Genetic risk shared across 24 chronic pain conditions: Identification and characterization with genomic structural equation modeling

Abstract

Chronic pain conditions frequently co-occur, suggesting common risks and paths to prevention and treatment. Previous studies have reported genetic correlations among specific groups of pain conditions and reported genetic risk for within-individual multi-site pain counts (≤7). Here, we identified genetic risk for multiple distinct pain disorders across individuals using 24 chronic pain conditions and genomic structural equation modeling (Genomic SEM). First, we ran individual genome-wide association studies (GWASs) on all 24 conditions in the UK Biobank (N ≤ 436,000) and estimated their pairwise genetic correlations. Then we used these correlations to model their genetic factor structure in Genomic SEM, employing both hypothesis- and data-driven exploratory approaches. A complementary network analysis enabled us to visualize these genetic relationships in an unstructured manner. Genomic SEM analysis revealed a general factor explaining most of the shared genetic variance across all pain conditions and a second, more specific factor explaining genetic covariance across musculoskeletal pain conditions. Network analysis revealed a large cluster of conditions and identified arthropathic, back, and neck pain as potential hubs for cross-condition chronic pain. Additionally, we ran GWASs on both factors extracted in Genomic SEM and annotated them functionally. Annotation identified pathways associated with organogenesis, metabolism, transcription, and DNA repair, with overrepresentation of strongly associated genes exclusively in brain tissues. Cross-reference with previous GWASs showed genetic overlap with cognition, mood, and brain structure. These results identify common genetic risks and suggest neurobiological and psychosocial mechanisms that should be targeted to prevent and treat cross-condition chronic pain.

1. Introduction

Chronic pain is a well-documented burden on the patient [1,26,46,79,98,109] and the healthcare system [114]. The costs of pain management are driven by an incomplete understanding of pain chronification mechanisms, which impedes effective prevention and treatment. It is now widely understood that chronic pain is complex and involves changes in brain pathways as well as peripheral mechanisms [94, 102, 113]. This understanding has led to the introduction of chronic primary pain disease codes in version 11 of the International Classification of Diseases (ICD-11) [123]. Nonetheless, with few exceptions [134], pain conditions are still classified largely based on the body site affected and either treated in the primary care setting [110] or referred to specialists by body site (back pain to orthopedists, irritable bowel syndrome to gastroenterologists, etc.). Unfortunately, the processes that drive chronic pain across conditions remain insufficiently understood, and most current treatments do not work for most patients [31,132]. There is an urgent need for a fundamentally different approach.

Recent work in mental health epidemiology has revealed extensive co-occurrence across disorders, leading to identification of common factors underlying multiple conditions [14,54,89]. Co-occurrence across pain conditions with different pathologies (e.g., migraine with irritable bowel syndrome [IBS]) has also been documented [3,67,74,106,116]. Furthermore, chronic pain conditions are heritable (with estimates up to 45% [47]), and genetic risks for pain are shared across conditions [84]. Two recent genome-wide association studies (GWASs) in the UK Biobank (UKBB) identified genetic variants related to multi-site pain [57,60]. These important studies, however, focused on pain widespreadness (quantitative 0–7 [57] or binarized: 1 versus 2 or more affected body sites [60]), without tracking whether genetic risks were shared across conditions. Widespreadness may either arise from co-occurring pain conditions or itself be a pain condition that affects a number of musculoskeletal body sites [20]. Given that five of the seven body sites used in both studies are musculoskeletal, including the most prevalent site, back pain), the genetic associations they assessed are likely driven by the genetic risk for widespread musculoskeletal pain and may not identify genes that cut across distinct disorders.

Here, we examine shared genetic risks across 24 distinct pain conditions in over 400,000 individuals in the UK Biobank (UKBB) [13]. We use Genomic Structural Equation Modeling (SEM) [43] to identify common genetic factors across conditions. Genomic SEM applies the traditional techniques of SEM, a widely used method for latent factor modeling [61], to genetic correlations estimated from genome-wide association statistics. Then we annotate the factors with associated biological pathways.

Figure 1 provides a graphical overview of the study. The specific questions we address are: (1) Is there a general, condition-agnostic genetic risk factor for chronic pain? (2) Are there additional genetic factors underlying subsets of pain conditions? (3) Does the genetic structure correspond to organization of pain by symptom location or hypothesized etiology? (4) What biological pathways and tissues are associated with these genetic factors? Addressing these questions may shed light on trans-diagnostic causes of chronic pain and potential targets for prevention and treatment.

Figure 1.

Scheme of study methods and analyses.

Abbreviations: GWAS, genome-wide association study; LDSC, linkage-disequilibrium score regression; EFA, exploratory factor analysis; CFA, confirmatory factor analysis; Genomic SEM, genomic structural equation modeling; CFI, comparative fit index.

2. Methods

The main steps in the analysis included: (1) GWAS, yielding association effect sizes for genetic variants (single-nucleotide polymorphisms, SNPs) with chronic pain conditions; (2) genetic correlations, which use the GWAS association effect sizes to estimate shared genetic risk across pain conditions; (3) Genomic SEM to identify common factors and compare factor models; (4) factor GWAS to identify SNPs associated with common factors; and (5) functional annotation of factor-associated SNPs in terms of likeliest implicated genes, pathways, and tissues.

2.1. Cohort

Participants in the UKBB were aged 40–69 and were recruited between 2006 and 2010 (UKBB data-request application 16651). The current standard in genetics is to limit analyses to samples of homogeneous ancestral background to avoid introducing confounds from population stratification [119]. We analyzed data from White Europeans, identified using UKBB-provided genomic principal components 1–4 [13], given that no other group had a sufficient sample size for our analyses (see Supplementary Table S2 for descriptive statistics of South Asians, the next highest sample size). Analyses in different ancestral groups will be a high priority when more data become available. Individuals who withdrew from the study by August 2020 were removed. Up to 435,971 people (54% female) were included in the analysis, with sample size varying by phenotype (Table 1).

Table 1:

Pain condition descriptive statistics

Condition	Full name	Cases	Controls	Preva-lence	h2SNP(SE)	Reported h2SNP	Citations
*aCMC	Arthropathy of carpometacarpal joint	1837	201439	0.009	0.09 (0.036)	NA	NA
arth	Arthropathies (non-specific, incl. osteoarthritis)	80737	157458	0.339	0.09 (0.005)	NA	NA
back	Back pain	119216	132641	0.473	0.09 (0.005)	0.11/0.12/0.076	[83, 35, 114]
chDs	Chest pain/discomfort	72156	359415	0.167	0.08 (0.004)
chPh	Chest pain during physical activity	2938	61044	0.046	0.13 (0.032)
*Crhn	Crohn’s Disease	1826	201030	0.009	0.13 (0.038)	0.47	[121]
crpl	Carpal tunnel	11912	424059	0.027	0.16 (0.011)	0.02/0.01	[135, 102]
CWP	Chronic widespread pain	6021	427884	0.014	0.14 (0.014)	0.10	[56]
cyst	Cystitis	15371	189253	0.075	0.03 (0.008)
*dbNr	Diabetic Neuropathy	772	435199	0.002	0.13 (0.051)	0.11	[85]
enLL	Enthesopathies of lower limb	7000	195713	0.035	0.06 (0.014)
enth	Enthesopathies	28754	175077	0.141	0.06 (0.007)
*FM	Fibromyalgia	2149	433822	0.005	0.10 (0.025)	0.14	[25]
gast	Gastritis	41746	179970	0.188	0.07 (0.006)
gout	Gout	15069	192253	0.073	0.20 (0.029)
hdch	Headache	40222	345292	0.104	0.13 (0.008)	0.21	[81]
hipA	Hip arthrosis	17676	193048	0.084	0.14 (0.012)
hipP	Hip pain	41907	381055	0.099	0.08 (0.005)	0.12	[83]
IBS	Irritable bowel syndrome	28419	182876	0.134	0.07 (0.008)
kneA	Knee arthrosis	31267	184763	0.145	0.14 (0.009)
kneP	Knee pain	78507	334812	0.190	0.10 (0.005)	0.08	[82]
legP	Leg pain	41484	108241	0.277	0.10 (0.008)
mgrn	Migraine	21586	189874	0.102	0.12 (0.009)	0.15	[39]
nksh	Neck/Shoulder pain	72952	329192	0.181	0.08 (0.004)	0.11	[84]
oesp	Oesophagitis	13003	195329	0.062	0.06 (0.010)
rhAt	Rheumatoid arthritis	8685	198125	0.042	0.08 (0.014)
**plrh	Polymyalgia rheumatica	2460	433511	0.006	0.09 (0.023)
pnjt	Pain in joint	12016	423955	0.028	0.05 (0.008)
*prst	Prostatitis	3604	199950	0.018	0.06 (0.020)
*seRA	Seropositive rheumatoid arthritis	839	201957	0.004	0.15 (0.064)
stmP	Stomach pain	21417	396116	0.051	0.08 (0.006)	0.14	[83]
**ulcC	Ulcerative colitis	4211	199773	0.021	0.12 (0.022)
*urCl	Urinary colic	4743	198679	0.023	0.06 (0.016)

h²_SNP is SNP heritability, variance in the phenotype explained by variance in genotypes (SNPs). S.E. is standard error. Reported h²_SNP is provided where available.

^*phenotypes that did not have a significant h²_SNP in either the odd or even autosome set.

^**phenotypes that did not load significantly onto either the common or specific factor in the EFA-informed CFA.

2.2. Phenotypes

The selected phenotypes were either chronic pain conditions, such as migraine or back pain lasting longer than three months, or conditions with persistent pain as a prevalent symptom, such as osteoarthritis. We drew an initial list of 91 phenotypes from five UKBB Categories: Medical conditions (100074), Health outcomes (713), Self-reported medical conditions (1003), Health and medical history (100036), and First occurrences (1712), downloaded in May 2021. We were able to greatly expand the number of pain conditions included in this study compared to previous studies [57,58,60,122], because of the First Occurrences dataset. This dataset gave researchers access to primary care and death register records to supplement self-reports and ICD-10 diagnoses that had been earlier available exclusively from hospital intake records. This updated UKKB dataset thus had, for the first time, more accurate case prevalences for a large number of conditions. We recoded these conditions into binary phenotypes (Supplementary Table S1) and pruned them to remove those that fell into one of the following categories: 1. heterogeneous disorders or groups of other conditions already included, such as ”Other diabetic polyneuropathies”; 2. branching traits (answers to questions dependent on endorsement of a previous question, with the exception of DF6159: ”Pain type(s) experienced in last month”, which was included as the branching question for pain experienced for more than three months); 3. disorders with case count < 500; 4. disorders that were not sufficiently related to genetics, with SNP heritability (see Section 2.2.1 below) less than or equal to 2 standard errors above zero (h²_SNP - 2 ∗ SE <= 0), (see Supplementary Note). Genomic SEM models genetic covariances, and traits with low heritability cannot show significant genetic covariance with other traits; hence they were not included here. This pruning left 33 heritable chronic pain conditions (Table 1), which were further reduced to 24 during factor analysis (see Section 2.2.2 below).

2.3. GWAS

To prepare the data for GWAS, we used Plink [17], a commonly used, open-source toolset for genetic analyses (details in Supplementary Note). For the next step, GWAS, we estimated associations between each SNP and each chronic pain condition (phenotype) of interest using logistic regression, with the SNP variant as the predictor and the pain condition as the outcome. For these analyses, we used Regenie (details in Supplementary Note).

2.4. Factor analysis and structural equation model

2.4.1. Heritability and genetic correlations

We used linkage disequilibrium score regression (LDSC) software [12] to determine SNP heritability, h²_SNP, and to estimate genetic correlations between every pair of pain conditions. For heritability, LDSC uses the pattern of SNP effects to estimate the variance in the phenotype that is attributable to all measured SNPs in aggregate (details in Supplementary Note). Table 1 reports the h²_SNP for each condition, on a liability scale (Supplementary Note). For genetic correlations, LDSC determines the extent to which the pattern of genetic associations for one phenotype is correlated with the pattern of genetic associations for another phenotype. The resulting matrix of pairwise genetic correlations across pairs of chronic pain conditions is in Figure 2A. These genetic correlations provide the basis for factor analysis in Genomic SEM.

Figure 2.

Genetic correlations, network, and Genomic SEM model.

Abbreviations: hdch, headache; mgrn, migraine; cyst, cystitis; back, back pain; nksh, neck/shoulder pain; hipP, hip pain; CWP, chronic widespread pain; pnjt, joint pain; gast, gastritis; oesp, oesphagitis; chDs, chest pain/discomfort; chPh, chest pain during physical activity; IBS, inflammatory bowel syndrome; stmP, stomach pain; hipA, hip arthrosis; crpl, carpal tunnel; enth, enthesopathies; rhAt, rheumatoid arthritis; arth, arthroses; kneA, knee arthrosis; kneP, knee pain; legP, leg pain; enLL, enthesopathies in lower limbs. (a) Genetic correlations for 24 pain conditions estimated using linkage disequilibrium score regression (LDSC) implemented in Genomic SEM. (b) Network of genetic correlations for 24 pain conditions, pruned for significance at FDR 0.01. The 19 conditions in yellow form a clique, complete subgraph. The 3 conditions in blue have the highest betweenness centrality, shortest path between 2 other nodes. Node size corresponds to strength, magnitude-weighted number of connections with other nodes. (c) EFA-CFA model for 24 pain conditions with residual covariances (~~) estimated for same body-site conditions (table in top left): hip arthrosis and pain; knee arthrosis and pain; headache and migraine; chest pain at baseline and during physical activity. F1 is the general factor with positive loadings from all conditions, and F2 is the musculoskeletal factor. CFI, comparative fit index; SRMR, standardized root mean squared residual. All loadings shown are significant at α=0.05. (d) Summary scores (overall measure of interconnectedness for each pain condition) obtained using F1 loadings from EFA-CFA and network strength and betweenness centrality, vector-normalized geometric means (y-axis). (More information on all conditions in Supplementary Table 1 (https://docs.google.com/spreadsheets/d/1S-vFvnwkD5iCP16La_iyjRDTxIqBoMRqAOV6SXpKhN8/edit#gid=0).

2.4.2. Factor analysis and Genomic SEM

We used Genomic SEM to test for evidence of shared genetic risk. SEM analyses test whether a hypothesized factor structure can adequately capture the observed correlations across a set of observed phenotypes (indicators). Genomic SEM applies the same method but to genetic correlations instead of phenotypic correlations; thus, the indicators are chronic pain conditions, but the data constitute correlations of genetic effects on phenotypes rather than phenotypic scores themselves. In confirmatory factor analysis (CFA), one specifies a series of hypothesis-driven models and compares their ability to accurately reproduce the pattern of correlations across conditions. Each model comprises a specification of which conditions load on which factors and whether the factors are themselves correlated. In exploratory factor analysis (EFA), one allows for the factor loadings to be determined in a data-driven fashion, identifying which groupings of conditions are best supported.

Our main analysis goal was to test for common factors without rigidly specifying groupings a priori. Thus, our main strategy combined EFA and CFA, using EFA to select the number of factors and their loadings and then evaluate the goodness of fit and compare with alternative models (see below for details). EFA-CFA, evaluated in [38] and recently used in [19,24,65], is a partially data-driven approach that captures observed groupings in the data while still permitting structure and inferences based on theory.

CFA can also be used to compare alternative theories about which types of common factors best explain observed correlations across chronic pain conditions. Here, we compared our EFA-guided model with two hypothesis-based CFA models: One based on shared genetic risks across conditions with similar body sites (‘anatomic’) and one based on shared risk for inflammatory conditions (‘etiologic’). The anatomic model included one general factor (on which all disorders load) and 6 specific factors that group conditions based on body site: Cranial, Gastrointestinal, Joint, Leg/Foot, Pelvic, and Torso. The etiologic model included the general factor and a specific factor for inflammatory conditions, which was the only putative etiology with a substantial number of representative conditions. We discussed a variety of other groupings, but as biological etiology is often unknown – a central problem in pain research – we did not reach clear consensus on additional etiological factors.

For all three models, the goal was to test a bifactor structure, which consisted of a general factor with loadings for all conditions and specific factors that are orthogonal to the general factor and have loadings for specific subsets of conditions. This type of model allowed us to test whether a common genetic factor underlay all tested pain conditions, while still allowing for shared variance for particular groups of conditions. Similar approaches have been used to model other multidimensional constructs, including personality [18] and psychopathology [9].

For the EFA portion of the EFA-CFA approach, we used the fa function in the ‘psych’ R package; a scree plot (Supplementary Figure S1) suggested three factors, allowing for correlations among factors using oblique rotation. The factor loadings were thresholded to define a CFA model for subsequent validation in Genomic SEM. The final model structure is shown in Figure 2C.

The use of EFA to guide model development requires validating the model’s fit on independent data, to avoid overfitting [33]. For this validation, we used a split-genome approach [34,42]. We developed a model using the EFA-CFA procedure described above in odd autosomes (1,3,...21), and assessed the fit of the final CFA in even autosomes (2,4,...22). This split created two independent sets, given that SNPs are not correlated across chromosomes (Mendel’s law of independent assortment).

Using an odd-even autosome split assumes that traits are polygenic (many genes contribute) and that relevant genes are distributed across autosomes, so that an estimate of genetic correlations from odd autosomes can be replicated in even ones. The polygenic nature of pain conditions, like other complex traits, was evident in this dataset (Supplementary Figure S2). However, to be conservative, we excluded conditions whose genetic associations were not heritable in both odd and even autosomes (Supplementary Note). This validation step led to a further exclusion of 7 conditions (arthropathy of carpometacarpal joint, diabetic neuropathy, Crohn’s disease, fibromyalgia, prostatitis, seropositive rheumatoid arthritis, and urinary colitis), whose heritability estimates were not significantly above 0 in at least one holdout set (Table 2). This final exclusion left 24 pain conditions for the validation step, which we also used in the main analysis and in the two hypothesis-driven approaches for consistency and comparability.

Table 2:

SNP heritability in whole genome, odd and even chromosomes

Condition	Whole genome h2SNP (SE)	Odd chroms. h2SNP (SE)	Even chroms. h2SNP (SE)
*aCMC	0.09 (0.036)	0.09 (0.026)	0.00 (0.023)
arth	0.09 (0.005)	0.05 (0.004)	0.04 (0.004)
back	0.09 (0.005)	0.04 (0.003)	0.05 (0.003)
chDs	0.08 (0.004)	0.04 (0.003)	0.04 (0.002)
chPh	0.13 (0.032)	0.09 (0.025)	0.05 (0.023)
*Crhn	0.13 (0.038)	0.08 (0.029)	0.05 (0.025)
crpl	0.16 (0.011)	0.07 (0.008)	0.08 (0.009)
cyst	0.03 (0.008)	0.01 (0.006)	0.02 (0.006)
*dbNr	0.13 (0.051)	0.02 (0.033)	0.11 (0.038)
enLL	0.06 (0.014)	0.03 (0.010)	0.03 (0.009)
enth	0.06 (0.007)	0.03 (0.005)	0.03 (0.005)
*FM	0.10 (0.025)	0.06 (0.017)	0.04 (0.019)
gast	0.07 (0.006)	0.03 (0.004)	0.04 (0.004)
CWP	0.14 (0.014)	0.07 (0.010)	0.07 (0.010)
gout	0.20 (0.029)	0.08 (0.013)	0.12 (0.024)
hdch	0.13 (0.008)	0.06 (0.004)	0.07 (0.007)
hipA	0.14 (0.012)	0.07 (0.009)	0.07 (0.008)
hipP	0.08 (0.005)	0.04 (0.003)	0.04 (0.003)
IBS	0.07 (0.008)	0.04 (0.005)	0.03 (0.004)
kneA	0.14 (0.009)	0.07 (0.005)	0.08 (0.007)
kneP	0.10 (0.005)	0.05 (0.003)	0.05 (0.003)
legP	0.10 (0.008)	0.05 (0.005)	0.05 (0.005)
mgrn	0.12 (0.009)	0.06 (0.006)	0.06 (0.007)
neck	0.08 (0.004)	0.04 (0.003)	0.04 (0.003)
oesp	0.06 (0.010)	0.03 (0.007)	0.03 (0.006)
rhAt	0.08 (0.014)	0.05 (0.010)	0.04 (0.010)
**plrh	0.09 (0.023)	0.05 (0.016)	0.03 (0.015)
pnjt	0.05 (0.008)	0.03 (0.006)	0.02 (0.005)
*prst	0.06 (0.020)	0.03 (0.014)	0.04 (0.013)
*seRA	0.15 (0.064)	0.17 (0.043)	−0.02 (0.047)
stmP	0.08 (0.006)	0.04 (0.004)	0.04 (0.005)
**ulcC	0.12 (0.022)	0.07 (0.016)	0.05 (0.015)
*urCl	0.06 (0.016)	0.02 (0.012)	0.03 (0.012)

h²_SNP is SNP heritability, variance in the phenotype explained by variance in genotypes (SNPs). S.E. is standard error.

^*phenotypes that did not have a significant h²_SNP in either the odd or even autosome set.

^**phenotypes that did not load significantly onto either the common or specific factor in the EFA-informed CFA. Condition definitions are in 1, and details are in Supplementary Table S1.

Models were evaluated using CFI (comparative fit index), which quantifies the extent to which the model fits better than a baseline model (one in which the variables are uncorrelated), and SRMR (standardized root mean residual), which quantifies the mean absolute difference between the observed correlations and the correlations predicted by the model [50]. A well-fitting model should generally have a CFI ≥ .95 and an SRMR ≤ .08 [50]. The models were additionally compared using AIC, (Akaike information criterion), which is a goodness-of-fit index favoring more parsimonious models (lower values indicate better fit) [61].

2.5. Factor GWAS

Common factors may have associations with genetic variants that are not detectable in analysis with individual conditions. To identify the SNPs with the largest associations with common chronic pain factors, we ran a factor GWAS in Genomic SEM (userGWAS function) (Supplementary Figure S3).

One issue to address when calculating GWAS on common factors is whether SNPs are associated primarily with the common factor or instead are more strongly associated with individual conditions that contribute to it. To assess each SNP for disproportionately strong or directionally opposing effects on a subset of conditions, we conducted a heterogeneity Q test [43,52] and discarded SNPs with heterogeneous effects as well as those correlated with them (Supplementary Note).

2.5.1. GWAS annotation

GWAS yields a list of SNPs significantly associated with a trait. Follow-up analyses are then needed to characterize those SNPs in terms of genes and pathways. Functional Mapping and Annotation (FUMA) of GWAS [137] is a platform developed to facilitate a number of standard GWAS follow-up analyses. To functionally characterize the genetic contributors to both individual phenotypes and the two factors, we submitted all GWAS results to FUMA for gene prioritization and functional annotation, using several integrated databases [137]. These analyses consisted of: 1. prioritizing SNPs based on their effect sizes and independence from each other; 2. mapping significant SNPs to genes as described below; 3. conducting a genome-wide gene-based association analysis using FUMA-implemented MAGMA (https://ctg.cncr.nl/software/magma) for gene analysis and gene property analyses; 4. gene set analysis for enrichment in known biological pathways; 5. gene property analysis, or testing for preferential expression of associated genes with 53 Gene-Tissue Expression repository (GTEX), version 8, tissues. We used FUMA’s default and standard significance thresholds and parameters, including p < 5×10⁻⁸ for lead SNPs (independent at r² < 0.1); p < 0.05 for all other SNPs; r² threshold for independent significant SNPs used for further annotations, including gene mapping: 0.6; reference panel population = UKB release 2b 10K European; minimum minor allele frequency = 0.01; maximum distance between LD blocks to merge into a locus = 250 kilobases. The r² threshold represents a squared pairwise correlation for SNP variant alleles. The sample sizes for the two factors (general and musculoskeletal) identified in the final EFA-CFA model were 422,752 and 468,929, respectively, calculated using the method described in [76]. Variants from the reference panel that were in LD with GWAS lead SNPs were included to increase the chance of including causal variants.

Mappings of independent significant (as defined in FUMA, p < 5×10⁻⁸ and r² < 0.6) SNPs onto genes was based on (1) positional distance (within 10 kilobases of gene start and stop coordinates); (2) statistical associations with transcription levels (expression quantitative trait locus, eQTL); and (3) chromatin interaction mapping, physical interactions with gene chromatin states (indicative of transcriptional accessibility). Only protein-coding genes were included, and the major histocompatibility (MHC) region was excluded from annotation. MAGMA analysis for gene-based (versus SNP-based) associations [23] was conducted with SNP assignment within windows of 10 kilobases of gene start and stop coordinates, and GTEx, version 8, [73] was used for gene expression analysis in 53 tissues. Given that the highest association statistic is not necessarily correlated with its relative importance (see this recent publication discussing negative selection as a mechanism for purging high-effect variants in critical gene loci [95]), our approach was to prioritize genes based on: 1. an a priori association p cut-off to ensure statistical rigor; and 2. convergent lines of evidence for functional importance, i.e. overlap in the three mapping approaches. In the resulting set, we interpret our findings in their entirety, without deference to the top association.

The gene-tissue expression analysis tested for association between highly expressed genes in 53 GTEx tissues and GWAS effect sizes for the same genes, which tests the relationship between the genes highly associated with the pain factors and highly expressed in different tissues (details in Supplementary Note). The parameters are summarized in Supplementary Table S3.

We used REVIGO [113] to assign, prune, and summarize biological pathways to the 25 genes with overlapping mappings (details in Supplementary Note).

2.6. Network analysis

While CFA has many strengths in permitting model comparison, some groups have emphasized that relationships among clinical conditions can have a complex causal structure that can be characterized in terms of networks of interacting variables [131]. We made no strong claims about the underlying causal structure and complemented the factor-analytic models with a network-based approach to characterize genetic relationships among conditions in terms of multiple local causes instead of a few latent causes. Network characterization and visualization was done in igraph in R [22]. Genetic correlations of the final 24 pain conditions were filtered for positive significant correlations, using a threshold of 0.01 false discovery rate (FDR)-corrected, calculated with fdrtool in R. We calculated two graph theoretic properties for each pain condition: (1) strength, calculated as the number of edges (genetic correlations with other pain conditions) weighted by their magnitude [7]; and (2) betweenness-centrality, the number of shortest paths between pairs of pain conditions that go through the pain condition in question) [11]. Strength identifies ‘hub’ conditions that are robustly genetically related to many other conditions and may thus be prominent indicators of multi-disorder susceptibility. Betweenness-centrality identifies ‘connector hubs’, conditions that are genetically related to multiple other conditions that are themselves less interrelated. ‘Connector hubs’ are thus key indicators of shared genetic vulnerability. These measures may themselves be correlated, and if so, combined into an overall index, as we did here (described below). At the network level, we estimated the largest clique, complete subgraph of intercorrelated pain conditions [29], which identifies a group of genetically interrelated conditions that may together serve as indicators of multi-disorder susceptibility.

2.7. Summary score

To summarize the evidence for which conditions are the most consistent key indicators of multi-disorder vulnerability, we combined results from Genomic SEM and network analysis, obtaining an overall measure of interconnectedness for each pain condition. Thus, we derived summary scores for all pain conditions using general factor loadings from EFA-CFA, network strength, and betweenness centrality, which are intercorrelated, r =.935 (general factor and strength), r =.614 (general factor and betweenness), r =.693 (strength and betweenness). We calculated a geometric mean of these 3 measures, after vector-normalizing them using the norm function in R.

3. Results

The work reported here is part of a project pre-registered on Open Science Foundation, OSF (Identifying and characterizing genetic susceptibility and its overlap with psychosocial traits, https://osf.io/4p5e3).

3.1. Univariate pain condition GWAS curation and annotation

We considered 91 potentially relevant pain phenotypes in the UK Biobank and selected 24 that (a) were indicative of chronic pain conditions, (b) had sufficient case counts (>500), and (c) were significantly heritable (see Methods; Table 1 and Supplementary Table S1). The sample size available for case assessment varied by condition and ranged from 63,982 (chest pain during physical activity) to 435,971 (several conditions). Prevalence ranged from 0.002 (772 cases, diabetic neuropathy) to 0.473 (119,216 cases, back pain). SNP heritability (variance in the phenotype explained by variance in the genotype) ranged from 0.03(SE, 0.008) for cystitis to 0.20(SE, 0.029) for gout.

Summaries of results from univariate GWAS are reported in Table 1 (SNP heritabilities), in Supplementary Figures S2 and S4 for Manhattan and quantile-quantile (QQ)-plots, and in Supplementary Table S14 for numbers of significant SNPs and genes.

3.2. Pain condition genetic correlations

Pairwise genetic correlations for the 24 pain conditions, Figure 2A, show a large cluster of interconnected vertices. This main cluster includes etiologically and anatomically diverse conditions, such as back pain, oesophagitis, irritable bowel syndrome (IBS), and carpal tunnel, suggesting shared genetic susceptibility among these disparate syndromes. Headache and migraine form a tight mini-cluster (top left), and cystitis, hip arthrosis, enthesopathies of the lower limb and gout show weaker correlations, suggesting more specific genetic risks for each of these four conditions.

A natural question is whether genetic correlations are potentially inflated when estimated in individuals with comorbid conditions. The answer is no. Conditions are primarily comorbid due to shared genetic risks or shared environmental risks. However, conditions with shared environmental risks would not correlate genetically, because alleles of different genes segregate independently from each other (Mendel’s law of independent assortment). Thus, given distinct genetic profiles for two conditions, the risk alleles for one condition would not correlate with the risk alleles for another. A similar genetic profile in many individuals with comorbid conditions indicates a true common predisposition.

3.3. Structural equation modeling

Using three approaches – hypothesis-driven anatomic (1) and etiologic (2), and largely data-driven exploratory-then-confirmatory (3) factor analyses (EFA-CFA) – we fit a bifactor model to test the loadings of all conditions onto a general factor, with differences in specific factor groupings in each approach. The anatomic model based on body site (Supplementary Figure S5, CFI= .875 and SRMR = .087) and the etiologic model, based on a grouping of inflammatory disorders (Supplementary Figure S6, CFI= .905, SRMR= .095) both had suboptimal fit (CFI≤ 0.95 and SRMR≥ .08), see Methods. The EFA-CFA model, shown in Figure 2C, produced an adequate overall fit (CFI= 0.956, SRMR = 0.075).

All pain conditions loaded positively and significantly onto the general factor. The specific factor had substantial positive loadings for arthropathies (which include osteoarthritis), carpal tunnel, enthesopathies of lower limb, other enthesopathies, hip arthrosis, hip pain, knee arthrosis, knee pain, leg pain, pain in joint, and rheumatoid arthritis. Given the pronounced musculoskeletal component among these indicators, we interpreted the specific factor as musculoskeletal. This factor is in line with the World Health Organization’s grouping of pain diseases of the musculoskeletal system, which groups conditions that affect joints, bones, muscles, the spine, and multiple body areas or systems [135]. The EFA-CFA was superior (AIC=4849.164) to both the anatomic (AIC=13184.43) and the etiologic (AIC=10024.93) models. In addition, the latter models had non-significant loadings on their specific factors (Leg/Foot, Pelvic, and Torso for the anatomic, Supplementary Figure S5, and Inflammatory for the etiologic, Supplementary Figure S6, suggesting that shared variance for those indicators was mainly explained by the general factor (details in the Supplementary Note). We validated this model by using the same approach, EFA on odd (CFI= .884 and SRMR= .123) and CFA testing on even (CFI= 0.903 and SRMR= .129) autosomes (details in the Supplementary Note). These comparable metrics in the odd/even and whole-genome datasets suggest that using EFA and CFA on the same dataset did not result in substantial overfitting.

3.4. Network analysis and central conditions

Network analysis provided additional evidence for substantial genetic overlap across pain conditions with a different theoretical model. Graph-theoretical properties of the network (Figure 2B) indicate shared genetic susceptibility, and node size corresponds to strength (magnitude-weighted number of connections). There is a complete subgraph of 19 interconnected conditions, highlighted in yellow: arthropathies, back pain, neck/shoulder pain, hip pain, knee pain, leg pain, chest pain (baseline and during physical activity), rheumatoid arthritis, knee arthrosis, joint pain, carpal tunnel, enthesopathies, widespread pain, gastritis, oesophagitis, stomach pain, headache, and IBS. Consistent with the CFA model, these conditions affect diverse body sites and span inflammatory and non-inflammatory as well as musculoskeletal and non-musculoskeletal forms of pain. Gout, hip arthrosis, enthesopathies of the lower limb, cystitis, and migraine lie outside the large cluster, but they still have more than 10 connections each. Overall, the network reveals a large core of pain syndromes with shared genetic vulnerability.

Some conditions were particularly central in the network, in several ways. Arthropathies, back, and neck/shoulder pain had the highest betweenness centrality (the highest number of shortest paths between node pairs that go through the index node), indicating that genetic associations between many conditions share genetic vulnerability with at least one of these three.

The summary score derived from the general factor, network node strength, and betweenness centrality, Figure 2D, reflects the highest degree of genetic overlap with other conditions. Once again, the top highest scorers were neck/shoulder pain, back pain, and arthropathies.

3.5. Factor GWAS and annotation

After running factor GWASs, we excluded QSNPs, which showed evidence of effects specific to certain pain conditions (not through the common factors), and conducted functional annotation of the GWAS output for each of these factors.

3.5.1. General Factor

The general factor GWAS yielded 33 genome-wide independent significant SNPs, Supplementary Table S4, Figure 3. FUMA mapped these to a total of 241 genes, using at least 1 of 3 methods (positional, eQTL, and chromatin interactions, see Methods), Supplementary Table S5: 26 by positional, 52 by eQTL, and 57 by chromatin interaction mappings. All 3 annotations were identified for 25 genes, highlighted in green in Supplementary Table S5.

Figure 3.

F1 factor GWAS output.

Genome-wide association study (GWAS) results for general pain factor (F1). SNP Manhattan (a) and quantile-quantile, QQ, (b) plots for F1 GWAS. (c) Gene-based genome-wide association Manhattan plot, with the top 31 associated genes labelled. Full gene names taken from the NCBI gene database, https://www.ncbi.nlm.nih.gov/gene/: ARPC5L, Actin-Related Protein 2/3 Complex Subunit 5-Like Protein; BSN, Bassoon Presynaptic Cytomatrix Protein; C6orf106, inflammation and lipid regulator with UBA-like and NBR1-like domains; CAMKV, CaM Kinase Like Vesicle Associated; CDHR4, Cadherin-Related Family Member 4; CTD-2330K9.3, Coats disease; DCC, Colorectal cancer suppressor; ERBB3, Erb-B2 Receptor Tyrosine Kinase 3; FOXP2, Forkhead Box P2; GNAT1, G Protein Subunit Alpha Transducin 1; IP6K1, Inositol hexakisphosphate kinase 1; IP6K3, Inositol hexakisphosphate kinase 3; LANCL1, LanC Like Glutathione S-Transferase 1; MAML3, Mastermind Like Transcriptional Coactivator 3; MARVELD3, MARVEL (Membrane-Associating) Domain Containing 3; MON1A, MON1 homolog A, secretory trafficking associated; MST1, macrophage stimulating 1; MST1R, macrophage stimulating 1 receptor; NCAM1, neural cell adhesion molecule 1; RBM5, RNA binding motif protein 5; RBM6, RNA binding motif protein 6; RERG, RAS like estrogen regulated growth inhibitor; RNF123, ring finger protein 123; ROBO2, roundabout guidance receptor 2; SAMD5, sterile alpha motif domain containing 5; SCAI, suppressor of cancer cell invasion; SEMA3F, semaphorin 3F; TRAIP, TRAF interacting protein; UBA7, ubiquitin like modifier activating enzyme 7; UQCC2, ubiquinol-cytochrome c reductase complex assembly factor 2; WDR38, WD repeat domain 38. (d) Gene property analysis for association between factor GWAS gene effects and gene expression levels in 53 specific tissues from GTEx, version 8.

REVIGO pathway analysis suggested that the pathways represented by these genes cover a broad range of biological processes, including organ development (gut, heart, muscle and brain), metabolism, catabolism, signaling, immunity, neuronal development, transcription, and DNA repair (Supplementary Table S6). FUMA gene set annotation showed a significant enrichment for a pathway involved in learned vocalization behavior or vocal learning (p = 8.93×10⁻⁰⁷, Bonferroni-corrected p = .0138). Additionally, this analysis showed a trend towards significance in several other pathways involved in mechanosensory behavior, several neuronal development processes, and several biosynthesis and calcium channel regulation processes (Supplementary Table S7). Although these pathways did not reach corrected significance, we note them, because they are supported by previous findings [58, 60, 83, 117] and may be useful for hypothesis-generation purposes. We note, additionally, that biological pathways have roles in multiple functions, and our results do not imply a direct link between pain and the functions, such as vocal learning, associated with these pathways.

MAGMA-based tissue expression analysis, as implemented in FUMA, tested for association between highly expressed genes in 53 GTEx tissues and GWAS effect sizes for the same genes (Supplementary Note). Associations were significant only in brain tissues: cortical regions (the cerebral cortex, dorsomedial prefrontal cortex BA9, and anterior cingulate cortex BA24), nucleus accumbens, basal ganglia, amygdala, hippocampus, hypothalamus, and cerebellum, Figure 3D.

Additionally we used FUMA to cross-reference SNPs and genes with other GWAS reports. Of note is the overlap in SNPs (Supplementary Table S8), and significant enrichment for genes reported to be associated with chronic pain conditions (back pain, Crohn’s disease, IBS, and multi-site chronic pain), brain structural traits, anthropometric traits, cognition and intelligence-related phenotypes, sleep-related phenotypes, neuroticism, and mood phenotypes (Supplementary Figure S7). Genetic overlap with non-pain conditions is suggestive of the complexity of factors contributing to chronic pain and suggest potential pathways of susceptibility to pain chronification. Furthermore, DCC, the top gene associated with the general factor, is also the top gene reported in a recent study of chronic overlapping pain conditions, which used pain for more than 3 months in different body sites from the UKBB (head, face, neck/shoulder, back, stomach, hip, knee, all over the body) [60]. Of the 241 genes mapped to independent significant SNPs from the general factor GWAS, FKBP5 is the only one previously targeted in a candidate gene study (as opposed to GWAS) for posttraumatic musculoskeletal pain [10, 71, 148].

3.5.2. Musculoskeletal Factor

The musculoskeletal factor GWAS yielded 7 genome-wide significant lead SNPs (Supplementary Table S9 and Supplementary Figure S8). Positional mapping yielded 5 unique genes; eQTL mapping yielded 18 genes; and chromatin interaction mapping yielded 19 genes, with 5 genes mapped using all 3 methods, green: DPYD, MAPK6, GLIS3, COL27A1, and SLC44A2 (Supplementary Table S10).

REVIGO pathway analysis showed associations with genes involved in bone and neuronal development, cell cycle, transcription regulation and signal transduction, Supplementary Table S11. Gene set annotation showed a Bonferroni-corrected significant enrichment for regulation of RNA biosynthetic process and nominally significant (p < 0.05) enrichment for several other regulatory processes, chromatin organization, cell migration involved in heart development, and DNA damage response (Supplementary Table S12). MAGMA tissue expression analysis found no significant association between gene expression and GWAS effect sizes for 53 tissues (Supplementary Figure S8D).

Cross-referencing with other GWAS reports identified previously reported SNP associations with anthropometric traits (height, hip circumference, offspring birth weight), hip or knee osteoarthritis, sleep-related phenotypes, and type 2 diabetes (Supplementary Table S13), and significant overlap with genes reported to be associated with inflammatory skin disease, palmitic and stearic acid levels (Supplementary Figure S9). None of the genes previously targeted in candidate gene studies for pain [146] mapped to independent significant SNPs for the musculoskeletal factor.

4. Discussion

We ran Genomic SEM on 24 pain conditions in the UKBB to examine the structure of their shared genetic risk and characterize the genetic variants common to them. Our results identify a general factor that explains substantial genetic variance in pain conditions with different suspected etiologies and anatomic presentations and points to their shared systemic pathophysiology. Additionally, a second factor explains some of the shared genetic variance across musculoskeletal conditions. The two-factor model explains the pattern of genetic associations among disorders better than either the anatomic or etiologic grouping of known inflammatory disorders. The shared genetic burden is also apparent in our network analyses.

We identified 184 novel targets for cross-condition chronic pain (Supplementary Tables S5). Two lines of evidence suggest that genes associated with our general pain factor have an important role in the central nervous system (CNS). First, the genes most associated with the general factor for pain also have the highest expression in brain tissues. Second, the pathways of these genes include CNS development and maintenance. Our study further adds to existing evidence for the role of DCC [60, 104, 118], an axonal guidance mediator [144], in chronic pain. Beyond CNS, the pathways of genes associated with the general pain factor also implicate a broad range of other functions, such as gut development, locomotion, and protein secretion, suggesting that susceptibility to chronic pain may involve other systemic biological changes. The new molecular targets we identify can be cross-referenced with animal models in ‘reverse-translation’ approaches to better understand the pathophysiology of chronic pain and develop novel treatments.

We note overlap between the genetic variants associated with our general factor and those previously reported in GWAS for cognitive, structural, mood, sleep, and personality traits, regulation of inflammation and neuroplasticity, and psychiatric disorders. This overlap underscores the highly multifaceted nature of pain as a biopsychosocial condition, while elucidating the key genes and systems involved [15, 20, 21, 77, 128, 129]. This pleiotropy, or the association of genes with multiple conditions, together with the polygenic nature of the general factor we identified, exemplifies the frequently observed many-to-many mapping between genes and traits [136]. Identifying links between polygenic risk profiles of different disorders can provide important information on susceptibility and treatment.

As might be expected, the genes associated with the musculoskeletal factor are fewer, and their pathways are less diverse. They implicate skeletal development, choline transport, signalling, and transcription machinery. Notably, they do not implicate the nervous system. Overlap with previous GWAS results suggests involvement of variants affecting anthropometric traits and thereby body-structural mechanisms. Similar associations have been shown for musculoskeletal pain conditions before: genetic overlap in osteoarthritis with height and BMI [28], back pain [35] and multi-site musculoskeletal pain [124] with structural-anatomic genes.

Our work builds on earlier genetic analyses of combinations of pain conditions selected based on anatomic proximity or hypothesized etiology [45,75,99,133,140]. While most studies have been conducted in twins, several large-scale chronic pain GWAS, with strengths complementary to twin studies [36], have been published on pain in the past three years [35, 82, 83, 85, 86, 87, 117]. These reports, which used earlier releases of the UKBB prior to primary care data availability, include three on multi-site pain [57, 58, 60]. Most recently, another study reported a polygenic risk score for pain spreading that was associated with a phenotypic profile dominated by mood, sleep, and neuroticism [118].

Studies of multi-site pain differ from our study in a critical way. They may show genes associated with only one of the constituent conditions to be associated with the average count of pain sites (widespreadness), if that condition is present in several combinations of comorbid pain sites. Thus, given a genetic variant associated with hip pain and two three-site combinations -- hip, knee, back pain and hip, neck, stomach pain -- this variant will appear to be associated with a widespreadness score of 3 without having associations with any of the other constituent conditions in this example. Analyses of association with multi-site pain are thus not designed to identify the genetic variants shared among multiple distinct conditions. On the other hand, Genomic SEM is designed to identify genetic variants that are truly common to different conditions, and it enabled us to capture genetic risk for chronic pain, regardless of etiology or symptomatology.

The existence of cross-condition genetic risk factors challenges the current clinical practice of grouping and treating chronic pain conditions based on location of symptoms on the body or suspected etiology [30]. Evidence for central processes beyond local pathophysiology has been accumulating. Biopsychosocial factors [20, 21, 77, 128, 129], neuroinflammation and neuroplasticity [5, 6, 27, 37, 53, 88, 92, 97, 125, 147], and neuroimaging traits [2, 4, 16, 44, 47, 62, 68, 69, 78, 91, 96, 101, 107, 143], have all been reported to modulate pain experience and chronic pain risk. This work has culminated in a new classification system for chronic pain in ICD-11, which shifts pain category assignment to a hierarchical approach: etiology, then pathophysiology, then body site [122]. It also includes chronic primary pain as a diagnosis “agnostic with regard to etiology” [93]. These changes are important steps toward aligning diagnosis with pathophysiology, and genetics is an important piece of the puzzle. Our Genomic SEM model suggests that, in addition to condition-specific genetic susceptibility, there is a genetically encoded pathophysiology common to different chronic pain conditions and supports the view of chronic pain as a disorder involving systemic pathology [141]. This study further identifies genetic risk markers that are shared across distinct pain conditions contributing to a new, biologically grounded way of conceptualizing chronic pain conditions. The strong association between expression of chronic pain risk genes in the brain and cross-condition chronic pain provides additional evidence that the common genetic risk markers we identified may be associated with central sensitization and chronic primary pain, which have both been linked to alterations in the brain and spinal cord (135,141).

4.1. Limitations

There are several notable limitations of this study. First, although the annotated genetic associations for the general factor suggest a combination of systemic biological and psychological predispositions, the precise mechanisms underlying these predispositions need to be elucidated in future studies before they can be translated to clinical applications. In service of this goal, the pathways implicated by the SNPs and genes we have identified, can be targeted in follow-up studies to further elucidate the systemic mechanisms that lead to chronic pain.

The second limitation lies in the reliance of annotations obtained from FUMA on the information available in existing data repositories, which may be restricted by insufficient resolution or small sample sizes. Thus, although we did not find associations with inflammatory cytokines, evidence for their role in pain is abundant [41, 56, 112] and should be investigated further.

Third, we do not expect the genetic patterns we identified to be either selective or sufficient to explain cross-condition chronic pain. This is due to both pleiotropy (the effects of one gene on several different conditions) and environmental effects, which are certain to play a large part in the development of chronic pain.

Fourth, the genetic scores of our common factors should be validated for association with chronic pain using either association analysis in an independent sample. This validation would require obtaining a polygenic score and is the aim of a follow-up study.

Fifth, given sample size limitations in the UKBB for non-European individuals, we were not able to test our model for generalizability across ancestral populations, which we attempted to do in the next largest sample: South Asians (Supplementary Table 2).

By establishing genetic risk factors in a large sample, this study paves the way for more detailed assessments of pain prognosis and treatment response in targeted studies. For example, the ongoing Acute to Chronic Pain Signatures (A2CPS) study aims to establish risk factors for post-surgical pain from genetic, multi-omics, psychosocial, and neuroimaging measures in another large sample (2,800 patients; a2cps.org). Our factor scores could be tested alongside previously identified genetic patterns for multi-site pain and compared to them as prognostic risk factors for chronic post-surgical pain.

4.2. Conclusion

In summary, our findings confirm that there is a genetic susceptibility common to a broad range of diverse chronic pain conditions. The shared pathophysiology for the conditions examined here appears to lie partly in the CNS and partly scattered across many different systems and functional processes. Additionally, there is a body-wide, suggestively musculoskeletal system-specific genetic factor. Our study calls for new ways to diagnose and treat chronic pain, whereby a given chronic pain condition is not considered as only a symptom of a localized somatic disease by the clinician specializing in it, but is seen as a manifestation of an underlying shared pathology with concurrent risk for other pain conditions and previously unexplored centralized treatment targets. Future work will go beyond pain conditions and explore genetic links with psychological and physical traits to help identify patients who would benefit most from specific interventions.