A computational genetic- and transcriptomics-based study nominates drug repurposing candidates for the treatment of chronic pain.

The transcriptomic imputation models supporting this study are available at https://predictdb.org/categories/downloads/ . Reference data from the 1000 Genomes Project can be downloaded from https://www.internationalgenome.org/home . The CMap LINCS Resource 2020 data release is available at https://clue.io/data/CMap2020#LINCS2020 . Previously published GWAS datasets analyzed in this study can be obtained as described in the original publications. The GWAS of chronic pain conducted in the Mount Sinai BioMe cohort for this study is available at https://doi.org/10.5281/zenodo.17204726 . Analysis code is available on GitHub under repository chronic-pain_drug-repositioning.

This study included three chronic pain disease expression signatures: 1) presence of chronic pain, 2) severity of chronic pain, and 3) resolution of acute pain. Presence of chronic pain was defined as the application of S-PrediXcan to summary statistics from a case–control GWAS of multisite chronic pain, with a case defined as an individual with more than one pain site lasting longer than 3 months (Ncase = 82,812). 14 Severity of chronic pain was defined as the application of S-PrediXcan to summary statistics from a continuous GWAS of number of chronic pain sites (N = 387,649). 15 Both GWAS were conducted using data from the UK Biobank and include participants of European genetic ancestry. The resolution of acute pain signature was from a published whole blood differential expression study of 98 patients with acute low back pain. 19 Gene expression was measured at two time points, 3 months apart. Patients reported substantial pain on a 0–10 numeric rating scale. Differential expression was measured both across timepoints and between patients with persistent (N = 49) or resolved (N = 49) pain. For signature mapping, we used the transcriptomic signature comparing timepoint 1 and timepoint 2 in patients whose pain resolved. There are few well-powered human differential expression studies of chronic pain or pain-related conditions (∼N < 150 cases). 20 , 21 Therefore, to identify chronic pain-associated genes for downstream signature mapping, we performed summary-based transcriptome-wide association studies (TWAS). We used S-PrediXcan 22 to impute genetically regulated gene expression (GREx) from genotype in two chronic pain phenotype GWAS: presence of multisite chronic pain and number of chronic pain sites. 14 , 15 The GWAS of number of chronic pain sites was filtered to include only single nucleotide polymorphisms; no other quality control steps were applied to the GWAS summary statistics before conducting TWAS. We performed TWAS across 17 blood, skeletal muscle, and central nervous system tissue models built from the Genotype-Tissue Expression Project (GTEx) Project, 22 Depression Genes and Networks (DGN), 22 the CommonMind Consortium (CMC), 23 and PsychENCODE 24 reference transcriptome datasets. The tissue models included: anterior cingulate cortex (BA24), amygdala, caudate (basal ganglia), cerebellar hemisphere, cerebellum, cortex, frontal cortex (BA9), hippocampus, hypothalamus, nucleus accumbens (basal ganglia), putamen (basal ganglia), skeletal muscle, spinal cord (cervical c-1), and substantia nigra GTEx v8 models, CMC dorsolateral prefrontal cortex, DGN whole blood, and PsychENCODE combined temporal and frontal cortex. Tissues were chosen due to previous GWAS enrichment of chronic pain for blood, skeletal muscle, and tissues of the central nervous system through gene-level association testing. 15 , 25 , 26 TWAS results were fine-mapped using the FOCUS 27 tool using tissue-specific reference weight databases imported from each respective PrediXcan model and LD estimated from 1000 Genomes European Phase 3 reference genotypes. We ran FOCUS with all default setting excluding the – p-threshold flag, which was set to a p-value of 0.01 as the minimum GWAS p-value required to perform TWAS fine-mapping. We used this more liberal p-value threshold because our goal was to explore global, distributed transcriptomic patterns linking chronic pain and drug effects. Including subthreshold SNPs allowed us to capture potentially associated genes that may not reach genome-wide significance, although we acknowledge this comes at the risk of introducing additional noise or non-causal signal into the analysis. We obtained drug expression signatures from the Connectivity Map (CMap) resource, a compendium of drug compound effects on gene expression across diverse cell lines. 28 The Expanded CMap LINCS Resource 2020 data release was downloaded from https://clue.io/data/CMap2020#LINCS2020 in April 2022. We used level 5 replicate-consensus drug signatures, filtering for high quality signatures as defined by the CMap resource as qc_pass = 1 and (median_recall_rank_spearman ≤ 5 or median_recall_rank_wtcs_50 ≤ 5). The final filtered dataset included 89,878 drug signatures across 7668 compounds and 156 cell types ( Fig. S1 ). In the last two decades researchers have proposed many connectivity scores for the comparison of drug and disease gene expression signatures, with few studies comparing the performance of each approach. 29 , 30 Connectivity scores generally fall into two categories: 1) enrichment tests or 2) similarity measures such as correlation or cosine similarity, and each connectivity score comes with its own limitations. By considering ranks rather than effect estimates, non-parametric enrichment approaches like the modified Kolmogorov–Smirnov test may overemphasize non-differentially expressed genes, leading to an inflated false positive rate, or miss potential associations by not considering the magnitude of differential expression. 31 Conversely, similarity measures are more liable to detection noise due to lowly-expressed genes, and there is no consensus regarding the appropriate number of top and bottom differentially-expressed genes to include in association testing. 31 Further comparison of available connectivity scores can be found in the comprehensive review by Samart et al. (2021). 31 There is some evidence that combining multiple individual connectivity scores into an ensemble score is effective in nominating drug repurposing candidates and improves performance above individual scores alone. 30 , 32 In this study we applied an ensemble connectivity score combining five popular individual connectivity scores, including both enrichment- and similarity-based scores: the weighted connectivity score 28 (WCS) and four eXtreme scores (Xsum, XCor, XSpe, and XCos). 33 Fig. S2 provides a flowchart of the connectivity score calculation. TWAS results at four significance thresholds (FDR 5%, 10%, 20% and the 90% fine-mapped credible set) and differential expression study results at three significance thresholds (FDR 1%, 5%, 10%) were included in the connectivity score. We applied a stricter significance threshold to the study of acute pain resolution to ensure a balanced representation of genes across the different disease signatures. The ensemble connectivity score is calculated as follows. 1) Individual connectivity scores (WCS, XSum, XCor, XSpe, XCos) are calculated for each disease signature using Z-scores from disease and drug gene expression signatures. 2) Individual connectivity scores are centered and scaled. 3) Individual scores are averaged across significance thresholds. 4) Mean scores across significance thresholds are averaged across score types to generate the ensemble score. Individual connectivity scores (WCS, XSum, XCor, XSpe, XCos) are calculated for each disease signature using Z-scores from disease and drug gene expression signatures. Individual connectivity scores are centered and scaled. Individual scores are averaged across significance thresholds. Mean scores across significance thresholds are averaged across score types to generate the ensemble score. We performed permutation testing to understand the range of ensemble scores we would expect due to chance and assign formal significance to our ensemble scores. We scrambled each chronic pain disease signature, randomly reassigning genes to association statistics and again performed the signature mapping analysis across 100 permutations for each tissue. We calculated the empirical p-value as the number of permuted scores less than the test ensemble connectivity score, adding 1 to the numerator and denominator to account for uncertainty in estimation of the p-value. 34 Fig. 1 provides a flowchart of the full signature mapping analysis framework. Fig. 1 Flowchart outlining the signature mapping analysis. Flowchart outlining the signature mapping analysis. We performed genome-wide association studies of chronic pain using the CBIPM-BioMe Biobank Program dataset, hosted under the Mount Sinai Data Ark. This data contains de-identified imputed genotype data and electronic health records. The electronic health records were collected and transformed to the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM) format, and includes demographics, encounters, labs, diagnoses, and medications, among other patient information. The genotyping data, before sample filtering performed in this study, included a total of 53,982 samples: 31,413 genotyped using the Regeneron Global Screening Array and 22,569 genotyped using the Sema4 Global Diversity Array. Imputation was performed separately for each array, with two rounds of imputation per array. First, imputation was conducted using the Michigan Imputation Server with the 1000 Genomes Project version 5 reference panel. Then, a second round of imputation was performed using the TOPMed Imputation Server with the TOPMed freeze 8 reference panel. Results were filtered for INFO score >0.7 and then combined. Further details regarding data acquisition and quality control procedures can be found at the CBIPM-BioMe Program web page. Diagnoses were determined using the Systematic Nomenclature of Medicine- Clinical Terms (SNOMED-CT) terminology, core codes within the OMOP Data Model. Given the limited pain phenotype information available in the Mount Sinai Biobank, we were only able to classify patients as having general chronic pain based on diagnostic codes. We derived two phenotypes for general chronic pain. • Chronic pain (narrow): Derived from a list of codes where chronic pain is central to the condition. This includes diagnoses where chronic pain is a primary diagnostic feature, such as chronic pain or chronic pain syndrome, central pain syndrome, chronic back pain, chronic headache, chronic migraine, complex regional pain syndrome, and phantom limb syndrome with pain. • Chronic pain (broad): Derived from a broader list of codes encompassing both chronic pain (narrow) conditions and additional chronic conditions where pain is a common symptom but not a defining feature of the condition. This includes diagnoses such as Crohn's disease, ulcerative colitis, arthritis, temporomandibular joint disorders, neuralgia, chronic fatigue syndrome, chronic interstitial cystitis, chronic ulcerative pancolitis, endometriosis, fibromyalgia, irritable bowel syndrome, neuritis, and vulvodynia. Chronic pain (narrow): Derived from a list of codes where chronic pain is central to the condition. This includes diagnoses where chronic pain is a primary diagnostic feature, such as chronic pain or chronic pain syndrome, central pain syndrome, chronic back pain, chronic headache, chronic migraine, complex regional pain syndrome, and phantom limb syndrome with pain. Chronic pain (broad): Derived from a broader list of codes encompassing both chronic pain (narrow) conditions and additional chronic conditions where pain is a common symptom but not a defining feature of the condition. This includes diagnoses such as Crohn's disease, ulcerative colitis, arthritis, temporomandibular joint disorders, neuralgia, chronic fatigue syndrome, chronic interstitial cystitis, chronic ulcerative pancolitis, endometriosis, fibromyalgia, irritable bowel syndrome, neuritis, and vulvodynia. To be classified as having a given condition, a patient required two instances of the corresponding SNOMED-CT code. A list of all SNOMED-CT codes used to define the narrow and broad chronic pain phenotypes is provided in Table S1 . SNPs with a minor allele frequency <0.05, genotyping imputation info score <0.7, a Hardy–Weinberg equilibrium test p 0.05 were removed. Participants with discordant self-reported and genotype-derived sex information, genotype missing rate >0.05, a heterozygosity rate differing more than 3 SD from the mean, and non-European genotype-derived ancestry were excluded from the analysis. We also identified pairs of related individuals, and excluded one member of each pair with kingship coefficient >0.0884 using the –king-cutoff parameter of the plink2 software. 35 The final GWAS cohorts included 1474 individuals diagnosed with narrowly-defined chronic pain, and 5091 individuals with broadly-defined chronic pain ( Table 1 ). Autosomal GWAS for broad and narrowly-defined chronic pain were run using plink –glm function, fitting logistic regression models adjusted for ten genotype PCs, age, sex, and sequencing chip. Table 1 Demographic information for final GWAS cohort in Mount Sinai health system. Full cohort Chronic pain (narrow) Chronic pain (broad) Sex: N Female (%) 9788 (51.7%) 852 (57.8%) 3047 (59.9%) Age (Years, Mean +SD) 61.7 (18.8) 66.9 (14.5) 66.2 (16.6) Total N 18,949 1474 5091 Demographic information for final GWAS cohort in Mount Sinai health system. We assessed genetic overlap among the two chronic pain phenotypes used in signature mapping analysis (number of chronic pain sites, 15 presence of multisite chronic pain 14 ) and the BioMe narrow chronic pain GWAS phenotype using multivariable LD score regression 36 (LDSC) within the GenomicSEM 37 R package. Files were munged using info score filter of 0.9 (note info score was not available for all three traits) and a MAF threshold of 0.01. Multivariable LDSC was then performed with sample prevalence parameters 7.7% (BioMe narrow chronic pain GWAS) and 50% (GWAS of presence of multisite chronic pain), and population prevalence as 20% for both (GWAS measuring number of chronic pain sites is quantitative, therefore sample prevalence and population prevalence were set as NA). We carried out unidirectional two-sample MR to test for causal effects of medication use on chronic pain. For this analysis we leveraged 23 previously published medication use GWAS representing 23 medication classes, 38 derived from the UK Biobank cohort, and summary statistics from our own GWAS of ‘narrow’ chronic pain in Mount Sinai BioMe (described in previous section) ( Fig. 2 ; Table 2 ). Fig. 2 Mendelian randomization analysis to estimate the causal effect of medication use on chronic pain. Table 2 GWAS summary statistics used in Mendelian randomization analysis. Trait Source Cohort N (N cases) N_SNP Chronic pain (narrow) Cote Mount Sinai BioMe 18,949 (1474) 7,288,503 Medication use traits Wu et al. 38 UK Biobank 39 78,808–305,913 4,653,539 N = total sample size or total sample size range across medication use traits, N_SNP = number of SNPs included in GWAS. Mendelian randomization analysis to estimate the causal effect of medication use on chronic pain. GWAS summary statistics used in Mendelian randomization analysis. N = total sample size or total sample size range across medication use traits, N_SNP = number of SNPs included in GWAS. We used three MR methods; MR-Egger, 40 MR-IVW (inverse variance weighted), 41 and MR-RAPS (robust adjusted profile score). 42 Both medication use and chronic pain are highly polygenic complex traits, and pleiotropy and other contributors to IV assumption violation are likely. These methods each account for this probably departure from perfect IV assumption adherence, increasing power to find potential causal relationships. MR-RAPS also accounts for (and power to find causal effects is increased by) inclusion of ‘weak’ (below genome-wide significance) instruments (SNPs). 42 To prepare summary statistics for MR, we used the R packages ‘TwoSampleMR’ 43 , 44 , ‘mr.raps’ 42 , ‘ieugwasr’ 45 and related 1000 Genomes European LD reference datasets, and performed LD clumping locally. For each medication use GWAS, we first used the TwoSampleMR function ‘read_exposure_data’ to begin initial processing GWAS summary statistics, followed by reserving SNPs associated with that medication class use at p < 5 x 10 −6 and then carrying out LD clumping 35 (clump_kb = 500, clump_r2 = 0.001, clump_p = 0.99) using ‘ieugwasr’ ‘ld_clump’ function, local copies of 1000 Genomes European genetic ancestry plink files, and ‘genetics.binaRies’ ‘get_plink_binary’. We then used ‘TwoSampleMR’ ‘read_outcome_data’ to read in the subset of chronic pain GWAS summary statistics corresponding to the SNPs remaining in the exposure (medication use) summary statistics following clumping. We harmonized using ‘TwoSampleMR’ ‘harmonise_data’ function, and performed MR-Egger and MR-IVW on the final harmonized dataset using ‘TwoSampleMR’ ‘mr’. We then performed MR-RAPS using ‘mr.raps’ ‘mr.raps.all’, retaining any warning messages for examination and then running each of the six models individually with their corresponding functions (see mr.raps vignette for detail) to obtain a p value (since this is not outputted by mr.raps.all). For models with robust loss function (Huber or Tukey) the recommended default k values were used. MR-RAPS involves fitting of a total of 6 models, and the most appropriate is reserved – in our analyses this was a model without overdispersion (indicating not all instruments are horizontally pleiotropic) and with robust loss function (robust loss functions Huber or Tukey are less sensitive to outliers). This resulted in a total of three MR outputs (Egger, IVW, MR-RAPS robust without overdispersion) per exposure-outcome analysis (23 analyses). We applied two levels of multiple testing correction: Bonferroni correction within-analysis (threshold 0.05/3 = 0.017) and FDR correction across all analyses (69 analyses total). We consider unadjusted MR p < 0.05 to indicate suggestive significance. This study used publicly available summary-level data and anonymized individual-level genotype and clinical information obtained through the IRB-approved CBIPM-BioMe Biobank Program. Separate ethics approval for this study was not required. The funders were not involved in the study design, data collection, analysis, interpretation, or writing of this article.

We imputed genetically-regulated gene expression (GREx) across 17 tissue types with previous chronic pain GWAS enrichment, 14 , 15 , 25 , 26 including tissues of the central nervous system, blood, and skeletal muscle. We identified 1976 gene–tissue associations (p fdr < 0.05) with severity of multisite chronic pain across 17 tissues and 1033 unique genes ( Supplementary File 1 ). We identified 944 gene–tissue associations (p fdr < 0.05) with presence of multisite chronic pain across 17 tissues and 516 unique genes ( Supplementary File 2 ). We see the largest number of significant gene associations for whole blood, the combined temporal and frontal cortex, skeletal muscle, cerebellum, and the dorsolateral prefrontal cortex. There is substantial overlap between results for the two chronic pain phenotypes, with many gene associations identified in TWAS for the presence of chronic pain also identified in TWAS for severity of chronic pain ( Table 3 ). Table 3 Number of shared and distinct TWAS associations (FDR 5%) across chronic pain phenotypes. Tissue Total tested genes Total significant genes (% of tested) Number of shared significant genes Severity of chronic pain Presence of chronic pain PsychENCODE_Cortex 14,342 586 (4.09%) 290 (2.02%) 203 DGN_Whole Blood 10,467 219 (2.09%) 103 (0.98%) 71 Muscle_Skeletal 7583 142 (1.87%) 83 (1.09%) 56 Cerebellum 6794 112 (1.65%) 74 (1.09%) 52 CMC_DLPFC 10,785 188 (1.74%) 68 (0.63%) 50 Cerebellar_Hemisphere 5753 90 (1.56%) 50 (0.87%) 38 Cortex 5500 87 (1.58%) 49 (0.89%) 34 Nucleus_accumbens_basal_ganglia 4851 86 (1.77%) 39 (0.80%) 33 Caudate_basal_ganglia 5004 71 (1.42%) 36 (0.72%) 26 Putamen_basal_ganglia 4436 72 (1.62%) 31 (0.70%) 23 Anterior_cingulate_cortex_BA24 3544 59 (1.66%) 32 (0.90%) 22 Frontal_Cortex_BA9 4563 80 (1.75%) 32 (0.70%) 22 Amygdala 2787 41 (1.47%) 12 (0.43%) 11 Hippocampus 3688 54 (1.46%) 11 (0.30%) 8 Hypothalamus 3652 47 (1.29%) 9 (0.25%) 8 Spinal_cord_cervical_c-1 3250 23 (0.71%) 14 (0.43%) 8 Substantia_nigra 2559 19 (0.74%) 11 (0.43%) 8 Number of shared and distinct TWAS associations (FDR 5%) across chronic pain phenotypes. We hypothesized that drugs with an opposite effect on gene expression to chronic pain, or similar effects to the resolution of acute pain, may be therapeutic for the prevention or treatment of chronic pain ( Fig. 3 ). Effective prevention and management of acute pain is an important factor in reducing the risk of chronic pain development. 46 Drugs with gene expression profiles similar to the resolution of acute pain (positive connectivity score) may be additional therapeutic candidates. To calculate the similarity or difference between drug and disease transcriptomic signatures, we applied a novel ensemble score, combining the five most popular single signature mapping scores. 28 , 29 Our signature mapping analysis nominated 894 putative drug candidates across all three disease signatures (resolution of acute pain: 284, severity of chronic pain: 460, presence of chronic pain: 367) (Presence of chronic pain and number of chronic pain sites: FDR 1%, Resolution of acute pain: FDR 5%) ( Table 4 ; Supplementary File 3 ). We chose different significance thresholds to ensure a balanced representation of drug candidates nominated by each disease signature. We observe greater similarity of disease signatures between presence and severity of chronic pain compared to with acute pain resolution ( Fig. S3 ), and as a result we also see greater overlap in drug candidates between those nominated using the severity or presence of chronic pain TWAS signatures. 25 drug candidates were nominated by all three disease signatures ( Table S2 ). We see the largest number of significant signature mapping results for transcriptomic signatures of drug exposures in A375, PC3, HA1E, and A549 cell lines, likely driven by the cell line composition of the Connectivity Map resource ( Fig. S1 ). We also observe significant signature mapping score in more CNS-relevant cell types; for example, neural progenitor cells are among the top five cell lines with the most significant signature mapping results for the resolution of acute pain signature. Fig. 3 Toy representation of example signature mapping result. Drugs with the opposite effect on gene expression as chronic pain demonstrate negative ensemble connectivity scores. Signature mapping scores typically show a sigmoidal distribution, where a subset of tested medications show strong concordant or opposite effects on gene expression as disease, corresponding to highly positive or negative connectivity scores, respectively. Table 4 Signature mapping analysis: Number of significant drug candidates by tissue and chronic pain phenotype. Cohort Tissue Severity of chronic pain (TWAS) Presence of chronic pain (TWAS) Resolution of acute pain (DEG) GTEx Anterior cingulate cortex BA24 0 0 Amygdala 1 0 Caudate basal ganglia 0 0 Cerebellar hemisphere 0 25 Cerebellum 2 20 Cortex 0 0 Frontal cortex BA9 0 186 Hippocampus 0 0 Hypothalamus 0 0 Nucleus accumbens 0 0 Putamen 43 0 Skeletal muscle 5 24 Spinal cord cervical c-1 0 0 Substantia nigra 0 0 CMC DLPFC 370 150 PsychENCODE Temporal + frontal cortex 7 3 DGN, DEG Study Whole blood 116 0 284 Total 544 408 284 Unique Drugs 460 367 284 Unique Drugs across all disease signatures = 894 Drugs shared across all disease signatures = 25 Intersection of unique drugs between disease signatures Severity of chronic pain (TWAS) Presence of chronic pain (TWAS) Resolution of acute pain (DEG) Severity of chronic pain (TWAS) 460 Presence of chronic pain (TWAS) 131 367 Resolution of acute pain (DEG) 60 51 284 Toy representation of example signature mapping result. Drugs with the opposite effect on gene expression as chronic pain demonstrate negative ensemble connectivity scores. Signature mapping scores typically show a sigmoidal distribution, where a subset of tested medications show strong concordant or opposite effects on gene expression as disease, corresponding to highly positive or negative connectivity scores, respectively. Signature mapping analysis: Number of significant drug candidates by tissue and chronic pain phenotype. 210/894 candidates are prescribable in the U.S. (resolution of acute pain: 63, severity of chronic pain: 121, presence of chronic pain: 92). Drug candidates were annotated according to the Anatomical Therapeutic Chemical (ATC) classification system using the DrugBank Online query tool. 47 Drug candidates span many broad drug classes, including antineoplastic and immunomodulating agents, psychoanaleptics, psycholeptics, drugs related to the genitourinary system and sex hormones, analgesics, and cardiovascular medications, among others ( Table S3 ). The overrepresentation of certain drug classes among our signature mapping results is likely affected both by the composition of drug types in the CMap resource and the relevance of the treatment to the biology of chronic pain. Among drugs recommended for chronic pain treatment by the CDC and FDA, and drugs tested in ongoing or completed clinical trials of chronic pain, 13 medications were nominated as drug candidates in our signature mapping analysis: celecoxib, clonazepam, dexamethasone, duloxetine, fluoxetine, lidocaine, nifedipine, omeprazole, paracetamol, pioglitazone, simvastatin, triamcinolone, and varenicline. . All three chronic pain traits show significant genetic correlation ( Table 5 ), indicating that the narrow chronic pain phenotype is similar to the more clinically-defined phenotypes of prior chronic pain GWAS. Chronic pain (narrow) shows significant genetic correlation albeit with large associated SE, likely because this trait GWAS is underpowered due to low sample size. This is also reflected in the non-significant liability SNP-h2 value for this trait. Table 5 Genetic correlation between chronic pain phenotypes. Chronic pain (narrow) Presence of chronic pain Severity of chronic pain SNP-h2 (SE) Chronic pain (narrow) 1 1.008 (0.306) 0.97 (0.235) 0.04 (0.13) Presence of chronic pain 14 −0.0086 (0.0084) 1 0.98 (0.04) 0.17 (0.0081) Severity of chronic pain 15 −0.0071 (0.0074) 0.738 (0.0082) 1 0.074 (0.0028) Upper: Genetic correlation results for multivariable LDSC. Standard error for genetic correlation coefficients is shown in parentheses. Lower: Cross-trait intercept (SE). SNP-h2 = Liability-scale SNP-heritability. Genetic correlation between chronic pain phenotypes. Upper: Genetic correlation results for multivariable LDSC. Standard error for genetic correlation coefficients is shown in parentheses. Lower: Cross-trait intercept (SE). SNP-h2 = Liability-scale SNP-heritability. While our signature mapping analysis of chronic pain traits successfully nominated novel drug repurposing candidates in a transcriptomics-informed and tissue-specific context, these drug hypotheses are largely based on associations and do not explicitly imply causal effects of drug exposure on a reduced risk of chronic pain. As a complementary analysis, we used mendelian randomization (MR) to test for potential causal effects of medication use on chronic pain. For this analysis we used previously published medication use GWAS for 23 medication ATC classes in UK Biobank to derive instrumental variables for medication exposure. To derive instrumental variables for the outcome (chronic pain) in a GWAS cohort without overlap with UK Biobank, we performed GWAS of broad chronic pain (case N = 5,091, control N = 13,858) and narrow chronic pain (case N = 1,474, control N = 17,475), as defined in the Methods, using the Mount Sinai BioMe cohort. GWAS for both narrow and broad chronic pain did not reveal any genome-wide significant associations (lowest SNP p-value: narrow = 1.35e-07, broad = 1.96e-06), which is to be expected given the small sample sizes of these cohorts ( Fig. S4 ). In addition to more significant SNP findings compared to broad chronic pain, we also find greater deviation from the null for the GWAS of narrowly-defined chronic pain, in line with previous research identifying greater power in genetic association studies for more homogeneous disease cohorts, 48 and thus used the “narrow” case definition GWAS for subsequent analysis. Narrow chronic pain, presence of multisite chronic pain 14 , and number of chronic pain sites 15 all showed significant correlation ( Table 5 ). MR analysis identified three medication classes with a suggestive (p-value < 0.05) causal effect on chronic pain ( Table S4 ), including N02A (Opioids), B01A (antithrombotic agents), and A02B (drugs for peptic ulcer and GERD) ( Table 6 ). N02A is also significantly associated after within-analysis Bonferroni correction (pval < 0.017). Results where A02B use is the exposure showed suggestively significant (p < 0.05) negative (beta < 0) effect of this medication on chronic pain (i.e., possibly causing improvement), however the MR Egger intercept was significantly different from 0 ( Table S5 p = 0.0053), suggesting significant directional pleiotropy and/or violation of the InSIDE assumption 49 among SNP instruments – IVW causal estimates are therefore likely biased overall for this drug class and should not be further interpreted. I GX 2 values ( Table 4 ) also suggest measurement error in MR Egger estimates leading to dilution (causal estimates will be downardly biased towards null). Table 6 MR results with suggestive significance. exposure method nsnp loss.function over.dispersion b se pval Q Q_pval I 2 A02B (Drugs for peptic ulcer and GERD) mr.raps 34 l2 FALSE −0.245 0.190 0.198 NA NA NA IVW 34 NA NA −0.237 0.182 0.193 32.8 0.48 0 MR Egger 34 NA NA −2.311 0.938 0.019 27.7 0.68 0 B01A (Antithrombotic agents) mr.raps 21 l2 FALSE 0.455 0.222 0.040 NA NA NA IVW 21 NA NA 0.439 0.255 0.086 28.0 0.11 0 MR Egger 21 NA NA −0.581 0.867 0.511 25.9 0.13 0 N02A (Opioids) mr.raps 20 l2 FALSE −0.477 0.173 0.006 NA NA NA IVW 20 NA NA −0.455 0.185 0.014 23.8 0.20 0 MR Egger 20 NA NA −0.203 1.213 0.869 23.8 0.16 0 b = beta value (causal estimate value), se = standard error of beta, pval = p value for beta, Q = heterogeneity measure (heterogeneity of individual SNP causal estimate values), Q_pval = p value associated with Q. IVW = Inverse variance weighted, I 2 = I GX 2 . Bolded values = should be interpreted further. MR results with suggestive significance. b = beta value (causal estimate value), se = standard error of beta, pval = p value for beta, Q = heterogeneity measure (heterogeneity of individual SNP causal estimate values), Q_pval = p value associated with Q. IVW = Inverse variance weighted, I 2 = I GX 2 . Bolded values = should be interpreted further. In the case of antithrombotic agents (B01A), the MR Egger intercept is not significantly different from zero ( Table S5 p = 0.46), and the MR Egger causal test p value > 0.05, indicating only the IVW and MR-RAPS estimates should be interpreted further. These causal estimates are greater than zero and suggestively significant (p < 0.05), indicating that use of this medication has a potentially positive causal effect on chronic pain and could contribute to pain symptomatology. Extensive follow-up, including experimental and clinical studies, is needed to clarify the biological mechanisms underlying this association and its clinical implications.

In this study we establish a genetics-informed framework for drug repurposing in chronic pain, leveraging existing genetic and transcriptomic resources. We identify FDA-approved drugs we hypothesize may be beneficial for prevention or treatment of chronic pain, detected through opposing expression patterns to chronic pain and association with reduced chronic pain through causal inference approaches. Signature mapping, previously effective in complex conditions like hyperlipidemia, hypertension 50 and Alzheimer's disease, 51 was expanded using a novel ensemble connectivity score and analyses across multiple pain phenotypes with both measured and genetically-predicted gene expression. We predict 210 U S.-prescribable drugs with significant reversed gene expression effects relative to chronic pain. Our analysis nominates known pain treatments, such as duloxetine, an antidepressant for neuropathic pain, and lidocaine, a topical or intravenous analgesic. It also highlights cardiovascular medications (calcium channel blockers, cardiac glycosides, ACE inhibitors, and beta blockers), corticosteroids, analgesics, antipsychotics, and antidepressant drugs as protective. Cardiovascular drugs are particularly notable, as chronic pain is comorbid with cardiovascular disease, with observational evidence suggesting a dose-dependent relationship between pain severity and cardiovascular disease. 52 A recent study modeling clinical record data showed cardiovascular medications demonstrate strong negative prediction across chronic, but not acute, pain conditions. 53 Family and twin analysis support shared genetic predisposition, 54 and Mendelian randomization indicates a possible causal role of chronic pain in cardiovascular disease development. 55 This may also extend to shared predisposition at the level of gene expression; a recent TWAS identified shared genes associated with both multisite chronic pain and cardiac dysrhythmia. 56 While further studies are needed to clarify the co-occurrence of these conditions and the mediating effects of factors like depression, socioeconomic status, and physical activity, our analysis suggests a complex interplay between chronic pain and cardiovascular disease and supports cardiovascular medications as potential chronic pain therapies. 11 medications from our analysis have been tested in ongoing or completed chronic pain trials ( Supplementary File 4 ). Pioglitazone, an anti-diabetes medication, has shown pain relief in both gut and neuropathic mouse pain models. 57 , 58 Other promising candidates include the antidepressant mirtazapine, which reduces pain symptoms and pain-related sleep disruption in human and animal chronic pain studies. 59 , 60 Statins like atorvastatin and rosuvastatin also display anti-inflammatory and pain-reducing effects when tested in rodent neuropathic pain models. 61 , 62 Lastly, atypical antipsychotics like ziprasidone, quietiapine, clozapine, and olanzapine have been previously investigated as secondary chronic pain treatments and all show reversed effects on pain-induced transcriptional changes in our analysis. 63 These findings generate genetically-informed hypotheses for future research of chronic pain treatments. Experimental follow-up in cell or animal models of pain, as well as investigation in clinical settings, will be essential to evaluate the therapeutic potential of these medications. One advantage of this approach is that signature mapping is largely hypothesis generating and agnostic to prior understanding of disease etiology or current therapeutics, making it particularly useful for diseases of unknown or heterogeneous etiology. 64 Additionally, this study leverages human data for testing as opposed to phenotyping exclusively in cell or animal disease models, potentially allowing for more complete modeling of the clinical complexity of chronic pain. Although not all disease effects are mediated through gene expression changes, this framework can be extended as multi-omic datasets become available and is an informative framework for diseases lacking effective, long-term treatment. MR analyses highlighted two medication classes with suggestive (p-value< 0.05) causal effects on chronic pain that meet MR assumptions and can be further interpreted. Antithrombotic agents show a potential positive causal effect on chronic pain, meaning exposure to these medications worsen treatment outcomes. Opioids show a negative causal effect on chronic pain, indicating therapeutic benefit, and is the only medication passing within-analysis Bonferroni correction (p-value < 0.017). These findings are in line with previous literature – opioid medications are powerful painkillers, though they carry a strong risk of misuse or addiction and unclear effectiveness in long term chronic pain treatment. 17 Antithrombotic agents, including vitamin K antagonists like warfarin, factor Xa inhibitors like rivaroxaban, and heparin, can have pain and chronic pain as side effects through various mechanisms. 65 However, extensive follow-up, including experimental studies, is needed to clarify the biological context in which antithrombotic agents may impact chronic pain. Ultimately, while this analysis recapitulates findings from the main drug signature mapping analysis, it also illustrates the utility of signature mapping for specific drug investigation alongside these more overarching and less powerful MR analyses. Ideally, a GWAS with large numbers of significant (p < 10ˆ–8) variants would be used to derive instruments for MR - our GWAS of narrow chronic pain was relatively underpowered compared to GWAS available for other chronic pain phenotypes. However, we account for weak instrument bias through use of methods such as MR-RAPS, where inclusion of many weak instruments actually increases power to find causal estimates. 66 Overall, as with any MR analysis (particularly those involving complex traits), causal relationship results are suggestive rather than definitive. Results of this study must be interpreted in the context of several limitations. First, the number of significant TWAS results per tissue is partly driven by the size of the reference panel used to build the transcriptomic imputation model. When comparing findings between tissues in this study, researchers should consider that results are influenced by the statistical power of each tissue model to estimate gene expression, and not necessarily biological relevance to chronic pain. In addition, these models are derived from post-mortem tissue, and do not consider environmental or developmental genetic effects on gene expression with potential disease relevance. Similarly, while CMap is a comprehensive resource testing many drug compounds, these expression signatures are largely derived from cancer cell lines and do not model the complexity of medication effects on chronic pain in vivo. Additionally, some drugs are tested at clinically infeasible dosages or durations. A medication may have similar transcriptomic effects across cell lines, or it may not, and results of this study must be interpreted cautiously. Further investigation in cellular or in vivo models of chronic pain is warranted. We also note that electronic health record data is not designed for research use, and misclassification of chronic pain remains possible despite testing specific and broad chronic pain phenotypes. Future investigation would benefit from more specific follow-up analysis accounting for comorbid conditions or differing therapeutic benefit in chronic pain subtypes. Finally, we emphasize results of this study are based on statistical modeling of population-level genetics, and do not demonstrate causal effects of drug exposure on protection against chronic pain development. Extensive follow-up is needed to understand therapeutic benefit, including candidate drug testing in cell or animal chronic pain models, as well as evaluation of side effects and safety in long-term treatment. There are many promising future avenues to further understanding of the effectiveness of these medications in chronic pain treatment. For researchers interested in selecting a candidate medication for experimental follow-up, we recommend beginning with drugs that have established safety protocols and an absence of severe side effects. Additional prioritization steps could include ranking candidates by the strength of their negative connectivity scores, selecting drugs nominated by multiple pain phenotypes, focusing on those tested in a relevant cell line (for example neural progenitor cells or neurons), and considering any preliminary evidence of effectiveness in pain treatment. One illustrative example is niclosamide, an anthelmintic drug used to treat tapeworm infections. In our study, this drug is nominated as a repurposing candidate by all three pain phenotypes and ranked among the top 50 medications in our signature mapping analysis of chronic pain severity. Prior research shows that niclosamide reduces macrophage-induced inflammation in endometriosis, 67 and reverses hyperalgesia in a rat model of neuropathic pain. 68 Beyond drug prioritization, there are also many opportunities to extend the computational framework developed in this study. Expanding drug transcriptomic resources to include more relevant cellular contexts, as opposed to cancer cell lines, would better recapitulate drug treatment effects and yield greater insights. Future studies can also test additional disease signatures as they become available. For example, the transcriptomic data available from mouse models of pain could be used to investigate cell-type specific signatures of chronic pain. 69 Moreover, the chronic pain TWAS used in this study are derived from individuals of European ancestry, and future work would benefit from the inclusion of diverse ancestral groups in the estimation of genetically-informed chronic pain expression signatures. This would facilitate more reliable generalization of findings to non-European genetic ancestry populations, as well as deepen understanding of genetic ancestry associated differences in chronic pain traits and related gene expression changes. Finally, while electronic health records are an extensive resource of patient interactions with the healthcare system, datasets with more precise chronic pain phenotyping may be additionally informative for accurate diagnosis and causal inference in future genetic studies.

LMH & ACC conceived and designed the study with input from CS and HY, LMH supervised and coordinated funding and project administration including data access. ACC performed all formal statistical analyses apart from MR and oversaw data curation. KJAJ carried out Mendelian randomization (MR) analyses and advised on chronic pain conditions and traits. ACC & KJAJ both directly accessed and verified the underlying data reported in the manuscript. CS reviewed significant drug results and provided clinical context. All authors contributed to drafting and reviewing manuscript for submission. All authors read and approved the final version of the manuscript.

Chronic pain, broadly defined as pain lasting longer than three months, is a common and understudied condition that affects an estimated 20–30% of the global population, 1 , 2 including over 50 million adults in the United States. 3 Chronic pain conditions low back or neck pain, musculoskeletal disorders, and migraine are among the top ten global leading causes of disability, 4 and the incidence of chronic pain exceed that of other major conditions, including diabetes, depression, and hypertension. 5 Chronic pain and many chronic pain conditions disproportionately affect women, older adults, individuals of low socioeconomic status, and those living in rural areas, and varies between different racial and ethnic groups. 6 , 7 Chronic pain is a highly complex condition with dynamic biological, social, and psychological impacts on disease risk and long-term outcomes; factors like comorbid psychiatric conditions, poor sleep, or poor social support systems can act as both a risk factor for and consequence of chronic pain development. 8 , 9 Although chronic pain can arise secondary to specific diseases or injuries, there is substantial evidence also supporting the study of chronic pain as a primary complex disease trait. 8 Many chronic pain conditions co-occur, with pain commonly reported in multiple areas of the body, suggesting some underlying shared mechanisms among chronic pain disorders. 10 Relationships between chronic pain severity and tissue damage are complex and non-linear within chronic pain conditions; for some conditions such as osteoarthritis there is no discernible relationship between pain and measurable tissue damage, in rheumatoid arthritis marked chronic pain may still be felt despite well-controlled inflammation, and in endometriosis excision surgeries often do not lead to full resolution of chronic pain. 11 , 12 , 13 Genome-wide association studies (GWAS) of chronic pain and pain traits agnostic of traditional diagnostic boundaries indicate that chronic pain is a complex trait with distinct genetic risk. The SNP-heritability is estimated at about 10–20% for chronic pain phenotypes, 14 , 15 and ranging from 3 to 20% for chronic pain conditions. 16 Chronic pain is most effectively treated using an interdisciplinary approach, with a combination of medication and non-drug alternatives such as cognitive behavioral therapy, exercise, or physical therapy. 9 However, these nonpharmacological treatments are not always covered by insurance, and thus not easily accessible to individuals of lower socioeconomic status seeking chronic pain treatment. Pharmacological treatments have historically relied on opioids, but given the risk of addiction, recent CDC and FDA guidelines emphasize non-opioid alternatives, including gabapentin, pregabalin, non-steroidal anti-inflammatory drugs (NSAIDs), antidepressants, anticonvulsants, and topical treatments such as lidocaine and capsaicin. There is insufficient evidence supporting the sustained effectiveness of these drugs for long-term treatment of chronic pain and these medications show small to moderate efficacy in short-term trials, 10 , 17 highlighting a continued need for drug alternatives. Given the complexity of chronic pain and limitations of traditional drug development pipelines, particularly those focused on single molecular targets, computational drug repurposing offers a promising alternative. By combining genetic and transcriptomic information about a disease, we can estimate existing medications with potential therapeutic benefit for complex conditions like chronic pain. Prior studies suggest that drugs with genetic evidence are more likely to proceed to late-stage clinical trials. 18 A phenotype-driven approach that considers the whole genome and transcriptome may yield additional insights beyond the single molecule- or gene-level investigation of traditional drug discovery pipelines. The aim of this study is to identify potential new treatments for chronic pain by applying a validated functional genomics drug repurposing method, signature mapping, to genetic and transcriptomic data. Specifically, we first generate chronic pain transcriptomic signatures by performing summary-based transcriptome-wide association studies (TWAS). We then use these chronic pain signatures, along with transcriptomic signatures from a published gene expression study of pain resolution, as inputs for signature mapping analyses. This approach nominates drugs with transcriptomic effects opposite those of chronic pain or like those of acute pain resolution as potential therapies. Finally, we complement these findings with Mendelian randomization analyses to assess causal relationships between drug classes and chronic pain. Through this genome-wide and transcriptome-wide framework, we seek to prioritize FDA-approved medications for further investigation as potential treatments for chronic pain.

None of the authors have interests to declare.