and ultimately drink the most alcohol. Further, the two behavioral scores relate to drinking on different timescales, with anxiety tracking naive Session-1 intake and social investigation tracking sustained intake, which suggests that an avoidance-like trait and a social-engagement trait contribute to distinct phases of escalation. Defeated mice in our study decrease the number of interaction bouts but paradoxically increase time spent interacting with a novel juvenile mouse. This is somewhat contradictory to published reports (Nestler and Russo 2024) suggesting that repeated SDS leads to reduced social interaction. One caveat of social interaction studies is that the underlying motivation for increased or decreased social interaction cannot be gleaned from behavior alone. Thus, increased social investigation may also reflect vigilance or threat assessment rather than increased sociability. Future studies that directly measure threat assessment and vigilance should be able to disentangle the motivational basis for these behaviors observed in this cohort.
Third, and most importantly, an unsupervised repertoire of dyadic behavioral modules learned from pose alone, without anxiety, sociability, or drinking labels, independently recovered the susceptibility and resilient phenotypes (Figure 9). Resilient mice showed increased occupancy in a following module (m4), the single module that significantly separated the phenotypes, and because this module uses information that never entered the clustering, its separation is a genuine cross-validation rather than a restatement of the cluster definition. The strongest finding of this study is the dissociation between behavioral modules associated with stress exposure (m13 and m1, distinguishing Control from SDS mice) and the module associated with stress vulnerability (m4, distinguishing Resilient from Susceptible mice). This demonstrates that the behavioral signatures of social defeat are distinct from those associated with stress-escalated drinking and that both can be identified from automated pose-derived behavior.
These results should be read against the study’s constraints. The cohort is small (n = 16) and the analyses are exploratory, with raw p-values and no multiple-comparison correction, so the thresholds are meant to surface candidate signatures rather than to provide confirmatory inference. One limitation of this study is that the resilient and susceptible groups were defined using anxiety, social investigation, and alcohol drinking. Therefore, differences between the groups in these measures are expected and primarily describe how the groups were formed. In contrast, the automated behavioral modules and social interaction bout analyses were not used to define the groups, making them independent measures that support the validity of the behavioral phenotype. Another limitation is that the same juvenile stimulus mice were used across multiple testing sessions. Although the focal mouse was the experimental unit in all analyses, future studies should account for juvenile identity as a random effect in mixed-effects models. Finally, the head-dip detector identifies head-dip behavior only, whereas the classification of Protected and Unprotected head dips was assigned after detection based on the location of the animal on the maze rather than being learned directly by the algorithm.
3.5 Conclusions
Overall, automated, pose-derived behavior is sufficient to define an individual-level resilient/susceptible phenotype of stress-escalated drinking, and an unsupervised behavioral repertoire learned from pose alone independently captures it. Together, these findings demonstrate that resilience and susceptibility to stress-escalated drinking are defined by coordinated changes in anxiety-like behavior, social investigation, and alcohol drinking, and that the behavioral signatures of stress exposure are distinct from those associated with stress vulnerability. These results establish automated, pose-derived behavior as a scalable approach for identifying individual differences in vulnerability to stress-escalated drinking. In summary, we identified behavioral signatures that predict an individual’s propensity to escalate alcohol drinking following social stress. Future studies should extend the automated ethogram beyond head-dip behavior, integrate behavioral and molecular datasets at the level of individual animals, and validate these behavioral signatures in independent cohorts (Section 5.5).
4 Study 2: Molecular Signatures of Social Stress Escalated Alcohol Consumption in the PVT
4.1 Introduction
Study 1 examined individual differences in stress-escalated drinking. In Study 2, we pivot from individual differences in behavior to molecular mechanisms by asking whether social defeat and alcohol recruit overlapping ensembles in the paraventricular thalamus (PVT), and what cell-type-specific transcriptional program defines the stress-plus-alcohol activated neurons.
Using bigenic Fos-2A-Cre:Ai14 mice that allow for activity-dependent genetic labeling and a mouse model of stress escalated alcohol consumption, we examined the overlap between cells that are activated by SDS and alcohol consumption in specific brain regions within the same animal. In these mice, the Fos promoter drives an inducible, tamoxifen-gated Cre recombinase such that in the presence of neuronal activity and the synthetic estrogen receptor ligand 4-hydroxytamoxifen (4-OHT), Cre recombinase is expressed in the nuclei of activated cells. In the bigenic mice, this Cre recombinase is used to drive the expression of a tdTomato reporter from the Ai14 allele (Guenthner et al. 2013; DeNardo et al. 2019) (Section 1.7).
Prior work in this model found that repeated SDS escalated drinking and produced robust neural activation across several regions, among which the PVT showed both high stress-tagging density and substantial overlap between stress- and alcohol-activated cells. The PVT is a midline thalamic hub that integrates homeostatic and arousal-related inputs and relays them to limbic targets to direct motivated behavior (Penzo and Gao 2021; Kirouac 2015), and both molecular and circuit evidence implicate it in alcohol and drug use (James and Dayas 2013; Hartmann and Pleil 2021), including a glutamatergic PVT-to-BNST projection that regulates binge drinking (Levine et al. 2021) and a neuropeptide-rich subtype organization (Curtis, Oakes, and Barson 2021; Gao et al. 2023). Prior work also showed that chemogenetic inhibition of the SDS-activated ensemble in the posterior PVT reduced alcohol consumption, which points to a functional role for this ensemble in regulating stress-escalated drinking.
These observations motivate the overarching hypothesis tested here, namely that cells activated by both stress and alcohol carry a unique transcriptional signature, distinct from cells activated by stress or alcohol alone, that underlies social stress-escalated drinking. This hypothesis was tested by activity-tagging the stress ensemble, then exposing the same animals to intermittent-access alcohol and profiling PVT nuclei by single-nucleus RNA sequencing, classifying them by tdTomato (stress) and immediate-early-gene (alcohol) status, and asking which cell types and genes distinguish the stress-plus-alcohol activated population.
4.2 Materials and Methods
4.2.1 Animals and cohorts
Study 2 used the same bigenic Fos-2A-Cre:Ai14 mice (C57BL/6J background, both sexes, singly housed on a 12 h light/dark cycle, IACUC-approved) described in Section 3.2.1, in separate and larger cohorts. The snRNA-seq cohort consisted of 16 pooled samples prepared from 31 mice in total (two to three SDS-exposed mice per sample) to clear the 10x minimum nuclei concentration (Section 4.2.4). Five samples were excluded before analysis because they contained little to no tdTomato signal in the PVT, leaving 11 samples, pooled from 22 mice, for the single-nucleus analysis: 8 from males and 3 from females.
4.2.2 Social defeat and alcohol exposure
Defeat and drinking followed the same resident-intruder and intermittent-access protocols described in Section 3.2.2 and Section 3.2.3. For male intruder cohorts, residents were aggressive, sexually experienced Swiss Webster (CFW) males paired with a CFW female, whereas for female intruder cohorts, residents were CFW females co-housed with a castrated CFW male for two to three weeks beforehand, with the male removed immediately before the session and a female Fos-2A-Cre:Ai14 intruder then introduced (Norman et al. 2015). This female-resident adaptation produces reliable attack behavior where standard same-sex defeat fails. The three-phase daily session (pre-defeat threat, up to five minutes of defeat capped at thirty bites, post-defeat threat), the ten-day schedule with resident rotation, the ten-day incubation period, and the six-session intermittent-access ethanol schedule are as in Study 1 (Figure 12). A parallel sucrose two-bottle-choice cohort was run as a non-drug palatable-fluid control, and elevated plus maze and open field tests were also run in the same animals where logistics allowed.
4.2.3 Activity tagging with Stress-TRAP
SDS-activated cells were permanently labeled using the Fos-2A-Cre × Ai14 (“TRAP”) system (Guenthner et al. 2013; DeNardo et al. 2019). Animals received three intraperitoneal injections of 4-hydroxytamoxifen (4-OHT) at 40 mg/kg on days 8, 9, and 10 of the SDS protocol, 30 minutes before the start of each day’s defeat session, which placed the active-Cre window over the defeat episodes so that tdTomato lineage-traces the cells engaged during the final SDS sessions. This timepoint was chosen from preliminary data showing robust cFos expression in the PVT at this stage of the protocol. Cre-mediated recombination peaks several hours after 4-OHT delivery and is functionally complete within a day, and the resulting tdTomato expression is permanent. After the last SDS session, mice were left undisturbed in their home cages for ten days and were then placed on the intermittent-access alcohol procedure. Animals were sacrificed four hours after the start of their sixth IA alcohol session (roughly three weeks after the 4-OHT injections), so that tdTomato marks the SDS-activated cohort while concurrent cFos protein (for histology) and endogenous IEG mRNAs (for snRNA-seq) mark the cells engaged by the most recent alcohol session (Figure 2). Non-defeated controls received an identical 4-OHT injection in their home cages, which allowed the contribution of handling and 4-OHT alone to be subtracted, and the 40 mg/kg dose was the lowest giving robust on-target labeling with minimal background.
4.2.4 Single-nucleus RNA sequencing
Tissue collection. Brains were removed at sacrifice and flash-frozen on dry ice. Coronal sections at 300 µm were cut on a cryostat over the PVT span (bregma -0.46 to -1.58), and anterior, intermediate, and posterior PVT were taken from those same sections with a 1.5 mm tissue punch (Stoelting), which preserved within-animal anterior-posterior contrasts.
Nuclei preparation and library construction. Nuclei were prepared by our collaborators at The University of Texas at Austin. Yield from bilateral PVT punches was about 200 nuclei/µl per mouse, so tissue from two mice was pooled per sequencing sample to reach roughly 400 nuclei/µl, comfortably above the 200 nuclei/µl minimum of the 10x Chromium platform (Zheng et al. 2017). Libraries were prepared with the 10x Genomics Chromium 3’ Single Cell Gene Expression v3 kit, targeting 5,000 nuclei per library.
Sequencing and alignment. Libraries were sequenced paired-end on an Illumina NovaSeq 6000, and reads were processed through Cell Ranger and 10x Cloud Analysis to produce cell-by-gene count matrices. Nuclei with more than 20% of UMI counts mapped to mitochondrial genes were dropped as compromised.
Sample-level QC outcome. Five samples had essentially no tdTomato signal in the PVT at sequencing, most likely reflecting failed activity-tagging, and these were excluded prior to integration. The integrated object used throughout the downstream analysis is the 11-sample dataset (about 45,000 nuclei).
4.2.5 Computational pipeline
The downstream analysis proceeded in six stages, in which cell-type identity was first established across all nuclei, then narrowed to neurons and the PVT, nuclei were classified by activity status (tdTomato and IEG signal), and finally cell-type-stratified differential expression was run with bootstrap support.
- All-cell object. The merged, SCTransformed (Hafemeister and Satija 2019), Harmony-integrated (Korsunsky et al. 2019) all-nuclei object was clustered in Seurat (Hao et al. 2021) over a resolution sweep (0.2–0.8) on the integrated embedding, and cell-type identities were assigned from canonical markers (nuclei with > 20% mitochondrial UMI excluded at QC).
- Neuronal subset. SCTransform was re-run on the neuron subset, the data were integrated with scVI (Lopez et al. 2018), and tdTomato status was visualized on the integrated UMAP (McInnes, Healy, and Melville 2018).
- PVT subset, anterior-posterior assignment, and neuronal subtypes. PVT cells were carried through SCTransform, PCA, scVI integration, and clustering, and five PVT neuronal subtypes were resolved, with each nucleus assigned to the anterior, intermediate, or posterior PVT from a section-of-origin annotation carried as sample metadata.
- IEG module scoring and tdTomato classification. The IEG panel (Fos, Fosl2, Nr4a1, Nr4a3, Ackr4, Wee1, Jun, Per1, Dusp6, Fosb, Egr1) was scored with
AddModuleScore(Tirosh et al. 2016) against an expression-matched set of control genes, and nuclei were split into IEG-high and IEG-low at the median of the module score. tdTomato-positive nuclei were counted as those with any detected tdTomato transcript, and crossing the two labels classified nuclei as stress-plus-alcohol (tdTomato+ IEG-high), stress-only (tdTomato+ IEG-low), alcohol-only (tdTomato− IEG-high), or neither (tdTomato− IEG-low). - Bootstrap statistics. The activated-cell counts and IEG-positive proportions were bootstrapped to obtain confidence intervals.
- Differential expression by activation status. For each activated population (stress alone, alcohol alone, or stress plus alcohol), differentially expressed genes were identified against the unactivated (neither) population by single-nucleus tests (Seurat
FindMarkers(Hao et al. 2021), MAST (Finak et al. 2015), on the log-normalized data; min.pct = 0.10, log-fold-change threshold 0.50). To guard against the imbalance between the small activated populations and the large control population, each contrast was wrapped in a 1,000-iteration control-cell bootstrap, in which on each iteration the control cells were resampled with replacement down to the size of the activated group, the test was re-run, and per-iteration FDR < 0.05 was applied. A gene was retained as a stable DEG if it was significant in at least 500 of the 1,000 iterations, and percentile (2.5–97.5%) confidence intervals were taken across iterations.
For motif enrichment HOMER was used (Heinz et al. 2010), and the Gene Ontology, Reactome, and KEGG annotations were taken from the same HOMER output.
4.2.6 Statistical analysis
Drinking and histology cohort effects were tested with two-way (mixed or ordinary) ANOVA and Sidak or Holm-Sidak post-hoc correction, and marker genes were identified by Wilcoxon rank-sum tests with FDR correction. For differential expression, single-nucleus MAST tests were used (Finak et al. 2015) (Seurat FindMarkers) with FDR correction, stabilized by a 1,000-iteration control-cell bootstrap with percentile confidence intervals, and genes significant in at least half of the bootstrap iterations were reported as DEGs.
4.3 Results
The Study 2 results move from the whole-brain stress screen that motivates the PVT focus, through the cell-type and subtype taxonomy of the PVT, to the distribution of activated populations and the differential-expression signature of the stress-plus-alcohol ensemble.
4.3.1 SDS activates cFos across multiple brain regions
In a whole-brain screen done in the Maiya lab of cFos expression in defeated mice, several regions showed appreciable cFos labeling, including the paraventricular thalamus (PVT), the paraventricular nucleus of the hypothalamus (PVN), the central amygdala, and the BNST. The PVT and PVN had the highest stress-tagging density, each labeling about 15% of total cells, and both were also reactivated by subsequent alcohol exposure. Direct quantification of cFos-positive cells across seven candidate regions confirmed a robust main effect of stress, with the PVT and PVN showing the largest increases and the only Control-versus-SDS comparisons to survive correction (Figure 13). The PVT and PVN were of particular interest because they are key nodes that integrate homeostatic and arousal-related inputs and relay them to limbic regions to direct motivated behaviors, including drug seeking (Kirouac 2015). This part of the thesis focuses on the molecular signatures of stress- and alcohol-activated cells in the PVT, motivated by the known role of the posterior PVT in alcohol intake and by the PVT-VGLUT2 to BNST circuit (Levine et al. 2021).
4.3.2 snRNA-seq defines major cell types in the PVT
Integrating all eleven samples (8 from males, 3 from females) and clustering the nuclei resolved the expected major cell classes of the PVT (neurons, oligodendrocytes, OPCs, astrocytes, microglia, ependymal cells, endothelial cells, and pericytes), each defined by canonical marker genes (Rbfox3, Mog, Pdgfra, Atp13a4, Cx3cr1, Dcdc2a, Cldn5, Pdgfrb) (Figure 14). The dissection captured the PVT alongside neighboring midline thalamic and habenular nuclei, which separate cleanly in the integrated embedding and allow the PVT to be isolated for the analyses that follow (Figure 15). Each region carries its own marker set, and the PVT was identified by Gck, Lama3, Fam20a, and Col12a1, among others (Table 5).
| Region | Gene | log2FC | pct.1 | pct.2 |
|---|---|---|---|---|
| PVT | Col12a1 | 3.78 | 0.48 | 0.08 |
| Fam20a | 3.41 | 0.63 | 0.09 | |
| Lama3 | 3.24 | 0.35 | 0.04 | |
| Gck | 3.13 | 0.28 | 0.03 | |
| AV | Col27a1 | 5.21 | 0.54 | 0.03 |
| Adgrg6 | 3.98 | 0.29 | 0.03 | |
| Klhl14 | 3.89 | 0.48 | 0.04 | |
| BSTpr | Reln | 3.15 | 0.40 | 0.09 |
| Slc6a1 | 2.91 | 0.40 | 0.09 | |
| CM | Ccdc192 | 5.63 | 0.67 | 0.02 |
| Ctxn3 | 5.09 | 0.50 | 0.02 | |
| HPF | Plekha2 | 5.96 | 0.41 | 0.01 |
| Neurod2 | 5.16 | 0.31 | 0.01 | |
| IMD | Apoe | 3.35 | 0.37 | 0.12 |
| Ttr | 2.88 | 0.33 | 0.08 | |
| LHb | Sox1ot | 4.28 | 0.36 | 0.02 |
| Adra1a | 3.22 | 0.35 | 0.05 | |
| MD | Tspear | 3.80 | 0.45 | 0.05 |
| Prkcd | 3.13 | 0.68 | 0.10 | |
| MHb | Ano1 | 7.04 | 0.88 | 0.02 |
| Tac2 | 6.41 | 0.47 | 0.01 |
4.3.3 Five transcriptionally distinct PVT neuronal subtypes
Subclustering the PVT neuronal nuclei (after scVI integration) resolved five transcriptionally distinct subtypes, PVT-1 through PVT-5 (Figure 16). Each subtype was identified by a characteristic marker, Drd2 (PVT-1), Slit2 (PVT-2), Lamb1 (PVT-3), Ankfn1 (PVT-4), and Adamtsl3 (PVT-5) (Figure 16), with full top-marker sets in Table 6. These subpopulations are consistent with previously described molecular and anatomical PVT subtypes (Curtis, Oakes, and Barson 2021; Gao et al. 2020, 2023; Shima et al. 2023).
| Subtype | Gene | log2FC | pct.1 | pct.2 |
|---|---|---|---|---|
| PVT-1 | Shroom3 | 1.05 | 0.62 | 0.32 |
| Tll1 | 1.01 | 0.70 | 0.36 | |
| Tafa4 | 0.89 | 0.47 | 0.26 | |
| PVT-2 | Slit2 | 2.90 | 0.36 | 0.09 |
| Trpm3 | 2.27 | 0.80 | 0.43 | |
| Cdh8 | 1.48 | 0.85 | 0.54 | |
| PVT-3 | Cadps2 | 4.08 | 0.63 | 0.10 |
| Pbx3 | 3.92 | 0.76 | 0.19 | |
| Lamb1 | 3.42 | 0.37 | 0.06 | |
| PVT-4 | Npsr1 | 2.63 | 0.28 | 0.07 |
| Ankfn1 | 2.36 | 0.36 | 0.13 | |
| Adamts2 | 2.10 | 0.34 | 0.12 | |
| PVT-5 | Spon1 | 1.69 | 0.61 | 0.34 |
| Satb2 | 1.33 | 0.33 | 0.12 | |
| Drd3 | 1.13 | 0.41 | 0.21 |
4.3.4 Anatomical specificity of PVT neuronal subtypes
Splitting the PVT along its anterior-intermediate-posterior axis revealed clear regional structure (Figure 16). Each subregion expresses a distinct marker set, namely Cadps2 and Pbx3 (anterior), Drd3 and Spon1 (intermediate), and Trpm3, Shroom3, and Col12a1 (posterior) (Figure 16; Table 7). These regional markers line up with the subtype markers, with PVT-3 sharing the anterior set, PVT-5 the intermediate set, and PVT-1 and PVT-2 the posterior set, which places the PVT-1 and PVT-2 subtypes in the posterior PVT.
| Region | Gene | log2FC | pct.1 | pct.2 |
|---|---|---|---|---|
| Anterior | Cadps2 | 4.66 | 0.51 | 0.07 |
| Pbx3 | 4.36 | 0.62 | 0.17 | |
| Glis3 | 3.96 | 0.55 | 0.12 | |
| Intermediate | Spon1 | 1.69 | 0.61 | 0.34 |
| Satb2 | 1.33 | 0.33 | 0.12 | |
| Drd3 | 1.13 | 0.41 | 0.21 | |
| Posterior | Trpm3 | 2.69 | 0.64 | 0.30 |
| Shroom3 | 2.39 | 0.59 | 0.18 | |
| Col12a1 | 2.24 | 0.63 | 0.21 |
4.3.5 IEG-score validation and distribution of activated populations across the PVT
The IEG module score was computed with Seurat’s AddModuleScore from a curated set of eleven immediate-early genes whose expression was correlated across nuclei. To confirm that this score reflects genuine neuronal activation rather than an arbitrary threshold, we ran three validation tests (Figure 17). First, we compared the distribution of IEG module scores between nuclei classified as IEG-negative and IEG-positive. As expected, IEG-positive nuclei had substantially higher module scores than IEG-negative nuclei, which confirms that the classification separates low- and high-activity cells.
Next, we asked whether the score reflects coordinated activation of multiple immediate-early genes. For each nucleus we counted how many of the curated IEGs were detectably expressed (0 to 9) and plotted this against that nucleus’s IEG module score. If the score truly captures activation, nuclei expressing more IEGs should have progressively higher scores, which is what we observed. The strong positive correlation (Pearson r = 0.87) shows that the module score scales with the coordinated recruitment of multiple immediate-early genes rather than the expression of any single transcript.
Finally, we asked whether this relationship simply reflects cells with higher overall transcriptional activity. We repeated the analysis with a randomly chosen set of nine genes in place of the curated IEG panel. The number of randomly expressed genes showed essentially no relationship with the IEG module score (Pearson r = −0.03). This negative control demonstrates that the correlation in the previous panel is not a byproduct of cells expressing more genes overall or of differences in sequencing depth, and that the IEG module score specifically reflects coordinated expression of activity-dependent genes.
Classifying PVT nuclei by IEG module score and tdTomato expression yielded the counts in Table 8. IEG-positive (alcohol-activated) nuclei were abundant across all subregions, whereas tdTomato-positive (stress-activated) and double-positive (stress-plus-alcohol) nuclei were rarer, and bootstrap confidence intervals were computed for the activated-cell counts.
AddModuleScore. Regional total counts are verified against the integrated object; the per-region activated subcounts are provisional pending a saved category × position crosstab.
| Region | Total Nuclei | IEG Only | tdTom Only | tdTom + IEG |
|---|---|---|---|---|
| Anterior | 6,209 | 2,247 | 216 | 172 |
| Intermediate | 4,627 | 1,929 | 141 | 132 |
| Posterior | 18,536 | 7,056 | 650 | 531 |
4.3.6 Stress-plus-alcohol cells in the posterior PVT have the most DEGs
For differential expression, each activated population (stress alone, alcohol alone, or stress plus alcohol) was contrasted against cells activated by neither (Figure 18). Within each PVT subregion, the number of differentially expressed genes (DEGs) was highest in the posterior PVT and was driven by the stress-plus-alcohol population. Pooled across the whole PVT, the same ordering held, as the stress-plus-alcohol population carried the most DEGs (257), more than stress alone (179) or alcohol alone (59) (Table 9). These results show that immediate-early genes (Fos, Nr4a1, Nr4a3, Per1) dominate the activated contrasts, alongside the secretory and neurotrophic genes Vgf, Pcsk1, and Bdnf.
| Activated population | Markers | DEGs vs. unactivated |
|---|---|---|
| Stress alone | tdTomato+ | 179 |
| Alcohol alone | IEG+ | 59 |
| Stress + alcohol | tdTomato+ IEG+ | 257 |
4.3.7 Top differentially expressed genes in stress-plus-alcohol cells
The top stress-plus-alcohol DEGs were dominated by neuropeptide and secretory genes (Vgf, Scg2, and Pcsk1), together with the neurotrophin Bdnf and a set of signaling and activity-dependent genes, including the orexin receptor Hcrtr2 and the phosphodiesterase Pde10a (Hcrtr2, Pde10a, Rgs20, Bsg, Hunk, Plk2) (Figure 19). Across the four activation groups these genes were most strongly upregulated in the stress-plus-alcohol cells, and the per-gene contrast against unactivated cells was significant for the leading genes (Figure 20). Gene Ontology analysis of the stress-plus-alcohol DEGs returned enrichment for response to oxidative stress, circadian regulation, glycolysis and glucose catabolism, autophagy, and mitochondrial respiration and NAD metabolism (Figure 21). However, these are transcriptomic findings only, and they were not confirmed by in situ methods in this project.
4.4 Discussion
Here, we combined activity tagging and single-nucleus RNA sequencing to ask whether social defeat stress and alcohol converge on shared PVT cells, and what transcriptional program defines the cells engaged by both.
Our results show that SDS and alcohol reactivate overlapping populations in the PVT. The whole-brain cFos screen identified the PVT (with the PVN) as one of the two regions most strongly engaged by SDS (Figure 13). Further, we found a high degree of overlap between stress and alcohol activated cells. This led us to hypothesize that stress biases PVT neurons to allocate the alcohol ensemble. These findings are consistent with long-standing observations that stressors and drugs of abuse recruit overlapping programs in shared neuronal populations (Berton et al. 2006; Cao et al. 2010; Chaudhury et al. 2013) (Section 1.5).
Second, these data indicate that the convergence is anatomically and transcriptionally specific. The snRNA-seq resolved five PVT neuronal subtypes (Figure 16), and mapping them along the anterior-intermediate-posterior axis placed PVT-1 and PVT-2 in the posterior PVT (Figure 16). The stress-plus-alcohol population in the posterior PVT carried the largest transcriptional response, with more DEGs than either single-activation population (Table 9). This localizes the bulk of the alcohol and stress-induced neuroadaptations to the posterior PVT, consistent with the known role of the posterior PVT in alcohol and stress-induced reward seeking (Levine et al. 2021; Matzeu and Martin-Fardon 2020; Hartmann and Pleil 2021).
Our results also demonstrate that the stress-plus-alcohol signature is dominated by neuropeptide and secretory machinery. Beyond the immediate-early genes expected of recently active neurons (Fos, Nr4a1, Nr4a3, Per1), themselves activity-dependent transcription factors with established roles in plasticity and memory (Hawk and Abel 2011), the leading stress-plus-alcohol DEGs were neuropeptide and secretory genes (Vgf, Scg2, Pcsk1) and the neurotrophin Bdnf (Figure 19). Each of these has independent ties to stress and alcohol, in that Bdnf signaling is required for social-defeat responses in the mesolimbic system and bidirectionally regulates alcohol drinking and anxiety-like behavior (Berton et al. 2006; Pandey et al. 2006; Logrip et al. 2015), Vgf is a stress- and activity-regulated secretory peptide with antidepressant-like actions (Thakker-Varia et al. 2007; Hunsberger et al. 2007), Scg2 belongs to the granin family of regulated neuropeptide precursors (Bartolomucci et al. 2011), and Pcsk1 encodes a proprotein convertase that processes those precursors into active neuropeptides (Seidah and Prat 2012). The signaling hits point the same way, as the orexin-2 receptor Hcrtr2 links the PVT to orexinergic control of alcohol seeking and stress-induced reinstatement (Lawrence et al. 2006; Brown, Khoo, and Lawrence 2013; James et al. 2017; Matzeu and Martin-Fardon 2020), and Pde10a is a striatal-type cyclic-nucleotide phosphodiesterase in motivated-behavior circuits (Menniti, Chappie, and Schmidt 2021), with other phosphodiesterase isoforms now being actively pursued as treatment targets for AUD (Grigsby et al. 2023). Together with the proteasomal (Patrick et al. 2023) and circadian (Spanagel et al. 2005; Logan and McClung 2019) programs suggested by the pathway analysis, this is consistent with the neuropeptide-rich organization of the PVT (Curtis, Oakes, and Barson 2021; Gao et al. 2023), and these data suggest neuropeptide processing, protein turnover, and circadian regulation as candidate mechanisms downstream of combined stress-plus-alcohol activation.
However, these conclusions are bounded by several limitations. Activity tagging captures cells only within a finite temporal window and admits some Cre leak, so tdTomato is a high-specificity but moderate-sensitivity marker of the stress ensemble (Section 1.7). The alcohol label depends on a module score of immediate-early genes split at its median, which is a sensitive but unselective threshold, and a stricter cut would trade sensitivity for specificity and is worth exploring. The differential-expression tests operate at the level of the nucleus rather than the animal, and the bootstrap stabilizes against the activated-versus-control imbalance but does not substitute for biological replication across subjects. The data represent a single time point, and although sex was included, the design is underpowered to stratify on it fully, while the per-region activated subcounts in Table 8 remain to be locked down against a saved crosstab. Finally, and most importantly, the DEGs are transcriptomic predictions, and none of them were confirmed with in situ methods (FISH, IHC, or RNAscope) in this project.
4.5 Conclusions
Overall, these results show that stress and alcohol converge on overlapping ensembles in the posterior PVT, and that the cells activated by both carry the largest and most distinctive transcriptional response, dominated by neuropeptide and secretory genes. The effect localizes to the PVT-1 and PVT-2 subtypes of the posterior PVT, which identifies these populations as a focal point for social stress-escalated drinking and nominates neuropeptide processing, proteasomal degradation, and circadian regulation as candidate downstream programs. Thus, validating these transcriptional predictions in situ, and testing the causal necessity of the posterior-PVT stress-plus-alcohol population, are the central next steps (Section 5.5).
5 General Discussion and Future Directions
5.1 Introduction
This thesis presented two studies using a mouse model of social defeat stress (SDS)-escalated alcohol drinking. Study 1 (Section 3) asked whether automated, pose-derived behavior can sort susceptible from resilient drinkers within a stressed cohort, while Study 2 (Section 4) asked which cells in the paraventricular thalamus (PVT) are engaged by both stress and alcohol, and what transcriptional program defines them.
5.2 General Discussion
5.2.1 Study 1: behavioral phenotyping
Study 1 showed that a resilient/susceptible split can be recovered from automated, pose-derived behavior. These results show that repeated defeat escalated drinking, but the escalation was carried by a subset of animals rather than a uniform shift, so the question of who escalates is not answered by the CTRL-versus-SDS comparison. Further, anxiety and social investigation, which were largely independent at baseline, became tightly coupled after defeat and related to drinking on different timescales, with anxiety tracking naive first-exposure intake while social investigation tracked sustained intake. Most importantly, an unsupervised repertoire of dyadic behavioral modules learned from pose alone, without anxiety, drinking, or the cluster labels, independently recovered the phenotype, and the modules that marked stress exposure (who was defeated) were dissociable from the module that marked stress vulnerability (who escalated). These data demonstrate that automated pose can read out both the stress exposure and the individual’s vulnerability, as separate behavioral signatures.
5.2.2 Study 2: PVT mechanisms
Study 2 showed that SDS and alcohol reactivate overlapping tdTomato-labeled populations in the PVT, which supports the hypothesis that stress and drinking share a PVT representation. Using snRNA-seq, the analysis resolved five PVT neuronal subtypes and placed PVT-1 and PVT-2 in the posterior PVT, where the cells activated by both stress and alcohol carried the largest transcriptional response. Earlier material had described an anterior-versus-posterior dissociation; that framing traced to a swapped anterior/posterior label, and it is replaced here with localization to the posterior PVT, consistent with the known role of the posterior PVT in alcohol and stress-induced reward seeking (Levine et al. 2021; Matzeu and Martin-Fardon 2020; Hartmann and Pleil 2021). The stress-plus-alcohol signature was dominated by neuropeptide and secretory genes (Vgf, Scg2, Pcsk1) and the neurotrophin Bdnf (Berton et al. 2006; Logrip et al. 2015; Thakker-Varia et al. 2007), which nominates neuropeptide processing, proteasomal degradation (Patrick et al. 2023), and circadian regulation (Spanagel et al. 2005) as candidate programs downstream of ensemble activation.
5.3 Limitations
Several limitations bound the Study 1 findings. The cohort is small (n = 16) and the analyses are exploratory, with raw p-values and no multiple-comparison correction, and because the resilient/susceptible clusters were defined on the anxiety × social × drinking space, the contrasts that reuse those axes are descriptive by construction, while the module and social-bout contrasts carry the evidential weight. Further, the learned model currently detects a single action class (head dips), with stretch-attend postures, locomotion, grooming, and rearing still to be added, and the Protected/Unprotected split is rule-based on zone occupancy rather than learned end-to-end. In addition, the juvenile social targets were reused across sessions, and the correlational link between behavior and drinking does not establish causality.
The Study 2 pipeline carries its own constraints. Activity tagging captures cells only within a finite temporal window and admits some Cre leak; the alcohol label is a median split of an immediate-early-gene module score, which is sensitive but unselective; the differential-expression tests are at the level of the nucleus rather than the animal, so the bootstrap stabilizes against the activated-versus-control imbalance but does not substitute for subject-level replication; the data represent a single time point; sex was included but the design is underpowered to stratify on it fully; and the snRNA-seq-derived DEGs were not confirmed with post-hoc FISH, IHC, or RNAscope in this project.
5.4 Overall Conclusions and Translational Impact
Overall, these results show that stress and alcohol converge on overlapping ensembles in the posterior PVT, and the cell-type-specific transcriptomic signatures resolved here, concentrated in the PVT-1 and PVT-2 subtypes, nominate neuropeptide, proteasomal, and circadian programs relevant to stress-driven AUD vulnerability. In parallel, the pose-based behavioral profiling offers a scalable path to individual-level classification of susceptible and resilient stress-escalated drinking. These data provide the basis for a twofold translational promise, in that the molecular signature points to candidate pathways and cell types for mechanistic and, eventually, therapeutic targeting of stress-escalated drinking, while the behavioral phenotyping offers an objective, high-dimensional stratification tool that could improve the design of mechanistic studies and inform biomarker discovery in human AUD subtypes.
5.5 Future Directions
Several concrete extensions follow from this work:
- Study 1. Future work will expand the learned dlc2action ethogram beyond head dips (stretch-attend postures, locomotion, grooming, rearing) on the full cohort and held-out validation videos; integrate complementary pose pipelines (Keypoint-MoSeq, multi-camera dyadic pose) as additional feature dimensions; and prospectively validate the susceptible/resilient classifier on independent cohorts. We will also examine how circuit manipulations and drugs that attenuate stress-escalated drinking affect behavioral signatures of susceptibility.
- Study 2. Future work will confirm the leading stress-plus-alcohol transcripts (e.g. Vgf, Bdnf, Pcsk1) in tdTomato/IEG-labeled posterior-PVT cells by multiplexed RNAscope or immunohistochemistry; tighten the IEG threshold and test the role of specific DEGs and molecular pathways in the posterior PVT in social stress escalated alcohol consumption.
- Study 1 × Study 2. Future work will perform snRNA-seq or TRAP-based profiling in behaviorally defined susceptible versus resilient mice, matching animal identities across assays, so as to test directly whether the posterior-PVT stress-plus-alcohol signature is stronger in the animals that the behavioral phenotype labels as susceptible.
Acknowledgments
I would like to thank my mentor, Dr. Rajani Maiya, for her guidance and support throughout this work. I am grateful to the members of the Maiya Lab and to our collaborators for their help with the behavioral, histological, and sequencing work that this thesis is built on, and to my family and friends for their support along the way.
Maiya Lab. Rajani Maiya, Fran Paliarin, Chelsea Duplantis, Samhita Basavanhalli, Emily Blaze, Emily Garcia
Collaborators.
- Mayfield Lab: Dayne Mayfield, Nihal Salem, Anna Warden
- Messing Lab: Bob Messing
- Gilpin Lab: Nick Gilpin, Andrea Jones
- Wills Lab: Tiffany Wills, Lucas Albrechet-Souza
- Salling Lab: Mike Salling
- Edwards Lab: Jessi Cucinello-Ragland
Funding.
- NIAAA R21AA030652 (RM)
- NIAAA R01AA031733 (RM)
List of Abbreviations
| Abbreviation | Definition |
|---|---|
| 4-OHT | 4-hydroxytamoxifen |
| ANOVA | Analysis of variance |
| AUD | Alcohol use disorder |
| BNST | Bed nucleus of the stria terminalis |
| cFos | c-Fos protein (immediate-early gene product) |
| DEG | Differentially expressed gene |
| DEKR | Disentangled keypoint regression (DeepLabCut architecture) |
| DLC | DeepLabCut |
| DSM-5 | Diagnostic and Statistical Manual of Mental Disorders, 5th ed. |
| EPM | Elevated plus maze |
| FDR | False discovery rate |
| g/kg | Grams of ethanol per kilogram body weight |
| GO | Gene Ontology |
| HRNet | High-resolution network (DeepLabCut architecture) |
| IA | Intermittent access |
| IACUC | Institutional Animal Care and Use Committee |
| IEG | Immediate-early gene |
| KEGG | Kyoto Encyclopedia of Genes and Genomes |
| MAST | Model-based Analysis of Single-cell Transcriptomics |
| MoSeq | Motion sequencing (Keypoint-MoSeq) |
| OPC | Oligodendrocyte precursor cell |
| PCA / PC | Principal component analysis / principal component |
| PVN | Paraventricular nucleus of the hypothalamus |
| PVT | Paraventricular thalamus (aPVT anterior, pPVT posterior) |
| ROI | Region of interest |
| SAP | Stretch-attend posture |
| scVI | Single-cell variational inference (integration) |
| SDS | Social defeat stress |
| SEM | Standard error of the mean |
| snRNA-seq | Single-nucleus RNA sequencing |
| TBC | Two-bottle choice |
| TRAP | Targeted Recombination in Active Populations |
| tdTom | tdTomato fluorescent reporter |
| UMAP | Uniform Manifold Approximation and Projection |
| UMI | Unique molecular identifier |
References
Analysis Scripts
All scripts used to analyze the behavioral data and generate the figures are provided in the thesis repository under scripts/, organized below by study, by where each step sits in the experimental pipeline, and by what it was used for. Acquisition and pose-estimation tools that are maintained as separate open-source projects are linked to their own repositories rather than duplicated here.
Study 1: Behavioral phenotyping
Acquisition and pose estimation
- Synchronized multi-camera video acquisition: SqueakShot (Raspberry Pi multi-camera capture for the dyadic social-interaction recordings).
- Two-dimensional → three-dimensional pose engine: Fus3dPose (Viterbi identity re-tracking, charuco calibration, and
aniposelibtriangulation for the social-interaction sessions). - Markerless 2D keypoint tracking: DeepLabCut (DEKR-W32 networks used for the EPM and social-interaction pose).
- Unsupervised behavioral-syllable reference: Keypoint-MoSeq.
- Head-dip action detection (dlc2action):
train_dlc2action.ipynbtrains the single-class head-dip detector;analyze_dlc2action_predictions.ipynbapplies it to detect dip bouts;classify_headdips.pypost-processes the detections.
Elevated plus maze: ROI kinematics and the anxiety score
01_convert_dlc_to_csv.ipynb: converts DeepLabCut pose output to per-frame keypoint CSVs.02_calibration_and_rois.ipynbandsvg_to_rois.py: define the five maze regions of interest (two closed arms, center, two open arms).03_roi_kinematics_analysis.ipynb,kinematics_utils.py,project_config.py: compute per-mouse open-arm time, open-arm entries, total distance, and mean speed.classify_headdip_subtypes.py: the point-in-polygon rule that splits head dips into Protected and Unprotected.anxiety_composite_sds84_reduced.py: builds the composite EPM anxiety score from its four sign-aligned, z-scored components.anxiety_metric_correlations.py: anxiety-to-drinking correlations.
Two-bottle-choice drinking
drinking_by_group.py: per-session ethanol intake and preference by group.thesis_plots.ipynb(shared): the cohort-level alcohol-consumption panels and the brain-region cFos figure.
Manuscript figure generation (poster pipeline)
These read the pre-computed CSVs from the analysis steps above and render the final figures.
poster_common.py: shared paths, palette, and save helpers.poster_score_definitions.py: the anxiety- and social-score definition figures.poster_epm_figures.py: drinking-by-session, preference, and per-subject drinking figures.poster_si_figures.py: anxiety × social scatter, the 3-D k-means phenotype, and the significant-module figure.poster_radar.py: the nine-axis phenotype fingerprint (radar).poster_module_umap.py: the 14-module behavioral UMAP.poster_densitymap.py: per-cluster and difference density maps over the embedding.poster_m4_cluster.py: the following/pursuit module by cluster.poster_rows.py: the per-metric resilient-versus-susceptible panel boxplots.module_stat_tables.py: the per-module occupancy statistics tables.poster_metric_tables.py: the per-component metric reference tables.
Study 2: Single-nucleus transcriptomics of the PVT
Single-nucleus analysis was carried out in R (Seurat); shared helpers are in Functions.R. The cohort-level drinking and cFos figures are generated by the shared thesis_plots.ipynb.
Cell-type and neuronal-subtype taxonomy
AllCells.qmd: all-nucleus SCTransform + Harmony clustering and cell-type assignment.Neurons.qmdandNeurons_Ident.qmd: neuronal subset (scVI), tdTomato visualization, and the anterior/intermediate/posterior region metadata.PVT.qmd: PVT subset, anterior-posterior assignment, and the five neuronal subtypes.markers.qmd: the subtype and region marker UMAPs, stacked violins, and marker tables.miscplots.qmd: additional cell-type plots.
Activation classification and differential expression
IEGPlots.qmd: IEG module scoring (AddModuleScore) and the tdTomato / IEG activation classification.CellCounts.qmdandPVT_IEGValidation.qmd: nuclei counts by activation category and subregion.Expression_BySample.qmd: differential expression across activation status.Bootstrapping.qmd: the MASTFindMarkerstest with the 1,000-iteration control-cell bootstrap that produces the stable differentially expressed gene sets.