Multi-ancestry genome-wide association and integrated multi-omics analyses of endometriosis and its clinical manifestations

Abstract

Endometriosis is a chronic systemic disease affecting ~10% of women, yet its genetic basis and molecular mechanisms remain poorly understood. Hence, we conducted a genome-wide association study of endometriosis and adenomyosis in ~1.4 million women, including 105,869 cases, aiming to expand loci discovery across ancestries, dissect symptom-specific effects, and integrate multi-omic data. We identified 80 genomic regions associated with endometriosis risk, including 37 novel loci and five also associated with adenomyosis. We identified putative causal variants underlying over 50 of these associations. Transcriptomic, epigenetic, and proteomic analyses across tissues linked endometriosis risk to pathways involved in cell differentiation, immune and hormonal regulation, tissue remodeling, and inflammation. Drug-repurposing analyses highlighted potential treatments currently used for breast cancer, contraception, and preterm birth prevention. Endometriosis polygenic risk interacted with abdominal pain, anxiety, migraine, and nausea. This study advances the understanding of genetic risk factors for endometriosis and molecularly supports several hypotheses on its pathogenesis.

Endometriosis is a chronic, systemic inflammatory disease characterized by the presence of endometrial-like tissue outside the uterine cavity¹. The disease, which affects 10% of the reproductive-age female population, carries a significant public health burden due to a debilitating, multi-system symptomatic profile that severely reduces quality of life². While endometriosis is a complex disease, it has a substantial genetic component, with a twin-based heritability (h²) estimated at 50%³ and a single nucleotide polymorphism (SNP)-h² of 8%^4,5. Genome-wide association studies (GWAS) have provided crucial insights into the molecular pathways involved in disease pathogenesis, implicating genes related to hormonal regulation, immunity, and cellular adhesion^4,6,7. The two most recent GWAS analyses of endometriosis, comprising 928,413 participants (including 44,125 cases; 30% non-European ancestry)⁷ and 762,600 participants of European descent (including 60,674 cases)⁴, identified up to 45 loci. Nevertheless, the complex genetic architecture of endometriosis remains to be fully characterized, especially related to the range of symptoms reported by affected women. Previous efforts to address endometriosis heterogeneity have included stratification by surgically defined subtypes, pain-related phenotypes, and the presence of comorbid conditions such as adenomyosis (a disease where endometrial tissue grows into the muscular wall of the uterus⁸) and infertility⁴. However, there is no international consensus on how to define or classify endometriosis. Existing surgical classification and staging systems show little or no correlation with patient outcomes, including pain severity and other symptoms⁹.

In this study, we conducted a large-scale cross-ancestry GWAS including 105,869 cases and 1,282,731 controls from eight cohorts to substantially expand the map of genetic risk factors for endometriosis and to further dissect its clinical and genetic heterogeneity. We also report multiple genome-wide significant loci for adenomyosis. Moving beyond traditional classifications, we performed genetic analyses on a wide range of clinical symptoms and explored the genetic overlap with key comorbidities, allowing for an in-depth investigation of pleiotropy and potential causal relationships. While ancestry-specific analyses have been previously conducted, our study is, to our knowledge, the first to implement a cross-ancestry polygenic risk score (PRS) framework, including individuals of six ancestry groups (African, AFR; Admixed American, AMR; Central/South Asian, CSA; East Asian, EAS; European, EUR; Middle Eastern, MID) to assess predictive performance and genetic transferability across global populations. To translate genetic associations into biological function, we performed comprehensive functional analyses for different definitions of endometriosis and adenomyosis. Overall, this work refines our understanding of the genetic architecture of endometriosis, providing new insights into the biology underlying its specific clinical features.

RESULTS

We conducted a GWAS meta-analysis of eight cohorts (Supplementary Table 1) with multiple phenotype definitions (Fig. 1). Ancestry-specific GWAS of endometriosis combined definition (i.e., endometriosis cases identified through self-reported information and electronic health records, EHRs) were performed for AFR, AMR, EAS, and EUR (Extended Data Fig. 1). For EUR, we also conducted GWAS of clinically diagnosed endometriosis based on EHRs (n_cases = 36,695), self-reported endometriosis (n_cases = 50,867), adenomyosis (n_cases = 8,753), and endometriosis without adenomyosis (n_cases = 6,627) (Extended Data Fig. 2). Details about heritability and genetic correlation between individual and meta-analyzed datasets are in the Supplementary Note and Supplementary Tables 2 and 3.

Figure 1 |. Phenotype definitions for endometriosis and adenomyosis.

African, AFR; Admixed American, AMR; Central/South Asian, CSA; East Asian, EAS; European, EUR; Middle Eastern, MID. Some graphical elements and icons were created using Canva Pro (https://www.canva.com/).

Gene discovery

We identified 80 LD-independent variants (r² < 0.1, based on 1000 Genomes Project EUR linkage disequilibrium (LD) reference panel¹⁰) showing genome-wide significant association (P < 5 × 10⁻⁸) with cross-ancestry endometriosis combined definition (Fig. 2 and Supplementary Table 4). Of these, 37 were novel (LD-independent from the previous GWAS findings). Among the new associations, the most significant was rs10514243 (β = −0.01, P = 5.9 × 10⁻¹³), located near CTD-2218K11.2, a predicted long non-coding RNA locus, ~280 kb downstream from XRCC4 and ~250 kb upstream from RPS23 and ATG10. The previously reported SYNE1 rs58415480 locus was significant for all endometriosis definitions and adenomyosis in EUR (cross-ancestry endometriosis: β = −0.04, P = 9.7 × 10⁻⁶²; EUR combined endometriosis: β = −0.03, P = 6 × 10⁻⁵⁷; EUR clinical endometriosis: β = −0.05, P = 8.1 × 10⁻⁵⁸; EUR self-reported endometriosis: β = −0.05, P = 1.7 × 10⁻²⁰; adenomyosis: β = −0.07, P = 6.2 × 10⁻¹³; endometriosis without adenomyosis: β = −0.09, P = 1.5 × 10⁻⁹). Of the 80 loci, WNT4 rs55938609 reached genome-wide significance (β = 0.1, P = 2.5 × 10⁻¹³) in EAS, which was also genome-wide significant for all EUR definitions, except for adenomyosis (Supplementary Table 4). We identified 63 genome-wide significant loci for the EUR endometriosis combined definition, one for EAS, 44 for EUR clinical endometriosis, 30 for EUR self-reported endometriosis, five for EUR adenomyosis, and five for EUR endometriosis without adenomyosis (Extended Data Figs. 1 and 2 and Supplementary Table 4). We did not identify any genome-wide significant SNP for the other population groups. Clinical endometriosis revealed four novel loci (most significant locus: rs12444457, β = 0.02, P = 1.1 × 10⁻⁸, located near RP11–467I17.1, ~160 kb downstream of WWOX and ~230 kb upstream of MAF) that were not genome-wide significant in the other endometriosis definitions (Supplementary Table 4). Additionally, there was one locus specific to self-reported endometriosis (rs10860902, β = 0.02, P = 1 × 10⁻⁸), located near LINC00485, ~135 kb upstream of IGF1 and ~145 kb downstream of PAH. Two novel endometriosis-associated loci, HRH1 rs184700 and SSPN rs7974900, were also genome-wide significant for adenomyosis (β = 0.06, P = 3.3 × 10⁻⁸ and β = 0.05, P = 6.3 × 10⁻⁹, respectively), but they did not reach significance for endometriosis without adenomyosis. Comparing effect sizes across disease definitions and ancestry groups (Extended Data Figs. 1 and 2), we observed several differences, suggesting the specificity of certain genetic effects (Supplementary Note and Supplementary Table 4).

Figure 2 |. Genome-wide association study (GWAS) of endometriosis (top) and adenomyosis (bottom).

Manhattan plots related to the cross-ancestry GWAS meta-analysis of endometriosis combined definition (n_cases = 105,869 and n_controls = 1,282,731) and adenomyosis (n_cases = 8,763 and n_controls = 415,718). Cases were defined by clinical or self-reported diagnosis, and controls were individuals with no recorded diagnosis of the respective condition. Loci highlighted in bold represent novel hits compared to previous endometriosis GWASs. Each point represents the P-value from a fixed-effects meta-analysis (weighted by sample size and P-value using METAL) of two-sided logistic regression tests of association between individual genetic variants and the phenotype across contributing cohorts, ordered by genomic position on the x axis and shown as −log₁₀(P-value) on the y axis. The dashed line represents the genome-wide significance threshold (P < 5 × 10⁻⁸), based on Bonferroni multiple testing correction.

Considering LD independent variants (r² < 0.1) from the index SNPs in the primary GWAS, the conditional analysis identified five additional loci for cross-ancestry endometriosis combined definition and eight additional loci for the EUR clinical phenotype (P_J < 5 × 10⁻⁸; Supplementary Table 5). Among these loci, two variants mapped to ESR1: rs74713348 (P_{Jcross-ancestry} = 6.97 × 10⁻¹⁵; P_{JEUR-combined} = 8.43 × 10⁻¹⁴; P_{JEUR-clinical} = 5.49 × 10⁻²¹) and rs111930025 (P_{Jcross-ancestry} = 7.9 × 10⁻¹⁰; P_{JEUR-combined} = 4.84 × 10⁻⁹; P_{JEUR-clinical} = 7.7 × 10⁻¹⁰).

Multi-omics characterization

To prioritize plausible causal variants within the LD-independent genome-wide significant loci, we conducted a fine-mapping analysis using PAINTOR¹¹, incorporating Bonferroni-significant functional annotations (Supplementary Table 6). For the EUR combined and cross-ancestry datasets, annotations included histone H3 trimethylated at lysine 4 (H3K4me3) and histone H3 acetylated at lysine 27 (H3K27ac) in uterine and ovarian tissues, super-enhancers in ovarian tissue, and differentially expressed enhancers in uterine tissue. For the EAS dataset, no annotations met the Bonferroni threshold, and fine-mapping was therefore performed without functional priors. Across ancestry-specific and cross-ancestry analyses of endometriosis definitions, we identified putative causal evidence (posterior probability, PP > 0.8) for 21 variants (Supplementary Table 7). Among these, rs28858973 (near AC003090.1) and VPS13B rs62532633 showed causal evidence in both EUR-specific analysis (PP = 0.91 and 1, respectively) and cross-ancestry analysis (PP = 0.95 and 1, respectively). Cross-ancestry analysis also identified rs7926666 (near ARL14EP), rs17428012 (near HOTTIP), and VEZT rs1964592 (PP = 1, 0.84, and 0.99, respectively), which showed suggestive evidence in EUR (PP = 0.79, 0.59, and 0.59, respectively). Among variants identified as causal considering EUR endometriosis clinical definition, rs12108558 (near RP11–530I17.1) showed suggestive causal evidence in EUR endometriosis combined definition (PP_EUR-clinical = 0.87; PP_EUR-combined = 0.70), while rs117914842 (near CDKN2B-AS1) showed suggestive evidence in the cross-ancestry analysis (PP_EUR-clinical = 0.9; PP_{cross-ancestry} = 0.60). In the EAS analysis, convergent suggestive evidence in EAS, EUR, and cross-ancestry combined definitions were observed for rs2501284 (near CDC42; PP = 0.59, 0.6, and 0.59, respectively). In total, suggestive causal evidence (PP > 0.5) was identified for 63 additional variants (Supplementary Table 7).

We identified 29 colocalization signals (PP of full colocalization, PPFC > 0.7) for 13 SNPs (PP for an individual SNP, PP_SNP > 0.5) showing shared causal effects on endometriosis and tissue-specific transcriptomic regulation (Supplementary Table 8). For the EUR clinical endometriosis definition, rs13100173 showed causal effect on endometriosis and HYAL3 transcription regulation in vagina (PPFC = 0.71, PP_SNP = 0.75), uterus (PPFC = 0.82, PP_SNP = 0.72), ovary (PPFC = 0.81, PP_SNP = 1), breast mammary tissue (PPFC = 0.81, PP_SNP = 0.85), sigmoid colon (PPFC = 0.81, PP_SNP = 0.94), and transverse colon (PPFC = 0.81, PP_SNP = 0.93). Broad cross-tissue colocalization was also observed for rs61926299 in the cross-ancestry endometriosis combined definition and ATF1 transcription regulation in ovary (PPFC = 0.71, PP_SNP = 0.56), breast mammary tissue (PPFC = 0.95, PP_SNP = 0.85), sigmoid colon (PPFC = 0.92, PP_snp = 0.58), and transverse colon (PPFC = 0.81, PP_SNP = 0.91), and LIMA1 in transverse colon (PPFC = 0.93, PP_SNP = 0.69). Convergent colocalization between endometriosis combined and clinical definitions was observed for rs9739834 with respect to VEZT transcriptomic regulation in sigmoid colon (combined definition: PPFC = 0.73, PP_SNP = 1; clinical definition: PPFC = 0.86, PP_SNP = 1). While the majority of the colocalization signals were observed for EUR and cross-ancestry associations, we also identified a colocalization for rs2388493 between EAS endometriosis and BAALC transcription regulation in sigmoid colon (PPFC = 0.81, PP_SNP = 0.56). To investigate possible cell-specific mechanisms, we used the Human Endometrial Cell Atlas¹², defining annotations related to endometrial and immune cell type-specific transcriptomic profiles. We observed nominally significant enrichment for decidual stromal cells and SOX9 functionalis for endometriosis SNP-h² (Supplementary Tables 9 and 10).

Colocalization analysis between endometriosis genetic associations and DNA methylation regulation identified shared causal effects (PPFC > 0.7 and PP_SNP > 0.5) across 568 SNP-CpG pairs (Supplementary Table 11). The majority of them (59%) overlapped among the different endometriosis definitions. For instance, rs4669752 mapping on GREB1 perfectly colocalized (PPFC = 1 and PP_SNP = 1) with methylation regulation of cg07314298, cg16908938, and cg13396044 considering cross-ancestry and EUR combined definitions and EUR clinical definition. Among the SNP-CpG pairs identified for only one endometriosis dataset, 114 of them were related to clinical endometriosis, with the strongest colocalization signal (PPFC = 1 and PP_SNP = 1) between rs635634 (upstream of ABO) and the methylation regulation of cg13506600, cg07241568, cg24267699, and cg14271713. We also identified 10 EAS-specific SNP-CpG colocalized pairs with six of them (rs17841114 colocalized with cg26397662 and cg09779576; rs74613634 colocalized with cg24363621, cg17067638, cg22835630, and cg04278444) located within SNHG14.

Colocalization analysis between endometriosis genetic association and plasma protein expression regulation identified shared causal effect (PPFC > 0.7 and PP_SNP > 0.5) linking rs11111182 to JUN and WASL proteins, with estimates that were consistent across cross-ancestry and EUR combined and clinical definitions (Supplementary Table 12). Considering a suggestive threshold (PPFC > 0.5), colocalization signals were observed for additional proteins (i.e., CCL2, DLK2, UNC5D, LYG2, ARL8B, FSHB, and GFAP).

Omics-wide analyses

We conducted a transcriptome-wide association study (TWAS) considering tissues implicated in endometriosis pathogenesis and symptomatology (i.e., uterus, ovary, and vagina), breast mammary tissue (due to shared hormone-responsive pathways), tissues potentially involved in endometriosis gastrointestinal manifestation (i.e., sigmoid and transverse colon), and whole blood (it can capture systemic and inflammatory pathways). This analysis identified 41 genes after Bonferroni multiple testing correction (P < 3.72 × 10⁻⁶; Fig. 3 and Supplementary Table 13). Top findings included a negative association of ARL14EP expression with endometriosis risk (P_multi-tissue = 1.36 × 10⁻⁴⁰; best tissue: breast mammary tissue), a positive association of CDC42 expression (P_multi-tissue = 3.76 × 10⁻²⁰; best tissue: colon sigmoid), and CCDC170 (P_multi-tissue = 4.33 × 10⁻¹⁷; best tissue: whole blood). Increased predicted expression of RBM6 was associated with a reduced risk of endometriosis in the multi-tissue analysis (P = 1.79 × 10⁻⁶), the primary tissues affected by the disease (P_uterus = 2.05 × 10⁻⁷; P_ovary = 5 × 10⁻⁷; P_vagina = 1.63 × 10⁻⁶), and whole blood (P = 3.17 × 10⁻⁶). Two additional genes identified in the multi-tissue TWAS showed positive associations with endometriosis (GMPPB: P_multi-tissue = 5.59 × 10⁻⁸, P_uterus = 1.92 × 10⁻⁷; CRISPLD1: P_multi-tissue = 9.26 × 10⁻⁸, P_uterus = 1.55 × 10⁻⁹). Among multi-tissue TWAS significant results, ovary- and vagina-specific positive associations with endometriosis were present for WASHC3 (P_multi-tissue = 2.11 × 10⁻⁹; P_ovary = 1.49 × 10⁻⁸; P_vagina = 6.7 × 10⁻⁸) and DTD1 (P_multi-tissue = 4.1 × 10⁻⁷; P_ovary = 1.47 × 10⁻⁸; P_vagina = 7.77 × 10⁻⁹). While they did not reach Bonferroni correction in the multi-tissue TWAS, we observed positive endometriosis associations of CISD2 (z = 4.78, P = 1.75 × 10⁻⁶) in the uterus-specific analysis, and lower expression of EIF2AK4 (z = −4.99, P = 6.05 × 10⁻⁷) and MSRA (z = −4.68, P = 2.81 × 10⁻⁶) in the vagina-specific analysis.

Figure 3 |. Transcriptome-wide (top) and proteome-wide (bottom) associations of endometriosis combined definition.

Each point represents the −log₁₀(P-value) from a two-sided z-test of association between the phenotype and the cis-genetically regulated expression level of a gene (TWAS) or plasma protein (PWAS), ordered by genomic position on the x axis. Black horizontal dashed lines indicate significance thresholds after multiple testing correction (P < 3.4 × 10⁻⁶ for TWAS; P < 3.6 × 10⁻⁴ for PWAS). In the top panel, point colors represent different tissues and are defined in the figure key, and gene labels correspond to the most significant tissue. TWAS was performed using S-PrediXcan and S-MultiXcan applied to endometriosis GWAS summary statistics (n_cases = 105,869 and n_controls = 1,282,731) derived from unrelated individuals. Tissue-specific gene expression models were drawn from GTEx v8 (uterus, n = 153; vagina, n = 170; ovary, n = 193; breast mammary tissue, n = 514; whole blood, n = 803; sigmoid colon, n = 419; transverse colon, n = 479); each GTEx donor represents an independent biological unit. PWAS was performed in individuals of European ancestry using pQTL imputation models for 1,318 plasma proteins trained in n = 7,213 independent individuals. No technical replicates were applied in either analysis.

Using protein quantitative trait locus (pQTL) weights derived from 1,318 plasma proteins¹³ to perform a proteome-wide association study (PWAS), we identified nine proteins whose genetically predicted plasma levels were significantly associated with endometriosis risk (FDR q < 0.05; Fig. 3 and Supplementary Table 14). The strongest proteome-wise association was observed for RSPO3, which showed a positive association with endometriosis in two distinct pQTL models (z = 7.46 and 7.38; P = 8.52 × 10⁻¹⁴ and 1.54 × 10⁻¹³, respectively). Other positive associations included SHBG (z = 5.38, P = 7.61 × 10⁻⁸), SAT2 (z = 4.43, P = 9.56 × 10⁻⁶), and PRSS57 (z = 3.57, P = 3.61 × 10⁻⁴). Negative associations included FN1 (z = −4.98, P = 6.36 × 10⁻⁷), CD109 (z = −4.88, P = 1.07 × 10⁻⁶), MANBA (z = −4.04, P = 5.42 × 10⁻⁵), DTD1 (z = −3.88, P = 1.06 × 10⁻⁴), and CYB5D2 (z = −3.76, P = 1.69 × 10⁻⁴).

In EUR, blood-based SMR meta-analysis using data derived from six cohorts identified 305 CpG sites associated with endometriosis (FDR q < 0.05, HEIDI P > 0.05), including 242 for the combined (Supplementary Table 15), 62 for the clinical (Supplementary Table 16), and 49 for the self-reported definitions (Supplementary Table 17). Five CpG sites within three loci reached FDR-significant associations without evidence of heterogeneity across the three definitions: cg10740902 and cg07098391 mapping on KDR; cg09995736 near VEZT; cg08907436 and cg17706972 mapping on ESR1. The latter associations are located within a genomic region that includes 12 additional FDR-significant CpG sites (top-result: cg23164938, P_combined = 5.49 × 10⁻¹¹). In the MHC region (chromosome 6: 29,600,203–31,129,828), there were 29 convergent FDR-significant effects (top-CpG association: cg14574237, P = 6.63 × 10⁻⁷; 14 related to endometriosis combined definition and 15 to self-reported endometriosis with five CpG sites shared between the two). Another locus with convergent SMR results was observed on chromosome 12 (25,886,475–27,260,441) mapping to several non-coding genes and three protein-coding genes (CBX3, SNX10, HNRNPA2B1, and NFE2L3), where 23 CpG associations related to the three endometriosis definitions (top-result: cg12942962, P_combined = 1.25 × 10⁻⁸). Because of the limited power of the adenomyosis SMR analysis, we identified only three FDR-significant CpG associations that were also identified in the endometriosis analyses: cg11466449 (P_adenomyosis = 1.37 × 10⁻⁶) mapping to ENSG00000289405 RNA gene and cg12868707 (P_adenomyosis = 1.83 × 10⁻⁸) and cg22105613 (P_adenomyosis = 7.23 × 10⁻⁷) both mapping to GDAP1 (Supplementary Table 18). The latter CpG site was also identified in combined (P = 4.88 × 10⁻¹⁰) and clinical (P = 9.63 × 10⁻⁷) endometriosis. While we did not identify FDR significant results in the EAS blood-based SMR analysis, nominally significant replication of EUR findings was observed for eight CpG sites (combined definition: cg00814883, cg02971253, cg02060096, cg07227024, cg09678539; clinical definition: cg14396995, cg13172367; self-reported definition: cg16650630).

In the endometrium-specific SMR, we identified 57 CpG sites associated with endometriosis (FDR q < 0.05; Supplementary Table 19). In the endometriosis combined definition, the top finding was cg03519931 mapping to WNT4 (P = 1.65 × 10⁻¹⁴). This CpG site was also FDR significant for the endometriosis clinical definition (P = 7.22 × 10⁻¹⁰) and adenomyosis (P = 1.43 × 10⁻⁵). Also, the other two CpG sites identified in the adenomyosis SMR analysis (cg13258528, P = 7.87 × 10⁻⁶; cg10414477, P = 1.96 × 10⁻⁵) were also significantly associated with endometriosis (P_{self-reported} = 2.98 × 10⁻¹⁰ and 1.03 × 10⁻⁵, respectively). Although there was consistency across endometriosis definitions and between endometriosis and adenomyosis for most associations, we also observed 20 CpG sites that reached FDR significance for endometriosis clinical definition but were null for self-reported endometriosis (P > 0.05; Supplementary Table 19). The top clinical vs. self-reported difference was cg05678033 (P = 1.31 × 10⁻⁶ vs. 0.239). Conversely, only two CpG sites were FDR significant for self-reported endometriosis but null for endometriosis clinical definition (cg06488957 mapping on RAI1 and cg22228116 mapping on HLCS; Supplementary Table 19). Regarding cross-tissue convergence, cg24126931 mapping on HCG2P7 was FDR significant in blood (P_combined = 5.69 × 10⁻⁵; P_{self-reported} = 4.03 × 10⁻⁵) and endometrium (P_combined = 6.41 × 10⁻⁵; P_{self-reported} = 4.73 × 10⁻⁶).

A summary of genes identified across multi-omics characterization and omics-wide analyses is available in Figure 4 and Supplementary Table 20.

Figure 4 |. Multi-omics convergence across endometriosis-associated loci.

Integration of GWAS, colocalization (eQTL, mQTL, pQTL), transcriptome-wide (TWAS), proteome-wide (PWAS), and summary-based Mendelian randomization (SMR-mQTL) analyses. Genes are grouped by major functional categories, including cellular morphogenesis, steroid and endocrine regulation, immune regulation, inflammation, and vascular development and remodeling. Some graphical elements and icons were created using Canva Pro (https://www.canva.com/).

Drug repositioning

We applied DRUGSETS¹⁴, a genetically informed drug repositioning framework, to identify therapeutic options for endometriosis. Considering EUR gene-based results related to combined endometriosis, we found two Bonferroni-significant drugs used for breast cancer treatment, neratinib and toremifene (P < 6.8 × 10⁻⁵; Supplementary Table 21). The cross-ancestry analysis identified three additional Bonferroni-significant candidates: progestins 17-hydroxyprogesterone-caproate and norgestimate and abemaciclib used for breast cancer treatment (P < 6.8 × 10⁻⁵; Supplementary Table 22). The enrichment signals related to these drugs were due to the significant association of ESR1 (P_{EUR gene-based} = 4.89 × 10⁻¹²; P_{cross-ancestry gene-based} = 1.01 × 10⁻¹³) and PGR (P_{EUR gene-based} = 5.19 × 10⁻⁵; P_{cross-ancestry gene-based} = 3.39 × 10⁻⁶) with endometriosis. In addition to these two genes, KDR (P_{EUR gene-based} = 2.56 × 10⁻¹¹; P_{cross-ancestry gene-based} = 7.46 × 10⁻¹³) contributed to neratinib enrichment, and CDKN2A (P_{cross-ancestry gene-based} = 1.95 × 10⁻⁶) and TP53 (P_{cross-ancestry gene-based} = 1.27 × 10⁻³) to abemaciclib enrichment.

We also used the Gene2drug approach¹⁵ to assess the overlap of endometriosis-enriched Gene Ontology (GO) terms (Supplementary Note and Supplementary Tables 23 and 24) with the transcriptomic profiles of molecular compounds available from the Connectivity Map¹⁶. While no result survived Bonferroni multiple testing correction (P < 3.82 × 10⁻⁵; Supplementary Table 25), we observed suggestive evidence (FDR q < 0.1) for thiostrepton (Enrichment Score (ES) = 0.2, P = 8.29 × 10⁻⁵), nortriptyline (ES = 0.2, P = 1.17 × 10⁻⁴), and verteporfin (ES = −0.2, P = 1.75 × 10⁻⁴).

Polygenic risk scoring

EUR endometriosis-PRS derived from non-overlapping cohorts was significantly associated with AoU-endometriosis (odds ratio (OR) = 1.38, 95%CI = 1.35–1.41) and UKB-endometriosis (OR = 1.3, 95%CI = 1.27–1.34) in EUR, and also with endometriosis across different population groups in AoU. Compared to individuals in the lowest PRS quartile, those in the highest quartile showed increased odds of endometriosis in EUR-AoU (OR = 2.19, 95%CI = 2.07–2.32), AMR-AoU (OR = 1.77, 95%CI = 1.54–2.04), EAS-AoU (OR = 2.75, 95%CI = 1.81–4.27), AFR-AoU (OR = 1.44, 95%CI = 1.23–1.69), CSA-AoU (OR = 1.92, 95%CI = 1.06–3.55), and MID-AoU (OR = 3.68, 95%CI = 1.27–12.20) (Fig. 5 and Supplementary Table 26). In addition to being affected by the sample size of the target samples, the magnitude of cross-ancestry PRS effects generally reflected the genetic distance between EUR and the other population groups.

Figure 5 |. Within-ancestry and cross-ancestry PRS associations of endometriosis.

logORs (centre) and corresponding 95% confidence intervals (error bars) are reported. Cases were defined by clinical or self-reported diagnosis, and controls were individuals with no recorded diagnosis of the respective condition. In the within-ancestry analysis, the EUR-endometriosis PRS (excluding EUR-AoU from the training sample, n_cases = 87,792 and n_controls = 980,864) was tested against the EUR-AoU sample (n_cases = 11,615 and n_controls = 112,670). In the cross-ancestry analysis, EUR-endometriosis PRS (with EUR-AoU in the training sample, n_cases = 99,407 and n_controls = 1,093,534) was tested against AoU samples of AFR (n_cases = 1,786 and n_controls = 35,470), AMR (n_cases = 1,731 and n_controls = 37,251), CSA (n_cases = 107 and n_controls = 2,358), EAS (n_cases = 217 and n_controls = 5,002), and MID (n_cases = 34 and n_controls = 655) descent.

We performed interaction analyses between endometriosis-related symptoms and the EUR endometriosis-PRS in AoU and UKB (Supplementary Table 27). Specifically, we tested whether the presence of individual symptoms modifies the effect of the endometriosis PRS on endometriosis risk. In AoU, results surviving FDR multiple comparisons included PRS × abdominal pain (OR = 0.94, 95%CI = 0.91–0.98), PRS × anxiety (OR = 0.93, 95%CI = 0.89–0.96), PRS × migraine (OR = 0.94, 95%CI = 0.91–0.98), PRS × nausea (OR = 0.90, 95%CI = 0.87–0.95), PRS × comorbidity burden (OR = 0.99, 95%CI = 0.98–0.99) in AoU (Fig. 6 and Supplementary Table 28). All showed attenuation of the PRS effect in individuals reporting the symptom. The AoU interactions were not statistically replicated in UKB.

Figure 6 |. Interaction effects between endometriosis PRS and endometriosis symptoms and comorbidities in AoU and UKB.

Forest plot illustrating the OR (centre) and 95% confidence intervals (error bars) for the interaction between endometriosis PRS and its various symptoms and comorbidities in two independent cohorts: AoU (circles; n_cases = 11,615 and n_controls = 112,670) and UKB (squares; n_cases = 6,611 and n_controls = 194,000). Cases were defined by clinical or self-reported endometriosis diagnosis, and controls were individuals with no recorded endometriosis diagnosis. Each row represents a different symptom or comorbidity, with the x-axis showing the OR for the PRS × trait interaction on endometriosis risk, estimated using a two-sided logistic regression model adjusted for age and the top 10 within-ancestry principal components. FDR multiple testing correction (q < 0.05) was performed for the 22 traits. Cohorts are indicated by point shape and significance levels by point color, as defined in the figure key.

Our phenome-wide pleiotropy analysis identified a wide range of genetic correlations, genetic overlaps, and potential cause-effect relationships linking endometriosis and adenomyosis with physical and mental health outcomes (Supplementary Note, Supplementary Figs. 1-3, and Supplementary Tables 29-34). Briefly, correlations were strongest with menstrual irregularities, infertility, ovarian cysts, uterine leiomyoma, MDD, and anxiety. Adenomyosis showed pleiotropy patterns only partially overlapping with endometriosis. For instance, only adenomyosis was genetically correlated with endometrial cancer. Genetically informed causal inference analyses supported endometriosis effect on reproductive traits, pain, and infertility, while chronic fatigue showed a causal effect on endometriosis.

DISCUSSION

This study represents the largest multi-ancestry endometriosis GWAS to date, including approximately 1.4 million women, of whom 105,869 are cases, nearly doubling the number of cases included in previous GWAS^4,7. The cross-ancestry analysis including six global population groups increased the number of genome-wide significant loci for endometriosis from 63 in the EUR-only analysis to 80, of which 37 are novel. Our GWAS meta-analysis of adenomyosis identified five genome-wide significant loci, with two of them novel (HRH1 and GDAP1) compared to a previous FinnGen study¹⁷. Genetic correlation analyses showed that endometriosis and endometriosis without adenomyosis had lower correlations with adenomyosis (r_g = 0.79 and 0.81, respectively) than between different endometriosis definitions, supporting the differences between these conditions¹⁸. However, in addition to limiting gene discovery, sample size differences between endometriosis and adenomyosis GWAS may have affected some of the difference observed in the follow-up analyses investigating molecular mechanisms and pleiotropic relationships.

Five loci were identified as genome-wide significant only in clinically confirmed endometriosis, despite the smaller sample size of this group. The previously found ZHX3 was identified only for clinical endometriosis, suggesting that the data source for endometriosis phenotyping can drive the specificity of gene discovery, as also reported before^4,6. This likely reflects more severe and symptomatic forms of the disease¹⁹, as women correctly self-report endometriosis about 84% of the time¹⁹. Our genetic correlation analysis showed a high r_g = 0.85 between clinical and self-reported endometriosis, consistent with phenotypic data, supporting the use of self-reported information in large-scale endometriosis studies¹⁹. We also demonstrate that European-derived PRS is statistically associated with endometriosis across population groups. While its predictive power is currently insufficient for diagnostic use, we highlight the need to develop ancestry-informed risk models and integrate genetic risk with clinical features.

In line with a previous endometriosis GWAS⁴, some loci (e.g., SYNE1 and GDAP1) showed effect size differences across endometriosis definitions that could reflect their relationship with disease severity or diagnostic heterogeneity. Among the novel loci, HRH1 rs184700*T allele was associated with increased adenomyosis and reduced endometriosis. While recent studies highlighted the potential link of HRH1 function to pain sensitivity and inflammatory processes related to endometriosis etiology²⁰, further research needs to focus on how HRH1 differentially impacts endometriosis and adenomyosis and how this contributes to the ongoing debate on whether these represent distinct diseases rather than different presentations of the same condition¹⁸.

In addition to increasing gene discovery power, the investigation of six ancestry groups permitted us to assess potential differences in the predisposition to endometriosis among worldwide populations. Comparing ancestry-specific effect size, we observed that HMGN1P19 rs993183*A and SYN3 rs10854627*A alleles were more strongly associated with endometriosis in EAS and EUR. While these loci have been previously associated with endometriosis^4,7, their role in the disease pathogenesis is still unclear. Our cross-ancestry GWAS meta-analysis found that seven of the 80 endometriosis-associated loci present genome-wide significant heterogeneous effects across population groups (i.e., WNT4 rs55938609, rs1430787, GREB1 rs1865573, KDR rs10517343, SYNE1 rs58415480, FSHB rs11031005, and VEZT rs4762308). This heterogeneity further supports the idea that, while endometriosis pathogenesis is expected to converge across individuals of diverse ancestral backgrounds, genetic variability may partially contribute to differences across worldwide regions.

Our integrative multi-omic analyses provide compelling evidence for shared genetic mechanisms influencing endometriosis risk through tissue-specific regulation of gene expression, protein abundance, and DNA methylation (Fig. 4). Convergent results across all six functional analyses confirmed a robust role for SEMA3F in endometriosis pathogenesis, consistent with prior GWAS⁷ and molecular²¹ findings. Additional genes implicated in cellular morphogenesis confirmed by three methods included EEFSEC²², UBE2D3²³, DTD1²⁴, and VEZT²⁵, with UBE2D3 representing a novel GWAS association. Although the remaining loci were previously identified by endometriosis GWAS analyses^4,7, our results offer the first detailed insights into their regulatory mechanisms. DTD1 showed a positive association with endometriosis in ovary- and vagina-specific TWAS, but a negative association in plasma-specific PWAS. Although it was not statistically significant, the blood-specific TWAS DTD1 effect on endometriosis was concordant with the one observed in the plasma-specific PWAS. This suggests tissue-specific differences in the mechanisms by which this locus contributes to endometriosis pathogenesis. Among other previously known genes, we observed extensive functional evidence for GREB1²⁶ and WNT4²⁷ (involved in steroid and endocrine regulation) and SKAP1²⁸ (involved in immune regulation), highlighting their broad regulatory roles beyond genetic association alone. We also identified novel GWAS and multi-omic hits supported by at least two independent approaches, including HOXA3²⁹, FN1³⁰, TMEM194A³¹, CALD1³², SPECC1³³, TFPI³⁴, XKR6³⁵, and ADK³⁶, implicating pathways in cellular morphogenesis, vascular development and remodeling, immune regulation, and inflammation. The functional categorization of implicated genes aligns with the main pathophysiological theories of endometriosis, including benign metastasis and impaired apoptosis, hormonal imbalance, immune dysfunction, chronic inflammation, and vascular dissemination. Many genes, especially those in the overrepresented cellular morphogenesis category, are also implicated in several cancers, highlighting breast cancer^37–41. In EAS, rs2388493 colocalized with BAALC transcriptomic regulation in sigmoid colon. To our knowledge, this is the first functional characterization of an endometriosis locus in a non-EUR population. Although BAALC is primarily studied in leukemia, its role in regulating cell proliferation and differentiation⁴² may point to altered stromal or epithelial remodeling as an EAS-specific mechanism in endometriosis pathophysiology. Our multi-omics characterization of endometriosis-associated loci was limited by reference datasets used to perform to estimate genetically regulated omic variation. In particular, most of them included both females and males, which may mask some female-specific regulatory mechanisms. Additionally, because endometriosis is driven by complex interactions among different tissues and organs, regulatory effects may be highly cell type-specific or influenced by hormonal or inflammatory states. Accordingly, future studies incorporating sex- and context-specific omic datasets will be essential for clarifying the molecular mechanisms of endometriosis and enhancing the resolution of genomic investigations.

Recent studies have highlighted new therapeutic targets and repurposing opportunities for endometriosis, including immune-related genes (e.g., GBP2, HCK)⁴³. Our genetically informed analyses identified five repurposed drugs, including three breast cancer agents (neratinib, toremifene, abemaciclib). Toremifene, a selective estrogen receptor modulator, has been studied for its effects on the uterus and vagina, demonstrating similar estrogenic activity in both tissues and raising the possibility of off-target impacts on endometrial tissue⁴⁴. Neratinib, an EGFR/HER2 inhibitor, has been shown to be a potential treatment option for certain patients with endometrial cancer⁴⁵. Furthermore, certain aromatase inhibitors (e.g., letrozole and anastrozole) have been shown to reduce pain and lesion size in postmenopausal endometriosis⁴⁶. The two progestins (17-hydroxyprogesterone caproate and norgestimate) are not currently used for endometriosis treatment but may help manage symptoms directly or indirectly related to the condition, such as heavy menstrual bleeding or preterm birth. The link of these drugs to endometriosis genetics was driven by the effect of ESR1 and PGR loci, which support the role of estrogen and progesterone signaling in the disease pathophysiology⁴⁷. The contribution of KDR to neratinib enrichment highlights the potential of targeting angiogenesis and vascular development as a therapeutic target for endometriosis⁴⁸. Similarly, CDKN2A and TP53 involvement in endometriosis-abemaciclib relationship suggests cell-cycle regulation⁴⁹ as a druggable target. The drug-repurposing analysis based on GO enrichments identified additional molecular compounds. The identification of thiostrepton converges with experimental evidence from animal models reporting its potential in inhibiting endometriotic lesions⁵⁰. Nortriptyline is an antidepressant also used to treat chronic pelvic pain off-label⁵¹, suggesting the importance of pain neuromodulation in endometriosis treatment. Finally, verteporfin has been linked to increased progestin sensitivity in cellular models⁵², which is often effective in the symptomatic management of endometriosis⁵³. The different targets identified by the drug-repurposing analyses likely reflect the underlying mechanisms captured by each method. DRUGSETS findings represent targets related to protein physical properties, whereas Gene2Drug results indicate targets associated with transcriptomic changes. Greater convergence between these methods is anticipated as the statistical power of endometriosis GWAS increases and the informativeness of experimental platforms for sex-specific biological changes improves.

Our symptom-based PRS interaction analyses suggest that genetic risk may attenuate in individuals with high systemic symptom burden and abdominal pain, anxiety, migraine, and nausea. However, these were observed only in AoU cohort, but were not replicated in the UKB. In line with our findings, a recent study reported that among women with endometriosis, those with a higher comorbidity burden had lower genetic predisposition⁵⁴. Hence, in the context of extensive symptomatology, non-genetic factors may play an important role in disease manifestation. However, symptom data were derived from both self-reported and EHR data, reflecting lifetime experiences rather than symptoms at recruitment, which may introduce temporal variability given the diagnostic delay typical of women’s health conditions. In addition to symptom-specific genetic interactions, our findings highlight broader genetic relationships of endometriosis and adenomyosis with reproductive, psychiatric, systemic, and pain-related traits. These results illustrate the systemic nature of these conditions, emphasizing shared genetic mechanisms with comorbidities, some of which may be causal. Notably, the causal effect of chronic fatigue on endometriosis is a novel finding. Chronic fatigue affects approximately 50% of individuals with endometriosis and significantly impacts quality of life⁵⁵. This comorbidity may share neuroimmune and neuroendocrine pathways with endometriosis, further reflecting the systemic nature of the disease⁵⁵.

Our findings offer genetic and molecular support for several leading theories of endometriosis pathogenesis, including benign metastasis, dysregulated immune system, cell cycle and differentiation, impaired apoptosis, inflammation, hormonal imbalance, vascular dissemination, and epigenetic alterations⁵⁶. These results suggest that the complex pathophysiology of endometriosis may be due to the interplay of several biological mechanisms, rather than a single causal pathway, highlighting the need for integrative approaches to fully understand this complex disease.

METHODS

Study populations

The GWAS meta-analysis included eight cohorts: UKB⁵⁷, AoU⁵⁸, MVP⁵⁹, Biobank Japan (BBJ)⁶⁰, International Endometriosis Genetics Consortium (IEGC)⁴, FinnGen¹⁷, EstBB⁶¹, and 23andMe ⁶² (Supplementary Table 1). As previously collected, deidentified data were used in the present study, no ethical approval was needed⁶³. In particular, we leveraged MVP, BBJ, IEGC, FinnGen, EstBB, and 23andMe genome-wide association statistics already available and generated novel genome-wide information from individual-level data available from UKB (AFR, AMR, CSA, EAS, MID ancestries) and AoU (all ancestries). The present study was additionally approved by the Norwegian Regional Committees for Medical and Health Research Ethics (REK 467496). The original cohorts were all approved by their institutional review boards, and all participants provided written informed consent. In total, our cross-ancestry meta-analysis included 105,869 cases and 1,282,731 female controls (Supplementary Table 1).

UKB is a large population-based study that collected genetic, health, and lifestyle information from over 500,000 participants across the United Kingdom⁵⁷. Self-reported (field code 1402) and clinical endometriosis (ICD-10: N80) were included in the analyses. Nine GWAS analyses were performed: ancestry-specific GWAS analyses for AFR, AMR, CSA, EAS, MID, and EUR from medical records, EUR from self-reported phenotype information, EUR with only adenomyosis cases (“endometriosis of uterus”, ICD-10: N80.0), and EUR endometriosis cases without adenomyosis (ICD-10: N80.1-N80.9; Supplementary Table 1). The EUR GWAS of endometriosis combined definition (based on EHRs and self-reported information) was included in the IEGC meta-analysis, hence it was not performed. In each ancestry group, pairwise relatedness was calculated, and one individual from each related group was retained prior to association testing. The PLINK2 (version 2.0)⁶⁴ Firth regression (i.e., a penalized method that reduces bias when data are sparse or unbalanced) was used for each GWAS to estimate genetic associations with covariates including age and the first ten within-ancestry principal components (PCs). The following quality control criteria were applied: biallelic variants with minor allele frequency (MAF) > 1%, Hardy-Weinberg equilibrium P < 10⁻⁶, variant call rate > 95%, and per-individual genotyping rate > 95%. Genetically inferred ancestry assignment was derived from the Pan-UKB Initiative⁶⁵.

The AoU Research Program is an ancestrally diverse US-based precision medicine initiative aiming to enroll over one million participants to reflect the diversity of the population, with rich genomic, clinical, and survey data⁵⁸. Ten GWAS analyses were performed using the same criteria as for UKB, with the addition of the EUR combined endometriosis definition (based on EHRs and self-reported information). Genetic associations were estimated using PLINK2 (version 2.0)⁶⁴ logistic regression, including age and the first ten within-ancestry PCs as covariates. The same quality control criteria were applied as in the UKB analysis. AoU genetically inferred ancestry assignment was previously described⁵⁸.

MVP is a U.S. Department of Veterans Affairs cohort that includes genetic and health data from over 900,000 U.S. military veterans to study the genetic basis of diseases⁵⁹. They generated the GWAS data from 16,231 and 32,192 female individuals of AFR and EUR ancestry, respectively. The genetic association analysis was performed using a PLINK 2⁶⁴ logistic regression model with covariates for age, sex, and the first ten within-ancestry PCs. Details of MVP GWAS have been previously described⁶⁶.

BBJ is a hospital-based biobank that collected DNA and clinical data from over 200,000 participants across Japan, focused on a wide range of common diseases⁶⁰. They conducted the GWAS by employing a generalized linear mixed model (GLMM) using SAIGE⁶⁷, age, and top 5 PCs were used as covariates, and it was generated from 82,761 female individuals of EAS ancestry.

The IEGC is a collaborative effort that aggregates data from multiple international studies to investigate the genetic factors of endometriosis⁴. They generated GWAS data from 474,160 individuals, of which 3,294 were of EAS ancestry. This meta-analysis included 23 cohorts with variable endometriosis case definitions including EHRs, surgically confirmed endometriosis, and self-reported information, and most studies applied logistic regression. The covariates used varied: several analyses included no adjustment, while others adjusted for age, sex, and a range of PCs, from using the first two only and up to 20. The meta-analysis was performed using the inverse variance weighting fixed-effects method implemented in METAL⁶⁸.

FinnGen is a public-private partnership combining genetic data from over 500,000 Finnish biobank participants with nationwide health registry data to advance understanding of disease genetics¹⁷. The FinnGen GWAS was performed using REGENIE⁶⁹, including covariates for age, sex, the top ten within-ancestry PCs, and genotyping batch. The GWAS data for endometriosis (ICD10: N80) and adenomyosis (ICD10: N80.0) were derived from Release 12, including 150,350 and 135,833 EUR individuals of EUR ancestry, respectively.

EstBB is a population-based cohort comprising over 200,000 individuals from Estonia, with comprehensive genetic and health data linked to national health records⁶¹. They performed the GWAS using REGENIE⁶⁹, including covariates for year of birth and the top ten within-ancestry PCs. Endometriosis GWAS data (ICD10: N80) were generated from 126,807 individuals of EUR ancestry.

23andMe is a direct-to-consumer genetics and research company with a cohort of research-consented customers, totaling millions of individuals, contributing genetic and self-reported health data to research⁶². The 23andMe endometriosis GWAS was performed using logistic regression models with current age, the top five PCs, and the genotyping platform as covariates⁶².

Given the different endometriosis phenotype definitions, we calculated genetic correlation among them using Linkage Disequilibrium Score Regression (LDSC)⁷⁰. LD scores were calculated using HapMap 3 variants⁷¹ and the 1000 Genomes Project EUR reference population¹⁰. These analyses were only performed in EUR due to the low sample size of other ancestry-specific GWASs.

Ancestry-specific and cross-ancestry meta-analyses

Before meta-analyzing the different datasets, the following quality control steps were applied: MAF > 1%, imputation INFO score > 0.8 when available, and exclusion of non-rs IDs and non-biallelic variants. For each meta-analysis, the fixed-effects inverse variance weighting method implemented in METAL (version 3)⁶⁸ was used, based on sample sizes and P-values to account for possible sample overlap, resulting in 39,620,082 SNPs. After meta-analysis, we excluded variants that were not present in more than 50% of the total effective sample size (n = 195,595)⁷², resulting in 7,865,243 remaining SNPs. Apart from the cross-ancestry meta-analysis, (n_cases = 105,869; n_controls = 1,282,731), additional meta-analysis were conducted for AFR (n_cases = 2,481; n_controls = 54,671), AMR (n_cases = 1,752; n_controls = 37,871), EAS (n_cases = 2,044; n_controls = 87,723), EUR (n_cases = 99,407; n_controls = 1,093,534), EUR clinical (n_cases = 36,695; n_controls = 567,775 controls), EUR self-reported (n_cases = 50,867; n_controls = 528,391), adenomyosis (n_cases = 8,753; n_controls = 415,718), and endometriosis without adenomyosis (n_cases = 6,627; n_controls = 283,316). LDSC metrics, including SNP-h², genomic inflation (lambda), mean chi-squared statistic, LDSC intercept, and confounding ratio, were calculated for each meta-analysis. FUMA⁷³ was used to define genome-wide significant variants.

We performed a joint analysis using GCTA-COJO (version 1.94.1)^74,75 to identify secondary association signals in the loci reaching genome-wide significance for cross-ancestry, EUR, and EAS GWAS of endometriosis combined definition and of EUR clinical definition. A stepwise model selection procedure was applied to identify independently associated SNPs. We used UKB ancestry-specific genome-wide data as the linkage disequilibrium (LD) reference panel. The default parameters were considered: 10 Mb for LD distance; 0.9 for multiple regression R² cutoff for collinearity; 0.2 for difference in allele frequency between the GWAS and LD reference. SNPs with joint-analysis P-value (P_J) < 5 × 10⁻⁸ were considered genome-wide significant. Among the SNPs identified by COJO, we further excluded those with LD r² ≥ 0.1 with any lead SNPs identified in the primary GWAS.

Functional characterization

Fine-mapping and colocalization.

To identify causal variants underlying the genome-wide significant associations, we conducted fine-mapping using PAINTOR (version 3), a probabilistic framework that incorporates functional annotations to model their effects on the magnitude of SNP associations and prioritize likely causal variants⁷⁶. We analyzed LD-independent endometriosis-associated loci in EUR, EUR-clinical, EAS, and cross-ancestry datasets using a ± 1-Mb window around the lead SNP at each locus. To select relevant functional annotations, we first estimated the enrichment of endometriosis SNP-h² in EUR, EUR-clinical, EAS, and cross-ancestry GWAS using partitioned SNP-h² analysis via LDSC and the baseline annotation set⁷⁷. Among the Bonferroni-significant annotations (n_tests = 97, P < 5 × 10⁻⁴), we retained those for which endometriosis-specific functional data were available in PAINTOR. To address multicollinearity among annotations, we clustered the annotation matrices separately for each ancestry-specific summary statistic. From each cluster, we selected a representative annotation, prioritizing distinct functional categories and maximizing dissimilarity from other clusters where possible. Clustering and variable selection were performed in R using the klaR package. Fine-mapping was conducted without pre-specifying the number of causal variants, using PAINTOR’s Gibbs sampling algorithm to sample from the posterior distribution directly.

Colocalization.

We conducted colocalization analysis to identify shared causal SNPs between endometriosis GWAS and transcriptomic, proteomic, and methylation regulation. Expression QTL (eQTL) data from six endometriosis-relevant tissues (i.e., uterus, ovary, vagina, breast mammary, and sigmoid and transverse colon) were derived from GTEx (v8)⁷⁸. For pQTLs, we used summary statistics from UKB and FinnGen. UKB includes 54,219 participants (mean age: 56.7 years (SD: 8.1), 54.3% female), mapping 2,923 proteins measured using the OLINK technology. FinnGen includes two platforms: OLINK (619 samples across 2,925 proteins) and SomaScan (828 samples across 7,596 proteins). Age and sex distribution data were not available. For mQTLs, we used summary data from Hatton et al.⁷⁹, comprising 3,701 EUR ancestry individuals from three cohorts and 2,099 EAS ancestry individuals from two cohorts, yielding 7,141 conditionally independent cis-mQTL variants. Information on age and sex was not reported. For each QTL dataset, we used HyPrColoc (version 0.0.2) deterministic Bayesian algorithm to estimate the PPFC⁸⁰. The colocalization analysis was conducted for cross-ancestry, EUR, and EAS GWAS of endometriosis combined definition and EUR clinical definition. We used PPFC > 0.7 as the threshold for identifying colocalization signals⁸⁰. In addition, we prioritized candidate causal SNPs with the threshold of proportion of the PPFC explained by the candidate causal variant (PP_SNP > 0.5). Estimates of regional association probability (P_R) were also calculated.

Partitioned heritability.

For combined, clinical, and self-reported endometriosis and adenomyosis, we performed partitioned heritability analyses using S-LDSC⁷⁷, testing the enrichment of tissue- and cell-type–specific transcriptomic profiles. We used single-cell RNA-seq (scRNA-seq) data from the Human Endometrial Cell Atlas⁸¹, which included 22 endometriosis patient samples and 13 healthy control samples (https://www.ebi.ac.uk/biostudies/ArrayExpress/studies/E-MTAB-14039/sdrf). For each cell type (e.g., epithelial, stromal, immune), gene expression in that cell type was compared against all other cell types separately in cases and controls to identify active genes. Genes were selected based on an absolute log₂ fold-change > 1.2 and FDR < 0.01⁸². These gene lists were then converted into LDSC annotations. Endometrial and immune cells were analyzed separately, as immune cells showed fewer detectable marker genes but represent a biologically relevant compartment for endometriosis. All annotations and LDSC results were cross-checked using CAD as a reference to validate the approach.

Transcriptome- and proteome-wide analyses.

Tissue-specific and cross-tissue TWAS of EUR endometriosis combined definition were performed using S-PrediXcan and S-MultiXcan approaches (version 0.8.1)^83,84. Leveraging GTEx v8 data⁷⁸, tissue-specific predictive models of gene expression and their corresponding LD covariance matrices were previously calculated (available at https://predictdb.org/categories/downloads/). We restricted the TWAS to seven tissues relevant to the pathogenesis of endometriosis and its symptoms: uterus (n = 153, 100% female), vagina (n = 170, 100% female), ovary (n = 193, 100% female), breast mammary tissue (n = 514, 35% female), whole blood (n = 803, 33% female), and sigmoid and transverse colon (n = 419, 36% female and n = 479, 36% female, respectively). Donors’ ages were 8.6% between 20–29 years, 7.9% between 30–39 years, 15.3% between 40–49 years, 32.2% between 50–59 years, and 35.9% between 60–70 years. Bonferroni multiple testing correction was applied to account for the number of genes tested (n = 13,457; P < 3.72 × 10⁻⁶).

We conducted a PWAS using the EUR GWAS meta-analysis to identify genes whose cis-regulated plasma protein abundance is associated with endometriosis risk. The analysis was performed using FUSION (version 1)⁸⁵, which integrates genome-wide association statistics with pre-trained protein imputation models. We specifically used the elastic-net (enet) models, due to their higher prediction accuracy compared to other approaches. We used pQTL weights derived from the plasma proteome study conducted by Zhang et al.¹³, which includes imputation models for 4,657 plasma proteins trained in 7,213 individuals of EUR ancestry (52% female) with approximately 25% aged ≤55, 26% aged 56–60, 25% aged 61–65, and 23% aged >65 years¹⁰. From these, imputation models were available for 1,318 plasma proteins with significant SNP-h² (P < 0.01) and were included in our PWAS analysis. The European subset of the 1000 Genomes Project¹⁰ was used as the reference panel. We applied a false discovery correction (FDR q < 0.05) accounting for the number of proteins tested.

mQTL-based SMR.

Because methylation models for omic-wide analyses are not available, we used the SMR approach (version 1.4.0)⁸⁶ to assess the effect of genetically regulated methylation changes on endometriosis and adenomyosis. Specifically, we integrated information regarding mQTL in blood (up to 256,524 CpG sites) and endometrial tissue (3,034 CpG sites) with 410,909 unique cis-SNPs available from genome-wide association statistics generated by the present study. The analysis was performed using ancestry-stratified information using matching reference populations available from the 1000 Genomes Project Phase 3. For EUR, blood mQTL datasets included the GoDMC catalog⁸⁷, comprising 36 cohorts with 27,750 EUR participants (53% female, mean age 42.5 years, range 8.3–79.1); the ARIES project⁸⁸, a pregnancy–mother–child cohort with 100% females (mean age 29.2 years, SD 4.4); Hatton et al.⁷, comprising five cohorts with 3,701 European ancestry individuals from three cohorts and 2,099 EAS individuals from two cohorts and 1,925 plasma proteins, with no sex or age reported; McRae et al.⁸⁹, including 1,980 individuals, with no data on sex or age; and Hannon et al.^90,91, with 1,193 individuals aged 28–98 years (55.9% female). In EUR, we also investigated an endometrium-specific mQTL dataset including data from 984 female participants with median age of 34.3 (range: 18–69)⁹². No other population-specific mQTL dataset was available for the other ancestry groups. SMR analyses were limited to cis-mQTLs (± 2 Mb of each CpG site), with thresholds applied for SMR testing (P < 5 × 10⁻⁸), LD pruning (r² between 0.05 and 0.9), and the HEIDI test (P > 0.05 using up to 20 top cis-SNPs). To enhance statistical power, in EUR blood-specific SMR results were meta-analyzed using sample size-weighted METAL⁶⁸, applying genomic control correction and sample overlap adjustment enabled via Z-statistic–based estimation. HEIDI test P > 0.05 in each of the datasets meta-analyzed was considered to exclude biases due to heterogeneity. False discovery rate was applied to account for the number of CpG sites tested (FDR q < 0.05).

Gene Ontology enrichment

To characterize endometriosis pathogenesis, we performed a gene-set enrichment analysis using the GSA-MiXeR framework (version 2.2.1). This allows to quantify h² enrichments with respect to predefined gene sets while accounting for LD structure⁹³. The primary GSA-MiXeR analysis was performed with the EUR GWAS meta-analysis of endometriosis combined phenotypes. We also investigated EUR adenomyosis to assess differences between these comorbid conditions and also compared EUR endometriosis clinical definition and EUR endometriosis self-reported definition. EUR populations available from 1000 Genomes Project were used as LD reference panel¹⁰. Gene sets were defined based on Gene Ontology (GO) annotations, including biological processes (BP), molecular functions (MF), and cellular components (CC). Bonferroni correction was applied to account for the number of GO terms tested (n = 79,223; P < 6.31 × 10⁻⁷). To reduce redundancy in GO term enrichments and identify representative functional themes, we employed REVIGO⁹⁴, using a semantic similarity cutoff of 0.7 and UniProt as the reference database.

Drug repurposing

We used a drug repositioning approach, DRUGSETS (version 1). Data were extracted from the Drug Gene Interaction Database and the Clue Repurposing Hub. A total of 1,201 drug-gene sets were defined for medications based on genes whose protein products are either direct molecular targets or known interactors with the respective drug. We performed competitive gene set analysis using MAGMA (v.1.10)⁹⁵, conditioning on a background gene set consisting of 2,281 drug target genes, to test for significant associations between drug–gene sets and endometriosis. Bonferroni correction was applied for multiple testing based on the number of drug-gene sets tested (n = 735, P < 6.80 × 10⁻⁵). Drug-gene sets were subsequently classified in three ways: by Anatomical Therapeutic Classification (ATC) III code, by clinical indication, and by the mechanism of action (MOA). A multiple linear regression approach was used to test whether membership in each drug group predicted stronger gene set associations, with group membership as the predictor and the MAGMA drug gene-set t-statistic as the outcome.

Additionally, we used Gene2Drug (version 1)¹⁵ to perform a drug repurposing analysis based on pathway-set enrichments that may be up- or down-regulated by drugs. Gene2Drug transforms gene expression data from the Connectivity Map⁹⁶ into pathway-level profiles and ranks pathways using the Kolmogorov-Smirnov test. Each drug is then assigned an enrichment score (EScore), reflecting both the direction and magnitude of pathway regulation. Semantically reduced Bonferroni-significant GO terms obtained from GSA-MiXeR and REVIGO were used as input.

Polygenic risk scoring

To investigate how endometriosis polygenic risk translates across ancestry groups, we derived PRS from the EUR endometriosis combined definition with and without the AoU cohort, and tested them in all populations in AoU, that is AFR, AMR, CSA, EAS, EUR, and MID. For the EUR analysis, we used summary statistics that excluded the AoU cohort. Posterior variant-level effect sizes for the EUR endometriosis combined definition were estimated using PRS-CS (version 1.1.0)⁹⁷, with the 1000 Genomes Project European population as the LD reference panel¹⁰. Endometriosis PRS were then computed using PLINK2 (version 2.0)⁶⁴, then the PRS estimates were normalized. Additionally, we evaluated the association of PRS quartiles with disease risk using logistic regression models, treating the first quartile as the reference, within each AoU ancestry group.

PRS were calculated in the same manner for UKB endometriosis cases and controls of EUR ancestry, using effect sizes derived from the EUR endometriosis combined definition, excluding the IEGC data, which includes UKB. Using data from EUR individuals in both AoU and UKB, we conducted a PRS interaction analysis using a logistic regression model to assess whether the association between the endometriosis PRS and endometriosis risk differs depending on the presence or absence of endometriosis systemic symptoms and comorbidities (i.e., self-reported and medical diagnoses of abdominal pain, anxiety, bloating, chronic pelvic pain, constipation, contraceptive use, depression, diarrhea, dysmenorrhea, dyspareunia, dysuria, eating disorders, fatigue, fibromyalgia, irritable bowel syndrome, infertility, length of menstrual cycle, low back pain, menorrhagia, migraine, nausea, pain severity, and the sum of all traits, comorbidity burden; Supplementary Tables 27 and 28). Age at recruitment and the top-10 within-ancestry PCs were added as covariates for all analyses.

Pleiotropy analysis

Genetic correlation.

Global r_g was calculated between EUR endometriosis (combined, self-reported, and clinical endometriosis, and adenomyosis) and 35 traits related to female health in seven categories (gynecological/reproductive, psychiatric, pain-related, metabolic, gastrointestinal, cancer, and systemic; Supplementary Table 29) were conducted using the LDSC method (v1.0.1)⁷⁰. Effective sample sizes ranged from 5,632 to 1,816,976. These datasets were selected to limit the sample overlap with the present GWAS meta-analysis to avoid biases in follow-up analyses such as Mendelian randomization. The genetic correlation analysis was limited to traits with a SNP-h² P < 0.05 (Supplementary Table 30). Compared to the standard SNP-h² z > 4, we used a relaxed threshold because of the reduced power of sex-specific GWAS. No bias inducing false-positive results is expected from this. An FDR multiple testing correction (FDR q < 0.05) was applied to define statistical significance. To evaluate heterogeneity across definitions, meta-analyzed endometriosis was compared to adenomyosis, and self-reported endometriosis was compared to endometriosis by clinical definitions using a Z-test. We also used MiXeR (version 1) mixture models⁹⁸ to estimate the number of shared variants between endometriosis and traits with FDR-significant global genetic correlations. As adenomyosis showed a significant genetic correlation with endometrial cancer, whereas endometriosis did not, we followed up this association using LAVA regression models⁹⁹ to identify the specific genomic regions showing the local genetic correlation between these traits. For both MixeR and LAVA analyses, LD was estimated using European populations from the 1000 Genomes Project Phase 3 as the reference panel¹⁰.

Mendelian randomization.

We conducted a bidirectional two-sample MR analysis implemented in the TwoSampleMR R package (version 0.7.0)¹⁰⁰ to investigate the potential causal relationship between endometriosis and genetically correlated traits (q < 0.05). The primary method used was the multiplicative random-effects inverse variance weighted (IVW) approach¹⁰¹. To avoid sample overlap, GWAS summary statistics excluding the shared sample were used for each trait. Only independent SNPs (r² < 0.001 in a 10,000-kb window) with genome-wide significance (P < 5 × 10⁻⁸) in the exposure were included. SNPs that reached significance in both the exposure and the outcome were excluded. Potential causal relationships were removed from the analyses when only one independent significant SNP remained as a valid instrument. P-values were adjusted for multiple comparisons using the FDR correction. We applied the MRlap approach¹⁰² to test the relationship between adenomyosis and endometrial cancer. MRlap applies an inverse-variance weighted framework that corrects for sample overlap, winner’s curse, and weak-instrument bias¹⁰², providing unbiased causal estimates even when the exposure and outcome datasets overlap as in the case of GWAS of adenomyosis and endometrial cancer. The same clumping parameters (r² < 0.001 within 10,000 kb) and genome-wide significance threshold (P < 5 × 10⁻⁸) were used as in the standard MR analyses.

For IVW results with q < 0.05, we performed sensitivity analyses to assess the robustness of the findings. The weighted median and weighed mode methods were used since they are less sensitive to outliers and remain reliable even when only a portion of the genetic variants are valid instruments^103,104. The MR-PRESSO global test was used to remove horizontal pleiotropic outliers and, if detected, the IVW method was rerun using the remaining instruments¹⁰⁵. To minimize the risk of reverse causation, Steiger filtering was used to exclude instruments that explained more variance in the outcome than in the exposure, and the IVW method was repeated on the remaining set of instruments¹⁰⁶. Additionally, if there was a sign of pleiotropy (P-value from MR-Egger intercept < 0.05), the I² statistic was calculated to quantify the strength of the violation of the NO Measurement Error (NOME) assumption¹⁰⁷. If the NOME assumption was not violated (I² > 0.9), the MR-Egger was applied, since it allows all genetic variants to have pleiotropic effects¹⁰⁸. When 0.6 < I² < 0.9, we applied the simulation extrapolation (SIMEX) method to generate an adjusted causal effect estimate¹⁰⁷. When I² < 0.6, MR-Egger results were not taken into consideration. Evidence of a causal relationship was considered if the same direction of effect as the IVW beta estimate and P < 0.05 was found for the (i) weighted median method, (ii) weighted mode method, (iii) IVW after MR PRESSO, (iv) IVW after Steiger filtering, and (v) in case of pleiotropy (MR-Egger intercept P < 0.05), MR-Egger or SIMEX if I² > 0.9 or 0.6 < I² < 0.9, respectively.

Latent causal variable analysis.

To triangulate causal evidence, we performed a LCV (version 1) analysis using LD scores derived from European populations in phase 3 of the 1000 Genomes Project. LCV provided estimates of the GCP, where the sign of the GCP indicates the direction of causality: positive values suggest that phenotype 1 influences phenotype 2, and negative values indicate the reverse. In the present study, we presented only positive GCP values, specifying the direction of the effect they correspond to. The direction of the LCV effect is reflected by the ρ statistic, with ρ > 0 indicating a positive effect and ρ < 0 indicating a negative effect. FDR q < 0.05 was considered to define significant results after multiple testing correction.

Extended Data

Extended Data Figure 1. Ancestry-specific genome-wide association study (GWAS) of endometriosis combined definition (cases identified through self-reported information and electronic health records).

Manhattan plots are shown for African (AFR, panel a; N_cases=2,481 and N_controls=54,671), Admixed American (AMR, panel b; N_cases=1,752 and N_controls=37,871), East Asian (EAS, panel c; N_cases=2,044 and N_controls=87,823), and European (EUR, panel d; N_cases=99,407 and N_controls=1,093,534) populations. Each point represents the p-value from a fixed-effects meta-analysis (weighted by sample size and p-value using METAL) of two-sided logistic regression tests of association between individual genetic variants and the phenotype across contributing cohorts, ordered by genomic position on the x axis and shown as −log10(p-value) on the y axis. The dashed line represents the genome-wide significance threshold (p<5×10⁻⁸), based on Bonferroni multiple testing correction.

Extended Data Figure 2. European (EUR) GWAS of endometriosis subtypes based on electronic health records (EHR) and self-reported data.

Manhattan plots are shown for clinically diagnosed endometriosis (panel a; N_cases=36,695 and N_controls=567,775), self-reported endometriosis (panel b; N_cases=50,867 and N_controls=528,391), and endometriosis excluding adenomyosis (panel c; N_cases=6,627 and N_controls=283,316). Each point represents the p-value from a fixed-effects meta-analysis (weighted by sample size and p-value using METAL) of two-sided logistic regression tests of association between individual genetic variants and the phenotype across contributing cohorts, ordered by genomic position on the x axis and shown as −log10(p-value) on the y axis. The dashed line represents the genome-wide significance threshold (p<5×10⁻⁸), based on Bonferroni multiple testing correction.

DATA AVAILABILITY

The genome-wide association statistics generated by the current study and the endometriosis polygenic score variants and weights are available in Zenodo (doi: 10.5281/zenodo.18983492). GWAS-data-related cohorts included in this study were derived from the following sources: AoU, https://workbench.researchallofus.org/; FinnGen, https://www.finngen.fi/en/access_results; MVP, https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs001672.v12.p1; UKB, https://www.ukbiobank.ac.uk/; BBJ, https://pheweb.jp/downloads; EstBB, https://www.ebi.ac.uk/gwas/; IEGC, https://www.ebi.ac.uk/gwas/; and 23andMe: full GWAS summary statistics for the 23andMe dataset will be made available through 23andMe to qualified researchers under an agreement with 23andMe that protects the privacy of the 23andMe participants, https://www.biorxiv.org/content/10.1101/401448v1.full.

CODE AVAILABILITY

The study used the following computational packages: METAL (version 3), https://github.com/statgen/METAL; GCTA-COJO (version 1.94.1), https://yanglab.westlake.edu.cn/software/gcta/#COJO; PAINTOR (version 3), https://github.com/gkichaev/PAINTOR_V3.0; LDSC (version 1), https://github.com/bulik/ldsc; PRS-CS (version 1.1.0), https://github.com/getian107/PRScs; PLINK2 (version 2.0), https://www.cog-genomics.org/plink/2.0/; MetaXcan (version 0.8.0), https://github.com/hakyimlab/MetaXcan; FUSION (version 1), https://github.com/gusevlab/fusion_twas; HyPrColoc (version 0.0.2), https://github.com/cnfoley/hyprcoloc; MixeR (version 1), https://github.com/precimed/mixer; LCV (version 1), https://github.com/lukejoconnor/LCV; TwoSampleMR (version 0.7.0): https://mrcieu.github.io/TwoSampleMR/; HyPrColoc (version 0.0.2): https://github.com/jrs95/hyprcoloc; MetaXcan (version 0.8.1): https://github.com/hakyimlab/MetaXcan; SMR (version 1.4.0): https://yanglab.westlake.edu.cn/software/smr/#Overview; GSA-MiXeR (version 2.2.1): https://github.com/precimed/gsa-mixer; DRUGSETS (version 1): https://github.com/nybell/drugsets; Gene2Drug (version 1): https://gene2drug.tigem.it/about.php.