A multi-ancestry genetic study of pain intensity in 598,339 veterans

Abstract

Chronic pain is a common problem, with more than one-fifth of adult Americans reporting pain daily or on most days. It adversely affects the quality of life and imposes substantial personal and economic costs. Efforts to treat chronic pain using opioids had a central role in precipitating the opioid crisis. Despite an estimated heritability of 25–50%, the genetic architecture of chronic pain is not well-characterized, in part because studies have largely been limited to samples of European ancestry. To help address this knowledge gap, we conducted a cross-ancestry meta-analysis of pain intensity in 598,339 participants in the Million Veteran Program, which identified 125 independent genetic loci, 66 of which are new. Pain intensity was genetically correlated with other pain phenotypes, level of substance use and substance use disorders, other psychiatric traits, education level and cognitive traits. Integration of the genome-wide association studies findings with functional genomics data shows enrichment for putatively causal genes (n = 142) and proteins (n = 14) expressed in brain tissues, specifically in GABAergic neurons. Drug repurposing analysis identified anticonvulsants, β-blockers and calcium-channel blockers, among other drug groups, as having potential analgesic effects. Our results provide insights into key molecular contributors to the experience of pain and highlight attractive drug targets.

Pain is an unpleasant sensory and emotional experience associated with, or resembling that associated with, actual or potential tissue damage¹. Acute pain typically lasts less than 4 weeks, whereas chronic pain lasts more than 3 months². More than 50 million US adults report experiencing pain on most days or every day³, making pain the most common reason for seeking medical treatment⁴. In the late 1980s, many medical and pain organizations adopted policies to ensure the adequate assessment of pain, which was designated ‘the fifth vital sign’². A dramatic increase in prescriptions for opioid analgesics resulted, contributing to the opioid epidemic and a doubling of opioid-related deaths in the 1990s^5,6.

Opioids are not efficacious in managing chronic noncancer pain⁷, and their long-term use is associated with addiction, sleep disturbances, opioid-induced hyperalgesia, endocrine changes and cardiac and cognitive effects^8,9. Nonopioid medications, such as nonsteroidal anti-inflammatory and antiepileptic drugs, are effective for only some types of pain and can have substantial adverse effects¹⁰. Thus, new therapeutic targets for chronic pain are needed to facilitate the discovery or repurposing of safe, effective analgesics.

Notably, drug development efforts informed by genetics can double the rate of success¹¹. Although the heritability (h²) of individual differences in the susceptibility to develop chronic pain is estimated to be 25–50% (ref. 12), the mechanisms that underlie it are poorly understood¹³. Genome-wide association studies (GWASs) of chronic pain in large samples have focused on specific bodily sites^13–17 or aspects of an individual’s sensitivity to experiencing and reporting pain^18–21. GWASs have identified genome-wide significant (GWS) loci for headache²², osteoarthritis²³, low back pain^16,17, knee pain¹⁴, neuropathic pain²⁴ and multisite chronic pain^18,19, with high genetic correlations among them²⁵. There are also significant genetic correlations between pain phenotypes and psychiatric, substance use, cognitive, anthropometric and circadian traits^{14,16–18,26}. This suggests that a common genetic susceptibility underlies a broad range of diverse chronic pain conditions²⁶ and common co-occurring conditions.

Most GWAS of pain traits have been conducted in predominantly European ancestry cohorts. However, biobanks linked to electronic health records (EHRs), with large, well-characterized, multi-ancestry samples, are now available for use in identifying genetic risk factors and therapeutic targets for chronic pain²⁷. The Million Veteran Program (MVP)²⁸, an observational cohort study and mega-biobank implemented in the US Department of Veterans Affairs (VA) healthcare system, includes data on routine pain screening. Pain ratings in the MVP use an 11-point ordinal numeric rating scale (NRS), which has been standard practice in VA primary care for more than a decade²⁹ and has been shown to be a consistent, valid measure of reported pain^30,31.

We conducted a cross-ancestry meta-analysis of the NRS in samples of African American (AA), European American (EA) and Hispanic American (HA) ancestries from the MVP (n = 598,339). Although the NRS is a report of pain intensity experienced at a specific point in time, we used the median of medians as a proxy for chronic pain. We also conducted secondary analyses in a subsample of 566,959 individuals that excluded participants with a lifetime opioid use disorder (OUD) diagnosis to assess potential confounding by OUD, in a subsample of 291,759 participants that excluded those with pain ratings = 0 and in males and females separately (sex-stratified; Extended Data Fig. 1).

Results

Description of the sample

The study sample comprised 598,339 individuals (AA = 112,968, EA = 436,683 and HA = 48,688), of whom 91.2% were male (Supplementary Table 1). The secondary analysis that excluded individuals with a lifetime OUD diagnosis was reduced by 5% across population groups (AA = 104,050, EA = 415,740 and HA = 46,169; Supplementary Table 1). The median ages were 61.4 (s.d. = 14.0) and 61.7 (s.d. = 14.1) in the full and non-OUD samples, respectively. Approximately half of individuals in both the full sample (51.2%) and the non-OUD sample (52.7%) reported a median NRS of 0, that is, no pain. Mild (NRS 1–3), moderate (NRS 4–6) and severe pain (NRS 7–10) were reported by 24.4%, 19.2% and 4.5%, respectively, in the full sample, and by 24.6%, 18.2% and 4.0%, respectively, in the non-OUD sample.

Identification of pain intensity risk loci

Our cross-ancestry meta-analysis of 15,895,579 imputed autosomal single-nucleotide polymorphisms (SNPs) in the AA, EA and HA samples identified 4,364 GWS variants represented by 162 linkage disequilibrium (LD)-clumped index variants (r² > 0.1; Fig. 1). No lead SNP showed evidence of heterogeneity across ancestries, based on the I² index (Supplementary Fig. 1). In a cross-ancestry meta-analysis of chromosome X (chrX), we identified one GWS variant (DASH2-rs195035; Supplementary Table 2). Analyses conditioned on the lead SNP left 125 (124 autosomal and 1 chrX) independent association signals (Supplementary Table 2), including 59 previously reported as pain-related loci^16,18 and 66 new loci (Fig. 1 and Supplementary Table 2). Seven independent variants are exonic, 81 reside within a gene transcript and 37 are intergenic. Of the 7 exonic variants, 2 have likely damaging effects (Poly-Phen > 0.5, Combined Annotation Dependent Depletion (CADD) > 15; SLC39A8-rs13107325 and WSCD2-rs3764002) and 4 are potentially deleterious (CADD > 15; ANAPC4-rs34811474, MIER-rs2034244, NUCB2-rs757081 and APOE-rs429358; Supplementary Table 2).

Fig. 1 |. Manhattan plot for the pain intensity cross-ancestry GWAS meta-analysis (n = 598,339).

This identified 125 independent index variants. The nearest gene to the 66 new loci (65 autosomal and 1 X-chromosomal) is annotated. SNPs above the red line are GWS after correction for multiple testing (P < 5 × 10⁻⁸).

We looked up the 125 independent variants in a recent meta-analysis of 17 United Kingdom Biobank (UKBB) pain-related traits²⁰, a common genetic pain factor consisting of 24 chronic pain conditions in the UKBB²⁶, and the human pain genes database (HPGD)³². Two variants (TCF4-rs618869 and SLC39A8-rs13107325) were GWS in the meta-analyses of pain traits (Supplementary Table 2). None of the 66 new pain associations reported here was GWS in the published studies^20,26 or the HPGD³², confirming their newness.

Because inflammation appears to have a role in pain susceptibility²⁰, we explored the 125 independent pain intensity loci (within a 1-Mb buffer) for pleiotropy with immune traits in the GWAS catalog³³. Of these, 23 (11 new) were pleiotropic with immune/hematopoietic traits, such as C-reactive protein levels and blood cell counts (Supplementary Table 3).

The GWAS using 7,069,962 imputed autosomal SNPs in EAs yielded 103 LD clumps (r² > 0.1) across 86 independent loci (Extended Data Fig. 2 and Supplementary Table 4). One chrX variant (DASH2-rs195035) was GWS in EAs (Supplementary Table 4). Of the 87 independent loci, 16 were not GWS in the cross-ancestry meta-analysis (Supplementary Table 4). We also identified 2 GWS variants in one locus (nearest gene PPARD; chr 6) in AAs (11,183,154 imputed SNPs) and 16 GWS variants in two loci (nearest genes RNU6-461P, chr 3 and RNU6-741P, chr 15) in HAs (5,859,313 imputed SNPs; Supplementary Table 5). No chrX locus was associated with pain intensity in AAs or HAs. We used a sign test to examine the 86 independent EA index autosomal variants in AAs and HAs, of which 57 and 74, respectively, were directly analyzed or had proxy SNPs in these populations (Supplementary Table 6). Most variants had the same direction of effect in both populations (n_SNPs AAs = 41, HAs = 61; sign test AAs P = 0.0013, HAs P = 1.39 × 10⁻⁸). Only 15 variants (n_SNPs AAs = 2, HAs = 13) were nominally associated (P < 0.05) and none survived multiple test correction (Supplementary Table 6). The cross-ancestry genetic-effect correlation (ρ_pe) was 0.71 (s.e. = 0.13, P = 2.12 × 10⁻²) between EAs and AAs and 0.74 (s.e. = 0.08, P = 6.81 × 10⁻⁴) between EAs and HAs. The cross-ancestry heritability estimates between AAs and HAs were too low to calculate ρ_pe between them.

Loci for non-OUD diagnosis, nonzero pain ratings and sex differences

The secondary cross-ancestry meta-analysis that excluded participants with a lifetime OUD diagnosis identified 3,400 SNPs in 101 LD-independent risk loci (Supplementary Table 7). Of these, 86 were GWS in the primary GWAS, whereas 13 were P < 10⁻⁶, including 11 that were new (Supplementary Table 7). We also identified 18 ancestry-specific loci (17 in EAs and 1 in AAs; Supplementary Tables 8 and 9).

The cross-ancestry meta-analysis of individuals with NRS > 0 identified 461 SNPs in 12 independent risk loci (Supplementary Table 10). Of these, 7 were associated with pain intensity in the primary GWAS, whereas 5 loci were not, including 3 (TNIK-rs189788533, GRK4-rs12500735, RP11-99C10.1-rs7124028 and FAM81A-rs149493877) new pain loci (Supplementary Table 10). In EAs and AAs, we identified 11 and 1 independent risk loci, respectively, including 3 (EA, TNIK-rs189788533 and RP11-404L6.2-rs6884145; AA, FAM81A-rs149493877) additional new associations (Supplementary Table 11).

In a cross-ancestry meta-analysis that comprised males only, we identified 97 independent risk loci (1 in chrX), including 10 that were not GWS in the primary GWAS and 7 new associations for pain traits (NASP-rs2991977, PDE11A-rs16865764, RP11-138I17.1-rs1726312995, CD14-rs2569190, RP11-572H4.1-rs1122665, SLITRK1-rs1331928 and SDK2-rs150636180; Supplementary Table 12). We also identified 75 independent risk loci (1 in chrX) in EAs—including 7 new for pain traits—(Supplementary Table 13) and 1 new locus each for AAs and HAs (Supplementary Table 14). No variant was GWS among females although 28 LD-clumped variants (r² > 0.1) showed a suggestive level of association (P < 5 × 10⁻⁶; Supplementary Table 15). The direction of allele effects was highly correlated between males and females (r = 0.68, P = 2.2 × 10⁻¹⁶; Supplementary Fig. 2), as was genetic correlation for pain intensity (EAs, r_g = 0.87, P = 1.41 × 10⁻³⁵ and AAs, r_g = 1, P = 6.8 × 10⁻³). SNP heritability was moderate for both females (EAs = 0.12 and AAs = 0.05) and males (EAs = 0.08 and AAs = 0.06).

Overall, the per-allele effect sizes of lead risk variants between the primary and secondary GWASs were high, ranging from 0.88 to unity (P < 2.2 × 10⁻¹⁶; Extended Data Fig. 3). The genomic inflation factor (λ_GC) of the fixed-effect meta-analysis across all GWASs ranges from 1.03 to 1.23 (Supplementary Fig. 3), as expected for a polygenic trait^34,35. Across all GWASs, the univariate linkage disequilibrium score regression (LDSC) intercept ranges from 0.99 to 1.2 (s.e. = 0.01), which, being close to 1.0, suggests that most of the genome-wide elevation of the association statistics comes from true additive polygenic effects rather than a confounder such as population stratification. LDSC genetic correlations (r_g) between the primary and secondary GWASs were high, ranging from 0.89 to unity, with overlapping confidence intervals (Extended Data Fig. 4). In the primary GWAS, the LDSC ratio between the intercept and mean χ² statistic (1.90) was 0.13, suggesting that 87% of the observed inflation in χ² statistic is due to the polygenicity of the pain trait.

SNP heritability and enrichment

The proportion of variation in pain intensity explained by common genetic variants (h²_SNP) was similar both for the primary (AAs: 0.06 ± 0.009, EAs: 0.08 ± 0.003 and HAs: 0.07 ± 0.011) and the non-OUD GWAS (AAs: 0.07 ± 0.009, EAs: 0.08 ± 0.003 and HAs: 0.07 ± 0.011; Supplementary Table 16).

Partitioning the SNP heritability for pain intensity revealed significant tissue-group enrichment in the central nervous system (CNS; P = 1.47 × 10⁻¹²), adrenal gland (P = 8.97 × 10⁻⁵), liver (P = 3.15 × 10⁻⁴), skeletal (P = 8.50 × 10⁻⁴), cardiovascular (P = 0.001) and immune/hematopoietic (P = 0.004) tissues (Fig. 2a,b and Supplementary Table 17). In gene expression datasets derived from multiple tissues, we observed predominant h²_SNP effects in brain (P = 2.87 × 10⁻⁵), including hippocampus (P = 1.00 × 10⁻⁴) and limbic system (P = 1.15 × 10⁻⁴; Fig. 2c,d and Supplementary Table 18). SNP-based heritability in histone modification data also showed robust enhancer (H3K27ac and H3K4me1) and active promoter (H3K4me3 and H3K9ac) enrichments in brain tissues, including the dorsolateral prefrontal cortex (dlPFC; P < 1.32 × 10⁻⁴), inferior temporal lobe (P < 3.09 × 10⁻⁴), angular gyrus (P = 8.42 × 10⁻⁵) and anterior caudate (P = 1.12 × 10⁻⁴; Fig. 2e, and Supplementary Table 19). Similar results were obtained for the partitioned heritability analysis of the non-OUD GWAS (Supplementary Tables 17 and 18), although it also included significant expression effects in the cortex and cerebellum.

Fig. 2 |. Enrichment of pain intensity in the brain.

a, Partitioning heritability enrichment analyses using LDSC showing enrichment for pain intensity in the CNS, adrenal, liver, cardiovascular, skeletal and immune/hematopoietic tissues. The dashed black lines indicate Bonferroni-corrected significance for multiple testing (P < 0.005). b, Proportion of heritability shows robust enrichment for SNPs in brain and immune-related tissues. c–e, Heritability enrichment analyses for gene expression (c and d) and chromatin interaction (top 35 annotations are shown in e; see Supplementary Table 17 for full details) using Genotype-Tissue Expression (GTEx) data show enrichment for pain intensity in brain regions previously associated with chronic pain. Bonferroni correction was applied within each tissue conditioned on the number of genes tested.

Although the SNP-based heritability and enrichment for the primary and non-OUD GWASs were similar, because the primary GWAS using the full sample yielded more risk loci, we based all downstream analyses (except r_g analyses) on the GWAS results from that sample.

Gene-set enrichment in tissue and cell types

We mapped GWAS variants to genes via expression quantitative trait locus (eQTL) association and assessed the tissue enrichment of mapped genes. After correcting for multiple testing (P = 9.25 × 10⁻⁴) in the cross-ancestry and EA-specific GWASs, we uncovered significant transcriptomic enrichment only in brain tissues (Extended Data Fig. 5). Consistent with previous findings of brain tissue enrichment across different pain phenotypes in EAs^18,21,23, both our EA and cross-ancestry analyses showed notable enrichment in the cerebellum (cross-ancestry, P = 1.84 × 10⁻⁷; EA, P = 2.90 × 10⁻⁶), cerebellar hemisphere (cross-ancestry, P = 3.32 × 10⁻⁷; EA, P = 6.23 × 10⁻⁶), cortex (cross-ancestry, P = 1.97 × 10⁻⁶; EA, P = 3 × 10⁻⁴) and frontal cortex (cross-ancestry, P = 1.83 × 10⁻⁶; EA, P = 4.17 × 10⁻⁴; Extended Data Fig. 5). Among AAs there were no significantly enriched tissues (Supplementary Table 20).

To investigate enrichment at the level of cell types in the EA GWAS results, we conducted cell-type-specific (CTS) analysis in a collection of 13 human brain scRNA-seq datasets. After adjusting for possible confounding due to correlated expression within datasets using a stepwise conditional analysis, we detected jointly significant cell-type enrichments (proportional significance (PS) > 0.5) for gamma-aminobutyric acid (GABAergic) neurons largely in the human adult midbrain (P = 0.003, β = 0.206, s.e. = 0.075, PS = 0.56) and to a lesser extent in the prefrontal cortex (P = 0.044, β = 0.045, s.e. = 0.016, PS = 0.39; Supplementary Table 21).

Prioritization of candidate genes

To facilitate the biological interpretation and identification of druggable targets, we used a combination of gene set and fine mapping, transcriptomic, proteomic and chromatin interaction models to prioritize high-confidence variants and genes that most likely drive GWAS associations. Assigning SNPs to genes using physical proximity, gene-based analyses identified 6 GWS genes in AAs, 203 in EAs and 154 in the cross-ancestry results (Extended Data Fig. 6 and Supplementary Table 22), but none in HAs. Gene-set analysis using cross-ancestry GWAS results identified significantly enriched biological processes in dopaminergic synaptic transmission (GO:0001963; Bonferroni, P = 0.026) and cellular components of the aryl hydrocarbon receptor complex (GO:0034751; Bonferroni, P = 0.024). Negative regulation of synaptic transmission (GO:0050805; Bonferroni, P = 0.016) was related to pain intensity in EAs (Supplementary Table 23).

For consistency with available reference data, we based the fine-mapping procedure on EA GWAS results using 78 genomic regions (spanning 103 index variants; Supplementary Table 24) defined by the maximum physical distance between the LD block of independent lead SNPs (Methods). Functional genomic prediction models used the full EA GWAS results (Extended Data Fig. 1).

We fine-mapped the 78 regions using the Bayesian method (Methods). For each region with independent causal signals (Supplementary Table 24), credible sets of variants (posterior probability (PP) > 0.5) were constructed to capture 95% of the regional PP (k ≤ 5; Supplementary Table 25). Of these regions, 4 harbored 1 SNP (potentially indicating the causal variant), 20 harbored 2 SNPs and 44 harbored 3 or more SNPs (Supplementary Table 25). In total, fine-mapping prioritized 76 unique credible variants (n = 108; Fig. 3a), including 26 independent lead SNPs and 18 new pain loci (Fig. 3b). Most (50 of 76) of the credible variants map to protein-coding genes and are mostly eQTLs (Supplementary Table 25) and 5 harbor missense variants, of which 3 (ANAPC4, APOE and SLC39A8) are known pain loci^18,23 and 2 (RYR2 and AKAP10) are new (Fig. 3b). This small proportion of missense variants and high eQTL enrichment are consistent with an increased probability that the credible variants influence liability to pain intensity through gene expression modulation.

Fig. 3 |. Gene prioritization for pain intensity.

a, Genomic annotation of credible sets using FINEMAP shows enrichment largely in noncoding regions and to a lesser extent in exons. PIP, posterior inclusion probability; UTR, untranslated region. b, Annotation of known and new credible genes. Dashed line indicates PP > 0.5. c, Number of overlapping genes across functional prediction models. d, Tissue enrichment of prioritized genes using SMR and GTEx data shows enrichment in brain regions. The size of the circle reflects −log₁₀(P). Bonferroni correction was applied within each tissue conditioned on the number of genes tested.

We performed transcriptome-wide association study (TWAS) and proteome-wide association study (PWAS) analyses to determine whether risk variants exert their effects via gene and/or protein expression. After correction for multiple testing, 196 unique genes (TWAS eQTL—294, TWAS sQTL—67 and PWAS—32) were associated with pain intensity (Supplementary Tables 26 and 27). Of these, 69 represent new associations (based on a window from the index GWAS locus >1 MB). PWAS showed significant associations in the dlPFC that overlapped for 22 unique genes across multiple brain tissues in the TWAS (eQTL—16 and sQTL—8; Fig. 3c).

Chromatin interaction mapping using Hi-C data in adult and fetal brain identified 512 unique significantly interacting genes (P = 2.84 × 10⁻⁸; Supplementary Table 28), of which 60 were associated with all six chromatin annotations (Supplementary Fig. 4) and 20 overlapped with TWAS and/or PWAS findings, including DPYSL5, KHK, MAPRE3, MST1R, NEK4, GNL3, GRK4, UHRF1BP1 and VKORC1 (Fig. 3c and Supplementary Tables 26–28).

Based on concordant evidence from colocalization analyses in TWAS and PWAS (COLOC PP4 > 0.80), 104 unique genes (TWAS eQTL—139, TWAS sQTL—20 and PWAS—14) were putatively causal for pain intensity (Supplementary Tables 26 and 27), of which 10 (including DPYSL5, GRK4, KHK and MST1R) were validated by summary-based Mendelian randomization (SMR) analysis (P_HEIDI > 0.05; Fig. 3d and Supplementary Table 29). Among the 104 genes, 6 (CHMP1A, GRIA1, GRK4, MST1R, STMN3 and TRAF3) captured 50% or more of the FINEMAP PP (Supplementary Table 25). Notably, the MST1R intronic locus (rs9815930), which is in a credible set that harbors four other variants in high LD with the new index variant rs2247036 (nearest gene—TRAIP; Extended Data Fig. 7), displayed the most robust causal effects from COLOC and SMR in more than one brain tissue (Fig. 3d).

We also explored the enrichment of causal genes and proteins in the dorsal root ganglia (DRG), which are important for the transduction of nociceptive signals from the periphery to the CNS. None of the causal genes or proteins (n = 104) was enriched in human or mouse DRG (enrichment score > 0.5; Supplementary Fig. 5a). Supporting results from TWAS and PWAS, 63 unique genes (human—38 and mouse—49) were primarily enriched in the CNS, of which 22 (including GRK4, GRIA1, MAPRE3, NEK4, STMN3 and TRAF3) showed common enrichment patterns across species (Supplementary Fig. 5b).

Integrating fine mapping, colocalization and SMR prioritized 156 high-confidence genes underlying the pain intensity GWAS association, of which 5 are exonic and missense (Supplementary Table 30), and 151 exert their effect via gene or protein expression.

Correlates of pain intensity

The strongest positive genetic correlations of pain intensity were with other pain phenotypes (for example, multisite chronic pain r_g = 0.789, osteoarthritis r_g = 0.710, neck/shoulder pain r_g = 0.669, back pain r_g = 0.697, hip pain r_g = 0.729 and knee pain r_g = 0.637; Fig. 4). Of 72 medical, anthropometric or psychiatric traits associated epidemiologically with pain severity, 56 were significantly genetically correlated with pain intensity in EAs (Bonferroni, P < 5.62 × 10⁻⁴; Fig. 4 and Supplementary Table 31).

Fig. 4 |. Genetic correlation.

Genetic correlation for pain intensity using LDSC. All points passing Bonferroni correction (P = 5.62 × 10⁻⁴ (0.05/89)) are plotted. The color of the circle indicates the phenotypic category. The vertical dashed line represents genetic correlation = 0. BMI, body mass index; HDL, high-density lipoprotein; T2D, type 2 diabetes; CUD, cannabis use disorder; CWP, chronic widespread pain; PAU, problematic alcohol use; MDD, major depressive disorder; ADHD, attention deficit/hypersensitivity disorder.

Notably, the liability to pain intensity was significantly positively genetically correlated with neuroticism, depression, insomnia, a variety of smoking-related measures, cannabis use disorder, alcohol dependence, OUD and being overweight or obese (Fig. 4). As in prior studies of other pain-related phenotypes^17,22,36, pain intensity was significantly negatively genetically correlated with educational attainment, cognitive performance, intelligence and age of smoking initiation (Fig. 4). Relevant to drug repurposing, pain intensity was also positively genetically correlated with the use of a variety of analgesic and anti-inflammatory drugs (Fig. 4). We also found significant r_gs with pain intensity for several medical conditions and health outcomes in the UKBB, including genitourinary disease, chronic bronchitis and angina (Bonferroni P < 3.72 × 10⁻⁵; Supplementary Table 35). In AAs, pain intensity was positively genetically correlated with post-traumatic stress disorder (PTSD)-related features (for example, re-experiencing and hyperarousal) and nominally associated (P < 0.05) with substance use traits (for example, maximum alcohol intake and smoking trajectory; Supplementary Table 33).

In the Yale–Penn sample, we calculated polygenic risk scores (PRSs) for 4,922 AAs and 5,709 EAs for use in phenome-wide association studies (PheWAS). Among AAs, no association survived Bonferroni correction (Supplementary Fig. 6 and Supplementary Table 34). In EAs, PheWAS identified 147 phenotypes, including 107 in the substance-related domain (40 opioid-related, 30 cocaine-related, 20 tobacco-related, 12 alcohol-related and 6 cannabis-related) and 39 in other domains (9 medical, 18 psychiatric (9 PTSD, 5 attention-deficit/hyperactivity disorder (ADHD), 2 conduct disorder and 2 antisocial personality disorder), 7 early childhood environmental and 5 demographic phenotypes) significantly associated with the pain PRS (Supplementary Fig. 7 and Supplementary Table 34). The most significant findings were a negative association of the pain PRS with educational attainment (P = 2.39 × 10⁻²⁶) and a positive association with the Fagerström test for nicotine dependence (P = 4.71 × 10⁻²⁵). Opioid dependence was also positively associated with the pain PRS (P = 3.87 × 10⁻¹²) and remained significant for a PRS based on the supplementary GWAS that excluded individuals with an OUD diagnosis (odds ratio (OR) = 1.27, P = 1.35 × 10⁻⁶).

In PMBB, we calculated PRS for 10,383 AAs and 29,355 EAs. In AAs, no association with the pain PRS survived Bonferroni correction (Supplementary Fig. 8 and Supplementary Table 35). In EAs, the pain PRS was associated with 63 phenotypes, including 7 pain phenotypes and 6 psychiatric disorders (that is, substance-, depression- and anxiety-related traits). The pain PRS was also associated with circulatory system traits (n = 11), infectious disease (n = 4), endocrine/metabolic traits (n = 8), genitourinary traits (n = 2), musculoskeletal traits (n = 3) and neoplasms (n = 4). The most significant findings were positive correlations with obesity (P = 1.97 × 10⁻⁴⁵) and tobacco use disorder (P = 1.55 × 10⁻²⁴) and a negative association with benign neoplasm of skin (P = 2.67 × 10⁻²⁶; Supplementary Fig. 9 and Supplementary Table 35). In females, the pain PRS was positively associated with sleep apnea and obstructive sleep apnea and negatively associated with disorders of refraction, degenerative skin conditions and astigmatism (P < 3.65 × 10⁻⁵; Supplementary Table 35). The pain PRS was negatively associated with elevated prostate-specific antigen in males (P = 1.03 × 10⁻⁶; Supplementary Table 35).

Two-sample Mendelian randomization (MR) between genetically correlated traits (n = 16) and pain intensity yielded nine traits with evidence of heterogeneity (Cochran, P < 0.05) and no horizontal pleiotropy (MR–Egger interval, P > 0.05), three of which were bidirectional (Supplementary Table 36). Genetically predicted higher depressed affect subcluster, neuroticism and smoking initiation had a significant positive bidirectional causal effect with pain intensity (Supplementary Table 36). Furthermore, increased opioid use (N02A) positively predicted pain intensity.

Genetically inferred drug repurposing

Of the 156 genes in EAs with evidence supporting causality from fine mapping and functional genomic prediction, 20 were present in the druggable genome database³⁷ (Supplementary Table 37). Of these druggable candidate genes, 11 (including GRIA1, GRK4 and MST1R) are tier-1 candidates, which include targets of licensed drugs and drugs in clinical trials, 4 genes (for example, NEK4 and RYR2) are in tier-2 and 4 are in tier-3 (Supplementary Table 30). Within tier-1, drugs that interact with GRK4 (a credible pain gene locus in moderate LD with the new index variant NOP14*rs71597204; Extended Data Fig. 8) are β-blockers (atenolol and metoprolol) and a calcium-channel blocking agent (verapamil; Fig. 5), which have analgesic effects in osteoarthritis^38,39 and migraine⁴⁰. Another tier-1 candidate gene—GRIA1—is targeted by anesthetics (sevoflurane, isoflurane and desflurane), antiepileptics (topiramate and perampanel), analgesics (methoxyflurane), psychoanaleptics (piracetam and aniracetam) and a diuretic (cyclothiazide; Fig. 5). Drug classes for pain intensity also included antihemorrhagic agents (for example, fostamatinib (tier-1, MST1R and FYN; tier-2, NEK4) and menadione (VKORC1); Fig. 5 and Supplementary Table 37).

Fig. 5 |. Drug repurposing.

Druggable targets and drug interactions for eight credible genes associated with pain intensity. For a full list of credible drug targets, see Supplementary Table 37.

Of the seven genes associated with pain intensity in AAs, PPARD, which harbors the new genetic signal discovered in this study, is a tier-1 druggable candidate with 30 interacting drug classes (Supplementary Table 37). The PPARD-negative modulator sulindac is an approved nonsteroidal anti-inflammatory and antirheumatic drug used to treat osteoarthritis.

Discussion

In the largest multi-ancestry, single-sample GWAS of pain intensity to date, cross-ancestry analyses identified 125 independent risk loci, including 124 autosomal and 1 X-chromosomal loci, including 66 new pain associations. Prior GWASs for chronic pain phenotypes identified 99 loci^16–20,24, although only in EA individuals. The diversity and size of the MVP sample enabled us to identify new association signals in both AAs (PPARD*rs9470000) and HAs (nearest genes RNU6-461P*rs146862033 and RNU6-741P*rs1019597899).

Findings from gene-set analysis, tissue enrichment and cell-type specificity highlight new biological pathways underlying pain, implicating the brain and providing genetic support to the current understanding of the pathophysiology of pain severity⁴¹. Genes predominantly expressed in the CNS, particularly in the cerebellum, cerebellar hemisphere and cortex, rather than in the DRG, are salient in modulating the intensity of pain as assessed here, consistent with prior associations of sustained chronic pain intensity with increased activity in these brain regions^42–44. Our findings are also consistent with prior reports^17,18,45,46 of enriched gene expression in brain that contributes to pain intensity in a dose- and time-dependent manner and may involve specific neuronal processes in brain regions implicated in emotional processing⁴¹.

Evidence that GABAergic neurons are cells of specific interest is a key new finding. GABA has long been implicated in the modulation and perception of pain^47–49, and specific GABAergic activity in the midbrain has been implicated as a modulator of pain and anxiety⁵⁰. Altered GABA levels have been reported in individuals with various types of pain^51,52 and have been associated with greater self-reported pain⁵³. Targeting GABA function in brain regions enriched for pain intensity could be a new therapeutic strategy.

Eleven of 156 prioritized genes encode druggable small molecules that are targets of licensed drugs or those in clinical trials, representing drug repurposing opportunities for treating chronic pain. We highlight MST1R, GRK4 and GRIA1, each with at least three lines of evidence supporting their involvement in chronic pain. MST1R encodes a cell-surface receptor with tyrosine kinase activity that mediates the inflammatory response. The MST1R inhibitor fostamatinib—prioritized by our drug repurposing analyses—is a possible therapeutic target for rheumatoid arthritis⁵⁴. Increased connectivity between frontal midbrain regions implicated in affective pain processing has been reported in patients with rheumatoid arthritis⁵⁵. Here we demonstrated that MST1R is highly expressed in the adult brain cortex, cerebellum and cerebellar hemisphere, suggesting that MST1R inhibitors may be effective in treating patients with inflammation-related pain.

GRK4 encodes G protein-coupled receptor kinase 4 and has been linked with hypertension⁵⁶, which is associated with chronic pain at the population level^57,58. Of note, we showed that GRK4 is significantly upregulated in the cerebellar hemisphere, fine maps to an intronic variant with >95% PP and is a target of β-blockers. The use of β-blockers has been associated with reduced osteoarthritis pain scores and prescription analgesic use³⁸, as well as fewer consultations for knee osteoarthritis, knee pain and hip pain³⁹. GRIA1 encodes an ionotropic glutamate receptor subunit, an excitatory neurotransmitter receptor at many synapses in the CNS. Loss-of-function mutations in GRIA1 are linked to neurodevelopmental impairments^59,60. The GRIA1 antagonist sevoflurane reduced pain in patients with chronic venous ulcer⁶¹. However, clinical trials of topiramate (another drug target for GRIA1) for treating neuropathic chronic pain were inconclusive⁶². Research on the mechanisms that underlie the biology of these potential drug targets for GRK4 and GRIA1 and their effects on the onset and severity of chronic pain are warranted.

Pain intensity was strongly genetically correlated with other chronic pain phenotypes, with the strongest genetic correlations with multisite chronic pain, followed by pain in specific bodily locations. In line with previous observations in GWASs of other pain-related phenotypes^{17,18,20,21,36}, there were also positive genetic correlations of pain intensity with psychiatric disorders, substance use and use disorders and anthropometric traits.

PheWAS findings in both the Yale–Penn sample—enriched for individuals with substance-related traits—and the PMBB—comprising a medical population—were prominent in EAs. These findings underscore the influence of co-occurring substance-related, psychiatric and medical pathology and educational achievement on the intensity of the pain experience. In contrast, the pain PRS yielded few associations in AAs in either of the target samples, which underscores the need for larger non-European samples for studies of pain intensity.

Two-sample MR analysis supported causal associations between pain and multiple traits. Smoking has previously been associated with greater pain intensity, but studies can be confounded by socioeconomic factors, and a bidirectional relationship has been proposed⁶³. Here we provide evidence for a bidirectional causal relationship between pain and smoking initiation. In line with previous findings^18,25, pain intensity had a bidirectional causal effect on the risk of both depressed affect subcluster and neuroticism, suggesting that greater pain could predispose individuals to increased risk for these psychiatric disorders and vice versa. Supporting the positive genetic correlation between opioid use and pain intensity, MR showed evidence of a causal effect of opioid use on pain intensity.

Pain intensity is a complex, polygenic trait with hundreds of genetic loci contributing to it. The evidence adduced here of pleiotropy of pain intensity with psychiatric traits such as neuroticism and depression reflects the contribution of nonphysical factors to the experience of pain intensity. This is consistent with the observed significant tissue-group enrichment in CNS, the predominant gene expression findings in brain (including the hippocampus and limbic system) and the SNP-based enhancer enrichments in histone modification in brain tissues (including the dlPFC, inferior temporal lobe, angular gyrus and anterior caudate).

The NRS phenotype, although a quantitative trait and thus more informative than a binary trait, is based on subjective reports. However, because the subjective experience of pain is a key defining feature of the clinical phenomenon^1,64, the phenotype has high public health significance. Pain scores recorded by clerks and nurses in a clinical setting may underestimate the patient’s response. In earlier work, self-reported pain from a direct patient survey correlated well with scores recorded in a VA clinical setting⁶⁵, despite lower scores recorded in the clinic. Nonetheless, the imprecise measurement of pain intensity likely yields lower power for gene discovery. We chose not to stratify the analyses using different types of pain (for example, secondary to osteoarthritis versus lower back pain versus peripheral neuropathy) to maximize statistical power but will examine different sources of pain in future analyses. We reduced a large number of pain assessments by taking individuals’ median of median NRS scores as a trait for GWAS. In subsequent analyses, we will evaluate other methods for characterizing pain severity (for example, pain trajectories). Our sample was also limited by being comprised predominantly of male veterans, which given sex differences in the experience and frequency of pain¹⁹ limits the application of the findings to the general population. Studies of pain intensity in large samples with more even sex distributions are needed. Although our sample was more diverse than prior GWAS of pain traits, analyses in the AA and HA samples were underpowered. Finally, we lacked a suitable replication sample, so efforts are needed to replicate the findings reported here.

Despite these limitations, the large MVP sample and the informative quantitative trait measured repeatedly within subjects, which provided a proxy for chronic pain, identified many new loci contributing to the trait. Downstream analyses localize the genetic effects largely to four CNS regions and using single-cell RNA-seq data link them specifically to GABAergic neurons. Combined with drug repurposing findings that implicate 20 druggable targets, this study provides a basis for studies of new, nonopioid medications for use in alleviating chronic pain.

Methods

Overview of analyses

We conducted ancestry-specific GWASs of pain scores using an 11-point ordinal NRS in (1) all AAs, EAs and HAs with pain ratings from the MVP, (2) a subset of these participants that excluded those with a lifetime OUD diagnosis, (3) a subset of participants that excluded those with pain ratings = 0 and (4) males and females separately (sex-stratified), each followed by a cross-ancestry meta-analysis. Details on phenotyping are provided below. Downstream analyses are based principally on the GWAS of pain scores in the full sample, complemented by the estimated heritability and genetic correlations (r_gs) for the sample exclusive of participants with OUD and stratified by sex.

MVP cohort

The MVP²⁸ is an EHR-based cohort comprising >900,000 veterans recruited at 63 VA medical centers nationwide. All participants provided written informed consent, a blood sample for DNA extraction and genotyping and approval to securely access their EHR for research purposes. The protocol and consent were approved by the Central Veterans Affairs Institutional Review Board (IRB) and all site-specific IRBs. All relevant guidelines for working with human participants were followed in the conduct of the study.

Phenotype description

As early as 2000, the VA recommended using the NRS to routinely measure pain in clinical practice as a ‘fifth vital sign’⁶⁶. Since that time, veterans have been asked to rate their pain severity in response to the question: ‘Are you in pain?’ They then rated their current pain on a scale of 0–10, where ‘0 is no pain and 10 is the worst pain imaginable’. Participants had at least one inpatient or outpatient pain rating in the EHR. We included 598,339 individuals with 76,798,104 NRS scores (median number of scores = 109 and interquartile range = 28–351) in the primary GWAS. To reduce the large number of pain observations, we calculated the median pain score by year for each participant and the median of the annual median pain scores. In a secondary GWAS, we excluded individuals with a documented International Classification of Diseases (ICD)-9/10 diagnosis code for OUD in the EHR, yielding a total of 566,959 study participants. Demographic characteristics for the secondary analysis sample are presented in Supplementary Table 1.

Genotyping and imputation

DNA samples were genotyped on the Affymetrix Axiom Biobank Array (MVP Release 4). For genotyped SNPs, standard quality control (QC) and subsequent imputation were implemented. Full details about SNP and sample QC by the MVP Genomics Working Group are published⁶⁷. Briefly, DNA samples were removed for sex mismatch, having seven or more relatives in MVP (kinship > 0.08), excessive heterozygosity or genotype call rate <98.5%. Variants were removed if they were monomorphic, had a high degree of missingness (call rate < 0.8) or a Hardy–Weinberg equilibrium (HWE) threshold of P < 1 × 10⁻⁶ both in the entire sample using a principal component analysis-adjusted method and within one of the three major ancestral groups (AA, EA and HA).

Genotype phasing and imputation were performed using SHA-PEIT4 (v.4.1.3)⁶⁸ and Minimac4 software⁶⁹, respectively. Biallelic SNPs were imputed using the African Genome Resources reference panel by the Sanger Institute (comprising all samples from the 1000 Genomes Project phase 3, version 5 reference panel⁷⁰, and 1,500 unrelated pan-African samples). Nonbiallelic SNPs and indels were imputed in a secondary imputation step using the 1000 Genomes Project phase 3, version 5 reference panel⁷⁰, with indels and complex variants from the second imputation merged into the African Genome Resources imputation.

We randomly removed one individual from each pair of related individuals (kinship > 0.08, n = 31,010). The HARE method⁷¹ was used to classify subjects into major ancestral groups (AA = 112,968, EA = 436,683 and HA = 48,688), and QC of imputed variants was performed within each ancestral group. Additional QC steps were carried out for the chrX analysis to reduce the risk of false-positive associations from sex-specific genotyping errors. For this, we excluded variants in pseudo-autosomal regions based on excessive heterozygosity rates. For both autosomal and chrX analysis, SNPs with imputation quality (INFO) score < 0.7; minor allele frequency (MAF) in AAs < 0.005, EAs < 0.001 and HAs < 0. 01; a genotype call rate < 0.95 or an HWE P < 1 × 10⁻⁶ was excluded.

Association analyses and risk locus definition

Genome-wide association testing was based on a linear regression model using PLINK (v.2.0)⁷² and was adjusted for sex, mean age of assessment and the first ten within-ancestry genetic principal components (PCs). For the chrX association analysis, we implemented a standard linear regression model using PLINK assuming x-inactivation (males were coded as 0/2 and females as 0/1/2), adjusting for the mean age of assessment and the first ten PCs. Both autosomal and X-chromosomal analyses were performed within ancestry groups in (1) all participants and (2) in males and females separately. Due to substantial differences in sample size across ancestral groups, meta-analyses were performed using a sample-size weighted method in METAL⁷³. Variants with P < 5 × 10⁻⁸ were considered GWS. Because the LD intercept (1.1, s.e. = 0.01) and attenuation ratio (0.13, s.e. = 0.01) of the LDSC showed minimal evidence of inflation or confounding, suggesting that none of the GWS lead SNPs showed evidence of heterogeneity across ancestries, we did not select the genomic control option in METAL.

To identify risk loci and their lead variants, we performed LD clumping in FUMA⁷⁴ at a range of 3,000 kb, r² > 0.1 and the respective ancestry 1000 Genomes reference panel⁷⁰. Following clumping, genomic risk loci within 1 MB of one another were incorporated into the same locus. We used GCTA-COJO⁷⁵ to define independent variants by conditioning them on the most significant variant within the locus. After conditioning, significant variants (P < 5 × 10⁻⁸) were considered independently associated. We performed a sign test to compare the direction of SNP effects across individual ancestral datasets. Independent lead variants in EAs were examined in AAs and HAs, and a binomial test was used to evaluate the null hypothesis that 50% of variants have the same effect direction across ancestries. For lead SNPs in EAs that were absent in AAs and HAs, we considered proxy GWS SNPs (P < 5 × 10⁻⁸) in high LD with the EA lead variant (r² ≥ 0.8).

To prioritize credible sets of variants driving our GWAS results, we used FINEMAP⁷⁶ to fine-map regions defined by LD clumps (r² > 0.1). Because fine mapping requires data from all markers in the region of interest⁷⁷, we merged LD clumps that physically overlapped (within a 1-MB window of the lead variant) and excluded SNPs in the major histocompatibility complex (MHC) region due to its complexity. FINEMAP credible set reports the likelihood of causality using the marginal PP, which ranges from 0 to 1, with values closer to 1 being most likely causal.

SNP-based heritability and functional enrichment

We used the LDSC regression⁷⁸ method to estimate the SNP-based heritability (h²_SNP) of pain intensity (in both the full and supplementary samples) in all ancestry groups based on common SNPs in HapMap3 (ref. 79). To ensure matching of the population LD structure, precalculated LD scores for EAs were derived from the 1000 Genomes European reference population (version 3)⁷⁰ using LDSC⁷⁸. In-sample LD scores for AAs and HAs were calculated from MVP AA and HA genotype data using cov-LDSC⁸⁰.

We used S-LDSC to partition the SNP heritability for pain intensity among EAs and explored the enrichment of the partitioned heritability by functional genomic categories^81,82 using the following three models: (1) a baseline-LD model that contains 75 overlapping annotations, including coding and regulatory regions of the genome and epigenomic features⁸¹; (2) a specific tissue model that examines ten overlapping cell-type groups derived from 220 cell-type-specific histone marks, including methylated histone H3 Lys4 (H3K4me1), trimethylated histone H3 Lys4 (H3K4me3), acetylated histones H3 Lys4 (H3K4ac) and H3K27ac⁸² and (3) a multi-tissue model based on gene expression and chromatin datasets generated by GTEx⁸³ and the Roadmap Epigenomics Mapping Consortium⁸⁴. For each model, we excluded multi-allelic and MHC region variants. Functional categories within each model were considered significantly enriched based on a Bonferroni-corrected P value.

Gene-set functional characterization

We applied multimarker analysis of genomic annotation (MAGMA) v.1.08 (ref. 85) in FUMA (v1.3.6a)⁷⁴ to identify genes and gene sets associated with the findings from the pain intensity GWAS and meta-analysis. Using the default setting in MAGMA, we mapped GWS SNPs to 18,702 protein-coding genes according to their physical position in NCBI build 37. We also used chromatin interaction (Hi-C) coupled MAGMA (H-MAGMA)⁸⁶ to assign noncoding (intergenic and intronic) SNPs to genes based on their chromatin interactions. H-MAGMA uses six Hi-C datasets derived from fetal brain, adult brain (n = 3), induced pluripotent stem cell (iPSC)-derived neurons and iPSC-derived astrocytes⁸⁷. We applied a Bonferroni correction (MAGMA, α = 0.05/18,702; H-MAGMA, α = 0.05/293,157/6) to identify genes significantly associated with pain intensity, correcting for all genes tested in each analysis (see Supplementary Tables 15 and 21 for full lists).

To determine the plausible tissue enrichment of mapped genes, we integrated our cross-ancestry and EA GWAS results with gene expression data from 54 tissues (GTEx v8) in FUMA⁷⁴. Next, we used FUMA to curate gene sets and Gene Ontology terms (from the Molecular Signature Database v.7.0; ref. 88). We corrected for gene size, density of variants and LD pattern between genes in each tissue (Bonferroni-corrected α = 0.05/54).

Enrichment for CTS transcriptomic profiles was performed in FUMA⁷⁴ using 13 human single-cell RNA-sequencing (scRNA-seq) datasets derived from brain⁸⁹ (see Supplementary Table 14 for a detailed list). FUMA estimates CTS transcriptomic enrichment from the scRNA-seq in the following three ways: (1) per selected dataset, (2) within datasets using a conditionally independent analysis (based on stepwise conditional testing of P values for each cell type that passes Bonferroni correction within the same dataset) and (3) across datasets (testing for PS across the results from step 2). PS reports the confidence level for observed cell-type enrichment as low significance: <0.5, jointly significant: 0.5–0.8 and independently significant: >0.8. We considered CTS enrichments with conditional independent signals (P < 0.05) and PS > 0.5 to be driven by joint/independent genetic signals in our pain intensity GWAS results.

Transcriptomic and proteomic regulation

To identify genes and proteins whose expression is associated with pain intensity, we integrated EA GWAS results with human brain transcriptomic (eQTL, n = 452; and sQTL, n = 452)^83,90 and proteomic (n = 722)⁹¹ data. We also obtained pretrained models of gene expression from GTEx v.8 for five brain tissues significantly enriched in MAGMA analyses—cerebellum, cerebellar hemisphere, cortex, frontal cortex and anterior cingulate cortex⁸³. Human brain transcriptomic and proteomic data for dlPFC were derived from the study in ref. 90. TWAS and PWAS analyses were performed using the FUSION pipeline⁹² with Bonferroni correction (α = 0.05/n genes tested) to account for multiple testing.

We used colocalization (coloc R package⁹³ in FUSION⁹²) as our primary method to identify SNPs that mediate association with pain intensity through effects on gene and protein expression and a posterior colocalization probability of 80% to denote a shared causal signal. To test the robustness of the colocalized signals, we also performed SMR analyses⁹⁴. We applied the HEIDI test⁹⁴ to filter out SMR signals (P_HEIDI < 0.05) due to LD between pain-associated variants and eQTLs/sQTLs. Human brain cis-eQTL and cis-sQTL summary data were obtained from ref. 95 and GTEx⁸³. For genomic regions containing multiple genes with significant SMR associations, we selected the top-associated cis-eQTL. We used Bonferroni correction to correct for multiple testing (α = 0.05/n genes tested).

To explore the enrichment of causal genes and proteins in the DRG, we accessed human and mouse RNA-seq data from 13 tissues (6 neural and 7 non-neural) from the DRG sensoryomics repository⁹⁶. The data contain relative gene abundances in standardized transcripts per million mapped reads and have been normalized to allow comparison across genes. The proportions of gene expression in the CNS (neural proportion score) and DRG (DRG enrichment score) in the context of profiled tissues were calculated, as described previously⁹⁶. Scores ranging from 0 to 1 were used to denote the strength of tissue enrichment.

Drug repurposing

We examined the drug repurposing status of genes in EAs (n = 156) with high causal probability from fine mapping and transcriptomic and proteomic analyses using the Druggable Genome database³⁷. For completeness, we also included the significantly associated genes mapped to GWS variants and MAGMA results in AAs (n = 7) and HAs (n = 2). The Druggable Genome database contains 4,479 coding gene sets with the potential to be modulated by a drug-like small molecule based on their nucleotide sequence and structural similarity to targets of existing drugs³⁷. This druggable genome was divided into three tiers. Tier-1 (n = 1,427) contains targets of licensed small molecules and biotherapeutic drugs (curated from the ChEMBL database⁹⁷) and drugs in clinical development. Tier-2 (n = 682) includes targets with verified bioactive drug-like small molecule binding partners and >50% identity with approved drug targets based on their nucleotide sequence. Tier-3 (n = 2,370) comprises targets or secreted proteins with more distant similarity with an approved drug and members of active protein complexes not included in tiers 1 and 2. All causal genes and those reported in any of the three tiers of the Druggable Genome were also examined for interaction with prescription drug targets in clinical development using the Drug–Gene Interaction Database⁹⁸, which compiles clinical trial information from the FDA, PharmGKB, Therapeutic Target Database and DrugBank databases, among others. We categorized each prescription drug identified using the Anatomical Therapeutic Chemical classification system, retrieved from the Kyoto Encyclopedia of Genes and Genomics (https://www.genome.jp/kegg/drug/).

Genetic correlation

We used LDSC⁷⁸ to calculate the r_g of pain intensity with (1) 89 other published pain, substance use, medication use, psychiatric and anthropometric traits from EA datasets selected using prior epidemiological evidence and (2) 12 psychiatric, substance use and anthropometric traits based on available AA GWAS summary data (see Supplementary Tables 24 and 26 for detailed lists). In EAs, all traits were tested using precomputed LD scores for HapMap3 (ref. 79), while in AAs, LD scores derived using cov-LDSC⁸⁰ from MVP AA genotype data were used. In a hypothesis-neutral manner, we also calculated r_gs of pain intensity with 1,344 published and unpublished traits from the UKBB using the Complex Trait Virtual Lab (CTG-VL; https://genoma.io/). CTG-VL is a free open-source platform that incorporates publicly available GWAS data that allow for the calculation of r_g for complex traits using LDSC⁹⁹. Each set of r_g analyses was Bonferroni corrected to control for multiple comparisons (α = 0.05/number of traits tested).

We also estimated the cross-ancestry r_gs for pain intensity between AAs, EAs and HAs using Popcorn¹⁰⁰, a computational method that determines the correlation of causal-variant effect sizes at SNPs common across population groups using GWAS summary-level data and LD information. Ancestry-specific LD scores were derived from the 1000 Genomes reference population⁷⁰.

PRS-based PheWAS

We calculated PRS for pain intensity and performed a PheWAS analysis in two samples—the Yale–Penn sample and the Penn Medicine Biobank (PMBB). The Yale–Penn sample¹⁰¹ was deeply phenotyped using the Semi-Structured Assessment for Drug Dependence and Alcoholism, a comprehensive psychiatric instrument that assesses physical, psychosocial and psychiatric aspects of SUDs and comorbid psychiatric traits^102,103. As described in detail previously¹⁰¹, genotyping was performed using the Illumina HumanOmni1-Quad microarray, the Illumina HumanCoreExome array or the Illumina Multi-Ethnic Global array, followed by imputation using Minimac3 (ref. 104) and the 1000 Genomes Project phase 3 reference panel⁷⁰ implemented on the Michigan imputation server (https://imputationserver.sph.umich.edu). SNPs with imputation quality (INFO) score <0.7, MAF <0.01, missingness >0.01 or an allele frequency difference between batches >0.04 were excluded, and individuals with genotype call rate <0.95 and related individuals with pi-hat >0.25 were excluded. PCs were used to determine genetic ancestry based on the 1000 Genomes Project phase 3 (ref. 70). The resulting dataset included 4,922 AAs and 5,709 EAs.

The PMBB¹⁰⁵ is linked to EHR phenotypes. PMBB samples were genotyped with the GSA genotyping array. Genotype phasing was performed using EAGLE¹⁰⁴, and imputation was performed using Minimac3 (ref. 104) on the TOPMed Imputation server⁶⁹. Following QC (INFO < 0.3, missingness > 0.95, MAF > 0.5 and sample call rate > 0.9), PLINK 1.90 was used to identify and remove related individuals based on identity by descent (pi-hat > 0.25). To estimate genetic ancestry, PCs were calculated using SNPs common to the PMBB and the 1000 Genomes Project phase 3 (ref. 70) and the smartpca module of the Eigensoft package (https://github.com/DReichLab/EIG). Participants were assigned to an ancestral group based on the distance of ten PCs from the 1000 Genomes reference populations. The resulting dataset included 10,383 AAs and 29,355 EAs.

PRSs for pain intensity were calculated in the Yale–Penn and the PMBB datasets using PRS-Continuous shrinkage software (PRS-CS)¹⁰⁶, with the default setting used to estimate the shrinkage parameters and the random seed fixed to 1 for reproducibility. To identify associations between the pain intensity PRSs and phenotypes, we performed a PheWAS in each dataset by fitting logistic regression models for binary traits and linear regression models for continuous traits. Analyses were conducted using the PheWAS v0.12 R package¹⁰⁷ with adjustment for sex, age at enrollment (in PMBB) or at interview (in Yale–Penn) and the first ten PCs within each genetic ancestry. We Bonferroni corrected each ancestry-specific analysis (Yale–Penn EAs and AAs, P < 7.87 × 10⁻⁵; PMBB EAs and AAs, P < 3.65 × 10⁻⁵).

MR

We used two-sample MR¹⁰⁸ to evaluate causal associations between genetically correlated traits and pain intensity among EAs only because the two other population groups provided inadequate statistical power for the analysis. Of the 56 traits that showed significant r_g, we removed traits with phenotypic similarity across each of the tested r_g categories (Supplementary Table 31), selected traits with higher r_g and excluded traits without known biopsychosocial associations with pain. This left 16 traits for MR analysis. Instrumental variants (IVs) were SNPs associated with exposure at P < 1 × 10⁻⁵ and a clumping threshold of r² = 0.01.

To quantify the strength of IVs, we calculated the F statistics of all genetic instruments using the per-allele effect size of SNP association with the phenotype (β) and s.e. using the following formula^109,110: F statistic = (β/s.e.)². IVs with F statistic estimates <10 were considered weak instruments that could bias results¹¹¹. We used Steiger’s test¹¹² to determine whether the SNP–outcome correlation is greater than the SNP–exposure correlation. SNPs that fail Steiger’s test may not be primarily associated with the exposure (Steiger, P > 0.05) and were filtered out before MR analysis. Because pleiotropy can bias MR findings¹¹³, we investigated its possible presence by assessing heterogeneity in the MR estimates across SNPs, using the I² index and Cochran’s Q heterogeneity test¹¹⁴. Finally, MR–Egger intercepts were used to assess the bias due to weak IVs and the possibility of horizontal pleiotropy. Potential causal effects were those for which at least two MR tests were significant after multiple correction (P = 3.13 × 10⁻³, 0.05/16) and did not violate the assumption of horizontal pleiotropy (MR–Egger intercept, P > 0.05).

Extended Data

Extended Data Fig. 1 |. Overview of the study.

Top left: primary GWAS analyses for pain intensity. Within ancestry, GWAS for African American (AA), European American (EA) and Hispanic American (HA) followed by cross-ancestry meta-analysis. These results were used for all downstream analyses. Top right: secondary GWAS analyses for pain intensity. Bottom: downstream analyses were conducted using the cross-ancestry, AA and EA GWAS results as indicated by color shadings: primary GWAS (green) and supplementary GWAS (brown).

Extended Data Fig. 2 |. Manhattan plot for the pain intensity in European American GWAS analysis.

Identified 87 independent risk loci. Novel loci (n = 52) are annotated in pink. The red line indicates GWS after correction for multiple testing (P < 5 × 10⁻⁸).

Extended Data Fig. 3 |. Effect-effect plot of cross-ancestry meta-analyses lead SNPs in the primary and secondary GWASs.

The magnitude and direction of the effect sizes are plotted for each GWAS. The results show significant (P < 2.2 × 10⁻¹⁶) high correlation (Pearson r test, two-sided) between the effect sizes (β) of pain intensity lead SNPs for primary GWAS and those for non-OUD (r = 1, a), non-zero (r = 0.97, b), males (r = 1, c) and females (r = 0.88, d).

Extended Data Fig. 4 |. LDSC genetic correlations for pain intensity primary and secondary GWAS.

African American: primary GWAS, n = 112,968; non-OUD GWAS, n = 104,050; non-zero GWAS, n = 61,499; male GWAS, n = 97,343; female GWAS, n = 15,625. European American: primary GWAS, n = 436,683; non-OUD GWAS, n = 416,740; non-zero GWAS, n = 202,784; male GWAS, n = 404,510; female GWAS, N = 32,173. Error bar is presented as 95% confidence interval.

Extended Data Fig. 5 |. MAGMA tissue enrichment for pain intensity in cross-ancestry and European American GWAS results.

Tissue enrichment analyses were conducted using FUMA. Bonferroni correction threshold (represented by the black dashed line) = 9.25 × 10⁻⁴ (0.05/54).

Extended Data Fig. 6 |. Gene-based Manhattan plots for cross-ancestry, European American and African American GWAS.

Gene-based association analyses were conducted using FUMA and genes that survive multiple correction are annotated (Bonferroni p = 2.67 × 10⁻⁶ [0.05/18,702]).

Extended Data Fig. 7 |. Regional plot for TRAIP*rs2247036 and MST1R*rs9815930 on chromosome 3.

Credible locus prioritized by FINEMAP (PP > 0.5) is annotated with red rings. The MST1R*rs9815930 locus is in high LD (r² > 0.8) with the lead variant TRAIP*rs2247036.

Extended Data Fig. 8 |. Regional plot for NOP14*rs71597204 and GRK4*rs2798303 on chromosome 4.

Credible locus prioritized by FINEMAP (PP > 0.5) is annotated with red rings. The GRK4*rs2798303 locus is in moderate LD (r² > 0.4) with the lead variant NOP14*rs71597204.

Data availability

The cross-ancestry and within-ancestry GWAS and meta-analysis summary-level association data will be available in the database of Genotypes and Phenotypes (dbGaP) (https://www.ncbi.nlm.nih.gov/gap/) under accession phs001672 ‘Veterans Administration (VA) MVP Summary Results from Omics Studies’. Registration and approval are needed following dbGaP’s data access process.