Introduction

Genetic variants that act as quantitative trait loci (QTLs) for gene regulatory features, such as DNA methylation (DNAm) and RNA expression (RNAexp) levels, are pervasive across the genome [1,2,3,4,5] and are enriched among disease-associated loci [6,29]. DNAm was measured using an Illumina Human MethylationEPIC BeadChip. As described before [11], DNAm data processing and QC were conducted using the R package minfi [30]. DNAm β-values were calculated and used in the meQTL analyses, representing the percentage of DNAm at each CpG (ratio of methylated intensities relative to the total intensity). See Supplementary Materials for more details.

In addition, the DNAm processing, sample and probe exclusions, and modeling information are described extensively in the Supplementary Materials of Markunas et al. [11].

Following RNA extraction, samples were sequenced using paired-end 100 bp reads on an Illumina HiSeq3000 at LIBD [11, 31]. RNA-seq samples were pseudomapped using Salmon v1.1.0 [32]. Transcript quantifications were aggregated to the gene-level counts using tximport v1.12.3 [33], resulting in 60,240 GENCODE v34 genes. Samples were excluded based on RNA-seq quality metrics, low RNA integrity (RIN) score, discrepancies between self-reported sex and chromosome Y gene expression, discrepancies between RNA and DNA derived genotype data, and missing genotype data. Sample-level QC removed 36 samples, resulting in a post-QC RNA-seq sample size of 203. Lowly expressed genes were then removed using the exclusion criteria of ≥90% of samples with ≤10 gene counts or ≤1 transcripts per million value. For eQTL map**, GRCh37 human genome reference coordinates were used for all gene annotations to align with the genotype data genome build. In total, 16,274 genes were considered for eQTL map**. See Supplementary Materials for more details.

Methylation and expression quantitative trait loci (meQTL/eQTL) map**

We performed cis-meQTL map** using imputed genetic variants with an MAF > 0.05 and DNAm intensity β-values of probes proximal, within 500 kilobases (kb) up or downstream, to these variants. Four different meQTL map** regression models were fit: (1) a “baseline” model to test for association between genetic variants and DNAm β-values across both smoking cases and controls; this model included age at death, sex, estimated non-neuronal cell-type proportion [34], principal component 1 (PC1) for DNAm array negative control probes, PC1 for imputed genotypes, and PC2 for imputed genotypes as covariates (see Supplementary Methods for model selection details); (2) a smoking cases–only model similar to the baseline model; (3) a smoking controls–only model similar to the baseline model; and (4) an interaction model to test for associations of a genetic variant-by-smoking status interaction with DNAm β-values. All models used rank-inverse normal transformed (RINT) DNAm β-values. The smoking cases–only and controls–only models are needed to generate summary statistics used to conduct stratified two degrees-of-freedom (2DF) tests. The stratified 2DF test jointly tests for genetic variant main effects and genetic variant-by-smoking status interaction effects [35]. Additional details are provided in the Supplementary Methods.

A similar framework was applied for cis-eQTLs. Baseline, smoking cases–only, smoking controls–only, and interaction eQTL models were fit for each imputed genetic variant and NAc expression levels of genes proximal (within 500 kb of gene body) to these variants. Counts for a given gene underwent median-of-ratios normalization [36] and RINT to obtain values used to represent the gene’s expression level. The covariates included in each model were age at death, sex, exon map** rate, ribosomal RNA map** rate, PC1 for imputed genotypes, PC2 for imputed genotypes, and four latent variables estimated by PEER v1.3 [37] to account for additional unmeasured sources of confounding (see Supplementary Methods).

To account for multiple hypothesis testing in the meQTL/eQTL models and stratified 2DF tests, p values were corrected using a conversative two-stage approach to mitigate inflation associated with single-stage approaches when applied to QTL map** [38]. This hierarchical approach first accounts for association tests across variants for a given DNAm probe/gene, then accounts for tests across all probes/genes. In the first stage, all nominal p values for a given probe/gene were adjusted using EigenMT [39]. In the second stage, EigenMT adjusted p values were further corrected by the number of probes/genes tested. Any initial two-stage adjusted p value smaller than a Bonferroni corrected p value was assigned the latter p value as the final two-stage adjusted p value. This ensures that the two-stage adjustment stringency does not exceed the family-wise error rate control of Bonferroni correction. A two-stage adjusted p value cutoff of 0.05 was used to identify statistically significant QTLs.

Gene Set Enrichment Analysis (GSEA) for eGenes and meQTL CpGs

We used the GSEAPreranked tool from GSEA v4.3.2 [40] to test for enrichment of smoking-associated eGenes and meQTL CpGs in the MSigDB v2023.1.Hs gene set collections C2 canonical pathways and C5 Gene Ontology gene sets. Only gene sets with 15–500 genes were included. CpGs were assigned to genes using the Infinium MethylationEPIC v1.0 B5 Manifest, resulting in 11,772 CpGs from the interaction meQTL map** model with at least one assigned gene. Each gene was ranked using the largest −log10-transformed, smoking interaction p value for a given gene from QTL map**. Weighted Kolmogorov–Smirnov–like statistics were computed for the enrichment scores, and p values were determined using 1000 gene set permutations. Significant pathways were selected based on false discovery rate (FDR) < 0.10, which was selected to maximize biological discovery and yet be more stringent than the default threshold of 0.25.

QTL enrichment testing for GSCAN-identified genetic variants

We conducted variant-based enrichment testing to assess whether genome-wide significant GSCAN loci were enriched for meQTLs or eQTLs. We obtained GSCAN summary statistics from the University of Minnesota’s Data Repository for U of M (https://doi.org/10.13020/3b1n-ff32), focusing on GSCAN’s GWAS results from 2019 to capture genetic loci with variants that commonly occur and have the largest effect sizes on smoking: N up to 1.2 million individuals, depending on the smoking trait analyzed [19]. For meQTL enrichment analysis, we compared the p value distributions from stratified 2DF tests (i.e., stratified 2DF meQTL map**) between GSCAN variants and a set of randomly matched variants. The random matched variant set was designed to be 10 times the size of the GSCAN variant set. The GSCAN variant set included linkage disequilibrium (LD)-pruned, genome-wide significant variants reported by GSCAN, across all four smoking traits, that were also available in our meQTL map** (361 variants, each representing an independent significant locus). The matched variant set included 3,600 LD-pruned variants selected using SNPsnap [41]. For variants with multiple stratified 2DF p values (i.e., proximal to multiple CpG sites), only the smallest p value was retained. The meQTL stratified 2DF p value distributions for the GSCAN and matched variant sets were tested for equality using a two-sided Kolmogorov–Smirnov test. The eQTL enrichment analysis followed the same procedure as the meQTL enrichment analysis. The GSCAN variant set included 305 variants because the overlap with variants from eQTL map** differed from meQTL map**. The SNPsnap-constructed, matched variant set included 3050 (305 × 10) variants.

Colocalization between GSCAN smoking GWAS and meQTL/eQTL map**s

We tested whether meQTL or eQTL signals from the baseline model QTL map** colocalized with GSCAN loci for smoking initiation, age at initiation, cigarettes per day, and cessation using the coloc v5.1.0 R package. We describe this analysis for meQTLs, but an equivalent framework was applied for eQTLs. All DNAm probes with an meQTL map** cis-window that overlapped with a GSCAN locus were considered for colocalization. For a given probe, the colocalization test region spanned all genetic variants that were (1) included in the meQTL map** for the probe and (2) tested in the GSCAN GWAS (Fig. S1). Summary statistics from GSCAN and the baseline model meQTL map** were used. Significant colocalization required a coloc posterior probability >0.8 for hypothesis 4 (both traits are associated and share a single causal variant) and a two-stage, adjusted p < 0.05 in the baseline meQTL analysis (Supplementary Materials).

For GSCAN loci that showed evidence of colocalization, HyPrColoc was applied to assess whether these colocalizations resulted from the same region of the locus [42]. Each HyPrColoc test included only one GSCAN trait, eGene, and CpG. For each HyPrColoc test, a genetic variant was included as input only if it had summary statistics available from the baseline model me/eQTL map** and GSCAN GWAS for the smoking trait. Significant colocalization was defined as a GSCAN trait–eGene–CpG triplet having posterior probability >0.8.

Results

Overview

Of the available 239 decedents, 201 RNA-seq samples and 220 DNAm samples, with genotype and smoking data, remained following QC, including 198 samples in both datasets (intersection) and 223 samples (52 cases and 171 controls) with either RNA-seq or DNAm data (union) (Table 1). Of the cases, 50% had African ancestry (AA) and 50% had European ancestry (EA) based on next-of-kin report and genotype confirmation. The manner of death differed slightly among cases and controls, but age, sex, and postmortem interval were similar (Table 1).

Table 1 Characteristics of decedents included for methylation and expression quantitative trait loci map** in nucleus accumbens.
Full size table

We generated single data-type QTL maps. Then each QTL type independently underwent colocalization analysis with GSCAN GWAS summary statistics, and joint colocalization analyses across data types were performed for significant me/eQTLs (Fig. 1).

Fig. 1: Analysis workflow.
figure 1

QTL analyses were conducted using DNA methylation, gene expression, and genotype data and three different QTL models to produce genome-wide cis-eQTL and cis-meQTL maps for nucleus accumbens. Two-stage multiple testing correction was applied to these QTL maps to identify significant QTLs that underwent further annotation. QTL maps were also integrated with GSCAN GWAS summary statistics to perform genetic variant enrichment testing and colocalization analyses. Smoking interaction QTL maps were used in conjunction with the Molecular Signatures Database to conduct gene set enrichment analysis and identify biological pathways enriched for CpGs/genes associated with smoking status and genotype.

Full size image

Genome-wide cis-meQTL maps

DNAm in NAc and genotype data were available for 52 smoking cases (26 EA, 26 AA) and 168 smoking controls (75 EA, 93 AA). In all, 11,206,899 variants were used in the initial analysis and 784,843 CpGs, resulting in 1,748,985,510 meQTL tests. After applying two-stage multiple testing correction, we identified 2,552,641 significant meQTL variants targeting 51,315 unique CpGs (Table S1) -which is restricted to the lead variant due to size; full results are available at synapse.org, https://doi.org/10.7303/syn50996324).

To identify the most robust signals, we performed post hoc filtering, kee** only meQTLs where the top variant had a minor allele frequency (MAF) ≥ 0.05 and missingness ≤0.10 in both ancestries, leaving 41,695 unique CpGs. The top five most significant meQTLs include CpGs that map to PITRM1, KDM3B, ARID1B, or MTL5 (Table S1).

Genome-wide cis-eQTL maps

Gene expression (RNA-seq) in NAc and genotype data were available for 47 smoking cases (24 EA, 23 AA) and 156 smoking controls (72 EA, 84 AA). These data were previously used to map eQTLs agnostic to the smoking phenotype [5]. In the present study accounting for variant-by-smoking interaction, we identified 83,095 significant eQTL variants from 1050 eGenes after multiple testing correction. Of these, 57,683 eQTL variants targeting 958 unique eGenes remained after post hoc filtering was applied to identify the most robust signals, kee** only eQTLs where the top variant had an MAF ≥ 0.05 and missingness ≤0.10 in both ancestries (Table S2). All significant results are shown in Table S2. Table S3 is filtered to the lead variant for each significant eGene (N = 1050). The 10 most significant eGenes were RPL9, ZSWIM7, GATD3B, RPS28, XRRA1, TMEM161B-AS1, CUTALP, NIPBL-DT, ZNF718, and SPATA7 (Table S3). Full results are available at synapse.org (https://doi.org/10.7303/syn50996324).

Both meQTLs and eQTLs were pervasive throughout the genome (Fig. 2). When we compared the eQTL to meQTL results, 562/1050 (54%) significant eGenes overlapped genes annotated to significant meQTL CpGs. Additionally, data from 655 (62%) robust eGenes (after MAF and missingness filtering) identified in our analysis were available in the NAc GTEx data. Of these 655, 509 (78%) met a Bonferroni correction significance threshold (0.05/655 = 7.6e-5) (Table S3, GTEx column: Nominal p-value). Thus, we observed a high correlation when comparing NAc eQTL results between our study and GTEx.

Fig. 2: Miami plot for stratified 2DF meQTL and eQTL maps.
figure 2

Stratified 2DF QTL map** −log10 nominal p values for significant (2DF adjusted p ≤ 0.05) genetic variant–CpG probe (top panel) and genetic variant-gene associations tests (bottom panel) are displayed as a function of genome position (x-axis). Genomic locations of significant GSCAN GWAS loci for four smoking phenotypes are denoted by the bars in the middle panel.

Full size image

Smoking interaction effects with meQTLs/eQTLs

To identify meQTLs/eQTLs that differed by smoking, we compared the 2DF QTL test results with those of the baseline models and interaction models. The meQTLs were primarily driven by main effects, with few showing evidence for interaction. Of the 41,695 significant unique CpGs, only five demonstrated strong evidence of a smoking interaction (Table 2, Fig. S2) based on Bonferroni correction (significance threshold of 0.05/41,695 = 1.2e-6). No eQTLs showed evidence of an interaction (significance threshold of 0.05/877 unique eGenes=5.7e-5) after filtering by MAF and missingness.

Table 2 Five meQTLs with evidence of significant variant × smoking interactions.
Full size table

Because use of stringent significance thresholds may miss subtle smoking interactions with individual variants, we performed GSEA using genes ranked based on QTL–smoking interaction p values to identify QTL-enriched biological processes and pathways that may be altered by smoking. For meQTL-smoking interactions, five pathways related to the synaptic cleft (neurotransmission site between pre- and postsynaptic membranes) were implicated. For eQTL-smoking interactions, cell cycle and wound response processes (cellular changes resulting from a stimulus indicated damage to an organism) were enriched (Table S4).

GWAS-identified variants that exert QTL effects on their target genes

To test the overlap** evidence of GWAS-identified variants as QTLs, we performed an enrichment analysis using a two-sample Kolmogorov–Smirnov test that compared NAc meQTL and eQTL p value distributions at GSCAN significant variants with a random set of variants. We found that GSCAN variants were significantly enriched for meQTLs (p = 0.005) but not for eQTLs (p = 0.3). This aligns with our observation that more meQTLs than eQTLs overlap GSCAN variants (Tables S5–S7). For comparison, using the same test with NAc eQTLs from GTEx, we found an enrichment p = 0.07.

Smoking meQTL, eQTL, and GWAS colocalization

We performed colocalization analyses of GSCAN’s GWAS results with our QTL maps to characterize heritable components of smoking that exert QTL effects. First, we performed pair-wise analyses (meQTL + GWAS, eQTL + GWAS), starting with GSCAN significant loci. Because a single locus may include more than one unique CpG or eGene, many colocalization analyses were performed per region. In general, we observed more colocalization of meQTLs than eQTLs with the GWAS loci (Tables S8–S10). Four genome-wide significant CpGs colocalized across two phenotypes (cigarettes per day and age of initiation): cg12293539 at MAML3 and cg00622170, cg11254171, and cg18236429 at the NOP14/NOP14-AS1 locus.

Next, we performed colocalization using a method (HyPrColoc) that can incorporate meQTLs, eQTLs, and smoking GWAS in a single analysis. Focusing on GSCAN GWAS loci that colocalized with both a significant eQTL and meQTL, we confirmed the GWAS–eQTL–meQTL colocalizing region at the GSCAN smoking initiation locus chr3: 52386605–54266212. Three variant–CpG–eGene combinations had significant colocalization, all involving ENSG00000243696 (predicted read-through of MUSTN1-ITIH4) as the eGene with (1) rs6445538 and cg25643088, (2) rs6445538 and cg19713033, and (3) rs4687672 and cg23815702 (Fig. 3A). The C allele of rs6445538 and the A allele of rs4687672 are associated with increased smoking initiation with very modest effect sizes: 0.023 for both [19, 43]. These alleles are associated with increased ENSG00000243696 gene expression (Table S2: Column: βbaseline), increased DNA methylation at cg25643088 and cg23815702, and decreased DNA methylation at cg19713033 (Fig. 3B, Table S1: Column: βbaseline).

Fig. 3: GSCAN smoking initiation GWAS, meQTL, and eQTL association test −log10 p values near GSCAN locus chr3: 52386605–54266212.
figure 3

A −log10 nominal p values for associations between genetic variants and smoking initiation, expression of three CpGs, and gene expression for the eGene identified by HyPrColoc with colocalization probability >0.8. Only genetic variants that overlapped across summary statistics for GSCAN, meQTL map**, and eQTL map** are plotted. Locations of genes included in the eQTL map** are annotated in the bottom of the panel. B Genotype-by-CpG plots for the CpGs in Fig. 3A showing CpG probe intensity levels partitioned by genotype.

Full size image

Discussion

Cigarette smoking remains highly prevalent and a leading cause of death globally, despite decades of research into the health consequences and public health campaigns to curb smoking [44, 45]. Addiction to cigarette smoking is a complex, multi-stage process involving a neuronal rewards system that includes the NAc region of the brain [21, 46]. The NAc is known to have a role in cognitive processing of motivation, reward, and reinforcement, which are essential to the first stage of addiction (binge/intoxication) [22]. This study employed methods to better understand the functional effects of heritable factors that influence smoking behaviors.

QTLs are pervasive throughout the genome and provide valuable insight into tissue-specific gene regulation. The GTEx project, widely used for exploring eQTLs, recently released a large-scale meQTL dataset encompassing nine human tissues [47]. However, this dataset did not capture brain tissues. To our knowledge, ours is the first genome-wide meQTL map in human NAc and is shared publicly as a new resource for the scientific community.

We found few significant variant-by-smoking interaction effects and conclude that most QTLs in the NAc may not differ by smoking. While some interactions with small effect sizes may exist, they may require larger sample sizes to detect. We investigated this possibility using a pathway analysis ranked by evidence of smoking interactions, and we identified pathways related to cellular damage, cell cycle, and the synaptic cleft, where the nicotinic acetylcholine receptors play an important role in regulating neurotransmission [48]. Five individual meQTLs differed significantly between smokers and nonsmokers, including Nudix Hydrolase 12 (NUDT12), Family with sequence similarity 53 member B (FAM53B), and Ring Finger Protein 39 (RNF39). NUDT12 plays a role in nicotinate and nicotinamide metabolism [49]. Interestingly, NUDT12 was identified in a transcriptome analysis of neurons following chronic nicotine exposure [50] and lies within a QTL interval for nicotine sensitivity in mouse studies [51, 52]. FAM53B has been associated with cocaine dependence [53]. Differential DNAm at FAM53B has been observed in COPD cases compared with controls [54]. RNF39 was found to be differentially methylated in a study of marijuana use [55] and other smoking-related DNAm changes [56, 57]. Together, these represent biologically plausible genes whereby smoking may alter genetically driven gene regulation in NAc.

We employed colocalization analyses [5] to identify trait-associated variants that act as regulators of DNAm or gene expression. We generally observed more overlap between the GWAS-associated variants and meQTLs than eQTLs.The weaker overlap of GWAS variants and eQTLs has been noted elsewhere [58], and this pattern was also supported by our enrichment analyses, and may relate to the magnitude of meQTL tests. However, the same pattern held when we looked only at QTLs that survived genome-wide multiple testing correction, whereby declaring meQTLs as statistically significant was based on a more stringent threshold than eQTLs and is consistent with studies of the prefrontal cortex [59], blood [60], and other tissues [47].

One genomic region showed robust evidence of colocalization with all three data types, highlighting novel functional evidence where changes to DNAm and gene expression may help explain the neurobiology underlying a smoking initiation associated heritable factor. The primary gene indicated is a predicted read-through of the Musculoskeletal, Embryonic Nuclear Protein 1 (MUSTN1) and inter-alpha-trypsin inhibitor, heavy chain 4 (ITIH4) genes. Little is known about the read-through transcript. MUSTN1 is known to have a role in skeletal muscle homeostasis, chondrocyte differentiation, and limb morphogenesis [61]. Schizophrenia-associated variants in ITIH4 have been shown to regulate expression of ITIH4 in the prefrontal cortex [62], and variants in ITIH4 is a biomarker for COPD [63, 64]. MUSTN1 and ITIH4 are expressed at low levels in the brain and moderately expressed in skeletal muscle [1, 65, 66]. There is evidence that changes to skeletal muscle homeostasis can influence the physiology of the brain [67, 68]. However, we did not detect a significant eQTL for MUSTN1 or ITIH4 in our eQTL analysis, indicating that the MUSTN1-ITIH4 read-through may warrant further functional interrogation to decipher how its role may differ from MUSTN1 and ITIH4.

This study has limitations to consider in interpreting the findings. First, because this study utilized an understudied brain tissue collected from decedents with a unique set of multi-omics data types and smoking status, sample availability was limited to 52 smoking cases and 171 smoking controls. This constrained sample size may have limited our ability to identify interaction effects, and it limited interrogation of ancestry- or sex-specific effects and extension into independent replication datasets. Also, QTL map** results in millions of tests, increasing the chance of type I error. We accounted for multiple testing using a two-stage correction strategy designed for this type of study [39]; however, the possibility of type I error remains. Furthermore, despite the concordance between eGenes identified from eQTL map** and those reported by GTEx, the two-stage correction strategy we applied has reduced sensitivity to detect eQTL relative to the single-stage FDR correction approach applied by GTEx, resulting in a lower number of reported eGenes in our study despite comparable sample size.

Finally, the QTL–GWAS colocalization was based on the first GSCAN meta-analysis, including up to 1.2 million individuals [19]. GSCAN recently released an updated meta-analysis with >3 million individuals [20]. Although comparison with the updated GSCAN meta-analysis might result in the identification of additional QTL–GWAS colocalization signals, given the sample size available with multi-omics data in our postmortem human brain study, we would have less statistical power to detect colocalization with the lower MAF variants in GSCAN2, and these variants are not well captured by our imputed genotype data. The present study focused on genetic loci with common variants with the largest effect sizes on smoking.

This study had several strengths. It is the first to provide a genome-wide meQTL map in human NAc, a relatively understudied brain tissue with an important role in the addiction cycle. Case definitions were carefully established based on corroborating evidence from several sources, including blood- and brain-based toxicology screens with confirmation by next-of-kin reports. Therefore, misclassification is unlikely. We used the 2DF test, which achieves power similar to that of a standard 1DF test when no interaction is present and simultaneously improves power when interaction effects are present [35]. We used conservative thresholds to identify the most robust signals, and the full results are provided for researchers to explore additional signals. To our knowledge, the present study represents the first large-scale formal testing of colocalization for GSCAN-identified loci with QTLs in human brain, identifying novel target genes and providing insight into the neurobiological function of smoking-associated heritable factors in relation to both DNAm and gene expression.

Overall, this multi-ancestry, multi-omics study of decedents with smoking status known and accounted for provides a unique resource for interrogating regions across the genome for their influence on gene regulation, cigarette smoking behaviors, and other complex conditions involving the NAc. Future studies may use these data to compare QTLs across other brain tissues to gain insights into tissue-specific regulation and to further investigate the neurobiology underlying other disease processes.