Integrative genomics approach identifies glial transcriptomic dysregulation and risk in the cortex of individuals with Alcohol Use Disorder

Abstract

Background:

Alcohol use disorder (AUD) is a prevalent neuropsychiatric disorder that is a major global health concern, affecting millions of people worldwide. Past studies of AUD used underpowered single cell analysis or bulk homogenates of postmortem brain tissue, which obscures gene expression changes in specific cell types. Therefore, we sought to conduct the largest-to-date single-nuclei RNAseq (snRNA-seq) postmortem brain study in AUD to elucidate transcriptomic pathology with cell type-specific resolution.

Methods:

Here we performed snRNA-seq and high dimensional network analysis of 73 post-mortem samples from individuals with AUD (N=36, N_nuclei= 248,873) and neurotypical controls (N=37, N_nuclei= 210,573) in the dorsolateral prefrontal cortex from both male and female donors. Additionally, we performed analysis for cell type-specific enrichment of aggregate genetic risk for AUD as well as integration of the AUD proteome for secondary validation.

Results:

We identified 32 distinct cell clusters and found widespread cell type-specific transcriptomic changes across the cortex in AUD, particularly affecting glial populations. We found the greatest dysregulation in novel microglial and astrocytic subtypes that accounted for the majority of differential gene expression and co-expression modules linked to AUD. Differential gene expression was secondarily validated by integration of a publicly available AUD proteome. Finally, analysis for aggregate genetic risk for AUD identified subtypes of glia as potential key players not only affected by but causally linked to the progression of AUD.

Conclusions:

These results highlight the importance of cell type-specific molecular changes in AUD and offer opportunities to identify novel targets for treatment on the single-nuclei level.

Introduction

Alcohol Use Disorder (AUD) is a prevalent neuropsychiatric disorder characterized by impaired ability to stop or control alcohol use despite adverse consequences. Worldwide over 100 million people had AUD in the last year and repeated alcohol abuse constitutes the seventh leading risk factor for death and disability worldwide(1,2). Specifically, in the United States, 28.6 million adults ages 18 and older (11.3%) had AUD in the past year(3). Despite its prevalence there are currently only a handful of treatments for AUD, with minimal efficacy in reducing harmful alcohol use. To develop more effective therapeutic targets, we must elucidate how alcohol combines with genetic risk to converge onto specific molecular processes that lead to the development and progression of AUD.

Although AUD is highly heritable, its genetic etiology is complex with additive effects of alcohol, combined with environmental and physiological factors(4). Assessment of AUD risk is also challenging due to its genetic architecture, complexity of behaviors and dynamic patterns of drinking across lifespan(4). Despite this genetic complexity, studies have identified consistent brain region-dependent alterations of genes involved in regulation of neuronal activity and neuroinflammation(5–8). Of note, most of these transcriptomic studies in human brain have been bulk analyses of tissue homogenates, which may obfuscate cell type-specific gene expression, particularly in minor populations. Moreover, it is difficult to verify whether subtle differences observed from tissue homogenates are explained by disease state or by differences in cell type composition between samples.

Developments in RNA-sequencing technologies, such as single-nucleus RNA sequencing (snRNA-seq), have permitted unprecedented resolution into the molecular architecture of the human brain and provide a solution for the drawbacks of bulk-tissue experiments(9,10). Specifically, this approach demonstrated that gene expression patterns in the brain are disease state- and cell type-specific, not only differentiating major cell types but also subpopulations(9,10). Previously, we profiled a small cohort of postmortem human dorsolateral prefrontal cortex (dlPFC) using snRNA-seq(11). Despite the small sample size, this study identified novel differentially expressed genes (DEGs) and found an enrichment of inflammatory genes in multiple glial subtypes in a region that is key in reward processing and the pathophysiology of addiction(11). To date, the only other published snRNAseq study in AUD is in the nucleus accumbens, which identified neuroinflammatory changes in glia that were underrepresented in bulk transcriptomes(11,12). Limitations of these studies include number of nuclei, sample size and representation of a single sex(11,12).

In this study, we leveraged snRNA-seq, GWAS and proteomic integration from postmortem human dlPFC of AUD and neurotypical controls (N=73) to define the transcriptomic architecture of AUD across 32 cell types. We identified DEGs in AUD, determined co-expression modules that correlated with DSM-5 diagnosis and identified key hub genes that may be implicated in the pathogenesis of AUD. The largest number of DEGs were detected in astrocytes, oligodendrocytes, and microglia. Finally, by integrating GWAS summary statistics we highlight causal cell types. This study provides the first comprehensive, well-powered single-nuclei study of the transcriptional changes in AUD and further highlights the clinical relevance of glial and immune changes in both the progression of AUD risk and pathogenesis.

Methods & Materials

Additional information can be found in Supplemental Information.

Cohort selection and postmortem tissue collection

Diagnosis of alcohol dependence was denoted as based on DSM-5 criteria. Postmortem brain samples were provided by the New South Wales Brain Tissue Resource Centre (NSWBTRC) at the University of Sydney. Cohort selection and processing of fresh frozen tissue samples were collected as previously described(8,13) (TableS1).

Droplet-based snRNA-seq

Libraries were prepared as previously described(11). Files were processed using Cellranger(v6.1.2) and aligned to the GRCh38 reference with default parameters before processing.

Filtering, normalization & visualization

Ambient RNA was corrected using SoupX(v1.3.0)(14). The corrected matrix was loaded into Seurat(v4.2.0). In depth quality control analyses were performed to ensure no samples were skewed for any metrics before pooling (FigS1–2,TableS1–2) Custom filtering thresholds and normalization were applied. We next integrated our cohort using Harmony(v1.1.0)(15) and followed a standard Seurat pipeline(FigS3). Principal component (PC) analysis was performed on the top 2000 highly variable features, the top 20 PCs were used to construct a nearest neighbor graph. Nuclei were then assigned to clusters using the Louvain algorithm with a resolution of 0.5.

Marker identification for clusters

The FindMarkers function was first applied after initial clustering to determine cell type annotation. Known marker genes from previous snRNA-seq studies were used as benchmarks for initial annotation(16–21). After the first round of clustering, marker genes were used to inspect and subcluster, resulting in 32 final clusters. Gene ontology (GO) enrichment analysis was performed using Metascape(22).

Cell type proportion

To determine if AUD changed proportions of cell type-specific nuclei we used sccomp(v3.20)(23).

Differential gene expression

DE testing was performed using Libra pseudobulking(v1.0.0)(24). First, nuclei corresponding to a given aggregate cell type were selected from the full dataset. Second, counts were summed to produce a pseudobulk matrix for each cell type. Lastly, DEGs between AUD and controls were detected using edgeR(v3.36.0) with correction for the following covariates: Sex, Age, PMI, AUD Classification and Sequencing Site. FDR<0.05 was used to detect DEGs within each cell type. Consistency was determined using a resampling approach(25).

High dimensional weighted gene correlation network analysis

Network analysis was performed with hdWGNCA(v0.1.1)(26). A soft power of 7 and a minimum module size of 25 genes was selected. This analysis included gene network construction, module-trait correlation, module eigengene by trait, and hub gene analysis. We selected DSM-5 and AUD Classification as variables for module eigengene expression.

External data sources

Data sources for glial subtypes can be found in Table S11 and Supplemental Information (27–35). Human AUD dlPFC datasets: bulk RNA-seq(8); proteome(36).

GWAS enrichment meta-analysis

Summary statistics for meta-analysis of alcohol dependence and AUDIT-C GWAS were obtained from the Million Veteran Program (N=318,694)(37). MAGMA(v1.10)(38,39) was used to perform gene-based association analysis for the GWAS summary statistics(8,11). We selected the European ancestry summary statistics since 72 out of 73 individuals were of European descent. The enrichment scores were calculated for: (1) cluster-specific DEGs; (2) aggregate DEGs; (3) hdWGCNA modules. A minimum of 10 genes were required to run MAGMA on a gene set.

Statistical analysis & Graphing

Module eigengene statistical comparisons were performed by two-tailed Welch’s t-test (P<0.05). Significance of overlapping gene sets was performed using a one-sided hypergeometric test (P<0.05). Graphs for figures were produced using R(v4.3.1) or Prism(v10.1.0) (GraphPad, San Diego, CA).

Ethical approval

The post-mortem tissue request for this study was received by the NSWBTRC and reviewed independently by an external Scientific Advisory Board.

Data Availability

Data is available in the Gene Expression Omnibus (GEO), accession ID GSE: GSE247416.

Results

Identification of refined subpopulations in the dlPFC of AUD individuals

To investigate the pathophysiology of AUD, we curated a cohort of 36 individuals with AUD and 37 controls across two institutions including both sexes: (9 AUD females, 10 control females; 27 AUD males, 27 control males), a range of ages (24-81 years, average 54.5 years) and postmortem interval (8.5-72 hours, average 30.8 hours), resulting in 459,446 nuclei (Fig 1A,TableS1). Transcriptionally distinct groups of nuclei were identified using unsupervised, graph-based clustering (Fig 1B). The final 32 clusters captured all major dlPFC cell types: excitatory and inhibitory neurons, microglia, astrocytes, oligodendrocyte precursor cells (OPCs), oligodendrocytes, endothelial cells, pericytes and vascular smooth muscle cells (VSMCs) (Fig 1B). Proportions of different cell types were comparable to other snRNA-seq studies(11,25) and similar across samples and sex, with neurons being over-represented relative to glia (50% neurons, 22% oligodendrocytes, 10% astrocytes, 8% endothelial cells, 5% microglia, 5% OPCs) (Fig 1C, FigS4). Spearman correlation of average expression between cell type-specific clusters showed distinct separation of neurons and glia (Fig 1D). We verified these annotations based on known marker gene expression patterns(16–18,20,25,40) (Fig 1E). Gene expression patterns previously linked to specific cortical layers(16) matched our clustering of excitatory neurons, with cluster-specific representation across layers I-VI (Fig 1F). Likewise, inhibitory cell types demonstrated cluster-specific interneuron marker expression(16) (Fig 1G).

Figure 1: Cell type identification in the dlPFC of control and AUD individuals.

(A) Schematic of sample cases and sequencing sites for the nuclei isolated from 73 brains. (B) UMAP plots of the 459,446 nuclei in our dataset demonstrating transcriptomically distinct cell type clusters determined by unsupervised clustering. (C) Fraction of nuclei represented by specific cell types/clusters. (D) Spearman correlation heatmap between clusters demonstrates transcriptomically distinct cell types and clustering. The color intensity represents the correlation value between 0.5 and 1. (E) Scaled mean expression levels of known cell type markers in the 32 clusters shown as a stacked dot plot. The size of the dots represents the proportion of cells expressing the gene (% Expressed), whereas the color intensity represents the average expression level (Avg Expression) (F) Cortical layer-specific markers varied in expression within the excitatory neuronal clusters. The violin plots depict the expression per cluster of layer-specific marker genes, from the more superficial layers (II) on the left to the deeper layers (V/VI) on the right. The color intensity represents the layer depth. (G) Known classes of inhibitory neurons were identifiable based on the expression pattern of inhibitory marker (GAD1), peptide genes (VIP, SST, CCK) and interneuron markers (RELN, LAMP5).

Defining glial subpopulations in AUD

Recent studies have demonstrated functionally significant immunomodulatory mechanisms and cell types in AUD(41). We hypothesized that AUD leads to relative changes in the expression of defined signatures in glial subtypes as a consequence of disease pathology(42,43). Therefore, we first defined glial subtypes of interest to determine subtype-specific shifts-in-state in AUD.

MG1 was enriched in known homeostatic markers and functions(27–29) (WP304(P2RY12,CX3CR1),−log₁₀ P=11.11,Fig 2C–D,TableS4). MG2 had a significantly higher proportion of expression in AUD (FDR=0.021,TableS3) and was enriched in disease-associated microglia (DAM:CLEC7A,MRC1,P<1.06e⁻⁶³), border-associated macrophages (BAM:TIMP2,PTPRC,P<0.014), and immune signaling (GO:0071345(IL1B,IRAK2),−log₁₀ P=17.44,Fig 2C,E,TableS4). These inflammatory and DAM/BAM markers suggest a shift in microglial phenotype, consistent with but expanding on previous studies highlighting the role of microglia in AUD(44–46).

Figure 2: Cell type-specific clustering reveals alcohol-responsive glial clusters in AUD.

(A) UMAP plot of microglia clusters 1 (MG1) and 2 (MG2). (B) Microglia cluster proportions between control (MG1: 0.94, MG2: 0.05) and AUD (MG1: 0.71, MG2: 0.28). (C) Violin plot of microglia subtype markers in MG1 and MG2: homeostatic markers (P2RY12, CX3CR1, CST3), inflammatory markers (TLR2, NFKBID, IL1B, IRAK2), border associated macrophage markers (PTPRC, TIMP2), and disease-associated microglia markers (CLEC7A, MRC1). Gene ontology enrichment for alcohol-responsive clusters (D) MG1 and (E) MG2 for the top 12 categories. Enrichment is a function of −log10 (P-value). (F) UMAP plot of astrocyte clusters 1 (Ast1), 2 (Ast2), and 3 (Ast3). (G) Astrocyte cluster proportions between control (Ast1: 0.58, Ast2: 0.15, Ast3: 0.27) and AUD (Ast1: 0.67, Ast2: 0.20, Ast3: 0.13). (H) Violin plot of astrocyte subtype markers in Ast1-3: canoncial markers (SLC1A2, SOX5, GRIA2), Ast1-2 specific astrocyte markers (RYR3, GPC5, CDH20, SLC4A4, MMD2), synaptic-astrocytic markers specific to Ast3 (CACNG8, MARCKS, SYNGR3, CPLX1), and inflammatory markers specific to Ast2 (JUN, CD44, VCAN, HMGCS1, ID2). Gene ontology enrichment for alcohol-responsive clusters (I) Ast1, (J) Ast2 and (K) Ast3 for the top 12 categories. Enrichment is a function of −log10 (P-value). (L) UMAP plot of oligodendrocyte clusters 1-5 (Oligo1-5). (M) Oligodendrocyte cluster proportions between control (Oligo1: 0.57, Oligo2: 0.40, Oligo3: 0.017, Oligo4: 0.009, Oligo5: 0.004) and AUD (Oligo1: 0.57, Oligo2: 0.33, Oligo3: 0.08, Oligo4: 0.01, Oligo5: 0.01). (N) Violin plot of oligodendrocyte markers in Oligo1-5: myleination markers (PLP1, MOBP, MBP), subtype markers (RBFOX1, LURAP1L-AS1, CDH19, OPALIN, LINC00844), and inflammatory/disease-associated markers (BCL2, NFKB1, MITF, EVOLV5, BCL6, JUNB, IFITM3). Gene ontology enrichment for alcohol-responsive cluster (O) Oligo2, (P) Oligo3, and (Q) Oligo5 for the top 12 categories. Enrichment is a function of −log10 (P-value). For the violin plots in C, H, N values extend from minimum to maximum, color is a function of individual genes.

There were also shifts in astrocyte subpopulation signatures in AUD (Ast1,FDR=0.026,TableS3). Ast1 (SLC1A2⁺/SOX5⁺) was enriched in canonical functions(47,48) (GO:0042063(GFAP,SOX9),−log₁₀P=11.71,Fig 2H–I,FigS5,TableS4). Ast2 (SLC1A2⁺/SOX5⁺) was also enriched in canonical functions as well as neuroinflammation(48) (WP5083(CD44,ID2),−log₁₀P=7.68,Fig 2H–J,FigS5,TableS4), consistent with previous studies(46,49,50). Ast3 expressed unique synaptic markers (SYNGR3,MARCKS) specific to the reported synaptic neuron and astrocyte program(19) (SNAP-a,Fig 2H,FigS5). Ast3 enrichment was equally split between astrocyte- and synapse-specific functions (GO:0050804(GRIA3,SLC1A3),−log₁₀P=67.30,Fig 2K,FigS5F,TableS4). SNAP-a enrichment showed strong overlap with Ast3 (FigS5G). Expression of SNAP-a genes with our astrocyte clusters specifically overlayed in Ast3, suggesting regulated neuron-astrocyte programs (FigS5H–I).

Proportions of oligodendrocyte subpopulation signatures also shifted in AUD (Oligo2,FDR=0.00048;Oligo5,FDR=0.0018,Fig 2L–M,TableS3). All oligodendrocyte clusters had strong enrichment of myelination genes in both AUD and controls (GO:0042552(PLP1,MBP),Oligo2,−log₁₀P=19.73;Oligo5,−log₁₀P=17.71,Fig 2N,TableS4). Subclusters were separately enriched in different oligodendrocyte functions such as lipid biosynthesis (GO:0008610(LIPA,QKI),Oligo2,−log₁₀P=8.81) or immune function (R-HSA-1280218(NFKB1,ITGB1),Oligo5,−log₁₀P=4.82,Fig 2O–Q,TableS4). Immune dysregulation in Oligo5 was consistent with recent work defining inflammatory subpopulations in white matter of individuals with multiple sclerosis(51,52). Taken together, these AUD-responsive gene signatures, suggest glial populations may be important regulators of AUD pathogenesis.

Glial clusters are primary drivers of AUD transcriptional signature

Next, we assessed transcriptomic dysregulation by performing pseudobulking, including covariates(24,53). We identified 3667 DEGs, with 1875 upregulated and 1792 downregulated in AUD (Fig 3A,TableS5). The majority of DEGs were identified in glial subpopulations (Fig 3A). We additionally tested consistency between our single-nuclei DEGs and bulk expression, finding strong consistency at a cell type-specific level(8) (Fig 3B).

Figure 3: Glial genes drive differential expression in AUD.

(A) Number of differentially expressed genes (DEGs) per cluster, with a total of 3667 DEGs. The y-axis represents the number of DEGs, positive values indicate upregulated genes and negative values represent downregulated genes. Colors represent specific clusters which map to the legend on the right. Aggregate cell type classifications are mapped on the x-axis. (B) Global consistency analysis of snRNA-seq versus bulk RNA-seq. Agreement or disagreement was estimated by deviation from random expectation (z-score) of single-nuclei DEG average rank scores in the ranked list of genes from bulk differential analysis. (C) Circle chart for upregulated and downregulated DEGs in snRNAseq analysis. Aggregate cell type classes are represented by color and map to the legend on the bottom of the graphs. Numbers represent the number of DEGs per aggregate cell type cluster. (D) Circle charts for upregulated and downregulated DEGs as a function of protein-coding versus non-coding. GO enrichment of (E) upregulated DEGs or (F) downregulated per cell type represented as a heatmap. Color represents enrichment as a function of −log10 P-value, with red indicating strong enrichment, light blue representing weak enrichment and dark blue representing no enrichment. Specific cell type clusters are labeled on the X-axis and map to the UMAP legend.

Specific glial clusters represented the majority of upregulated DEGs, such as Ast1, MG1 and Oligo2, which also showed significant shift-in-state proportional changes in AUD (Fig 3C). Excitatory neurons had the second largest fraction of upregulated DEGs after astrocytes, like our pilot study(11) (Fig 3C). Approximately 17% of upregulated DEGs were non-coding genes (Fig 3D), reinforcing their known role in AUD(5,54). Pathway enrichment of upregulated DEGs reinforced the role of inflammatory signaling in AUD in glial cell types, especially microglia (Fig 3E,TableS6). Microglia-specific DEGs were enriched in immune response (GO:0001819(CD14,IL1A),−log₁₀P=8.04) and novel type-II interferons (WP619(OAS1,ISG15),−log₁₀P=5.59,TableS6). Astrocyte-specific upregulated DEGs showed strong enrichment of cellular response to stress (R-HSA-2262752(MAPK14,DCTN1),−log₁₀P=7.25) and extracellular matrix (ECM) (R-HSA-1474244(NCAN,BCAN),−log₁₀P=3.71,Fig 3E,TableS6). Excitatory and inhibitory neuron upregulated DEGs had strong enrichment in regulation of the synapse (GO:0050807(GRM3,LRRC4C),−log₁₀P=4.63,Fig 3E,TableS6). Moreover, most DEGs/pathway enrichments were unique to each cluster, suggesting transcriptional shifts in AUD are primarily cell type-specific (Fig 3E,TableS5–6). Examples of overlapping DEG targets from AUD literature are available in TableS7(55).

Downregulated DEGs were distributed evenly across cell type-specific clusters (Fig 3A–C,TableS5). Glial clusters represented the majority of downregulated DEGs with excitatory neurons having the third largest fraction after astrocytes/oligodendrocytes (Fig 3A–C,TableS5). Approximately 23% of downregulated DEGs represented non-coding genes (Fig 3C). Astrocyte downregulated DEGs were enriched in gliogenesis (GO:0042063(POU3F2,VIM),−log₁₀P=4.13) and trans-synaptic signaling (GO:0099537(KCNQ3,SYNJ1),−log₁₀P=6.27), suggesting dysregulation of astrocytes and synaptic regulatory function in AUD (Fig 3F,TableS6). Microglial downregulated genes were enriched in regulation of cell proliferation (GO:0008285(TP53,IFNLR1),−log₁₀P=3.48) and lipid processes (GO:0046467(PLPP1,ELOVL5),−log₁₀P=2.12,Fig 3F,TableS6). Oligodendrocyte downregulated genes were enriched in senescence (R-HSA-2559586(BCL6,TXNIP),−log₁₀P=9.32,Fig 3F,TableS6). Excitatory neurons and inhibitory neurons downregulated genes were enriched in chromosome organization (R-HSA-73886(CENPC,TERF2IP),−log₁₀P=2.07). Downregulated DEGs had some cell type-specific patterns but had more overlap between general functions downregulated in AUD (Fig 3F,TableS6). In addition to AUD-responsive pathways in common between both sexes, such as ECM and immune signaling, we also observed sex-specific changes in gene expression (FigS6–7,TableS8), consistent with previous observations(56,57).

Microglia hdWGCNA modules identify inflammatory and immune targets in AUD

Because the majority of DEGs were observed in glia, we leveraged hdWGCNA to determine disease and subtype-specific gene co-expression network changes. The microglial network contained 8 modules, 3 of which had increased eigengene expression in AUD (Fig 4A,TableS9). Microglia module MG-M2, 360 genes, associated with cluster MG1, was more highly expressed in mild/moderate AUD (Fig 4B–E). Module MG-M2 was also positively correlated with total drinking years (R=0.19,Fig 4C,TableS10). Module MG-M2 was highly enriched in DAM (−log₁₀P=10.07) and interferon response microglia (IRM:−log₁₀P=6.95,Fig 4F,TableS11) and canonical immune functions (28,29) (GO:0071345(NFKB1,IRAK2),−log₁₀P=14.80,Fig 4G,TableS12). Top hub genes for module MG-M2 included known immune targets for AUD such as TLR2, IKBKE, and IL1B(5,55,58) (Fig 4H).

Figure 4: Microglia co-expression modules highlight immune dysfunction and loss of homeostatic regulation in AUD.

Co-expression modules from microglia clusters represented by dot plot as the module eigengene (hME) average expression (color of dot) and the percent of genes expressed (% Expressed) in each module (size of dot) as a function of (A) AUD classification or (B) cell type cluster. Red indicates increased average expression, blue indicates decreased average expression of the module. (C) Module-trait correlation heatmap for specific traits: Alcohol intake (g/day), DSM-5, Total drinking years, AUD Classification, Liver class and Sex. Color is a function of correlation strength, with red representing positive correlation and blue representing negative correlation with module eigengene expression, *R > |0.2|. (D) UMAP plot colored by modules eigengenes for microglia co-expression module MG-M2. (E) hME expression of microglia co-expression module MG-M2 across DSM-5 categories. (F) MG-M2 enrichment for microglia subtypes (DAM, IRM, PAM and HLA). (G) Top 10 GO category enrichments for MG-M2. (H) Visualization of top hub genes for MG-M2. Genes marked with an asterisk (*) represent hub genes that are also differentially expressed in microglia clusters. (I) UMAP plot colored by modules eigengenes for MG-M3. (J) hME expression of microglia co-expression module MG-M3 across DSM-5 categories. (K) MG-M3 enrichment for microglia subtypes (HLA, DAM, WAM). (L) Top 10 GO category enrichments for MG-M3. (M) Visualization of top hub genes for MG-M3. (N) UMAP plot colored by modules eigengenes for microglia co-expression module MG-M7. (O) hME expression of microglia co-expression module MG-M7 across DSM-5 categories. (P) MG-M7 enrichment for homeostatic microglia subtype. (Q) Top 10 GO category enrichments for MG-M7. (H) Visualization of top hub genes for MG-M7. Enrichment is a function of −log10 (P-value). For the violin plots in E, J, O values extend from minimum to maximum, with a white boxplot to show the median value. Color is a function of DSM-5 category. Two-tailed Welch’s t-test for DSM-5 module eigengene expression comparisons, * P<0.05, **P<0.01, ***P<0.001, ****P<0.0001.

Similarly, module MG-M3, 103 genes, was highly expressed in cluster MG1 and significantly increased in mild/moderate/severe AUD (Fig 4A,B,I,J). Module MG-M3 overlapped with antigen-presenting microglia (HLA:−log₁₀P=13.49), DAM (−log₁₀P=11.64) and white matter-associated microglia (WAM:−log₁₀P=8.82,Fig 4K,TableS11)(28,29,31). Module MG-M3 also had strong enrichment in Parkinson’s disease (hsa05012(VCP,UQCRC1),−log₁₀P =40.66) and cellular response to stress (R-HSA-9711123(LAMTOR5,COX5B),−log₁₀P=15.75,Fig 4L,TableS12). Hub genes included neurodegenerative targets such as PARK7, which is a potential biomarker of AUD(59) (Fig 4M,TableS7). Module MG-M7, 79 genes, was expressed in cluster MG1 and upregulated in mild/moderate AUD (Fig 4A,B,N,O,TableS9). Module MG-M7 positively correlated with total drinking years (R=0.25,Fig 4C,TableS10). Module MG-M7 only overlapped with homeostatic microglia markers (−log₁₀P=10.04,Fig 4P,TableS11) and was enriched in canonical pathways such as the complement system(60) (WP5090(CX3CR1,C3),−log₁₀P=2.70,Fig 4Q,TableS12), suggesting a regulatory role of microglia in synapse elimination in AUD(61,62). Hub genes included homeostatic and complement markers (TGFB1,CX3CR1,C3,Fig 4K,TableS9). The identification of microglia subtypes, suggests novel microglial disease-states and networks in AUD(63). Furthermore, identified co-expression pathway targets such as interferon signaling are promising since interferon signaling is downstream of targets that regulate behavior(64–66).

Astrocyte and oligodendrocyte modules highlight roles for synapse regulation and inflammation in AUD

We additionally constructed networks for the other major glial cell types that were the primary drivers of differential expression, the astrocyte network consisted of four modules, all of which had increased eigengene expression in AUD (Fig 5A,TableS9). Of interest was module Ast-M1, containing 2005 genes, primarily expressed in Ast3, upregulated in AUD and correlated with total drinking years (R=−0.21) and liver class (R=0.27,Fig 5B–E,TableS10). Module Ast-M1 was enriched in modulation of chemical synaptic transmission (GO:0050804(YHWAG,GRIA3),−log₁₀P=59.92,Fig 5F,TableS12), consistent with the SNAP-a signature(19). Ast-M1 was also enriched in immune response categories (R-HSA-1280218(PAK1,CDC42),−log₁₀P=27.96,TableS12). Hub genes for module Ast-M1 highlight SNAP-a markers such as SNAP25 and SYT1 (Fig 5G). Another module of interest was module Ast-M3, 70 genes, which was primarily expressed in cluster Ast2, upregulated in AUD and positively correlated with total drinking years (R=0.25,Fig 5B–C,H–I,TableS9–10) and highly enriched in the ECM (M5884,−log₁₀P=5.59;R-HSA-1638091,−log₁₀P=5.10,Fig 5J,TableS12). Hub genes included ECM components such as VCAN, TNC and GPC6 (Fig 5K,TableS9), reinforcing the role of the ECM in AUD.

Figure 5: Astrocyte co-expression modules indicate inflammatory phenotype with synapse dysregulation in AUD.

(A) Co-expression modules from astrocyte clusters represented by dot plot as the module eigengene (hME) average expression (color of dot) and the percent of genes expressed (% Expressed) in each module (size of dot) as a function of (A) AUD classification or (B) cell type cluster. Red indicates increased average expression, blue indicates decreased average expression of the module. (C) Module-trait correlation heatmap for specific traits: Alcohol intake (g/day), DSM-5, Total drinking years, AUD Classification, Liver class and Sex. Color is a function of correlation strength, with red representing positive correlation and blue representing negative correlation with module eigengene expression, *R > |0.2|. (D) UMAP plot colored by modules eigengenes for astrocyte co-expression module Ast-M1. (E) hME expression of astrocyte co-expression module Ast-M1 across DSM-5 categories. (F) Top 10 GO category enrichments for Ast-M1. (G) Visualization of top hub genes for Ast-M1. Genes marked with an asterisk (*) represent hub genes that are also differentially expressed in astrocyte clusters. (H) UMAP plot colored by modules eigengenes for Ast-M3. (I) hME expression of astrocyte co-expression module Ast-M3 across DSM-5 categories. (J) Top 10 GO category enrichments for Ast-M3. (K) Visualization of top hub genes for Ast-M3. For the violin plots in E, I values extend from minimum to maximum, with a white boxplot to show the median value. Color is a function of DSM-5 category. Two-tailed Welch’s t-test for DSM-5 module eigengene expression comparisons, * P<0.05, **P<0.01, ***P<0.001, ****P<0.0001.

The oligodendrocyte network consisted of 5 modules, two of which had increased eigengene expression in AUD and correlated with alcohol phenotypes (Fig S8A–C,TableS9). Of these two modules, Oligo-M2 was enriched in neurodegenerative pathways (hsa05012(CALM1,SOD1),−log₁₀P=55.19,FigS8D–G,TableS9,12) and Oligo-M4 was enriched in homeostatic oligodendrocyte functions (GO:0042552(MBP,MOBP),−log₁₀P=17.62,FigS8H–K,TableS9,12).

Excitatory neuronal modules show modest changes in synaptic organization

As the second largest class of DEGs, we constructed gene co-expression networks for excitatory neurons. Clusters Excit1-7 showed a distinct signature in comparison to Excit8-11, which had no significantly correlated modules (FigS9). Excit1-7 had nine gene co-expression modules, six of which had increased eigengene expression in AUD with one module correlating with total drinking years (Excit-M3,R=0.42,FigS10A–C,TableS9–10). Module Excit-M3 was enriched in Fragile X syndrome (WP4549(BRAF,CUL5),−log₁₀P=8.95) as well as alcoholism (hsa05034(BDNF,CAMKK1),−log₁₀P=2.47,FigS10D–F,TableS9,12). Hub genes include genes previously implicated in AUD such as GRIN2A and HOMER1(67) (FigS10G). The inhibitory neuron network consisted of twelve gene co-expression modules, none of which correlated with alcohol traits (FigS11,TableS9–10). The modest correlations between alcohol traits and neuronal gene co-expression networks in contrast to the strong correlations observed in glial networks further suggests the importance of glial co-expression networks in AUD and their therapeutic potential.

GWAS enrichment meta-analysis highlights role for glia in AUD risk

We next used MAGMA to identify which cell types harbor aggregated genetic risk for AUD. Notably, most cell types enriched for AUD-associated SNPs were glial (Fig 6,TableS13), with the strongest glial enrichments observed for differential expression in microglia and OPCs (Fig 6A–B). Astrocytes and neurons were enriched for AUD risk in the context of specific hdWGCNA modules (Fig 6C). By integrating genetic variants that show a causal role in genomic analyses and exhibit transcriptomic differences between AUD and controls we are able to identify gene expression changes within these cell types that are potentially causal to AUD risk.

Figure 6: GWAS meta-analysis enrichment for alcohol phenotypes highlights glial risk for AUD disease progression.

Heatmaps of GWAS enrichment using MAGMA for alcohol dependence and consumption AUDIT (AUDIT-C). (A) Enrichment of specific cell type clusters for significant differentially expressed genes (DEGs) (B) Enrichment of aggregate cell type classes for significant DEGs. (C). Enrichment of hdWGCNA modules. Color is a function of −log10 (P-value), with dark blue representing stronger enrichment, light blue representing weaker enrichment and white representing no enrichment. *P<0.05, **P<0.01.

Proteome secondary validation replicates broad glial transcriptomic changes observed in AUD

Finally, we validated our snRNA-seq data with a recent AUD dlPFC proteome(36). We determined a significant degree of overlap between concordant differentially abundant proteins and DEGs(68) (FigS12B–C,TableS14). We additionally tested consistency of overlap between the AUD proteome and our data, which showed modest consistency (FigS12D–H,TableS14). Pathway analysis of the overlapping genes/proteins highlighted categories observed in both our dataset and the alcohol field such as the role of the ECM, inflammatory response and dysregulation of neurotransmitter release (FigS12I–J,TableS14). As well as additional novel categories such as interferon response, which is consistent with our differential expression and microglia network analysis (FigS12H).

Discussion

The findings presented here substantially refine our understanding of AUD molecular pathology and risk in human brain, with a high-resolution dataset of the dlPFC, the largest study of this kind to date. The quality and power of the dataset allowed us to investigate cell type-specific dysregulation of gene expression in AUD across 32 clusters, capturing significantly more DEGs in more subpopulations than previous single-nuclei studies(11,12). Furthermore, we are the first to perform hdWGCNA analysis to identify co-expression networks in individuals with AUD on the single-nuclei level. These co-expression networks identified novel regulatory players that correlated with alcohol-specific traits. A significant overlap of these genes were validated using available AUD proteome data, suggesting these gene expression changes are long-lasting and can affect behavior(36). Finally, the integration of single-nuclei clusters with GWAS data revealed specific glial subtypes that are potentially causal in development of AUD risk.

A notable finding of this study is the subtypes of astrocytes and microglia identified in AUD and the robust differential gene expression and network changes in these cell types. Because both cell types have already been linked to alcohol behaviors, dissecting the subtypes proved a powerful tool to identify novel immunomodulatory mechanisms and therapeutic targets.

In astrocytes, we identified alcohol-responsive subtypes: homeostatic, inflammatory and SNAP-a. Of note, a novel finding is the identification of the recently reported SNAP-a program(19). Specifically, decreased expression of SNAP-a in AUD indicated a novel loss of homeostatic control of synaptic regulation in favor of inflammatory functions in astrocytes. The changes in homeostatic and inflammatory astrocytes have previously been identified in both our pilot snRNA-seq study and mouse models of alcohol dependence(11,69). One of the largest categories of DEGs in astrocytes upregulated in AUD was CSPGs and other components of the ECM, which play a key role in astrocyte-neuron communication. CSPG and ECM enrichment was also observed in module Ast-M3, which correlated with total drinking years. Local ECM remodeling, dependent on CSPG response to inflammation, has been posited as a crucial component underlying neuropsychiatric disorders(70,71). In individuals with AUD, ECM dysregulation has recently been reported in the orbitofrontal cortex on the protein level(72). Animal models have also found CSPGs are elevated after binge drinking(73) and have a regulatory role in aversion-resistant(74) and intermittent drinking(75). Importantly, inflammatory mediators such as interferons and TNFs modulate ECM remodeling, which can impact behavior. This capacity for regulation of behavior may stem from ECM remodeling’s role in regulation of synaptic plasticity(71). Similarly, in the dlPFC from individuals with opioid use disorder, there was significant enrichment in factors that control ECM remodeling, highlighting a common neuroimmune-ECM interaction in substance use disorders(76). Furthermore, proteome integration replicated several key ECM proteins and demonstrated overlapping enrichment of ECM function(36). Taken together this suggests a strong link between neuroinflammation, ECM remodeling, synaptic regulation, and AUD pathology.

In microglia, we identified two clusters that were alcohol-responsive and represented loss of homeostatic control leading to an inflammatory phenotype—a common microglial mechanism observed in other neurological disorders(31,63,77). Beyond a simple dichotomy, we were able to identify novel microglia subtypes in AUD, including BAMs and DAMs. BAMs are localized between the brain parenchyma and the periphery, leading to a pivotal role in immune surveillance and inflammatory regulation(27). Recently, BAMs have been implicated in the pathogenesis of neurodegenerative disorders via regulation of immune cell infiltration into the parenchyma(78). Similarly, DAMs have also been implicated in the pathogenesis of neurodegenerative disorders due to their downregulation of homeostatic genes and upregulation of genes involved in lysosomal, phagocytic and lipid metabolism(29,63). Network analysis further uncovered the importance of this microglia subtype in regulation of co-expression. Microglia module MG-M2, which correlated with total drinking years and was upregulated in AUD, was highly enriched in loss of homeostatic microglia and upregulation of DAM and interferon-responsive microglia. Hub genes in MG-M2 have been shown to regulate alcohol behaviors and synaptic functions. TLR2 decreases alcohol intake in mouse models(79) and IL1B increases GABA release via proinflammatory pathways in alcohol-dependent mice(80). NEAT1, a lncRNA upregulated in alcohol-dependent mice and a hub gene in MG-M2, has also been shown to mediate and promote inflammation in cell culture and animal models(69,81,82). Proteome integration replicated the importance of DAM/BAM subpopulations as well as novel pathways like interferon signaling(36). The recruitment of these microglial subpopulations in AUD may have a causative effect since ablation of microglia blunts escalation of drinking after alcohol dependence(44). Moreover, the enrichment of AUD risk in cluster MG1 further causally connects these shifting inflammatory microglial subtypes with regulation of both AUD risk and pathogenesis. However, future studies will be required to link inflammatory changes to specific subtypes to develop microglia subtype-specific treatments.

Although most alcohol-responsive populations and DEGs were observed in glial clusters, we ultimately consider these glial effects as a concerted neural response leading to neuronal changes that alter behavior. These changes likely occur via alcohol’s direct interactions with neurons and the neuroinflammatory effect of alcohol on glial populations that alter neuronal firing(55,67). Inhibitory and excitatory neurons represented ~25% of DEGs and the categories enriched in these subclusters were canonical to AUD such as glutamatergic/GABAergic signaling, as well as novel links to Fragile X Syndrome. One previous study linked ethanol-dependent changes in GABA_BR and dendritic plasticity to FMRP, suggesting FMRP is an important regulator of synaptic plasticity following alcohol exposure(83). Additionally, genetic risk for AUD was enriched in inhibitory clusters/modules as well as the excitatory modules mentioned above. How these neuronal firing changes are linked with neuroinflammatory mechanisms that effect AUD risk and pathology will be investigated in future.

Our study is not without limitations. Almost all individuals were of European ancestry(n=71), limiting the generalizability of our results. Most individuals included in our study were also male(N=54), with a small cohort of females(N=19). We included sex as a covariate in the differential expression analysis as to not bias our results, however, sex-specific gene expression was identified when analyzed separately. Additionally, although our cohort is pre-screened to remove comorbid mood disorders, mild depression and anxiety was observed in AUD individuals at a higher proportion than controls, which may alter or worsen the transcriptomic dysregulation observed in AUD(84). Technical limitations with snRNA-seq of human brain samples have been previously described, including the effect of batch correction(85). We, like others, found a much greater proportion of neurons compared to glia but due to the sample size and extraction technique were able to capture more non-neuronal cells than typically reported. Recent studies reported ~1.6-2.1% microglia(16,86); whereas, we captured 4.9% microglia, allowing us to report on novel subpopulations in AUD for the first time. Moreover, despite droplet-based snRNA-seq not capturing genes expressed at very low levels, we were still able to identify differential gene expression for thousands of genes in defined cell types.

Our data illustrates the utility of this type of complex, single-nuclei data in understanding the molecular etiology of AUD pathology and risk. We have demonstrated individuals with AUD have broad transcriptomic dysregulation in multiple cell types, specifically in glia subpopulations. Moreover, our results show a specific role for novel glial subtypes, especially DAM/BAM microglia and SNAP-a astrocytes, underlying genetic risk of AUD. We expect that this dataset will be crucial for designing informed experiments to identify the biological mechanisms and novel drug targets in AUD in future.