Plasma Proteome-Wide Mendelian Randomization with Multi-Layer Triangulation Identifies Druggable Targets for Uterine Leiomyoma

Uterine leiomyomas (fibroids) are benign monoclonal smooth-muscle tumors of the myometrium that affect up to 70% of women by the age of 50 and are the single leading indication for hysterectomy. 1 , 2 They cause heavy menstrual bleeding, pelvic pain, bulk symptoms, subfertility and pregnancy complications, and impose a substantial economic and quality-of-life burden. Diagnosis rests on pelvic imaging: transvaginal ultrasound is the first-line modality and magnetic resonance imaging is used for mapping and problem-solving, and both also serve to separate a benign fibroid from the rare uterine sarcoma and other pelvic malignancy—a distinction that imaging narrows but does not settle, since reported sensitivities for uterine sarcoma are 56.4% for standard ultrasound, 88.4% for expert ultrasound and 87.5% for magnetic resonance imaging. 3 , 4 Despite this, medical therapy is limited and largely hormonal: gonadotropin-releasing hormone (GnRH) agonists and antagonists and selective progesterone-receptor modulators reduce symptoms but are constrained by hypo-estrogenic side-effects, ceiling durations of use and symptom recurrence after discontinuation. 2 , 5 Surgery—myomectomy or hysterectomy—remains the definitive treatment. There is a clear unmet need for non-hormonal, mechanism-anchored medical therapies, yet systematic, genetically grounded discovery of druggable targets for fibroids has been lacking. Human genetic evidence approximately doubles the probability of success in drug development, and drug-target Mendelian randomization (MR) has emerged as a scalable framework to exploit it. 6 , 7 By using cis-acting protein quantitative trait loci (cis-pQTL) as instruments, drug-target MR approximates the effect of lifelong pharmacological modulation of a protein on disease risk, while colocalization guards against confounding by linkage disequilibrium (LD) by testing whether the protein and the disease share a single causal variant. 8 Applied proteome-wide, the approach simultaneously estimates causality and nominates tractable targets and has repeatedly recovered established and novel drug targets across diseases. 8–11 Two developments now make a proteome-wide drug-target MR of uterine leiomyoma feasible. First, large plasma proteogenomic resources—the UK Biobank Pharma Proteomics Project (UKB-PPP) and deCODE—provide robust cis-pQTL for thousands of proteins. 9 , 12 Second, well-powered fibroid GWAS are available in two large populations: FinnGen and the recent multi-ancestry meta-analysis of Kim et al; 13 , 14 as set out in Methods, these two are not independent of each other, because the Kim meta-analysis incorporates an earlier FinnGen release. Prior causal-genetic work on fibroids has operated largely at the DNA-variant and transcript level—GWAS loci, transcriptome-wide association and expression-based summary-data MR 14–22 —while protein-level MR of fibroids has so far been confined to small candidate panels (eg, cathepsins or inflammatory cytokines) 23 , 24 or to multi-disease plasma-proteome screens in which uterine fibroids appeared only as one outcome among many on a single platform (for example, a deCODE-based 18-disease screen that reported THBS3). 25 , 26 A dedicated, comprehensive proteome-wide drug-target MR of uterine fibroids—integrating two proteomic platforms, dual-cohort colocalization and orthogonal triangulation—has not been reported. Here we integrate cis-pQTL MR, dual-cohort colocalization, SMR, directionality and pleiotropy controls, two proteomic platforms, a phenome-wide safety scan, disease-tissue expression, single-cell localization and formal druggability annotation to nominate causal, druggable proteins for uterine leiomyoma. To our knowledge this is the first dedicated, comprehensive proteome-wide drug-target MR of uterine fibroids, distinct from multi-disease plasma-proteome screens in which the disease featured only as one outcome among many.

We performed two-sample cis-pQTL (drug-target) MR with colocalization as the primary analysis, followed by a pre-specified triangulation across orthogonal methods and data sources. Reporting follows the STROBE-MR guideline 27 ( Additional file 1 ). This study analyzed only publicly available, de-identified, summary-level statistics from previously published human genetic and transcriptomic studies; no individual-level data were accessed, no participants were recruited and no primary data were collected. Every parent study had obtained ethical approval and informed consent from its own participants, and all datasets used here are released for open secondary use, so no separate permission from the data owners was required. In accordance with items 1 and 2 of Article 32 of the Measures for Ethical Review of Life Science and Medical Research Involving Human Subjects—issued jointly by the National Health Commission, the Ministry of Education, the Ministry of Science and Technology and the National Administration of Traditional Chinese Medicine of the People’s Republic of China on 18 February 2023 (Guo Wei Ke Jiao Fa (2023) No. 4), which exempt from ethical review research conducted using lawfully obtained public data and research conducted using anonymized information data—ethical review and informed consent were exempted for this study, and institutional review board approval was therefore not sought. The study was conducted in accordance with the Declaration of Helsinki. Table 1 lists every dataset used, with its population, ancestry, sample size, case definition and role in the analysis. This is a two-sample, summary-level Mendelian randomization: we enrolled no participants and had access to no individual-level records, so participant-level baseline characteristics could not be tabulated. Table 1 therefore reports the published characteristics of the contributing study populations, which is the corresponding description for this design, and states how many uterine leiomyoma cases contribute to each analysis layer. Table 1 Characteristics of the Data Sources Dataset (Accession) Population and Ancestry Sample Size Phenotype or Measurement Role in This Study UKB-PPP, Olink Explore 3072 (Synapse syn51365301) 9 UK Biobank participants, European ancestry 54,219 participants; 2923 proteins assayed; 1459 with a valid cis instrument Normalized plasma protein abundance Exposure: cis-pQTL instruments FinnGen R12, CD2_BENIGN_LEIOMYOMA_UTERI 13 Finnish women 42,107 cases/239,957 controls Leiomyoma of the uterus, ICD-coded in national registers Discovery outcome Kim et al 2025, European stratum (GCST90461958) 14 European; contributing cohorts include UK Biobank and FinnGen release 8 53,711 cases/380,441 controls Two or more ICD-9/ICD-10 fibroid codes, or self-reported physician diagnosis Second outcome cohort a Kim et al 2025, East/South/Central Asian stratum (GCST90461959) 14 East Asian and Central South Asian 14,905 cases/69,609 controls As above Independent cross-ancestry look-up Kim et al 2025, African stratum (GCST90461960) 14 African 5678 cases/15,760 controls As above Independent cross-ancestry look-up deCODE (SomaScan) 12 Icelandic 35,559 participants Aptamer-based plasma protein abundance Cross-platform pQTL INTERVAL (SomaScan) 28 UK blood donors, European 3301 participants Aptamer-based plasma protein abundance Cross-platform pQTL eQTLGen 29 European, whole blood Up to 31,684 participants Blood gene expression Blood cis-eQTL GTEx v8 30 Multi-tissue post-mortem donors v8 release (uterus and smooth-muscle-containing tissues) Tissue gene expression Tissue cis-eQTL look-up FinnGen R12, N14_MESNRUIRREG 13 Finnish women 50,120 cases/130,160 controls Excessive, frequent and irregular menstruation Symptom-level extension FinnGen R12, N14_ENDOMETRIOSIS 13 Finnish women 20,190 cases/130,160 controls Endometriosis, ICD-coded Specificity control GSE128229 31 Women undergoing surgery for fibroids 15 leiomyoma and 15 matched myometrium (paired) RNA sequencing Disease-tissue expression GSE169255 31 Women undergoing surgery for fibroids 6 fibroid and 6 myometrium RNA sequencing Replication of tissue expression Human Protein Atlas, single-cell resource 32 , 33 Human tissues Public release Cell-type gene expression Cell-type localization Notes : a The Kim et al European stratum is not independent of the discovery outcome or of the exposure: it incorporates FinnGen release 8 of the same endpoint (28,365 cases/162,514 controls in that release) and UK Biobank. The East/South/Central Asian and African strata contain no FinnGen participants and no European-ancestry UK Biobank participants. This is a two-sample, summary-level Mendelian randomization. The authors enrolled no participants and had no access to individual-level records, so participant-level baseline characteristics could not be tabulated; the table reports the population characteristics of each contributing study as published. Characteristics of the Data Sources Notes : a The Kim et al European stratum is not independent of the discovery outcome or of the exposure: it incorporates FinnGen release 8 of the same endpoint (28,365 cases/162,514 controls in that release) and UK Biobank. The East/South/Central Asian and African strata contain no FinnGen participants and no European-ancestry UK Biobank participants. This is a two-sample, summary-level Mendelian randomization. The authors enrolled no participants and had no access to individual-level records, so participant-level baseline characteristics could not be tabulated; the table reports the population characteristics of each contributing study as published. Sentinel cis-pQTL for 2923 plasma proteins were obtained from UKB-PPP (Sun et al, Olink Explore 3072; n=54,219 European-ancestry participants). 9 Instruments were cis (within ±500 kb of the gene body), genome-wide significant (P10); the lead cis variant was used per protein and palindromic variants were removed. Variant positions were harmonized to GRCh38. Of the 2923 assayed proteins, 1459 had a valid cis instrument (a genome-wide-significant cis-pQTL surviving F>10 and allele harmonization) and constituted the tested set; this instrument filter was applied before any outcome association was examined, and Benjamini–Hochberg FDR was computed across these 1459 tested proteins. Discovery used FinnGen release R12, endpoint CD2_BENIGN_LEIOMYOMA_UTERI (42,107 cases/239,957 controls; Finnish; women only), in which cases are ascertained from the nationwide hospital-discharge, cause-of-death and specialist out-patient registers by ICD codes for leiomyoma of the uterus, and controls are the remaining female participants. 13 Replication used the European-ancestry subset of the multi-ancestry meta-analysis of Kim et al 2025 (GWAS Catalog GCST90461958; 53,711 cases/380,441 controls), which assigns each contributing cohort to the ancestry stratum matching its participants; across the four strata the meta-analysis draws on BioVU, eMERGE, CARDIA, the Black Women’s Health Study, UK Biobank, FinnGen, Biobank Japan and All of Us, with cases defined by two or more ICD-9/ICD-10 fibroid codes or a self-reported physician diagnosis. 14 Both are GRCh38. The two outcome datasets are not independent of each other. The Kim European meta-analysis incorporates FinnGen release 8 of the same endpoint (28,365 cases/162,514 controls in that release), so up to about half of its European cases are also part of the FinnGen R12 discovery set, and it further includes UK Biobank, the source of the pQTL exposure. We therefore treat agreement between the two cohorts as a stringency filter that removes cohort-specific and release-specific artefacts, not as independent replication, and we rely on the East/South/Central Asian and African analyses of Kim et al—which contain no FinnGen participants and no European-ancestry UK Biobank participants—for outcome evidence that is independent of both the discovery cohort and the exposure. Per-protein causal estimates were computed by the Wald ratio (single cis variant) after allele harmonization. Statistical significance across the 1459 testable proteins was controlled by the Benjamini–Hochberg false discovery rate (FDR<0.05). Drug-target MR rests on the instrument-relevance, independence and exclusion-restriction assumptions, applied here following published guidelines. 6 , 34–37 Effect estimates are reported as odds ratios (OR) for leiomyoma per standard-deviation higher genetically predicted protein. The three instrumental-variable assumptions were each checked explicitly. Relevance (IV1) was enforced by admitting only genome-wide-significant cis-pQTL with an F-statistic above 10; the prioritized instruments had F between 229 and 769 ( Table 2 ), so weak-instrument bias is negligible. Independence (IV2), that the instrument is unrelated to confounders of the protein–fibroid relationship, is supported by the use of cis variants, whose action on the protein is local and mechanistically direct, by the restriction of the primary analysis to European-ancestry samples whose association statistics were adjusted for genetic principal components, and by inspection of the phenome-wide scan for associations with plausible confounders. The exclusion restriction (IV3), that the variant affects fibroid risk only through the protein, cannot be tested directly with one instrument and was addressed by four indirect checks: colocalization, which separates a single shared causal variant from two distinct variants in linkage disequilibrium; a cis-pleiotropy screen counting how many other plasma proteins each lead variant also instruments; the phenome-wide scan; and Steiger filtering, which verifies that the instrument explains more variance in the protein than in the outcome. Reverse MR provided an additional directional check. These procedures follow published guidance for drug-target MR and the STROBE-MR reporting framework. 6 , 27 , 34–37 Table 2 The Six Colocalized Causal Proteins for Uterine Leiomyoma Target Direction OR [95% CI] FinnGen OR [95% CI] Kim coloc PP4 (FinnGen/Kim) SMR P (FinnGen/Kim) F FSHB Risk 1.79 [1.56–2.07] 1.76 [1.53–2.03] 0.92/0.98 5.6×10 −14 /5.3×10 −15 409 ITGA11 Protective 0.76 [0.71–0.82] 0.80 [0.75–0.86] 1.00/1.00 5.4×10 −12 /1.2×10 −10 769 RSPO3 Risk 1.33 [1.21–1.47] 1.26 [1.15–1.38] 0.98/0.99 2.4×10 −8 /3.8×10 −7 505 EIF2AK3 Protective 0.71 [0.63–0.81] 0.72 [0.63–0.82] 0.97/1.00 1.4×10 −6 /3.6×10 −7 229 WNT9A Risk 1.31 [1.16–1.48] 1.24 [1.11–1.39] 0.96/0.88 2.4×10 −5 /1.4×10 −4 322 PARP1 Risk 1.23 [1.11–1.37] 1.27 [1.14–1.41] 0.95/0.96 2.4×10 −4 /2.1×10 −5 247 Notes : All instruments have F>10 and are Steiger-correct; all reverse-MR inverse-variance-weighted estimates are non-significant (P≥0.08); none is cis-pleiotropic. Abbreviations : OR, odds ratio per standard-deviation higher genetically predicted protein; CI, confidence interval; PP4, posterior probability of a shared causal variant; SMR, summary-data-based Mendelian randomization; F, instrument F-statistic. The Six Colocalized Causal Proteins for Uterine Leiomyoma Notes : All instruments have F>10 and are Steiger-correct; all reverse-MR inverse-variance-weighted estimates are non-significant (P≥0.08); none is cis-pleiotropic. Abbreviations : OR, odds ratio per standard-deviation higher genetically predicted protein; CI, confidence interval; PP4, posterior probability of a shared causal variant; SMR, summary-data-based Mendelian randomization; F, instrument F-statistic. All 31 FDR-significant proteins underwent colocalization: coloc.abf (default priors) was applied across the cis region (±500 kb around the lead cis-pQTL) against each outcome, requiring a posterior probability of a shared causal variant PP4>0.8 in both FinnGen and the Kim European meta-analysis. 8 Multiple-causal-variant (SuSiE) colocalization 38 is not applicable to single cis-variant instruments and was not performed. Because the posterior probabilities returned by coloc.abf depend on the prior probability p12 that a variant affects both traits, robustness to that prior was assessed in three complementary ways. First, PP4 was recomputed for all 31 FDR-significant proteins in both cohorts across p12 = 1×10 −6 to 1×10 −4 . Second, for the prioritized targets PP4 was recomputed across a finer grid from 1×10 −7 to 1×10 −4 and we recorded, for each target and cohort, the weakest prior at which PP4 still exceeded 0.8. Third, we computed the conditional probability PP4/(PP3+PP4)—the posterior support for one shared causal variant relative to two distinct causal variants, conditional on both traits having a signal in the region, which is far less sensitive to p12 than PP4 itself. SMR was computed for each prioritized target in both outcome cohorts using the lead cis-pQTL. Because the HEIDI heterogeneity test requires multiple cis-SNPs and is undefined for single-variant instruments, colocalization (above) rather than HEIDI served as the shared-causal-variant test distinguishing causality from LD linkage. 39 With a single cis instrument, SMR re-expresses the Wald-ratio estimate and so confirms the direction and magnitude of the MR association rather than providing fully independent evidence. Steiger filtering confirmed that each instrument explained more variance in the protein than in the outcome. 40 Reverse MR (leiomyoma liability → protein) used 64 independent genome-wide leiomyoma variants (FinnGen R12; P<5×10 −8 , r 2 <0.001 within 1 Mb) with inverse-variance-weighted, weighted-median and MR-Egger estimators. 41 cis-pQTL pleiotropy was screened by counting other proteins associated with each instrument. Each protein is instrumented by a single cis variant and the causal estimate is a Wald ratio, so no multi-instrument regression is fitted and collinearity between instruments cannot arise. The analogous concern—that the pQTL signal and the fibroid signal are driven by two different but correlated variants in the same linkage-disequilibrium block—is precisely what colocalization tests, and it was required in both outcome cohorts. In the reverse-MR analysis, which does use multiple instruments, correlation between instruments was removed by LD clumping to r 2 <0.001 within 1 Mb. Weak-instrument bias is negligible (F 229–769). Because instrument–exposure effect sizes were taken from the dataset in which they were discovered, causal magnitudes are subject to winner’s curse and are likely to be overestimated in absolute value, although the direction of effect is unaffected. 42 The exposure (UK Biobank) and the replication outcome (which includes UK Biobank and FinnGen release 8) share participants; participant overlap in two-sample MR biases estimates towards the confounded observational association in proportion to the degree of overlap and to instrument weakness, so with F≥229 the expected bias here is small, 43 and colocalization, which compares patterns of association across a region rather than pooling two effect estimates, is not a two-sample estimator and is not affected by overlap. Population stratification was addressed by the parent studies’ principal-component adjustment and by restricting the primary analysis to European-ancestry data. Odds ratios are reported per standard-deviation higher genetically predicted protein on the non-collapsible odds scale, and index causal direction and relative support rather than the effect expected from any particular drug dose. Neither outcome GWAS is a study of pelvic symptoms in general: both define cases by a recorded clinical diagnosis of uterine leiomyoma, which in routine practice is made on pelvic imaging and, when surgery follows, confirmed histologically. Endometriosis, adenomyosis and gynecological malignancy carry their own diagnostic codes and are not counted as fibroid cases, and occult uterine sarcoma is found in only about 1 of every 280 hysterectomies performed for presumed leiomyoma (125 of 34,728 in a population-based cohort). 44 Residual misclassification cannot be excluded and would be expected to bias estimates towards the null. To test empirically whether the prioritized signals are specific to fibroids rather than shared across reproductive-tract disease, all six targets were additionally colocalized with endometriosis (FinnGen R12 endpoint N14_ENDOMETRIOSIS). cis-pQTL from deCODE (Ferkingstad et al; SomaScan, n=35,559) 12 and INTERVAL (Sun et al 2018; SomaScan, n=3,301) 28 were used to repeat MR/colocalization for proteins represented on the SomaScan panel and to assess effect-direction concordance. Blood cis-eQTL MR used eQTLGen (n≤31,684). 29 Tissue cis-eQTL were queried in GTEx v8, including uterus and well-powered smooth-muscle–containing tissues, and benchmarked against a positive-control eGene. 30 Each target instrument was screened across the OpenGWAS database to characterize on-target/repurposing and potential adverse phenotypes. 41 , 45 Fibroid-versus-myometrium differential expression was assessed in two transcriptomic datasets— GSE128229 (15 leiomyoma vs 15 matched myometrium, paired RNA-seq; DESeq2) and GSE169255 (6 fibroid vs 6 normal myometrium), both obtained from NCBI GEO 31 —and cell-type localization from the Human Protein Atlas single-cell resource. 32 , 33 Effect-direction concordance for the prioritized targets was examined in the East/South/Central Asian and African subsets of Kim et al (GWAS Catalog GCST90461959 and GCST90461960); because the pQTL exposure is European, this is an outcome-direction look-up at the European cis lead variant rather than a within-ancestry MR. Colocalization with heavy menstrual bleeding used FinnGen R12 endpoint N14_MESNRUIRREG (50,120 cases/130,160 controls). Target tractability, modality, known drugs and clinical-trial status were annotated with Open Targets, DrugBank, DGIdb and the druggable-genome framework. 46–49 Analyses used R 4.6.0 (coloc 5.2.3, MendelianRandomization 0.10.0, data.table 1.18.4, ieugwasr 1.1.0, ggplot2 4.0.3; versions from sessionInfo()); colocalization used coloc.abf with default priors (p1=p2=1×10 −4 , p12=1×10 −5 ) plus a p12 sensitivity grid (1×10 −6 –1×10 −4 for all 31 proteins; 1×10 −7 –1×10 −4 for the six prioritized targets); reverse MR used genome-wide leiomyoma instruments (P<5×10 −8 , LD-clumped r 2 <0.001 within 1 Mb); fibroid-versus-myometrium differential expression used DESeq2 with a paired design (~patient + tissue) and Benjamini–Hochberg adjustment. Code, the per-protein instrument table and intermediate results: https://doi.org/10.5281/zenodo.20831756 (Zenodo).

Of 1459 proteins with valid cis instruments, 31 showed FDR-significant MR associations with uterine leiomyoma ( Figures 1 and 2a ; Additional file 2 ). Effect sizes spanned protective and risk-increasing directions, motivating a colocalization filter to distinguish causal proteins from LD artefacts. Figure 1 Study design and analytical pipeline. Plasma cis-pQTL instruments (UKB-PPP) were tested proteome-wide against uterine leiomyoma (FinnGen R12), prioritized by colocalization (PP4>0.8), replicated in the Kim 2025 European meta-analysis and on the deCODE/INTERVAL SomaScan platform, and triangulated by SMR, Steiger, reverse MR, phenome-wide safety scanning, disease-tissue expression and druggability annotation. A flowchart detailing the process of plasma cis-pQTL analysis and drug-target identification. Figure 2 ( a ) Volcano plot of proteome-wide cis-pQTL MR (1,459 proteins; x-axis, log OR per SD higher protein; y-axis, −log10 P); the 31 FDR-significant proteins are highlighted and the six colocalized targets labeled in italics. ( b ) Two-cohort forest plot of OR (95% CI) for the six targets in FinnGen and the Kim European meta-analysis. Image A presents a scatter plot with categories: Not significant, FDR < 0.05 and Colocalized target. The x-axis is ′MR effect (log OR per SD higher protein)′ from -1.0 to 0.6 and the y-axis is ′-log10(P)′ from 0 to 16. Notable points: EIF2AK3 (-0.35, 6.5), ITGA11 (-0.25, 12), RSPO3 (0.30, 8), WNT9A (0.25, 5), PARP1 (0.20, 4), FSHB (0.55, 15). A dashed line is at x = 0. Unlabeled points cluster near x = 0 and y = 0 to 2. Image B shows a forest plot with y-axis categories: FSHB, ITGA11, RSPO3, EIF2AK3, WNT9A, PARP1. The x-axis is ′OR (95% CI) per SD higher protein′ with ticks at 0.6, 0.8, 1.0, 1.3, 1.8. A dashed line is at 1.0. Legend: Cohort, FinnGen, Kim 2025. Each protein has two diamond markers with horizontal 95% confidence intervals (FinnGen, then Kim 2025): FSHB 1.79 and 1.76; ITGA11 0.76 and 0.80; RSPO3 1.33 and 1.26; EIF2AK3 0.71 and 0.72; WNT9A 1.31 and 1.24; PARP1 1.23 and 1.27. A scatter plot and a forest plot showing protein associations and odds ratios for six labeled targets. Study design and analytical pipeline. Plasma cis-pQTL instruments (UKB-PPP) were tested proteome-wide against uterine leiomyoma (FinnGen R12), prioritized by colocalization (PP4>0.8), replicated in the Kim 2025 European meta-analysis and on the deCODE/INTERVAL SomaScan platform, and triangulated by SMR, Steiger, reverse MR, phenome-wide safety scanning, disease-tissue expression and druggability annotation. ( a ) Volcano plot of proteome-wide cis-pQTL MR (1,459 proteins; x-axis, log OR per SD higher protein; y-axis, −log10 P); the 31 FDR-significant proteins are highlighted and the six colocalized targets labeled in italics. ( b ) Two-cohort forest plot of OR (95% CI) for the six targets in FinnGen and the Kim European meta-analysis. All 31 FDR-significant proteins were carried to colocalization against both outcome GWAS. Six colocalized (PP4>0.8) in both FinnGen and the Kim European meta-analysis—ITGA11, RSPO3, EIF2AK3, WNT9A, PARP1 and FSHB ( Table 2 ; Figures 2b and 3 ; the shared regional signal at the FSHB locus is shown in Figure 4 )—with concordant effect directions ( Figure 2b ), whereas the remaining 25 did not. Notably, several (eg, SULT2A1, NUB1, CWC15, DNM1) colocalized in only one cohort and would have been false positives under a single-cohort design, underscoring the stringency added by requiring colocalization in both cohorts ( Additional file 3 ; Figure 3 ). The targets converge on three pathways central to fibroid biology—Wnt signaling (RSPO3, WNT9A), extracellular-matrix/fibrosis (ITGA11) and reproductive-hormone signaling (FSHB)—alongside DNA repair (PARP1) and the PERK/integrated-stress response (EIF2AK3). FSHB, encoding the β-subunit of follicle-stimulating hormone, recapitulates the mechanism of approved gonadotropin-axis therapy and serves as an internal positive control. Figure 3 Colocalization posterior probabilities (PP4) for all 31 FDR-significant proteins against FinnGen and the Kim European meta-analysis. Cell shading and the printed value give PP4 in each cohort. A star (★) after a gene symbol marks the six proteins that exceed PP4>0.8 in both GWAS; these lie above the dashed coral line, and the remaining 25 proteins lie below it, several colocalizing in only one cohort. A heatmap displays gene colocalization with PP4 values in two studies: FinnGen and Kim 2025. Genes with PP4 > 0.8 in both studies are marked with a star and listed above a dashed line. These include ITGA11 (1.00, 1.00), RSPO3 (0.98, 0.99), EIF2AK3 (0.97, 1.00), PARP1 (0.95, 0.96), FSHB (0.92, 0.98) and WNT9A (0.96, 0.88). Below the line, genes like SULT2A1 (0.93, 0.60) and DNM1 (0.47, 0.87) have varying PP4 values. CWC15 (0.87, 0.19) and NUB1 (0.82, 0.04) exceed 0.8 in FinnGen only; the remaining genes show lower colocalization. The legend indicates PP4 ranges from 0.0 to 1.0. Colocalization heatmap: ITGA11-IL18 vs FinnGen & Kim 2025; probabilities displayed. Figure 4 Regional colocalization at the FSHB locus. Variant-level association (−log10 P ) across the FSHB cis region for the plasma FSHB cis-pQTL (UKB-PPP; top) and uterine leiomyoma (FinnGen; bottom); the lead cis-pQTL variant rs11031006 (diamond; shaded band) coincides in both signals, consistent with a single shared causal variant (coloc.abf PP4 = 0.92). Regional plots for PARP1 (PP4 0.95) and RSPO3 (PP4 0.98) are provided as Supplementary figures S2 and S3 , respectively ( Additional file 7 ). Two vertically stacked scatter plots share a horizontal axis. Top plot: Vertical axis labeled ′FSHB cis pQTL (UKB-PPP)′ (minus log subscript 10), ranging 0-100. Horizontal axis is chromosome 11 position (Mb), spanning 29.8 to 30.6. Points cluster near 0, with a dense group around 30.2-30.4, including a highlighted diamond, ′lead rs11031006,′ near 30.2 at about 90. A shaded band centers near 30.2. Bottom plot: Vertical axis labeled ′Uterine leiomyoma GWAS (FinnGen)′ (minus log subscript 10), ranging 0-17.5 with 2.5 intervals. Horizontal axis labeled ′Chromosome 11 position (Mb),′ with ticks at 29.8, 30.0, 30.2, 30.4, 30.6. Points cluster near 0, with higher concentration around 30.2-30.4. A highlighted diamond, ′rs11031006,′ sits near 30.2 at about 15-16. A shaded band aligns with the top plot. A two scatter plot figure showing FSHB cis pQTL and uterine leiomyoma genome wide association across chromosome 11. Colocalization posterior probabilities (PP4) for all 31 FDR-significant proteins against FinnGen and the Kim European meta-analysis. Cell shading and the printed value give PP4 in each cohort. A star (★) after a gene symbol marks the six proteins that exceed PP4>0.8 in both GWAS; these lie above the dashed coral line, and the remaining 25 proteins lie below it, several colocalizing in only one cohort. Regional colocalization at the FSHB locus. Variant-level association (−log10 P ) across the FSHB cis region for the plasma FSHB cis-pQTL (UKB-PPP; top) and uterine leiomyoma (FinnGen; bottom); the lead cis-pQTL variant rs11031006 (diamond; shaded band) coincides in both signals, consistent with a single shared causal variant (coloc.abf PP4 = 0.92). Regional plots for PARP1 (PP4 0.95) and RSPO3 (PP4 0.98) are provided as Supplementary figures S2 and S3 , respectively ( Additional file 7 ). Every target passed each applicable orthogonal test ( Figure 5 ): SMR was significant in both cohorts (all FDR<0.05); Steiger filtering confirmed the protein→fibroid direction for all six (all P≤1.9×10 −17 ); reverse MR provided no significant evidence that leiomyoma liability influences the proteins (all IVW P≥0.08; MR-Egger intercepts non-significant); although Cochran’s Q indicated heterogeneity across the 64 polygenic reverse instruments—as expected for a complex-trait liability—the IVW, weighted-median and MR-Egger estimators concordantly returned null effects; instruments were strong (F 229–769) and free of cis-pleiotropy (each lead variant a pQTL for only its own protein). Sensitivity of the colocalization results to the choice of prior is reported separately below. Figure 5 Convergent-evidence matrix: each target against each orthogonal test (MR in two cohorts, colocalization in two cohorts, SMR in two cohorts, Steiger direction, reverse-MR non-significant, cross-platform SomaScan, instrument strength, cis-pleiotropy). Symbols, as printed in the tiles: ✓ (green) the test supports the target; ~ (amber) direction-concordant only; n/a (grey) not applicable, because the protein is not on the SomaScan panel or the SomaScan analyte differs. Every tile in the matrix carries one of these three symbols. Heatmap matrix with 6 target rows labeled, FSHB, ITGA11, RSPO3, EIF2AK3, WNT9A, PARP1. Column labels, MR FinnGen, MR Kim, Coloc FinnGen, Coloc Kim, SMR FinnGen, SMR Kim, Steiger, Reverse MR n.s., Cross platform, Strong IV F greater than 10, No cis pleiotropy. No axis units shown. Tile symbols are check mark, tilde, or n slash a. Row FSHB: check marks in all columns except Cross platform is n slash a. Row ITGA11: check marks in all columns except Cross platform is n slash a. Row RSPO3: check marks in all columns except Cross platform is tilde. Row EIF2AK3: check marks in all columns except Cross platform is n slash a. Row WNT9A: check marks in all columns except Cross platform is n slash a. Row PARP1: check marks in every column including Cross platform. A heatmap matrix of orthogonal tests by targets FSHB to PARP1, with tiles showing symbols. Convergent-evidence matrix: each target against each orthogonal test (MR in two cohorts, colocalization in two cohorts, SMR in two cohorts, Steiger direction, reverse-MR non-significant, cross-platform SomaScan, instrument strength, cis-pleiotropy). Symbols, as printed in the tiles: ✓ (green) the test supports the target; ~ (amber) direction-concordant only; n/a (grey) not applicable, because the protein is not on the SomaScan panel or the SomaScan analyte differs. Every tile in the matrix carries one of these three symbols. Colocalization posteriors depend on the prior probability p12 that a variant affects both traits, so we asked how much of the classification survives a change in that prior. Across the grid applied to all 31 FDR-significant proteins in both cohorts ( Additional file 3 ), exactly the six prioritized proteins colocalized in both cohorts at the conventional default (p12=1×10 −5 ) and none of the other 25 did; under the most permissive prior (1×10 −4 ) two further proteins (SULT2A1 and DNM1) also crossed the threshold in both cohorts, and under the most stringent (1×10 −6 ) only ITGA11 and RSPO3 retained PP4>0.8 in both, with FSHB and EIF2AK3 falling below 0.8 in FinnGen and WNT9A and PARP1 in both. The finer grid quantifies this ( Table 3 ): PP4>0.8 is maintained in both cohorts down to p12=2×10 −7 for ITGA11, 1×10 −6 for RSPO3, 2×10 −6 for EIF2AK3, 5×10 −6 for FSHB and PARP1, and only at the default 1×10 −5 for WNT9A. A prior-insensitive comparison is more uniform: the posterior for two distinct causal variants (PP3) was small throughout (0.0005–0.076), so conditional on both traits having a signal in the region the support for a single shared variant, PP4/(PP3+PP4), was 0.92–0.9995 for every target in both cohorts. We therefore treat ITGA11 and RSPO3 as prior-robust, EIF2AK3, FSHB and PARP1 as intermediate, and WNT9A as supported only at the default prior, and we carry that ranking into the interpretation of each target below. Table 3 Robustness of Colocalization to the Choice of Prior Target PP4 FinnGen/Kim (Default Prior) PP3 FinnGen/Kim PP4/(PP3+PP4) FinnGen/Kim Weakest p12 Giving PP4>0.8 in Both Cohorts Classification FSHB 0.924/0.983 0.0757/0.0173 0.924/0.983 5×10 −6 Intermediate ITGA11 0.996/0.999 0.0043/0.0005 0.996/0.999 2×10 −7 Prior-robust RSPO3 0.982/0.992 0.0181/0.0080 0.982/0.992 1×10 −6 Prior-robust EIF2AK3 0.973/0.998 0.0269/0.0020 0.973/0.998 2×10 −6 Intermediate WNT9A 0.963/0.877 0.0150/0.0209 0.985/0.977 1×10 −5 Default prior only PARP1 0.950/0.958 0.0488/0.0404 0.951/0.960 5×10 −6 Intermediate Notes : PP3, posterior probability of two distinct causal variants; PP4, posterior probability of one shared causal variant; p12, prior probability that a variant affects both traits (default 1×10 −5 ). PP4/(PP3+PP4) is the support for a shared causal variant conditional on both traits having a signal in the region, and is far less sensitive to p12 than PP4 itself. Computed with coloc.abf over the cis region (±500 kb around the lead cis-pQTL) in each outcome cohort; the full grid for all 31 FDR-significant proteins is in Additional file 3 . Robustness of Colocalization to the Choice of Prior Notes : PP3, posterior probability of two distinct causal variants; PP4, posterior probability of one shared causal variant; p12, prior probability that a variant affects both traits (default 1×10 −5 ). PP4/(PP3+PP4) is the support for a shared causal variant conditional on both traits having a signal in the region, and is far less sensitive to p12 than PP4 itself. Computed with coloc.abf over the cis region (±500 kb around the lead cis-pQTL) in each outcome cohort; the full grid for all 31 FDR-significant proteins is in Additional file 3 . On the SomaScan platform, PARP1 colocalized in deCODE (PP4 0.97) with concordant direction—independent confirmation on a second proteomic technology. RSPO3 was direction-concordant across deCODE and INTERVAL (INTERVAL cis-pQTL MR OR 1.14, P=1.4×10 −5 ), although its deCODE colocalization was weaker (PP4 0.26). ITGA11, WNT9A and EIF2AK3 are not represented on SomaScan. The FSHB SomaScan reagent targets the FSH heterodimer rather than the FSHB subunit and is therefore not directly comparable. None of the six genes was a significant cis-eGene in GTEx uterus, in well-powered smooth-muscle–containing tissues, or (for FSHB) in pituitary, whereas a positive-control eGene was readily detected; blood cis-eQTL (eQTLGen) corroborated none of the targets, because the relevant genes are not expressed, or not genetically regulated at the transcript level, in these bulk tissues. This pattern—strong cis-pQTL without a detectable bulk-tissue cis-eQTL—indicates regulation at the protein or post-transcriptional level, or in specific cell types diluted in bulk tissue, and underscores why a plasma-proteome (rather than transcriptome) instrument is informative for these targets. The target instruments showed coherent, mechanism-consistent phenome-wide profiles ( Figure 6 ; Additional file 4 ). FSHB associated with reproductive phenotypes (menstrual-cycle length, age at menarche and menopause, hysterectomy, endometriosis); ITGA11 with uterine fibroids and insulin-like growth factor 1; and PARP1 with platelet count and plateletcrit—the latter recapitulating the on-target thrombocytopenia of clinical PARP inhibition. RSPO3 (bone mineral density), EIF2AK3 (height/skeletal) and WNT9A (lipids) showed broader pleiotropy; for cis-pQTL instruments such associations largely reflect the proteins’ downstream (vertical) biology, and colocalization addresses horizontal pleiotropy at the fibroid locus, but these three targets warrant correspondingly greater caution. Figure 6 Phenome-wide/safety scan: top trait associations for each target’s cis-pQTL instrument (OpenGWAS). On-target/safety-relevant traits are highlighted, including PARP1–platelet indices and FSHB–reproductive phenotypes. Trait names are reproduced verbatim from OpenGWAS, which holds separate records for near-synonymous traits (for example plateletcrit and platelet crit). Six dot plots analyze trait associations for FSHB, ITGA11, RSPO3, EIF2AK3, WNT9A and PARP1. Each plot uses x-axis as -log10(P) and y-axis for traits. FSHB′s strongest trait is menstrual cycle length at 49. ITGA11′s strongest is uterine fibroids at 12. RSPO3′s strongest is heel bone mineral density at 230, the highest value across all plots. EIF2AK3′s strongest is height at 60. WNT9A′s strongest is triglyceride at 11.8. PARP1′s strongest is telomere length at 44. Six dot plots showing trait association significance for FSHB, ITGA11, RSPO3, EIF2AK3, WNT9A and PARP1. Phenome-wide/safety scan: top trait associations for each target’s cis-pQTL instrument (OpenGWAS). On-target/safety-relevant traits are highlighted, including PARP1–platelet indices and FSHB–reproductive phenotypes. Trait names are reproduced verbatim from OpenGWAS, which holds separate records for near-synonymous traits (for example plateletcrit and platelet crit). In fibroid vs matched myometrium, PARP1 and RSPO3—two of the colocalized risk-increasing targets—were up-regulated, concordant with their MR direction (PARP1 BH-adjusted P=2.1×10 −13 and replicated P=6.2×10 −4 ; RSPO3 adjusted P=7.3×10 −5 ), and effect directions were consistent across both transcriptomic datasets ( Figure 7 ; Additional file 5 ). WNT9A was lower in fibroid tissue, ITGA11 and EIF2AK3 were not significantly differential, and FSHB was not expressed in uterus, consistent with an endocrine rather than local action. Single-cell localization placed ITGA11 in fibroblasts/fibro-adipogenic progenitors and RSPO3 in decidual stromal cells—the mesenchymal compartment of fibroid biology—and FSHB in pituitary gonadotrophs. Figure 7 Fibroid vs adjacent myometrium expression for the six targets ( GSE128229 , 15 matched pairs; boxes show median and interquartile range, whiskers the 1.5×interquartile range, and points individual samples). Symbols give the DESeq2 Benjamini–Hochberg adjusted P for fibroid vs myometrium: *** <0.001, ** <0.01, ns ≥0.05; panels annotated “not expressed in uterus” carry no test. PARP1 and RSPO3 are up-regulated in fibroid; WNT9A is significantly down-regulated (discordant with its risk-increasing MR direction; see Discussion); FSHB is not expressed in uterus. Six box plots are displayed in two rows and three columns, comparing Myo. and Leio. expression levels. Top left: FSHB shows no expression in the uterus. Top middle: ITGA11, with ′ns′ annotation, has Myo. median at 3.7 and Leio. median at 3.56. Top right: RSPO3, marked with three asterisks, shows Myo. median at 3.15 and Leio. median at 3.33. Bottom left: EIF2AK3, annotated ′ns′, has Myo. median at 2.56 and Leio. median at 2.49. Bottom middle: WNT9A, with two asterisks, shows Myo. median at 2.15 and Leio. median at 1.85. Bottom right: PARP1, marked with three asterisks, has Myo. median at 2.95 and Leio. median at 3.18. Each plot′s y-axis is labeled ′Expression log10(norm+1)′ with varying ranges and increments. Six box plots showing expression differences between myometrium and leiomyoma for six targets. Fibroid vs adjacent myometrium expression for the six targets ( GSE128229 , 15 matched pairs; boxes show median and interquartile range, whiskers the 1.5×interquartile range, and points individual samples). Symbols give the DESeq2 Benjamini–Hochberg adjusted P for fibroid vs myometrium: *** <0.001, ** <0.01, ns ≥0.05; panels annotated “not expressed in uterus” carry no test. PARP1 and RSPO3 are up-regulated in fibroid; WNT9A is significantly down-regulated (discordant with its risk-increasing MR direction; see Discussion); FSHB is not expressed in uterus. Because the pQTL exposure is European, this is a cross-ancestry outcome-direction look-up at the European lead variant rather than a within-ancestry MR. In the independent East/South/Central Asian fibroid meta-analysis (n≈84,500), all six lead variants were direction-concordant with the European estimate, with nominally significant fibroid associations (FSHB P=2.3×10 −7 , ITGA11 P=4.4×10 −11 , RSPO3 P=7.1×10 −3 , EIF2AK3 P=2.2×10 −5 , WNT9A P=1.9×10 −3 , PARP1 P=7.4×10 −5 ); in the smaller African subset (n≈21,400) FSHB, ITGA11 and RSPO3 were direction-concordant (FSHB P=2.2×10 −3 ) while the remaining three were underpowered (all P>0.1, no true reversal). Extending to the cardinal symptom, FSHB colocalized strongly with heavy menstrual bleeding (FinnGen N14; OR 2.31, P=1.5×10 −32 , PP4 0.99) and RSPO3 showed a suggestive signal (PP4 0.70), whereas the tumor-biology targets (PARP1, ITGA11, WNT9A, EIF2AK3) did not colocalize with bleeding (PP4<0.25)—consistent with FSHB acting through the gonadotropin axis on both tumor risk and bleeding, and the others on tumor biology. All six targets carry some pharmacological tractability annotation ( Table 4 ; Additional file 6 ), but tractability is heterogeneous. PARP1 is an approved-drug target (olaparib, niraparib, rucaparib, talazoparib) 50 and the most immediately actionable repurposing candidate, with risk-increasing genetics indicating inhibition. The gonadotropin axis upstream of FSHB is clinically validated in fibroids, although the approved agents relugolix and elagolix act on the pituitary GnRH receptor rather than on FSHB directly. 51 RSPO3 has antibody-modality precedence (the anti-RSPO3 antibody rosmantuzumab reached Phase 1). 52 WNT9A is addressable only non-selectively via porcupine inhibitors. 53 The two protective targets, EIF2AK3 (PERK) and ITGA11, would require augmentation, for which only inhibitory tool compounds currently exist; they are therefore hypothesis-generating. Table 4 Druggability and Repurposing Hypotheses for the Six Colocalized Targets Target MR-Implied Action Tractability Agents/Status PARP1 Inhibit Small molecule: approved PARP inhibitors (olaparib etc), approved in oncology FSHB Lower FSH Axis druggable (upstream) GnRH antagonists relugolix/elagolix (approved for fibroids, act on GnRHR); anti-FSHβ antibody preclinical RSPO3 Inhibit Antibody: Phase 1 precedence Anti-RSPO3 antibody (rosmantuzumab, Phase 1, discontinued) WNT9A Inhibit Pathway-level only Porcupine inhibitors (non-selective; Phase 1/2 oncology) ITGA11 Augment Antibody-tractable surface Neutralizing anti-integrin antibodies (opposite direction; preclinical) EIF2AK3 Augment (caution) Small molecule (druggable) PERK activators preclinical; chronic activation toxic Notes : For the two proteins with a protective Mendelian-randomization direction (ITGA11, EIF2AK3), the available pharmacology acts opposite to the direction required, so neither is a repurposing candidate. Druggability and Repurposing Hypotheses for the Six Colocalized Targets Notes : For the two proteins with a protective Mendelian-randomization direction (ITGA11, EIF2AK3), the available pharmacology acts opposite to the direction required, so neither is a repurposing candidate.

Colocalized, triangulated plasma proteins implicate druggable Wnt, extracellular-matrix and hormonal pathways in uterine leiomyoma. ITGA11 and RSPO3 are the colocalization signals most robust to the choice of prior; PARP1, whose colocalization is prior-sensitive but which is confirmed on a second proteomic platform, up-regulated in fibroid tissue and specific to fibroids rather than shared with endometriosis, is the candidate best supported for experimental evaluation, with RSPO3 a second, antibody-tractable lead. The oncology-grade hematological and embryo-fetal toxicity of approved PARP inhibitors means that this is a hypothesis for mechanistic and preclinical testing, not a therapy that could be offered to reproductive-age women as these drugs are currently formulated. Confirmation in independent and non-European populations is required.

In this proteome-wide drug-target MR of uterine leiomyoma, six plasma proteins were causal, colocalized in both outcome cohorts and supported by orthogonal evidence: ITGA11, RSPO3, EIF2AK3, WNT9A, PARP1 and FSHB ( Figure 5 ). They are not, however, equally well supported, and the discussion that follows is organized around that difference rather than around the methods used to reach it. The targets map onto the core biology of fibroids, and their convergence on myometrial proliferation is summarized in Supplementary Figure S1 ( Additional file 7 ). RSPO3 and WNT9A implicate Wnt/β-catenin signaling, a central driver of myometrial stem-cell self-renewal and leiomyoma growth; 54–56 our data refine the Wnt signal—previously detected at the WNT4 locus in transcriptome-wide analysis 14 —to specific, secreted, druggable Wnt effectors. RSPO3 and WNT9A are instrumented by independent cis variants on different chromosomes, each without effect on the other’s protein, indicating two distinct Wnt-axis signals rather than one. ITGA11, a collagen-binding integrin marking myofibroblasts and fibro-adipogenic progenitors, implicates extracellular-matrix remodeling and fibrosis, the histological hallmark of fibroids; it is expressed in human fibroid stroma. 57 , 58 FSHB nominates the gonadotropin axis and, by recapitulating the mechanism of approved GnRH-axis therapy, provides an internal positive control validating the pipeline. 59 , 60 Because follicle-stimulating hormone acts principally on the ovary to sustain the estrogen-rich endocrine milieu that drives fibroid growth, the FSHB signal also anchors fibroid risk to ovarian endocrine function. PARP1 is the standout repurposing candidate: risk-increasing genetics indicate inhibition, it is targeted by four approved PARP inhibitors, it colocalized across two proteomic platforms, and it is up-regulated in fibroid vs myometrium in two datasets. 61 PARP inhibitors have previously been proposed for the rare fumarate-hydratase (FH)-deficient leiomyoma subtype, in which oncometabolite accumulation impairs homologous-recombination repair and confers a synthetic-lethal PARP dependence; 62 our germline cis-pQTL MR is mechanistically distinct, extending a PARP1 rationale from that rare subtype to common fibroid risk in the general population rather than invoking synthetic lethality. PARP1 sustains DNA repair and genomic stability, 50 , 63 plausibly licensing the clonal expansion of fibroid precursors that frequently carry MED12 or other driver alterations; 64 , 65 its phenome-wide association with platelet indices mirrors the on-target thrombocytopenia of clinical PARP inhibition. The safety barrier to repurposing PARP inhibitors for a benign, often fertility-desiring, reproductive-age population is, however, substantial and quantifiable. In a meta-analysis of 31 randomized trials in solid tumors, anemia was the commonest adverse event at any grade (49.2%) and at grade 3 or worse (25.0%), and PARP inhibition raised the risk of grade≥3 anemia (risk ratio 5.43), thrombocytopenia (5.42) and neutropenia (1.70). 66 Secondary myelodysplastic syndrome and acute myeloid leukemia are rarer but often fatal: a meta-analysis of 18 placebo-controlled trials found an excess with PARP inhibition (0.73% vs 0.47%; Peto odds ratio 2.63, 95% CI 1.13–6.14), with a median latency of 17.8 months and death in 45% of pharmacovigilance cases, 67 although the more recent solid-tumor meta-analysis did not detect an increase. 66 Decisively for this indication, the approved agents are embryo-fetal toxic: the olaparib label requires pregnancy testing before treatment and effective contraception during treatment and for six months after the last dose, and reports myelodysplastic syndrome or acute myeloid leukemia in about 1.2% of exposed patients, mostly fatal. 68 A drug that must be paired with six months of contraception is close to unusable in a woman whose reason for seeking fibroid treatment is subfertility, and the balance of benefit and harm that is acceptable in ovarian cancer is not the balance acceptable in a benign tumor. Any translation would therefore need either a population in whom fertility is not the goal, or a delivery strategy that confines exposure to the uterus, neither of which has been tested; PARP1 is therefore best framed as a genetically and tissue-supported hypothesis constrained by oncology-grade toxicity and an as-yet-unproven fibroid PARP-dependency, not a ready-to-deploy therapy, and the MED12-related clonal-expansion mechanism we invoke is speculative. RSPO3 offers a second, antibody-tractable candidate with Phase-1 modality precedence. Three of the six targets point one way in MR and the other way in tissue, and this deserves a direct answer rather than a caveat. The two estimands are not the same quantity. MR estimates the consequence of a lifelong, genetically set difference in circulating protein, present from conception and acting before any tumor exists; tissue differential expression compares an established tumor with adjacent myometrium at a single moment and therefore captures whatever the tumor has done to its own transcriptional program. A protein whose lifelong abundance lowers the probability that a fibroid ever forms can be reactively up-regulated inside one that has formed, and the two observations are then consistent rather than contradictory. Three further considerations apply here. Our MR instruments plasma protein, whereas the tissue data measure local transcript, and for secreted, endocrine or paracrine molecules these need not move together—FSHB is the clearest case, acting on the ovary and not expressed in the uterus at all. The tissue datasets are small (15 and 6 matched pairs) and cross-sectional, so they establish association with prevalent disease, not direction of causation. And the discordance is not uniform: it is statistically significant only for WNT9A, whereas ITGA11 and EIF2AK3 were not significantly differential in our own data and the pro-fibrotic evidence for them comes from published tissue and mouse work rather than from this analysis. We conclude that the tissue data neither confirm nor refute the MR estimates for these three proteins, and we weight them accordingly: PARP1 and RSPO3, where MR and tissue agree, are the two nominations we regard as supported by both lines of evidence, and the concordant direction in fibroid tissue is one of the reasons PARP1 leads. ITGA11 and EIF2AK3 are protective in our MR (higher protein, lower fibroid risk), yet in tissue and mouse models both behave pro-fibrotically (ITGA11 protein is reported elevated in fibroids, and PERK promotes myofibroblast reprogramming). 57 , 69 , 70 For ITGA11 specifically, a further caveat applies: its sentinel cis-pQTL is a protein-altering missense variant (rs2306022, Val433Met) rather than a regulatory variant, so the plasma readout may partly reflect altered antibody-epitope binding rather than true abundance—a recognized caveat for cis-pQTL at protein-altering variants—and ITGA11 is not covered by SomaScan for orthogonal confirmation; colocalization nonetheless establishes the ITGA11 locus as causally shared with fibroid risk, and we interpret it as a coding-variant/locus signal accordingly. The discordance nonetheless tempers confidence in these two targets. Two consequences follow, and we state them plainly. Both proteins would have to be augmented to reproduce the protective direction, whereas every agent that exists for either of them—neutralizing anti-integrin antibodies for ITGA11, PERK inhibitors for EIF2AK3—acts in the opposite direction, so there is at present no pharmacological route to test either hypothesis in the direction our data imply. Chronic PERK activation is moreover toxic, which makes EIF2AK3 augmentation unattractive even if an agent existed. ITGA11 and EIF2AK3 are therefore hypothesis-generating targets for mechanistic work, not repurposing candidates, and we do not present them as the latter. A related discordance affects WNT9A, which is risk-increasing in MR yet significantly lower in fibroid tissue (the only statistically significant tissue/MR discordance among the six); as a secreted Wnt ligand whose local transcript level need not track its systemic, lifelong causal action, this discordance—together with WNT9A’s weaker, prior-sensitive colocalization—similarly tempers confidence in this target. Our study asks a different question from the largest fibroid GWAS to date (Kim et al; 74,294 cases across ancestries). That study advanced gene discovery at the level of DNA variants and predicted transcripts; it did not analyze proteins, pQTL or drug targets, and none of our six targets appeared among its reported loci or transcriptome-wide genes. 14 We use its European stratum as a second outcome cohort and its Asian and African strata as independent cross-ancestry evidence, so the two studies are complementary rather than competing. By moving from associated variants and predicted mRNA to causal, circulating, druggable proteins, drug-target MR resolves nominations that genome- and transcriptome-level analyses could not. Earlier plasma-proteome MR that touched fibroids did so only within multi-disease or phenome-wide screens on a single platform—a deCODE-based 18-disease screen that reported THBS3 (not any of our targets), and a multi-ancestry proteome-phenome-wide atlas spanning hundreds of phenotypes—in which fibroids, if examined at all, were one outcome among many and were not the analytical focus; 25 , 26 protein-level MR of fibroids has otherwise been limited to small candidate panels. 23 , 24 Our dedicated, two-platform, multi-method design, requiring colocalization in two fibroid cohorts, is to our knowledge the first of its kind for this disease. RSPO3 and FSHB have previously been nominated by plasma-proteome MR as drug targets for endometriosis, 71 , 72 a condition genetically correlated with uterine fibroids; a reviewer might therefore ask whether their fibroid signals merely reflect shared reproductive-tract architecture. To address this directly, we colocalized all six targets with endometriosis (FinnGen). FSHB (PP4 0.999) and ITGA11 (0.945) shared a colocalizing signal with endometriosis, and RSPO3 was borderline (0.808); these three are therefore best regarded as shared reproductive-tract targets (consistent with, rather than artefacts of, their shared genetic architecture), whereas PARP1 (PP4 0.006), WNT9A (0.063) and EIF2AK3 (0.005) were fibroid-specific. Notably, PARP1, our flagship repurposing candidate, is fibroid-specific and is therefore not explained by shared endometriosis biology. We accordingly frame RSPO3, FSHB and ITGA11 as shared reproductive-tract targets and PARP1, WNT9A and EIF2AK3 as fibroid-specific. RSPO3’s standing as a second lead therefore rests on its prior-robust dual-cohort colocalization and fibroid-tissue up-regulation, with its shared endometriosis signal and weaker cross-platform (SomaScan) colocalization (PP4 0.26) acknowledged as trade-offs relative to the fibroid-specific PARP1. The design’s main strength is that prioritization was hypothesis-free but the threshold for being carried forward was high, and that the pipeline recovered a target whose biology is already clinically validated in this disease (FSHB) without being told to look for it.

First, the two outcome datasets are not independent of each other: the Kim European meta-analysis includes FinnGen release 8 of the same endpoint and UK Biobank, so agreement between the cohorts removes cohort- and release-specific artefacts but is not replication in an unrelated sample, and the exposure (UK Biobank) overlaps the outcome for the same reason. Colocalization is not a two-sample estimator and is unaffected by overlap, and the Asian and African analyses contain neither FinnGen nor European-ancestry UK Biobank participants; at the European cis lead variant, outcome directions were concordant in an independent Asian meta-analysis (a single-variant look-up, not within-ancestry MR or formal replication; FSHB, ITGA11 and RSPO3 also concordant in African ancestry), offering limited support for generalizability, while the pQTL exposure remains European and the African subset was underpowered. Second, the six prioritized targets are instrumented by single cis variants, so multi-SNP pleiotropy diagnostics (MR-Egger, Cochran’s Q, MR-PRESSO) are mathematically undefined; with one instrument, horizontal pleiotropy cannot be tested directly and is addressed only indirectly through colocalization, a cis-pleiotropy screen and PheWAS. Instrument weights were taken from the discovery proteomic dataset, so effect-size magnitudes are subject to winner’s curse, and the odds ratios index causal direction and relative support rather than the effect expected from a given drug dose (Methods). Third, no target is a bulk-tissue eGene, so we could not perform tissue cis-eQTL colocalization; the disease-tissue evidence is expression-level (descriptive) rather than causal. Fourth, the SomaScan platform does not cover three targets and measures a different analyte for FSHB. Fifth, RSPO3, EIF2AK3 and WNT9A show appreciable phenome-wide pleiotropy, and WNT9A (significantly down-regulated in fibroid tissue despite a risk-increasing MR estimate), ITGA11 and EIF2AK3 show tissue/MR direction discordance, warranting caution. Sixth, colocalization at the default prior is the basis of our classification, and two targets (WNT9A and, to a lesser degree, PARP1) do not survive a ten-fold more stringent prior, in either cohort, so their standing rests on the default prior together with their orthogonal evidence rather than on colocalization alone. Finally, MR estimates lifelong average effects in a European-ancestry population; the nominations require confirmation in independent and non-European samples, in tissue, and ultimately in clinical studies before any of them can be regarded as validated.