Multivariate, Multi-Omic Analysis in 799,429 Individuals Identifies 134 Loci Associated with Somatoform Traits

Abstract

Introduction

Somatoform traits (e.g., health anxiety, somatic preoccupation, and bodily distress symptoms) are prevalent and pose challenges to clinical practice. Understanding their genetic basis could improve diagnostic and therapeutic approaches.

Methods

Using available summary statistics, we conducted a multivariate genome-wide association study (GWAS) and multi-omic analysis of four somatoform traits – fatigue, irritable bowel syndrome, pain intensity, and health satisfaction – in 799,429 individuals genetically similar to European reference panels.

Results

The GWAS identified 134 loci associated with a somatoform common factor, including 44 loci not significant in the input GWAS and 8 novel loci for somatoform traits. Novel loci were mechanistically informative, mapping to the DNM1 gene and the protocadherin gene cluster (PCDHA1-4), which are involved in nociceptor sensitization and synaptogenesis, respectively. Gene-property analyses highlighted an enrichment of genes involved in synaptic transmission and enriched expression in 11 brain tissues and the pituitary. Across two brain transcriptomic datasets, we identified 16 high-confidence genes whose expression in enriched tissues was associated with somatoform traits. There was substantial polygenic overlap (76–83%) between the somatoform and externalizing, internalizing, and general psychopathology factors. Somatoform polygenic scores were associated with obesity, type 2 diabetes, and tobacco use disorder in independent biobanks. Drug repurposing analyses suggested potential therapeutic targets, including MEK inhibitors, while Mendelian randomization analyses indicated potentially protective effects of gut microbiota.

Discussion

Consistent with emerging medical and genetic knowledge, somatoform traits have a shared etiology and considerable polygenic overlap with psychopathology. The biological insights from drug repurposing and Mendelian randomization analyses could provide promising avenues for treatment development.

Introduction

Persistent physical symptoms (PPS) adversely impact quality of life, increase healthcare utilization, and contribute to disability [1]. These symptoms may occur alone or as part of functional somatic syndromes, such as irritable bowel syndrome (IBS). Up to 10% of the population is affected by at least 1 functional somatic syndromes [2], with higher rates at the symptom level. In a study of over 70,000 patients, nearly half reported at least one somatic symptom with no identified organic cause [3]. Despite their prevalence and impact, the etiology of PPS remains poorly understood.

Genome-wide association studies (GWASs) of individual PPS have implicated gene expression and biological processes in the immune and central nervous systems. However, phenotypic and genetic correlations across PPS and between PPS and psychopathology suggest a partially shared etiology [4, 5]. This aligns with the Hierarchical Taxonomy of Psychopathology (HiTOP) framework, which proposes a somatoform spectrum that reflects the common liability to PPS [6, 7]. Genomic structural equation modeling (gSEM) has provided further evidence for a shared genetic architecture that underlies chronic and nociplastic pain conditions [8, 9]. Extending this work to somatoform traits more broadly could identify common biological pathways and improve the assessment and treatment of PPS.

We leveraged gSEM to examine the shared genetic architecture of fatigue, health satisfaction, IBS, and pain intensity – traits that index a potential somatoform spectrum [7]. Headache/migraine was initially included, but it was not retained due to poorer fit on the latent factor. Next, we performed downstream analyses using the GWAS summary statistics (Fig. 1). These included gene prioritization efforts, gene-property analyses, MiXeR, genetic correlations, and phenome-wide association studies (PheWAS) in BioVU, Penn Medicine BioBank (PMBB), and Yale-Penn. Finally, we conducted drug repurposing and Mendelian randomization (MR) of gut microbiome taxa, given the significance of the gut-brain-axis [10]. With this approach, we aimed to deepen our understanding of the genetic basis of somatoform traits.

Fig. 1.

Overview of analyses. The somatoform factor N reflects the effective sample size. IBS, irritable bowel syndrome.

Methods

Summary statistics were chosen to correspond to the HiTOP somatoform spectrum [6], consisting of five components: (1) bodily distress symptoms (e.g., fatigue and pain), (2) conversion symptoms (e.g., motor/sensory impairments without an identified medical basis), (3) health anxiety, (4) disease conviction (i.e., persistent belief that one has an illness despite medical reassurance), and (5) somatic preoccupation (i.e., heightened monitoring and attention to sensations) [11]. We selected five GWAS from individuals genetically similar to Europeans (EUR) [12]: fatigue (N = 350,580; http://www.nealelab.is/uk-biobank/), headache/migraine [13] (N = 360,391; not retained due to poorer fit), health satisfaction (N = 119,567; http://www.nealelab.is/uk-biobank/; reverse coded), IBS [14] (N = 486,601), and pain intensity [15] (N = 436,683). Online supplementary materials (for all online suppl. material, see https://doi.org/10.1159/000551114) provide details on phenotype selection. Participants’ written informed consent and Institutional Review Board (IRB) approval was obtained by the original GWAS study authors.

Using the GenomicSEM R package [16], linkage disequilibrium score regression (LDSC) was implemented to estimate genetic covariance matrices. GenomicSEM accounts for sample overlap through estimation of the off-diagonal elements of a sampling covariance matrix. The resulting matrices were used to perform confirmatory factor analysis for a common factor model and to compare a two factor model (online suppl. materials). Model fit was evaluated based on chi-square, Akaike information criterion (AIC), comparative fit index (CFI), and standardized root mean square residual (SRMR) values. We also examined the proportion of variance explained by the common factor to ensure that each trait was sufficiently represented (i.e., standardized loading ≥0.30). In addition, we considered the relative loading of traits and their genetic correlations with one another to determine the optimal final model.

To perform GWAS, we regressed each SNP on the somatoform factor using diagonally weighted least squares estimation. Details on model specification can be found in the online supplementary materials. Q_SNP was used to identify and remove SNPs with heterogeneous effects across the traits. Clumping of the GWAS results was performed using PLINK 1.9 [17] with an r² threshold of 0.1 and a physical distance threshold of 3,000 kb. The novelty of loci was based on the lead SNPs positional overlap (within ± 1,000 kb) with genome-wide significant (GWS) variants from previous GWAS of the included traits and pain traits not included in the factor. For novel SNPs, we performed SNP-level PheWAS using GWAS Atlas [18] and the ieugwasr R package [19].

Gene Mapping and Functional Annotation

In FUMA v1.5.2 [20], SNPs were mapped to protein-coding genes based on: (1) position (≤10 kb), (2) eQTL (BrainSeq [21], PsychENCODE [22], BRAINEAC [23], and GTEx v8 [24] brain), and (3) chromatin interaction mapping (Hi-C brain) [25, 26]. SNPs were functionally annotated using ANNOVAR [27], combined annotation dependent depletion (CADD) [28], and RegulomeDB [29]. We considered SNPs with CADD > 20 to be potentially deleterious and those > 12.37 to be likely pathogenic [30].

Gene-Based Enrichment

We performed gene-set and gene-property analyses using MAGMA v1.08 in FUMA v1.5.2 [31]. SNPs were positionally mapped to protein-coding genes. Using the resulting gene-level p values, analyses were performed for MsigDB v7.0 [32] gene sets and gene ontology terms. Gene-property analyses were performed for 54 tissue types (GTEx v8) [24] and 11 developmental stages (BrainSpan) [26]. These analyses test whether genes expressed in a particular tissue or developmental period are more likely to be associated with the somatoform factor, adjusting for average expression. Cell-type-specificity analyses were conducted across 16 human brain datasets to investigate whether specific cell types were implicated, with a set of conditional analyses performed to correct for associations across datasets (see suppl. materials for details).

eQTL and pQTL Association Analyses

To identify SNPs with associations mediated by gene and protein expression, we performed summary-data-based Mendelian randomization (SMR) analyses [33]. The heterogeneity in dependent instruments (HEIDI) test was used to distinguish pleiotropic effects from linkage. We identified associations that had a P_SMR < 0.05 after Bonferroni correction and a P_HEIDI >0.05. We used the MetaBrain [34] cis-eQTL database from 7 brain regions (n = 8,613). MetaBrain provides coverage of cortical data relevant to symptom perceptions [35] and chronic pain [36]. For blood pQTL analyses, we used two cis- and trans-pQTL databases: (1) UK Biobank Pharma Proteomics Project [37] and (2) deCODE Consortium [38]. Blood pQTL datasets provide information on circulating systemic protein levels, which have relevance to pain and fatigue [39–41].

Transcriptome-Wide Association Studies

We conducted two transcriptome-wide association studies (TWAS) using MetaXcan [42, 43]. First, we used S-MultiXcan to examine associations across GTEx v8 tissues for which gene expression was enriched based on MAGMA results. We complemented this approach with S-PrediXcan using data from PsychENCODE [44], which is enriched for individuals with psychiatric conditions that are often comorbid with PPS [45]. PsychENCODE data includes tissue from the prefrontal cortex, temporal cortex, and cerebellum [44]. TWAS were conducted in addition to SMR as these approaches provide unique but complementary information [46, 47]. SMR uses observed single-eQTL or pQTL associations to test whether expression levels are causally associated with a trait, accounting for LD structure. In contrast, MetaXcan examines predicted gene expression using multiple eQTLs (and tissues, in the case of S-MultiXcan) to enhance power but does not infer causality.

Polygenic Overlap with Psychopathology

Bivariate causal mixture models estimate genetic overlap between traits, even when causal variants have opposite directions of effect [48]. The Dice coefficient estimates the proportion of SNPs that overlap with another trait out of the total estimated causal SNPs for the two traits. We performed bivariate MiXeR to estimate polygenic overlap with externalizing (EXT), internalizing (INT), and EXT+INT factors [49], which are in the HiTOP model [6] and highly comorbid with somatoform traits [50].

Genetic Correlations

We performed genetic correlations with 1,426 phenotypes from publicly available GWAS using the Complex-Trait Genetics Virtual Lab (CTG-VL) [51]. Phenotypes spanned biological variables, physical diseases, and psychiatric disorders. Complex-Trait Genetics Virtual Lab uses LDSC software [52, 53] with 1000 Genomes Project phase 3 [54] EUR data as LD references. We applied a Bonferroni correction to identify significant correlations.

Lab- and Phenome-Wide Association Studies

We conducted PheWAS in BioVU [55], PMBB [56], and Yale-Penn [57] using the PheWAS v0.12 R package. BioVU, PMBB, and Yale-Penn received approval from their respective IRBs, and participants provided written informed consent. Lab-wide association studies (LabWAS) were performed in BioVU to examine associations with lab results and biomarkers [58]. For eight somatoform-related phecodes, we evaluated the performance of the somatoform factor PGS compared to each univariate GWAS PGS. PGS associations for the univariate and somatoform GWAS were compared using a pairwise Z-test for the equality of regression coefficients (log-odds). Significant Z-values indicated one PGS was a stronger predictor than another. We calculated PGS using PRS-CS software [59], applying the default settings. Models included age, sex, and ten ancestry principal components as covariates, with a Bonferroni correction applied to identify associations. Given that both BioVU and PMBB are EHR datasets, we meta-analyzed their results for each phenotype (n = 1,442) by calculating a weighted average. Cohort details are in the online supplementary materials.

Drug Repurposing

To perform drug repurposing, we used the Library of Integrated Network-Based Cellular Signatures L1000 database, yielding 829 compounds and expression profiles from five neuronal cell lines. We matched medication signatures to gene expression signatures obtained from a somatoform TWAS. Weighted Pearson correlations were calculated between each transcriptome association and the compound signatures [60], with genes weighted by the proportion of heritability explained using the metafor R package (v.3.8–1). Each compound was included as a fixed effect, incorporating the weighted effect size and sampling variability across different times, cell lines, and doses. To account for repeated assessments of the same gene as different transcripts across multiple tissue types, we included brain region as a random effect. Significance was evaluated using a Bonferroni correction. We also used a second drug repurposing method, drug-gene set analysis (DRUGSETS) [61] (online suppl. materials).

MR with Gut Microbiota

To identify potential causal effects of the gut microbiome on somatoform traits, we conducted MR analyses using two GWAS of gut microbiota abundance. The first included 211 taxa measured across various cohorts (n = 14,306) [62]. The second included 207 taxa, measured in a smaller, but more consistently genotyped, sample (n = 7,738) [63]. We extracted instruments at p < 1e−5 and performed clumping using EUR 1000 Genomes to ensure independence. Matching SNPs were extracted from the somatoform GWAS as outcomes. The inverse variance weighted estimate was the primary MR method used. We performed sensitivity analyses to evaluate the robustness of the results (online suppl. materials).

Results

LDSC identified genetic correlations between the traits (online suppl. Fig. 1). However, genetic correlations with headache/migraine were lower than the other traits (rg_average = 0.27 vs. 0.58). A two-factor confirmatory factor analysis model, although an excellent fit to the data (χ²(4) = 8.79, AIC = 30.79, CFI = 1, SRMR = 0.02), yielded highly correlated factors (r = 0.88) and was not retained. In the common factor model, although the loading of headache/migraine met our pre-specified minimum of 0.30, it was substantially lower than the other traits (all >0.65), suggesting that headache/migraine was not adequately represented by the common factor. Thus, we excluded it, which improved model fit (online suppl. materials; Δχ²(3) = 12.31, p = 0.006; fit: χ²(2) = 9.30, p = 0.01, AIC = 25.30, CFI = 1, SRMR = 0.03; Figure 2a).

Fig. 2.

Factor analysis and GWAS results for the somatoform factor. a Confirmatory factor analysis. b Manhattan plot of the somatoform genome-wide association study (GWAS). Lead variants for loci not identified in the input GWAS are annotated. Gold diamonds indicate that the variant was not in a genome-wide significant (GWS) locus in any of the input GWAS, and yellow diamonds indicate that the variant was not in a GWS locus in previous GWAS of pain traits.

GWAS identified 134 associated loci (N_eff = 799,429; online suppl. Table 1), of which 44 were not GWS in any of the input GWAS, and 8 had also not been associated with any pain traits (Fig. 2b). The lead SNPs from the 8 loci were associated with physical, immunological, and mental health measures (online suppl. Table 2; online suppl. Fig. 2). Eight loci exhibited heterogeneous effects across the somatoform traits (online suppl. Table 3).

Gene Mapping and Functional Annotation

MAGMA identified 874 genes based on position, eQTL, and chromatin interactions (online suppl. Fig. 3). One-third (33.98%) were mapped by more than one approach, and 11.21% were mapped by all three. Candidate SNPs were enriched for intronic, intergenic, non-coding RNA intronic, 3′ UTR, and 5′ UTR categories. Many candidate SNPs were likely deleterious (4.59%) or regulatory (59.17%) based on CADD and RegulomeDB scores (online suppl. Tables 4, 5).

Gene-Based Enrichment

Gene-set analyses showed enrichment for genes involved in negative regulation of synaptic transmission (b = 0.71, SE = 0.14, p = 3.24e−07). There was enrichment in 11 of 13 brain tissues (online suppl. Fig. 4), particularly the cerebellar hemisphere (b = 0.05, SE = 0.01, p = 1.43e−12) and cerebellum (b = 0.05, SE = 0.01, p = 2.63e−12). Gene expression was enriched in the early and late mid-prenatal stages. Seven cell types were associated with the somatoform factor (online suppl. Fig. 5). Three (GABAergic neurons in the prefrontal cortex at gestational week 26, GABAergic neurons in the human midbrain, and inhibitory neurons from PsychENCODE adult brain samples) were independently associated, with the others jointly explained by their association with the independent cell types. GABAergic neurons at gestational week 26 and inhibitory adult neurons were collinear, suggesting that their associations are driven by similar genetic signals.

eQTL and pQTL Association Analyses

We identified 28 genes whose expression levels in the brain exerted putatively causal effects on somatoform traits (online suppl. Fig. 6a). Among these were UHRF1BP1, which interacts with a key regulator of DNA methylation and is involved in cell apoptosis, and HLA-DRB1, part of the human leukocyte antigen (HLA) family of genes that initiate immune responses. Using data from the UKB Pharma Proteomics Project, we identified 117 genes whose protein levels were associated with somatoform traits, including members of the CD300 family of genes involved in inflammatory responses (online suppl. Fig. 6b). Additionally, genes involved in neural development and synaptic processes were significant (e.g., HS6ST1 and LRRN1). Using plasma proteomic data from deCODE, we identified 57 genes whose effects on protein abundance were associated with somatoform traits (online suppl. Fig. 6c). Across the two pQTL datasets, 6 genes were consistently identified: CD300A, CD300C, CLEC4G, HS6ST1, LRRN1, and RNASET2.

Transcriptome-Wide Association Studies

We identified 158 genes with altered expression levels associated with somatoform traits (online suppl. Fig. 7a). In a second TWAS in psychiatric cases and controls [64], we identified 131 genes (online suppl. Fig. 7b). Across the two TWAS, 34 genes were consistently identified.

Across the two TWAS and the SMR eQTL analysis, five genes (CCDC144CP, PPP6C, SCAI, UHRF1BP1, USP32P3) were implicated by all three methods. In addition, SEMA3F, ZNF646, HLA-DRB1, KAT8, ILRUN, DCC, C17orf58, CDK2AP1, OLFML2A, SNCA-AS1, and VPS33B-DT were implicated by SMR and one (but not both) of the TWAS. Based on a search of the TWAS Atlas [65], three (SCAI, SEMA3F, and OLFML2A) were identified in previous pain TWAS, and six others were identified by SMR analyses for pain.

Polygenic Overlap with Psychopathology

MiXeR models estimated 11,321 causal SNPs for the somatoform factor (SD = 562.64). Somatoform and EXT shared an estimated 76% of causal variants (SD = 0.18), while 79% of somatoform’s causal variants were shared with INT (SD = 0.12). Estimates were similar for EXT+INT (Dice = 0.83, SD = 0.07; online suppl. Fig. 8).

Genetic Correlations

The somatoform factor was genetically correlated with 646 phenotypes (online suppl. Table 6; online suppl. Fig. 9). One of the top correlations was with “Symptoms, signs, and abnormal clinical and laboratory findings, not elsewhere classified” (r_g = 0.80, SE = 0.02, p = 1.69e−242), which corresponds to ICD-10 R00–R99 codes that capture clinical presentations that do not meet criteria for a specific diagnosis. These codes are commonly used to indicate general symptom burden. Approximately 70 associations (10.84%) were with pain-related conditions and medications. Results also identified genetic correlations with obesity, socioeconomic status, and mental and physical health.

Lab- and Phenome-Wide Association Studies

After meta-analyzing PMBB and BioVU, there were 229 associations (online suppl. Fig. 10). The top associations included obesity (beta = 0.18, SE = 0.01, p = 2.76e−76), tobacco use disorder (TUD; beta = 0.18, SE = 0.01, p = 9.69e−72), and type 2 diabetes (beta = 0.17, SE = 0.01, p = 3.67e−69). There were also associations with pain conditions. Study-specific results are in online supplementary Tables 7 and 8.

The somatoform PGS consistently predicted somatoform-relevant phecodes (Fig. 3), including somatoform disorder (OR = 1.21 [1.08–1.35] p = 7.08e−4), chronic pain (OR = 1.13, [1.10–1.16], p = 2.83e−22), functional digestive disorders (OR = 1.06 [1.03–1.09], p = 1.79e−4), and fatigue (OR = 1.03 [1.02–1.05], p = 6.62e−5). In contrast, the individual GWAS PGS were best at predicting their respective phenotype (e.g., the IBS PGS predicted IBS: OR = 1.17 [1.12–1.22], p = 9.08e−13). After FDR correction for multiple testing, the somatoform PGS was a stronger predictor of chronic pain than the fatigue (Z = 2.89, p_FDR = 0.03), health satisfaction (Z = 3.22, p_FDR = 0.01), and IBS PGS (Z = 4.01, p_FDR = 0.0005). In contrast, the IBS PGS was a stronger predictor of IBS than the somatoform PGS (Z = −3.08, p_FDR = 0.008). For other outcomes, effect sizes were not significantly different.

Fig. 3.

Polygenic score predictive utility for somatoform-relevant traits in the Penn Medicine Biobank and BioVU. Outcomes are defined based on phecodes. PGS, polygenic score; IBS, irritable bowel syndrome.

LabWAS identified 40 biomarker associations (online suppl. Fig. 11), including elevated C reactive protein levels (beta = 0.03, SE = 0.005, p = 4.43e−15), white blood cell counts (beta = 0.05, SE = 0.004, p = 1.58e−37), and hemoglobin A1c levels (beta = 0.03, SE = 0.005, p = 4.43e−15), as well as lower levels of iron (beta = −0.04, SE = 0.01, p = 1.75e−6) and Vitamin D (beta = −0.06, SE = 0.01, p = 5e−19). In the Yale-Penn sample, there were 229 associations (online suppl. Fig. 10b; online suppl. Table 9), including lower education (beta = −0.17, SE = 0.01, p = 1.15e−38), poor self-reported health (beta = −0.18, SE = 0.02, p = 1.29e−25), more emergency room visits (beta = 1.18, SE = 0.15, p = 5.13e−15), and greater childhood adversity (beta = 0.28, SE = 0.04, p = 1.16e−12).

Drug Repurposing

We identified 324 perturbagens (online suppl. Table 10) that included analgesics (e.g., diclofenac and ibuprofen), antidiarrheals (e.g., loperamide), and antidepressants (e.g., bupropion). Two targeted MAP2K1 and reversed the transcriptomic signature found in the PsychENCODE TWAS. Both (pd-0325901 and selumetinib) are kinase inhibitors. Additionally, ten had gene targets that mapped to GWS SNPs (online suppl. Fig. 12).

MR with Gut Microbiota

Adlercreutzia equolifaciens (beta = −0.04, SE = 0.01, p = 5.13e−5) and Ruminococcus bromii (beta = −0.05, SE = 0.01, p = 6.65e−4) exhibited putatively causal protective effects on the somatoform factor (online suppl. Fig. 13). The genus Adlercreutzia (beta = −0.04, SE = 0.01, p = 5.14e−5) was also potentially protective. In the MiBioGen cohort, Adlercreutzia was not associated but did show a consistent direction of effect as that seen in the Dutch Microbiome Project. In MiBioGen, the effects of Ruminococcus were in a matching direction and significant prior to FDR correction (beta = −0.01, SE = 0.006, p = 0.04).

Discussion

Although often considered distinct in clinical practice, our findings support a common genetic factor that contributes to somatoform traits. Consistent with this, a somatoform PGS was broadly associated with somatoform-relevant phecodes. This aligns with the emerging view of PPS as part of a broader spectrum linked by etiological pathways [66]. In an effective sample of 799,429 EUR individuals, we identified 134 loci associated with somatoform traits, including 44 that were not significant in the input GWAS and 8 never associated with PPS. One novel lead SNP, rs3003575, is an intronic variant in DNM1, which encodes dynamin-1. Disrupting DNM1 expression in dorsal root ganglia neurons reversed mechanical and thermal allodynia and suppressed nociceptor sensitization in mice [67], suggesting a role for the gene in pain maintenance. A second novel lead SNP, rs2879086, is an exonic SNP of the protocadherin gene cluster (PCDHA1-4) that is involved in synaptogenesis and neuronal arborization [68, 69]. Other novel lead SNPs mapped to LRBA, HMGA2, ATP11B, RP11-678D18.1, RNU6-408P, and RNU6-1096P. These findings highlight the utility of a multivariate approach to identify novel, mechanistically informative loci for somatoform traits.

Functional annotation indicated enrichment of candidate SNPs in regulatory regions, with most (59.17%) involved in binding and gene expression, providing evidence of the role of gene regulatory mechanisms in the etiology of PPS. We also identified eight loci with heterogeneous effects across the somatoform traits. These loci generally had stronger associations with pain intensity than the other traits, suggesting possible pain-specific mechanisms. Further characterization could help disentangle broader somatoform genetic influences from those specific to pain.

Gene expression was enriched across 11 brain tissue types and the pituitary, with the strongest associations for the cerebellum and cerebellar hemisphere. The cerebellum may modulate the emotional and cognitive elements of pain through communication with subcortical and cortical regions [70]. Individuals with chronic pain show altered activation in the cerebellum during pain [71], and functional connectivity between the cerebellum and other brain regions is correlated with pain intensity ratings [72]. The cerebellum may also be involved in salience processing [70, 73], which refers to the integration of sensory information and is implicated in chronic pain [73–75] and IBS [76]. Cell-type analyses implicated inhibitory GABAergic neurons in the midbrain and prefrontal cortex. Mouse models also support their role in regulating sensory sensitivity [77], and they were implicated in a GWAS of pain intensity [15, 77]. Thus, dysregulation of inhibitory control via GABAergic neurons may contribute to heightened sensitivity to PPS.

We identified potential causal genes using brain transcriptomic data, with 16 genes implicated by both SMR analyses and TWAS. Among these, two (CCDC144CP and USP32P3) are pseudogenes [78], while the others have roles in immune functioning (e.g., HLA-DRB1 and ILRUN), axon guidance (e.g., SEMA3F and DCC), epigenetic and transcriptional regulation (e.g., KAT8 and ZNF646), and cell cycle control (e.g., PPP6C, SCAI, CDK2AP1). Seven of the 16 high-confidence genes were not previously identified in TWAS or SMR analyses of somatoform traits, and of the nine genes with previous associations, all were pain-related [65]; thus, our results provide insights into their broader role in somatoform symptomatology. We also identified six genes whose effects on protein levels mediated their association with somatoform traits, including four involved in immune regulation (CD300A, CD300C, CLEC4G, and RNASET2) and two in neuronal development (HS6ST1 and LRRN1). Preclinical research is needed to examine the mechanisms by which these prioritized genes influence PPS.

We found polygenic overlap between somatoform and INT, EXT, and INT+EXT [49]. Family-based studies have also identified genetic overlap between somatoform and INT [79–81]. For example, a Swedish twin study of chronic widespread pain, chronic fatigue, IBS, and headache identified a latent affective factor that accounted for their shared liability and comorbidity with INT [80]. Their finding aligns with the substantial genetic overlap we observed with INT traits. Genetic correlations further highlighted the shared etiology of somatoform traits with physical and mental health. In three biobanks, somatoform PGS were associated with diabetes, TUD, mood disorders, obesity, post-traumatic stress disorder, and sleep disorders. Collectively, these findings point to shared mechanisms among these disorders, which, if better understood, could help identify broad therapeutic targets.

Addressing shared risk factors for PPS may improve outcomes across health domains. Across independent datasets, R. bromii had the strongest support for potential protective effects on somatoform traits. A keystone bacterial species for their ability to metabolize resistant starch [82], reduced abundance of R. bromii is associated with chronic pancreatitis [83] and Crohn’s disease [84]. We also identified compounds that may have promise for treating PPS, including ten that target genes mapped by GWS variants. Four were dopamine receptor antagonists, including atypical antipsychotics (nemonapride, melperone, and benperidol) and a peripherally active D2 receptor antagonist with antihypertensive properties (carmoxirole) [85]. Two compounds (PD-0325901 and selumetinib) reversed the transcriptomic expression signature of the somatoform factor by targeting MAP2K1. Whereas mitogen-activated protein kinase (MAPK) signaling pathways are involved in pain sensitization and their inhibition reduces pain in animal models [86], our findings suggest they may have applications for treating other PPS as well.

Limitations

This study has several limitations. First, the summary statistics were exclusively from EUR individuals, potentially limiting generalizability to other populations. Second, although genetic correlation and pheWAS analyses underscored complex interactions between physical and mental health conditions, they do not provide information on underlying mechanisms. Third, gSEM captures shared genetic variance, but inherently lacks specificity for individual traits. Thus, the trade-off between power and specificity should be considered when interpreting findings. Fourth, few well-powered GWAS of somatoform traits are available. The included traits reflect theoretical and practical considerations (see online suppl. materials). As a result, the common factor does not represent all components of HiTOP’s somatoform spectrum, and some components are represented by multiple traits. As larger GWAS of somatoform-relevant traits are conducted, further exploration of their shared genetic underpinnings will be warranted and feasible. Compared to other HiTOP domains, the somatoform spectrum remains among the least genetically characterized, and integration with broader psychopathology models, including those that capture thought disorder and detachment spectra, will be essential for evaluating and refining transdiagnostic frameworks.

Conclusions

By identifying a common genetic factor that contributes to somatoform traits, our findings highlight shared biological pathways that link conditions considered distinct in clinical practice. The genetic correlations and substantial polygenic overlap with psychopathology align with dimensional psychopathology models. These findings enhance our understanding of somatoform traits and underscore the need for approaches that address the interplay between physical and mental health.

Acknowledgments

We acknowledge the Penn Medicine BioBank (PMBB) for providing data and thank the patient-participants of Penn Medicine who consented to participate in this research program. We would also like to thank the Penn Medicine BioBank team and Regeneron Genetics Center for providing genetic variant data for analysis. We also thank the Veterans who participated in the Million Veteran Program (MVP) for their contributions to this research.

Statement of Ethics

This study was determined exempt by the University of Pennsylvania Institutional Review Board (protocol #25-0704). This study is based on summary-level data extracted from: UK Biobank (https://www.nealelab.is/uk-biobank/), dbGaP accession phs001672 for Million Veteran Program (https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs001672), and European Bioinformatics Institute GWAS Catalog (https://www.ebi.ac.uk/gwas/) under accession No. GCST90016564.

Conflict of Interest Statement

Dr. Kranzler is a member of advisory boards for Dicerna Pharmaceuticals, Sophrosyne Pharmaceuticals, Enthion Pharmaceuticals, Clearmind Medicine, and Altimmune; a consultant to Sobrera Pharmaceuticals; the recipient of research funding and medication supplies for an investigator-initiated study from Alkermes; and a member of the American Society of Clinical Psychopharmacology’s Alcohol Clinical Trials Initiative, which was supported in the last 3 years by Alkermes, Dicerna, Ethypharm, Lundbeck, Mitsubishi, Otsuka, and Pear Therapeutics; Prof. Joel Gelernter and Dr. Sandra Sanchez-Roige were members of the journal’s Editorial Board at the time of submission. The other authors have nothing to disclose.

Funding Sources

Veterans Integrated Service Network 4, Mental Illness Research, Education and Clinical Center, National Institute on Alcohol Abuse and Alcoholism grants R01 AA030056 and K01 AA028292 (to RLK), and National Human Genome Research Institute grant T32 HG009495 (to KLF). The PMBB is supported by Perelman School of Medicine at University of Pennsylvania, a gift from the Smilow family, and the National Center for Advancing Translational Sciences of the National Institutes of Health under CTSA award number UL1TR001878. Vanderbilt University Medical Center’s BioVU biorepository is supported by institutional funding, private agencies and federal grants, including the NIH-funded S10RR025141 instrumentation award and Clinical and Translational Science Award grants UL1TR002243, UL1TR000445 and UL1RR024975; genomic data are also supported by investigator-led projects that include U01HG004798, R01NS032830, RC2GM092618, P50GM115305, U01HG006378, U19HL065962 and R01HD074711, as well as the additional funding sources listed at https://victr.vanderbilt.edu/pub/biovu/.

Author Contributions

Christal N. Davis conceived the study, conducted the data analyses, visualized the results, and drafted the manuscript. Sylvanus Toikumo contributed to data analysis and study conceptualization. Alexander S. Hatoum contributed to data analysis and conceptualized the drug repurposing analyses. Yousef Khan contributed to data analysis and drafting of the manuscript. Benjamin K. Pham, Shreya R. Pakala, and Kyra L. Feuer contributed to data analysis. Joel Gelernter and Henry R. Kranzler secured funding and contributed to data collection and supervision. Sandra Sanchez-Roige contributed to supervision and data analysis. Rachel L. Kember provided supervision. All authors read and approved the final manuscript.

Funding Statement

Veterans Integrated Service Network 4, Mental Illness Research, Education and Clinical Center, National Institute on Alcohol Abuse and Alcoholism grants R01 AA030056 and K01 AA028292 (to RLK), and National Human Genome Research Institute grant T32 HG009495 (to KLF). The PMBB is supported by Perelman School of Medicine at University of Pennsylvania, a gift from the Smilow family, and the National Center for Advancing Translational Sciences of the National Institutes of Health under CTSA award number UL1TR001878. Vanderbilt University Medical Center’s BioVU biorepository is supported by institutional funding, private agencies and federal grants, including the NIH-funded S10RR025141 instrumentation award and Clinical and Translational Science Award grants UL1TR002243, UL1TR000445 and UL1RR024975; genomic data are also supported by investigator-led projects that include U01HG004798, R01NS032830, RC2GM092618, P50GM115305, U01HG006378, U19HL065962 and R01HD074711, as well as the additional funding sources listed at https://victr.vanderbilt.edu/pub/biovu/.

Data Availability Statement

The summary statistics generated for the somatoform factor are available at the following link: https://upenn.box.com/s/8wiiu0oi0up07dlsaikt9mr7eyhmubf0.