Identifying novel gene dysregulation associated with opioid overdose death: a meta-analysis of differential gene expression in human prefrontal cortex

Abstract

Only recently have human postmortem brain studies of differential gene expression (DGE) associated with opioid overdose death (OOD) been published; sample sizes from these studies have been modest (N = 40–153). To increase statistical power to identify OOD-associated genes, we leveraged human prefrontal cortex RNA-seq data from four independent OOD studies and conducted a transcriptome-wide DGE meta-analysis (N = 272). Using a unified gene expression data processing and analysis framework across studies, we meta-analyzed 20, 098 genes and found 335 significant differentially expressed genes (DEGs) by OOD status (false discovery rate < 0.05). Of these, 66 DEGs were among the list of 303 genes reported as OOD-associated in prior prefrontal cortex molecular studies (e.g., genes/gene families OPRK1, NPAS4, DUSP, EGR). The remaining 269 DEGs were not previously reported (e.g., NR4A2, SYT1, HCRTR2, BDNF). There was little evidence of genetic drivers for the observed differences in gene expression between opioid addiction cases and controls. Enrichment analyses for the DEGs across molecular pathway and biological process databases highlight an interconnected set of genes and pathways linked to orexin and tyrosine kinase receptors through MEK/ERK/MAPK signaling to affect neuronal plasticity.

Introduction

The opioid epidemic continues to be a tremendous burden on our society and communities around the world. In the United States, 5.6 million people ages 12 and older misused opioids during 2021 [1, 2]. In the same year, the United States saw the highest 12-month count of opioid overdose deaths (OOD) recorded, >80,000, a 40% increase since 2019 [1]. Although our understanding of the neurobiology of opioid addiction remains limited, increased sample sizes from recent meta-analyses of genome-wide association studies (GWAS) have begun to identify genetic loci robustly associated with opioid addiction, including loci in or near OPRM1, FURIN, and SCAI/PPP6C/RABEPK [3–5]. However, such robust findings for critical features of gene regulation in the human brain, and gene expression in particular, have yet to emerge.

Only recently have human postmortem brain studies of differential gene expression (DGE) associated with OOD been published: Corradin et al. [6]; Mendez et al. [7]; Seney et al. [8]; and Sosnowski et al. [9]. All four of these studies used human postmortem dorsolateral prefrontal cortex (DLPFC) brain tissue from donors identified as dying from OOD through toxicology assays administered by forensic scientists and/or phenotypic evidence of opioid addiction, unavoidably combining the potential gene dysregulation associated with acute overdose and chronic exposure to opioids. Each of these independent studies had modest sample sizes (N = 40–153) and compared bulk RNA-seq data from individuals who died from OOD to individuals who died from non–drug use causes. The DLPFC region of the brain involves the preoccupation/ anticipation component of the addiction cycle, which affects craving, impulsivity, and executive function [10]. The DLPFC has also been linked to DGE levels studies of OPRM1 [11] and associated with anxiety and impulsive neurological disorders (ADHD, schizophrenia, bipolar disorder, etc.). Addictive and reward-response phenotypes are also traits reported to be linked to the DLPFC from human and animal model studies, which can lead to drug-seeking behavior.

Each DLPFC DGE study reported OOD-associated genes with plausible biological links to addiction. However, with their sample sizes, statistical power to robustly identify OOD-associated DGE is limited. Here, we uniformly processed the bulk RNA-seq data across the four OOD studies and performed a DGE meta-analysis including 272 samples (case N = 137, control N = 135), making this the largest transcriptome-wide analysis of OOD to date. Our findings include novel differentially expressed genes (DEGs) in addition to confirming a subset of previously reported OOD-associated genes. We link our meta-analysis DEGs to different biological processes and pathways, notably the orexin receptor signaling system and signaling by receptor tyrosine kinases. Furthermore, we investigated genetic regulation of OOD-associated DEGs and assessed shared genetics between these genes and 47 GWAS traits through partitioned heritability and colocalization analyses [12–14].

Methods

Contributing study cohorts

Characteristics of the four cohorts contributing to this meta-analysis are provided in Table 1, reflecting post-QC numbers. Each of the contributing studies defined cases as decedents whose death was attributed to opioid overdose based on toxicological analyses by coroners’ offices and/or phenotypic evidence of a history of opioid misuse or opioid addiction. Details of decedent identification, tissue processing, and data generation are provided in the original studies. Here we briefly review each study’s OOD case and control definitions. Corradin study: The study included opioid cases (N = 51) and unaffected controls (N = 51). The cause of death was determined through medico-legal investigations, including autopsy findings, medical records, and toxicology reports. Opioid cases had documented opioid abuse history, positive toxicology, and forensic confirmation of opioid-related death. Controls were age-matched, with no history of drug use and negative toxicology. Exclusion criteria included suicide, psychiatric disorders, or chronic illness [6]. Mendez study: The study included OUD cases (N = 29) and controls (N = 18). Postmortem brains were obtained from UTHealth Brain Collection in collaboration with the Harris County Institute of Forensic Science. Demographics, medical records, autopsy, and toxicology reports were reviewed. A psychological autopsy was conducted via next-of-kin interviews. OUD diagnosis was assigned using DSM-5 criteria by an independent panel. RNA-seq data was generated for a subset of decedents (N = 27 OUD, N = 14 controls). Most OUD cases tested positive for opioids at death. Causes of death included overdose (N = 22), cardiovascular disease (N = 4), and other causes [7]. Seney study: The study included OUD cases (N = 20) and unaffected controls (N = 20), matched by sex and age. OUD cases had a diagnosis for at least five years. DSM-IV diagnoses were determined through psychological autopsy, structured family interviews, and medical record reviews. Controls had no lifetime psychiatric or neurologic disorders [8]. Sosnowski study: The study included opioid cases (N = 72), psychiatric controls (N = 53), and non-psychiatric controls (N = 28). Among opioid cases, 66 had OUD, two had cocaine use disorder, and four did not meet criteria for substance use disorder despite opioid overdose. Non-psychiatric controls had no psychiatric or substance use diagnoses, with negative toxicology. Psychiatric controls had DSM-V diagnoses, possibly including substance use disorders, but did not die of opioid overdose. A retrospective clinical review was conducted using autopsy, toxicology, psychiatric records, and family informant interviews. Diagnoses were determined by two board-certified psychiatrists [9].

Table 1.

Samples from all four studies used in the meta-analysis.

	Corradin et al.		Mendez et al.		Seney et al.		Sosnowski et al.
	Control (N = 23)	Case (N = 21)	Control (N = 13)	Case (N = 27)	Control (N = 20)	Case (N = 20)	Control (N = 79)	Case (N = 69)
Self-reported Sex
Male	20 (87.0%)	17 (81.0%)	11 (84.6%)	15 (55.6%)	10 (50.0%)	10 (50.0%)	44 (55.7%)	50 (72.5%)
Female	3 (13.0%)	4 (19.0%)	2 (15.4%)	12 (44.4%)	10 (50.0%)	10 (50.0%)	35 (44.3%)	19 (27.5%)
Age at Death
Mean (SD)	35.3 (12.4)	36.8 (10.2)	54.8 (15.2)	39.1 (13.2)	47.3 (9.49)	46.9 (7.27)	37.3 (9.72)	33.6 (8.85)
Median [Min, Max]	35.0 [18.0, 56.0]	39.0 [21.0, 51.0]	61.0 [17.0, 74.0]	36.0 [19.0, 70.0]	48.0 [23.0, 60.0]	45.5 [35.0, 59.0]	36.8 [18.1, 49.7]	33.4 [18.6, 49.8]
Race/Ethniciy
White	14 (60.9%)	11 (52.4%)	7 (53.8%)	22 (81.5%)	13 (65.0%)	19 (95.0%)	54 (68.4%)	58 (84.1%)
African American	9 (39.1%)	10 (47.6%)	3 (23.1%)	5 (18.5%)	7(35.0%)	1(5.0%)	25(31.6%)	10(14.5%)
Asian	0 (0%)	0 (0%)	1 (7.7%)	0 (0%)	0 (0%)	0 (0%)	0 (0%)	0 (0%)
Hispanic	0 (0%)	0 (0%)	2(15.4%)	0 (0%)	0 (0%)	0 (0%)	0 (0%)	0 (0%)
Multi-Racial	0 (0%)	0 (0%)	0 (0%)	0 (0%)	0 (0%)	0 (0%)	0 (0%)	1 (1.4%)

All control sample patients had no history with opioid use disorder. All cases sample patients had a record of prior exposer to opioids before the overdose event.

Information pertaining to patient OUD history and/or case control descriptions can be found in the methods section of the manuscript.

Data processing

Paired-end RNA-seq FASTQ files were obtained from four publicly available datasets (Sequence Read Archive study ID SRP324812, Sosnowski et al. [9], Gene Expression Omnibus accession IDs GSE174409, Seney et al. [8] and GSE182321, Mendez et al. [7], and dbGaP study phs002724.v1.p1, Corradin et al. [6]) and processed through a unified workflow. All studies used ribo-depletion prior to RNA sequencing or PolyA enrichment. Briefly, adapters were trimmed and reads filtered using Trimmomatic v0.39 [15]. Reads were then pseudo-mapped using Salmon v1.1.0 [16] in selective alignment mode [17] using the GENCODE v30 comprehensive gene annotation as the transcriptome index and the full GRCh38 primary genome assembly as a selective alignment decoy sequence. Salmon transcript quantifications (mapping percentage > 30%) were aggregated to the gene level using tximport v1.12.3 [18]. Quality metrics for raw and post-Trimmomatic reads (retained reads percentage > 60%) were generated using FASTQC v0.11.8 [19], and reads were aligned to the GRCh38 genome using HISAT2 v2.1.0 [20] to generate additional quality metrics. All quality metrics and read mapping statistics were aggregated using MultiQC v1.7 [21]. Samples were then retained based on several quality control criteria (e.g., effective sequencing depth > 10 M, read GC content between 35 and 65%, transcriptome mapping perfectage > 50%, RNA integrity number (RIN) score > = 5, mitochondrial genome mapping percentage < 50%, and Ribosomal RNA mapping percentage < 1%. The sample sizes provided in Table 1 are the samples available after applying QC filters.

Differential gene expression analysis

For each dataset, a regression model was fit using limma with voom-transformed count data [22]. Prior to model fitting, lowly expressed genes were removed (<10 gene counts in the proportion of samples that comprise the smaller OOD status group). Applying a uniform set of regression covariates to analyses of each cohort resulted in poorly controlled false positives (lambda > 12; see Supplementary Fig. 1A). To account for dataset-specific characteristics, covariates included in the regression model for each dataset reflected those from their respective published studies. This approach substantially improved control of false positives (lambda < 3; see Supplementary Fig. 1B). All models included OOD status, age, sex, post-mortem interval (PMI), and RIN as covariates. Additionally, the Corradin model included race, sequencing batch, and 12 surrogate variables estimated by SVA [23]. The Mendez model included cerebellar pH. The Seney model included race and brain tissue pH. The Sosnowski model included race, cocaine/amphetamine toxicology report status, ribosomal RNA mapping rate, gene assignment rate, mitochondrial RNA mapping rate, concordant read pair mapping rate, overall mapping rate, External RNA Control Consortium spike-in error rate, and 10 quality control surrogate variables [23, 24].

Meta-analysis

To combine evidence of differential expression across datasets, we use the samples sizes and differential gene expression analysis p-values from our independent analyses of each study to conduct a weighted Fisher’s meta-analysis implemented by the R package metapro [25]. MetaPro combines gene-level p-values across studies using a statistical framework that supports multiple aggregation approaches; in our analysis, we selected weighted Fisher’s method, which calculates a combined test statistic from study-specific p-values and sample sizes. The resulting statistic follows a chi-square distribution under the null hypothesis of no differential expression, allowing us to detect genes that show consistent signals of differential expression across studies, even when effect sizes may vary. This method assumes independence between studies and is well-suited for integrating results from multiple transcriptomic datasets.

Any genes that were only tested in one study (because they were lowly expressed in the others) were removed, and a Benjamini-Hochberg FDR threshold of < 0.05 was applied to declare a gene as significantly differentially expressed.

Gene set overrepresentation analysis

The gene set overrepresentation analysis was conducted using the ToppGene Suite, ToppFunn:Transcriptome, ontology, phenotype, proteome, and pharmacome annotations-based gene list functional enrichment analysis [26], to identify GO terms, biological pathways, and disease-annotated gene sets enriched for meta-analysis DEG. GO terms included gene sets from the molecular function, biological process, and cellular components categories [27, 28]. Biological pathways were from the MsigDB C2 BIOCARTA collection (v7.5.1) [29–31]. Disease-annotated gene sets were sourced from DisGeNET (BeFree &Curated) [32] (Supplementary Table 1). The R package simplifyEnrichment [33] was used to do semantic similarity-based hierarchical clustering of GO biological process terms using the Wang et al. [34] distance metric. Binary cut was used to determine GO term clusters. Nine other clustering methods were compared against the binary cut method (Supplementary Fig. 2), but were outperformed in similarity scores and cluster quantity by binary cut.

Partitioned heritability analysis

In our study, we utilized sLDSC [35] to assess the heritability of 47 GWAS traits captured by the genomic regions spanning the OOD-associated differentially expressed genes (DEGs). Of these traits, 41 have been previously examined in relation to OUD [5] and 6 in relation to sleep (i.e., insomnia and differential sleep duration) [36, 37] (Supplementary Table 2) For each of the 335 meta-analysis DEGs, we extracted the genomic region encompassing the gene body and 100 kilobases upstream and downstream of the gene ends. These regions, along with GWAS summary statistics for the 47 traits, were provided as input for sLDSC. We utilized the GENCODE v30 [17] comprehensive gene annotation GTF file for GRCh37 to obtain gene start and end coordinates and converted the GWAS summary statistics for each trait to the build GRCh37 format. An LD reference panel derived from the 1000 Genomes Phase 3 EUR superpopulation was used for the LD scores. The munge_sumstats.py script from the LDSC GitHub repository was used to ensure consistency in the formatting of the GWAS summary statistics.

Differential cell-type proportion testing

Cell types were inferred using the BISQUE deconvolution tool [38] using DLPFC single-nuclei RNA-seq data [39] as the reference panel. Several different cell-type deconvolution methods (meta vs. mega-analysis frameworks), reference datasets [39, 40] (Tran et al. and Brenner et al.), and modeling approaches (beta regression and arcsin regression) were investigated prior to selecting the described method below. The proportions of the nine cell types present in the reference (astrocytes, GABAergic neurons, excitatory neurons, macrophages, microglia, mural cells, oligodendrocytes, oligodendrocyte precursors, and T cells) were estimated in the bulk RNA-seq data from the four studies. To determine whether cell-type proportions differed by OUD status for each study, cell-type proportions were arcsin transformed, and a linear regression model was fit for each cell type with the transformed proportions as the outcome variable. The explanatory variables for all models included age, sex, RIN, PMI, and OOD status. For each study, additional covariates were included (Corradin: sequencing batch and race; Mendez: cerebellar pH; Sosnowski: cocaine/amphetamine toxicology report status, ribosomal RNA mapping rate, concordant reads mapping rate, overall mapping rate, External RNA Control Consortium spike-in error rate, and race). Two-sided t-tests were conducted within each dataset to assess study-specific associations between OOD and cell type proportion. Finally, a differential cell-type proportion meta-analysis p-value was calculated using the Fisher’s method as implemented with the sumlog function from metapro v1.8 R package [41]. Cell-type proportions were considered as significantly different by OOD status for Bonferroni-adjusted meta-analysis p-value < 0.05 (Supplementary Table 3).

Study reported gene list

To construct a list of DEGs identified from prior DGE studies the thresholds used in the prior studies and significance reported by each respective study were applied, and the DEG lists were then concatenated (N = 303): Corradin et al., [6] (Bonferroni corrected p-value < 0.05 [N = 10]); Mendez et al., [7] (FDR p-value < 0.05 and |FC| > 1.5” [N = 29]); Seney et al., [8] (FDR p-value < 0.01 and log₂ FC > + 0.26 [i.e., FC + 1.2 or 20% expression change], focused on the top 250 genes of 567 [N = 250]; Sosnowski et al., [9] (FDR corrected p-value < 0.10 and log₂ FC < −1.5, with the addition of three extra genes [N = 4]).

Expression quantitative trait loci look-up for meta-analysis DEGs

We used the eQTL mapping summary statistics from the GTEx single-tissue cis-eQTL datasets [42, 43] “Brain_Frontal_Cortex_BA9.v8.signif_variant_gene_pairs.txt” (significant frontal cortex variant-gene associations in GTEx v8) and “Brain_Forntal_Cortex_BA9.v7.signif_variant_gene_pairs.txt” (significant frontal cortex variant-gene associations in GTEx v7) for colocalization analysis with summary statistics from an opioid addiction GWAS. Summary statistics were standardized to the GRCh37 reference coordinates. Ensembl IDs for the 335 meta-analysis genes were used as look-up IDs for genes in the eQTL summary statistics. The eQTL summary statistics for matched genes included only variant-gene associations with permutation-based p-values < 0.05.

Colocalization of loci for cross-brain region eGenes and GWAS traits

To perform colocalization analysis between cross-brain region eGenes and 37 GWAS traits (Supplementary Table 4, a subset of phenotypes used for the partitioned heritability analysis), through SYNAPSE (https://www.synapse.org/Home:x) we obtained publicly available summary statistics from Hibar et al. cross-brain region eQTL meta-analysis and the corresponding variant coordinate annotations [44]. For each meta-analysis DEG, we tested for colocalization of eQTL and GWAS signal. Colocalization was conducted using the coloc v3.2.0 R package [45]. All summary statistics used were in genome build GRCh37/hg19. Using the coloc.abf() function, the eQTL data were specified as a quantitative trait while the GWAS data were specified as case-control for binary traits or quantitative otherwise. For the Synapse summary statistics, p-value and European MAF (1000 Genome Phase 3) were provided to Coloc (beta and variance is not available). Due to unit scaling in the Synapse eQTL meta-analysis methods, sdY = 1 was provided for every gene. For the GWAS summary stastistics, beta and variance were provided. Only variants which overlapped the eQTL summary statistics, the GWAS statistics, and 1000 G MAF reference were used (Supplementary Table 5). Colocalization hypothesis 4 (both traits have a causal SNP in the region, and the two causal SNPs are the same) posterior probability > 0.7 was considered as a significant colocalization [46].

Enrichment of meta-analyzed DEGs in MAGMA

To evaluate whether meta-analyzed DEGs were enriched for GWAS associations, we obtained publicly available summary statistics for 34 GWAS traits (same input traits used in colocalization analysis) and processed them through Functional Mapping and annotation of genetic assiociations (FUMA) [46] to run a multi -marker analysis of GenoMic Annotation (MAGMA) [47] gene-based tests. For each trait, Benjamini–Hochberg (BH) correction was applied to generate FDR values, and genes with FDR < 0.05 were considered significant. We then constructed 2 × 2 contingency tables (DEG vs. non-DEG × MAGMA significant vs. not) for each GWAS and performed Fisher’s exact tests. Enrichment testing was possible in 14 of the 22 GWAS traits represented in the MAGMA results, and Fisher’s test p-values were corrected across traits using BH.

Results

Differential gene expression meta-analysis

To increase statistical power and further our understanding of OOD, we leveraged the combined sample sizes of the four studies in a meta-analysis (N = 272, cases = 137, controls = 113). We used the gene expression data generated from our uniform RNA-seq data processing workflow and fit regression models consistent with those used by the original studies (see Methods). Next, we combined resulting DGE summary statistics (yielding N = 20 098 genes) and combined p-values using a weighted Fisher’s method. Most of the tested genes were present in all four studies with the remaining tested genes occurring in two or three of the four studies (N = 20 098). Genes present in only one study were excluded from the meta-analysis (N = 1 781; see Supplementary Fig. 3). The meta-analysis resulted in 335 significant DEGs (false discovery rate [FDR] < 0.05) as seen in Fig. 1A and B. This approach is standard in meta-analytic frameworks, where the goal is to combine evidence of consistent differential expression across multiple datasets even when the individual effect sizes (SMDs) may be small. It is important to note that a gene with a small SMD can still be statistically significant (i.e., meet the FDR threshold) if the signal is consistent across datasets and the variance is low. Conversely, large SMDs may not reach significance if they are inconsistent or highly variable across studies. Thus, the presence of DEGs with SMDs close to zero reflects the use of a meta-analytic approach based on p-value aggregation rather than effect size magnitude alone.

Fig. 1. Differentially expressed gene results from meta-analysis.

A. Volcano plot (FDR-BH corrected p-value < 0.05) for genes tested in the differential gene expression meta-analysis. Differential expression magnitude is represented here using the standardized mean differences (SMD) across studies. A total of 169 up-regulated (red triangle) and 166 down-regulated (blue triangle) significant genes; (B). MA plot (FDR-BH corrected p-value < 0.05) using the SMD; (C). Biotypes of the 335 significant genes from meta-analysis partitioned into percentages of the various different types of genetic biomarker types; (D). Venn diagram representing how many genes were reported genes (significantly expressed genes documented in the four studies described above), meta-analysis genes (significant differentiated expressed genes Bonferroni corrected), and the 66 consistent DE genes (number of genes that are present in both previously reported gene list and meta-analysis gene list).

Categorically, 85.2% were protein coding and 10.6% were long noncoding RNA genes (Fig. 1C). Comparing the list of 303 DEGs reported in the prior DGE studies with the 335 DEGs from this meta-analysis, only 66 genes from the prior studies [6–9] were also statistically significant in the meta-analysis, including DUSP2, DUSP4, DUSP6, EGR1, EGR4, ARC, and NPAS4, which are all genes linked and associated to OUD (Supplementary Table 5 & 6) (Fig. 1D). Contribution of each cohort to these results is provided in Supplementary Table 6.

Cell type deconvolution

Previous studies have indicated that opioid use can affect cell-type proportion within the brain [8, 48]. To investigate cell-type proportions in this meta-analysis, we used a single-cell DLPFC reference dataset, Tran et al. [39], to deconvolve the cell-type proportions for each sample across the four independent datasets for nine cell types (astrocytes, GABAergic neurons, excitatory neurons, macrophages, microglia, mural cells, oligodendrocytes, OPCs, and T cells) (Supplementary Table 3). Significant differences in microglia cell-type proportions were observed within the Sosnowski dataset (Bonferroni p-value = 0. 0.002) and in the meta-analysis of cell-type differences across all four studies (Bonferroni p-value = 0.039) (see Supplementary Fig. 4). Although microglia proportions are significantly different between OOD cases and controls, our analysis shows varying directions of effect across cohorts and that the statistical significance in the meta-analysis is primarily driven by the Sosnowski et al. [9] results. Although limited by the reference panel for predicting cell-type proportions in DLPFC, these results suggest that cell type proportions played a minimal role in the observed DGE.

DEG characterization through enrichment analyses

Gene ontology terms

Gene Ontology (GO) enrichment analysis was performed on the significant meta-analysis genes to better understand the roles these genes may play in OOD. The 335 genes were tested for enrichment in GO molecular function (MF), cellular component (CC), and biological process (BP) terms, resulting in 78 significant terms (FDR-BH p-value < 0.05). Of the 78, three were GO MF terms relating to signal transduction and transcription factor activity including “MAP Kinase phosphatase activity” (p-value = 0.0412), “nuclear glucocorticoid receptor binding” (p-value = 0.0412), and “DNA-binding transcription factor activity, RNA polymerase 2 specific” (p-value = 0.0412). Ten of the enriched GO terms were GO CC terms and 65 were GO BP terms that reflected functions in synaptic signaling, chromatin regulation, and synaptic development, all of which are important in synaptic plasticity [49, 50]. Semantic clustering was performed on the enriched CC and BP terms separately. Within the GO CC terms, two clusters were identified: the largest being “vesicles dense cores,” which include synaptic vesicles followed by “acetyltransferase histone complex.” Of the GO BP terms, eight clusters were identified based on z-score; “morphogenesis,” “transcription regulation RNA,” “transsynaptic signaling synaptic,” “memory,” and “response hormone” with the largest z-score (Fig. 2A). The largest cluster of GO BP terms enriched were terms relating to morphogenesis, projection, and neuronal development.

Fig. 2. Comprehensive gene ontology cluster methods results.

A. Heatmap of FDR significant p-values < 0.05 toppgene biological process GO terms N = 65, clustering of semantic similarity, the larger the characteristic, the greater the correlation. B. Dot plot of Reactome databases results from differentially expressed gene list. Entities (nucleic acids, proteins, complexes, vaccines, anti-cancer therapeutics, and small molecules) participating in reactions form a network of biological interactions and are grouped into pathways.

Reactome pathways

To expand biological characterization of the 335 DEGs, with a focus on relationships among signaling and metabolic molecules, we conducted a Reactome pathway enrichment analysis [51]. Of the 335 submitted genes, 182 were present in the Reactome database. Five pathways were significantly enriched for our DEGs (FDR < 0.05; Fig. 2B). The primary pathway was Signaling by Receptor Tyrosine Kinase (RTK): 32 DEGs among the 625 in this pathway, entity ratio 0.04. The remaining four pathways are all nested under the primary pathway, with the enrichments driven by subsets of the 32 DEGs linked with the Signaling by RTK pathway, Signaling by NTRKs, Signaling by NTRK1, Nuclear Events (kinase and transcription factor activation), and Neuronal Growth Factor stimulated transcription. Signaling by RTK is itself a subfamily of the Signal Transduction pathway, constituting a class of cell surface proteins enabling ligand binding for stimulating intracellular cascades. Among other functions, these RTK pathways mediate synaptic plasticity that may be affected by opioid addiction or OOD (Supplementary Table 7 & 8).

When comparing the complete meta-analysis gene list (N = 335) to the 66 genes also previously reported as DEGs via GO similarity clustering, disease association (presented below), and pathway association, an increase in gene count within each annotation was observed (e.g., GO biological processes; Fig. 3A). This result indicates a strengthening of these GO term enrichments with the increase in statistical power under the meta-analysis (Supplementary Fig. 2) (Fig. 3A).

Fig. 3. Meta-analysis DEGs biological inference.

A. Bar plot of overlapping biological process GO terms from both complete meta analysis gene lists (Bonferroni-adjusted p-value < 0.05) and retained genes that were previously reported in the four studies. B. Only Bonferroni corrected p-value < 0.05, enriched gene pathway (orexin receptor pathway) gene list (13 genes) from complete meta-analysis gene list Up regulation (red) and Down regulation (blue). C. Heat plot of enriched disease from complete meta-analysis gene list. Presented diseases were reduced from 29–15 significantly enriched diseases to focus on neurological and sleep-related disorders. D. Dot plot representing significant colocalization analysis results. 9 traits were correlated to 21 genes from the 335 significant DEGs list. All Posterior Probability hypotheses were > 0.7.

Pathway enrichment: orexin

Overall functional enrichment analyses of the genes significant in meta-analysis using ToppFunn (MsigDB biocarta v7.5.1) show the orexin receptor pathway [52–54] as the only significantly enriched gene pathway (Bonferroni p-value = 8.39 × 10⁻⁴). The orexin receptor pathway includes genes involved in the orexin receptor signaling system that encompasses a wide range of functions such as regulating the sleep-wake cycle, energy homeostasis, neuroendocrine function, glucose metabolism, reward seeking, and drug addiction. Within the significant meta-analysis genes, two primary gene categories are present—increased expression in OOD cases of genes associated with sleep deprivation (DUSP4, EGR1, ARC, NR4A3, NR4A1, PDP1, EGR2) and the increased expression in OOD cases of one of two primary receptors for orexin A and orexin B (HCRTR2) (see conceptual model; Fig. 3B). Although these genes have a variety of functions, these categories were used in the descriptions within the framework of the orexin receptor signaling pathway. Also, it is important to note that the direction of effect seen in these genes are opposite of what is associated with wakefulness and coincides with sleep/sleeplessness phenotypes (Supplementary Fig. 5).

Disease enrichment

Using the meta-analysis gene list, we conducted a disease enrichment analysis to find associations between the observed DEGs and more than 11 diseases that could be linked to OOD. Using the DisGeNET database [32], we identified 29 diseases with FDR p-values < 0.05 (Fig. 3C) and 4 diseases with Bonferroni corrected p-values < 0.05: cocaine abuse (p-value = 2.29 × 10⁻²), stress-psychological (p-value = 4.88 × 10⁻²), cocaine-related disorders (p-value = 2.29 × 10⁻²), and hypertensive disease (p-value = 4.88 × 10⁻²), as seen in Table 2. When grouped based on symptom, morphology, and localization characteristics, five categories are observed: psychiatric disorders, substance abuse disorders, cardiovascular diseases, cancer diseases, and miscellaneous. Overall, the enriched diseases span psychiatric, substance use, and cardiovascular conditions, all of which can contribute to or be exacerbated by the hypoxia and cardiorespiratory collapse caused by asphyxiation during opioid overdose, linking our molecular findings to the primary mechanism of death.

Table 2.

Differentially expressed genes significantly enriched for a spectrum of diseases.

Psychiatric disorders	Substance abuse disorders	Cardiovascular diseases	Cancer diseases	Miscellaneous
Pain	Addictive behavior	Hypotensive disease	Nasopharyngeal carcinoma	Bejel
Neurodegenerative Disorders	Cocaine abuse	Cardiovascular disease	Malignant neoplasm of endometrium	Cartilage—hair hypoplasia
Mental disorders	Cocaine-related disorders	Cerebral infarction	Brain neoplasms	Nonvenereal endemic syphilis
Stress, psychological	Cocaine dependence	Myelodysplastic syndrome	Small cell carcinoma of lung
Neuropathy		Acute myocardial infarction	Secondary malignant neoplasm of liver
Anxiety and fear		Transient ischemic attack	Colorectal neoplasms
Excessive daytime sleepiness		Myocardial ischemia	Carcinoid heart disease
		Hypertensive nephropathy

Significantly enriched diseases that had FDR p-values < 0.05 can be seen (N = 29). Bonferroni corrected p-values < 0.05 can be seen in red (N = 4).

Genetics and differentially expressed genes

Partitioned heritability with published GWAS

We examined whether DEGs associated with OOD were enriched for genetic signals associated with opioid addiction and other related phenotypes using stratified linkage disequilibrium score regression (s-LDSC). We evaluated the 335 meta-analysis gene set against the 38 phenotype traits previously reported in Gaddis et al., [5] for opioid addiction GWAS. Given the enrichment of DEGs for the orexin pathway and its known role in sleep, we added two phenotypes from Jansen et al., [37] for insomnia GWAS, three phenotypes from Dashti et al., [36] for sleep duration GWAS, and one phenotype from Lane et al., 2019 for sleep duration. After filtering for FDR-adjusted p-values < 0.05, no phenotypes met the significance threshold (Supplementary Table 2). This indicates that genetic signal for the 37 phenotypes were not enriched within loci for the 335 OOD meta-analysis genes.

Colocalization analysis of differential expressed gene and GWAS signals

Despite a lack of heritability enrichment for GWAS traits across loci for the 335 meta-analysis genes, we sought to understand whether specific genes exhibited genetic regulation that aligned with GWAS signals. We conducted a colocalization analysis to determine whether any known eQTLs associated with the 335 significant DEGs were associated with phenotypes of interest. Our colocalization analyses used the coloc R package [45, 55] on 37 traits, some previously reported in Gaddis et al. [5] for opioid addiction GWAS and others from sleep cycle studies (Supplementary Table 4). We identified significant colocalization between 21 genes and 9 different phenotypes (posterior probability H4 > 0.7). FADS1, NAV1, FLT3LG, RGS4, CENPF, SUDS3, lnc-OXR1, DST, PKD1P3, RNF112, and VPS39 have been linked to neuropsychiatric disorders and sleep (Supplementary Table 9 & Supplementary Fig. 6) [56, 57]. Overall, these results suggest that a subset of genes linked to OOD are regulated by proximal genetic variation that is also associated with neurobiological or psychiatric traits.

Enrichment of meta-analyzed DEGs in MAGMA

Meta-analyzed DEGs did not show significant enrichment among MAGMA-significant genes in any trait after FDR correction across traits. While several traits (e.g., Schizophrenia, Bipolar Disorder, Opioid Addiction) showed nominal overlap, none reached significance (Supplementary Table 10).

Discussion

The goal for this study was to extend discovery and assess the robustness of differences in gene expression associated with OOD by conducting a meta-analysis of four recently published, independent transcriptome-wide case/control studies of the DLPFC. Our meta-analysis identified 335 DEGs, of which 269 were novel and 66 were previously reported by the independent studies [6–9]. Gene set enrichment analyses indicated several biological features that suggest that these DEGs dysregulate biology important to addiction. GO term enrichment results characterized OOD-associated DGE as functionally related to synaptic signaling and morphogenesis (Fig. 2A). Follow-up of DEGs in reactome annotations found enrichment in 5 signal transduction pathways (Fig. 2B). Additionally, the set of meta-DEGs were significantly enriched for the orexin pathway (Fig. 3B) and associated with several diseases across the spectrum of substance use, psychiatric disorders, cardiovascular diseases, and cancers (Fig. 3C). There was little evidence that the observed DGE resulted from differences in cell-type proportion between cases and controls, with a difference for microglia observed in one of four cohorts and differing directions of effect across cohorts. Similarly, there was limited evidence of genetic drivers associated with addiction being concentrated around DEGs. Four phenotypes shared heritability with DEGs (educational attainment, schizophrenia, bipolar disorder, and Alzheimer’s disease), but only when setting a less stringent significance cutoff (p-value < 0.1). Colocalization testing between eQTLs for DEGs and 37 GWAS traits did identify genetic loci linked to both OOD genes and several GWAS phenotypes, the most notable and significant can be seen in Fig. 3D. The colocalization analysis suggests that genetic variants linked to Alzheimer’s, Schizophrenia, Sleep Duration, Neuroticism, Smoking, Cigarettes per Day, and Years of Education may also influence gene expression related to OOD, highlighting shared genetic risk and potential biological mechanisms [58]. Notably, the orexin pathway was significantly enriched in the downstream DGE and pathway analysis, aligning with its known role in wakefulness, addiction, and cognitive function, all of which are relevant to opioid use and overdose vulnerability. FADS1 has been indirectly linked to the orexin pathway and RGS4 is both associated with Alzheimer’s, Schizophrenia, and bipolar disorder. Alternatively, RGS4 is known to regulate airway constriction [59], which is linked to asphyxiation, the physiological response that causes overdose death. The presence of Smoking, Cigarettes per Day, and Years of Education in the colocalization results, while less immediately intuitive, suggests possible pleiotropic genetic effects, wherein a single genetic variant influences multiple traits. Given that dopamine and reward-processing pathways contribute to both nicotine dependence and opioid addiction, the shared genetic signals could reflect a broader susceptibility to substance use disorders. Additionally, Years of Education is a widely recognized proxy for cognitive function, executive control, and socioeconomic status, all of which are critical determinants of substance use behaviors, stress response, and access to treatment [60]. The overlap between opioid overdose risk and neuropsychiatric traits such as Schizophrenia and Neuroticism further supports the hypothesis that genetic factors governing mental health, impulsivity, and cognitive processing may play a key role in opioid-related mortality. Meanwhile, the connection to Alzheimer’s Disease may point to shared neurodegenerative or inflammatory pathways that warrant further exploration. Neuro-inflammation has a known byproduct of opioid use and overdose symptoms. These finding show supporting links between eQTLs linked to other neurological based diseases and significant differentially expressed genes. This helps with understanding conceptually how all of the genes play a cruel role in complex diseases or associated pathways known to active in DLPFC region of the brain.

Figure 4 provides a conceptual model linking DEGs and enriched pathways from across our analyses to suggest a coherent hypothesis of gene dysregulation associated with OOD in DLPFC. We detail these connections below. When we investigated all DEGs using the reactome database, the five pathways that were enriched converged on BDNF-activated TRK receptor signaling and FGF2 activated FGFR signaling (both nested within the RTK pathway), both of which lead to transcriptional regulation via CREB TF activation, neurite outgrowth, and plasticity (Fig. 2B & Fig. 4) [61]. This is important because both BDNF and FGF2 have well-established associations with drug addiction [62–68]. However, what is not always clear is which of the three downstream signaling cascades BDNF & FGF2-RTKs activates: (1) PLC-gamma, (2) PI3K-AKT, or (3) MEK/ERK/MAPK signaling [61, 69]. Our data indicates, downstream activation of MEK/ERK signaling because SHC1, a DEG within the enriched pathways, binds to GRB:SOS when phosphorylated, which activates MAPK signaling [70]. GPCR signaling (enriched pathway) can activate PLC mediated upregulation of intracellular Ca 2 + , PLC mediated activation of MAPK via RAS GRP/GRP (Fig. 4). TRK-MEK/ERK signaling has been shown to result in neuronal growth and plasticity [71]. The observed up-regulation of FGF2 could mediate transcriptional activation causing synaptic plasticity and growth as well [72]. Together our functional enrichments of the meta-analysis DEGs show functions of TRK signaling and associated downstream events, which helps to elucidate the specific signaling cascade by which opioid overdose death, and posssibly opioid addiction, affects the neuronal plasticity within the DLPFC. Accumulated evidence suggests that neuroplasticity is a central aspect of the biology of addiction [73]. Our DGE analyses identify significantly reduced gene expression from BDNF down through the TRK-MEK/ERK signaling pathway. Reduced expression of these genes [74–76] and this signaling pathway [77, 78] are consistent with reduced neural plasticity consequent to chronic opioid exposure, which may reflect decreased dendritic spine density [79], other neuronal morphologic changes, or synaptic connectivity.

Fig. 4. Conceptual diagram of cell signaling pathways enriched.

Schematic including the DEGs within the orexin receptor pathway, GPCR signaling, and Tyrosine receptor Kinase (Trk) signaling enriched using ToppFunn, Gene Ontology (GO), Reactome enrichment analyses respectively. Receptors and gene products from the hit list for each pathway are included in this depiction. Solid arrows indicate direct regulation while dashed arrows indicate indirect regulation. Genes are colored to represent significant up regulation in green and down regulation in red. Additional pathway related genes that are included are colored blue. Within this diagram we see themes of synaptic plasticity in the pathways leading to actin/microtubule organization, neurite outgrowth, and GPC receptor recycling. We also see transcriptional plasticity, which is important response in addiction through the differential expression and regulation of activity regulated TFs (NPAS4, BDNF, ARC, DUSP2/4/6, ETV5, EGR1/2/5), chromatin remodelers (KAT6A, SUDS3, KCTD21). Together this diagram represents potential mechanisms by which these signaling pathways may be affected by opioid overdose. Created with bioRender.com.

Another key finding that emerged in this meta-analysis is significant enrichment of genes in the orexin receptor pathway (MsigDB biocarta, v7.5.1). Orexin signaling occurs via activation of G-protein coupled hypocretin receptors ORX1 and ORX2. The neuropeptides orexin A (ORXA) and orexin B (ORXB) can bind and activate ORX2, whereas ORX1 can be activated by ORXA only [53]. Orexin neurons are mostly located in hypothalamus, but their projections span many brain regions including the DLPFC [69, 80, 81]. Our data show decreased gene expression of the ORX2 receptor [82], which activates MAPK and ERK signaling and regulates the sleep-wake cycle [83]. Additionally, the activation of the ORX1 or ORX2 (GPCR) from ORXB (enriched DEG) mediates PLC activation which then can activate RAS GRP/GFP from the CAMKII enzyme. The CAMK2B gene that regulates CAMKII is a nominally significant DEG in our meta-analysis (wfisher p-value = 0.00138, adj p-value = 0.0619) and is known to interact with the ARC and RAS genes (enriched DEGs), and also ERK/MAPK pathways [69, 84]. This activation of the RAS GRP/GFP can link GPCRs to RAS which is also a target of RTKs (Fig. 4). It is important to note that orexins are dynamic with the circadian rhythm and disruption of this rhythm is associated with OUD [85]. Indeed, a recent study by Xue et al. [85], identified broad alterations in gene transcription rhythmicity (transcription abundance changing with circadian rhythm) that were associated with OUD in postmortem human DLFC and NAc tissue. Comparing the 335 DEGs from our meta-analysis of DGE in DLPFC to those identified by Xue et al. for circadian rhythm disruption in DLPFC (n = 339) [85], a small number of genes were associated with OUD here and alterations in rhythmicity: three genes, TRIM3, RFX3, AL117339.4, “gained” rhythmicity in OUD (OUD > control) and six genes, GPR19, NACC2, EGR4, PLAT, PPAT, ZFAND5 “lost” rhythmicity in OUD (OUD < control). Notably, the orexin genes were not significantly associated with transcript rhythmicity in Xue et al [58, 85]. However, in neurotypical brains, orexins are up-regulated during wake cycles and down-regulated during sleep cycles [83]. A disruption of ORX2 has been shown in animal models to cause narcolepsy-like symptoms [86]. Although orexin signaling has mainly been studied under the lens of the sleep-wake cycle, newer studies have shown associations with other neurological diseases including Parkinson’s disease [87] and addiction [88]. It has been shown that patients with narcolepsy have low to undetectable orexin-producing neurons in their brains and there are low rates of substance use among narcoleptics, which is attributed to reduced activation of the mesolimbic reward circuitry [83, 86]. On the other hand, patients with OUD tend to have much higher levels of orexin-producing neurons, and one of the primary withdrawal symptoms of opioid use is insomnia [83]. It is important to note that the decedents in this analysis are in a state of drug satiation (opioid overdose), which may explain the lower expression of genes in the orexin pathway in this study and findings of higher orexin levels in patients who are in withdrawal. Several clinical trials of Suvorexant and Lemborexant, orexin receptor agonists, will help to elucidate the relationships between orexin, opiate use, and sleep [89–92]. These studies and ours suggest that orexin levels are not only correlated with sleep regulation but also the brain’s reward circuitry and specifically opioid use.

Overall, this study has shown statistically robust evidence of RTK-mediated synaptic plasticity and significant down-regulation of the orexin signaling pathway in OOD brains. Interestingly, studies have also established associations between orexin signaling, BDNF, and FGF2. Several studies that test orexins as therapeutics for neurological diseases, including anxiety [93], depression [94], and Parkinson’s disease [87], have shown that orexin treatment is associated with the up-regulation of BDNF. Most notably, one study demonstrated that orexin A treatment increases BDNF protein in several neuronal subtypes, and orexin treatments in addition to blocking Pi3K/Akt signaling, which is known to upregulate BDNF transcription, did not result in BDNF up-regulation. Thus, they concluded that orexin A up-regulates BDNF through PI3K/Akt signaling. This paired with our data suggests that in opioid overdose brains, orexin mediates regulation of BDNF, which in turn activates TRK signaling and downstream transcriptional and synaptic plasticity. GPCR signaling were also enriched with HTR1B, EDN1, ORX2, and ADM genes which are also associated with the orexin pathway and link to Ras and MEK/ERK/MapK signaling, from the PLC mediated activation of CAMKII (Fig. 4) [69].

We employed an additional approach to characterize the DEGs by assessing their enrichment for established disease associations. This analysis revealed significant enrichment of DEGs associated with substance use disorders, specifically cocaine-related disorders, and a range of neurological disorders encompassing mental health conditions, pain syndromes, neurodegenerative diseases, and excessive daytime sleepiness. Of particular interest, a subset of enriched genes exhibited associations with multiple diseases, as illustrated in Fig. 3C Genes like HTR1B, ARC, FOSB, FGF2, EGR1, BDNF, and GCC1 demonstrated overlaps in their disease associations. Unexpectedly, we also observed an enrichment of DEGs associated with cardiovascular diseases and cancer among those related to OOD. These cardiovascular conditions spanned a spectrum from brain-related disorders such as hypertensive diseases, cardiovascular disease, cerebral infarction, hypertension nephropathy, myocardial ischemia, transient ischemic attack, and acute myocardial infarction, among others.

A common thread among the majority of these diseases was disruption in blood flow and their connection to addictive responses. Given that OOD typically results from suffocation, we noted intriguing gene overlaps with diseases associated with asphyxiation. These overlaps encompassed factors such as blood flow disturbances, cerebral nutrient deficiencies, oxygen deficits in the brain, damage to the cardiovascular lining, and stroke-related processes.

Limitations of the study

A recurring issue within this field of OUD research using postmortem human brains is that we cannot distinguish between the effects of acute overdose death and chronic OUD: our results likely include signatures of both phenotypes. Follow-up analyses in functional paradigms with model organisms likely need to differentiate these biological processes. Although this study is the largest of DGE and OOD to date, by virtue of meta-analysis, it remains likely that it is underpowered to detect all dysregulated gene expression associated with OOD in the DLPFC. The relatively modest number of genes that were observed from the prior four independent cohort studies in the meta-analysis suggests the need for larger sample sizes and more diverse samples to further enhance the robustness of our gene expression findings. The improved statistical power to detect DGE of the current meta-analysis over the individual cohort studies can be seen in the analyses presented in Supplementary Fig. 7 which shows a drop in detectable fold change from approximately 1.4–1.1 with 80% power under the meta-analysis. To reach 80% power to detect a 1.0 fold change will require approximately triple the current meta-analysis size. Another potential limitation is the use of bulk-tissue RNA-seq. Single-cell sequencing data would potentially offer more biologically informative insight within specific cell types. However, our meta-analysis of existing data has identified novel findings, despite not being sensitive to cell type-specific gene expression. Finally, the genetic heterogeneity between samples is low, consisting almost entirely of people of European descent. This does not accurately represent the population in the United States or the population of individuals with OOD. Therefore, our results miss neurobiological gene expression patterns that differ across a range of ancestral populations.

Summary

Our primary objective was to expand our understanding of the neurobiology associated with OOD and shed light on OUD by conducting a meta-analysis of four recent, independent transcriptome-wide investigations of the DLPFC. Through this meta-analysis, we identified 335 DEGs, 269 of which were novel discoveries and 66 were identified in the prior independent studies. This meta-analysis increased the pool of DEGs associated with OOD and underscored the growing robustness of these findings with larger sample sizes and contributions from independent cohorts. Significantly, the involvement of MAPK and ERK signaling in the DLPFC through a number of pathways, including the orexin pathway, GPCR signaling, and RTK signaling, emerged as a central discovery, providing new insights into the neurobiological mechanisms of opioid overdose death and opioid addiction. DEGs exhibited significant functional enrichment in biological processes related to addiction, such as synaptic signaling and morphogenesis. Furthermore, our functional enrichment analyses shed light on the signaling cascades associated with these genes, revealing likely down-regulation of RTK-mediated synaptic plasticity and a notable down-regulation of the orexin signaling pathway. These findings offer a foundation for future research into potential therapeutic interventions for OUDs by targeting the interplay between orexin (GPCR signaling), BDNF, FGF2, (RTK signaling) (Fig. 4).

Author contributions

Javan K. Carter: CN, DC, FA, I, SP, V, M, PA, VZ, WOD, WRE Bryan C. Quach: CN, FA, I, M, S, SP, V, VZ, WER Caryn Willis: FA, S, SP, WRE Melyssa S. Minto: M, S, WRE Dana B. Hancock: FAQ, SP, I, PA, Janitza Montalvo-Ortiz: FAQ, R, WRE Olivia Corradin: FAQ, R, WRE Ryan W. Logan: FAQ, R, WRE Consuelo Walss-Bass: FAQ, R, WRE Brion S. Maher: FAQ, R, WRE Eric Otto Johnson: CN, FAQ, R, SP, I, PA, WOD, WRE. Conceptualization (CN), Curation (DC), Formal Analysis (FA), Funding Acquisition (FAQ), Investigation (I), Methodology (M), Project Administration (PA), Resources (R), Software (S), Supervision (SP), Validation (V), Visualization(VZ), Writing – Original Draft Preparation (WOD) & Writing – Review & Editing (WRE).

Data availability

All data analyzed in this study were obtained from previously published studies and publicly available repositories. Detailed information regarding data sources, inclusion criteria, cohort characteristics, and data processing procedures is provided in the Contributing Study Cohort and Data Processing section of this manuscript. All processed data supporting the findings of this study are available in the Supplementary Files. No new primary datasets were generated. Additional information is available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

Our work was determined to be exempt by RTI’s Institutional Review Board given that the data analyzed were derived from postmortem human brain which are not considered human subjects.

Supplementary information

The online version contains supplementary material available at 10.1038/s41398-026-03977-9.