Network-informed multi-trait genomic analysis decodes the shared genetic architecture and therapeutic landscape of pelvic floor disorders Authors: Jiawen Wang, Yupeng Chen, Lingfeng Meng, Ruijie Yao, Jiong Zhang, Qingguo Zhu, Jinfeng Wu, Liefu Ye, Yaoguang Zhang Published: 2026 ABSTRACT AbstractPelvic floor disorders, including pelvic organ prolapse and urinary incontinence, represent a common health burden with substantial clinical comorbidity, but their shared genetic architecture remains incompletely understood. We performed a multi-trait genomic analysis of six pelvic-floor-related phenotypes using publicly available GWAS summary statistics from FinnGen R12 and the GWAS Catalog, with Pan-UKBB summary statistics used only for cross-ancestry validation. Linkage disequilibrium score regression and Genomic Structural Equation Modeling identified two broad latent genetic dimensions: a structural factor related to anatomical prolapse and a functional factor related to urinary and bowel dysfunction. MAGMA gene-based analysis identified 267 significant genes enriched for extracellular matrix organization and urogenital developmental pathways, including WNT4, LOXL1, ESR1, WT1, HNF1B, and FGFR2. An exploratory network-informed machine-learning framework incorporating protein-protein interaction topology improved gene prioritization over a baseline Random Forest model and highlighted biologically plausible hub genes. Cross-ancestry analysis of Pan-UKBB female genital prolapse supported directional portability of European-discovered signals and identified WNT4 as a directionally concordant locus. These findings provide a systems-level map of pelvic-floor-related genetic architecture and prioritize candidate genes for future functional validation.Subject terms: Computational biology and bioinformatics, Diseases, Genetics, Medical research, Urology Introduction Pelvic floor disorders (PFDs)—encompassing pelvic organ prolapse (POP), urinary incontinence, and functional defecatory dysfunction—represent a pervasive global health burden, affecting nearly 25% of adult women and significantly impacting quality of life1,2. Despite their distinct clinical presentations, these conditions frequently co-occur, suggesting a shared underlying susceptibility rooted in the structural compromise of the pelvic support system3. While epidemiological evidence has long pointed to a strong heritable component, our understanding of the specific genetic architecture driving these disorders remains fragmented4. Previous genome-wide association studies (GWAS) have largely been confined to single phenotypes or limited by modest sample sizes, failing to capture the pervasive pleiotropy that characterizes pelvic floor pathology5. Furthermore, the translation of statistical associations into actionable biological insights has been hindered by the limitations of traditional linear mapping approaches, which often struggle to prioritize causal genes within complex, non-linear biological networks6,7. To address these challenges, we implemented a comprehensive systems biology framework that integrates multi-trait genomic analysis with network-informed machine-learning gene prioritization. Rather than generating new individual-level GWAS, this study used publicly available GWAS summary statistics from FinnGen R12 and the GWAS Catalog for six clinically relevant pelvic-floor-related phenotypes. This mixed-source design was adopted because no single public database contained adequately powered summary statistics for all phenotypes of interest. By applying genomic structural equation modeling (Genomic SEM), we moved beyond simple disease classification to resolve the phenotypic complexity into two robust latent dimensions: a structural factor driving anatomical prolapse and a functional factor underlying urinary dysfunction. This distinction provides the first genetic evidence decoupling the pathways of mechanical support failure from neuromuscular dysfunction8,9. To bridge the gap between genetic loci and candidate therapeutic targets, we implemented an exploratory network-informed gene-prioritization framework that integrated GWAS-derived gene-level evidence, pleiotropy metrics, protein-protein interaction topology, and druggability annotations. This approach was designed to prioritize biologically plausible genes that may not be ranked highly by association P values alone. The model was evaluated using curated OMIM/HPO-supported pelvic-floor-related genes and independent validation sets, including cross-ancestry directionally concordant loci, pre-GWAS literature candidates, and known drug targets. These in silico predictions should be interpreted as hypothesis-generating and require experimental validation. Here, we present a high-resolution map of the genetic architecture of PFDs. We identify transcriptome-wide significant loci dominated by regulators of extracellular matrix (ECM) organization and urogenital development, including WNT4 and LOXL110,11. We further evaluated the cross-ancestry portability of selected European-discovered signals using Pan-UKBB ancestry-stratified summary statistics, providing supportive directional evidence for selected loci while recognizing the limited non-European case numbers. Finally, via exploratory drug-target interaction mapping, we highlight 30 potential repurposing candidates that may warrant future experimental and clinical evaluation. This study not only elucidates the shared genetic basis of PFDs but also establishes a scalable network-informed computational framework for decoding the complex genetics of structural diseases. Results Study design and multi-trait analytic framework To systematically dissect the genetic architecture of pelvic-floor-related disorders, we implemented a multi-trait genomic analysis framework integrating GWAS summary statistics harmonization, LDSC-based genetic correlation analysis, GSEM, MAGMA gene-based association testing, Mendelian randomization, and network-informed gene prioritization (Fig. 1). Discovery-stage analyses used publicly available GWAS summary statistics from FinnGen R12 and the GWAS Catalog across six phenotypes: POP, female genital prolapse, urinary incontinence, benign prostatic hyperplasia, bladder dysfunction, and constipation. Pan-UKBB summary statistics were used only for cross-ancestry validation and were not included in the discovery-stage analyses. The exact data source, accession or phenocode, sample size, genome build, and download link for each GWAS dataset are listed in Supplementary Table 1. This design was adopted because no single public resource provided well-powered GWAS summary statistics for all pelvic-floor-related phenotypes analyzed in this study. Analytical framework integrating publicly available GWAS summary statistics, multi-trait genomic analysis, and network-informed gene prioritization for pelvic-floor-related disorders. Discovery-stage GWAS summary statistics were obtained from FinnGen R12 and the GWAS Catalog; Pan-UKBB was used only for cross-ancestry validation. LDSC linkage disequilibrium score regression, GSEM genomic structural equation modeling, MTAG multi-trait analysis of GWAS, MAGMA multi-marker analysis of genomic annotation, MR Mendelian randomization, GNN graph neural network, PPI protein–protein interaction. Shared genetic architecture and latent structure of pelvic floor disorders We first quantified the shared genetic basis of PFDs using linkage disequilibrium score regression (LDSC). LDSC intercepts were close to 1.0 for all traits (range: 0.991–1.108), confirming minimal confounding by population stratification. Mean χ² statistics ranged from 1.036 (Bladder) to 1.364 (FemaleProlapse), with genomic inflation factors (λ_GC) between 1.037 and 1.321. The bar plot illustrates the SNP-based heritability (h2) on the liability scale for the six studied phenotypes, with error bars representing standard errors. Consistent with clinical observations, Female Genital Prolapse exhibited the highest heritable component (h2=0.027±0.005), followed by Benign Prostatic Hyperplasia (h2=0.022) and general POP (h2=0.018). In contrast, functional bladder disorders displayed lower heritability estimates (h2=0.003), indicating a greater influence of environmental or non-additive genetic factors on these traits. Pairwise genetic correlation analysis revealed significant shared genetic architecture: 14 of 15 pairwise genetic correlations were statistically significant (P < 0.05; range rg = 0.16–0.95), with only POP–Bladder failing to reach significance (rg = 0.16, P = 0.091) (Fig. 2A and Supplementary Fig. 1). Pairwise analysis demonstrated a near-perfect genetic correlation between general POP and female-specific prolapse (rg = 0.95, P < 0.001), alongside a distinct cluster of urinary traits where incontinence showed moderate-to-strong correlations with benign prostatic hyperplasia (rg = 0.45) and bladder disorders (rg = 0.38). To systematically disentangle these correlations, we applied GSEM, exploratory factor analysis identified two factors with eigenvalue > 1 (eigenvalue₁ = 2.74, variance explained = 55.6%; eigenvalue₂ = 1.31, 29.4%), consistent with the Kaiser criterion. We compared three CFA models (Supplementary Table 1): a single-factor model (χ² = 94,134, df = 9, CFI = 0.743, SRMR = 0.159), a two-factor model separating FemalePelvic and Urinary dimensions (χ² = 16,862, df = 3, CFI = 0.948, SRMR = 0.125), and a three-factor model (χ² = 50,287, df = 7, CFI = 0.863, SRMR = 0.041). The two-factor model provided the best balance of fit (highest CFI) and resolved the phenotypic complexity into two discrete latent dimensions (Fig. 2B and Supplementary Fig. 2). The first factor, reflecting structural pelvic integrity, was predominantly defined by Female Prolapse (standardized loading λ = 1.97) and POP (λ = 0.48), while the second factor captured a spectrum of functional urinary dysfunction driven by Incontinence (λ = 0.68), Bladder disorders (λ = 0.67), and benign prostatic hyperplasia (BPH, λ = 0.43). POP and FemaleProlapse showed very high genetic correlation (rg = 0.95, P = 2.4 × 10⁻¹⁷³), consistent with their overlapping clinical definitions. The standardized loading of 1.97 with negative residual variance (−0.078) constitutes a Heywood case, indicating that the latent factor accounts for more than the total observed genetic variance, likely reflecting the near-identity of POP and FemaleProlapse at the genetic level. Notably, the negligible loading of Incontinence on the structural factor (λ = −0.09) reinforces the existence of genetically distinct pathways driving anatomical prolapse versus functional incontinence, despite their frequent clinical comorbidity. Shared genetic architecture and latent factor structure of pelvic floor phenotypes. A Heatmap of pairwise genetic correlations (rg) estimated by LDSC across the six pelvic-floor-related phenotypes. Color intensity and numerical values represent the magnitude of genetic correlation; asterisks denote significant correlations (P < 0.05). B Path diagram of the best-fitting two-factor genomic structural equation model (CFI = 0.948, SRMR = 0.125). Ovals represent latent genetic factors (FemalePelvic and Urinary); rectangles represent observed phenotypes; single-headed arrows represent standardized factor loadings (λ); double-headed arrows represent residual variances. POP pelvic organ prolapse, FemaleProlapse female genital prolapse, Incontinence urinary incontinence, BPH benign prostatic hyperplasia, Bladder bladder dysfunction, Constipation constipation. Transcriptome-wide association identifies key developmental and structural regulators Prior to gene-level analysis, we evaluated the reliability of MTAG-boosted summary statistics. MTAG maxFDR was below 0.08 for five of six traits (POP: 0.002, BPH: 0.014, FemaleProlapse: 0.002, Incontinence: 0.067, Constipation: 0.077), indicating that the multi-trait boosted associations are unlikely to be false discoveries. Bladder showed a higher maxFDR of 0.223, consistent with its low heritability (h² = 0.003), and MTAG results for this trait should be interpreted with caution. To translate variant-level associations into actionable gene targets, we conducted a gene-based association analysis using MAGMA, aggregating SNP signals while accounting for linkage disequilibrium and gene density. This gene-based scan identified MAGMA-significant genes exceeding the Bonferroni-corrected threshold P<2.5×10−6, revealing a genetic landscape dominated by regulators of connective tissue integrity and urogenital development (Fig. 3). The complete MAGMA gene-level results, including phenotype-specific ranks, gene symbols, Entrez IDs, chromosomal positions, mapped SNP counts, MAGMA Z statistics, gene-level P values, and phenotype-specific association patterns, are provided as Supplementary Data 1 and Supplementary Data 2. These supplementary Excel files were added to allow independent verification of the MAGMA-identified genes, with corresponding titles and legends provided in the Supplementary Information PDF. The Manhattan plot displays distinct genomic peaks with high-confidence biologically plausible candidates, including the estrogen receptor ESR1 and the elastic fiber component LOXL1, underscoring the hormonal and structural etiology of PFDs. Notably, the analysis highlighted a cluster of developmental transcription factors—such as WNT4, HNF1B, WT1, and FGFR2—whose signals significantly exceed background noise (P < 10⁻¹⁵), suggesting that susceptibility to pelvic floor dysfunction may stem from subtle perturbations in embryonic urogenital tract morphogenesis rather than solely from acute mechanical failure. Transcriptome-wide association analysis identifies key developmental and extracellular matrix regulators. Manhattan plot displaying MAGMA gene-based association results. The y-axis shows −logsub10/sub(P) values; the red dashed line indicates the Bonferroni-corrected significance threshold (P < 2.5 × 10−6). Labeled genes represent biologically plausible candidates exceeding genome-wide significance, including WNT4, LOXL1, ESR1, WT1, HNF1B, and FGFR2. Superior gene prioritization via graph neural network ensemble To address the limitations of association P-value-based prioritization, we implemented an exploratory network-informed machine-learning framework that integrated gene-level GWAS evidence, pleiotropy metrics, protein–protein interaction topology, and druggability annotations. Each gene was treated as one observation. Training labels were derived from a curated set of OMIM/HPO-supported pelvic-floor-related genes, while genes without curated evidence were treated as background genes. The baseline Random Forest model used 200 trees, a maximum tree depth of 6, balanced class weights, and five-fold stratified cross-validation. The reported Random Forest AUC of 0.638 represents the mean cross-validated ROC-AUC for distinguishing curated positive genes from background genes. A graph-based ensemble incorporating network topology achieved a cross-validated AUC of 0.698, suggesting that PPI network information provided additional prioritization value beyond gene-level GWAS signals alone. Detailed model features, training labels, hyperparameters, and validation procedures are described in the Methods section. Independent validation underscored the model’s robust ability to identify biologically relevant signals: prioritized genes were significantly enriched for cross-ancestry replicated loci (observed median rank = 3 vs. expected 4551; P < 1 × 10⁻⁴) and pre-GWAS literature candidates (P < 1 × 10⁻⁴), while successfully retrieving known drug targets (P = 0.0095). The top-ranked genes (Fig. 4) formed a coherent pathogenic core, placing the fibroblast growth factor receptor FGFR2 (Rank 1) and developmental regulator WT1 (Rank 2) at the apex, followed by HNF1B and WNT4. The horizontal bar chart ranks the importance of node features utilized in the gradient boosting classifier, which represents the final stage of the ensemble model. Among the features, GWAS-derived summary statistics, particularly the maximum Z-score (max_z) and the mean Z-score (mean_z) across phenotypes, emerged as the dominant predictors of disease association. Additionally, topological features such as PageRank and Betweenness centrality, although contributing minor influences, provided valuable and non-redundant insights. These topological features aided in the prioritization of network hubs that exhibit moderate statistical association signals. Feature importance analysis (Supplementary Fig. 3) revealed that while GWAS effect sizes (Z-scores) remained the dominant predictor (importance = 0.24), network centrality measures also contributed significantly to classification power, allowing the identification of “silent” hub genes that may be overlooked by effect size alone. Network-informed machine-learning framework prioritizes candidate genes. A Ranked bar chart of the top 20 prioritized genes from the graph-based ensemble model, ordered by ensemble prioritization score. Gene symbols are shown on the y-axis; bar color indicates functional annotation category (developmental, ECM, signaling, other). B Feature importance plot from the gradient boosting classifier (final stage of the ensemble), ranking the contribution of individual node features to classification performance. GWAS-derived summary statistics (max_Z, mean_Z) and pleiotropy metrics dominate as predictors, while topological features (PageRank, Betweenness) provide complementary, non-redundant information. AUC area under the receiver operating characteristic curve, PPI protein–protein interaction. Cross-ancestry portability and core genetic determinants We next evaluated the cross-ancestry portability of European-discovered signals using publicly available Pan-UKBB summary statistics. Because ancestry-stratified Pan-UKBB data were consistently available for female genital prolapse, this validation was restricted to the N81 phenotype and should not be interpreted as replication of the full multi-trait discovery framework. Projection of European-discovered variants into the African ancestry stratum showed 76.4% directional concordance, with a statistically significant sign-test result and a positive correlation in effect estimates. The WNT4 locus showed consistent effect directionality across European, African, and Central/South Asian strata, supporting it as a directionally concordant cross-ancestry locus. However, because the non-European case numbers were limited, especially in the African and Central/South Asian strata, these findings should be interpreted as supportive evidence of portability rather than definitive trans-ethnic replication (Fig. 5). Cross-ancestry directional portability of European-discovered risk loci. A Scatter plot comparing effect allele effect sizes (beta) for European-discovered variants between the European discovery cohort (x-axis) and the Pan-UKBB African ancestry stratum (y-axis). Each point represents one variant; the red diagonal line indicates perfect concordance. Directional concordance = 76.4%; sign test P < 0.05. B Forest plot of effect sizes for the top cross-ancestry concordant loci (WNT4, WT1, ESR1, LOXL1) across European, African, and Central/South Asian ancestry strata. Points represent beta estimates; horizontal lines represent 95% confidence intervals. WNT4 showed consistent effect directionality across all three ancestry strata. Functional genomic integration and pathway convergence To elucidate the biological mechanisms driven by these prioritized loci, we integrated expression quantitative trait loci (eQTL) colocalization and protein interaction network analyses (Fig. 6). Pharmacogenomic network and drug repurposing candidates targeting the PFD pathogenic core. The protein–protein interaction subnetwork of top-prioritized genes is displayed, with nodes colored by functional category (developmental regulators, ECM components, signaling hubs) and edges representing STRING experimental and database interactions (combined score ≥400). Drug repurposing candidates mapping to hub nodes are annotated with DrugBank identifiers. ECM extracellular matrix, TGF-β transforming growth factor-beta. Colocalization analysis provided strong evidence for functional regulation in disease-relevant tissues: WNT4 significantly colocalized with expression signals (defined as PPH4 > 0.8) in the uterus, vagina, and connective tissue. In our evaluation, the notation “Score 5/5” indicates that these stringent colocalization criteria were successfully met in all 5 specific prioritized tissue types tested within that category. Similarly, WT1 and ESR1 exhibited specific regulatory effects in urogenital (Score 5/5) and reproductive tissues (Score 4/5), respectively, directly linking genetic variation to tissue-specific transcriptomic changes. Furthermore, mapping the top-prioritized genes onto the PPI network revealed a dense functional module centered on ECM organization and TGF-β signaling. This network is orchestrated by key hub genes, including SMAD3, TGFBR2, and COL1A1, confirming that genetic liability for PFDs converges on the disruption of collagen fibrillogenesis and connective tissue maintenance pathways. Causal inference via mendelian randomization Two-sample bidirectional MR using IVW, weighted median, and MR-Egger methods identified several significant causal relationships among PFD phenotypes (Fig. 7). The strongest evidence supported a bidirectional causal relationship between BPH and Incontinence: genetically predicted BPH increased Incontinence risk (IVW β = 0.168, P = 1.03 × 10⁻¹¹; 38 instruments) with consistent estimates across all sensitivity methods and no evidence of horizontal pleiotropy (Egger intercept P = 0.774). The reverse direction was also significant (IVW β = 0.331, P = 2.77 × 10⁻⁸; 5 instruments), though limited instrument count warrants cautious interpretation. POP showed a unidirectional causal effect on Incontinence (POP: IVW β = 0.086, P = 2.36 × 10⁻³; FemaleProlapse: IVW β = 0.077, P = 2.86 × 10⁻³), while no reverse causation was detected (Incontinence → POP: P = 0.466; Incontinence → FemaleProlapse: P = 0.350), supporting the hypothesis that anatomical prolapse mechanically contributes to bladder dysfunction. A nominally significant negative effect of POP on BPH (β = −0.071, P = 8.01 × 10⁻³) may reflect sex-specific genetic architecture, though evidence of horizontal pleiotropy (Egger intercept P = 0.021) suggests this association should be interpreted with caution. Bidirectional Mendelian randomization reveals causal relationships among pelvic floor disorder phenotypes. A Circos-style summary of significant MR associations (IVW P < 0.05). Arrow direction indicates the forward MR exposure→outcome direction; arrow thickness is proportional to the IVW beta estimate. Red arrows denote bidirectional associations; blue arrows denote unidirectional associations. B Scatter plot of SNP-exposure effects versus SNP-outcome effects for the BPH→Incontinence analysis (IVW β = 0.168, P = 1.03 × 10−11). Each point represents one instrumental variant; the slope of the fitted line represents the IVW causal estimate. Error bars represent standard errors of variant-specific effects. IVW inverse-variance weighted, MR Mendelian randomization, BPH benign prostatic hyperplasia, POP pelvic organ prolapse. Discussion In this study, we present a systems biology framework to dissect the genetic architecture of pelvic-floor-related disorders using publicly available GWAS summary statistics from FinnGen R12 and the GWAS Catalog. By integrating six clinically relevant phenotypes, we resolved the genetic correlation structure into two broad latent dimensions—structural pelvic integrity and functional urinary or bowel-related dysfunction—and prioritized candidate genes enriched for ECM organization and urogenital developmental pathways. Importantly, no individual-level genotype or phenotype data from UK Biobank, FinnGen, deCODE, or other biobanks were accessed in this study; all analyses were based on publicly available summary statistics12,13. Our approach extends traditional single-phenotype association analyses by incorporating network-informed gene prioritization, suggesting that protein–protein interaction topology can help prioritize biologically plausible hub genes such as FGFR2 and WT1 that may not be ranked highly by association P values alone. Furthermore, the directional concordance of selected loci, particularly WNT4, across Pan-UKBB ancestry strata provides supportive evidence for cross-ancestry portability, although larger non-European datasets will be required for definitive trans-ethnic replication14. A pivotal finding of our genomic structural equation modeling is the clear delineation of PFDs into two genetically distinct axes: a “structural” factor driven by POP and a “functional” factor encompassing urinary incontinence and bladder dysfunction. Clinically, these conditions frequently co-occur, leading to the historical assumption of a single “pelvic floor failure” syndrome15,16. However, our data suggests that while they share substantial pleiotropy (rg up to 0.95), the primary genetic drivers diverge significantly. The structural factor is heavily loaded by genes involved in connective tissue tensile strength (e.g., LOXL1, ELN), reflecting a mechanical failure of the pelvic support system17,18. In contrast, the functional factor likely captures neuromuscular deficits or urothelial signaling abnormalities, as evidenced by the negligible loading of incontinence on the structural dimension19,20. This genetic uncoupling has profound clinical implications: it suggests that surgical repair of anatomical defects (prolapse) may not address the underlying genetic susceptibility to functional symptoms (incontinence), potentially explaining the high persistence of urinary symptoms post-surgery21,22. Future therapeutic strategies may need to stratify patients based on their “structural” versus “functional” polygenic liability to optimize treatment outcomes. Our factor structure revealed a clear sex-stratified genetic architecture: Factor 1 (“FemalePelvic”) was defined by female-specific conditions (POP, FemaleProlapse), while Factor 2 (“Urinary”) captured male-predominant and sex-shared conditions (BPH, Bladder, Incontinence). The cross-loading of Incontinence on both factors (λ_FemalePelvic = −0.09, λ_Urinary = 0.68) suggests it represents a convergent phenotype bridging sex-specific pathophysiological pathways. Future sex-stratified GWAS with individual-level data would be needed to formally decompose shared versus sex-specific genetic effects. Our transcriptome-wide analysis unequivocally positions ECM organization and urogenital development as the central biological engines of PFD risk23. The identification of LOXL1—a critical enzyme for elastin fiber assembly—as a top signal reinforces the “connective tissue hypothesis,” wherein genetic variations in collagen fibrillogenesis compromise the visceral fascia’s ability to withstand intra-abdominal pressure24–26. More intriguingly, the prominence of developmental transcription factors, specifically WNT4, WT1, HNF1B, and HOXA13, supports a “developmental origin” etiology for adult PFDs. WNT4 is essential for Müllerian duct formation and the patterning of the female reproductive tract27–29. Its directional concordance across European, African, and Central/South Asian Pan-UKBB strata suggests that subtle perturbations in embryonic organogenesis may contribute to pelvic-floor vulnerability, although this observation requires confirmation in larger non-European datasets, predisposing individuals to mechanical failure decades later under the stress of childbirth or aging30,31. This aligns with the eQTL evidence showing WNT4 and WT1 exert regulatory effects specifically in uterine and vaginal tissues, providing a direct functional link between developmental variants and adult tissue homeostasis32,33. A key methodological component of this work is the use of network-informed machine learning for exploratory gene prioritization. Traditional post-GWAS approaches often rely on genomic proximity, gene-level P values, or simple overlap with annotation resources, which may miss genes with moderate statistical evidence but strong network relevance. In our framework, the baseline Random Forest model achieved a mean five-fold cross-validated ROC-AUC of 0.638, while the graph-based ensemble achieved a cross-validated AUC of 0.698. These results suggest that protein-protein interaction topology provides complementary information to GWAS-derived gene-level statistics. Nevertheless, the model should be interpreted as a hypothesis-generating prioritization tool rather than a clinically validated predictor or proof of causality. The prioritization of FGFR2, WT1, HNF1B, and WNT4 highlights biologically plausible developmental and extracellular-matrix-related pathways that warrant further experimental validation. Ultimately, the goal of genomic discovery is clinical translation. By mapping our GNN-prioritized genes to the druggable genome, we identified 30 potential repurposing candidates. The network highlights the estrogen signaling pathway (ESR1) and TGF-β signaling (TGFBR2, SMAD3) as key modulation points34–36. This provides a genetic rationale for the clinical use of local estrogen therapy, which is currently employed empirically. More novel is the identification of FGFR2 inhibitors (e.g., Nintedanib) and modulators within the network37. While systemic use of such potent agents is likely inappropriate for benign conditions due to side effects, this finding directs attention toward localized delivery systems or downstream effectors that could enhance pelvic tissue regeneration without systemic toxicity. The Mendelian randomization results further suggest potential causal relationships among pelvic-floor-related phenotypes, particularly between prolapse-related traits and urinary dysfunction, supporting the need to consider structural and functional components jointly in future mechanistic studies. From a translational perspective, WNT4 (ensemble rank 4, functional evidence score 5/5) represents a compelling candidate for preventive screening. Its established role in Müllerian duct development and cross-ancestry directional concordance, with statistical significance in European data and supportive directionality in African ancestry data, suggest it may represent a candidate risk locus for prolapse pending further validation. WT1 (ensemble rank 2, functional evidence score 5/5), with its known function in urogenital development and eQTL evidence in urogenital tissues, could inform post-surgical recurrence risk stratification. The convergence of multiple top-ranked genes (SMAD3, TGFBR2, COL1A1) on the TGF-β/ECM axis suggests this pathway as a potential therapeutic target for combinatorial intervention. We acknowledge that using known drug targets as one validation set introduces circular reasoning risk. However, our validation strategy employed three independent gene sets: (1) cross-ancestry replicated genes (WNT4, WT1), which are agnostic to druggability; (2) FDA-approved drug targets (ESR1, AGTR1, VEGFA, PGR, PTGS2); and (3) pre-GWAS literature genes known before 2010 (LOXL1, ELN, FBLN5, COL3A1). The convergent enrichment across all three sets (permutation P < 0.01 for each) supports the model’s ability to prioritize biologically relevant genes beyond drug-target bias. The strengths of this study lie in its multi-trait design, the integration of deep learning, and the incorporation of cross-ancestry validation. However, several limitations must be acknowledged. First, the discovery analyses combined GWAS summary statistics from FinnGen R12 and the GWAS Catalog rather than from a single homogeneous cohort. This mixed-source design was necessary because no single public database provided well-powered summary statistics for all pelvic-floor-related phenotypes. However, the datasets differed in ancestry composition, phenotype definition, control selection, genome build, imputation panel, and study design. Although harmonization procedures and sensitivity analyses were applied where possible, residual heterogeneity cannot be fully excluded. Second, because only summary statistics were used, we could not re-define phenotypes, perform uniform individual-level quality control, adjust covariates consistently across cohorts, or directly quantify sample overlap. Third, cross-ancestry validation was restricted mainly to Pan-UKBB female genital prolapse summary statistics and was limited by small non-European case numbers; therefore, these analyses should be interpreted as directional portability assessments rather than definitive independent replication. Finally, the heritability estimates (h2) indicate that a significant portion of phenotypic variance remains unexplained, suggesting that rare variants or environmental interactions not captured in this study play a substantial role. Additionally, non-European sample sizes (AFR N_cases = 106, CSA N_cases = 181) were insufficient for reliable independent discovery, and the cross-ancestry analysis was limited to a single phenotype (N81, Female Genital Prolapse); these results should be interpreted as directional evidence of generalizability rather than definitive trans-ethnic replication. Sensitivity analysis excluding FemaleProlapse yielded a two-factor model with acceptable fit (CFI = 1.00, SRMR = 0.003), though POP then exhibited a Heywood case (loading = 1.19, residual variance = −0.41), indicating that these two phenotypes share a common latent genetic dimension. These results support the robustness of the two-factor structure while acknowledging the near-redundancy of POP and FemaleProlapse. In summary, this study provides the most detailed map of the genetic landscape of PFDs to date. By decoupling the genetic drivers of structural versus functional phenotypes and applying network-informed gene prioritization, we identified a set of developmental and extracellular-matrix-regulating genes that may inform future functional studies and therapeutic exploration. These findings advance our understanding of pelvic-floor-related genetic architecture and demonstrate the value of integrating multi-trait genomics with network-informed computational prioritization. Methods GWAS summary statistics and phenotype definition: This study was based exclusively on publicly available GWAS summary statistics. No individual-level genotype or phenotype data from UK Biobank, FinnGen, deCODE, or any other biobank were accessed or analyzed. Discovery-stage analyses included six pelvic-floor-related phenotypes: POP, female genital prolapse, urinary incontinence, benign prostatic hyperplasia, bladder dysfunction, and constipation. Summary statistics were obtained from FinnGen R12 and the GWAS Catalog. FinnGen R12 was used for benign prostatic hyperplasia, bladder dysfunction, constipation, and female genital prolapse. GWAS Catalog datasets were used for POP and urinary incontinence because these resources provided larger or better harmonized publicly available summary statistics for these traits. The exact source, accession identifier or phenocode, sample size, genome build, file name, and download link for each dataset are provided in Supplementary Table 1. A mixed-source GWAS design was used because no single public database contained well-powered summary statistics for all pelvic-floor-related phenotypes included in this study. To minimize inconsistencies across datasets, summary statistics were harmonized to a standardized format containing chromosome, position, effect allele, non-effect allele, effect size, standard error, P value, allele frequency where available, and rsID where available. Genome builds and allele codings were checked before downstream analyses. Pan-UKBB summary statistics were used only for cross-ancestry validation and were not included in the discovery-stage analyses. This study was based exclusively on publicly available GWAS summary statistics and did not access individual-level genotype, phenotype, or clinical data. Therefore, additional institutional review board approval and additional informed consent were not required for the present secondary analysis. All original studies and biobank resources had obtained ethics approval and informed consent from participants according to their respective protocols. UK Biobank received ethical approval from the North West Multi-centre Research Ethics Committee, with the current ethics approval reference number 21/NW/0157. FinnGen and GWAS Catalog datasets were generated under the ethics approvals and consent procedures described by the original contributing studies. All research involving human participants was conducted in accordance with the Declaration of Helsinki. GWAS Summary-Statistics Harmonization and Multi-trait Analysis. Because this study used publicly available summary statistics only, no individual-level genotype quality control, phenotype extraction, or de novo UK Biobank SAIGE GWAS was performed. Instead, all downloaded GWAS summary statistics were harmonized before analysis. For FinnGen R12 files, the alternate allele was treated as the effect allele and the reference allele as the non-effect allele. For GWAS Catalog files, effect allele definitions were extracted from the original summary-statistics metadata and harmonized to the same format. Variants with missing essential fields, ambiguous allele coding that could not be resolved, or invalid P values were removed. Where necessary, genome coordinates were lifted to a common build before integration. To leverage cross-trait genetic sharing and improve downstream gene-level discovery, MTAG was applied to the harmonized phenotype-specific summary statistics. Genomic SEM and LDSC analyses were performed using pre-MTAG summary statistics to avoid inflating genetic correlation estimates through multi-trait adjustment. MTAG-boosted summary statistics were used only for downstream MAGMA gene-based association analyses. We utilized LDSC to estimate SNP-based heritability (h²) and quantify pairwise genetic correlations (rg) between phenotypes. Analyses were based on pre-computed European LD scores from the 1000 Genomes Project Phase 3. Genomic structural equation modeling To elucidate the latent structure underlying the observed phenotypic correlations, we applied Genomic SEM. We constructed the genetic covariance matrix (S) and its sampling covariance matrix (V) using the LDSC outputs. We first performed an Exploratory Factor Analysis (EFA) to determine the optimal number of factors. Two factors with eigenvalue > 1 were identified (eigenvalue₁ = 2.74, variance explained = 55.6%; eigenvalue₂ = 1.31, 29.4%), consistent with the Kaiser criterion. We then conducted confirmatory factor analysis (CFA) using the diagonally weighted least squares (DWLS) estimator, comparing three models: a single-factor model, a two-factor model separating FemalePelvic and Urinary dimensions, and a three-factor model adding a Bowel dimension (see Supplementary Table 1 for fit indices). Transcriptome-wide association analysis Gene-level associations were computed using MAGMA v1.10. Harmonized GWAS summary statistics were mapped to genes using the MAGMA gene-location file NCBI37.3.gene.loc and the 1000 Genomes European reference panel. SNPs were assigned to protein-coding genes according to genomic position, and MAGMA gene-based tests were performed while accounting for local linkage disequilibrium. Statistical significance was defined using a Bonferroni-corrected threshold of P<0.05/20,000≈2.5×10−6. Genomic SEM analyses used pre-MTAG summary statistics, whereas MTAG-boosted summary statistics were used exclusively for downstream MAGMA analyses to maximize gene-level discovery power for individual phenotypes. We did not perform GWAS directly on latent factors. The Genomic SEM factor structure was used to interpret shared genetic architecture, while gene-level discovery was performed at the individual-phenotype level. The complete MAGMA gene-based association results are provided as Supplementary Data 1 and 2, with corresponding titles and legends included in the Supplementary Information PDF. Network-informed machine-learning gene prioritization Gene prioritization was performed using an exploratory network-informed machine-learning framework. Each gene was treated as one observation. Input features included gene-level GWAS evidence, cross-phenotype pleiotropy, STRING protein-protein interaction topology, and druggability annotations. GWAS-derived features included the minimum MAGMA gene-level value across phenotypes, normalized−log10(P), maximum and mean MAGMA Z statistics, and the number of mapped SNPs. Pleiotropy was defined as the number of phenotypes with nominal gene-level association P|0.05. Network features, including degree, betweenness, closeness, PageRank, and clustering coefficient, were calculated from STRING v11.5 using a combined confidence score threshold of 400. Druggability annotations were obtained from DGIdb v5.0 and DrugBank v5.1. Training labels were defined using a curated set of OMIM/HPO-supported pelvic-floor-related, urogenital-development, urinary-dysfunction, and connective-tissue-related genes. Genes not included in the curated positive set and with complete feature information were treated as background genes. Because of class imbalance, balanced class weights were used during model training. A Random Forest classifier was used as the baseline model with 200 trees, a maximum depth of 6, balanced class weights, and five-fold stratified cross-validation. The reported Random Forest AUC of 0.638 represents the mean cross-validated ROC-AUC for distinguishing curated positive genes from background genes. A graph-based ensemble incorporating STRING protein-protein interaction topology was then used to generate a relative gene-prioritization score and achieved a mean cross-validated AUC of 0.698. The final scores were used only for candidate-gene ranking and biological interpretation, not as formal statistical evidence of causality or clinical predictive probabilities. Detailed feature definitions, hyperparameters, training labels, and validation procedures are described above in this Methods section. Cross-ancestry replication and functional mapping Cross-ancestry validation was performed using publicly available Pan-UKBB summary statistics. Pan-UKBB was used only for validation and was not included in the discovery-stage GWAS summary-statistics analysis. Because ancestry-stratified summary statistics were consistently available for female genital prolapse, cross-ancestry validation focused on the N81 phenotype across European, African, and Central/South Asian ancestry strata. Directional concordance and effect-size correlation were calculated by comparing European-discovered variants with ancestry-specific effect estimates. Given the small non-European case numbers, especially in the African and Central/South Asian strata, these analyses were interpreted as directional portability assessments rather than definitive independent replication. Colocalization analysis was performed using the MTAG-boosted summary statistics for female genital prolapse and GTEx v8 eQTL data. The analysis was conducted using the coloc R package to assess the posterior probability that GWAS and eQTL signals share a single causal variant (hypothesis 4, PPH4). A significant colocalization was defined by a stringent threshold of PPH4 > 0.8 across prioritized relevant tissues: Uterus, Vagina, and Connective Tissue (fibroblasts). High-priority genes were mapped to the DrugBank v5.1 database for exploratory drug-repurposing analysis. Candidates were ranked based on a composite score of the GNN priority, the number of targeted network hubs, and existing clinical indications. Two-sample Mendelian randomization was conducted using the inverse-variance weighted method to infer putative causal relationships among pelvic-floor-related phenotypes. Supplementary information Supplementary Information Acknowledgements This work was supported by the Fujian Provincial Natural Science Foundation of China (Grant No. 2025J08052); the Joint Funds for the Innovation of Science and Technology, Fujian Province (Grant Nos. 2024Y9080 and 2024Y9019); the Noncommunicable Chronic Diseases-National Science and Technology Major Project (Grant No. 2024ZD0525700); and the Provincial Specialized Disease Diagnosis and Treatment Center Construction Project—Pelvic Floor Dysfunction Diagnosis and Treatment Center, Fujian, P.R.C. Author contributions Jiawen Wang (J.W.), Yupeng Chen (Y.C.), and Lingfeng Meng (L.M.) led the formal analysis and conceptualization of the study and developed the software used in the analysis. Ruijie Yao (R.Y.) was primarily responsible for drafting the manuscript, with substantial support from Jiong Zhang (J.Z.) and Qingguo Zhu (Q.Z.). Jinfeng Wu (J.W.), Yaoguang Zhang (Y.Z.), and Liefu Ye (L.Y.) provided supervision and contributed to revising the manuscript for important intellectual content. All authors contributed significantly to this research. Data availability The data used in this study are publicly available GWAS summary statistics from FinnGen R12, the GWAS Catalog, and Pan-UKBB. FinnGen R12 summary statistics are available at https://r12.finngen.fi/. GWAS Catalog summary statistics are available at https://www.ebi.ac.uk/gwas/summary-statistics/. Pan-UKBB summary statistics are available at https://pan.ukbb.broadinstitute.org/. The exact phenotype-specific source, accession identifier or phenocode, file name, genome build, sample size, and download link for each GWAS dataset are provided in Supplementary Table 1. Code availability Analysis code is available at https://github.com/LI-993/pelvic_floor_gwas. The repository contains the study-specific scripts required to reproduce the final analyses, including GWAS summary-statistics download or source documentation, harmonization, LDSC, Genomic SEM, MTAG, MAGMA, Mendelian randomization, cross-ancestry validation, and network-informed machine-learning gene prioritization. Earlier exploratory scripts that attempted alternative data-access routes were not used for the final analyses and have been removed or clearly labeled as deprecated. The final results reported in this manuscript were generated from the study-specific workflow described in the repository and the Methods section of this manuscript. AI-assisted tools, if used, were limited to code formatting, annotation, or drafting support; all analysis scripts were reviewed, executed, and verified by the authors, and the reported results derive from the reproducible workflow described above. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Jiawen Wang, Yupeng Chen, Lingfeng Meng. Contributor Information Qingguo Zhu, Email: zuqinguo@163.com. Jinfeng Wu, Email: wjf54007@fjmu.edu.cn. Liefu Ye, Email: yeliefu@fjmu.edu.cn. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41514-026-00430-5. References Associated Data Supplementary Materials Supplementary Information Data Availability Statement The data used in this study are publicly available GWAS summary statistics from FinnGen R12, the GWAS Catalog, and Pan-UKBB. FinnGen R12 summary statistics are available at https://r12.finngen.fi/. GWAS Catalog summary statistics are available at https://www.ebi.ac.uk/gwas/summary-statistics/. Pan-UKBB summary statistics are available at https://pan.ukbb.broadinstitute.org/. The exact phenotype-specific source, accession identifier or phenocode, file name, genome build, sample size, and download link for each GWAS dataset are provided in Supplementary Table 1. Analysis code is available at https://github.com/LI-993/pelvic_floor_gwas. The repository contains the study-specific scripts required to reproduce the final analyses, including GWAS summary-statistics download or source documentation, harmonization, LDSC, Genomic SEM, MTAG, MAGMA, Mendelian randomization, cross-ancestry validation, and network-informed machine-learning gene prioritization. Earlier exploratory scripts that attempted alternative data-access routes were not used for the final analyses and have been removed or clearly labeled as deprecated. The final results reported in this manuscript were generated from the study-specific workflow described in the repository and the Methods section of this manuscript. AI-assisted tools, if used, were limited to code formatting, annotation, or drafting support; all analysis scripts were reviewed, executed, and verified by the authors, and the reported results derive from the reproducible workflow described above.