{"paper_id":"e78f5c67-98b1-41f9-9a44-84fb467208e7","body_text":"1 \n \nDecoding the regulatory genetic architecture of 1 \nendometriosis using AlphaGenome 2 \n 3 \nSanu Bifal Maji 1, Apostol Apostolov 1,2,3,4, Alberto Sola-Leyva 1,2,3, Lucia Blanco-4 \nRodriguez1,5, Amruta D. S. Pathare1,2,3*, Andres Salumets1,2,3,5* 5 \n1Celvia CC AS, Tartu, Estonia. 6 \n2Division of Obstetrics and Gynaecology, Department of Clinical Science, Intervention and Technology, 7 \nKarolinska Institutet, Stockholm, Sweden.  8 \n3Department of Gynaecology and Reproductive Medicine, Karolinska University Hospital, Huddinge, 9 \nStockholm, Sweden. 10 \n4Department of Biotechnology, Institute of Molecular and Cell Biology, University of Tartu, Tartu, Estonia. 11 \n5Department of Obstetrics and Gynaecology, Institute of Clinical Medicine, University of Tartu, Tartu, Estonia. 12 \n*Co-last authors 13 \nCorresponding author – Andres Salumets, andres.salumets@ki.se 14 \n 15 \n 16 \n 17 \n 18 \n 19 \n 20 \n 21 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \nNOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.\n\n2 \n \nAbstract 22 \nBackground 23 \nEndometriosis is a complex, estrogen-dependent disease with a strong genetic component. 24 \nAlthough genome-wide association studies (GWAS) have identified multiple susceptibility 25 \nloci, most associated variants reside in noncoding regions, limiting biological interpretation 26 \nand causal gene identification. Moreover, GWAS gene prioritization is limited by incomplete 27 \ntissue-specific annotation coverage (e.g., GTEx, ENCODE, fine-mapping, Mendelian 28 \nrandomization, and network-based methods). We therefore applied the AlphaGenome 29 \nartificial intelligence framework to prioritize endometriosis-associated variants based on 30 \npredicted uterus-specific regulatory effects. 31 \nMethods 32 \nWe analysed the top 10,000 endometriosis-associated single-nucleotide polymorphisms 33 \n(SNPs) identified by previously published GWAS by Rahmioglu et al , using AlphaGenome 34 \nacross multiple genomic output types. Uterus-specific predictions with high-confidence 35 \neffects (|quantile score| ≥  0.90) were grouped into major regulatory modalities. 36 \nAlphaGenome-prioritized SNPs within ±500 kb of known GWAS loci were classified into 37 \ntiers based on the number of supported regulatory modalities, with broader support indicating 38 \nstronger multilayer regulatory evidence. Effect allele frequency, linkage disequilibrium (LD), 39 \nand overlap with previously published endometriosis-associated variants were also assessed. 40 \nResults 41 \nAlphaGenome generated uterus-specific, 147,033 high-confidence signals across 10,000 42 \nendometriosis-associated variants, spanning six regulatory modalities including gene 43 \nexpression, promoter activity, chromatin accessibility, transcription factor binding, histone 44 \nmodification, and RNA splicing. Within the 42 established endometriosis GWAS loci, 45 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n3 \n \nAlphaGenome identified 42 alternative sub-threshold SNPs with stronger predicted uterus-46 \nspecific regulatory effects than the published GWAS lead variants. Nineteen AlphaGenome-47 \nprioritized SNPs were classified as tier 1, showing support across all six regulatory 48 \nmodalities, compared with five GWAS lead SNPs. Linkage disequilibrium analysis identified 49 \neight tier 1 SNPs with weak-to-low LD (r² < 0.5) relative to the corresponding GWAS lead 50 \nvariants, regulating majority of genes involved in estrogen-driven proliferation and 51 \ninflammatory signalling, highlighting their potential relevance to endometriosis pathogenesis. 52 \nAdditionally, we identified 167 genome-wide significant SNPs outside 42 published GWAS 53 \nlead SNP loci including six tier 1 SNPs (rs1482061, rs7772579, rs6557140, rs2982571, 54 \nrs12631337 and rs79626929), encompassing genes nearby ESR1/ 6q25.1, substantiating 55 \nbiological relevance for endometriosis pathogenesis. 56 \nConclusions 57 \nAlphaGenome-based regulatory prioritization refined endometriosis-associated genome-wide 58 \nassociation study loci by identifying variants with stronger predicted uterus-specific 59 \nfunctional relevance. These findings provide a regulatory framework for prioritizing 60 \ncandidate variants and genes for downstream functional validation in endometriosis. 61 \nKeywords: AlphaGenome; endometriosis; GWAS; regulatory variant prioritization; uterus-62 \nspecific regulation; non-coding variants; multimodal genomics; RNA splicing. 63 \n 64 \n 65 \n 66 \n 67 \n 68 \n 69 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n4 \n \nBackground 70 \nEndometriosis is a chronic inflammatory disorder affecting approximately 5-10% of women 71 \nof reproductive age and is characterized by the presence of endometrial-like tissue outside the 72 \nuterus, most commonly on pelvic organs [1]. The disease is associated with debilitating 73 \npelvic pain, infertility, and reduced quality of life, and diagnosis is often substantially delayed 74 \nbecause definitive confirmation typically requires surgical visualization of lesions [2,3]. 75 \nCurrent treatment options remain limited, relying mainly on hormonal suppression or 76 \ninvasive approaches to surgically remove the lesions, both of which have limitations and 77 \nimportant impact on women health. Biologically and clinically, endometriosis is considered a 78 \nheterogeneous condition characterized by variability in lesion types, disease stage, infertility, 79 \nand pain manifestations, suggesting that multiple pathogenic mechanisms contribute to 80 \ndisease susceptibility and presentation.  81 \nGenetic factors make a major contribution to endometriosis risk, with heritability estimated at 82 \napproximately 50%, with substantial proportion attributed to common genetic variation [4,5]. 83 \nRecent large-scale genome-wide association studies (GWAS) have expanded understanding 84 \nof the genetic architecture of endometriosis [4,6–13]. In particular, a meta-analysis including 85 \n60,674 cases and 701,926 controls identified 42 genome-wide significant endometriosis-86 \nrelated loci [3]. Fine-mapping of these loci further resolved six high-confidence candidate 87 \ncausal variants, including variants at or near SYNE1, HOXA10, HOXC10, LINC00629, ESR1, 88 \nand LNC-LBCS genes, all located in non-coding regions [3]. More recently, a large multi-89 \nancestry GWAS and integrated multi-omics analysis identified 80 genomic regions associated 90 \nwith endometriosis risk, including 37 newly reported loci [14]. Multi-omics integrative 91 \nanalyses in several tissues have further linked endometriosis genetic risk to pathways 92 \ninvolved in cell differentiation, immune and hormonal regulation, tissue remodeling, and 93 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n5 \n \ninflammation [3,14]. Consistent with these findings, many implicated loci map near genes 94 \ninvolved in hormone signaling, uterine development, immune regulation, adhesion, and 95 \nangiogenesis [1,7,9,10,12,13,15–25], which are vital biological processes related to the 96 \npathogenesis of endometriosis. However, despite these advances, the regulatory 97 \nconsequences of most endometriosis-associated variants remain incompletely understood, 98 \nparticularly because the majority reside in non-coding regions, making gene prioritization and 99 \nmechanistic interpretation from GWAS signals alone inherently challenging. As a result, 100 \ncurrent GWAS-based approaches largely resolve association signals at the locus level rather 101 \nthan identifying causal genes or regulatory mechanisms, particularly in hormonally regulated 102 \ntissues such as the uterus, where long-range and context-dependent gene regulation is 103 \nprominent. To address this limitation, there is a need for approaches that can systematically 104 \nlink non-coding genetic variation to downstream regulatory effects across disease-relevant 105 \ntissues and molecular layers. 106 \nIn this context, recent advances in deep learning models such as AlphaGenome represent a 107 \nmajor step forward in sequence-to-function modeling by enabling simultaneous prediction of 108 \nthousands of functional genomic output tracks at base-pair resolution across multiple 109 \nregulatory modalities, including gene expression, chromatin accessibility, transcription factor 110 \n(TF) binding, histone modifications, promoter activity, three-dimensional chromatin 111 \norganization and RNA splicing [26]. Unlike earlier models like SpliceAI [27], BPNet [28], 112 \nProCapNet [29], among others, which were designed for more specific regulatory tasks or 113 \nwere constrained by trade-offs between input sequence length and output resolution, 114 \nAlphaGenome integrates long-range genomic context of up to 1 Mb with high-resolution 115 \noutput, enabling the modelling of both local and distal regulatory effects within a unified 116 \nframework [26]. The model has demonstrated state-of-the-art or near state-of-the-art 117 \nperformance across multiple benchmarking tasks and has shown strong ability to predict the 118 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n6 \n \nmolecular consequences of non-coding variants across diverse regulatory layers [26]. 119 \nNotably, AlphaGenome has also been shown to recapitulate known disease-associated 120 \nregulatory mechanisms, including non-coding variant effects near the TAL1  oncogene [30], 121 \nwhere it successfully captured coordinated changes in TF binding, chromatin accessibility, 122 \nand gene expression associated with oncogenic activation. Such performance highlights its 123 \nvalue as a framework for linking non-coding genetic variation to downstream molecular 124 \nconsequences across multiple regulatory modalities [26]. Given that the majority of disease-125 \nassociated GWAS variants as in endometriosis, lie in non-coding regions [3,5], tools such as 126 \nAlphaGenome may provide an important opportunity to bridge the gap between genetic 127 \nassociation and biological mechanism. This is particularly important for loci where fine-128 \nmapping, expression and methylation quantitative trait locus, or sub-phenotype analyses 129 \nsuggest functional relevance, yet the precise molecular direction and breadth of regulatory 130 \nperturbation remain unresolved. 131 \nIn this study, we integrated endometriosis GWAS risk loci [3,14] with uterus-specific 132 \nAlphaGenome predictions to investigate the regulatory architecture of disease-associated 133 \nvariants across multiple molecular layers including transcription, promoter, chromatin 134 \naccessibility, TF binding, histone modification, three-dimensional chromatin organization 135 \nand RNA splicing. The variants based on this multimodal approach were scored to prioritize 136 \nloci which could predict regulatory interpretation of endometriosis. By combining large-scale 137 \ngenetic association data with deep learning-based regulatory prediction, this work seeks to 138 \nrefine the functional interpretation of endometriosis risk loci and provide new insight into the 139 \ntissue-relevant regulatory mechanisms underlying disease susceptibility and symptom 140 \nheterogeneity. 141 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n7 \n \nMethods 142 \nData source and variant processing 143 \nA total of 10,000 SNPs (p < 3.36 × 10/i2 5) were selected from the endometriosis GWAS meta-144 \nanalysis reported in study by Rahmioglu et al , [3] (Additional file 2: Table S1). The initial 145 \nvariant information extracted from the GWAS source included chromosome, genomic 146 \nposition in hg19, and the corresponding effect and non-effect alleles. To ensure compatibility 147 \nwith AlphaGenome, which requires variant coordinates in the GRCh38/hg38 reference 148 \ngenome, all SNPs were converted from hg19 to hg38 using the Ensembl REST API. Variants 149 \nwere mapped individually, and only successfully converted variants were retained for 150 \ndownstream analysis. RsIDs were then annotated using Ensembl variation records based on 151 \nhg38 genomic coordinates (Additional file 2: Table S2). For AlphaGenome prediction, 152 \nvariants were encoded such that the alternate allele corresponded to the GWAS effect allele, 153 \nand the reference allele corresponded to the GW AS non-effect allele, after checking 154 \ncompatibility with the hg38 genomic position. The processed variants were formatted into a 155 \ntab-separated input file compatible with AlphaGenome, including variant ID, chromosome in 156 \nchr format, hg38 genomic position, reference/non-effect allele, and alternate/effect allele 157 \n(Fig. 1, Additional file 1 and Additional file 2: Table S2). Results were further cross-158 \nreferenced with a recent multi-ancestry endometriosis GWAS dataset [14] to investigate the 159 \nregulatory modalities associated with these genetic signals. 160 \n 161 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n8 \n 162 \nFigure 1. AlphaGenome workflow for prioritizing endometriosis-associated SNPs . A 163 \ntotal of 10,000 endometriosis-associated GWAS SNPs ( p < 3.36×10 /i1/i1 ) were converted 164 \nfrom hg19 to hg38 and analyzed using AlphaGenome with a 1 Mb sequence context across 11 165 \nregulatory genomic output types. Predictions were restricted to uterus-specific biosamples, 166 \nand high-confidence signals (|quantile score| ≥  0.90). Genomic output types were grouped 167 \ninto six regulatory modalities like gene expression, promoter activity, chromatin accessibility, 168 \ntranscription factor binding, histone modification, and RNA splicing. Each SNP was assigned 169 \na tier based on regulatory modalities (Tier 1 = all six modalities; Tier 5 = ≤ 2 modalities). 170 \nTiered SNPs were evaluated across three groups: published GWAS lead SNPs at 42 genome-171 \nwide significant loci, AlphaGenome-prioritized alternative SNPs within ±500 kb of these 172 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n9 \n \nloci, and outside-locus SNPs. Results were further cross-referenced with a recent multi-173 \nancestry endometriosis GWAS [14] to validate novel regulatory signals. 174 \n 175 \nAlphaGenome-based variant effect prediction 176 \nVariant effect prediction was performed using the AlphaGenome Python SDK version 0.6.1 177 \nwithin a Conda-based computational environment. The model was configured for Homo 178 \nsapiens using a 1 Mb sequence context window, allowing both proximal and distal regulatory 179 \nelements surrounding each variant to be incorporated into the prediction [26]. For each SNP, 180 \na genomic interval centered on the variant position was extracted and used as input to the 181 \nmodel. Variant effects were computed by comparing predicted regulatory signals between the 182 \nreference and alternate alleles using the variant-scoring framework implemented in 183 \nAlphaGenome. The AlphaGenome generates predictions across 11 raw genomic output types 184 \nbased on experimental techniques such as RNA sequencing (RNA-seq), Precision Run-On 185 \ncoupled with cap analysis (PRO-cap), Cap Analysis of Gene Expression (CAGE), Assay for 186 \nTransposase-Accessible Chromatin sequencing (ATAC-seq), DNase sequencing (DNase-seq), 187 \nTranscription Factor Chromatin Immunoprecipitation Sequencing (TF ChIP-seq), Histone 188 \nChromatin Immunoprecipitation sequencing (Histone ChIP-seq), splice sites, splice junctions, 189 \nsplice-site usage, and contact maps. For each variant-track combination, AlphaGenome 190 \nproduced annotations describing the regulatory context and predicted effect size. These 191 \nincluded variant-level information, such as variant_id and the scored genomic interval; gene-192 \nlevel annotations including gene_id, gene_name, gene_type, and gene_strand; assay-specific 193 \nattributes including output type, assay title, track name, and track strand; and biosample 194 \nmetadata including biosample name, biosample type, life stage, and ontology terms. 195 \nAdditional regulatory descriptors, such as TF identity, histone mark type, and tissue 196 \nannotations, including GTEx tissue labels, were retained where available. Quantitative 197 \noutputs included raw prediction scores and normalized quantile scores, representing the 198 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n10 \n \nmagnitude and direction of variant-associated regulatory effects. All predictions were retained 199 \nfor downstream filtering and analysis. 200 \nBiosample-level exploratory analysis 201 \nTo characterize the distribution of AlphaGenome predictions across biological contexts, 202 \nprediction outputs were summarized at the biosample level using metadata from each 203 \ngenomic output types (Additional file 2: Table S3). For each biosample, aggregate metrics 204 \nwere calculated to describe prediction abundance, strength, and regulatory breadth, including 205 \nthe total number of variant-track signals, the number of high-confidence signals defined by 206 \nabsolute quantile score ≥  0.90, positive and negative high-confidence signal counts, the 207 \nnumber of available genomic output types, distinct output types, represented regulatory 208 \nmodalities, and quantile-score summary statistics such as maximum and mean absolute 209 \nquantile score (Additional file 2: Table S4). Because biosamples differed in the number of 210 \navailable AlphaGenome tracks, signal counts were also normalized by track availability. 211 \nNormalized total and high-confidence signal burdens were calculated by dividing the 212 \ncorresponding signal counts by the number of available genomic output types for each 213 \nbiosample, providing track-adjusted estimates of prediction burden. Biosamples were first 214 \nsummarized across all biosample types and then filtered to retain ‘tissue biosamples’ for the 215 \nmain tissue-level exploratory analysis, reducing potential over-representation of cell lines or 216 \nisolated cell types (Additional file 2: Table S4). Tissues were ranked by total and normalized 217 \nsignal burden, while regulatory breadth was assessed using the number of distinct output 218 \ntypes and regulatory modalities. These summaries guided tissue-context selection for 219 \ndownstream uterus-specific analyses. 220 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n11 \n \nUterus-specific filtering and high-confidence prediction selection 221 \nGiven the uterine origin of endometriosis, AlphaGenome predictions were filtered using 222 \nbiosample annotations. Predictions corresponding to uterus-related biosamples were retained 223 \nbased on metadata fields such as biosample_name, tissue annotation, and associated ontology 224 \ninformation. AlphaGenome quantile scores were used to quantify the relative magnitude of 225 \npredicted variant effects across different assays and biosamples. The quantile score represents 226 \nthe relative strength of a variant-induced predicted signal change compared with a 227 \nbackground distribution of model predictions. Positive quantile scores indicate an increase in 228 \npredicted signal associated with the alternate allele, whereas negative quantile scores indicate 229 \na decrease relative to the reference allele. Given the large scale and heterogeneity of the 230 \ngenerated prediction output, a stringent high-confidence threshold was applied. Only 231 \npredictions with an absolute quantile score ≥  0.90 were retained for the main uterus-specific 232 \nregulatory analyses. This threshold was used to prioritize strong predicted regulatory effects 233 \nwhile reducing noise across the multi-modal AlphaGenome output (Fig. 1, Additional file 1 234 \nand Additional file 2: Table S5).  235 \nAnnotation of uterus-specific AlphaGenome predictions 236 \nThe uterus-specific prediction table was annotated by mapping each AlphaGenome variant_id 237 \nto its corresponding rsID using the hg38 SNP reference file generated during variant 238 \npreprocessing (Additional file 2: Table S2). Locus information was incorporated by matching 239 \nannotated rsIDs to the curated lead-SNP locus table derived from the 42 genome-wide 240 \nsignificant endometriosis loci reported in the published GWAS study [3] (Additional file 2: 241 \nTables S6 and S7). To characterize the spatial relationship between predicted regulatory 242 \nvariants and their associated genes, the distance between each SNP and the corresponding 243 \ngene body was calculated in base pairs using hg38 gene annotation coordinates. For each 244 \nvariant-gene pair, the SNP position was compared with the annotated start and end 245 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n12 \n \ncoordinates of the matched gene. Variants located within the gene body were assigned a 246 \ndistance of 0 bp, whereas intergenic variants were assigned the shortest linear distance to the 247 \nnearest gene boundary. Additional gene-level features, including gene start, gene end, and 248 \ngene strand, were incorporated into the final annotated uterus-specific prediction table 249 \n(Additional file 2: Table S8). 250 \nRegulatory category grouping and tier classification 251 \nFor SNP-level downstream interpretation, uterus-specific high-confidence AlphaGenome 252 \npredictions were collapsed into integrated regulatory modalities rather than analyzed as 253 \nindividual raw genomic output types. Although AlphaGenome generated predictions across 254 \n11 raw genomic output types, only the genomic tracks represented after uterus-specific 255 \nbiosample filtering and high-confidence filtering were used for tier-based classification. 256 \nAccordingly, six uterus-specific regulatory modalities were defined: gene expression, 257 \npromoter activity/transcription initiation, chromatin accessibility, TF binding, histone 258 \nmodification, and RNA splicing. RNA-seq predictions were assigned to the gene expression 259 \ncategory, reflecting predicted transcript abundance. CAGE predictions were assigned to 260 \npromoter activity, reflecting transcription start site-associated regulatory activity. ATAC-seq 261 \nand DNase-seq predictions were grouped under chromatin accessibility, representing 262 \npredicted open chromatin regions. TF ChIP-seq predictions were assigned to TF binding, 263 \nwhereas histone ChIP-seq predictions were assigned to histone modification, representing 264 \nchromatin-state-associated regulatory signals. Splicing-related predictions, such as splice 265 \njunctions and splice-site usage, were grouped under RNA splicing, capturing predicted effects 266 \non splice junctions and splice-site usage (Fig. 1). 267 \nOther AlphaGenome predicted genomic output types, including PRO-cap, splice sites, and 268 \ncontact maps, were retained in the global AlphaGenome output summary but were not present 269 \namong the uterus-specific high-confidence tracks used for downstream SNP interpretation. 270 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n13 \n \nTherefore, these outputs were not included in the uterus-specific tiering framework (Fig. 2). 271 \nTier classification was based only on the six collapsed regulatory modalities described above. 272 \nEach SNP was summarized according to the number of distinct uterus-specific high-273 \nconfidence regulatory modalities in which it showed predicted regulatory effects. The SNPs 274 \nwere then classified into five tiers according to regulatory-modality breadth: Tier 1, support 275 \nacross all six modalities; Tier 2, support across five modalities; Tier 3, support across four 276 \nmodalities; Tier 4, support across three modalities; and Tier 5, support across two or fewer 277 \nmodalities. This six-regulatory modality tiering framework was applied consistently to the 278 \npublished GWAS lead SNPs, AlphaGenome-prioritized SNPs, and outside-GW AS-lead-SNP 279 \nloci (Fig. 1). 280 \nLocus-based grouping of 42 lead GWAS SNPs and selection of AlphaGenome-281 \nprioritized representative SNPs 282 \nTo compare uterus-specific AlphaGenome predictions with the published endometriosis 283 \nGWAS architecture, the 42 lead SNPs were assigned to the 42 genome-wide significant loci 284 \nidentified in the GWAS meta-analysis [3] (Additional file 2: Table S7). In this GWAS 285 \nframework, loci were defined around lead SNPs representing the most significantly 286 \nassociated variant within a regional association signal, using a ±500 kb window ( p < 5 × 10 -287 \n8). Following the same locus-based approach, the 42 published GW AS lead SNPs were used 288 \nto construct ±500 kb locus windows in hg38 coordinates. All uterus-filtered SNPs with high-289 \nconfidence AlphaGenome predictions were assigned to these loci based on genomic position. 290 \nWhere a SNP fell into overlapping locus windows, it was assigned to the nearest locus based 291 \non distance to the center of the lead-SNP region. For each locus, retained SNPs were 292 \nsummarized according to their AlphaGenome-predicted regulatory modalities. This included 293 \nthe number of unique genomic track-level predictions, the number of supported regulatory 294 \nmodalities, and the maximum absolute quantile score. SNP-level GWAS association p values 295 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n14 \n \nwere mapped back to retained SNPs using chromosome and hg19 genomic position from the 296 \noriginal GW AS summary dataset (Additional file 2: Table S1). Within each locus, SNPs were 297 \nranked according to GWAS significance and AlphaGenome regulatory support (Additional 298 \nfile 2: Table S9). 299 \nTo identify AlphaGenome-prioritized representative SNPs within the same GWAS loci, one 300 \nalternative SNP was selected per locus using the following criteria: the SNP had to be 301 \ngenome-wide significant at p < 5 × 10/i2/i2 , distinct from the SNP with the smallest p value in 302 \nthat locus, and supported by high-confidence uterus-specific AlphaGenome predictions. 303 \nPrioritization was based first on the number of genomic track-level predictions, followed by 304 \nthe number of supported regulatory modalities, maximum absolute quantile score, and GWAS 305 \np value (Additional file 2: Table S10). These SNPs are further referred as “AlphaGenome-306 \nprioritized SNPs” (Fig. 1 and Additional file 1). The final locus-level summary table 307 \nincluded, for each locus, the published GWAS lead SNP and the AlphaGenome-prioritized 308 \nrepresentative SNP, together with genomic position, effect and non-effect alleles, effect allele 309 \nfrequency, and GWAS association p value (Additional file 2: Table S11). Separate prediction 310 \nsummary tables were also generated containing the full uterus-specific AlphaGenome 311 \nregulatory profiles for the published GWAS lead SNPs and AlphaGenome-prioritized SNPs 312 \nrepresenting the 42 loci (Additional file 2: Tables S12 and S13). 313 \nIdentification and characterization of outside-locus SNPs from the 10,000 314 \nendometriosis-associated GWAS SNPs 315 \nSNPs from the 10,000 GWAS-ranked variant set that did not fall within any of the 42 ±500 316 \nkb published GWAS lead SNP locus windows (n = 3,454) were classified as outside-locus or 317 \nunassigned SNPs (Fig. 1, Additional file 1 and Additional file 2: Table S14). These SNPs 318 \nwere retained for a separate exploratory regulatory-track analysis. Genome-wide significant 319 \noutside-locus SNPs were identified using the same threshold applied in the locus-based 320 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n15 \n \nanalysis, p < 5 × 10 /i2/i2 . Outside-locus SNPs were then filtered for the presence of at least 321 \none high-confidence uterus-specific AlphaGenome prediction (absolute quantile score ≥  322 \n0.90). Each retained outside-locus SNP was summarized according to the number of 323 \nsupported regulatory modalities, number of unique genomic track-level predictions, 324 \nmaximum absolute quantile score, and GWAS p value. The same six-category framework and 325 \ntier-classification system described above were applied to the outside-locus SNPs. SNP-level 326 \ninformation, including rsID, genomic position, GW AS p value, regulatory categories, 327 \nmaximum absolute quantile score, and tier assignment, was recorded in the outside-locus 328 \nsummary (Additional file 2: Table S15). 329 \nLinkage disequilibrium analysis between published endometriosis-associated GWAS 330 \nlead SNPs and AlphaGenome-prioritized SNPs 331 \nTo evaluate whether AlphaGenome-prioritized SNPs represented the same association signals 332 \nas the corresponding published GWAS lead SNPs, pairwise linkage disequilibrium (LD) was 333 \ncalculated for each published GWAS lead SNP-AlphaGenome-prioritized SNP pair (Fig. 1 334 \nand Additional file 1). LD was estimated using the 1000 Genomes Phase 3 European 335 \nreference panel to remain consistent with the ancestry-specific LD framework used in the 336 \noriginal GWAS study. For each of the 42 loci, the corresponding published GWAS lead SNP 337 \nand AlphaGenome-prioritized SNP were extracted from the European reference panel and 338 \nanalyzed as an SNP pair. Both r 2 and D′  were calculated for each pair. The r 2 value was used 339 \nas the primary classification metric because it reflects how well one SNP predicts another, 340 \nwhereas D ′  was retained as an additional LD descriptor but was not used for category 341 \nassignment. SNP pairs were classified into four LD categories based on r²: strong LD, 342 \ndefined as r² ≥  0.8; moderate LD, defined as 0.5 ≤  r2 < 0.8; weak LD, defined as 0.2 ≤  r2 < 343 \n0.5; and low LD, defined as r 2 < 0.2. For summary-level interpretation, strong and moderate 344 \nLD pairs were considered genetically linked to the published GWAS association signal, 345 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n16 \n \nwhereas weak and low LD pairs were considered potentially distinct or partially independent 346 \nregulatory candidates within the same GWAS-defined locus. The resulting LD table included 347 \nthe published GWAS lead SNP, the AlphaGenome-prioritized SNP, genomic positions, r2, D′ , 348 \nand LD category for each locus (Additional file 2: Table S16). 349 \nCross-reference with multi-ancestry endometriosis-associated GWAS SNPs 350 \nTo determine whether the outside-locus SNP set contained variants subsequently implicated 351 \nin endometriosis, the outside-locus/unassigned SNP list was compared with novel lead SNPs 352 \nreported in a more recent multi-ancestry GWAS study [14] (Fig. 1 and Additional file 1). 353 \nReported novel SNPs were matched against the original 10,000-SNP AlphaGenome input 354 \ndataset using rsID and, where needed, chromosome-position information. For each matched 355 \nSNP, locus assignment status was checked to determine whether it fell within one of the 356 \noriginal 42±500 kb published GW AS lead SNP windows or belonged to the outside-357 \nlocus/unassigned SNP group. Matched SNPs were then annotated with their AlphaGenome 358 \nregulatory-category support, maximum absolute quantile score, number of unique track-level 359 \npredictions, GWAS p value from the 10,000-SNP dataset, nearest gene reported in the new 360 \nGWAS, and tier classification. The final comparison table included overlap status with the 361 \noriginal 10,000-SNP dataset, outside-locus status, regulatory categories, and predicted uterus-362 \nspecific AlphaGenome effects (Additional file 2: Table S17). 363 \n 364 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n17 \n \nResults 365 \nGlobal AlphaGenome prediction landscape across 10,000 endometriosis-associated 366 \nGWAS SNPs 367 \nAlphaGenome analysis of the 10,000 endometriosis-associated GWAS SNPs generated a total 368 \nof 325,451,290 variant-track prediction signals across 11 raw regulatory output types, in all 369 \navailable biosamples, without restricting the initial analysis to uterus, which were further 370 \ngrouped into seven broader functional regulatory modalities for interpretation (Fig. 2 and 371 \nAdditional file 2: Table S18), including gene expression, promoter activity/transcription 372 \ninitiation, chromatin accessibility, TF binding, histone modification, RNA splicing, and three-373 \ndimensional chromatin organization. The landscape of global predicted genomic outputs was 374 \nstrongly dominated by RNA-seq, which accounted for 224,292,024 signals, corresponding to 375 \n68.9% of the total AlphaGenome output (Fig. 2). The next most represented output types 376 \nwere TF ChIP-seq (9.9%), splice junctions (7.2%), and histone ChIP-seq (6.9%). In contrast, 377 \nCAGE (3.4%), DNase-seq (1.9%), A TAC-seq (1.0%), splice-site usage (0.7%), contact maps 378 \n(0.1%), PRO-cap (0.1%), and splice sites contributed smaller proportions of the total 379 \nprediction space (<0.1%) (Fig. 2 and Additional file 2: Table S18). When raw output types 380 \nwere collapsed into broader regulatory modalities, gene expression represented the largest 381 \ncomponent of the prediction landscape (68.9%), followed by TF binding (9.9%), RNA 382 \nsplicing (7.9%), histone modification (6.9%), promoter activity/transcription initiation 383 \n(3.4%), chromatin accessibility (2.9%), and three-dimensional chromatin organization (0.1%) 384 \n(Fig. 2). 385 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n18 \n \n 386 \nFigure 2. Global AlphaGenome prediction landscape across the 10,000-SNP dataset.  387 \nNested donut chart showing the total number of AlphaGenome variant-track prediction 388 \nsignals across 11 raw genomic output types in outer circle collapsed into seven regulatory 389 \nmodalities in the inner circle.  390 \n 391 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n19 \nUterus-specific AlphaGenome predictions and prioritization reveals regulatory 392 \nheterogeneity across endometriosis GWAS loci 393 \nBiosample-level summarization of 10,000 endometriosis-associated GWAS SNPs revealed 394 \nsubstantial variation in AlphaGenome prediction signal representation across several tissues 395 \n(Additional file 2: Tables S4). The highest total number of variant-track prediction signals 396 \nwere observed in stomach, spleen, adrenal gland, ovary, testis, liver, sigmoid colon, kidney, 397 \nspinal cord, and urinary bladder, indicating that raw signal abundance differed markedly 398 \nacross tissues. This variation likely reflected differences in the number and diversity of 399 \navailable AlphaGenome output types for each tissue. After normalization by the number of 400 \navailable output types, differences in signal burden across tissue remained evident, showing 401 \nthat the observed variation was not explained solely by output type availability. High-402 \nconfidence signals were also unevenly distributed across tissues, indicating biological 403 \nvariability in predicted regulatory activity (Additional file 2: Table S4).  404 \n 405 \n 406 \n 407 \n 408 \n 409 \n 410 \n 411 \n 412 \n 413 \n 414 \n 415 \nFigure 3. Biosample-level AlphaGenome signal burden and regulatory breadth.  416 \nEach point represents a biosample plotted by the total number of AlphaGenome variant-track 417 \nprediction signals for 10,000 endometriosis-associated GWAS SNPs and the number of 418 \ndetected regulatory modalities. Reproductive tissues are highlighted and annotated. Point 419 \ncolor indicates regulatory breadth among the highlighted tissues, ranging from one to six 420 \ndetected modalities. Dashed vertical and horizontal lines define the burden and breadth 421 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n20 \n \nthresholds used to separate biosamples into low burden/narrow breadth, low burden/broad 422 \nbreadth, high burden/narrow breadth, and high burden/broad breadth quadrants.  423 \nAmong reproductive tissues, ovary and uterus were both strongly represented (Fig. 3). Ovary 424 \ncontained 2,675,024 total prediction signals and 259,856 high-confidence signals across 425 \nseven genomic output types, whereas uterus contained 1,432,412 total prediction signals and 426 \n147,033 high-confidence signals across eight genomic output types spanning six regulatory 427 \nmodalities (Fig. 4) such as gene expression (RNA-seq), promoter activity/transcription 428 \ninitiation associated activity (CAGE), chromatin accessibility (combined ATAC-seq and 429 \nDNase-seq), histone modification (histone ChIP-seq), TF binding (TF ChIP-seq), and RNA 430 \nsplicing (combined splice junctions and splice-site usage) (Additional file 2: Table S19). 431 \nRNA-seq was the dominant uterus-specific output type, contributing 1,132,788 predictions, 432 \nfollowed by splice junctions, CAGE, TF ChIP-seq, histone ChIP-seq, ATAC-seq, DNase-seq, 433 \nand splice-site usage. 434 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n21 \n \n 435 \nFigure 4. Uterus Specific AlphaGenome prediction landscape across the 10,000-SNP 436 \ndataset. Nested donut chart showing the total number of AlphaGenome variant-track 437 \nprediction signals across 8 raw genomic output types represented in outer circle collapsed 438 \ninto six regulatory modalities in inner circle. 439 \n 440 \nAlthough ovary showed higher total and high-confidence signal counts, the uterus provided 441 \nthe most disease-relevant tissue context for endometriosis and captured a broader regulatory 442 \nprediction landscape, with eight genomic output types compared with seven in the ovary. 443 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n22 \n \nTherefore, the uterine context was selected for downstream locus-level analyses despite the 444 \noverall signal distribution remaining largely dominated by transcript abundance–related 445 \npredictions.  446 \nMoreover, following tissue-specific prioritization, uterus-filtered high-confidence SNPs were 447 \nmapped to the 42 genome-wide significant endometriosis loci previously identified by GWAS 448 \nmeta-analysis [3]. Using regulatory prioritization strategy, AlphaGenome identified 42 449 \nalternative “AlphaGenome-prioritized SNPs” that differed from the published GWAS lead 450 \nvariants and were used for below comparative locus-level analysis (Fig. 5a). 451 \nComparison of uterus-specific regulatory signal retention between GWAS lead and 452 \nAlphaGenome-prioritized SNPs 453 \nBefore high-confidence filtering (absolute quantile score ≥  0.90), both the published GWAS 454 \nlead SNPs and the AlphaGenome-prioritized SNPs retained predictions across the major 455 \nuterus-relevant regulatory modalities with broadly comparable total signal distributions (Fig. 456 \n5b and Additional file 2: Table S19). Consistent with this pattern, paired locus-level statistical 457 \ntesting showed no significant differences ( p > 0.05) between the two SNP sets across the 458 \ntested modalities after false-discovery-rate (FDR) correction (Additional file 2: Table S20). 459 \nThis indicates that, at the level of raw uterus-filtered prediction coverage, the two matched 460 \n42-SNP sets showed similar regulatory representation. 461 \nAfter applying the high-confidence threshold (absolute quantile score ≥  0.90), a clearer 462 \ndifference emerged between the two SNP sets. The AlphaGenome-prioritized SNPs retained 463 \nmore uterus-specific high-confidence regulatory signals than the published GW AS lead SNPs 464 \nacross most major regulatory modalities (Fig. 5b and Additional file 2: Table S19). Paired 465 \nWilcoxon signed-rank tests across the 42 matched loci showed significantly higher high-466 \nconfidence signal burdens for AlphaGenome-prioritized SNPs in RNA-seq, CAGE, ATAC-467 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n23 \n \nseq, DNase-seq, TF ChIP-seq, and histone ChIP-seq tracks after FDR correction (p < 0.05) 468 \n(Additional file 2: Table S20). In contrast, splice junctions and splice-site usage showed 469 \nnominal differences but did not differ significantly between the two SNP sets. Although 470 \nRNA-seq remained the dominant output type in both SNP sets, AlphaGenome-prioritized 471 \nSNPs retained more high-confidence non-expression regulatory signals than GW AS lead 472 \nSNPs, particularly for CAGE, ATAC-seq, DNase-seq, TF ChIP-seq, and histone ChIP-seq 473 \n(Fig. 5b and Additional file 2: Table S19). This shift was also visible after collapsing tracks 474 \ninto regulatory modalities: compared with GW AS lead SNPs, AlphaGenome-prioritized SNPs 475 \nshowed lower relative gene expression contribution (78.6% to 64.9%) and higher 476 \ncontributions from promoter activity (3.3% to 6.7%), chromatin accessibility (6.0% to 477 \n11.1%), TF binding (2.7% to 5.4%), and histone modification (4.9% to 6.8%) after high-478 \nconfidence filtering (Fig. 5c). These findings suggest that AlphaGenome prioritization 479 \nenriched variants with stronger uterus-relevant high-confidence regulatory effects beyond 480 \ngene expression alone. 481 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n24 \n 482 \nFigure 5. Uterus-specific AlphaGenome prioritization of endometriosis-associated SNPs. 483 \n(a) Manhattan plot showing GWAS association strength for SNPs across chromosomes. Grey 484 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n25 \n \npoints represent other SNPs in the 10,000-SNP dataset, red points indicate the 42 published 485 \nGWAS lead SNPs, and blue points indicate the 42 AlphaGenome-prioritized subthreshold 486 \nSNPs selected within corresponding GWAS loci. Selected loci are labelled, and the dashed 487 \nhorizontal line indicates the genome-wide significance threshold of p = 5 × 10 /i2/i2 . (b) 488 \nAbsolute number of retained uterus-specific AlphaGenome variant–track prediction signals 489 \nacross major genomic output types for the two matched 42-SNP sets. Bars show counts for 490 \npublished GWAS lead SNPs after uterus tissue filtering, published GWAS lead SNPs after 491 \nhigh-confidence filtering, AlphaGenome-prioritized SNPs after uterus tissue filtering, and 492 \nAlphaGenome-prioritized SNPs after high-confidence filtering. High-confidence filtering was 493 \ndefined as an absolute quantile score ≥  0.90. Asterisks indicate output types in which 494 \nAlphaGenome-prioritized SNPs showed significantly higher high-confidence signal burden 495 \nthan published GW AS lead SNPs after FDR correction ( p < 0.05; paired Wilcoxon signed-496 \nrank test). (c) Bubble plot showing the proportional distribution of retained uterus-specific 497 \nprediction signals across regulatory modalities for the same two SNP sets and filtering stages. 498 \nGenomic output types were collapsed into six regulatory modalities. Bubble size and 499 \npercentage labels indicate the contribution of each modality to the total retained prediction 500 \nsignals within each dataset/filtering stage. 501 \n 502 \nAlphaGenome-prioritized SNPs show broader uterus-specific multi-modal regulatory 503 \nsupport than published GWAS lead SNPs 504 \nUterus-specific AlphaGenome predictions showed that all 42 published GWAS lead SNPs 505 \nand alterative 42 AlphaGenome-prioritized SNPs retained at least one high-confidence 506 \nregulatory signal in the uterine context (Fig. 6a-d and Additional file 2: Table S12, S13). Tier-507 \nbased stratification demonstrated substantial heterogeneity among the 42 GWAS lead SNPs, 508 \nwith RNA-seq representing the most consistently supported regulatory category in the uterine 509 \ncontext (Fig. 6a). Overall, only 5 of the 42 published GWAS lead SNPs belongs to tier 1 510 \ncategory, while 4 additional lead SNPs were represented in tier 2 category. Most published 511 \nlead SNPs were assigned to tier 4 or tier 5, indicating that many statistically defined GWAS 512 \nlead SNPs showed limited predicted regulatory breadth in the uterine context (Fig. 6b). On 513 \ncontrary, alternative AlphaGenome-prioritized lead SNPs demonstrated substantially stronger 514 \nuterus-specific regulatory support (Fig. 6c). Notably, 35 out of 42 AlphaGenome-prioritized 515 \nlead SNPs were classified within tier 1-3 categories, reflecting strong regulatory support 516 \nacross transcriptomic, epigenomic, and splicing-related datasets in the uterus (Fig. 6d and 517 \nTable 1). This indicates a broader and more integrated regulatory landscape for the 518 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n26 \n \nAlphaGenome-prioritized SNPs compared with the GW AS lead SNPs (Additional file 2: 519 \nTable S13). On the other hand, the GWAS lead SNPs exhibited limited regulatory overlap 520 \nwith fewer regulatory modalities predictions. Moreover, allele frequency comparisons 521 \nshowed no systematic difference between AlphaGenome-prioritized and published GWAS 522 \nlead SNPs: prioritized SNPs had higher effect allele frequencies at 22 loci (0.426 ± 0.253) 523 \nand lower at 20 loci (0.385 ± 0.223; Additional file 2: Table S11). This balance suggests that 524 \nthe richer regulatory signal observed in AlphaGenome-prioritized SNPs was not simply an 525 \nartifact of allele frequency. Collectively, these findings suggest that AlphaGenome-based 526 \nprioritization can refine GWAS loci by identifying variants with stronger predicted functional 527 \nrelevance in the uterine regulatory context. 528 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n27 \n529 \nFigure 6. Uterus-specific regulatory profiles of GWAS lead and prioritized SNPs.  ( a) 530 \nHeatmap showing high-confidence AlphaGenome predictions, defined as absolute quantile 531 \nscore ≥  0.90, across six integrated regulatory modalities for the 42 published GWAS lead 532 \nSNPs and the (b)  UpSet plot showing the overlap of six integrated uterus-specific 533 \nAlphaGenome regulatory modalities among the 42 published GWAS lead SNPs. (c) Heatmap 534 \nshowing high-confidence AlphaGenome predictions, defined as absolute quantile score ≥  535 \n0.90, across six integrated regulatory modalities for the 42 AlphaGenome-prioritized 536 \nrepresentative SNPs. (d)  UpSet plot showing the combinatorial distribution of regulatory 537 \nsupport among the 42 AlphaGenome-prioritized SNPs selected within ±500 kb of the original 538 \nGWAS lead loci. In the heatmaps (a, c), positive predicted effects of the alternative allele are 539 \nshown in green, negative predicted effects in purple, and the absence of retained high-540 \nconfidence predictions in grey. Colour intensity corresponds to the magnitude of the quantile 541 \nscore. In the UpSet plots (b, d), the six integrated regulatory categories are highlighted in red 542 \n \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n28 \n \ntext, and bar heights represent the number of SNPs supported by each individual or combined 543 \nregulatory modality. 544 \n 545 \n Table 1. Uterus-specific regulatory tier classification of endometriosis-associated loci. 546 \n 547 \nTier Number of \nregulatory \nmodalities \nAlphaGenome predictions for the \n42 lead SNPs representing the \nGWAS loci \nRepresentative loci from \nAlphaGenome-selected SNPs  \nNumber \nof loci \nLoci Number \nof loci \nLoci \nTier 1 6 5 GREB1/2p25.1, \nSYNE1/6q25.1*, \nABO/9q34.2, \nMLLT10/10p12.31, \nHOXC10/12p13.13* \n19 ABO/9q34.2, \nCD109/6q13, \nCEP112/17q24.1, \nDNM3/1q24.3, \nGDAP1/8q21.11, \nGREB1/2p25.1, \nHOXC10/12p13.13*, \nIGF1/12q23.2, \nKDR/4q12, \nMLLT10/10p12.31, \nPDLIM5/4q22.3, \nPTPRO/12p12.3, \nRIN3/14q32.12, \nRNLS/10q23.31, \nSKAP1/17q21.32, \nSYNE1/6q25.1*, \nVEZT/12q22, \nVPS13B/8q22.2, \nWNT4/1p36.12 \nTier 2 5 4 7p15.2/7p15.2, \nHOXA10/7p15.2*, \nRIN3/14q32.12, \nVEZT/12q22 \n11 7p15.2/7p15.2, \nBSN/3p21.31, \nCDKN2B-AS1/9p21.3, \nFRMD7/Xq26.2, \nFSHB/11p14.1, \nHOXA10/7p15.2*, \nID4/6p22.3, \nNGF/1p13.2, \nSLC19A2/1q24.2, \nSRP14-AS1/15q15.1, \nand WT1/11p14.1 \nTier 3 4 4 PDLIM5/4q22.3, \nHEY2/6q22.31, \nCDKN2B-AS1/9p21.3, \nFSHB/11p14.1 \n5 7p12.3/7p12.3, \nBMPR2/2q33.1, \nFAM120B/6q27, \nHEY2/6q22.31, \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n29 \n \nTEX11/Xq13.1 \nTier 4 3 9 EBF1/5q33.3, \nFAM120B/6q27, \nFRMD7/Xq26.2, \nGDAP1/8q21.11, \nKCTD9/8p21.2, \nSKAP1/17q21.32, \nSLC19A2/1q24.2, \nVPS13B/8q22.2, \nWNT4/1p36.12 \n4 DLEU1/13q14.2, \nEBF1/5q33.3, \nETAA1/2p14, \nKCTD9/8p21.2 \nTier 5 2 14 7p12.3/7p12.3, \nACTL9/19p13.2, \nCD109/6q13, \nCEP112/17q24.1, \nDLEU1/13q14.2, \nDNM3/1q24.3, \nETAA1/2p14, \nID4/6p22.3, \nIGF1/12q23.2, \nKDR/4q12, \nLINC00629/Xq26.3*, \nNGF/1p13.2, \nPTPRO/12p12.3, \nSRP14-AS1/15q15.1 \n2 ACTL9/19p13.2, \nLINC00629/Xq26.3* \n1 6 ASTN2/9q33.1, \nBMPR2/2q33.1, \nBSN/3p21.31, \nRNLS/10q23.31, \nTEX11/Xq13.1, \nWT1/11p14.1 \n1 ASTN2/9q33.1 \n 548 \nLegends: Loci were stratified according to the number of independent regulatory modalities 549 \nsupported by uterus-specific AlphaGenome predictions.  Six regulatory modalities were 550 \nconsidered: gene expression (RNA-seq), promoter activity (CAGE), chromatin accessibility 551 \n(ATAC-seq and/or DNase-seq), histone modification (histone ChIP-seq), transcription factor 552 \nbinding (TF ChIP-seq), and RNA splicing (splice junction and/or splice-site usage). For each 553 \nlocus, the presence of at least one high-confidence signal, defined as absolute quantile score ≥  554 \n0.90, within a regulatory modality was counted as one unit of evidence. Each regulatory 555 \nmodality was counted once per locus, regardless of the number of individual tracks, genes, or 556 \ntranscripts detected within that category. Loci were classified into five tiers based on total 557 \nregulatory support: Tier 1, six categories; Tier 2, five categories; Tier 3, four categories; Tier 558 \n4, three categories; and Tier 5, two or fewer categories. Higher-tier loci represent variants 559 \nwith broader multi-layer regulatory support across transcriptional, chromatin-associated, and 560 \npost-transcriptional processes. Asterisks indicate loci that were previously resolved by GWAS 561 \nfine-mapping as high-confidence candidate causal loci and were also identified in the present 562 \nAlphaGenome analysis. 563 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n30 \n \nAlphaGenome-prioritized SNPs are mostly linked to GWAS lead SNPs, while 14 loci 564 \nshow potential independent regulatory candidates for endometriosis 565 \nLD analysis identified 14 loci in which the AlphaGenome-prioritized SNP showed weak-to-566 \nlow LD with the corresponding published GWAS lead SNP, defined as r² < 0.5 (Additional 567 \nfile 2: Table S16). These included nine loci with moderate LD, defined as 0.2 ≤  r² < 0.5: 568 \nCD109/6q13, ETAA1/2p14, GDAP1/8q21.11, KDR/4q12, RIN3/14q32.12, SLC19A2/1q24.2, 569 \nSRP14-AS1/15q15.1, TEX11/Xq13.1, and VEZT/12q22. Five additional loci showed low LD, 570 \ndefined as r² < 0.2: SYNE1/6q25.1, CDKN2B-AS1/9p21.3, LINC00629/Xq26.3, 571 \nGREB1/2p25.1, and WNT4/1p36.12. These 14 weak-to-low LD loci may represent distinct or 572 \npartially independent regulatory candidates within the same GWAS-defined regions. 573 \nImportantly, among these, eight AlphaGenome-prioritized SNPs were classified as tier 1, 574 \nthree as tier 2, and one each as tier 3, tier 4, and tier 5, showing high-confidence support 575 \nacross all six integrated uterus-specific regulatory modalities. 576 \nIn contrast, the remaining 28 of 42 loci showed at least moderate LD, defined as r² ≥  0.5, 577 \nbetween the published GWAS lead SNP and the AlphaGenome-prioritized SNP. Of these, 22 578 \nloci showed strong LD, defined as r² ≥  0.8, whereas six loci showed moderate LD, defined as 579 \n0.5 ≤  r² < 0.8. Among these 28 strong- or moderate-LD loci, 11 AlphaGenome-prioritized 580 \nSNPs were classified as tier 1, eight as tier 2, four as tier 3, three as tier 4, and two as tier 5. 581 \nThese results indicate that, in most loci, AlphaGenome prioritization identified SNPs that 582 \nremained genetically linked to the original GWAS association signal while providing 583 \nstronger predicted uterus-specific regulatory support. At the same time, several low-LD 584 \nprioritized SNPs also showed broad regulatory support, suggesting potential additional 585 \nregulatory candidates within the same GWAS-defined regions. The complete list of published 586 \nGWAS lead SNPs, AlphaGenome-prioritized SNPs, r² values, D ′  values, and LD 587 \ninterpretation and tier-based categories is provided in Additional file 2: table S16. 588 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n31 \n \nOutside-locus analysis identifies 167 additional genome-wide significant SNPs with 589 \nuterus-specific regulatory support 590 \nAmong SNPs located outside the 42 published GWAS lead SNP locus windows, 167 SNPs 591 \nreached genome-wide significance (p  < 5 × 10 /i2/i2 ) and showed high-confidence 592 \nAlphaGenome predictions in at least one uterus-specific regulatory modality (Additional file 593 \n2: Table S14-S15). Of these outside-locus SNPs, six were classified as tier 1, including 594 \nrs1482061, rs7772579, rs6557140, rs2982571, rs79626929, and rs12631337. Five of these 595 \nSNPs mapped to chromosome 6 within the 6.q25.1 cluster (hg38) which spans a regulatory 596 \nblock including genes such as SYNE1, ESR1, CCDC170, AKAP12, ZBTB2, ARMT1, and 597 \nRMND1 (Fig. 7a and Additional file 2: Table S21). Notably, these variants were not captured 598 \nwithin the predefined 42 GWAS lead SNP locus windows yet exhibited tier 1 regulatory 599 \nsupport in the uterine context, suggesting that the AlphaGenome-based prioritization 600 \nidentifies additional high-confidence regulatory variants within known disease-relevant 601 \ngenomic regions that are not recovered by conventional locus-based GWAS lead SNP 602 \ndefinitions.  603 \n The remaining tier 1 SNP out of six outside-locus SNPs (rs12631337) mapped to 604 \nchromosome 3 at chr3:50,161,104 in hg38 coordinates and was located within a gene-rich 605 \nregion containing multiple protein-coding genes, including SEMA3F, GNAI2, HYAL2, RBM5, 606 \nRBM6, TUSC2, UBA7, TMEM115, IP6K1, APEH, SEMA3B, RASSF1, IFRD2, SLC38A3, 607 \nMAPKAPK3, and NPRL2 (Fig. 7b and Additional file 2: Table S21). Other genome-wide 608 \nsignificant outside-locus SNPs were distributed across lower regulatory tiers, with 9 SNPs 609 \nclassified as tier 2, the 14 SNPs as tier 3, the 38 SNPs as tier 4, and 100 SNPs as tier 5. 610 \nWithin tier 5, 51 SNPs showed support across two regulatory modalities, whereas 49 SNPs 611 \nretained only one high-confidence regulatory modality. Thus, although most genome-wide 612 \nsignificant outside-locus SNPs showed narrower regulatory support, a smaller subset 613 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n32 \ndisplayed broad multi-layer uterus-specific regulatory evidence. These findings indicate that, 614 \nbeyond the predefined published GWAS loci, other genome-wide significant variants may 615 \ncarry uterus-specific regulatory information relevant for downstream functional prioritization 616 \n(Additional file 2: Table S14-S21). 617 \n 618 \nFigure 7. Genomic positions of endometriosis-associated SNPs on chromosomes 6 and 3. 619 \nSchematic representation of selected endometriosis-associated genomic loci showing GWAS 620 \nlead SNPs, AlphaGenome-prioritized SNPs, outside-locus SNPs, and nearby Ensembl-621 \nannotated AlphaGenome predicted protein coding genes using GRCh38 coordinates. (a) 622 \nChromosome 6q25.1–6q25.2 locus containing the ESR1/SYNE1 region. (b) Chromosome 623 \n3p21.31–3p21.2 locus containing genes surrounding the selected chromosome 3 SNPs. 624 \nOutside-locus SNPs are shown in blue, the GWAS lead SNPs are shown in orange, and the 625 \nAlphaGenome-prioritized SNPs are shown in purple. Gene boxes indicate Ensembl-annotated 626 \ngenes, and genomic positions are shown in megabases according to GRCh38. 627 \n 628 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n33 \n \nCross-reference with multi-ancestry GWAS SNPs reveals high-confidence uterus-629 \nspecific regulatory support for newly reported endometriosis loci 630 \nTo assess whether uterus-specific AlphaGenome predictions supported recently reported 631 \nendometriosis-associated variants, the AlphaGenome-ranked 10,000-SNP dataset was cross-632 \nreferenced with 37 novel lead SNPs reported in a recent GW AS study [4,5]. Of these, 22 633 \nvariants were present in the original 10,000-SNP dataset, whereas 15 were absent (Additional 634 \nfile 2: Table S17). All 22 overlapping SNPs were located outside the original 42 published 635 \nendometriosis GWAS loci, based on the ±500 kb lead-SNP window definition, and none 636 \ncorresponded to the original GWAS lead SNPs. The overlapping SNPs showed variable 637 \ndegrees of uterus-specific regulatory support. Based on the number of high-confidence 638 \nAlphaGenome regulatory modalities detected, one SNP was classified as tier 1, three as tier 2, 639 \nfour as tier 3, five as tier 4, and nine as tier 5 (Additional file 2: Table S17). Notably, 13 of 640 \nthe 22 overlapping SNPs belonged to tiers 1-4, indicating support across three to six 641 \nregulatory modalities. The strongest signal was observed for rs17 99293, reported near 642 \nARHGAP26, which showed high-confidence predictions across all six integrated categories, 643 \ntier 1 (Fig. 8). 644 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n34 \n 645 \nFigure 8. Cross-reference of multi-ancestry endometriosis GW AS SNPs with uterus-specific 646 \nAlphaGenome-predicted regulated genes.  Multi-ancestry endometriosis GW AS SNPs overlapping 647 \nthe top 10,000 AlphaGenome-ranked variants are categorized into tier 1-4 according to the number of 648 \nsupported high-confidence regulatory modalities. The left column shows the SNP identifier and 649 \nAlphaGenome tier. The middle column shows the nearest gene for the SNPs in the multi-ancestry 650 \nGW AS, and the right column shows genes or transcripts predicted by AlphaGenome to be regulated in 651 \nthe uterus-specific dataset. Highlighted genes indicate common genes identified both as multi-652 \nancestry GW AS-reported genes and as AlphaGenome-predicted regulated targets.  653 \n 654 \nImportantly, several genes reported as nearest genes in the recent GWAS were also detected 655 \nas AlphaGenome-predicted regulated genes in the uterus-specific output, supporting 656 \nconcordance between the external GW AS annotation and the present regulatory predictions 657 \n(Fig. 8). For example, rs2290358 was reported near CALD1, and AlphaGenome also 658 \npredicted regulation of CALD1, together with additional transcripts including AKR1B1, 659 \nTMEM140, BPGM, AKR1B15 , and nearby unannotated Ensembl transcripts. Similarly, 660 \nrs1799293 was reported near ARHGAP26 and showed AlphaGenome-predicted effects 661 \ninvolving ARHGAP26 as well as FGF1. The rs7974900 variant, reported near the SSPN 662 \nregion, was also linked to predicted regulation of SSPN and additional transcripts including 663 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n35 \n \nBHLHE41, ITPR2,  and RASSF8 . Other examples included rs223346 near UBE2D3, 664 \nrs62246055 near FOXP1, rs10824194 near ADK, and rs2834747 near RUNX1, for which the 665 \nGWAS-reported nearby gene was also represented among the AlphaGenome-predicted 666 \nregulated genes or transcripts (Fig. 8 and Additional file 2: Table S22). Together, this cross-667 \nreference indicates that a subset of recently reported multi-ancestry endometriosis GWAS 668 \nSNPs already present in the 10,000-SNP AlphaGenome dataset showed high-confidence 669 \nuterus-specific regulatory evidence. In several cases, AlphaGenome not only recovered the 670 \nGWAS-reported nearby gene but also identified additional potentially regulated transcripts 671 \nand multilayer regulatory modalities, providing functional support for these newly reported 672 \nendometriosis-associated loci. 673 \n 674 \nDiscussion 675 \nEncyclopedia of DNA Elements (ENCODE) project data and subsequent Mendelian 676 \nrandomization studies have consistently shown that ~90% of GWAS signals across complex 677 \ndiseases map to non-coding functional elements [31,32]. Endometriosis is not an exception. 678 \nThe majority of GWAS signals identified for endometriosis map to non-coding or intergenic 679 \nregions of the genome [33,34] suggesting that these variants confer disease risk through 680 \nregulatory rather than protein-altering mechanisms [35]. However, functional interpretation 681 \nof these loci remains challenging because the lead SNP at a GW AS locus selected based on 682 \nstatistical association strength rather than functional evidence, may not be necessarily the 683 \ncausal or most functionally relevant variant. In regions of LD, the lead associated SNP may 684 \nact as a proxy for nearby variants with stronger functional effects on gene regulation. 685 \nMoreover, assigning GWAS loci to causal genes remains challenging because the 686 \nfunctionally affected gene is not always the nearest gene to the associated variant. This is 687 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n36 \n \nparticularly relevant for non-coding risk variants, whose functional impact and target genes 688 \nare often difficult to define [36], because such variants may influence distal regulatory 689 \nelements, chromatin interactions, or non-coding RNA-mediated regulatory networks. 690 \nEndometriosis, being a heterogenous disorder involving multiple molecular mechanisms, is 691 \nalso characterized by dysregulated non-coding RNA networks, including miRNAs, lncRNAs, 692 \nand circRNAs in endometriotic lesions, eutopic endometrium, and peripheral immune 693 \ncompartments [37–39]. These regulatory networks interact with epigenetic modifications, 694 \nchromatin remodeling, and TF activity to coordinate tissue-specific gene expression. 695 \nImportantly, because many non-coding RNA loci and regulatory elements reside within non-696 \ncoding genomic regions, GWAS SNPs mapping to these regions may contribute to disease 697 \nsusceptibility through disruption of regulatory and post-transcriptional mechanisms that are 698 \nnot captured by protein-coding-centric analyses. Consequently, integrative computational 699 \nframeworks capable of modeling long-range regulatory effects are increasingly required for 700 \nfunctional interpretation of disease-associated non-coding variants.  701 \nWithin this context, application of the AlphaGenome framework [26] enabled systematic 702 \nfunctional annotation of 10,000 endometriosis-associated GWAS SNPs in a uterus-specific 703 \nregulatory landscape. By leveraging up to 1 Mb of surrounding genomic sequence context for 704 \neach variant, AlphaGenome facilitated evaluation of variant effects across multiple regulatory 705 \nmodalities, including gene expression, promoter activity, chromatin accessibility, TF binding, 706 \nhistone modification, and RNA splicing. We used a tier-based classification framework for 707 \nintegration of these predictions which further enabled refinement of uterus-specific regulatory 708 \nprioritization across GWAS loci. Collectively, these findings support the utility of long-range 709 \nsequence-based modeling approaches for improving identification and prioritization of 710 \npotentially functional non-coding variants underlying endometriosis susceptibility. 711 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n37 \n \nPrevious GWAS follow-up studies have largely relied on experimentally derived functional 712 \nannotation resources, including expression quantitative trait locus datasets such as GTEx, 713 \nregulatory annotations from ENCODE and related epigenomic consortia, fine-mapping, 714 \nMendelian randomization, and computational causal-gene prioritization methods [40]. More 715 \nrecent network-based approaches, such as SigNet, integrate within-locus evidence with cross-716 \nlocus gene and regulatory network information to prioritize likely causal genes at GWAS loci 717 \n[36]. Although these approaches are valuable for moving from association signals toward 718 \ncandidate genes, they remain constrained by the availability, resolution, and tissue relevance 719 \nof the underlying annotation datasets, particularly for less studied diseases like endometriosis. 720 \nIn particular, many resources have incomplete coverage of disease-relevant uterine context 721 \nand often prioritize genes rather than directly modelling variant-level regulatory effects 722 \nacross multiple molecular layers. In contrast, our analysis enabled integrated evaluation of 723 \nendometriosis-associated variants across multiple predicted regulatory layers within a uterus-724 \nspecific context. This provides a complementary framework for interpreting non-coding 725 \nGWAS signals beyond statistical association alone. 726 \nThe comparison between published GWAS lead SNPs [3] and AlphaGenome-prioritized 727 \nSNPs showed that all published GWAS lead SNPs retained at least one predicted regulatory 728 \nsignal in the uterus-specific analysis, indicating that the published GW AS loci remain 729 \nfunctionally relevant. This was consistent with partial concordance with independent fine-730 \nmapping evidence reported by Rahmioglu et al . [3], where four of six high-confidence non-731 \ncoding variants also showed multi-layer regulatory support in the present analysis. However, 732 \nAlphaGenome-based prioritization identified alternative SNPs within the same GWAS-733 \ndefined regions that exhibited broader multimodal regulatory support, suggesting improved 734 \nfunctional resolution beyond statistical association alone. Notably, only five published GWAS 735 \nlead SNPs were classified as tier 1, whereas 19 AlphaGenome-prioritized SNPs were 736 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n38 \n \nassigned to tier 1, reflecting a substantially higher proportion of variants with convergent 737 \nsupport across all six regulatory categories following re-prioritization. These results are 738 \nconsistent with previous work showing that LD can make it difficult to distinguish lead 739 \nassociated variants from nearby causal or functionally relevant variants within the same 740 \nGWAS locus [40]. 741 \nA subset of loci showed low LD between published GWAS lead SNPs (42 loci) and 742 \nAlphaGenome-prioritized SNPs, indicating that the most statistically significant GWAS SNPs 743 \ndo not always correspond to the strongest predicted regulatory effect. Notably, among 14 loci 744 \nwith weak-to-low LD (r² < 0.5), eight loci ( RIN3/14q32.12, GDAP1/8q21.11, KDR/4q12, 745 \nVEZT/12q22, CD109/6q13, WNT4/1p36.12, GREB1/2p25.1, SYNE1/6q25.1) were classified 746 \nas tier 1 with six regulatory modalities. Collectively, these findings highlight the potential of 747 \nregulatory annotation to refine causal inference and identify potential functional variation 748 \nbeyond standard LD-based interpretation of GWAS signals.  749 \nAmong these prioritized tier 1 loci, the WNT4/1p36.12,VEZT/12q22 and KDR/4q12 loci are 750 \nof particular relevance, as they have been consistently identified among the most significant 751 \nSNPs associated with endometriosis [41–47]. These loci also showed weak-to-low LD (r² < 752 \n0.5) between the AlphaGenome-prioritized SNPs and the corresponding GWAS lead variants, 753 \nsuggesting that the strongest association signals do not necessarily reflect the most likely 754 \ncausal regulatory variants. Together, these features provide informative examples of how 755 \nregulatory annotation can refine causal variant prioritization beyond published GWAS lead 756 \nSNPs. The WNT4/1p36.12  contains a high-affinity estrogen receptor alpha-binding site at the 757 \nlocus, suggesting direct hormonal regulation of its expression [48]. WNT4 protein expression 758 \nis reduced in both eutopic and ectopic endometrium in endometriosis compared to normal 759 \nendometrium [49], it also regulates stromal cell proliferation, apoptosis, and decidualization, 760 \nprocesses essential for endometrial tissue remodelling. These alterations may disrupt 761 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n39 \n \nendometrial homeostasis and promote aberrant tissue remodelling, contributing to the 762 \ndevelopment and persistence of endometriotic lesions. Moreover, the expression of WNT4 in 763 \nnormal peritoneum suggests that endometriosis arising via metaplasia cannot be completely 764 \nexcluded [50]. Similarly, the VEZT/12q22 locus is characterised by increased VEZT 765 \nexpression in ectopic compared with eutopic endometrium in women with endometriosis 766 \n[51], and its mRNA and protein expression shows a menstrual cycle stage-specific increase in 767 \nendometrial glands during the secretory phase [52]. The VEZT/12q22 locus contains an NF-768 \nκ B binding site, suggesting that risk variants may influence inflammatory signaling pathways 769 \nin endometriosis through modulation of NF-κ B–mediated transcriptional regulation [52]. The 770 \nKDR/4q12 locus encodes VEGFR2 gene , the principal mediator of VEGF-driven 771 \nangiogenesis and endothelial proliferation [47]. The risk variant (rs17773813[G]) at 772 \nKDR/4q12 locus is associated with disease severity, with larger effect sizes in stage III/IV 773 \nrelative to stage I/II endometriosis [47] [3]. Collectively, these loci represent functionally 774 \ncoherent examples spanning hormonal regulation ( WNT4/1p36.12), inflammatory signalling 775 \n(VEZT/12q22), and angiogenesis ( KDR/4q12) involved in endometriosis related pathways. 776 \nThus, at well-established endometriosis susceptibility loci, AlphaGenome consistently 777 \nprioritized alternative SNPs, classifying them as tier 1 across all six regulatory modalities. 778 \nNotably, these SNPs were absent from the original GWAS reports and showed weak-to-low 779 \nLD (r² < 0.5) with the corresponding lead variants, underscoring the potential of regulatory 780 \nannotation to refine causal variant identification beyond conventional GW AS-based 781 \napproaches.  782 \nThe alternate SNP was also suggested at the GREB1/2p25.1 locus, which was likewise among 783 \nthe eight tier 1 loci showing weak-to-low LD with the published GWAS lead SNP , and 784 \nencodes an early estrogen-responsive gene and a key co-regulator of estrogen receptor 785 \nactivity in hormone-driven tissues. Increased GREB1 expression has also been reported in 786 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n40 \n \nperitoneal endometriotic lesions compared with eutopic endometrium, suggesting a role in the 787 \nestrogen-driven proliferative environment of ectopic lesions [53]. Similarly, the 788 \nSYNE1/6q25.2 locus, another low-LD tier 1 locus identified by AlphaGenome, encompassing 789 \nESR1, is one of the most architecturally complex susceptibility regions in endometriosis 790 \ngenetics. This locus is regulated through long-range chromatin interactions involving ESR1, 791 \nCCDC170, ARMT1 , and SYNE1 genes, with distal regulatory elements and transcription 792 \nfactors coordinating ESR1 expression [54]. This suggests that intronic variants  in 793 \nSYNE1/6q25.2 locus may function as non-coding regulatory elements rather than protein-794 \naltering variants. Consistently, the multimodal regulatory activity identified at the 795 \nSYNE1/6q25.2 locus by AlphaGenome supports the functional relevance of this region in 796 \nestrogen-dependent endometriosis pathogenesis.  797 \nFinally, the CD109/6q13 locus, also belonging to the subset of eight tier 1 loci in which the 798 \nAlphaGenome-prioritized SNP exhibited weak-to-low LD with the published GWAS lead 799 \nvariant, encodes a GPI-anchored glycoprotein that negatively regulates TGF- β  signalling 800 \n[55,56]. Given the well-established involvement of TGF- β  signalling in endometriosis 801 \npathogenesis including inflammation, epithelial-mesenchymal transition, angiogenesis, and 802 \nfibrosis [57], CD109 represents a key modulator of these processes. Hence these 803 \nAlphaGenome-prioritized tier 1 loci, particularly those showing low LD with published 804 \nGWAS lead SNPs, represent examples of improved functional prioritization beyond statistical 805 \nassociation alone and further support the contribution of regulatory mechanisms to 806 \nendometriosis susceptibility.  807 \nAnother important finding in our study was the outside-locus analysis, which extended the 808 \nAlphaGenome interpretation beyond the published 42 GWAS lead SNP windows. Although 809 \nthe primary analysis was restricted to SNPs located within ±500 kb of each lead SNP , 3,454 810 \nSNPs lay outside these predefined regions, including 167 genome-wide significant SNPs with 811 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n41 \n \nhigh-confidence uterus-specific regulatory predictions. While most of these variants showed 812 \nlimited regulatory support, six SNPs (rs1482061, rs7772579, rs6557140, rs2982571, 813 \nrs79626929, and rs12631337) were classified as tier 1 variants. Five of these SNPs 814 \n(rs1482061, rs7772579, rs6557140, rs2982571, and rs79626929) are located on chromosome 815 \n6 and are predicted to regulate genes within ESR1 /6q25.1 and SYNE1/6q25.2 loci including 816 \nESR1, SYNE1, AKAP12, ZBTB2, ARMT1 and RMND1 (Fig. 7a). In our study, these variants 817 \nare designated as outside-locus SNPs because they fall outside the 500 kb window used for 818 \nlocus definition, despite being in proximity to the published GWAS lead SNP rs71575922 819 \nand the AlphaGenome-prioritized SNP rs147631975 regulating SYNE1 gene. Notably, these 820 \nvariants cluster near the outside-locus tier 1 SNPs, suggesting the presence of an extended 821 \nregulatory region spanning the ESR1/6q25.1-SYNE1/6q25.2 locus.  822 \nAmong the outside-locus SNPs, tier 1 variant rs7772579 is in strong LD (r² = 0.98) with the 823 \npreviously reported endometriosis-associated variant rs2206949 in ESR1 gene [46].  824 \nAdditionally, tier 2 SNP rs1971256, located in CCDC170  has also been reported as an 825 \nendometriosis risk variant [46]. The identification of multiple prioritized variants within 826 \nESR1/6q25.1-SYNE1/6q25.2 region, further supports that this is a robust and consistently 827 \nreplicated endometriosis susceptibility loci. ESR1 gene, encoding estrogen receptor α  showed 828 \nrobust genome-wide significant association across multiple meta-analyses, with several 829 \nindependent signals at this locus further substantiating its central role in estrogen-driven 830 \nlesion growth and inflammation [8,46]. Nearby genes, including SYNE1, CCDC170, 831 \nAKAP12, ZBTB2, ARMT1, and RMND1 have also been implicated through GWAS and 832 \ntranscriptome-wide analyses, showing expression correlations with ESR1 and cell-specific 833 \nexpression in endometrial epithelial cells during the secretory phase, implicated in 834 \nimplantation impairment, often associated with endometriosis [8,46,58]. AKAP12, while not 835 \nmapped to published GWAS lead SNPs, showed negative correlation with ESR1 and 836 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n42 \n \nfunctions in cytoskeletal and signalling regulation, supporting its potential modulatory role 837 \nwithin this network [59].  838 \nSimilarly, tier 1 outside-locus SNP rs12631337 predicted to regulate tumor suppressor gene 839 \ncluster on chromosome 3, including SEMA3F, GNAI2, HYAL2, RBM5, RBM6, TUSC2, 840 \nUBA7, TMEM115, IP6K1, APEH, SEMA3B, RASSF1, IFRD2, SLC38A3, MAPKAPK3, and 841 \nNPRL2 (Fig. 7b) . This SNP is in proximity to the BSN /3p21.31 locus together with the 842 \nGWAS lead SNP rs1352889. However, rs12631337 falls outside the predefined ±500 kb 843 \nlocus boundary and was therefore classified as an outside-locus variant despite its likely 844 \nregulatory connection to this region. Among the genes predicted to be regulated by this SNP, 845 \nRASSF1 is particularly notable because its tumor-suppressor isoform, RASSF1A, has been 846 \nreported to show consistent epigenetic dysregulation in endometriosis, including promoter 847 \nhypermethylation in both ectopic and eutopic endometrium, reduced expression, and 848 \nassociation with disease severity [60]. 849 \nThese findings indicate that additional genome-wide significant variants outside conventional 850 \nGWAS locus boundaries may also carry strong regulatory relevance in uterine tissue. Given 851 \nthat GW AS locus definitions are constrained by lead SNP selection and fixed window sizes, 852 \nthese findings suggest that outside-locus regulatory prioritization can complement standard 853 \nlocus-based analyses by identifying potentially functional variants not captured within 854 \nestablished GWAS regions. The comparison of AlphaGenome-prioritized SNPs with multi-855 \nancestry endometriosis GWAS SNPs [14] also supports the broader interpretation of 856 \nregulatory modalities. Of the 37 multi-ancestry lead SNPs, 22 were already present in the 857 \n10,000 SNP GWAS dataset [3] and all belonged to the outside-locus category rather than to 858 \nthe 42 GWAS lead SNP windows. This suggests that SNPs located outside the original locus 859 \nframework may subsequently be identified as endometriosis-associated signals as larger and 860 \nmore comprehensive GWAS datasets become available. Moreover, while in the multi-861 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n43 \n \nancestry GWAS study, these SNPs were mapped to specific genes, AlphaGenome predicted 862 \nmultiple regulatory target genes at the same loci, including the previously mapped genes, 863 \nthereby broadening the functional interpretation of these regions. Among these multi-ancestry 864 \nSNPs, rs1799293, previously mapped to ARHGAP26 gene was classified as tier 1 with 865 \nsupport across six regulatory modalities. AlphaGenome additionally identified FGF1 as  866 \nregulatory target gene for this  SNP alongside the previously mapped gene ARHGAP26. 867 \nPromoter polymorphism -1385A/G (rs30411) at FGF1  has been associated with 868 \nendometriosis risk, with the A allele showing reduced frequency in endometriosis cases, 869 \nimplicating altered angiogenic signaling [61]. In parallel, ARHGAP26 is downregulated in 870 \nectopic and eutopic endometrium compared with normal endometrium and was negatively 871 \nassociated with the severity of menorrhagia [62]. Together, the outside-locus and multi-872 \nancestry GWAS comparison results suggest that functional interpretation may benefit from 873 \nextending beyond previously defined lead-SNP windows. While the 42 published loci remain 874 \nthe central framework of this study, the outside-locus findings highlight additional genome-875 \nwide significant variants with predicted uterine regulatory effects. These variants may 876 \nprovide useful candidates for future fine-mapping, replication, and functional validation as 877 \nlarger GW AS datasets continue to refine the genetic architecture of endometriosis.   878 \nConclusions 879 \nThis study demonstrates that integrating GWAS signals with uterus-specific regulatory 880 \npredictions from AlphaGenome can refine the functional interpretation of endometriosis risk 881 \nloci. By extending analysis beyond published lead SNPs to include additional GWAS-ranked 882 \nand surrounding variants, we show that non-lead variants can exhibit stronger predicted 883 \nregulatory effects and multimodal support across loci. This re-prioritization framework, 884 \nsupported by tissue-specific filtering and multi-layer regulatory annotation, improves the 885 \nidentification of likely causal non-coding variants and provides a more nuanced view of the 886 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n44 \n \ngenetic architecture underlying endometriosis. However, the findings are based on 887 \ncomputational predictions and require experimental validation. In addition, uterus-specific 888 \nfiltering may not capture all disease-relevant cell types, and the strict prioritization threshold 889 \nmay exclude variants with moderate functional effects. The fixed locus window and LD 890 \nassumptions may also miss long-range or ancestry-specific regulatory mechanisms. Overall, 891 \nthis approach provides a refined and scalable framework that can be incorporated into 892 \nstandard GWAS analysis pipelines to enhance variant prioritization and functional 893 \ninterpretation, thereby supporting more systematic identification of candidate causal variants 894 \nin endometriosis and related complex traits. 895 \nAbbreviations 896 \nATAC-seq: Assay for Transposase-Accessible Chromatin sequencing 897 \nCAGE: Cap Analysis of Gene Expression 898 \nDNase-seq: DNase sequencing 899 \nGTEx: Genotype-Tissue Expression 900 \nGWAS: Genome-wide association study 901 \nPRO-cap: Precision Run-On coupled with cap analysis 902 \nRNA-seq: RNA sequencing 903 \nSNP: Single-nucleotide polymorphism 904 \nTF ChIP-seq: Transcription factor chromatin immunoprecipitation sequencing 905 \neQTL: Expression quantitative trait locus 906 \nLD: Linkage disequilibrium 907 \n3D: Three-dimensional 908 \nDeclarations 909 \nEthics approval and consent to participate 910 \nNot applicable. 911 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n45 \n \nConsent for publication 912 \nNot applicable. 913 \nData Availability 914 \nAll data analysed and presented in this study are included in this published article and its 915 \nadditional files. Additional global AlphaGenome raw prediction outputs generated from the 916 \nanalysis of the 10,000 SNPs and the analysis code are available from the corresponding 917 \nauthor upon reasonable request. 918 \nCompeting interests 919 \nThe authors declare that they have no competing interests. 920 \nFunding 921 \nThe present study was funded by: Horizon Europe grant (NESTOR, grant no. 101120075) of 922 \nthe European Commission; Novo Nordisk Foundation (grant no. NNF24OC0092384); 923 \nEstonian Research Council (grants nos. PSG1082, PRG1076); Swedish Research Council 924 \n(grant no. 2024-02530); Estonian Ministry of Education and Research Centres of Excellence 925 \ngrant TK214 name of CoE and the Sigrid Jusélius Foundation. 926 \nAuthors’ contributions 927 \nSanu Bifal Maji conceived the study design, performed the AlphaGenome-based variant 928 \nannotation and prioritization analyses, conducted data processing, generated figures and 929 \ntables, interpreted the results, and drafted the manuscript. Alberto Sola-Leyva, Apostol 930 \nApostolov, Lucia Blanco-Rodriguez contributed to data interpretation, methodological 931 \nrefinement, and manuscript revision. Andres Salumets and Amruta D. S. Pathare supervised 932 \nthe study, contributed to conceptualization and interpretation of the results, and critically 933 \nrevised the manuscript. All authors read and approved the final manuscript. 934 \n 935 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n46 \n \nReferences 936 \n 1. Zondervan KT, Becker CM, Koga K, Missmer SA, Taylor RN, Viganò P. Endometriosis. 937 \nNat Rev Dis Primers. 2018;4:9. https://doi.org/10.1038/s41572-018-0008-5 938 \n2. Nnoaham KE, Hummelshoj L, Webster P, d’Hooghe T, De Cicco Nardone F, De Cicco 939 \nNardone C, et al. Impact of endometriosis on quality of life and work productivity: a 940 \nmulticenter study across ten countries. Fertility and Sterility. 2011;96:366-373.e8. 941 \nhttps://doi.org/10.1016/j.fertnstert.2011.05.090 942 \n3. Rahmioglu N, Mortlock S, Ghiasi M, Møller PL, Stefansdottir L, Galarneau G, et al. The 943 \ngenetic basis of endometriosis and comorbidity with other pain and inflammatory conditions. 944 \nNat Genet. 2023;55:423–36. https://doi.org/10.1038/s41588-023-01323-z 945 \n4. Treloar SA, O’Connor DT, O’Connor VM, Martin NG. Genetic influences on 946 \nendometriosis in an Australian twin sample. Fertility and Sterility. 1999;71:701–10. 947 \nhttps://doi.org/10.1016/S0015-0282(98)00540-8 948 \n5. Saha R, Pettersson HJ, Svedberg P, Olovsson M, Bergqvist A, Marions L, et al. Heritability 949 \nof endometriosis. Fertility and Sterility. 2015;104:947–52. 950 \nhttps://doi.org/10.1016/j.fertnstert.2015.06.035 951 \n6. GTEx Consortium. Genetic effects on gene expression across human tissues. Nature. 952 \n2017;550:204–13. https://doi.org/10.1038/nature24277 953 \n7. Ruth KS, Beaumont RN, Tyrrell J, Jones SE, Tuke MA, Yaghootkar H, et al. Genetic 954 \nevidence that lower circulating FSH levels lengthen menstrual cycle, increase age at 955 \nmenopause and impact female reproductive health. Hum Reprod. 2016;31:473–81. 956 \nhttps://doi.org/10.1093/humrep/dev318 957 \n8. Marla S, Mortlock S, Houshdaran S, Fung J, McKinnon B, Holdsworth-Carson SJ, et al. 958 \nGenetic risk factors for endometriosis near estrogen receptor 1 and coexpression of genes in 959 \nthis region in endometrium. Molecular Human Reproduction. 2021;27:gaaa082. 960 \nhttps://doi.org/10.1093/molehr/gaaa082 961 \n9. Hautakangas H, Winsvold BS, Ruotsalainen SE, Bjornsdottir G, Harder A VE, Kogelman 962 \nLJA, et al. Genome-wide analysis of 102,084 migraine cases identifies 123 risk loci and 963 \nsubtype-specific risk alleles. Nat Genet. 2022;54:152–60. https://doi.org/10.1038/s41588-964 \n021-00990-0 965 \n10. Jones A V , Hockley JRF, Hyde C, Gorman D, Sredic-Rhodes A, Bilsland J, et al. Genome-966 \nwide association analysis of pain severity in dysmenorrhea identifies association at 967 \nchromosome 1p13.2, near the nerve growth factor locus. Pain. 2016;157:2571–81. 968 \nhttps://doi.org/10.1097/j.pain.0000000000000678 969 \n11. Painter JN, Anderson CA, Nyholt DR, Macgregor S, Lin J, Lee SH, et al. Genome-wide 970 \nassociation study identifies a locus at 7p15.2 associated with endometriosis. Nat Genet. 971 \n2011;43:51–4. https://doi.org/10.1038/ng.731 972 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n47 \n \n12. Johnston KJA, Adams MJ, Nicholl BI, Ward J, Strawbridge RJ, Ferguson A, et al. 973 \nGenome-wide association study of multisite chronic pain in UK Biobank. Zhu X, editor. 974 \nPLoS Genet. 2019;15:e1008164. https://doi.org/10.1371/journal.pgen.1008164 975 \n13. Ruth KS, Campbell PJ, Chew S, Lim EM, Hadlow N, Stuckey BG, et al. Genome-wide 976 \nassociation study with 1000 genomes imputation identifies signals for nine sex hormone-977 \nrelated phenotypes. Eur J Hum Genet. 2016;24:284–90. 978 \nhttps://doi.org/10.1038/ejhg.2015.102 979 \n14. Koller D, He J, Løkhammer S, Aranda S, Qiu D, Davtian D, et al. Multi-ancestry 980 \ngenome-wide association and integrated multi-omics analyses of endometriosis and its 981 \nclinical manifestations. Nat Genet [Internet]. 2026 [cited 2026 May 9]; 982 \nhttps://doi.org/10.1038/s41588-026-02582-2 983 \n15. Cheng J, Deng Y , Zhou J. Role of the Ubiquitin System in Chronic Pain. Front Mol 984 \nNeurosci. 2021;14:674914. https://doi.org/10.3389/fnmol.2021.674914 985 \n16. Ding Y-Q, Luo H, Qi J-G. MHCII-restricted T helper cells: an emerging trigger for 986 \nchronic tactile allodynia after nerve injuries. J Neuroinflammation. 2020;17:3. 987 \nhttps://doi.org/10.1186/s12974-019-1684-0 988 \n17. Ilad RS, Fleming SD, Bebington CR, Murphy CR. Ubiquitin is associated with the 989 \nsurvival of ectopic stromal cells in endometriosis. Reprod Biol Endocrinol. 2004;2:69. 990 \nhttps://doi.org/10.1186/1477-7827-2-69 991 \n18. García-Gómez E, Vázquez-Martínez ER, Reyes-Mayoral C, Cruz-Orozco OP, Camacho-992 \nArroyo I, Cerbón M. Regulation of Inflammation Pathways and Inflammasome by Sex 993 \nSteroid Hormones in Endometriosis. Front Endocrinol. 2020;10:935. 994 \nhttps://doi.org/10.3389/fendo.2019.00935 995 \n19. Gregus AM, Levine IS, Eddinger KA, Yaksh TL, Buczynski MW. Sex differences in 996 \nneuroimmune and glial mechanisms of pain. Pain. 2021;162:2186–200. 997 \nhttps://doi.org/10.1097/j.pain.0000000000002215 998 \n20. Lenert ME, Avona A, Garner KM, Barron LR, Burton MD. Sensory Neurons, 999 \nNeuroimmunity, and Pain Modulation by Sex Hormones. Endocrinology. 2021;162:bqab109. 1000 \nhttps://doi.org/10.1210/endocr/bqab109 1001 \n21. Fung JN, Girling JE, Lukowski SW, Sapkota Y , Wallace L, Holdsworth-Carson SJ, et al. 1002 \nThe genetic regulation of transcription in human endometrial tissue. Human Reproduction. 1003 \n2017;32:893–904. https://doi.org/10.1093/humrep/dex006 1004 \n22. O’Mara TA, Spurdle AB, Glubb DM, Endometrial Cancer Association Consortium. 1005 \nAnalysis of Promoter-Associated Chromatin Interactions Reveals Biologically Relevant 1006 \nCandidate Target Genes at Endometrial Cancer Risk Loci. Cancers. 2019;11:1440. 1007 \nhttps://doi.org/10.3390/cancers11101440 1008 \n23. Zondervan KT, Cardon LR, Kennedy SH. The genetic basis of endometriosis: Current 1009 \nOpinion in Obstetrics and Gynecology. 2001;13:309–14. https://doi.org/10.1097/00001703-1010 \n200106000-00011 1011 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n48 \n \n24. Nathan A, Reinhardt P, Kruspe D, Jörß T, Groth M, Nolte H, et al. The Wilms tumor 1012 \nprotein Wt1 contributes to female fertility by regulating oviductal proteostasis. Human 1013 \nMolecular Genetics. 2017;26:1694–705. https://doi.org/10.1093/hmg/ddx075 1014 \n25. Jones MR, Goodarzi MO. Genetic determinants of polycystic ovary syndrome: progress 1015 \nand future directions. Fertility and Sterility. 2016;106:25–32. 1016 \nhttps://doi.org/10.1016/j.fertnstert.2016.04.040 1017 \n26. Avsec Ž, Latysheva N, Cheng J, Novati G, Taylor KR, Ward T, et al. Advancing 1018 \nregulatory variant effect prediction with AlphaGenome. Nature. 2026;649:1206–18. 1019 \nhttps://doi.org/10.1038/s41586-025-10014-0 1020 \n27. Jaganathan K, Kyriazopoulou Panagiotopoulou S, McRae JF, Darbandi SF, Knowles D, 1021 \nLi YI, et al. Predicting Splicing from Primary Sequence with Deep Learning. Cell. 1022 \n2019;176:535-548.e24. https://doi.org/10.1016/j.cell.2018.12.015 1023 \n28. Avsec Ž, Weilert M, Shrikumar A, Krueger S, Alexandari A, Dalal K, et al. Base-1024 \nresolution models of transcription-factor binding reveal soft motif syntax. Nat Genet. 1025 \n2021;53:354–66. https://doi.org/10.1038/s41588-021-00782-6 1026 \n29. Cochran K, Yin M, Mantripragada A, Schreiber J, Marinov GK, Shah SR, et al. 1027 \nDissecting the cis-regulatory syntax of transcription initiation with deep learning [Internet]. 1028 \nGenomics; 2024 [cited 2026 Apr 13]. https://doi.org/10.1101/2024.05.28.596138 1029 \n30. Mansour MR, Abraham BJ, Anders L, Berezovskaya A, Gutierrez A, Durbin AD, et al. An 1030 \noncogenic super-enhancer formed through somatic mutation of a noncoding intergenic 1031 \nelement. Science. 2014;346:1373–7. https://doi.org/10.1126/science.1259037 1032 \n31. Cano-Gamez E, Trynka G. From GWAS to Function: Using Functional Genomics to 1033 \nIdentify the Mechanisms Underlying Complex Diseases. Front Genet. 2020;11:424. 1034 \nhttps://doi.org/10.3389/fgene.2020.00424 1035 \n32. Maurano MT, Humbert R, Rynes E, Thurman RE, Haugen E, Wang H, et al. Systematic 1036 \nlocalization of common disease-associated variation in regulatory DNA. Science. New York, 1037 \nN.Y .; 2012;337:1190–5. https://doi.org/10.1126/science.1222794 1038 \n33. Garibaldi-Ríos AF, Rodríguez-Gutiérrez PG, García-Díaz JM, Zúñiga-González GM, 1039 \nFiguera LE, Gómez-Meda BC, et al. Regulatory Effects of Endometriosis-Associated Genetic 1040 \nVariants: A Multi-Tissue eQTL Analysis. Diseases. 2025;13:248. 1041 \nhttps://doi.org/10.3390/diseases13080248 1042 \n34. Rahmioglu N, Nyholt DR, Morris AP, Missmer SA, Montgomery GW, Zondervan KT. 1043 \nGenetic variants underlying risk of endometriosis: insights from meta-analysis of eight 1044 \ngenome-wide association and replication datasets. Human Reproduction Update. 1045 \n2014;20:702–16. https://doi.org/10.1093/humupd/dmu015 1046 \n35. Hindorff LA, Sethupathy P, Junkins HA, Ramos EM, Mehta JP, Collins FS, et al. 1047 \nPotential etiologic and functional implications of genome-wide association loci for human 1048 \ndiseases and traits. Proc Natl Acad Sci USA. 2009;106:9362–7. 1049 \nhttps://doi.org/10.1073/pnas.0903103106 1050 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n49 \n \n36. Mousavi Z, Arvanitis M, Duong T, Brody JA, Battle A, Sotoodehnia N, et al. 1051 \nPrioritization of causal genes from genome-wide association studies by Bayesian data 1052 \nintegration across loci. Rigoutsos I, editor. PLoS Comput Biol. 2025;21:e1012725. 1053 \nhttps://doi.org/10.1371/journal.pcbi.1012725 1054 \n37. Ghafouri-Fard S, Dashti S, Taheri M. PCAT1: An oncogenic lncRNA in diverse cancers 1055 \nand a putative therapeutic target. Experimental and Molecular Pathology. 2020;114:104429. 1056 \nhttps://doi.org/10.1016/j.yexmp.2020.104429 1057 \n38. Panir K, Schjenken JE, Robertson SA, Hull ML. Non-coding RNAs in endometriosis: a 1058 \nnarrative review. Human Reproduction Update. 2018;24:497–515. 1059 \nhttps://doi.org/10.1093/humupd/dmy014 1060 \n39. Abbaszadeh M, Karimi M, Rajaei S. The landscape of non-coding RNAs in the 1061 \nimmunopathogenesis of Endometriosis. Front Immunol. 2023;14:1223828. 1062 \nhttps://doi.org/10.3389/fimmu.2023.1223828 1063 \n40. Long E, Williams J, Zhang H, Choi J. An evolving understanding of multiple causal 1064 \nvariants underlying genetic association signals. The American Journal of Human Genetics. 1065 \n2025;112:741–50. https://doi.org/10.1016/j.ajhg.2025.01.018 1066 \n41. Meidyana S, Isfandiary S, Primariawan RY . WNT4 (rs7521902 and rs16826658) 1067 \npolymorphism and its association with endometriosis – A systematic review and meta-1068 \nanalysis. European Journal of Obstetrics & Gynecology and Reproductive Biology. 1069 \n2024;295:111–7. https://doi.org/10.1016/j.ejogrb.2024.01.038 1070 \n42. Albertsen HM, Chettier R, Farrington P , Ward K. Genome-Wide Association Study Link 1071 \nNovel Loci to Endometriosis. Huang L-M, editor. PLoS ONE. 2013;8:e58257. 1072 \nhttps://doi.org/10.1371/journal.pone.0058257 1073 \n43. Mafra F, Catto M, Bianco B, Barbosa CP, Christofolini D. Association of WNT4 1074 \npolymorphisms with endometriosis in infertile patients. J Assist Reprod Genet. 1075 \n2015;32:1359–64. https://doi.org/10.1007/s10815-015-0523-1 1076 \n44. Meidyana S, Isfandiary S, Primariawan RY . WNT4 (rs7521902 and rs16826658) 1077 \npolymorphism and its association with endometriosis – A systematic review and meta-1078 \nanalysis. European Journal of Obstetrics & Gynecology and Reproductive Biology. 1079 \n2024;295:111–7. https://doi.org/10.1016/j.ejogrb.2024.01.038 1080 \n45. Painter JN, Anderson CA, Nyholt DR, Macgregor S, Lin J, Lee SH, et al. Genome-wide 1081 \nassociation study identifies a locus at 7p15.2 associated with endometriosis. Nat Genet. 1082 \n2011;43:51–4. https://doi.org/10.1038/ng.731 1083 \n46. Sapkota Y , Steinthorsdottir V , Morris AP , Fassbender A, Rahmioglu N, De Vivo I, et al. 1084 \nMeta-analysis identifies five novel loci associated with endometriosis highlighting key genes 1085 \ninvolved in hormone metabolism. Nat Commun. 2017;8:15539. 1086 \nhttps://doi.org/10.1038/ncomms15539 1087 \n47. Steinthorsdottir V , Thorleifsson G, Aradottir K, Feenstra B, Sigurdsson A, Stefansdottir L, 1088 \net al. Common variants upstream of KDR encoding VEGFR2 and in TTC39B associate with 1089 \nendometriosis. Nat Commun. 2016;7:12350. https://doi.org/10.1038/ncomms12350 1090 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n50 \n \n48. Pavlič ev M, McDonough-Goldstein CE, Zupan AM, Muglia L, Hu Y-C, Kong F, et al. A 1091 \ncommon allele increases endometrial Wnt4 expression, with antagonistic implications for 1092 \npregnancy, reproductive cancers, and endometriosis. Nat Commun. 2024;15:1152. 1093 \nhttps://doi.org/10.1038/s41467-024-45338-4 1094 \n49. Liang Y , Li Y , Liu K, Chen P, Wang D. Expression and Significance of WNT4 in Ectopic 1095 \nand Eutopic Endometrium of Human Endometriosis. Reprod Sci. 2016;23:379–85. 1096 \nhttps://doi.org/10.1177/1933719115602763 1097 \n50. Gaetje R, Holtrich U, Engels K, Kissler S, Rody A, Karn T, et al. Endometriosis may be 1098 \ngenerated by mimicking the ontogenetic development of the female genital tract. Fertility and 1099 \nSterility. 2007;87:651–6. https://doi.org/10.1016/j.fertnstert.2006.07.1533 1100 \n51. Meola J, Rosa E Silva JC, Dentillo DB, Da Silva WA, Veiga-Castelli LC, De Souza 1101 \nBernardes LA, et al. Differentially expressed genes in eutopic and ectopic endometrium of 1102 \nwomen with endometriosis. Fertility and Sterility. 2010;93:1750–73. 1103 \nhttps://doi.org/10.1016/j.fertnstert.2008.12.058 1104 \n52. Holdsworth-Carson SJ, Fung JN, Luong HTT, Sapkota Y , Bowdler LM, Wallace L, et al. 1105 \nEndometrial vezatin and its association with endometriosis risk. Hum Reprod. 2016;31:999–1106 \n1013. https://doi.org/10.1093/humrep/dew047 1107 \n53. Pellegrini C, Gori I, Achtari C, Hornung D, Chardonnens E, Wunder D, et al. The 1108 \nexpression of estrogen receptors as well as GREB1, c-MYC, and cyclin D1, estrogen-1109 \nregulated genes implicated in proliferation, is increased in peritoneal endometriosis. Fertil 1110 \nSteril. 2012;98:1200–8. https://doi.org/10.1016/j.fertnstert.2012.06.056 1111 \n54. Lung DK, Reese RM, Alarid ET. Intrinsic and Extrinsic Factors Governing the 1112 \nTranscriptional Regulation of ESR1. Horm Cancer. 2020;11:129–47. 1113 \nhttps://doi.org/10.1007/s12672-020-00388-0 1114 \n55. Nilufer R, Karina B, Paraskevi C, Rebecca D, Genevieve G, Ayush G, et al. Large-scale 1115 \ngenome-wide association meta-analysis of endometriosis reveals 13 novel loci and 1116 \ngenetically-associated comorbidity with other pain conditions [Internet]. Genomics; 2018 1117 \n[cited 2026 May 19]. https://doi.org/10.1101/406967 1118 \n56. Galarneau G, Fontanillas P, the Celmatix Research Team, the 23andMe Research Team, 1119 \nClementi C, Hu-Seliger T, et al. Genome-wide association studies on endometriosis and 1120 \nendometriosis-related infertility [Internet]. Genetics; 2018 [cited 2026 May 19]. 1121 \nhttps://doi.org/10.1101/401448 1122 \n57. Young VJ, Brown JK, Saunders PTK, Duncan WC, Horne AW. The Peritoneum Is Both a 1123 \nSource and Target of TGF- β  in Women with Endometriosis. Dettman R, editor. PLoS ONE. 1124 \n2014;9:e106773. https://doi.org/10.1371/journal.pone.0106773 1125 \n58. Guare LA, Das J, Caruth L, Rajagopalan A, Akerele AT, Brumpton BM, et al. Expanding 1126 \nthe genetic landscape of endometriosis: Integrative -omics analyses implicate key genes and 1127 \npathways in a multi-ancestry study of over one million women [Internet]. Obstetrics and 1128 \nGynecology; 2024 [cited 2026 May 19]. https://doi.org/10.1101/2024.11.26.24316723 1129 \n59. Li H. Physiologic and pathophysiologic roles of AKAP12. Sci Prog. 1130 \n2022;105:368504221109212. https://doi.org/10.1177/00368504221109212 1131 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint \n\n51 \n \n60. Wu Y , Zhang M, Zhang X, Xu Z, Jin W. Methylation status and protein expression of 1132 \nRASSF1A in endometriosis. Oncology Letters. 2016;11:4107–12. 1133 \nhttps://doi.org/10.3892/ol.2016.4512 1134 \n61. Kang S, Li S-Z, Wang N, Zhou R-M, Wang T, Wang D-J, et al. Association between 1135 \ngenetic polymorphisms in fibroblast growth factor (FGF)1 and FGF2 and risk of 1136 \nendometriosis and adenomyosis in Chinese women. Human Reproduction. 2010;25:1806–11. 1137 \nhttps://doi.org/10.1093/humrep/deq128 1138 \n62. Jiang C, Gong W, Chen R, Ke H, Qu X, Yang W, et al. RhoA/ROCK/ARHGAP26 1139 \nsignaling in the eutopic and ectopic endometrium is involved in clinical characteristics of 1140 \nadenomyosis. J Int Med Res. 2018;46:5019–29. https://doi.org/10.1177/0300060518789038 1141 \n 1142 \n . CC-BY-NC-ND 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)\nThe copyright holder for this preprint this version posted June 30, 2026. ; https://doi.org/10.64898/2026.06.27.26356730doi: medRxiv preprint","source_license":"CC0","license_restricted":false}