Introduction

Substantial progress has been made in our biological understanding of alcohol use disorder (AUD). AUD is influenced by a constellation of genetic factors [1] as well as multiple tissues and cell types predominantly in the brain [2]. Scant evidence exists on whether, or how the molecular and cellular findings from animal models of heavy alcohol use map onto the neurobiological processes of human AUD [3]. A deeper understanding of how human AUD data correlate with pre-clinical alcohol use paradigms could pinpoint conserved biological entities of AUD and chronic alcohol use.

A rich array of mammalian alcohol use models are utilized for investigating specific bio-behavioral components of human AUD ranging from chronic ethanol exposure [4] self-administration [5] and binge drinking [6]. Various species and strains of animals are used that differ in the degree of volitional alcohol consumption and may differ in how they model human AUD.

Multi-omic datasets on human and animal alcohol use traits afford the opportunity to systematically explore the resemblance, or distinctions, of molecular patterns of alcohol use across species, tissues, and behavioral models. Most publicly available datasets include brain RNA data in the reward and anti-reward pathways in the brain that govern various addictive behaviors. The cellular make-up, neuroanatomical organization, and molecular features (i.e., RNA expression) in the brain are fairly conserved across species [7,8,9].

Recently, Friske et al. [10] performed a meta-analysis of transcriptome-wide data in humans and animal models of AUD, highlighting neuroimmune dysregulation and related molecular pathways across the brain. While this study underscored core transcriptional changes conserved across species, it did not address the overlap of gene networks, single-cell deconvolution nor systematically compare distinct preclinical paradigms of alcohol use.

Consequently, the present study extends that work by integrating both bulk and single-cell data from multiple mammalian models designed to capture diverse drinking phenotypes, allowing us to pinpoint cell-type-specific expression and identify conserved gene co-expression networks in AUD. We hypothesized that human AUD RNA expression and cell type abundance would show consilience with animal models of alcohol use, but that the degree and nature of conservation would depend on 1) species, 2) brain region, and 3) behavioral paradigm. Specifically, we predicted that biological correlates of human AUD would overlap most with 1) primate models of alcohol use, 2) within limbic brain regions, and 3) in behavioral paradigms of binge drinking. Lastly, we predicted that genes associated with AUD and alcohol use across species would be enriched for genes that comprise the heritability of alcohol use traits.

Methods

Samples

Clinical and preclinical transcriptomic brain data on alcohol use was used in this paper. In total, we used bulk RNA-seq and microarray data on 92 humans, 53 primates, and 90 mice (see Table 1).

Table 1 Descriptive Information on the Chronic Alcohol Use Samples Used in the Study.
Full size table

Bulk transcriptomic data

Two human AUD post-mortem brain samples were used – a discovery and replication sample. The discovery dataset included RNA-sequencing (RNA-seq) samples from the New South Wales Brain Tissue Resource Center from a previous publication [11]. We acquired RNA-seq data from superior prefrontal cortex (PFC), nucleus accumbens (NAc), and central nucleus of the amygdala (CeA; [11]). These samples were collected within three days of death from 30 individuals diagnosed with AUD and 30 social/non-drinking controls. We included two independent human replication sample used microarray data from the New South Wales Brain Tissue Resource Center using superior PFC tissue in an independent sample of 17 individuals with AUD and 15 matched controls [12]. The second human replication sample, were ascertained from the Lieber Institute for Brain Development and used dlPFC tissue (BA 46/9) from 61 individuals with AUD and 61 matched controls [13].

Various preclinical alcohol use models were used that incorporated wide-ranging behavioral models, and a couple of mammalian species that used transcriptomic data in similar brain regions. Male primate (Rhesus macaque) RNA-seq data stemmed from four cohorts in the Monkey Alcohol Tissue Research Resource, originating from two separate studies. The first study used primate brain data from the PFC (cortical area 32) and the CeA [14]. The second primate study focused on NAc core [15]. Primates had four months of schedule-induced polydipsia to provoke daily drinking. After induction, primates underwent two 6-month sessions of alcohol self-administration where alcohol was freely available for over 90% of the day. Prior literature has demonstrated that primate alcohol consumption is comparable to humans with AUD within this paradigm [16]. Primates show different levels of alcohol consumption and were split into multiple groups, including controls (alcohol naïve; NAc only), low drinkers, high drinkers, very high drinkers (ethanol consumption > 3 g/kg for > 20% of days), or binge drinkers (CeA and PFC only). In the NAc, drinking groups were compared with alcohol naïve controls. For CeA and PFC data, drinking groups were compared with the low drinkers (PFC and CeA – as no alcohol naive controls existed in these data). After outlier removal (due to low RNA-seq read count; z-score < −2 of scale factor of normalized read counts), we used a total of 83 monkey brain samples (nNAc = 23; nCeA = 30; nPFC = 30).

Two studies investigated four drinking traits within mouse models. One used the high drinking in the dark (HDID) binge drinking paradigm in 68 male and female genetically diverse mice that were bred for drinking or controls (Heterogenous Stock-Northport mice; [17]). The HDID binge drinking mouse sample used RNA-seq of NAc tissue. The drinking in the dark paradigm starts three hours into the dark phase and includes a four-day self-administration paradigm (20% ethanol) with the first three days consisting of three separate 2 h sessions and the fourth day consisting of a single 4 h session. Mice were sacrificed three weeks after their last session. The second mouse sample leveraged summary data from a study on adult male C57BL/6 J mice (n = 8–11 per group; [18]) assessing alcohol self-administration, chronic intermittent ethanol exposure (CIE), and the combination of CIE and alcohol self-administration. Alcohol self-administration involved a two-week acclimation followed by six weeks of a two-hour, two-bottle choice paradigm with 15% ethanol or non-alcoholic solution. CIE, modeling non-voluntary effects of chronic alcohol exposure, consisted of inhaling ethanol or air through a vapor chamber for 16 h/day over four days, followed by a three-day abstinence period, totaling four cycles. CIE mice typically exhibit blood ethanol concentrations of 175–225 mg/dL [19]. Combined CIE and alcohol self-administration involved a two-week acclimation, followed by alcohol self-administration for five days after each CIE cycle, capturing an escalated drinking phenotype. Mouse brains were harvested 24 h after the last session, with ethanol-naïve mice serving as controls for differential expressions.

Single-cell transcriptomic data

Single-nucleus and single-cell RNA sequencing data (sn/scRNA-seq), provided by 10x Genomics, were utilized from three studies involving humans [20], primates [21], and mice [22]. These studies analyzed brain regions corresponding to those examined in the bulk transcriptomic samples. Human brain regions included the dorsal-lateral PFC (n = 3), NAc (n = 8), and amygdala (n = 5). Mid-fetal primate scRNA-seq data included dorsal-lateral PFC (n = 2), striatum (n = 2), and amygdala (n = 2). Lastly, young adult male (8–10-week-old) C57BL/6 J mice included the NAc (n = 11). The original cell-type identities were used from the original publications, but data underwent further quality control filtering out cells/nuclei with >5% mitochondrial reads and those with unique feature counts over 10,000 or less than 200 using Seurat version 5 [23]. All sn/scRNA-seq data were normalized via negative binomial regression corrected for mitochondrial counts using SCTransform [23]. Cell types were identified using differentially expressed genes.

Data processing

All bulk RNA-seq samples underwent uniform processing pipeline. Specifically, we preprocessed samples with Trimmomatic version 0.39 [24] to filter out reads < 36 bp long, leading or trailing reads < Phred score of 3 and allowing a maximum of 2 mismatches per read). After pre-processing, samples underwent alignment to the respective genomes (hg19, mmul_10, mm10) using STAR aligner version 2.5.3.a ([25]; -outSAMstrandField intronMotif –outFilterScoreMinOverLread 0.3 -outFilterScoreMinOverRread 0.3 and –twopassMode Basic) after preprocessing with Trimmomatic version 0.39 ([24]; reads < 36 bp long, leading or trailing reads < Phred score of 3 and allowing a maximum of 2 mismatches per read). RNA-seq read alignment yielded an average of 76,181,104 reads in humans (s.d. = 29,529,191; M%Alignment = 86.26%), 34,551,920 reads in primates (s.d. = 8,202,258; M%Alignment = 79.71%), and 32,317,723 (s.d. = 6,271,112) reads in mice (M%Alignment = 88.90%). Aligned reads were counted using featureCounts [26].

Differential expression

DESeq2 was employed for differential expression analysis in humans, primates, and HDID mice. Differential expression analyses in humans controlled for sex, age, post-mortem index (PMI), and brain pH level, which together accounted for the principal technical and biological sources of variance in these data. In primates, we controlled for cohort (accounting for age and sex), and in mice we controlled for sex since other variables were balanced across groups. For all species, DESeq2’s independent filtering procedure – which leverages the average normalized counts per gene - was applied to account for genes with consistently low read counts prior to model fitting. We estimated transcriptome-wide correlations across differential expression results – using only homologous genes across species (via biomaRt; [27]). To quantify the transcriptional overlap across species, we conducted Pearson Product Moment correlations on differential expression test statistic (i.e., Wald statistic) – to balance both effect size and error estimates in differential expression analyses. To adjust for multiple testing, we performed multiple-testing correction using Benjamini-Hochberg false discovery rate.

Single-cell deconvolution

We estimated the relative abundance of cell types in bulk RNA-seq datasets using CIBERSORTx [28]. Specifically, we employed the standard cell-type proportion estimation module with 100 permutations, rather than the batch correction or cell-specific expression imputation features, to maintain comparability across bulk datasets and avoid overfitting given the modest sample sizes. Signature matrices were constructed from the sn/scRNA-seq reference datasets described above. Using Wilcoxon tests, we identified the top 100 marker genes for each cell-type using FindAllMarkers in Seurat. To ensure group differences were not due to confounding factors we correlated estimated cell abundances with covariates used in differential expression analyses and found no substantial associations.

Gene Co-expression network analyses using WGCNA

We performed gene network modeling via weighted gene co-expression analyses (WGCNA; [29]). First, we created signed hybrid WGCNA networks from human post-mortem brain tissue- separately for each brain region examined. Before WGCNA gene network construction, we filtered genes that had less than 10 read counts in more than 90% of the samples and performed variance-stabilizing transformations as per WGCNA creators’ recommendations (https://horvath.genetics.ucla.edu/html/CoexpressionNetwork/Rpackages/WGCNA/faq.html). WGCNA models included 18,341, 16,474, and 16,755 genes for the PFC, NAc, and CeA, respectively. We weighted correlations between genes by using soft thresholds (6, 10, and 18) that were elected to maximize fit and satisfy WGCNA modeling assumptions (all scale free topology > 0.851). With this weighted correlation matrix, we created an adjacency matrix and then a distance matrix, which was hierarchically clustered and split into discrete gene networks using a dynamic tree cutting algorithm. Gene networks were arbitrarily assigned a color (e.g. the magenta gene network in the PFC is different than the magenta network in the CeA).

Conservation, association, and characterization of gene co-expression networks

We assessed the reproducibility of human gene co-expression networks and their conservation in primate and mouse models using the modulePreservation function. This function determines the pair-wise concordance of gene-gene correlations within a gene co-expression network between input and test data [30]. As recommended by Langfelder et al., [30], we evaluated the Zsummary metric, which is a composite of 7 different preservation statistics. A Zsummary value > 10 represents strong reproducibility and was used as our threshold for conservation. First, we performed module preservation analyses on 100 bootstrapped human brain samples. All WGCNA networks had a Zsummary score > 10.65 indicating robust and stable gene co-expression networks.

Using an effect-size approach, we investigated whether WGCNA networks were associated with alcohol outcomes by permutating the average absolute value of differential expression test statistic of genes in gene co-expression network (10,000 permutations). Significant WGCNA associations were required to have an BH-FDR adjusted p-value < 0.05. Gene co-expression networks were tested for enrichment in KEGG pathways, biological processes, molecular functions, and cellular component ontological processes using enrichR [31] as well as human brain cell-types via cell-specific enrichment analysis (CSEA; [32]). We characterized hub genes using two approaches: expression-based Phenome-Wide Association Studies (ePheWAS; https://www.systems-genetics.org/ephewas) and querying the NHGRI-EBI catalogue of human genome-wide association studies (GWAS; https://www.ebi.ac.uk/gwas/).

Partitioned heritability

We performed partitioned heritability analyses using Linkage Disequilibrium Score Regression (LDSC; [33, 34]). Partitioned LDSC estimates whether single nucleotide polymorphisms (SNPs) in and around a gene-set account for a significant proportion of the heritability of a trait proportional to the amount of variants included. To determine behavioral specificity, we tested whether the individual heritability of multiple human alcohol and tobacco use traits were enriched for all genes within conserved WGCNA networks associated with human and animal chronic alcohol use (across all brain regions). The specific GWAS used in our analyses included publicly available data on AUD [1], problematic alcohol use [35], drinks per week [36], cigarettes per day [36], cigarette cessation [37] and a negative control trait – wears glasses (https://nealelab.github.io/UKBB_ldsc/). Our analyses used LDSC’s default gene window size (100 kb) to assign SNPs to genes based on the expectation that up to 80% of local gene regulatory regions occur within 100 kb of a gene [38].

Results

Using transcriptome-wide samples of 92 humans, 53 primates, and 90 mice in three brain regions (see Table 1 and see Supplementary Table S1-S5 for differential expression results), we assessed the overlap of AUD RNA signature with pre-clinical chronic alcohol use. Using PFC, NAc, and CeA RNA-seq data on a sample of 60 humans, our analyses revealed significant correlations between differential expression results of human AUD and mammalian models of alcohol use (see Fig. 1A; all padj < 0.05). Cross-species correlations differed by tissue, species, and preclinical model. The most significant and positive correlations with human AUD occurred in primates and mouse models of binge drinking. Some correlations were negative. For instance, mouse chronic intermittent ethanol exposure (CIE) and alcohol self-administration data exhibited significant negative correlations with AUD in the PFC, but these models showed no significant correlations with AUD in the NAc or CeA. It is important to note that regional specificity in brain regions across species are not completely identical.

Fig. 1: AUD Brain RNA Signature Robustly Correlated with Animal Models of Alcohol Use.
Fig. 1: AUD Brain RNA Signature Robustly Correlated with Animal Models of Alcohol Use.
Full size image

Transcriptome-wide correlations of differentially expressed genes across traits, brain regions and species. A Heatmap of AUD RNA correlations with mouse and primate models of alcohol use; **** p < 0.001; blank = p > 0.05; NA not applicable (B) Scatterplot showing reproducibility of brain RNA correlations using an independent sample of AUD; R = Pearson correlation coefficient. Self Admin = ethanol self-administration; CIE chronic intermittent ethanol exposure; HDID high drinking in the dark, PFC pre-frontal cortex, NAc nucleus accumbens; CeA central nucleus of amygdala.

To examine the reproducibility of RNA correlations, we used an independent microarray sample of 32 individuals with AUD and matched controls in the PFC [12]. AUD RNA signatures correlated in the expected direction across samples, although with a small effect size (r = 0.15, p < 0.0001). All but the heavy drinking primate model showed significant correlations with AUD (see Fig. 1B), suggesting our findings are reproducible. Notably, the very heavy drinking primate model showed correlations that aligned with human AUD at a level comparable to 39% of the correlation observed relative to the human AUD correlation with itself across different samples (ie., r = 0.16 for AUD-AUD; r = 0.063 for VHD-AUD; see Fig. 1B). To further assess the robustness of our findings, we correlated differentially expressed genes in AUD with an independent sample [13] and observed a significant correlation (p < 0.05; see Supplementary fig. S1), reinforcing the reproducibility of AUD-associated transcriptional signatures across cohorts.

Next, we used publicly available snRNA-sequencing data from post-mortem human and fetal primate PFC (11,196–12,803 cells after QC), striatum (12,243–19,862 cells after QC) and amygdala (13,422–13,984 cells after QC) as well as mouse NAc (41,519 cells after QC) studies to investigate the cellular basis of AUD and mammalian alcohol use. Only bulk RNA-seq data and not microarray data (i.e., CIE and alcohol self-administration) were integrated with snRNA-seq data to avoid methodological biases. Note that many, but not all, brain cell-types were similar across species (see Fig. 2). Using single-cell deconvolution techniques (CIBERSORT), we identified that the proportion of oligodendrocytes was decreased in human AUD vs controls in the PFC and the NAc (see Fig. 2). Similarly, the proportion of oligodendrocytes were decreased in the NAc of a mouse model of binge drinking (high drinking in the dark [HDID]; [17]) and oligodendrocyte progenitor cells were decreased in the PFC of primate models of alcohol use (all P < 0.05; see Fig. 2). A few other trends were observed in AUD and showed similar effects in primates and mice. For instance, the relative proportion of NAc dopamine receptor 2 medium spiny neurons (MSN_D2) were trending to decrease in AUD and were decreased in HDID binge drinking mice. Additionally, the relative proportion of CeA excitatory neurons were trending to be decreased in AUD and were also decreased in binge drinking primates. Cell type abundance changes were not significantly correlated with human meta-data (all p < 0.1; see Supplementary fig. S2). Notably, the directionality was consistent across species and some proportion of cell-types showed a dose-response, but the lack of significance could reflect low statistical power.

Fig. 2: Single-cell Deconvolution of AUD and Alcohol Use Brain Signatures Across Animal Models and Brain Regions.
Fig. 2: Single-cell Deconvolution of AUD and Alcohol Use Brain Signatures Across Animal Models and Brain Regions.
Full size image

In silico cytometry of human AUD and preclinical alcohol use across traits, brain regions, and species. A-C shows clustering of cell-types by brain region and species. D-F shows the proportion of each cell type by group across species, brain regions and pre-clinical models. HDID high drinking in the dark, Oligo oligodendrocytes, Excit excitatory neuron, Inhib inhibitory neuron; NascIntN nascent interneuron, MSN_D2 dopamine D2 receptor medium spiny neuron; OPC oligodendrocyte progenitor cell; PFC pre-frontal cortex, NAc nucleus accumbens, CeA central nucleus of amygdala; **** p < 0.001; * p < 0.05; # p < 0.1.

Next, using the 60 RNA-seq samples of AUD and matched controls, we created /gene co-expression networks via weighted gene co-expression network analyses (WGCNA) gene in each brain region individually and tested for their conservation across species. WGCNA uses pairwise RNA correlations of all gene pairs, clusters similar correlated gene pairs together and creates distinct gene co-expression networks arbitrarily assigned to a color. We tested whether human gene co-expression networks were conserved across species (and within brain region) using permutation analyses and a recommended threshold outlined in prior work (Zsummary > 10; [30]). We found that 6, 16, and 8 networks were highly conserved across species in the PFC, NAc and CeA, respectively (Zsummary > 10; see Supplementary fig. S3). Of the 30 highly conserved WGCNA networks across species, 80% were highly conserved in primates and 46.67% were highly conserved in mice. Eight conserved gene co-expression networks were associated with both human AUD samples and at least one alcohol use outcome in animals (see Fig. 3). All gene co-expression networks were associated with AUD and mammalian alcohol traits in the same direction. All gene co-expression networks were robustly enriched for certain brain cell types and all but one was significantly enriched for gene ontology functions (see Table 2).

Fig. 3: Conserved Gene Co-expression Networks Associated with Human AUD and Animal Models of Alcohol Use.
Fig. 3: Conserved Gene Co-expression Networks Associated with Human AUD and Animal Models of Alcohol Use.
Full size image

Gene co-expression networks associated with alcohol use across samples and species. X-axis shows the WGCNA gene co-expression network and the y-axis shows the mean differential expression test statistic by brain region and trait. All gene networks in this plot were significant via permutation analyses (padj < 0.05). WGCNA weighted gene co-expression network analysis, CIE chronic intermittent alcohol exposure, HDID high drinking in the dark. Note color for WGCNA networks are arbitrary and the magenta gene network in the PFC is distinct from the magenta gene network in the CeA.

Table 2 WGCNA Gene Networks were Conserved Across Species and Associated with Human AUD and Chronic Alcohol Use.
Full size table

To identify key regulators that may be driving these physiological processes, we sought to find and characterize hub genes in these WGCNA gene co-expression networks. Selecting the top 2.5% of module membership in each gene network, we found a total of 63 hub genes (see Fig. 4; 57 unique genes). We characterized these genes with two online tools: NHGRI and ePheWAS and reported these results in Table 3. In short, we found these genes may be implicated in various drug use traits, as well as impulsivity, motivation, stress, and other molecular and morphological outcomes.

Fig. 4: Conserved Gene Co-expression Networks Associated with Human AUD and Animal Models of Alcohol use.
Fig. 4: Conserved Gene Co-expression Networks Associated with Human AUD and Animal Models of Alcohol use.
Full size image

Image created via Biorender.com. Note enriched cell-types are displayed next to WGCNA networks and dashed lines between genes are for visual purposes only and not derived from gene-gene connectivity metrics. Bottom right two gene networks are in the CeA, the top left is in the PFC and the rest are in the NAc.

Table 3 Hub genes from WGCNA Gene Networks that were Conserved Across Species and Associated with Human AUD and Animal Chronic Alcohol Use.
Full size table

Lastly, we used partitioned heritability analyses to test for genetic enrichment. The heritability of drinks per week and problematic alcohol use, but not tobacco nor a non-substance use trait (wears glasses), were significantly enriched for the genes from the conserved gene co-expression networks associated with AUD and alcohol use (see Fig. 5).

Fig. 5: Conserved AUD & Alcohol Use Networks Enriched for GWAS Associations of Alcohol Use.
Fig. 5: Conserved AUD & Alcohol Use Networks Enriched for GWAS Associations of Alcohol Use.
Full size image

Results from partitioned heritability analyses. DNA variants within 100 kb of the genes from conserved WGCNA networks that were associated with human AUD and mammalian models of alcohol use were included in these analyses. * P < 0.05; NS = P > 0.05.

Discussion

This study found partial support for our hypotheses. RNA expression and cell-type abundances associated with human AUD generally correlated with animal models of alcohol use, but the extent of which depended on species, brain region, and behavioral trait. The strongest overlap of brain RNA patterns was observed between humans and primates, which matches the close phylogenetic relationship between species. Mouse binge drinking models showed substantial overlap with human AUD gene expression, which is consistent with independent biological data on human and mouse alcohol use phenotypes [39]. Importantly, brain correlations with primate and mouse alcohol use were reproducible with an independent human AUD sample [12] highlighting the reproducibility of these findings.

Single-cell transcriptomic data provided insights into the cellular basis of AUD and mammalian chronic alcohol use. The abundance of oligodendrocyte cell types were decreased in AUD among humans and in the animal models of alcohol use in the PFC and NAc. These observations are consistent with snRNA-seq data in AUD [40], as well as postmortem cellular and in vivo neuroimaging data, suggesting that prolonged alcohol use is linked to oligodendrocyte and white-matter damage [41]. We also observed parallel trends suggesting decreased abundance of dopamine MSN_D2 NAc medium spiny neurons (MSN) in human AUD and mouse binge drinking. D2 MSNs in NAc are thought to mediate reward, motivation, and addictive behaviors. Inhibiting NAc MSN D2 neurons has been shown to experimentally promote alcohol intake [40]. While these analyses had limited statistical power, it was reassuring to detect consistent directionality across species and mirror similar observations in the literature.

Utilizing gene co-expression network analyses, we were able to identify conserved systems across species. These analyses suggested that dopamine and acetylcholine neurons, as well as a variety of glial cells, play a role in cross-species consilience in various brain regions and alcohol use traits. Corroborating results highlighted by Friske et al., [10], we found that pathways and gene-sets involved in the inflammation, myelination and synaptic plasticity were enriched among these gene-networks. Expression of the hub genes from these gene networks were linked with relevant traits in mice via ePheWAS: motivation, impulsivity and locomotor response [1] [2], as well as pertinent traits via human GWASs: alcohol, tobacco and cannabis use. Accordingly, genes from these co-expression networks accounted for ~20% of the heritability of alcohol use traits but were not enriched in a negative control trait (wearing glasses). In sum, despite the challenges inherent to harmonizing such disparate data sources, there is significant face validity and coherence with known addiction biology to suggest that the approach used here has utility.

Our results support findings from Friske et al. [10] and extend these results by assessing the cross-species cell-type abundance and gene-network effects as well as pinpointing the magnitude and direction of cross-species gene-expression correlations by behavioral paradigm. The findings of the current study not only replicated convergent neuroimmune processes but also refined them by highlighting oligodendrocyte depletion and altered dopamine receptor 2 medium spiny neurons - processes with potential translational relevance for AUD research.

Nevertheless, these observations should be interpreted with caution. The heterogeneity across studies complicates interpretations. For instance, we are unable to decipher the degree to which an association – or lack thereof – is due to the degree of alcohol exposure, genetics, mis-match of brain regions, or technical factors. Sometimes associations of differentially expressed genes across species were negative, which is in line with previous cross-species research on drug use traits but is not understood well [39, 42]. Data used in the current paper lacks diversity in genetics and sex and may limit the generalizability of our findings, and RNA-seq samples were collected at different timepoints following ethanol exposure, which may further contribute to variability in transcriptomic signatures. Cell-type abundance analyses should be interpreted with caution in light of small (underpowered) samples, containing few cells and sometimes using less relevant samples (i.e., fetal primate data). In particular, the use of fetal primate reference data may introduce developmental differences when compared to adult tissue, although the consistency of directionality in cell-type shifts across species suggests that these limitations did not obscure the broader cross-species patterns. Notwithstanding the limitations, our paper provides a data-driven approach to highlight the biological intersection between human AUD and animal alcohol use and systematically compares pre-clinical paradigms of alcohol use and their relevance to human AUD. Future efforts would benefit from harmonizing data with larger samples, and more traits/substances and incorporating different techniques to synergistically assess the biological overlap across species, cell types, and brain regions/subregions.