{"paper_id":"6373b4e5-a749-46d5-acd4-01cee1aaa3f8","body_text":"1 \nTitle 1 \nFull Tit le: Enhancing Genetic Association Power i n Endometrio sis through U nsupervised Clustering 2 \nof Clinical Subtypes Identified fr om Ele ctronic Health Records 3 \nShort Title: The genetics of clinical end ometriosis subtypes 4 \nAuthors 5 \nLindsay Guare 1 , L eigh Ann Hu mphre y 2 ,  Margaret R ush 2 ,  Meredi th Pollie 2 ,  Yuan L uo 3 ,  Chunhua  6 \nWeng 4 , Wei-Qi Wei 5 , Lea h Kottyan 6 ,  Gail Jarvik 7 , Noemie Elh adad 4 , P enn Medicine Biobank, 7 \nRegeneron Genetics Center, Krina Z on dervan 8 , Stace y Missmer 9 , Marijana V ujkovic 10 , Digna Velez-8 \nEdwards 5 , Suneeta Senapati 2 , Shefali Setia-Verma 11 * 9 \n*Corre sponding  Author: Shefali Setia-Verma (shefali.setiaverma@pennme dici ne.upenn.e du)  10 \nAffiliations  11 \n1. Genomics and Computational Biolog y, University of Pennsylvania 12 \n2.  Department of  Obstetrics and Gynecology, Hospital of  the University of  Pennsylvania  13 \n3. Northwestern University  14 \n4. Columbia University 15 \n5. Vanderbilt University Medica l Cente r 16 \n6. Cincinnati Children’s Hospital Medical Center 17 \n7. University of Washington   18 \n8. University of Ox fo rd 19 \n9. Michigan State University  20 \n10. University of Pennsylvania 21 \n11. Department of Pathology and Labo ratory Medicine, University of Pennsylvania 22 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \nNOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.\n\n2 \nAbstract 23 \nBackground: Endometriosis af fects 10% of repr oductive-age women, and yet, it goes undiagnosed 24 \nfor 3.6 years on average  after sympto ms onset. Despite large GWAS meta-a nalyses (N >  750 ,000),  25 \nonly a few d ozen causal loci have been identified. We hypothesized th at the cha llenges in 26 \nidentifying causal genes for endometriosis stem fr om heterogeneity across clinical and biological 27 \nfactors underlying end ometriosis diag nosis.  28 \nMethods: We extracted known endometriosis risk factors, symptoms, and co ncomitant conditions  29 \nfrom the Penn Medicine Biobank (PMBB) and perfo rmed unsupervised spectr al clust ering on 4,078 30 \nwomen with endometriosis. The 5 clusters were characterized by utilizin g additional electronic 31 \nhealt h record (EHR ) variables, such as  endometriosis-related comor bidities a nd confi rmed surgical 32 \nphenotypes. From four EHR-linked ge netic da tasets,  PMBB, e MERGE, AOU , a nd UKBB, we extracted  33 \nlead variants and tag variants 39 known end ometriosis loci  f or as sociati on testing. We meta-34 \nanalyzed ancestry-stratified case/cont rol tests for each locus and cluster in addition to a positive 35 \ncontrol (T otal N endometriosis cases  = 10,108) .  36 \nResults: We have designated the five  subtype clust ers as pain comorbidities, uterine disorders, 37 \npregnancy complications, c ardiomet abolic comorbidities, and EHR-asy mptomatic based  on  38 \nenriched features from each group.  One locus, RNLS,  surpassed the  gen ome-wide significan t  39 \nthreshold in the positive control. Thirteen more loci reached a Bon ferr oni t hreshold of 1.3 x  10 -3  40 \n(0.05 /  39) in  the  positive control.  The  cluster-stratified  tests yield ed more si gnificant associations  41 \nthan t he positive  control for anywhere from 5 to 15 loci depending on t he cl uster. Bonferroni  42 \nsignificant loci were identified for f o ur out of five clusters, including WN T4  and GREB1  f o r  t h e  43 \nuterine disorders cluster , RNLS for the  cardiometabolic c luster,  FSHB for the pregnancy 44 \ncomplications cluster, and SYNE1  a n d  CDKN2B -AS 1  f or the EHR-asymptoma tic cluster. This study  45 \nenhances our understanding of  the clinical presentation pat terns of end ometriosis su btypes, 46 \nshowcasing the inn ovative approach employed to investigate this complex disease.  47 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n3 \nAbbreviations 48 \nAOU  All of Us Bio bank  49 \neMERGE electronic medical record and genomic s network  50 \nEHR  electronic healt h record 51 \nGWAS  genome-wide ass ociation study 52 \nICD   international classification of  diseases  53 \nPMBB  Penn Medicine biobank  54 \nUKBB   United Kingdom bi obank  55 \nIntroduction 56 \nEndometriosis, a complex gynecological condition affects 10% of women of reproductive  57 \nage globally and more than 50% of wo men with infertility (1), ye t it often go es either undiagnosed 58 \nor misdiagno sed, leading to delay ed diagnoses and delivery of  eff ective th erapy (2,3). 59 \nEndometriosis is  primarily characteriz ed by the presence of endometrial-like tissue outside of the  60 \nuterus. For managing the condition without surgery, th e main trea tments i nclude pain reli ef and 61 \nhormone-based therapies, neither of  which are curative. A notable nu mber o f w omen with  62 \nendometriosis receive opioids for pai n management , despite  th e need for more sustainable and 63 \neffective treatm ent options  (4,5). On the other hand, hormonal therapies m ay have  limi tations to 64 \nutiliza tion due to severe side effects o r a desire to become pregnant. Typically , the trea tment for  65 \nendometriosis often includes b oth medical and surgical approaches, however 30-50% o f patients  66 \nwith severe endome triosis may require a second surgery within 3-5  years (6). The most 67 \ncomprehensive surgical management  i nvolves a hysterectomy  with  bilatera l salpingoophorectomy  68 \n(7).Treatment and  hea lth  care visits accumul ate many direc t and indirec t costs for w omen with 69 \nendometriosis. The estimated economic cost o f endometriosi s in the US is ~$ 10k per patient which  70 \nis ~14% higher than that of diabetes ( 8), and does not include the costs patients incur by having t o  71 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n4 \nmiss wo rk days because o f their symptoms. In total, endome triosis pres ents a high economic  72 \nburden that  exceeds $22  billion in  th e U.S. a lone (9).  The condition not o nly imposes significant 73 \ncosts but also involves severe  sym ptoms, del ayed di agnosis, limited  tr eatment  options, and 74 \nfinancial strain:  challenges tha t could be significantly mitigated with a more detaile d  75 \nunderstanding of  the disease.  76 \nElectronic Healt h Records (EHRs) represent a rich,  yet underuti lized , da ta source for  77 \ncapturing the  phenotypic spec tru m of endometriosis (10). Although the  symptoms for  78 \nendometriosis  can be quite severe , inc luding chronic debilita ting pain,  dyspa reunia, and inferti lity , 79 \nthe average ti me to diagnosis is 4.5 years (11), in part because the only way to definitively diagnose 80 \nendometriosis is by surgical observati on o f endometrial lesions gro wing out side of the uterus (e.g.  81 \nabdominal cavity , pel vis, ovaries, e tc.)  (12). The variabilit y in symptoms an d disease presentation 82 \nadds to the  difficulty of diagnosis and h inders the optima l use of ele ctronic he alth records (EHRs) in  83 \nresearch for accurately identifying affected individua ls and control subjects (13–15), which is  84 \ncritical for understanding the disease a nd advancing treatmen t strategies. T he  depth and breadt h of  85 \nEHR data provide a unique opport unity to apply unsupervised learnin g techniques fo r the  86 \nidentification of distinct phenotypic clusters that ma y correspond to  clinical subtypes of  87 \nendometriosis. Such an approach a li gns with precision medicine's goal t o tailor diagnosis and  88 \ntreatment  strategies to individu al pati ent charac teristics, pot ential ly re veali ng novel insights into  89 \nthe disease's pathophysiology. 90 \nBetter understanding of the disease m echanisms of endometrio sis c ould lea d to improved  91 \ndiagnostic practices, reducing costs  to the heal thcare system and improving quality of life through  92 \ntreatment  and earl ier diagnosis fo r patients. In spite of the  preval e nce and severity  of  93 \nendometriosis, etiology of endometr iosis is still poorly understood. The pursuit thus far of 94 \nbiomarkers and drug  targets base d on genetic contributions  of  dise ase in patients with  95 \nendometriosis has mainly include d g enome-wide association studies to id entify genetic variants 96 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n5 \ncontributing to the  disease (16 ,17).  Twin studies have  estimat ed the heritabili ty of endometriosis to 97 \nbe 47.5% (18),  and common variants  are estima ted to contributed 26% of phenotypic variance  98 \n(19), but the largest GWAS to-date ( N > 750,00 0, 60 ,674 cases) has only explained 9% of  the 99 \nphenotypic variance (17). Although these recent advances in genomic st udies have promised 100 \ninsights into the underlying genetic mechanisms of endometriosi s, yet th e heterogeneity of the 101 \ndisease presentation has consistently  complicated these efforts. Tradition al genetic association 102 \nstudies have  struggled to untangle  th e  intricate web of genotypic and pheno typic diversity  within 103 \nendometriosis patients, leading to a critical need for innovative approaches to dissect the disease's  104 \ncomplexity.  105 \nWe hypothesized that underlying  clinical heterogeneity is obscuring the genetic  106 \nmechanisms and preventing large-scale genetic studies from explaining mo re of  the heritability . 107 \nEndometriosis causes a  wide range of symptoms  and conc omitant conditions, including severe  108 \nchronic pain, gastrointestinal inflam mation, and infertilit y. Addi tionally ,  many symptoms of  109 \nendometriosis are shared between ot her gynecological diseases su ch as pr imary dysmenorrhea,  110 \novarian cysts, and pelvic infla mmato ry disease; making symptom-based diagnosis challenging  111 \n(20,21). Recent studies hav e highlight e d the importance of  complex disease s ubtyping in improving  112 \nour understanding of the genetic  me chanisms underlying endome triosis. For example, a  recent  113 \nstudy on polycystic ovary syndrom e (PCOS) used unsupervised clust ering to identify three  114 \nsubtypes of PCOS based on lab and biometric valu es before conducting genome-wide associatio n  115 \nstudy for each subt ype (22).  This ap proach allowed for a more nu anced understanding of the  116 \ngenetic basis o f PCOS and could be a pplied to endometriosis to identify  s ubtypes with distinct 117 \ngenetic mechanisms. Building on the premise that a more nuanced understanding of  endometriosi s  118 \nsubtypes could unlock new genet ic a ssociations, our study leverages unsu pervised, phenotypic  119 \nclustering analysis of EHR data to s ystematica lly iden tify and characteriz e clinical subtypes o f  120 \nendometriosis. By dissecting the he te rogeneity inherent in th e disease, we  aim to increase the  121 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n6 \npower o f genetic ass ociation analyses, facilitating the iden tification of su btype-specific disease  122 \nmechanisms. This approach not only promises to enhance our understan ding of end ometriosis  123 \ngenetics but also to refine diagnostic  criteria and inform more targeted a nd effective treat ment 124 \nstrategies.  125 \nIn conclusion, the complex nature of end ometriosis, with its  diverse s ymptoms and  126 \noverlapping features with other gynecological diseases, presents challenges for under standing its  127 \ngenetic mec hanisms. In this  manuscrip t, we  detai l the methodology and  findings of our study, which  128 \nintegrates unsupervised phenotypic c lustering with subsequent genetic association analyses fo r  129 \neach iden tified endometriosis subtyp e. By doing so, we  aim to bridge the  gap between c linical  130 \nobservations and genetic research in endometriosis, providing a roadmap for future studies to 131 \nexplore the gene tic und erpinnings of this complex disease wit h renewed  clar ity and  precision. This 132 \ndeeper understanding ma y pave the  way for more targeted and personalized ap proaches to 133 \ndiagnosis, treatment,  and manage men t of this debilitating condition. Furthe r research and large-134 \nscale genetic studies are needed to fully elucidat e the genetic archite cture of endometriosis and its  135 \nsubtypes, ultima tel y leading to improv ed outcomes f or a ffected individuals.  136 \nMethods 137 \nDatasets Used for Sub-phenotyping and Genetic Association 138 \nThe Penn Medicine Biobank (PMB B)  is the University of Pennsylvania’s healt h system-139 \nbased bio bank which consists of a bout 250,000 consented par ticipants, with 4 3,624 of those having 140 \nimputed genotype da ta (imputed to TOPMED reference panel) linked with t heir electronic heal th 141 \nrecord (EHR) history. The P MBB is an electronic he alt h record (EHR)-linked biobank that integrates  142 \na wide variety of health-related information, including diagnosis codes, lab oratory measurements, 143 \nimaging data, and lifestyle information, with genomic and bi omarker data. T he PMBB is one o f the  144 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n7 \nmost diverse medical  biobanks, with approximately 30% of participants being of non -European  145 \nancestry. This diversity is crucia l for ensuring that research findings are applicable to a broad range  146 \nof populations. The biobank also bene fits from a median of seven years of lo ngitudinal da ta in t he 147 \nEHR, providing valuable information  on participants' health histories (2 2). For our  study, we 148 \ntreated th e PMBB as two distinct datasets: those without and those with ge notype data. EHR data  149 \nfrom the non-genotyped PMBB were used for cluster derivat ion whereas the genotype d PMBB 150 \ncohort was  used in the genetic analyses. 151 \nThe Electronic Medical Re cords and Genomics (eMERGE ) network i s a National Human 152 \nGenome Research Insti tute-funded  consortium engaged in the  deve lopment  of methods and  best 153 \npractices for using the electronic me dical record as a tool for genomic re search.  The e MERGE 154 \nnetwork is a publicly-availabl e datase t  with contributions fr om multiple  hea l th systems within the 155 \nUnited Stat es which contains about 100,000 partici pants with linked h ealt h records and imputed 156 \ngenomic data (imputed  to HRC reference panel) (23). The eMERGE consortium validat ed th e 157 \nhypothesis that clinica l data derived from electronic medical records  can b e used successfully for  158 \ncomplex genomic analysis of disease susceptibility across diverse patient  populations (24). The  159 \neMERGE network has shown the effici ency that can resul t from the use of el ectronic health record 160 \ndata.  161 \nThe All  of Us (AOU) Researc h Program is an ini tiativ e crea ted by the  N IH to recruit  162 \ndemographically  diverse indi viduals  to the  largest US-based biobank to-date. Recruitment  began in  163 \n2018, and since th en, over 400, 000 pe ople have signed up and submitted ba seline questions (25).  164 \n245,388  of them have short-read whol e genome sequence dat a, col lecti vely  representing over one 165 \nbillion genetic variants (26). Participa nts’ EHRs are contributed to the AOU data processing center 166 \nusing the Sync for Science platform (27 ), which works with EHR vendors such as Epic and Cerner to  167 \ncollate structured pati ent data for rese arch use (28). 168 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n8 \nThe UK Bio bank (UKBB ) is  a large and  comprehensive dataset tha t provides v aluable 169 \nresources for  researchers  studying a w ide range o f health-related topics. The UKBB is  a p opulation-170 \nbased publicly available datase t consis ting of  about 500,000 UK citizens with EHR data, heal th 171 \nsurvey data, and imputed genotypes. T he UK Bio bank has per fo rmed genome- wide genotyping on 172 \nall participants using the UK Bio bank A xiom A rray (29). This  array directly measures 173 \napproximately 850,000 varian ts, and more than 90 million variants are imp uted using the 174 \nHaplotype Reference Cons ortium and UK10K + 1000 Genomes  reference pan els.  175 \nAll four of the biobanks mentioned above (PMBB, eMERGE, AOU, and UK BB) utilize th e 176 \nObservational Medica l Out comes Par tnership Common Data Model (OMO P-CDM) to represent  177 \nstructured EHR data in a harmonized  format (30). For this study, we utiliz ed women with ICD-178 \ndiagnosed endometriosis in the  non-genotyped PMBB cohort (N endo  = 4,0 78) as t he d erivation 179 \ndataset for the clinical subtypes. For deeper characterization of our subtypes, we perfo rmed chart-180 \nreviews on 682 randomly selected end ometriosis cases from the genotyped PMBB. Then, we meta-181 \nanalyzed  women from the  genotyped  PMBB (N =  20 ,697 , N endo  =  1, 198),  six non-pediatric sites 182 \nw i t h i n  t h e  e M E R G E  n e t w o r k  ( N  =  5 1 , 8 0 0 ,  N endo  =  2 ,243),  the  AOU  research p rogram (N = 108 ,098 ,  183 \nN endo  = 2 ,126),  and UKBB (N = 2 61,8 24,  N endo  = 4,45 1) to form our main genetic analysis test  set (N  =  184 \n442,419 , N endo  = 10,01 8).  185 \nEach of the  biobanks projected thei r samples onto th e t housand genomes reference 186 \npopulation and performed clustering t o assign genetically inferred ancestry l abels corresponding t o  187 \nthose from  the thousand genomes pr oject (31). We restricted our  genetic association analyses to  188 \nthe groups which had substanti al sam ple sizes, whic h were those with high similarity t he AFR and  189 \nEUR thousand genomes superp opulati ons. We will refer t o those groups  usi ng AFR and EUR fr om  190 \nhere on out. 191 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n9 \nExtraction of Endometriosis-Related Clinical Features 192 \n Patients wit h endometriosis have  he t erogeneous clinical presentations; t he re are a wide 193 \nvariety of ass ociated symptoms, risk f actors, and comor bidities. We fir st determined participants’ 194 \ncase-control status o f endometriosi s us ing structured EHR data: ICD-9 and I C D-10 billing codes  617 195 \nand N80, respectiv ely . Then for endom etriosis cases, we determined whether each individual h ad a 196 \nhistory of end ometriosis- related clini cal features. In tota l, we extracted 3 9 ICD-based features 197 \n(Table S1): 9 ICD -based anatomical subtypes, 14 comorbidities, 8 sympto ms, and 8 pregnancy-198 \nrelated p henotypes.  We sele cted onl y symptoms, comorbidities, and pregna ncy-related conditions 199 \nfor clustering, removing th e 9  an atomical subtyp es to be  used do wnstream in cl uster 200 \ncharacterization.  We further restrict ed these conditions to those with a  prevalence  amongst 201 \nendometriosis cases in the subtype dataset of at least 5%, lea ving us wit h 17 features for the 202 \nclustering analysis (Figure S1).  203 \nUnsupervised Clustering 204 \n We tested  four popular methods for unsupervised clus tering: spectral clus tering, density-205 \nbased spatial clustering of  applications with noise  (D BSCAN), hierarchical agg lomerative clust ering, 206 \nand k-means clustering. Spectral clust ering identifies clusters by decomp osi ng a dataset’s affinity 207 \nmatrix into its eigenvectors and then clustering in the eigenvector space using QR clustering  208 \nalgorithm (32,33) . DBSCAN is an algorithm which i dentifies dense regions of data points to discover 209 \nclusters (34). Hi erarchical  agglomerat ive cl ustering is an unsupervised c lassification method th at  210 \nuses a pairwise distance matrix to  iterativel y merge nearby points to gether (35). K-means 211 \nclustering randomly initi alizes centroids for each cluster and then alterna t es between assigning  212 \ndata points to their nearest centroid  an d adjusting the centroids until converg ence (36).  213 \nIn addition to choosing an algorithm , a common struggle with unsupervise d clustering is 214 \nchoosing a target number of clusters in a non-arbitrary way. We used several empirical metrics for  215 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n10 \nthis: silhouette score, distortion score, and a metric we developed to represent the “evenness” of  216 \nclusters. The si lhouet te score is a metr ic which considers both intra- and inter-cluster distances to 217 \nassess tightness within a c luster and distance between c lusters; high er silh ouette scores indicate  218 \nbetter quality clust ers. The distortion score  is the sum o f s quared erro r s with respect to the  219 \ncentroid of each clust er, t hus it  is desir ed to minimiz e distortion. Our evennes s metric, optimi zed by 220 \nminimization, was defined as the fractional difference between the size o f th e largest and smallest 221 \nclusters. We measured th ese metrics across tests for 2 -20 clusters for each  of the f our clustering 222 \nmethods (except f or DBSCAN which au tomatical ly infers the optimal number of clusters).  223 \nCharacterization of Unsupervised Clusters 224 \n After identifying clinical clusters wit hin our ob servation dataset, our obj ective was to  225 \ndelineat e th eir characteristics . We p erformed two-p opulation z-score pro portion tests (37) to  226 \ndetermine if the rat es of input conditions were sig nificantly different on a cluster-vs-other-clusters  227 \nbasis. For our training set (th e non-genotyped PMBB, N endo  = 4 ,078),  we examined two sets of 228 \nfeatures for the z-score tests: the 1 7 input features as wel l as ICD-based anatomical subtypes of 229 \nendometriosis including adenomyosi s, endometrioma, superficial lesions, and deep lesion s  230 \n(Supplementary Table  S2). For c haracterizing our cl usters, we also uti li zed a  chart-reviewed  231 \ndataset of 682 genotyped PMBB patients with endometriosis IC D codes. Th e features considered  232 \nhere were confirmed endometrio sis  and adenomyosis status and chart-abstracted symptoms, 233 \ncomorbidities, and surgical phenotyp es (Supplementary Tabl e S3). By co nsidering the cl uster-234 \nspecific dif ferences in these EHR-der ived features among the two datasets, we could observ e  235 \npatterns in clinical presentation. Based on these patterns, we as signed labels t o each cluster.  236 \nCluster-Stratified Candidate Gene Association Testing 237 \nTo identify genetic heterogeneity amo ng the varied clinical presentations of  endometriosis, 238 \nwe performed cluster-stratified,  ancest ry-stratified candidate  gene association studies.  Using P LINK 239 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n11 \n2.0 (38),  we extract ed single  nucl eotid e polymorphisms (SNPs) in LD (kb dist ance < 0.5  Mb and R 2  > 240 \n0.1) with 39 au tosomal lea d SNPs rep orted in the  most recent endometriosis GWAS(17). LD was  241 \ncomputed based on the thousand geno mes reference panel (31). Cluster phenotypes were assigned  242 \nfor  PMBB, eMERGE, and AOU using  a K-Nearest neighbor s’ classifier (39) wi th K=3 on the same 17  243 \nICD-based features. For each stud y,  we employed  a linear mixed mod el regression method 244 \nemployed in  SAIGE (40) to t est for associations between genotypes and case -control status. Cases 245 \nwere females with  endometriosis from one cluster and  controls were biological femal es with  no ICD  246 \nhistory of endometrio sis. In the regres sion models we included the first four principal components, 247 \nage, and batch indicators (eMERGE o nly) as covariates. The ancestry-stratified results of  these 248 \nstudies were then met a-analyz ed usin g Plink 1.9 (41) for each of the cluster -phenotypes. We also  249 \ntested a baseline  overall  endometriosis (cases from all c lusters combined) a s a positive control to 250 \nidentify how many known loci we wer e able to replicate . Because m ultip le g enetic ancestry groups  251 \nwere included , we c hose a random-effects meta-anal ysis, which  is more robust to heterogeneity  252 \n(42). 253 \nResults 254 \nDerivation, Study, and Validation Datasets 255 \nThis study uti lized  five datase ts t o investigate th e genetic mechanis ms underlying 256 \nendometriosis and its su btypes. The datasets used were endometriosis cases  in the non-gen otyped 257 \nPMBB for the deriva tion of clusters,  a c hart-reviewed endome triosis cohort to help  charac terize  the  258 \nclusters, the  genotyped PMBB, six site s within t he e MERGE network, AOU , a nd UKBB for genome-259 \nwide association analyses (See Met h ods). The sample siz es for each coh ort, the mean age  at 260 \ndiagnosis, the number o f cases and  c ontrols, and the mean age at the tim e of data pull for  each 261 \ncohort are shown in Ta ble 1. See Me t hods for details on each of the four d atasets. By leveraging 262 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n12 \nthese da tasets,  t he st udy aimed to i dentify endometriosis subtypes and  gain insights into t he 263 \ngenetic factors as sociated with endom etriosis and its  subtypes. 264 \n 265 \n  266 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n1 3\nTable 1:  Cohort sample size and average age of cases and controls for the datasets used in this 267 \nanalysis. Age was considered the age as of when the EHR data were collected. 268 \nDatase t  Endometriosis N (AFR / EUR) Mean Age (SD) \nCluster Derivation Set: \nNon-Genotyped PMBB Cases 4,078 (NA) 49.9 (13.3) \nGenetic Association Sets: \nAOU Cases 2,126 (542 / 1,584) 52.2 (12.8) \nControls 108,099 (31,435 / 76,664) 56.8 (16.8) \neMERGE Cases 2,243 (353 / 1,890) 59.9 (14.6) \nControls 49,557 (9,934 / 39,623) 59.7 (23.4) \nPMBB Cases 1,198 (562 / 636) 54.2 (12.9) \nControls 19,493 (6,524 / 12,969) 60.0 (17.8) \nUKBB Cases 4,541 (112 / 4,429) 51.5 (7.5) \nControls 257,283 (4,524 / 252,759) 56.6 (8.0) \nMeta-Analysis Totals: \nMETA Cases 10,108 (1,569 / 8,539) 53.9 (11.3) \nControls 434,432 (52,417 / 382,015) 57.1 (13.6) \nDerivation of Unsupervised of Clusters 269 \nUnsupervised clustering was performed in non-genotyped PMBB dataset of 4,078 wome n270 \nwith EHR-diagnosed endometriosis using 17 clinical features (supplementary figure S2). We teste d271 \nfour methods for unsupervised clustering as well as 19 values for the number of clusters (K=2-20 )272 \nand measured three metrics to empirically choose a clustering method and number of cluster s273 \n(Figure 1). 274 \n275 \nFigure 1: testing various clustering algorithms and K-values to empirically choose an optimal 276 \nmethod. The three metrics shown are (a) Manhattan-distance-based silhouette score, (b) distortio n277 \n3  \nn  \nd  \n)  \ns  \n \nn  \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n14 \nor sum of  s quared erro rs, and (c ) even ness represented by  the difference in  f r action between the 278 \nlargest and smallest clusters. Based on these tests, we chose spectral clustering with K=5. 279 \n 280 \n Based on these tests, we  first elimina t ed DBSCAN because the  inferred num ber of clusters  281 \nwas 131, a far too complex model to be useful or interpretable. Next, we e l iminated hierarchic al 282 \nclustering because the sizes of the resulting clusters were more uneven th an the other methods.   283 \nSpectral clustering and k-means clustering were ultimately more difficult to choo se between, but 284 \nwhen we focused on the sh apes of the  distortion curves across the va lues of K, we observed that  k-285 \nmeans lacked an “ elbow” to show a  clear optimal K va lue wh ereas spec tral clus tering clearly  286 \nindicated 5 as an idea l K with a local minimum. Thus , we chose spectral cl u stering with K=5 as our  287 \nunsupervised subtyping model.  The si zes of the final clust ers were: (1) 441 - 11%, (2) 686 - 17%,  288 \n(3) 1,151- 28%, (4) 796 - 20%, and (5) 1,004 - 25%. Figure 2 ill ustrates t he eigenvectors of the  289 \naffinity matrix which were used fo r clustering the data points. 290 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n1 5\n291 \nFigure 2: pairwise scatter plots of the first five eigenvectors of the affinity matrix used for spectral 292 \nclustering, colored by cluster. This five-dimensional eigenvector space was used for clustering. The 293 \ndiagonal shows kernel density estimator plots for each of the five eigenvectors. 294 \n  295 \n 296 \n5  \n \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n1 6\nData-Driven Cluster Characterization 297 \n After clustering, we aimed to characterize these clusters by observing patterns in clinica l298 \npresentation (prevalence) amongst the input features. We performed two sets of z-scor e299 \nproportion tests comparing prevalence of each feature between each cluster and the other fou r300 \nclusters in our training set. The first set of tests was performed on the original cluster derivatio n301 \ncohort, and the features included were the 17 input features (symptoms and comorbidities wit h302 \nprevalence > 5%) as well as ICD-defined anatomical subtypes of endometriosis (Figure 3).  303 \n304 \nFigure 3:  feature tests for the non-genotyped PMBB training set. Shown are (a) z-scores for the 305 \ndifference in proportion tests, annotated with p-values that are significant and (b) feature 306 \nprevalence by cluster to provide context for the z-score tests. 307 \n 308 \nAmong the five clusters identified in the training set, there were many input features an d309 \nICD-based anatomical subtypes with significantly different proportions. To identify distinguishin g310 \n6  \nl  \ne  \nr  \nn  \nh  \n \nd  \ng  \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n17 \nfeatures between the clust ers, we focu s on phenotypes which were signi fica ntly enriched and h ad  311 \nthe highest prevalenc e in that cluster.  Cluster one had the highest rates of  (and was  signi ficantly  312 \nenriched for)  dysuria (Z=8.9), migraine (Z=10.6), IBS (Z=10.3), fibromyalgia (Z=15.3), asth ma 313 \n(10.3), abdominal pelvic pain (Z=13 . 6), and shortness o f b reath (Z=13.5).  Cluster two had the 314 \nhighest rates of the following signif icantly enriched traits: dysmenorrhea (Z=21.9), infertilit y 315 \n(Z=5.9), irregular menstruation (Z =31.75),  leiomyoma of uterus (Z =21.9), and ut erine 316 \nendometriosis defined by IC D-9 617.0* or I CD-10 N80.0* (Z=13.4). Cluster thr ee’s defining features  317 \nwere high risk pregnancy supervision ( Z=7.1), superficial lesions defined by ICD-9 617.3* or ICD -10  318 \nN80.3* (Z=7 .1), and lower abdomin al pain (Z=14. 6). Individ uals in cl uster four had highest  319 \nprevalence of abnormal cholesterol (Z=33.1) and hyp ertension (Z=33.9), whi le clust er five was only  320 \nenriched for  unspecified endometrio sis defined as  ICD -9 617.9* or I CD-10 N8 0.9* (Z=7.0).  321 \nThe second set o f tests wa s perf ormed on a sub set of  endometriosis  cases (N=682) from  the 322 \ngenotyped PMBB for whom chart reviews were performed by OB-GYN clinical fel lows at th e 323 \nUniversity of Pennsylvania Hospital System. The feat ures tested were gold  standard confirme d  324 \ndiagnoses (endometriosis, adenomyo s is, fibr oids, and any I CD false po sitives), surgical subtypes, 325 \nhormone use at the time of con firmation procedure, and symptoms identifie d from a  combination  326 \nof structured data and notes (Figure 4 ) . 327 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n1 8\n328 \nFigure 4: feature tests for the chart reviewed PMBB dataset. Shown are (a) z-scores for the 329 \ndifference in proportion tests, annotated with p-values that are significant and (b) feature 330 \nprevalence by cluster to provide context for the z-score tests. 331 \n Because the size of our chart-reviewed dataset was limited, there were fewer significan t332 \ntests. For cluster one, the phenotypes which were most significantly prevalent were interstitia333 \ncystitis (Z=3.8) and fibromyalgia (Z=6.9). For cluster two, the defining features were confirme d334 \nadenomyosis status (Z=3.7), confirmed uterine fibroids (Z=7.1), and symptomatic bleeding (Z=5.3)335 \nCluster three’s most highly enriched features were pelvic pain (Z=3.5) and hormone use at the tim e336 \nof surgery (Z=4.1).  Considering the enriched features for each cluster among the two sets of tests337 \nwe defined the following labels for 5 clusters: (1) pain comorbidities, (2) uterine disorders, (3 )338 \npregnancy complications, (4) cardiometabolic comorbidities, and (5) EHR-asymptomatic. 339 \n8  \n \nt  \nl  \nd  \n. \ne  \n, \n)  \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n19 \nCandidate Gene Association Testing Stratified by Phenotypic Cluster 340 \n We applied the subtype classifications observed in  our  derivation set to  our fo ur genetic 341 \nassociation datasets, PMBB, eMERGE, AOU, and UKBB. We used a  K-nearest  n eighbors  model with 342 \nthe same 17 EHR-derived features to a ssign endometrio sis cases  to  the five p henotypes (Table 2). 343 \nTable 2:  c ounts and pr opo rtions of  en dometriosis cases  in each cluster by  dataset.  344 \nDatase t  P a i n  \nComorbidities \nUterine \nDisorders \nPregnancy \nComplications \nCardiometabolic \nComorbidities \nEH R -\nAsymptomatic \nCluster Derivation Set:  \nTraining  441 (10.8%) 686 (16.8%) 1,151 (28.2%)  796 (19.5%) 1,004 (24.6%)  \nGenetic Association Sets:  \nAOU 713 (21.8%)  690 (21.1%) 723 (22.1%) 783 (23.9%) 362 (11.1%) \neMERGE 495 (22.1%) 505 (22.5%) 382 (17.0%) 709 (31.6%) 152 (6.8%) \nPMBB 200 (16.7%) 222 (18.5%) 273 (22.8%) 366 (30.6%) 137 (11.4%) \nUKBB  231 (5.1%) 607 (13.4%) 842 (18.5%) 285 (6.3%) 2,576 (56.7%)  \nMeta-Analysis Totals: \nMETA 1,639 (14.6%)  2,024 (18.0%)  2,220 (19.7%)  2,143 (19.0%)  3,227 (28.7%)  \n 345 \n The smallest clust er was the pain com o rbidities cluster, with only 14.6% of  total 346 \nendometriosis cases  being  assigned t o this cluster. The EHR-asymptomatic cl uster was the largest 347 \ncluster overall.  The other three cluster s occurred i n relatively even proportio ns in the overall 348 \nmeta-analysis group at 18.0% (uterine disorders), 19.7% (pregnancy complic ations), and 19.0% 349 \n(cardiometabolic como rbidities). 350 \n To establish a reference for the expected leve l of signal replication, we began with a positive  351 \ncontrol test. We conducted associatio n tests on 39 established gene tic locat ions (autosomes only)  352 \nknown t o be  linked to endometrio sis. ( Figure 5). 353 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n2 0\n354 \nFigure 5: results for our endometriosis case vs control positive control association tests at each of 355 \nthe 39 known loci. Shown are the lead SNPs from the Rahmioglu et al 2023 GWAS as well their tag 356 \nSNPs in LD (kb distance < 0.5 Mb and R 2  > 0.1). X-axis labels are from the known GWAS. 357 \n 358 \n Our positive control test resulted in fourteen replicating loci. Only one was genome-wid e359 \nsignificant, RNLS/10q23 .31  (P = 1.91x10 -9 , rs792212:T). Thirteen were significant at a Bon Ferroni -360 \ncorrected threshold of 0.05 / 39: WNT 4/1p3 6.1 2  (P = 9.12x10 -8 , rs2235529:T), DNM 3/1q 24.3  ( P  =361 \n3.54x10 -5 , rs655853:C), GREB1/2p25 .1  (P = 6.55x10 -4 , rs34532804:A), PDLI M5/4q22. 3  (P = 8.58x10362 \n4 , rs1493112:T), EBF1/5q33. 3  (P = 7.64x10 -4 , rs1878936:C), SYNE1 /6q 25.1  (P = 3.20x10 -4363 \nrs13206045:C), GDAP1/8q21.11  (P = 4.19x10 -7 , rs10957712:T), CDKN2B-AS 1/ 9p21.3  (P = 1.75x10 -6364 \nrs10122243:T), ASTN2/9q3 3.1  (P = 9.01x10 -4 , rs62576127:A), ABO/9q34.2  (P = 5.87x10 -4365 \nrs495828:G), FSHB/1 1p14 .1  (P = 1.17x10 -3 , rs11031006:A), WT1/11p 14.1  (P = 2.85x10 -4366 \nrs72638188:T), DLEU 1/13q14 .2 (P = 2.24x10 -4 , rs9568417:G). 367 \nTo test whether stratifying by clinical presentation allowed for greater resolution in geneti c368 \nassociations, we performed case-control candidate gene association studies for the five phenotypi c369 \nclusters by meta-analyzing ancestry-stratified summary statistics from four EHR-linked geneti c370 \ndatasets: PMBB, eMERGE, AOU, and UKBB. We observe 18 / 39 loci (46%) significantly associatin g371 \nwith one or more clusters (Figure 6a, 6b). Also, for up to 15 loci, the cluster-stratified phenotype s372 \n0  \n \ne  \n-\n=  \n-\n, \n, \n, \n, \nc  \nc  \nc  \ng  \ns  \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n2 1\nyield stronger associations than the positive control despite having smaller sample sizes (Figur e373 \n6c). 374 \n375 \nFigure 6: Phenotype-specific association test results. The top panel (a) indicates which known loci 376 \nwere significantly replicated by the positive control and five clusters. The bottom left panel (b) 377 \nshows the number and names of statistically significant associations for each phenotype. The 378 \nbottom right panel (c) shows the number of loci for which each phenotype had a more significant 379 \nassociation than the baseline. 380 \n The smallest clust er, cluster one, with high rates of pain comorbidities, was not significantl y381 \nassociated with any known loci, but it was more significantly associated than the positive contro l382 \nfor eight loci as shown in Figure 6c. The uterine disorders cluster (two) was significantly associate d383 \nwith four loci, WNT4/1 p36. 12 , DNM 3/ 1q24.3 , GREB1 /2p2 5.1 , and GDAP1 /8q 21.11 . Out of the seve n384 \n1  \ne  \n \ny  \nl  \nd  \nn  \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n22 \nloci significantly associated with  the  p regnancy complications clust er (three ), three of th em were 385 \nnot signi ficantly associated with any o ther clusters or the positive contr ol: K DR/4q1 2 , 7p1 5.2 , and  386 \nKCTD9/8 p21. 2 . Cluster f our, enrich ed for  cardiometabolic c omor bidities, was sig nificantly 387 \nassociated with one locus,  RNLS/10q2 3.31 , the strongest hit from the positi ve control. RNLS w a s  388 \nalso significantly associated with clust ers three and five. El even loci were si gnificantly associated 389 \nwith the EHR-asymptomatic clus ter,  and six  o f those ( BSN /3p2 1.31 , ID 4 /6p22 .3 , CD 109 /6q13 , 390 \nMLLT10/ 10p1 2.31 , IGF1/ 12q23 .2 , and SKAP1/17q 21.3 2 ) had no other associ ations, even with the  391 \npositive control. 392 \nDiscussion 393 \n Endometriosis  presents with hetero geneous symptoms ranging fr om s evere pain to  394 \ninfertility, contributing to varying pa tient experiences and treatment responses. Several large  395 \ngenome-wide ass ociation studies and  meta-analyses have been performed for  endometriosis  to-396 \ndate.  However, the  genomic underpi nnings of endometriosis remain inco mplete ly understood, 397 \nlargely due to the clinica l het erogeneity and the limita tions of  traditional genome-wide association  398 \nstudies (GWAS) that aggregate all cas es into a single analysis pool. This a pproach may obscure  399 \ngenetic variations specific to  different  endometriosis phenotypes, thus necessitating more re fined  400 \nstratification techniques. There are various approaches to phenotyping participants for these  401 \nstudies including surgical notes and electronic health records. While there h ave also been analyses  402 \nwhich account for  disease pr ogres sion  (17), there have not been any  genom e-wide investigations  403 \ninto the genetics underlying the heterogeneous presentation patterns of  endo metriosis. 404 \n In this stu dy,  we aimed  to investiga t e the  genetics of het erogeneity in en dometriosis by  405 \ndefining data-driven subtypes in wom en from the non -genotyped PMBB end ometriosis population  406 \n(N=4,078) . We extrac ted c linical  features known to be associated wit h endometriosis and  407 \nperformed unsupervised spectral clust ering, identifying five clusters. Unsupe rvised clustering was 408 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n23 \nan ideal approach for this stud y becau se it a way to find pat terns in the d at a without introducing 409 \nprior kn owledge or  bias. We cho se spe ctral clustering with five clusters based on empirical metrics 410 \nmeasured by comparing four  diffe rent unsupervised clustering methods acro ss a range  o f K  values 411 \n(2-20 clusters). This method h ad th e best tradeoff between s quared error, silhouette score, and  412 \ncluster evenness.  413 \nTo understand the c linical  presentatio n patterns of each of the c lusters, we  compared the 414 \nrates of the inpu t feat ures, diagnoses, and chart-reviewed p henotypes am ongst them.  Based on 415 \nstatistical enrichment testing across the features,  the  clust ers were labeled as (1) pain  416 \ncomorbidities, (2) uterine disorders, (3) pregnancy complications,  (4) cardiometabolic  417 \ncomorbidities, and (5) EHR-asymptomatic. This nuance d phenotyping,  which diverges from  418 \ntraditional classifications, allows for a  deeper understanding of the pathophysiological variations 419 \nwithin endometriosis and  highlights the necessity of  tailored therapeutic approaches. 420 \n After deriving and characterizing the clusters in the non-genotyped PMBB, we used a k-421 \nnearest neighbors’ model to trans fer the subtypes to the other four EHR-li nked genetic datasets,  422 \nPMBB, eMERGE, AOU, and UKBB. We performed ancestry-stratified candidate gene testing for each  423 \nof the clusters using SAIGE and ide ntified eight genome-wide significant  signals. The gen etic 424 \nanalysis of these clusters yiel ded intr iguing results. While 4 6% of previou sly known GWA S loci  425 \nwere replicated in our study, signific ant differences in loci  association s ac ross  the clusters were 426 \nobserved. For instance, genes like WN T4  and GREB1  showed specific associations with the  uterine  427 \ndisorders and EHR asymptomatic clus ters, suggesting that these genes migh t play distinc t roles in 428 \nthe pathogenesis of these phenotypic presentations o f endometrio sis. C on versely, th e BSN gene,  429 \nalthough not statisti cal ly significant, d emonstrated greater significance in the pain and pregnancy 430 \ncomplications clusters, indicating a possible link to neurovascular or in flammatory mechanisms 431 \nthat could exacerbate these conditions. 432 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n24 \n Renalase (RNLS) is the protein associated with our only genome-wi de significant  433 \nassociation fr om the positive control,  RNLS/10q 23.3 1 . At the  Bonferroni si gn ificance threshold,  the  434 \nassociation with RNLS was si gnifi cant for  three out o f five sub-ph enotypes: pregnancy  435 \ncomplications, cardiometabolic comorbidities, and EHR-asymp tomatic.  It  was the onl y significant  436 \nassociation with the cardiometabolic cluster. RNLS is highly expressed in the heart and contributes  437 \nto regulating blood pressure (43). I n genetic association studies, RNLS has been previously 438 \nassociated with t ype 1 diabetes (44) a nd smoking initiation (45). Smoking is a known risk factor o f  439 \nendometriosis. 440 \n Cluster three, with high rates of pre gnancy-related complicat ions such as infertility and 441 \nhigh-risk pregnancy, was sig nificantly associated with seven loci including FSHB (P = 1.8 x  10 -4 ). 442 \nThe FSHB gene codes fo r the beta-subunit of follicle-stimulating hormone (FSH). FSH is essential for 443 \nfemale fertilit y and has been shown to regulate m yometrial contracti le ac tivity (46).  FSHB w a s  444 \nsignificantly associated with the positi ve control as well, but not with any of t he other clusters.  445 \n Cluster five, whic h was large ly asymp t omatic in th e EHR,  was the  largest c lu ster. Over half 446 \n(56%) of UK BB endometriosi s patients were assigned to this cluster . It is possible that those  447 \nassigned to this cluster fr om any dataset have symptoms that were not reco rded in the structured 448 \ndata which we had access to. Two well-known endometriosi s loci from  both o f the last major  449 \nGWASs are SY N E 1  and CDKN2B -AS1  ( 16,17), both of which were significant ly associated with  the  450 \npositive control and the  EHR-asymptomatic c luster.  Six loci were associated with this  clust er and  no 451 \nother phenotypes:  BSN , ID4 , CD109, MLLT10 , IGF1,  and SKAP1. MLLT10  and BSN  hav e been 452 \npreviously associated with pain perc eption and maintenance (17) . Serum  levels of IGF-1 are 453 \nsignificantly elevated in women with e ndometriosis (47). Gene e xpressi on of  ID4  is  down -regulated  454 \nin eutopic and ectopic endometria l tis sue of women with endometriosi s (4 8). CD109I  and SKAP1  455 \nhave been previously associated with e ndometrial cancers (49,50).  456 \n   457 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n25 \n Subtyping complex diseases, like endo metriosis, is crucial fo r advancing prec ision medicine. 458 \nThe findings f rom our  study underscore the uti lity of EHR as a rich resource  for disease subtyping  459 \nand genetic rese arch. T he linkage of detaile d clinic al data  with gene tic inf ormation enables the 460 \nidentification of phenotype-genotype c orrelations that are  often dilu ted in  broader GWAS analyses. 461 \nFurthermore, the use of spectral cluste ring helps elucidate t he het erogeneity within endometriosis, 462 \nproviding a framework f or understanding the m ultiface ted na ture of the  dise ase and facilit ating t he 463 \ndevelopment of personalized medicine.  464 \nHowever, it is essential to ackno wledge the limita tions of  our study. One significant  465 \nconstraint was the sample size, whic h was particularly li mited for some of th e smaller cl usters and 466 \nfor  individuals of non- European ancest ry. This limitation could potentially int roduce bias and  af fect 467 \nthe generali zability of our findings. Additionally , our study re lies on struct ured elec tronic healt h 468 \ndata only,  which m ay not capt ure th e  full clinic al pic ture and cou ld be subject to inaccuracies or 469 \nincomplete records. Lastl y, this gene tic association analyses in this stud y  only focused on the  470 \ncandidate genes that are previously known to  be as sociated with endometriosis. This appr oach  471 \nmight hav e restricted  our ability to discover novel genetic loci poten tial ly r elevant  to th e specific 472 \nclusters identified . Despite  these  li mitations, our stud y marks a  meani ngful advancemen t in  473 \nunderstanding the  genetic  factors that may contribute  to t he hetero geneity observed in 474 \nendometriosis. By focusin g on genetic associations gleaned fr om electronic health records, we offer  475 \na novel perspective tha t could be instrumental in future research and treat ment approaches. To 476 \nexpand upon the current findings, future research should aim to perform comprehensive GWAS 477 \nacross the identified endometriosis subtypes. This will enable the de tection of novel loci that could 478 \nbe crucial fo r understanding the di stinct mechanisms underlying each subtype. Additionall y,  479 \nintegrating multi-omics data (such a s transcriptomic, proteomic, and metabolomic data) could  480 \nfurther refine the molecu lar signatu res associated with each c luster,  en hancing the biological 481 \ninterpretability of the genetic associat ions. Another p romising  avenue is the longitudinal study of  482 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n26 \nthese cl usters to assess disease progr ession and treatment outcomes,  whic h could inform more  483 \neffective, personalized therapeu tic strategies.  484 \nIn conclusion, our research  highlig hts th e importance  of subtype-specific studies in  485 \nelucida ting the  genetic  basis of endometriosis. By leveraging t he c apabilities of EHR-linked  486 \nbiobank s and employing advanced  clu stering techniques,  we pav e t he wa y for more targeted an d  487 \neffective approaches to understanding  and managing this complex di sease. 488 \nAcknowledgements 489 \nResearch reported in this publicatio n was supported by the Eunice Kennedy Shriver National 490 \nInstitute of Child Health and Hu man Development of the National Instit utes of Health under award  491 \nnumber R01HD110567.  492 \n 493 \nWe acknowledge t he Penn Me dicine BioBank (PMBB) f or providing data a nd thank  the  pati ent-494 \nparticipants of Penn Medicine who c o nsented to participate in this research program. We w ould 495 \nalso like to thank the Penn Medicine BioBank team and Regeneron Genetics Center fo r providing  496 \ngenetic variant data for analysis. T he PMBB is appr oved under IRB p rotocol# 813913 and 497 \nsupported by Perelman School o f Me dicine at University of Pennsylvania, a gift fr om the Smilow  498 \nfamily, and t he National  Center for Advancing Translational Sciences  of th e Na tional Insti tutes of 499 \nHealth under CTSA award numbe r UL1TR001878.  500 \n 501 \nThis phase of the eMERGE Network w as initiated and funded by the NHGRI through the following  502 \ngrants:  U01HG008657 (Group Health Cooperative/University of Washington); U01HG008685 503 \n(Brigham and Women’s Hospital); U01HG0086 72 (Vanderbilt Univer sity Medica l Center); 504 \nU01HG008666 (Cincinnati Chil dren’s  Hospital Medical Cent er); U01HG0 06379 (Mayo Clinic);  505 \nU01HG008679  (Geisinger Clinic); U01HG008680  (Columbia Universit y Heal th Scienc es); 506 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n27 \nU01HG008684 (Chil dren’s Hospital of Philadelp hia); U01HG008 673 (Nort hwestern University); 507 \nU01HG008701 (Vanderbilt Uni versit y Medical Center serving as the Coo rdinating Center) ;  508 \nU01HG008676  (Partners Hea lthcare /Broad Institute); and U 01HG00866 4 (Baylor College of 509 \nMedicine).  510 \n 511 \nWe gratefully  acknowledge A ll  of U s participants for their contributions, without  whom this 512 \nresearch would not have been p ossi ble. We also thank the National Institu te s o f Health’s All of Us  513 \nResearch Program for  making available the participant data examined in this study.  514 \n 515 \nThis research has been conducted u sing the UK Biobank Resource unde r Application Number 516 \n32133.  517 \nReferences 518 \n1.  Gao X, Outley J, Botteman M, Spalding J,  Simon  JA, Pashos CL. Ec onomic burde n of  519 \nendometriosis. Fertil Steril. 2006 Dec 1;86(6):1561–72 .  520 \n2.  Burney RO, Giudice LC. Pathogenesis a nd pathophysiology o f endometrio sis. Fertil Steril. 2012 521 \nSep 1;98(3):511–9.  522 \n3.  Greene R, Stratton P, Cleary SD, Ballweg ML, Sinaii N. Diagno stic experience a mong 4,334 523 \nwomen repo rting sur gically diagnosed endometriosis. Fertil Steril. 2009 Jan;9 1(1):32–9.  524 \n4.  As-Sanie S, Soliman AM, Evans K, Erpel ding N, Lanier RK, Katz NP. Short-actin g and Long -acting 525 \nOpioids Utilization among W omen Dia gnosed with E ndometriosis  in the United States: A 526 \nPopulation-based Claims Study. J Minim Invasive Gynecol. 2021 Feb 1;28(2): 297-306.e 2.  527 \n5.  Soliman AM, Surrey E, Bonafede M, Nelson JK, Castelli-Haley J . Real-World Evaluation of Direct 528 \nand Indirect Economic Burden Amo ng Endometriosis  Patients in the United States. Adv Ther. 529 \n2018 Mar 1;35(3):408–23.  530 \n6.  Shakiba K, Bena  JF, McGill KM, Minger J, Falcone T. Surgical Treatment of  Endo metriosis: A  7-531 \nYear Follow-up on  the Requirement for  Further Surgery. Obstet Gynecol. 2008 532 \nJun;111(6):1285.  533 \n7.  Singh SS, Suen MWH. Surgery  f or end o metriosis: bey ond medical therapies. F ertil Steril. 2017 534 \nMar 1;107(3):549–54.  535 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n28 \n8.  Ellis K, Munro D, Clarke J. Endometrio s is Is Undervalued: A Call to Action. Fr o nt Glob Womens  536 \nHealth [Internet]. 202 2 [cited 2022 Dec 1];3. Available from:  537 \nhttps://www.frontiersin.or g/articles/10.3389 /fgwh.2022. 902371  538 \n9.  Simoens S, Dunselman G, Di rksen C, Hummelshoj L, Bokor  A, B randes I, et al. The burden o f 539 \nendometriosis: c osts and  quality of life  of women with endometriosis  and trea ted in referral 540 \ncentres. Hum Reprod Ox f Engl. 2012 M ay;27(5):1292–9 .  541 \n10.  Penrod N, Okeh C, Velez Edwards DR, B arnhart K, Senapati S, Verma S S. Leveraging electronic 542 \nhealt h record data fo r endometrio sis r esearch. Front Digit Health [Internet]. 2023 Jun 5 [cited 543 \n2024 Apr 12];5. Available from:  https://www.fr ontiersin.org/j ournals/digital-544 \nhealt h/articl es/10.3 389/fdgth. 2023 .1 150687/ful l  545 \n11.  Soliman AM, Fuldeore M, Snabes MC. F actors Ass ociated with Time to Endom etriosis Diagnosi s 546 \nin the United States. J Womens Health. 2017 Jul;26(7):788 –97 .  547 \n12.  Becker CM, B okor  A, Heikinheimo O, H orne A, Jansen F, Kiesel L, et al. ESHRE guideline: 548 \nendometriosis†. Hum Reprod Open. 2022 Jan 1;2022(2):hoac009.  549 \n13.  Wykes CB, Clark TJ, Khan KS. Accuracy of laparo scopy in  the diagnosis  o f endo metriosis: a 550 \nsystematic quantita tive review. BJOG Int J Obstet Gynaecol. 2004 Nov;111(11):1204–12.  551 \n14.  Nisolle M, Paindav eine B, Bourdon  A, B erlière M, Casanas-Roux  F, Don nez J. Histologic study o f 552 \nperitoneal endometriosis in  infertile w omen. Fertil Steril. 1990 Jun;5 3(6):984–8.  553 \n15.  Fauconnier A, Fritel X, Chapron C. [End ometriosis and  pelvic pain: epidemiological evidence of 554 \nthe relationship and implications]. Gynecol Obstet Fertil. 2009 Jan;37(1):57– 6 9.  555 \n16.  Sapkota Y, Steinthorsdottir V, Mor ris A P, Fassbender A, Rahmioglu N, De Vivo I, et al. Me ta-556 \nanalysis identifies five n ovel loci associ ated with endometriosis highlighting key genes involved 557 \nin hormone  metabolism. Nat Commun. 2017 May 24;8(1):15539.  558 \n17.  Rahmioglu N, Mortlock S, Ghiasi M, Møller PL, Stefansdottir L, Galarneau G, et al. The genetic 559 \nbasis of  endometriosis  and com orbidit y with other pain and in flammatory co nditions. Nat 560 \nGenet. 2023 Mar;55(3):423–36 .  561 \n18.  Saha R, Pettersson HJ, Svedberg P, Olov sson M, Bergqvist A, Marions  L, et al. Heritability of 562 \nendometriosis. Fertil Steril. 2015 Oct 1;104(4):947–52 .  563 \n19.  Lee SH, Harold D, Nyholt DR, Goddard  ME, Zondervan KT, Williams J, et al. Estimation and 564 \npartitioning o f polygenic variation  captured by common  SNPs for  Alzheimer’s disease, multip le 565 \nsclerosis and  endometriosis. Hum Mol Genet. 2013 Feb 15;22(4):832–41 .  566 \n20.  Brawn J, Mo rotti M, Zondervan KT, Bec ker CM, Vincent K. Central changes  ass ociated with 567 \nchronic pelvic pain and endometriosi s. Hum Reprod Update. 2014;20(5):737 – 47.  568 \n21.  Nezha t C, Li A, Abed S, Balassiano E, Sol iemannjad R, Nezha t A, et al. Strong As sociation 569 \nBetween Endometriosis  and Symptom atic Leiomyomas. JSLS. 2016;20(3):e20 16.0005 3.  570 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n29 \n22.  Verma A, Damrauer SM, Naseer N, Weaver J, Kripke CM, Guare L, et al. The Penn Medicine 571 \nBioBank:  To wards a  Genomics -Enabled Learning Healthcare System to Accelerate Precision 572 \nMedicine in a Diverse P opulation. J Pers Med. 2022 Dec;12(12):1974.  573 \n23.  Gottesman O, Kuivaniemi H, Tromp G, Faucet t WA, Li R, Manolio TA, et al. The Electronic 574 \nMedical Records and Genomics  (eMER GE) Network:  past, present, and future. Genet Med Of f J 575 \nAm Coll Med Genet. 2013 Oct;15(10):761–71 .  576 \n24.  GWAS and  enrichment analyses of  non -alcoholic fatty liver disease identify new trait-associated 577 \ngenes and pathways acros s eMERGE  Network |  Sprin gerLink [Internet]. [cited  2023 Aug 14]. 578 \nAvailable from: https://link.springer.com/article/10 .1186/s12 916-019-136 4-z 579 \n25.  Data Bro wser | All o f Us  Public Data Br owser  [Internet]. [cited 2024 Mar 30]. Available from: 580 \nhttps://databrowser.researchallofus.org/ 581 \n26.  Genomic data in the All of  Us Research Program. Nature. 2024;62 7(8003):340 –6.  582 \n27.  The “All of Us” Research Pr ogram. N En gl J Med. 2019 Aug 15;381(7):668–76 .  583 \n28.  Sync For  Science [Internet]. [cited 202 4 Mar 30]. Available from:  https://syncfor.science/  584 \n29.  Bycroft C, Freeman C, Petkova D, Band G, Elliott LT, Sharp K, et al. The UK Bi ob ank resource 585 \nwith deep phenotyping and genomic d ata. Na ture. 2018 Oct; 562(7726):203 – 9.  586 \n30.  Hallinan CM, Ward R, Hart GK, Sullivan C, Pratt N, Ng AP, et al . Seamless EMR data access: 587 \nIntegrated governance, digital hea lth a nd the OMOP-CDM. BMJ Health Care Inform. 2024 Feb 588 \n21;31(1):e100953 .  589 \n31.  An integrated map of  genetic variation from 1,092 human genomes. Nature. 2012 Nov 590 \n1;491(7422):56–6 5.  591 \n32.  Jia H, Ding S, Xu X, Nie R. The latest res earch progress  on  spectral clustering. Neural Comput 592 \nAppl. 2014 Jun 1;24(7):1477– 86.  593 \n33.  Damle A, Minden V, Ying L. Simple, direct and efficient multi-way spectral clustering. Inf 594 \nInference J IMA. 2019 Mar 15;8(1):181–203.  595 \n34.  Khan K, Rehman SU, Aziz K, Fon g S, Sar asvady S. DBS CAN: Past, present and fu ture. In: The Fifth 596 \nInternational Conference on the Applications o f Digital Information  and Web  Technologies 597 \n(ICADIWT  2014) [Internet]. 2014 [cited 2024 Mar 30]. p. 232–8. Avail able fro m: 598 \nhttps://ieeexplore.ie ee.org/abstract/d ocument/68146 87  599 \n35.  Algorithms fo r hierarchical clustering: an overview -  Murtagh - 2012 -  WIREs  Data Mining and 600 \nKnowledge Di scovery -  Wiley Online Li brary [Internet]. [cited 2024 Mar 30]. Available from: 601 \nhttps://wires.onlinelibrary.wiley .com/ doi/full/10.1 002/widm. 53  602 \n36.  Kanungo T, Mount DM, Netanyahu NS, Piatko CD, Silverman R, Wu AY. An  ef fic ient k-means 603 \nclustering algorithm: analysis and implementation. IEEE Trans Pattern Anal Mach Intell . 2002 604 \nJul;24(7):881 –92.  605 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n30 \n37.  Watts V. 9.5 Statistica l Inference fo r T wo Population Pr oportion s. 2022 Sep 1  [cited 2024 Apr 606 \n16]; Available from:  https://ecampusontario.pressbo oks.pub/introstats/chapter/9-5-607 \nstatistical-inference-for-two -population-propo rtions/  608 \n38.  Chang CC, Chow  CC, Tellier LC, Vattikut i S, Purcell SM, Lee JJ. Second-generation PLINK: rising  to 609 \nthe chal lenge of larger and  richer datas ets. GigaScience. 2015 Dec 1;4(1):s13742-015-0047– 8.  610 \n39.  Islam MJ, Wu QMJ, Ahmadi M , Sid-Ahmed MA. Investigating the Performance of Naive- Bayes  611 \nClassifiers  and K - Nearest Neighbor  Classifiers. In:  2007 International Con ference on 612 \nConvergence Inf ormation Technology ( ICCIT 2007) [Internet]. 2007 [cited 2024 Apr 16]. p. 613 \n1541–6 . Available from: https://ieeexplore.ieee.org/abstract/document/ 442 0473  614 \n40.  Zhou W, Nielsen JB, Fritsche LG, Dey R, Gabrielsen ME, W olford  BN, et al. Effici ently controlling 615 \nfor  case-contr ol imbalance and sample relatedness in large-scale genetic assoc iation studies. 616 \nNat Genet. 2018 Sep;50(9):1335 –41.  617 \n41.  Purcell S, Nea le B, Todd-Br own K, Tho mas L, Ferreira MAR, Bender D, et al. PLINK: A To ol Set 618 \nfor  Whole-Genome  Ass ociation and Po pulation-Based Linkage Analyses. Am J Hum Genet. 2007 619 \nSep;81(3):559–75.  620 \n42.  Borenstein M, Hedges LV, Higgin s JPT, Rothstein HR. A basic intr oduction to  fi xed-effect and 621 \nrandom-effects  models f or  meta-analys is. Res Synth Methods. 2010;1(2):97–1 11.  622 \n43.  Heydarpour M, Parksook  WW, H opkins  PN, Pojoga LH, Williams GH, Williams JS. A candidate 623 \nlocus in the renalase gene and suscepti bility to blood  pressure  respon ses to  the dietary salt. J 624 \nHypertens. 2023 May;41(5):723 .  625 \n44.  Onengut-Gumuscu S, Chen WM, Burren  O, Co oper NJ, Quinlan AR, Mychaleck yj JC, et al. Fine 626 \nmapping of  type 1 diabetes susceptibility loci and evidence for  colocalization of causal variants 627 \nwith lymphoid gene enhancers. Nat Genet. 2015 Apr;47(4):381–6 .  628 \n45.  Saunders GRB, Wang X, Chen F, Jang  SK , Liu M, Wang C, et al. Genetic diversity fuels gene 629 \ndiscovery f or to bacco and  alcohol use. Nature . 2022 Dec;612(7941):720– 4.  630 \n46.  Stilley JAW, Segaloff DL. FSH  Actions a nd Pregnancy: Looki ng Beyond Ovaria n FSH Receptors. 631 \nEndocrinology. 2018 Dec  1;159(12):4033–42 .  632 \n47.  Heidari S, Kolahdouz-Mohammadi R, K hodaverdi S, Tajik N, Delbandi AA. Expr ession levels of  633 \nMCP-1, HGF, and IGF-1 in end ometriotic patients compared with non -endome triotic controls. 634 \nBMC W omens Health. 2021 Dec 20;21(1):422.  635 \n48.  Amirteimouri S, Ashini M, Ramazanali F, Aflatoonian R, Af sharian P, Shahhoseini M. Epigenetic 636 \nrole of  the nuclear factor NF-Y on I D ge ne family in endometrial tissues o f wo men with 637 \nendometriosis: a  case contr ol study. Reprod Bi ol Endoc rinol. 2019 Mar 15;17( 1):32.  638 \n49.  Al-kuraishy HM, Al-Maiahy TJ, Al-Garee b AI, Alexiou A, Papadakis M, Saad HM, et al. The 639 \npossible r ole furin  and furin inhibito rs in endometrial adenocarcinoma: A  nar rative review. 640 \nCancer Rep. 2024;7(1):e1920.  641 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint \n\n31 \n50.  Kho PF, Wang X, Cuéllar-Partida G, Dör k T, Go ode EL, Lambrechts D, et al. Multi-tissue 642 \ntranscriptome-wide ass ociation study i dentifies eight candidate genes and tiss ue-specific gene 643 \nexpression  underlying endometrial cancer susceptibility. Commun Biol. 2021 Oct 21;4(1):1–8.  644 \n 645 \nAll rights reserved. No reuse allowed without permission. \n(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. \nThe copyright holder for this preprintthis version posted April 22, 2024. ; https://doi.org/10.1101/2024.04.22.24306092doi: medRxiv preprint","source_license":"CC0","license_restricted":false}