{"paper_id":"793558a2-ba6f-43ba-af61-10875713c632","body_text":"1 \n \nSite-dependent transcriptomic signatures of endometriosis are conserved 1 \nacross hormonal states 2 \nDouglas E. Henze1, George Crowley1, Stephen R. Quake1,2# 3 \nAffiliations: 4 \n1 Department of Bioengineering, Stanford University, Stanford, CA, USA 5 \n2 Department of Applied Physics, Stanford University, Stanford, CA, USA 6 \n# Correspondence to steve@quake-lab.org 7 \n 8 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n2 \n \nAbstract 9 \nEndometriosis, a chronic condition in which endometrial tissue grows at other sites in the body, 10 \nproduces lesions whose gene expression profiles vary by anatomical location and menstrual 11 \ncycle stage. The extent to which these location-dependent transcriptional differences persist 12 \nacross hormonal states has remained unknown. Here, we compile a comprehensive single-cell 13 \nRNA sequencing atlas of 672,051 cells across 112 donors from publicly available sources 14 \nspanning the eutopic endometrium, menstrual effluent, ovaries, peritoneum, and ectopic 15 \nendometrium from peritoneal and ovarian sites. By training machine learning classifiers on 16 \nreference tissue signatures, we identify a subset of cells within ectopic lesions that retain a 17 \nuterine transcriptomic identity, which we term “core” lesion cells, and which are distinguished 18 \nfrom surrounding host tissue populations. We show that transcriptomic differences between core 19 \nlesion cells at both ectopic sites are robust across all phases of the menstrual cycle and in patients 20 \nreceiving exogenous hormonal therapy, and validate these findings in  independent datasets. Cell 21 \ntype- and tissue-specific gene signatures derived from these core populations are sufficient to 22 \nclassify disease status and lesion subtype in independent bulk tissue data across 168 patients. 23 \nThese findings establish that although the core lesion cells maintain elements of eutopic 24 \nendometrium identity, they also contain information about the anatomical site of the lesion.  This 25 \nanatomic-specificity provides a potential framework for cycle-independent endometriosis 26 \ndiagnosis and subtyping. 27 \n 28 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n3 \n \nMain 29 \nEndometriosis, a chronic inflammatory disease affecting approximately 10% of reproductive-age 30 \nwomen, is characterized by the growth of endometrial tissue outside of the uterus 1,2. The clinical 31 \npresentation is heterogeneous, ranging from dysmenorrhea  (painful periods) and chronic pelvic 32 \npain to infertility, and diagnosis , which requires laparoscopic surgery with histological 33 \nconfirmation, is delayed by an average of 6 -10 years3,4. Over this interval, disease can progress 34 \nfrom minimal to severe stages as classified by the revised American Society for Reproductive 35 \nMedicine (rASRM) system5. Beyond staging, endometriosis can be typed by lesion location, most 36 \ncommonly as ovarian endometriomas or peritoneal lesions, which are thought to represent distinct 37 \npathological entities with divergent structural architectures and potentially distinct pathogeneses6. 38 \nHowever, no adequate non-invasive tools currently exist to distinguish one subtype from the other. 39 \nThe growth and maintenance of ectopic lesions is driven in large part by estrogen signaling and 40 \ndisrupted progesterone responsiveness 7–9, with additional menstrual hormones implicated in 41 \ndisease progression 10,11. These hormonal dependencies have motivated the use of ovulation -42 \nblocking therapies as the primary medical treatment  for endometriosis  and have shaped the 43 \nassumption that hormonal cycling is a major driver of transcriptomic heterogeneity within ectopic 44 \ntissue. Indeed, the eutopic endometrium undergoes dramatic transcriptomic remodeling across the 45 \nmenstrual cycle12,13, and this responsiveness is at least partially preserved in ectopic lesions6,14. To 46 \ninvestigate the cellular and molecular landscape of endometriosis across these hormonal states, 47 \nsingle-cell RNA sequencing (scRNA -seq), a technique that measures gene activity in individual 48 \ncells, has been applied to profile eutopic and ectopic tissues 14–17. These studies have revealed 49 \ntranscriptomic heterogeneity across tissue sites and cell types, yet hormonal variation across the 50 \nmenstrual cycle and between treated and untreated patients remains deeply confounded with 51 \nectopic site, leaving it unclear which variable principally shapes the molecular identity of lesion 52 \nsubtypes.  53 \nHere, we integrate publicly available scRNA -seq datasets spanning the eutopic endometrium, 54 \nmenstrual effluent, ovaries, peritoneum, and ectopic endometrium from both peritoneal and 55 \novarian sites to construct a comprehensive multi -tissue reference atlas. Using cell type -level 56 \nclassifiers trained on reference tissue signatures, we identify the core endometrial-like populations 57 \nwithin each ectopic site and characterize how these populations differ from the surrounding host 58 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n4 \n \ntissue microenvironment. We demonstrate that ectopic site, rather than hormonal state, is the 59 \nprimary axis of transcriptomic variation among core lesion mesenchymal lineage cells, indicating 60 \nthe robustness of potential RNA biomarkers across the menstrual cycle. Finally, we leverage the 61 \ncell type- and tissue-specific signatures of core ectopic cells to stratify independent bulk tissue 62 \nsamples by disease status and lesion location, establishing a molecular framework for non-invasive 63 \nendometriosis diagnosis that is independent of hormonal condition. 64 \nResults 65 \nConstruction of a comprehensive multi-tissue reference atlas for endometriosis 66 \nTo define cell types across endometriosis lesions in the context of their potential lesion sites, we 67 \nassembled an integrated reference atlas comprising 672,051 cells from 112 donors spanning 68 \neutopic endometrium, menstrual effluent, ovaries, peritoneal biopsies, and ectopic endometrium 69 \nfrom both peritoneal (EcP) and ovarian (EcO) sites (Figure 1a,b). The atlas was constructed by 70 \nharmonizing publicly available single -cell RNA sequencing (scRNA -seq) datasets across these 71 \ntissues. Specifically, to represent eutopic and menstruating endometrial tissue, we incorporated the 72 \nHuman Endometrial Cell Atlas (HECA) 17, which provides harmonized metadata across publicly 73 \navailable eutopic endometrium samples, together with menstrual effluent profiles from Shih et al.18 74 \nthat include healthy controls, symptomatic individuals, and endometriosis-positive individuals. To 75 \nrepresent ovarian tissue, we integrated healthy ovarian datasets from Jones et al.19, Tabula Sapiens 76 \n2.020, and Fonseca et al 15. To include ectopic tissue and enable direct comparisons between both 77 \nperitoneal and ovarian ectopic sites without introducing sampling bias, we restricted inclusion to 78 \nstudies that profiled both EcP and EcO lesions . This included Fonseca et al. and Tan et al .16. To 79 \nincorporate a peritoneal tissue reference, we also included the EcP-adjacent samples from Tan et 80 \nal., and the unaffected ovary samples from Fonseca et al. as an additional ovary reference . 81 \nAlthough these datasets originated from different laboratories and collection time points, they 82 \nexhibited comparable read depth and gene detection rates (Supplementary Figure 1a,b), supporting 83 \ntheir suitability for joint integration. 84 \nTo establish unified cell identities across these diverse tissues, we performed large language model 85 \n(LLM)-based de novo cell type annotation21, generating broad cell class labels across all integrated 86 \ndatasets (Figure 1c; Methods). The resulting LLM annotations mapped closely to the expert -87 \ncurated annotations in the HECA ( Supplementary Figure 1c) and aligned well with canonical 88 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n5 \n \nmarker gene expression for each cell type (Figure 1d), validating the unified labels. The most 89 \nnotable discrepancies between the LLM -based and HECA annotations involved T cell 90 \nsubpopulations and red blood cells; the LLM -guided approach classified a greater proportion of 91 \ncells as blood cells, consistent with a conservative interpretation of the strong hemoglobin signal 92 \nobserved broadly across the atlas  in previous studies  (Figure 1d).  Beyond re producing known 93 \nannotations, integration coupled with LLM-based labeling highlighted additional populations and 94 \ntransitional states not emphasized in the source studies. These included myofibroblast-like cells in 95 \nthe eutopic endometrium (Figure 1d, Supplementary Figure 1c) and transitional epithelial cell 96 \nstates co -expressing canonical epithelial and stromal markers. Additionally, we identified a  97 \nmesothelial cell population, which was most abundant in the ectopic peritoneal tissue (Figure 1c, 98 \nSupplementary Figure 1d). 99 \nTaken together, atlas integration substantially reduced batch effects across tissues, datasets, and 100 \ndisease states (Supplementary Figure 1d-f) while yielding unified cell type annotations spanning 101 \nall sampled compartments. This comprehensive multi-tissue reference atlas provides a foundation 102 \nfor defining tissue-specific transcriptomic signatures and for deconvolving cell type-level signals 103 \nin endometriosis. 104 \nSomatic mutations link ectopic lesions to eutopic endometrium 105 \nThe prevailing hypothesis for the pathogenesis of endometriosis is that retrograde menstruation 106 \nleads to endometrial tissue implant ing within the peritoneal space 22. If ectopic lesions originate 107 \nfrom displaced eutopic endometrium, then both tissues should share a common mutational history. 108 \nTo assess this, we identified 1,064 high -confidence somatic variants across paired eutopic (300 109 \nvariants) and ectopic (764 variants) endometrium samples in the Fonseca et al. dataset (Methods) 110 \n(Supplementary Figure 2a). We first compared the mutational processes operating in each tissue 111 \nby analyzing their mutation spectra. The 6-class mutation spectrum revealed similar proportional 112 \ncontributions in both tissues ( Supplementary Figure 2a), and the 96 -class trinucleotide spectrum 113 \nconfirmed broad similarity between the two sites (r = 0.87) ( Supplementary Figure 2 b). Both 114 \nspectra exhibited a prominent SBS5 signature, characterized by frequent C>T and T>C mutations, 115 \nconsistent with prior observations across eutopic and ectopic endometrium23,24. 116 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n6 \n \nHaving established that both tissues undergo similar mutagenic processes, we next asked whether 117 \nthe identified somatic mutations were subject to selection. We computed the ratio of 118 \nnonsynonymous to synonymous mutations (dN/dS) and evaluated whether the observed ratio 119 \nexceeded random expectations by generating null distributions across 100,000 permutations. In 120 \neach permutation, an alternative base was randomly assigned , weighted by the observed 121 \ntrinucleotide spectrum, at the sequence-specific context of each coding variant position (Methods). 122 \nThe observed dN/dS was indistinguishable from the null distribution ( Supplementary Figure 2c), 123 \nindicating that somatic mutations accumulate without detectable genome -wide selection for or 124 \nagainst protein-altering changes. Together with the shared mutagenic signatures, this suggests that 125 \nboth tissues harbor a set of passenger mutations amenable to lineage tracing. 126 \nWe therefore asked whether any somatic variants were shared between paired eutopic and ectopic 127 \nsamples, and identified sets of shared mutations in two patients ( Supplementary Figure 2d). To 128 \nassess whether these shared variants reflected clonal relationships, we analyzed the clonal cell 129 \nfraction, defined as the proportion of cells confidently carrying each mutation (Methods). In both 130 \npatients, the shared mutations were present in comparably small fractions of cells at both sites 131 \n(Supplementary Figure 2e), consistent with an oligoclonal structure in which the ectopic lesion 132 \ndoes not arise from a single massively expanded clone. To determine whether any clonal expansion 133 \nnonetheless occurred at ectopic sites, we compared variant allele frequencies (VAFs) of all called 134 \nmutations across paired eutopic and ectopic samples in all four patients ( Supplementary Figure 135 \n2f). Across four of five paired samples, ectopic sites exhibited modestly elevated VAFs, suggesting 136 \na degree of clonal expansion following the initial seeding event. 137 \nFinally, to understand whether somatic mutations accumulate differentially across cell types, we 138 \nmapped called variants back to their cell type of origin (Supplementary Figure 2g). Consistent with 139 \nprevious results25, the stromal and fibroblast fraction carry variant alleles at a relatively abundant 140 \nrate. This result is not consistent with prior bulk observations that epithelial cells harbor the most 141 \nmutations, likely owing to lower cell numbers that led to reduced sequencing depth in the bulk 142 \nanalysis. Notably, the distribution of mutations between eutopic and ectopic endometrium varied 143 \nsubstantially across cell types, motivating a cell-type-resolved analysis of the ectopic lesion.  144 \nCell type-level classifiers distinguish endometrial -like from host tissue cells within ectopic 145 \nlesions 146 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n7 \n \nDuring atlas integration, we observed that certain ectopic endometrium stromal cell populations 147 \nco-clustered with ovarian stromal cells (Supplementary Figure 1d), raising the question of whether 148 \nan underlying transcriptomic signature for each reference tissue could be leveraged to classify cell 149 \ntypes within the ectopic endometrium as endometrial  derived or host -tissue derived with high 150 \nconfidence (Figure 1a). To test this, we trained logistic regression classifiers on differentially 151 \nexpressed, highly cell type-specific genes to distinguish endometrial (eutopic/menstrual effluent) 152 \nfrom non-endometrial reference (ovary/peritoneum) cell types (Methods). The classifiers achieved 153 \nstrong performance across the most abundant cell types (Figure 2a) and accurately assigned the 154 \nmajority of cells as endometrial or non -endometrial in origin ( Supplementary Figure 3a), with a 155 \npooled false discovery rate of 1.5% across all cell types . Notably, the feature genes driving the 156 \nmajority of our cell type-level classifiers were enriched for endometrium-specific genes as defined 157 \nby RNA expression  in the Human Protein Atlas (HPA) 26 (Supplementary Figure 3b), further 158 \nvalidating the biological relevance of the discriminative signal. 159 \nGiven that endometriosis is defined by the growth of endometrial-like tissue outside the uterus, we 160 \nnext sought to determine whether cells within resected ectopic sites exhibited an endometrial 161 \ntranscriptomic identity. Applying our classifiers across the entire dataset ( Supplementary Figure 162 \n3c), we found that the majority of cells at ectopic sites were transcriptomically more similar to 163 \ntheir host tissue of residence than to the endometrium. The proportion of lesion cells classified as 164 \nendometrial-like, hereafter referred to as \"core\" lesion cells , varied by lineage: endothelial cells 165 \nshowed the lowest fraction (5%), followed by mesenchymal (8%), myeloid (11%), epithelial 166 \n(15%), and lymphoid (33%) lineage cells. These modest proportions are consistent with prior 167 \nobservations that stromal cells of endometrial origin constitute a minority of endometrioma tissue14 168 \nand that immune populations within lesions represent an admixture of host - and endometrium-169 \nderived cells 27. Comparing across ectopic sites, we observed broadly comparable rates of 170 \nendometrial-like classification, with the most notable divergence occurring among immune 171 \nlineages, where EcO myeloid cells and EcP lymphoid cells  were more frequently classified as 172 \nendometrial-like ( Supplementary Figure 3 d,e). The relative enrichment of endometrial -like 173 \nmyeloid cells in EcO is consistent with previous observations that the predominant macrophage 174 \npopulation in endometriomas exhibits a tissue -resident phenotype that is also enriched in the 175 \neutopic endometrium of patients with endometriosis16. 176 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n8 \n \nTo test whether core populations faithfully recapitulate the eutopic endometrial microenvironment, 177 \nwe performed differential cell -cell interaction analysis 28 comparing the eutopic endometrium 178 \nacross all endometriosis patients against the core cells of both the EcO and the EcP sites (Methods). 179 \nIn the EcO, core cells retained conserved interactions among mesenchymal populations relative to 180 \nnon-core cells ( Supplementary Figure 4a), suggesting preservation of structural cell 181 \ncommunication networks. The EcP maintained this mesenchymal interaction, but it also displayed 182 \na greater number of enriched core interactions involving epithelial populations, more closely 183 \nresembling the eutopic endometrium than the EcO ( Supplementary Figure 4b). This divergence 184 \nlikely reflects the altered structural dynamics between EcO and EcP, as well as the reduced 185 \nprevalence of epithelial tissue within endometrioma samples 14–16. Consistent with prior reports , 186 \ncore EcP cells also exhibited well-conserved epithelial-to-stromal crosstalk29. 187 \nTo directly assess transcriptomic similarity to eutopic endometrium, we performed  differential 188 \nexpression analysis between diseased eutopic endometrium and both core and non -core 189 \npopulations within each ectopic site separately (Supplementary Figure 5a,b; Methods). Across cell 190 \ntypes, core populations consistently exhibited fewer differentially expressed genes (DEGs) relative 191 \nto diseased eutopic endometrium than non-core populations at the same site (Figure 2c), suggesting 192 \nthat core cells are transcriptomically closer to endometrial tissue. We also observed a higher DEG 193 \nburden per cell type in EcO than in EcP core cells, suggesting that endometrial-like populations in 194 \nendometrioma may be more divergent from diseased eutopic endometrium than those in peritoneal 195 \nlesions.  196 \nTo interpret the biological basis of these differences, we partitioned DEGs for each cell type into 197 \ncore-specific, non-core-specific, and shared gene programs between EcO and EcP. We focus on 198 \nmacrophages, which exhibited amongst the largest DEG burdens at both sites (Supplementary 199 \nFigure 5a,b), to illustrate the functional differences between ectopic sites . In EcO, non -core-200 \nspecific macrophage DEGs were enriched for cell migration pathways (Supplementary Figure 5c), 201 \nconsistent with evidence that macrophages can traffic from host tissue into endometrial lesions 30. 202 \nThe core-specific and shared programs in EcO macrophages highlighted cytokine response and 203 \npositive regulation of smooth muscle proliferation , with the core -specific program additionally 204 \nincluding activation pathways (Supplementary Figure 5d,e), aligning with the inflammatory milieu 205 \nof endometriosis and macrophage contributions to lesion vascularization 31,32. In EcP, non -core-206 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n9 \n \nspecific genes reflected a macrophage expansion response similar to that in EcO, but with the 207 \naddition of  local proliferation pathways rather than solely migration, and with a stronger 208 \nextracellular matrix organizing component, consistent with prior results33 (Supplementary Figure 209 \n5f). EcP shared and core -specific macrophage signatures showed enrichment for golgi lumen 210 \nacidification and positive regulation of cell death , respectively (Supplementary Figure 5g,h), the 211 \nlatter raising the possibility of macrophage-associated dysfunction within the lesions. 212 \nHaving established that core -specific DEGs capture distinct biology characterizing ectopic cells 213 \nrelative to eutopic tissue, we next asked how conserved these transcriptomic programs were 214 \nbetween lesion sites. Strikingly, Jaccard indices between EcO and EcP core-specific gene sets were 215 \nrelatively low across individual cell types (Figure 2 d), with the greatest site -conserved changes 216 \noccurring within immune populations. Grouping core-specific DEGs by immune lineage revealed 217 \nthat the majority of enriched programs were site - and lineage -specific. Specifically, l ymphoid 218 \nDEG pathways reflected chemotaxis in EcP and response to cytokines  in EcO, while myeloid 219 \npathways reflected differentiation programs in EcP and catabolic processes in EcO (Figure 2f). 220 \nThese site-dependent transcriptomic signatures suggest that the local tissue environment makes a 221 \nsubstantial contribution to the molecular identity of endometrial-like cells. 222 \nMenstrual cycle stage and exogenous hormones modulate ectopic endometrium in a cell type-223 \nspecific manner 224 \nFluctuating estrogen and progesterone levels across the menstrual cycle are widely thought to 225 \nmodulate the function of  the endometrial lesion-resident cell types, analogous to hormone -226 \nresponsive programs in the eutopic endometrium12,17. We asked whether these changes within the 227 \nectopic endometrium could explain the varying degrees of pain experienced throughout the 228 \nmenstrual cycle by endometriosis patients . These patients  are known to experience primary 229 \ndysmenorrhea, which is an acute symptom limited to the perimenstrual window and driven by  230 \nmyometrial hypercontractility, vasoconstriction and ischemic pain 34. In addition to this primary 231 \npain, endometriosis patients experience a secondary dysmenorrhea, which can begin  in the 232 \nsecretory phase of the menstrual cycle 35,36, indicative of a pain mechanism which is entirely 233 \nseparate from the hypercontractility in primary dysmenorrhea37. We first quantified the number of 234 \ngenes varying along the menstrual cycle for each cell type with respect to its compartment of origin 235 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n10 \n \n(Methods). This revealed substantial differences in hormone responsiveness between the eutopic 236 \nendometrium and the ectopic sites for both immune and stromal cell compartments 237 \n(Supplementary Figure 6a). 238 \nGiven that the menstrual cycle comprises two distinct phases (proliferative and secretory) 239 \nassociated with different levels of endometriosis-related pain, and that exogenous hormones 240 \nconstitute an additional hormonal state, we next sought to identify phase -dependent expression 241 \npatterns within each cell type. We calculated mean expression profiles of gene –cell type–tissue 242 \ntriplets across the three hormonal strata and clustered these profiles into six groups with similar 243 \nexpression dynamics (Supplementary Figure 6b; Methods). To identify the underlying gene 244 \nprograms, we annotated the clusters using gene ontology biological processes (Methods), which 245 \nrevealed correspondence with established endometrial biology. Because endometriosis-associated 246 \npain is initiated in the secretory phase35, we focused on the three clusters with expression peaks or 247 \ntroughs in this interval. Cluster 3 , which peaks in the secretory phase, was characterized by the 248 \nresponse to zinc ion, an element known to fluctuate throughout the menstrual cycle and accumulate 249 \nduring the secretory phase 38. Conversely, Cluster 1, which troughs in the secretory stage, was 250 \nenriched for extracellular matrix assembly, consistent with post -menstrual stromal regeneration 251 \nand matrix deposition39, while Cluster 6, which also reached its lowest expression in the secretory 252 \nphase, was defined by antigen presentation and alpha beta T cell activation with simultaneous 253 \nsuppression of NK cell -mediated cytotoxicity, inconsistent with prior observations of immune 254 \ntolerance in the secretory state40,41. This failure to adopt the tolerogenic immune state that normally 255 \ncharacterizes the secretory phase 41 may reflect a blunted progesterone response, a hallmark of 256 \nendometriosis driven by loss of the stimulatory PR-B isoform42,43. 257 \nWe partitioned the cycle-responsive genes in the peaking and troughing groups by their tissue and 258 \ncell type of origin (Supplementary Figure 6c , d). The greatest number of cycle -responsive genes 259 \noccurred in mesenchymal populations, including stromal cells, fibroblasts, and vascular smooth 260 \nmuscle cells. Notably, across all cell types, these genes exhibited site -specific signatures, 261 \nindicating that eutopic and ectopic endometrium diverge in their hormonal response. 262 \nTo characterize the functional implications of secretory -phase dynamics restricted to ectopic 263 \ntissue, we took the union of EcO and EcP genes across cell types. The secretory -peak programs 264 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n11 \n \nwere most strongly enriched for macrophage immune response and fibroblast ECM development 265 \n(Supplementary Figure 6e), consistent with the heightened inflammatory and fibrotic states 266 \nreported in endometriotic lesions33,44. Programs troughing in the secretory state further reinforced 267 \nan inflammatory signature, with macrophages displaying elevated cytokine production 268 \n(Supplementary Figure 6f). Unexpectedly, NK cells exhibited increased cytotoxicity, diverging 269 \nfrom their canonical behavior during the secretory phase45 (Supplementary Figure 6f), reinforcing 270 \nthe position that the ectopic immune compartment becomes inflammatory at the initiation of 271 \nsecondary dysmenorrhea 37. In contrast, monocytes contributed an anti -inflammatory profile, 272 \nlargely aligned with prior evidence that macrophage ontogeny and recruitment status give rise to 273 \nfunctionally divergent populations within endometriotic lesions 27,46. The convergence of these 274 \ninflammatory and cytotoxic profiles within fibrotic lesions provides a molecular framework for 275 \nthe secondary dysmenorrhea of endometriosis that is mechanistically distinct from primary. In 276 \nplace of an acute myometrial contraction, the lesion ’s immune cells sustain a hormonally driven 277 \ninflammatory state linking hormonal fluctuations to the modulation of pain across the menstrual 278 \ncycle in endometriosis patients. 279 \nEctopic site modulates cell type -specific transcriptomic programs independently of 280 \nhormonal state 281 \nWhile our analysis shows that ectopic endometrial cells are hormone responsive, it also  revealed 282 \npronounced, site-specific transcriptomic divergences from the eutopic endometrium among core 283 \nectopic cell types. We therefore asked which effect was larger: whether the transc riptomic 284 \nvariation in the ectopic endometrium primarily reflects changes in hormonal state or whether the 285 \nectopic site itself imposes an independent and dominant effect on cell state (Figure 3a). 286 \nTo disentangle these two variables, we computed Spearman correlations between the mean batch-287 \ncorrected principal component  coordinates of each core cell type under two conditions: first, 288 \nbetween EcO and EcP cells in matched hormonal states, and second, between naturally cycling 289 \ncells and those from patients receiving exogenous hormonal treatment at matched sites. A greater 290 \nnumber of cell types exhibited higher correlation between different hormonal states than between 291 \ndifferent sites (Figure 3b), indicating that ectopic site exerts a stronger influence on transcriptomic 292 \nidentity than hormonal status. This effect was not uniform across compartments. The gap between 293 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n12 \n \nsite and hormonal correlations was widest for mesenchymal populations, including stromal cells, 294 \nmyofibroblasts, and vascular smooth muscle cells, whereas immune cell types were comparatively 295 \nstable across both perturbations, with slightly higher correlation between sites than hormonal states 296 \n(Figure 3b). This finding is notable given the longstanding emphasis on interpreting hormonal 297 \ncycling as a primary driver of transcriptomic variation in ectopic endometrial tissue . To 298 \nsystematically characterize this site -dependent heterogeneity, we identified co -expressed gene 299 \nmodules within the same cell types across all donors, ectopic sites, and hormonal states (Methods). 300 \nThis analysis yielded thirty-three gene expression modules, twenty-six of which were significantly 301 \nenriched for at least one Gene Ontology Biological Process term and were retained for annotation 302 \n(Figure 3c; Supplementary Data Table 1; Methods). These modules were variably used across cell 303 \ntypes, with each module showing its highest mean usage in a single cell type (Figure 3c). 304 \nWe next restricted analysis to modules whose usage was specific to a cell type , which yield ed 305 \ntwenty-two modules, and t hree of the cell types had no s pecific module and were assigned the 306 \nmost highly used annotated modu le, giving twenty-five cell type-module pairs spanning twenty-307 \nthree unique modules  (Supplementary Data Table 2; Methods) . We then tested whether module 308 \nusage differed between EcO and EcP independently of hormonal state  (Methods). Eleven of the 309 \ntwenty-five cell type-module pairs showed significant site -dependent deviation, and these were 310 \npredominantly associated with structural and endothelial populations, with only  two immune-311 \nassociated program reaching significance , suggesting  that the structurally important cell types 312 \nwithin the lesions are what drive site dependence (Figure 3d; Supplementary Data Table 3). Gene 313 \nset enrichment analysis of these site-dependent modules converged on several recurring functional 314 \naxes. Endothelial cells preferentially engaged cell migration programs involving CDH5, CLDN5, 315 \nand ROBO4 (Supplementary Data Table 1) , consistent with established roles for endothelial 316 \nangiogenesis47 in endometriosis . Lymphatic endothelial cells showed increased enrichment of 317 \nblood vessel morphogenesis programs in EcP across both the Secretory and Exogenous Hormones 318 \nstates, in line with reports of increased lymphatic vessel density in peritoneal endometriosis48. 319 \nEpithelial lineage cells diverged between ectopic sites through  intermediate filament assembly  320 \nprograms involving KRT19, KRT17, and KRT16 (Supplementary Data Table 1),  genes implicated 321 \nin recognized hallmarks of endometriosis such as  tissue remodeling and epithelial-mesenchymal 322 \ntransition49. Among mesenchymal cell types, extracellular matrix (ECM) organization programs 323 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n13 \n \nincluded genes such as GAS6  (Supplementary Data Table 1), a known regulator of endometriosis-324 \nrelated fibrosis, ECM remodeling, and immunosuppression in other disease contexts 50,51. 325 \nMyofibroblasts showed deviations in pathways involving IGFBP1, IGFBP2, and IGFBP4 326 \n(Supplementary Data Table 1 ), which exhibit differential expression across lesion sites 52 and 327 \nregulate fibrosis53 and stromal cell proliferation54 in endometriosis. Stromal populations displayed 328 \nincreased usage of programs related to cytoplasmic translation and ECM organization, reflecting 329 \nthe extensive fibrosis and remodeling characteristic of different lesion sites6,14. Among the immune 330 \ncell programs with significant site dependency, CD8 T cells were related to the unfolded protein 331 \nresponse, encompassing heat shock proteins indicative of the oxidative stress environment in 332 \nendometriosis55. Collectively, these divergences reveal a previously unappreciated site 333 \ndependency of endometriosis -associated transcriptomic programs spanning structural, epithelial, 334 \nand immune cell compartments. 335 \nAlthough these programs demonstrated site -specific enrichment, their usage varied across 336 \nhormonal states, prompting us to ask whether the site -dependent differences persisted across 337 \ndifferent hormonal milieus. Remarkably, the EcO-to-EcP program deviations across all cell types 338 \nwere conserved regardless of whether individuals were receiving exogenous hormones or 339 \nmenstrual cycle  stage (Supplementary Figure 7a), indicating that the observed site -specific 340 \ndifferences were independent of hormonal state. To validate this finding, we performed differential 341 \nexpression analysis between EcO and EcP mesenchymal lineage core lesion cells from donors 342 \nreceiving exogenous hormones in the Tan et al. dataset  and applied hierarchical clustering of the 343 \ntop differentially expressed genes to the same cell types across a range of hormonal states in the 344 \nentirely held-out Fonseca et al. dataset. Strikingly, EcO and EcP cells clustered apart from one 345 \nanother regardless of hormonal state or cell type ( Supplementary Figure 7b), demonstrating that 346 \nsite-driven transcriptomic differences are robust across the full spectrum of hormonal conditions. 347 \nThe genes driving this separation carry mechanistic relevance in endometriosis. TNXB, MME, and 348 \nRHOB56–58 have each been implicated in disease pathogenesis through ECM remodeling or 349 \nepithelial-mesenchymal transition, while TCF21 and IL32 are recognized regulators of the 350 \ninflammatory phenotype in endometriosis 59,60. ID4 has been identified through genome -wide 351 \nassociation studies (GWASs) as having an increased association with severe relative to mild 352 \nendometriosis61, raising the possibility that these site -dependent signatures may also capture 353 \nvariation in disease severity independent of rASRM staging. 354 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n14 \n \nCell type lineages exhibit transcriptomic responses varying with endometriosis severity 355 \nThe identification of a severity-associated gene within our site-dependent signatures prompted us 356 \nto ask more broadly whether cell type lineages in ectopic lesions exhibit differential transcriptomic 357 \nresponses to increasing disease severity, and whether the differentially expressed genes shared 358 \nacross severity stages differ by lesion site. To address this, we performed differential expression 359 \nanalysis comparing control healthy reference tissue against non -core lesion cells and eutopic 360 \nendometrium against core lesion cells for each lineage, then assessed whether the resulting effect 361 \nsizes scaled with disease severity (Methods). Comparing these effect sizes between mild and 362 \nsevere endometriosis across lineages, we observed differential directionality: vascular and 363 \nepithelial cell populations displayed decreasing effect sizes with greater severity, whereas immune 364 \npopulations showed increasing effect sizes ( Supplementary Figure 8a), suggesting lineage -365 \ndependent transcriptomic remodeling as disease progresses. 366 \nGiven these divergent lineage responses, we next examined whether  the intersections between 367 \ndifferentially expressed genes in mild and severe endometriosis at each lesion site were  distinct. 368 \nThe Ec P sites exhibited a more conserved pan severity signature than the Ec O sites 369 \n(Supplementary Figure 8b, c), indicating less overall transcriptomic difference associated with the 370 \nEcP. Among the shared differentially expressed genes, both sites displayed similar proportions of 371 \ncell type-specific and tissue-specific signatures, comprising 30 cell type-specific genes identified 372 \nas cell type -specific in Tabula Sapiens 62 alongside 14 genes with endometrium, ovary, or 373 \nperitoneum specificity as annotated by the Human Protein Atlas 63 (Supplementary Figure 8b,c; 374 \nMethods). This co-occurrence of both cell type - and tissue-specific genes within severity -shared 375 \nsignatures raised the question of whether such signatures carry diagnostic utility beyond our single-376 \ncell framework. 377 \nCell type- and tissue-specific gene signatures are sufficient to stratify endometriosis 378 \nsubtypes 379 \nTo assess the diagnostic potential of the cell type- and tissue-specific signatures identified through 380 \nthe severity analysis, we turned to independent bulk tissue microarray data spanning eutopic 381 \nendometrium and peritoneum from both endometriosis patients and control individuals, as well as 382 \nbulk lesion samples from deep infiltrating lesions, superficial peritoneal lesions, and 383 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n15 \n \nendometriomas64 (Figure 4a; Supplementary Figure 8a,b). Projecting the bulk expression of both 384 \nthe cell type - and tissue -gene sets into PCA space revealed separation among tissue types 385 \n(Supplementary Figure 9a,b), and we hypothesized that a combined signature of both cell type - 386 \nand tissue-specific genes would further improve stratification (Figure 4a). 387 \nWe therefore tested whether these signatures could classify disease status and lesion origin across 388 \nthree clinically relevant comparisons (Methods). We first created a Random Forest classifier which 389 \ndistinguishes between healthy and diseased  eutopic endometrium. For this classifier , the tissue -390 \nspecific genes , which included very few endometrium -specific markers , yielded a statistically 391 \ninsignificant AUC of 0. 592 [0.492-0.691], whereas the cell type -specific signature achieved an 392 \nAUC of 0.821 [0.745-0.883] (Figure 4 b), demonstrating that cell type -specific signatures are 393 \nsufficient to deconvolve endometriosis status in the eutopic endometrium. A similar hierarchy 394 \nemerged for a classifier built to identify control versus diseased peritoneal tissue, where the tissue-395 \nspecific classifier achieved a  modest AUC of 0. 696 [0.564-0.826] and the cell type -specific 396 \nclassifier reached 0.829 [0.702-0.930] (Figure 4 c). T he improved tissue -specific performance 397 \nrelative to the endometrium comparison suggests a stronger contribution of tissue-level signatures 398 \nin distinguishing diseased peritoneum from control peritoneum. We then created a third classifier 399 \nfor lesion subtype classification , where we identified if tissue came from a deep infiltrating 400 \nperitoneal lesion, a superficial peritoneal lesion or an endometrioma , the cell type-specific (AUC 401 \n= 0.904 [0.855-0.944]) performed worse than the combined tissue and cell type signature (AUC = 402 \n0.936 [0.891-0.970]) classifier (Figure 4d), reflecting an increased contribution of tissue -specific 403 \ngenes in this context. Consistent with this, the tissue -specific signature alone was most effective 404 \nin classifying lesion type of origin (AUC = 0. 908 [0.863 -0.944]) (Figure 4d), further reinforcing 405 \nthe presence of an underlying tissue -of-residence signature distinguishing lesion subtypes. 406 \nNotably, combining cell type - and tissue -specific signatures into a unified gene set yielded 407 \nminimal gains in classification accuracy  compared to just the cell -type specific classifier , 408 \nconsistent with the possibility that cell type -specific genes already encode sufficient tissue -level 409 \ninformation to distinguish disease states and lesion subtypes. 410 \nSeveral genes driving these classifiers have established biological relevance in endometriosis. 411 \nAcross both healthy compared to patient tissue- and cell-type-specific classifiers, the stromal cell 412 \ngenes CLDN11, previously proposed as an endometriosis biomarker65, and SERPINE2, a stromal-413 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n16 \n \nspecific gene, were among the highest -contributing features (Supplementary Figure 9a-c). Both 414 \ngenes showed a stepwise increase in expression from control to patient endometrium and 415 \nperitoneum, and from these tissues to lesions (Figure 4 e), with both genes reaching their highest 416 \nlevels in endometrioma, consistent with the dense stromal compartment characteristic of this lesion 417 \nsubtype. In the eutopic endometrium classifier, the top -contributing genes included  immune-418 \nspecific  CD2, GZMH, and IL7R (Supplementary Figure 10a), consistent with prior evidence that 419 \nan enhanced inflammatory state in the eutopic endometrium may serve as an indicator of 420 \nendometriosis66. The peritoneum classifier also showed high importance of immune cell-specific 421 \ngenes including RUNX3 and CCL4L2 (Supplementary Figure 10b), reflecting the established 422 \ninflammatory environment in the peritoneum of endometriosis patients67,68. In the lesion classifier, 423 \nstromal cell-specific genes such as SERPINE2 and CDLN11 (Figure 4e; Supplementary Figure 424 \n10c) were among the strongest contributors, consistent with the site -dependent mesenchymal 425 \nsignatures identified in our earlier analyses. Notably, the lesion classifier assigned high importance 426 \nto EMX2, an endometrium -specific gene, with less or negligible importance for  some ovary-427 \nspecific genes ( GREB1 and OGN respectively) and the peritoneum -specific gene CFD 428 \n(Supplementary Figure 10c), suggesting that endometrial identity is one of the primary 429 \ndiscriminative features for lesion classification. 430 \nAltogether, these results demonstrate that cell type - and tissue-specific transcriptomic signatures 431 \nderived from the integrated single-cell atlas are sufficient to classify diseased from control tissue 432 \nin both the endometrium and peritoneum, while also capturing the spatial signatures that 433 \ndistinguish lesion subtypes , establishing a molecular framework with potential translational 434 \napplications for endometriosis diagnosis and subtyping. 435 \nDiscussion 436 \nEndometriosis affects roughly one in ten women of reproductive age worldwide, yet after nearly a 437 \ncentury of research there are still no adequate tools for non -invasive diagnosis 1,2. This delay is 438 \ndriven in large part by the persistent failure to identify robust non-invasive biomarkers at the blood, 439 \nurinary, or menstrual effluent level69. A fundamental barrier to biomarker discovery is the cellular 440 \nheterogeneity of ectopic lesions, which comprise an admixture of endometrial -derived cells and 441 \nhost tissue constituents whose signals confound bulk molecular analyses. To address this, we 442 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n17 \n \ncompiled, across multiple data sources, a comprehensive multi-tissue single-cell atlas, across 112 443 \ndonors, of 672,051 cells across the eutopic endometrium, menstrual effluent, ovaries, peritoneum, 444 \nand ectopic endometrium from both peritoneal and ovarian sites. This enabled, for the first time, a 445 \nsystematic deconvolution of endometrial -like transcriptomic signatures  from the surrounding 446 \nmicroenvironment across multiple tissues, hormonal states, and disease severities. 447 \nA central finding of this study is that the majority of cells within resected ectopic lesions are 448 \ntranscriptomically more similar to their host tissue of residence than to the eutopic endometrium. 449 \nThis observation carries important implications for how single -cell studies of endometriosis are 450 \ninterpreted. Previous comparisons of transcriptomic profiles between ectopic sites, without first 451 \ndeconvolving the fraction of cells most similar to host tissue, are, in effect, comparisons between 452 \ntwo different tissue microenvironments rather than between the most disease-relevant endometrial-453 \nlike populations. By training cell type-level classifiers on reference tissue signatures and applying 454 \nthem across the ectopic endometrium, we were able to identify the subset of “core” lesion cells 455 \nwhich most closely resemble eutopic endometrial cells and to characterize them separately from 456 \nthe host tissue compartment. This analytical framework is conceptually analogous to methods 457 \ndeveloped for resolving cell -type location in single -cell data from  spatially dependent gene 458 \nexpression70,71, but applied here entirely within a single -cell context to resolve tissue -of-origin 459 \nidentity. 460 \nThe identification of core endometrial -like cells at both ectopic sites is broadly consistent with 461 \nSampson's retrograde menstruation hypothesis 22, which posits that shed endometrial tissue 462 \nimplants at ectopic locations  to cause endometriosis . Recent genomic evidence has further 463 \nsupported this model through the identification of shared somatic mutations between eutopic 464 \nendometrium and matched ectopic lesions 72,73. Our own analysis of paired eutopic and ectopic 465 \nsamples from the Fonseca et al. dataset similarly identified shared variants between matched 466 \ntissues, alongside highly concordant mutational spectra dominated by the SBS5 signature 467 \nindicative of common mutagenic processes.  Together, these shared mutations and spectral 468 \nconcordance provide genomic evidence that cells in ectopic lesions descend from uterine 469 \nprecursors, and the observation that variant allele frequencies are elevated in ectopic tissue relative 470 \nto eutopic endometrium suggests a modest clonal expansion at ectopic sites. The small clonal cell 471 \nfractions of shared mutations additionally point to an oligoclonal lesion structure, likely reflecting 472 \nboth intrinsic oligoclonality of the seeded population and substantial contributions from the 473 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n18 \n \nsurrounding tissue microenvironment.  Our transcriptomic data complement this genomic 474 \nevidence: the retention of a uterine gene expression identity in core lesion cells, despite their 475 \nresidence in a foreign tissue microenvironment , is precisely what the retrograde menstruation 476 \nmodel predicts. At the same time, the fact that the majority of cells at ectopic sites resemble their 477 \nhost tissue rather than the endometrium indicates that lesion establishment involves extensive 478 \nremodeling and recruitment of local cell populations, consistent with a model in which retrograde-479 \nseeded endometrial cells serve as a nidus that co-opts the surrounding tissue environment. 480 \nOur classifier -based approach provides a complementary transcriptomic perspective, 481 \ndemonstrating that the endometrial-like fraction within resected tissue is modest, ranging from 5% 482 \nin endothelial cells to 33% in lymphoid cells, and that the non -core compartment carries distinct 483 \ntranscriptomic programs. Notably, we observed differential immune cell distributions between the 484 \ncore and non-core compartments across ectopic sites: non-core macrophages at ovarian sites were 485 \nenriched for programs related to cell migration, while those at peritoneal sites showed greater 486 \nenrichment for proliferative programs. The migration signature in ovarian non -core macrophages 487 \nis consistent with evidence that endometriosis triggers continuous monocyte recruitment to lesion 488 \nsites, with macrophages of distinct origins exerting opposing effects on lesion establishment74. We 489 \nalso observed differences in cell -cell communication preservation between the eutopic 490 \nendometrium and the two ectopic sites, likely reflecting the differences in macro-level structural 491 \narchitectures that characterize ovarian endometriomas, which form enclosed pseudocysts through 492 \ninvagination of the ovarian cortex, versus  those in peritoneal implants6,75. Taken together, these 493 \nresults demonstrate that core endometrial -like cells carry known transcriptomic signatures of 494 \nendometriosis and provide insight into how the two lesion microenvironments diverge through 495 \ndisease progression. 496 \nThe eutopic endometrium undergoes dramatic transcriptomic remodeling across the menstrual 497 \ncycle12,13,17, and ectopic lesions have been shown to retain some degree of hormonal 498 \nresponsiveness6,8,14. Accordingly, hormonal variation has long been considered a primary driver 499 \nof transcriptomic heterogeneity in endometriotic tissue, and this assumption has influenced both 500 \nexperimental design and therapeutic development. Our results refine this assumption , 501 \ndemonstrating that the mesenchymal cell types are more transcriptomically similar across 502 \nhormonal states than across ectopic sites and that hierarchical clustering of differentially expressed 503 \ngenes in an entirely held -out dataset separates EcO and EcP cells from one another regardless of 504 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n19 \n \nhormonal condition. Critically, however, this site dominance is itself cell -type-dependent and it 505 \nholds most strongly for the mesenchymal and structural populations, whereas immune cell types 506 \nretain greater hormonal responsiveness and modulate their functional state with cycle phase . The 507 \nstability of these site -dependent signatures echoes observations from other biological contexts 508 \ndemonstrating that the local tissue microenvironment is the dominant determinant of cell state, as 509 \nhas been established for specific tissue-resident macrophages76,77 and cross-tissue populations of 510 \nother cell types more broadly20. Our findings extend this principle to endometriosis, suggesting 511 \nthat the ectopic tissue niche imprints a durable transcriptomic identity  on endometrial -like 512 \nmesenchymal cells that supersedes hormonal cycling, even as the immune compartment continues 513 \nto track hormonal state. This compartment split reconciles the dominance of site in subtype 514 \ndiscrimination with the hormonally tuned immune programs relevant to cyclical symptom 515 \nvariation. 516 \nThis hormone-independent site specificity has direct implications for biomarker development. A 517 \nlongstanding challenge in endometriosis diagnostics has been the confounding effect of menstrual 518 \ncycle phase on candidate biomarker performance, necessitating cycle-matched sampling that limits 519 \nclinical scalability. Our observation that mesenchymal gene signatures for discriminating between 520 \nsubtypes of endometriosis are conserved across hormonal states suggests that robust biomarkers 521 \ncan, in principle, be identified for endometriosis subtyping without requiring cycle -phase 522 \nstratification. We demonstrated this by showing that cell type-specific signatures derived from all 523 \ncell types in combined single-cell atlases are sufficient to classify disease status in independent 524 \nbulk tissue microarray data for both the eutopic endometrium and the peritoneum, while tissue -525 \nspecific signatures can only effectively distinguish lesion subtypes. These complementary scales 526 \nof molecular identity may correspond to distinct diagnostic modalities.  527 \nThe plasma cell -free transcriptome has been shown to represent a linear combination of tissue -528 \nspecific RNA contributions 62,78, with recent work demonstrating that cell -type-of-origin can be 529 \nresolved from circulating cell -free RNA at single -cell resolution 62. High tissue -specific 530 \nclassification accuracy for lesion subtype in our analyses suggests that peripheral blood cell -free 531 \nmRNA could, in principle, be leveraged to identify and classify endometriosis subtypes non -532 \ninvasively, as has been demonstrated for cancer detection and subtyping 79. In parallel, 533 \nendometriosis, by virtue of its transcriptomic signatures in the eutopic endometrium , is uniquely 534 \namenable to diagnostic approaches based on menstrual effluent, a non -invasive medium that has 535 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n20 \n \nshown promise for endometriosis biomarker discovery 18,80,81. Given that the cell type -specific 536 \nsignatures are sufficient to distinguish diseased from healthy bulk eutopic endometrium, the cell 537 \ntype-specific fraction of cell -free mRNA within menstrual effluent represents a potential avenue 538 \nfor future diagnostic development. 539 \nIn summary, this study provides a multi -tissue, multi -condition single -cell reference atlas for 540 \nendometriosis that enables deconvolution of endometrial -like cells from their host tissue 541 \nmicroenvironment. We demonstrate that ectopic site, rather than hormonal state, is the dominant 542 \ndriver of transcriptomic heterogeneity among the mesenchymal core lesion cells. The cell type - 543 \nand tissue-specific signatures identified here are sufficient to stratify disease status and lesion site 544 \nin independent bulk data, establishing a molecular framework that can be extended to non-invasive 545 \ndiagnostics through peripheral blood cell -free transcriptomics and menstrual effluent analysis. 546 \nMore broadly, this re -analysis paradigm  of deconvolving spatial and cell type -specific disease 547 \nsignatures from heterogeneous tissue samples is generalizable to any disease with a resident tissue 548 \ncomponent, and offers a scalable approach for robust classifier development across heterogeneous 549 \npatient populations.  Given that site -dependent transcriptomic signatures persist regardless of 550 \nhormonal state, the retrograde menstruation model gains further support from this analysis: lesions 551 \nat different anatomical sites maintain distinct molecular identities that reflect a combination of 552 \ntheir originating tissue and their local tissue context.  553 \nMethods 554 \nSingle cell transcriptomics data integration and cell typing 555 \nWe collected processed scRNA -seq datasets from previously published studies on human 556 \nendometrial tissue, human ovary tissue, and human endometriosis lesions. These datasets included: 557 \nthe Human Endometrial Cell Atlas (Mareckova et al)17, Tabula Sapiens 2.0 Ovary Data20, Jones et 558 \nal19, Fonseca et al endometriosis lesion cells  and unaffected ovary samples 15, Tan et al 559 \nendometriosis lesion cells and peritoneal cells16, and menstrual effluent single cell data from Shih 560 \net al 18. The Human Endometrial Cell Atlas also included endometrial biopsy data from the 561 \nfollowing studies: Wang et al 12, Garcia-Alonso et al 82, Tan et al 16, Lai et al 83, Fonseca et al 15, 562 \nHuang et al84. For the endometriosis lesions datasets only studies which included both peritoneal 563 \nand ovarian endometriosis lesions were included. 564 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n21 \n \nWe integrated the data from different sources using Harmony85 using both dataset and individual 565 \nas our batch correction variables. The resulting Harmony -corrected principal components were 566 \nutilized to estimate a neighbor graph and generate a uniform manifold approximation and 567 \nprojection (UMAP) visualization. Following embedding of the samples on a UMAP, we followed 568 \nthe AnnDictionary automated cell type labeling pipeline21 using Claude Sonnet 3.5. This gave us 569 \npreliminary broad cell type labels. These labels, if they contained cell type names which were too 570 \nbroad, such as “T Cell” were further annotated down to CD4, CD8, and T reg identities using the 571 \nsame marker genes used in the HECA atlas. Additional manual cell typing was applied on a 572 \nlineage-to-lineage basis, where each individual cell type lineage was re -clustered independently 573 \nand more fine-grained cell typing was performed on the AnnDictionary assigned cell types using 574 \nliterature-based marker gene expression. 575 \nSomatic Mutation Calling and Analysis 576 \nSomatic mutations were called de novo from scRNA-seq data from the paired eutopic and ectopic 577 \nsamples found in Fonseca et al 10X 3’ single cell RNA sequencing data using SComatic86. In brief, 578 \nthe raw .fastq files, which were downloaded from GEO accession GSE213216, were processed in 579 \nCell Ranger 10.0.0, using GRCh38,  and produced aligned BAM files. These files were then split 580 \naccording to their cell type annotation and SComatic was used to pileup bases across individual 581 \ncell types, compare variants to the reference genome, annotate the variants, and filter them. The 582 \nvariant filtering process removed sites which were present in a panel of normals and also removed 583 \nknown RNA editing sites as described previously86. 584 \nTo assess selection pressure on coding mutations we computed the ratio of nonsynonymous to 585 \nsynonymous mutations (dN/dS) using the amino acid consequences of the observed mutations, by 586 \nintersecting variant positions with GENCODE CDS annotations from GRCh38. To generate a null 587 \nexpectation under neutrality, we performed a permutation test across 100,000 iterations. In each 588 \npermutation, at each of the 213 coding variant positions, a random alternative base was sample d 589 \nwith probabilities weighted by the observed trinucleotide spectrum at the mutation’s context. This 590 \nsampling was performed across all mutation sites followed by the calculation of the dN/dS. 591 \nSignificance was assessed by computing a z -score and a p -value from the standard normal 592 \ndistribution. 593 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n22 \n \nTo identify shared mutations between eutopic and ectopic samples we used the pooled variants at 594 \neach patient-tissue pair to identify the same somatic changes. To further identify the clonal cell 595 \nfraction, the percentage of cells carrying a given mutation, we ran the de novo somatic variant 596 \ncalling on the cell type BAM files independently, without bulking across cells. This resulted in 597 \nhigh confidence mutation calling for a select group of cells, these high con fidence single cell 598 \nmutation calls, were then used to estimate the fractional contribution of the shared mutations to 599 \nthe overall population of cells. The pooled variants across our samples were utilized to call both 600 \nthe variant allele frequency and accumulation of mutations per cell type. 601 \nSingle cell transcriptomics-based endometrium classifier 602 \nAfter integrating the full single cell object we trained cell type specific binary classifiers to identify 603 \nthe most endometrium-like cells in the lesions. To train this model we utilized the reference data 604 \nfrom the HECA, menstrual effluent, ovarian datasets, and peritoneum data from the Tan et al study. 605 \nFor each annotated cell type we separated our data into the cells from the eutopic endometrium, or 606 \nmenstrual effluent, and those from another tissue source. Cell types were included only when both 607 \nclasses had 200 or greater cells per class.  608 \nWe created ElasticNet classifier models trained on log -normalized gene expression for each cell. 609 \nFeatures were selected in a fold -specific manner to avoid information leakage. Within each 610 \ntraining split, we performed differential gene expression analyses with a Mann-Whitney-Wilcoxon 611 \ntest comparing the endometrium to all other tissues. We filtered the genes for the classifier by 612 \nBenjamini Hochberg adjusted p -value, minimum absolute log2 fold -change (1), and minimum 613 \ndetection frequency (20%). To reduce the influence of broadly expressed or ambiguously 614 \nannotated features, we excluded mitochondrial and ribosomal genes and additional low-specificity 615 \ngene families and identifiers using a fixed pattern -based filter. We also included the opportunity 616 \nfor a tissue-specificity filtering step for the differentially expressed genes, via a cell type level tau 617 \ncalculation87 across the different tissues . This was computed from tissue -wise mean expression 618 \nwithin the training split. 619 \nClassifiers were ElasticNet models which tuned regularization strength and sparsity by grid search 620 \nover multiple values of the inverse regularization parameter and the ElasticNet L1 ratio. 621 \nPerformance was evaluated using nested cross-validation with donor-disjoint outer folds to ensure 622 \nthat training and testing sets contained non -overlapping donors. Hyperparameters were selected 623 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n23 \n \nwithin each outer training fold using an inner cross -validation loop, and the inner folds were also 624 \ndonor-aware. To mitigate donor imbalance in some configurations, we applied donor -balanced 625 \nsample weighting within training folds, reweighting cells inversely proportional to the number of 626 \ncells contributed by each donor within each class. 627 \nTo make the downstream calls equivalent to a probability, we calibrated predicted probabilities 628 \nusing sigmoid calibration fit on the training data within each outer fold, and then generated 629 \ncalibrated probabilities for the held -out validation split. For each outer validation fold, we 630 \nquantified discrimination using area under the precision –recall curve and area under the ROC 631 \ncurve. We additionally reported balanced accuracy, F1 score, and confusion matrices using an 80% 632 \nprobability cut off. For each configuration and cell type, we trained a final model on all available 633 \ncells using the same feature -construction strategy, tuned hyperparameters using donor -aware 634 \ncross-validation when possible, and serialized the fitted model together with the selected feature 635 \nlist and cross-validation summary. 636 \nSelection of the final model was performed through a parameter sweep across feature strategies 637 \n(differential-expression genes with or without tau filtering), feature-set sizes, filtering thresholds, 638 \nand negative -class definitions. To select a single model per cell type, we aggregated cross -639 \nvalidation metrics across runs and ranked candidates. We prioritized configurations meeting a 640 \nminimum F1-performance threshold of 0.9 in the tau constrained models, using non-tau filtered 641 \nmodels as a fallback when the tau-based models did not meet the minimal F1 threshold. Selected 642 \nruns were consolidated into a final model set and summarized with diagnostic plots comparing 643 \nperformance, calibration, and feature-set size across cell types. 644 \nDifferential Gene Expression 645 \nDifferential expression was performed using a two -part hurdle model (MAST) 88 to identify 646 \ntranscriptional changes within each cell type. For each cell type lesion cells were stratified using 647 \nthe classifier derived probability, lesions were divided into cells above or below a fixed probability 648 \nthreshold of 80%. For comparisons involving eutopic endometrium and lesion cells, reference 649 \nendometrium was additionally partitioned to just the samples having endometriosis. For severity-650 \nstratified analyses, samples were categorized into mild or severe disease based on stage groupings, 651 \nand comparisons were performed within severity strata.  652 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n24 \n \nTo minimize confounding by menstrual cycle state  across donors , differential expression was 653 \nperformed using menstrual cycle stage as a covariate. Cells were grouped into discrete cycle-state 654 \nbins, and only cycle states represented in both case and control groups were analyzed. Differential 655 \nexpression was then run independently within each cycle-state subset. For each gene, we extracted 656 \na single, combined significance value from the hurdle component of the model to test for 657 \ndifferential expression between case and control. Effect sizes were reported as log fold changes 658 \nfrom the combined model and converted to log2 scale. Multiple testing correction was performed 659 \nusing the Benjamini -Hochberg procedure to control the false discovery rate. For each run, the 660 \noutput table included per -gene log2 fold change with confidence intervals, test statistics and P 661 \nvalues, and FDR-adjusted P values. 662 \nComparing core and non-core cells to the eutopic endometrium 663 \nTo first compare the core and non -core cells of the lesion cells in different sites to the eutopic 664 \nendometrium, we performed a stratified analysis by performing differential communication and 665 \ndifferential expression analysis. The differential communication analysis was performed with 666 \nscDiffComm28, in which both the ectopic peritoneal and ovarian core and non -core cells were 667 \ncompared to the eutopic endometrium for a total of 4 separate comparisons. Then using the output 668 \ntables from scDiffComm we isolated the “FLAT ” interactions, those which were not changing 669 \nfrom the eutopic endometrium, with an FDR adjusted p -value of below 0.05 and an odds ratio  670 \n(OR) of greater than 1. Then to compare the core versus non-core interaction changes within each 671 \nof our lesion types we subtracted the ORA Score, combined adjusted p-value and log2(OR), of the 672 \nnon-core cells from the core cells to identify stronger FLAT interactions in the core cell or non -673 \ncore cell populations. 674 \nAdditionally, we compared the core and non -core cells to the same cell types in the eutopic 675 \nendometrium of individuals with endometriosis using the strategy outlined above. We were then 676 \nable to make cross core and non-core comparisons by looking for unique genes. We also took the 677 \nJaccard index between these core -specific gene sets for every cell type across the two separa te 678 \nectopic sites. Finally, utilizing gseapy, we searched for the biological process gene ontology terms 679 \nin the differentially expressed genes for specific lineages across the sites. This was performed, in 680 \nbrief, by taking the union of the different lineage differentially expressed genes, running the gseapy 681 \nenrich function, and identifying gene programs in which 3 or more genes were present and 682 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n25 \n \nremoving gene sets that did not contain at least a 10% overlap of gene terms, and removing gene 683 \nsets which were redundant with 50% or more of overlapping genes. 684 \nIdentifying Hormone Sensitive Gene Modules 685 \nIn order to identify gene modules in endometrial cells which are responsive to different stages of 686 \nthe menstrual cycle and hormone therapy we divided our dataset into tissue-cell type pairs, where 687 \ntissue is either eutopic, the core endometrial peritoneal lesion cells or the core endometrial ovarian 688 \nlesion cells. We then tested for differentially expressed genes across hormonal states (Proliferative, 689 \nSecretory, or Exogenous Hormones) using MAST, with donor included as a fixed effect . Cycle 690 \nresponsive genes were identified by thresholding the genes for a Benjamini-Hochberg adjusted p-691 \nvalue of less than 0.05 and a maximum log2 fold change of greater than or equal to one, where the 692 \npeak-versus-trough fold-change was computed as the difference between the highest- and lowest-693 \nexpressing phases for each gene. 694 \nTo identify recurrent trajectory programs across the hormonal status which span the eutopic and 695 \nectopic tissue  we pooled all gene -tissue-cell type triplets. The mean expression of the cy cle-696 \nresponsive genes for these triplets was then z-scored and subject to k-means clustering. We swept 697 \nthe number of clusters from 3 to 10, where k=6 yielded the highest silhouette score (0.540). To 698 \nannotate each cluster with its dominant biological program, we took the union of genes whose 699 \ntriples were assigned to that cluster and ran gene set enrichment analysis for biological processes 700 \nusing gseapy. The resulting term list was filtered to adjusted p -value < 0.05 and  greater than or 701 \nequal to 2 overlapping genes, then a greedy redundancy filter was applied that iteratively selected 702 \nterms by adjusted p -value while excluding terms whose gene content overlapped previously -703 \nselected terms by more than 50%. The top three non -redundant terms per cluster were used as 704 \ncluster labels. 705 \nFor each cell type, cluster-assigned genes were partitioned into mutually-exclusive site categories 706 \nbased on which tissue contexts they were called cycle-responsive in: eutopic-only, EcO-only, EcP-707 \nonly, or shared (assigned in two or more tissues). Per-celltype site-partition counts were visualized 708 \nas stacked bars. We then identified the clusters which reached a peak or had their minimum in the 709 \nSecretory phase of the menstrual cycle. For each cluster group, we identified genes whose triples 710 \nwere assigned to that cluster in at least one ectopic tissue (EcO or EcP) but not in eutopic , the 711 \nectopic-only gene set per cell type. Ectopic -only genes from EcO and EcP were pooled per cell 712 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n26 \n \ntype, and gene set enrichment  analysis was run separately for the secretory -peak (C3) and 713 \nsecretory-trough (C1 or C6) pooled gene sets. Cell types with fewer than two genes in the ectopic-714 \nonly set were excluded from enrichment. 715 \nCore Endometrial Cell Modules 716 \nTo identify co -expressed gene modules  across the core endometrial cells of both ectopic sites , 717 \nconsensus non-negative matrix factorization (cNMF) was applied to the raw gene expression count 718 \nmatrix across matrix ranks ranging from 10 to 55, with 1000 iterations per K. The optimal number 719 \nof factors, K = 33, was determined by evaluating the trade -off between reconstruction error and 720 \nstability across K values. All factorization outputs from the 10 00 iterations at K = 33 were 721 \naggregated and subjected to density filtering, yielding 33 robust gene modules.  722 \ncNMF Analysis 723 \nTo interpret cNMF programs, we performed pathway enrichment on the top program genes using 724 \ngseapy. For each cNMF factor, the top 100 genes with the highest weights were selected for gene 725 \nontology (GO) analysis. For 100 genes in each cNMF factor, pathways with an adjusted p-value < 726 \n0.01 were retained. Among these, up to 10 pathways were selected per factor , a candidate term 727 \nwas discarded if more than 25% of its overlapping genes were shared with any single higher -728 \nranked term retained for the same factor. We further enforced global non -redundancy by 729 \ndiscarding terms whose overlapping gene sets shared more than 25% of their genes with a term 730 \nalready retained for a different factor. Because these filters operate on leading-edge gene sets rather 731 \nthan on term identity, the same GO term may be retained for more than one factor when the 732 \nunderlying genes are distinct. This procedure returned 47 terms across 26 of the 33 factors, with 733 \none to four terms per factor. Where a factor was annotated with multiple terms, the term with the 734 \nlowest adjusted P value was used as its representative label. The complete mapping is provided in 735 \nSupplementary Data Table 1.  Annotated terms and associated gene sets were compiled into a 736 \nprogram-to-pathway mapping used for labeling downstream visualizations and summarization.  737 \nTo summarize cell -type specificity, each cell's program usage vector was normalized to sum to 738 \none across all 33 programs. We then computed mean normalized usage of each annotated program 739 \nwithin each of 19 cell types and z-scored these means across cell types within each program (Figure 740 \n3c). Each program was assigned to the cell type at which its z -scored usage was maximal.  For 741 \ndownstream analysis we retained only programs whose usage was specific to a single cell type, 742 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n27 \n \ndefined as a margin of at least 0.5 standard deviations between the highest and second-highest cell-743 \ntype z-score; 22 of the 26 annotated programs met this criterion. Each cell type was allocated a 744 \nnumber of programs equal to the number of specific programs assigned to it. Three cell types had 745 \nno specific program and were instead allocated the annotated program with the highest raw mean 746 \nusage within that cell type; for two of these, the resulting program was already assigned to another 747 \ncell type and is shown under both. This yielded 25 cell -type-program pairs spanning 23 unique 748 \nprograms (Supplementary Data Table 2). Because specificity is defined by relative enrichment 749 \nacross cell types whereas the fallback rule uses absolute usage within a cell type, one program is 750 \nplaced with stromal cells in Figure 3c and with glandular epithelial cells in Figure 3d. 751 \nTo compare program activity across biological contexts, we defined condition labels as the 752 \ncombination of lesion site class (EcO or EcP) and cycle -state bin. Within each cell type and 753 \ncondition, we computed mean normalized usage of each program  across the cells of the 754 \ncorresponding cell type within each of the six conditions and plotted as raw mean usage. We tested 755 \nfor differences in program usage between lesion sites while adjusting for hormonal state and 756 \naccounting for within-donor correlation. For each cell-type- program pair, we fit an ordinary least 757 \nsquares model predicting per-cell program usage from lesion site and hormonal state (usage ~ site 758 \n+ state) on the cells of that cell type, using cluster -robust standard errors with donor as the 759 \nclustering variable. Because the model is additive, the site coefficient estimates a single state-760 \nadjusted site effect; the reported contrast is EcO minus EcP, with positive values indicating higher 761 \nusage in EcO . Note that clustering adjusts standard errors for within -donor correlation but does 762 \nnot remove donor-level confounding of the point estimates, as donor is nested within lesion site . 763 \nMultiple-hypothesis correction was performed using the Benjamini-Hochberg procedure across all 764 \nreal P-values and results are contained within Supplementary Data Table 3. 765 \nTo quantify global transcriptional differences between conditions while accounting for cell -type 766 \ncomposition, we computed pairwise distances in program -usage space using a cell -type–aware 767 \ncentroid approach. For each cell type and condition, we computed the centroid of selected program 768 \nusage vectors (mean across cells) and calculated pairwise condition distances within each cell type 769 \nusing mean absolute difference across program dimensions. We then aggregated distances across 770 \ncell types using the mean values across the conditions. To summarize site differences within each 771 \ncycle state, we extracted the cell type specific distances between matched site pairs within each 772 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n28 \n \nstate and compared distances across states using paired Wilcoxon signed -rank tests across cell 773 \ntypes, followed by Benjamini Hochberg correction. 774 \nHierarchical clustering of hormonally controlled DEGs 775 \nTo assess the ability of differentially expressed genes in one hormonal state to stratify lesion 776 \nsubtype in other states we first removed the Fonseca et al dataset from the overall single cell object. 777 \nWe then performed differential expr ession analysis on the core lesion cells in the peritoneal 778 \nendometriosis samples against the core lesion cells in the endometrioma samples in the Tan et al 779 \nhormonally controlled samples. We then restrict the differentially expressed genes using an 780 \nadjusted p-value threshold of 0.01 and a log2fc threshold of 0. 75. Amongst these genes for each 781 \nof our mesenchymal cell types, we isolated the top 15 positive and negative direction genes for 782 \neach of the 4 cell types, including Stromal Cells, Vascular Smooth Muscle Cells, Fibroblasts, and 783 \nMyofibroblast like cells. We then observed the expression of these genes on the scaled log 784 \nnormalized expression data from the Fonseca et al core lesion cells across a multitude of menstrual 785 \ncycle stages. We then applied hierarchical clustering to the pseudobulk expression for donor -786 \ntissue-celltype groups and visualized the dendrogram and the heatmap together.  787 \nCreation of tissue and cell type specific gene sets 788 \nTo compare differential expression patterns across both disease severity and tissue groupings we 789 \ncalculated differential gene expression between reference ovarian tissue and our non -core 790 \nendometrioma cells, then between reference peritoneum tissue and our non-core peritoneal lesion 791 \ncells, and between the eutopic endometrium of patients with endometriosis and the core 792 \nendometrioma or peritoneal lesion cells. If a gene for a comparison had an adjusted P value of less 793 \nthan 0.05 it was kept for effect size analysis. To identify endometrioma lineage based effect sizes, 794 \nwe grouped the individual cell types in both endometrioma based comparisons by lineage and 795 \ncalculated the 250 largest magnitude differential expression values. This process was repeated for 796 \nboth severity classes and the same was repeated for the peritoneal based differential expression 797 \nvalues. The lineage level comparisons were performed using a Mann-Whitney-Wilcoxon test with 798 \na BH correction of the P values. 799 \nTo assess whether severity strata shared similar transcriptional responses, we filtered the 800 \ndifferentially expressed genes to only include those with a log2 fold change of magnitude 1 or 801 \ngreater and computed the intersection between the mild and severe differential expression genes 802 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n29 \n \nfor each cell type in each comparison in a direction aware manner. We then took the union of the 803 \nDEG sets across all cell types within each severity grouping and plotted one venn diagram for each 804 \ntissue comparison to show a broad mild to severe gene overlap. 805 \nTo interpret the ability of the intersecting genes to be deconvolved in a non -invasive manner we 806 \nquantified the number of tissue enriched/enhanced genes (HPA)26 which also had a Gini coefficient 807 \nof >0.6, and cell type specific gene sets as defined by the Gini coefficient of  >0.8 for differentially 808 \nexpressed genes between cell types  in Tabula Sapiens v.2.0 as described previously62. For tissue 809 \nspecificity, we only included genes which were specific for the endometrium or ovaries, and the 810 \nperitoneum specific set was approximated as being tissue  specific genes for tissues in the 811 \nperitoneal cavity, such as the bladder, fallopian tubes, adipose tissue, and intestines . For each 812 \nintersection, we recorded whether the gene was cell type specific for any of its mapped cell types. 813 \nTo directly compare tissue and cell type specificity within the shared mild to severe genes we 814 \nclassified each gene as tissue or cell type specific. For each comparison, we visualized category 815 \nproportions using pie charts. We then saved the union of the cell type specific gene across all of 816 \nour comparisons, we also saved the union of tissue specific genes across all comparisons for 817 \ndownstream analysis. 818 \nMachine Learning Classification of Endometriosis Lesions 819 \nMicro-array data was obtained from the Gene Expression Omnibus (GEO accession: 820 \nGSE141549)59, comprising batch-corrected, normalized microarray expression profiles from 205 821 \nsamples across seven tissue categories: control endometrium, patient endometrium, control 822 \nperitoneum, patient peritoneum, deep infiltrating endometriosis, peritoneal lesions, and ovarian 823 \nendometrioma. We employed a patient-aware data partitioning approach during model evaluation, 824 \nmaking sure no individual donor was included in both the training and the validation samples. 825 \nWe created three separate models for every comparison, a tissue -enriched gene set model 826 \n(described above), a cell type specific gene set model (described above), and a combined gene set 827 \nmodel. Genes not present in the expression dataset were excluded, and duplicate gene identifiers 828 \nwere resolved by retaining the first occurrence. We first calculated the first and second princip al 829 \ncomponents of the micro -array data for the three separate gene sets and observed visualization 830 \nacross the different tissues present. 831 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n30 \n \nWe then created machine learning models for each of our gene sets, using a  Random Forest  832 \nclassifier, to distinguish between (1) control and patient endometrium, (2) control and patient 833 \nperitoneum, and (3) between deep infiltrating endometriosis, peritoneal lesions, and ovarian 834 \nendometrioma. To train the model features were standardized to zero mean and unit variance prior 835 \nto model fitting. Hyperparameters were tuned via grid search and the class weights were also 836 \nbalanced to account for unequal class frequencies. Nested cross validation was performed using 837 \nstratified group k -fold cross -validation (k=5) on both the inner and out er folds, ensuring all 838 \nsamples from a given patient remained within the same fold. Hyperparameter selection was 839 \nperformed via nested cross -validation where the inner loop was used for identifying optimal 840 \nparameters and the outer loop evaluation providing performance estimates. 841 \nModel performance was assessed using the area under the receiver operating characteristic curve 842 \n(ROC-AUC) for the out -of-fold prediction generated from the nested cross -validation. For our 843 \nmulti-class classifier a macro-averaged one-versus-rest ROC curve was computed for every gene 844 \nset. Confidence intervals for ROC curves and AUC values were estimated via block bootstrapping 845 \nwith the patient as the resampling unit. In short, for each of  the 1,000 bootstrap iterations, N 846 \npatients were drawn with replacement from the N unique patients in the analysis, and all out -of-847 \nfold predictions belonging to the sampled patients were aggregated to form a bootstrap sample (so 848 \na patient drawn twice contributed all of their samples twice).  The 2.5th and 97.5th percentiles of 849 \nthe bootstrap distribution defined the confidence bounds . The coefficients of a Random Forest 850 \nmodel trained across the entire dataset were then used to rank feature importance. 851 \nData Availability 852 \nAll data utilized in this experiment was from publicly available repositories which did not require 853 \napproval. The HECA was downloaded from  a web portal which originated from its parent 854 \npublication: https://www.reproductivecellatlas.org/endometrium_reference.html. Shih et al 855 \nmenstrual effluent data was available through National Center for Biotechnology 856 \nInformation/Gene Expression Omnibus (GEO) accession number GSE203191. Tan et al lesion 857 \nand peritoneal data were available through a web portal: 858 \nhttps://singlecell.jax.org/datasets/endometriosis-2022. Fonseca et al lesion and ovary data w ere 859 \navailable through GEO accession number GSE213216. Jones et al single cell ovary data was 860 \navailable through GEO accession number GSE260685. Tabula Sapiens 2.0 could be accessed 861 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n31 \n \nthrough the CellxGene Web portal: https://cellxgene.cziscience.com/collections/e5f58829-1a66-862 \n40b5-a624-9046778e74f5. 863 \nCode Availability 864 \nAll code for the work and the manuscript are available on GitHub at: 865 \nhttps://github.com/doughenze/endometriosis_sites. 866 \nAcknowledgments 867 \nThe authors thank all members of the Quake Lab for valuable feedback and discussions throughout 868 \nthe study.  869 \nAuthor Information 870 \nDepartment of Bioengineering, Stanford University, Stanford, CA, USA 871 \nDouglas E. Henze, George Crowley, and Stephen R. Quake 872 \nDepartment of Applied Physics, Stanford University, Stanford, CA, USA 873 \nStephen R. Quake 874 \nAuthor’s Contribution 875 \nD.E.H, G.C., and S.R.Q conceptualized and designed experiments. D.E.H collected data. D.E.H 876 \nand G.C analyzed the data. D.E.H and S.R.Q discussed and interpreted the data. D.E.H designed 877 \nthe figures. D.E.H and S.R.Q wrote the manuscript. D.E.H  and S.R.Q supervised, directed, and 878 \nmanaged the study. All authors discussed the results and commented on the manuscript. 879 \nCompeting Interest 880 \nThe authors declare no competing interests. 881 \nCorrespondence 882 \nCorrespondence to Stephen R. Quake 883 \n 884 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n32 \n \nReferences 885 \n1. Zondervan, K. T., Becker, C. M. & Missmer, S. A. Endometriosis. N. Engl. J. Med. 382, 886 \n1244–1256 (2020). 887 \n2. Giudice, L. C. Endometriosis. N. Engl. J. Med. 362, 2389–2398 (2010). 888 \n3. Taylor, H. S. Reimagining Endometriosis. Med 2, 481–485 (2021). 889 \n4. Zondervan, K. T. et al. Endometriosis. Nat. Rev. Dis. Primer 4, 9 (2018). 890 \n5. American Society For Reproductive Medicine. Revised American Society for 891 \nReproductive Medicine classification of endometriosis: 1996. Fertil. Steril. 67, 817–821 892 \n(1997). 893 \n6. Nisolle, M. & Donnez, J. Peritoneal endometriosis, ovarian endometriosis, and 894 \nadenomyotic nodules of the rectovaginal septum are three different entities. Fertil. 895 \nSteril. 68, 585–596 (1997). 896 \n7. Bulun, S. E. et al. Endometriosis. Endocr. Rev. 40, 1048–1079 (2019). 897 \n8. Zeitoun, K. M. & Bulun, S. E. Aromatase: a key molecule in the pathophysiology of 898 \nendometriosis and a therapeutic target. Fertil. Steril. 72, 961–969 (1999). 899 \n9. Vercellini, P., Viganò, P., Somigliana, E. & Fedele, L. Endometriosis: pathogenesis and 900 \ntreatment. Nat. Rev. Endocrinol. 10, 261–275 (2014). 901 \n10. La Marca, A., Carducci Artenisio, A., Stabile, G., Rivasi, F. & Volpe, A. Evidence for 902 \ncycle-dependent expression of follicle-stimulating hormone receptor in human 903 \nendometrium. Gynecol. Endocrinol. Off. J. Int. Soc. Gynecol. Endocrinol. 21, 303–306 904 \n(2005). 905 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n33 \n \n11. Sapkota, Y. et al. Meta-analysis identifies five novel loci associated with endometriosis 906 \nhighlighting key genes involved in hormone metabolism. Nat. Commun. 8, 15539 907 \n(2017). 908 \n12. Wang, W. et al. Single-cell transcriptomic atlas of the human endometrium during the 909 \nmenstrual cycle. Nat. Med. 26, 1644–1653 (2020). 910 \n13. Talbi, S. et al. Molecular phenotyping of human endometrium distinguishes menstrual 911 \ncycle phases and underlying biological processes in normo-ovulatory women. 912 \nEndocrinology 147, 1097–1121 (2006). 913 \n14. Liu, S. et al. Single-cell and spatial transcriptomic profiling revealed niche interactions 914 \nsustaining growth of endometriotic lesions. Cell Genomics 5, 100737 (2025). 915 \n15. Fonseca, M. A. S. et al. Single-cell transcriptomic analysis of endometriosis. Nat. 916 \nGenet. 55, 255–267 (2023). 917 \n16. Tan, Y. et al. Single-cell analysis of endometriosis reveals a coordinated transcriptional 918 \nprogramme driving immunotolerance and angiogenesis across eutopic and ectopic 919 \ntissues. Nat. Cell Biol. 24, 1306–1318 (2022). 920 \n17. Marečková, M. et al. An integrated single-cell reference atlas of the human 921 \nendometrium. Nat. Genet. 56, 1925–1937 (2024). 922 \n18. Shih, A. J. et al. Single-cell analysis of menstrual endometrial tissues defines 923 \nphenotypes associated with endometriosis. BMC Med. 20, 315 (2022). 924 \n19. Jones, A. S. K. et al. Cellular atlas of the human ovary using morphologically guided 925 \nspatial transcriptomics and single-cell sequencing. Sci. Adv. 10, eadm7506 (2024). 926 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n34 \n \n20. Quake, S. R. & The Tabula Sapiens Consortium. Tabula Sapiens reveals transcription 927 \nfactor expression, senescence effects, and sex-specific features in cell types from 28 928 \nhuman organs and tissues. Preprint at https://doi.org/10.1101/2024.12.03.626516 929 \n(2024). 930 \n21. Crowley, G. et al. Benchmarking cell type and gene set annotation by large language 931 \nmodels with AnnDictionary. Nat. Commun. 16, 9511 (2025). 932 \n22. Sampson, J. A. Metastatic or Embolic Endometriosis, due to the Menstrual 933 \nDissemination of Endometrial Tissue into the Venous Circulation. Am. J. Pathol. 3, 93-934 \n110.43 (1927). 935 \n23. Moore, L. et al. The mutational landscape of normal human endometrial epithelium. 936 \nNature 580, 640–646 (2020). 937 \n24. Ólafsson, S. et al. Endometriosis lesions are oligoclonal structures derived from the 938 \nnormal endometrium. Preprint at https://doi.org/10.64898/2026.02.25.708037 (2026). 939 \n25. Suda, K. et al. Different mutation profiles between epithelium and stroma in 940 \nendometriosis and normal endometrium. Hum. Reprod. 34, 1899–1905 (2019). 941 \n26. Uhlén, M. et al. Proteomics. Tissue-based map of the human proteome. Science 347, 942 \n1260419 (2015). 943 \n27. Henlon, Y. et al. Single-cell analysis identifies distinct macrophage phenotypes 944 \nassociated with prodisease and proresolving functions in the endometriotic niche. 945 \nProc. Natl. Acad. Sci. U. S. A. 121, e2405474121 (2024). 946 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n35 \n \n28. Lagger, C. et al. scDiffCom: a tool for differential analysis of cell–cell interactions 947 \nprovides a mouse atlas of aging changes in intercellular communication. Nat. Aging 3, 948 \n1446–1461 (2023). 949 \n29. Burns, G. W. et al. Spatial transcriptomic analysis identifies epithelium-macrophage 950 \ncrosstalk in endometriotic lesions. iScience 28, 111790 (2025). 951 \n30. Greaves, E. et al. Estradiol is a critical mediator of macrophage-nerve cross talk in 952 \nperitoneal endometriosis. Am. J. Pathol. 185, 2286–2297 (2015). 953 \n31. Ono, Y. et al. CD206+ macrophage is an accelerator of endometriotic-like lesion via 954 \npromoting angiogenesis in the endometriosis mouse model. Sci. Rep. 11, 853 (2021). 955 \n32. Bacci, M. et al. Macrophages are alternatively activated in patients with endometriosis 956 \nand required for growth and vascularization of lesions in a mouse model of disease. 957 \nAm. J. Pathol. 175, 547–556 (2009). 958 \n33. Wu, M.-H., Hsiao, K.-Y. & Tsai, S.-J. Endometriosis and possible inflammation markers. 959 \nGynecol. Minim. Invasive Ther. 4, 61–67 (2015). 960 \n34. Dawood, M. Y. Primary dysmenorrhea: advances in pathogenesis and management. 961 \nObstet. Gynecol. 108, 428–441 (2006). 962 \n35. Stratton, P. & Berkley, K. J. Chronic pelvic pain and endometriosis: translational 963 \nevidence of the relationship and implications. Hum. Reprod. Update 17, 327–346 964 \n(2011). 965 \n36. ACOG Committee Opinion No. 760: Dysmenorrhea and Endometriosis in the 966 \nAdolescent. Obstet. Gynecol. 132, e249–e258 (2018). 967 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n36 \n \n37. Begum, I. A. The connection between endometriosis and secondary dysmenorrhea. J. 968 \nReprod. Immunol. 168, 104425 (2025). 969 \n38. Deuster, P. A., Dolev, E., Bernier, L. L. & Trostmann, U. H. Magnesium and zinc status 970 \nduring the menstrual cycle. Am. J. Obstet. Gynecol. 157, 964–968 (1987). 971 \n39. Bruner, K. L., Eisenberg, E., Gorstein, F. & Osteen, K. G. Progesterone and transforming 972 \ngrowth factor-β coordinately regulate suppression of endometrial matrix 973 \nmetalloproteinases in a model of experimental endometriosis. Steroids 64, 648–653 974 \n(1999). 975 \n40. Obeagu, E. I. & Obeagu, G. U. From menses to ovulation: the immunomodulatory role 976 \nof monocyte cytokines in menstrual cycle physiology and reproductive health - a 977 \nnarrative review. Ann. Med. Surg. 87, 8453–8459 (2025). 978 \n41. Wira, C. R., Rodriguez-Garcia, M. & Patel, M. V. The role of sex hormones in immune 979 \nprotection of the female reproductive tract. Nat. Rev. Immunol. 15, 217–230 (2015). 980 \n42. Attia, G. R. et al. Progesterone receptor isoform A but not B is expressed in 981 \nendometriosis. J. Clin. Endocrinol. Metab. 85, 2897–2902 (2000). 982 \n43. Burney, R. O. et al. Gene expression analysis of endometrium reveals progesterone 983 \nresistance and candidate susceptibility genes in women with endometriosis. 984 \nEndocrinology 148, 3814–3826 (2007). 985 \n44. Vissers, G., Giacomozzi, M., Verdurmen, W., Peek, R. & Nap, A. The role of fibrosis in 986 \nendometriosis: a systematic review. Hum. Reprod. Update 30, 706–750 (2024). 987 \n45. Gong, H. et al. The regulation of ovary and conceptus on the uterine natural killer cells 988 \nduring early pregnancy. Reprod. Biol. Endocrinol. 15, 73 (2017). 989 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n37 \n \n46. Hogg, C. et al. Macrophages inhibit and enhance endometriosis depending on their 990 \norigin. Proc. Natl. Acad. Sci. 118, e2013776118 (2021). 991 \n47. Rocha, A. L. L., Reis, F. M. & Taylor, R. N. Angiogenesis and endometriosis. Obstet. 992 \nGynecol. Int. 2013, 859619 (2013). 993 \n48. Reichelt, U. et al. High lymph vessel density and expression of lymphatic growth 994 \nfactors in peritoneal endometriosis. Reprod. Sci. 19, 876–882 (2012). 995 \n49. Zhang, Q., Duan, J., Olson, M., Fazleabas, A. & Guo, S.-W. Cellular Changes Consistent 996 \nWith Epithelial-Mesenchymal Transition and Fibroblast-to-Myofibroblast 997 \nTransdifferentiation in the Progression of Experimental Endometriosis in Baboons. 998 \nReprod. Sci. 23, 1409–1421 (2016). 999 \n50. Pušić, M. N. et al. Growth Arrest-Specific Protein 6 Is Elevated in Endometriosis but 1000 \nShows Poor Diagnostic Performance. Int. J. Mol. Sci. 26, 8348 (2025). 1001 \n51. Zhai, X. et al. Gas6/AXL pathway: immunological landscape and therapeutic potential. 1002 \nFront. Oncol. 13, 1121130 (2023). 1003 \n52. Klemmt, P. A. B., Carver, J. G., Kennedy, S. H., Koninckx, P. R. & Mardon, H. J. Stromal 1004 \ncells from endometriotic lesions and endometrium from women with endometriosis 1005 \nhave reduced decidualization capacity. Fertil. Steril. 85, 564–572 (2006). 1006 \n53. Wang, X. et al. The role of insulin-like growth factor binding proteins in TGF-β1-induced 1007 \nfibroblast-myofibroblast transition during endometriosis fibrosis. Cell. Signal. 140, 1008 \n112362 (2026). 1009 \n54. Giudice, L. C., Dsupin, B. A., Gargosky, S. E., Rosenfeld, R. G. & Irwin, J. C. The insulin-1010 \nlike growth factor system in human peritoneal fluid: its effects on endometrial stromal 1011 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n38 \n \ncells and its potential relevance to endometriosis. J. Clin. Endocrinol. Metab. 79, 1284–1012 \n1293 (1994). 1013 \n55. Clower, L., Fleshman, T., Geldenhuys, W. J. & Santanam, N. Targeting Oxidative Stress 1014 \nInvolved in Endometriosis and Its Pain. Biomolecules 12, 1055 (2022). 1015 \n56. Harrington, D. J. et al. Tenascin is differentially expressed in endometrium and 1016 \nendometriosis. J. Pathol. 187, 242–248 (1999). 1017 \n57. Ma, R. et al. Identification of key genes associated with endometriosis and endometrial 1018 \ncancer by bioinformatics analysis. Front. Oncol. 14, 1387860 (2024). 1019 \n58. Yotova, I. et al. LINC01638 promotes epithelial-to-mesenchymal transition in 1020 \nendometriosis epithelial cells by up-regulating RHOB via HDAC1 suppression. Reprod. 1021 \nBiomed. Online 51, 104942 (2025). 1022 \n59. Ganieva, U. et al. Involvement of Transcription Factor 21 in the Pathogenesis of Fibrosis 1023 \nin Endometriosis. Am. J. Pathol. 190, 145–157 (2020). 1024 \n60. Lee, M.-Y. et al. Role of interleukin-32 in the pathogenesis of endometriosis: in vitro, 1025 \nhuman and transgenic mouse data. Hum. Reprod. 33, 807–816 (2018). 1026 \n61. Rahmioglu, N. et al. Genetic variants underlying risk of endometriosis: insights from 1027 \nmeta-analysis of eight genome-wide association and replication datasets. Hum. 1028 \nReprod. Update 20, 702–716 (2014). 1029 \n62. Vorperian, S. K. et al. Cell types of origin of the cell-free transcriptome. Nat. 1030 \nBiotechnol. 40, 855–861 (2022). 1031 \n63. Moufarrej, M. N. et al. Early prediction of preeclampsia in pregnancy with cell-free RNA. 1032 \nNature 602, 689–694 (2022). 1033 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n39 \n \n64. Gabriel, M. et al. A relational database to identify differentially expressed genes in the 1034 \nendometrium and endometriosis lesions. Sci. Data 7, 284 (2020). 1035 \n65. Li, Q. et al. Identification of Candidate Gene Signatures and Regulatory Networks in 1036 \nEndometriosis and its Related Infertility by Integrated Analysis. Reprod. Sci. 29, 411–1037 \n426 (2022). 1038 \n66. Donnez, J. & Cacciottola, L. Endometriosis: An Inflammatory Disease That Requires 1039 \nNew Therapeutic Options. Int. J. Mol. Sci. 23, 1518 (2022). 1040 \n67. Lousse, J.-C. et al. Peritoneal endometriosis is an inflammatory disease. Front. Biosci. 1041 \n4, 23–40 (2012). 1042 \n68. Ferrero, S., Vellone, V. G. & Barra, F. Pathophysiology of pain in patients with peritoneal 1043 \nendometriosis. Ann. Transl. Med. 7, S8 (2019). 1044 \n69. Nisenblat, V. et al. Blood biomarkers for the non-invasive diagnosis of endometriosis. 1045 \nCochrane Database Syst. Rev. 2016, CD012179 (2016). 1046 \n70. Halpern, K. B. et al. Single-cell spatial reconstruction reveals global division of labour 1047 \nin the mammalian liver. Nature 542, 352–356 (2017). 1048 \n71. Moor, A. E. et al. Spatial Reconstruction of Single Enterocytes Uncovers Broad Zonation 1049 \nalong the Intestinal Villus Axis. Cell 175, 1156-1167.e15 (2018). 1050 \n72. Suda, K. et al. Clonal Expansion and Diversification of Cancer-Associated Mutations in 1051 \nEndometriosis and Normal Endometrium. Cell Rep. 24, 1777–1789 (2018). 1052 \n73. Anglesio, M. S. et al. Cancer-Associated Mutations in Endometriosis without Cancer. 1053 \nN. Engl. J. Med. 376, 1835–1848 (2017). 1054 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n40 \n \n74. Murray, P. J. & Wynn, T. A. Protective and pathogenic functions of macrophage subsets. 1055 \nNat. Rev. Immunol. 11, 723–737 (2011). 1056 \n75. Donnez, J. et al. Peritoneal endometriosis and “endometriotic” nodules of the 1057 \nrectovaginal septum are two different entities. Fertil. Steril. 66, 362–368 (1996). 1058 \n76. Lavin, Y. et al. Tissue-resident macrophage enhancer landscapes are shaped by the 1059 \nlocal microenvironment. Cell 159, 1312–1326 (2014). 1060 \n77. Gautier, E. L. et al. Gene-expression profiles and transcriptional regulatory pathways 1061 \nthat underlie the identity and diversity of mouse tissue macrophages. Nat. Immunol. 1062 \n13, 1118–1128 (2012). 1063 \n78. Koh, W. et al. Noninvasive in vivo monitoring of tissue-specific global gene expression 1064 \nin humans. Proc. Natl. Acad. Sci. 111, 7361–7366 (2014). 1065 \n79. Larson, M. H. et al. A comprehensive characterization of the cell-free transcriptome 1066 \nreveals tissue- and subtype-specific biomarkers for cancer detection. Nat. Commun. 1067 \n12, 2357 (2021). 1068 \n80. Warren, L. A. et al. Analysis of menstrual effluent: diagnostic potential for 1069 \nendometriosis. Mol. Med. 24, 1 (2018). 1070 \n81. Wilson, T. R. et al. Analysis of menstrual effluent uncovers endometriosis-specific cell 1071 \npopulations and impaired cellular pathway processes. Preprint at 1072 \nhttps://doi.org/10.1101/2025.08.21.671582 (2025). 1073 \n82. Garcia-Alonso, L. et al. Mapping the temporal and spatial dynamics of the human 1074 \nendometrium in vivo and in vitro. Nat. Genet. 53, 1698–1711 (2021). 1075 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n41 \n \n83. Lai, Z.-Z. et al. Single-cell transcriptome profiling of the human endometrium of 1076 \npatients with recurrent implantation failure. Theranostics 12, 6527–6547 (2022). 1077 \n84. Huang, X. et al. Single-cell transcriptome analysis reveals endometrial immune 1078 \nmicroenvironment in minimal/mild endometriosis. Clin. Exp. Immunol. 212, 285–295 1079 \n(2023). 1080 \n85. Korsunsky, I. et al. Fast, sensitive and accurate integration of single-cell data with 1081 \nHarmony. Nat. Methods 16, 1289–1296 (2019). 1082 \n86. Muyas, F. et al. De novo detection of somatic mutations in high-throughput single-cell 1083 \nprofiling data sets. Nat. Biotechnol. 42, 758–767 (2024). 1084 \n87. Yanai, I. et al. Genome-wide midrange transcription profiles reveal expression level 1085 \nrelationships in human tissue specification. Bioinformatics 21, 650–659 (2005). 1086 \n88. Finak, G. et al. MAST: a flexible statistical framework for assessing transcriptional 1087 \nchanges and characterizing heterogeneity in single-cell RNA sequencing data. Genome 1088 \nBiol. 16, 278 (2015). 1089 \n 1090 \n1091 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n42 \n \nFigures 1092 \n 1093 \n 1094 \n 1095 \nFigure 1: Integrated cell atlas across endometriosis related tissues 1096 \nA, Schematic overview of the data analysis pipeline, building references and training classifiers 1097 \non reference tissue and applying them to lesions. B, Donor distribution, number of unique donors 1098 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n43 \n \nfor every tissue-dataset pair. C, UMAP plot across the different tissues colored by broad cell 1099 \ntype. D, Marker gene expression across the broad cell types. 1100 \n 1101 \n 1102 \n 1103 \n 1104 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n44 \n \nFigure 2: Tissue-of-origin classifiers reveal heterogeneous endometrial-like and host tissue 1105 \npopulations within ectopic lesions. 1106 \nA,  Grouped bar plot showing the relative performance amongst each of the cell type-specific 1107 \nclassifiers. Sorted by macro F1. B,  Percentage of lesion cells for each lineage which mapped to 1108 \nbeing endometrium-like (green) or not (gray). C, Empirical cumulative distribution function to 1109 \nshow the relative percent of differentially expressed genes from the eutopic endometrium; lines 1110 \nare colored by site and core or non-core category. D, Ranked Jaccard Indices between the core-1111 \nspecific differentially expressed genes for each of the cell types across the endometrioma and 1112 \nperitoneal endometriosis samples E, Gene ontology biological processes terms for the 1113 \ndifferentially expressed genes from the eutopic endometrium for myeloid and lymphoid cells 1114 \nwithin each of the sites. 1115 \n 1116 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n45 \n \n 1117 \n 1118 \nFigure 3: Transcriptomic programs are more associated with site than hormonal status 1119 \nA, Schematic image of the primary axes of lesion cell variation being hormonal or spatial. B, 1120 \nDumbbell plot showing the spearman correlation across sites in matched hormone cases or 1121 \nacross hormones in matched site cases, with line meaning perfect agreement. C, Diagonalized 1122 \nheatmap showing the z-scored usage of different NMF modules across cell types. D, Bar plot 1123 \nshowing the mean usage for each site-hormone pair across all cell types. * represents a 1124 \nstatistically significant deviation between sites via a linear mixed model.  1125 \n 1126 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n46 \n \n 1127 \nFigure 4: Severity-associated cell type- and tissue-specific signatures classify disease status 1128 \nand lesion origin. 1129 \nA, PCA embedding of combined cell type- and tissue-specific gene signature in the bulk tissue 1130 \nmicro array data. B, Performance of each of the different gene sets to diagnose bulk eutopic 1131 \nendometrium samples as having endometriosis or not as shown via a receiver operator 1132 \ncharacteristic (ROC) curve. C, Performance of each of the different gene sets to diagnose bulk 1133 \nperitoneum samples as having endometriosis or not as shown via a ROC curve. D, Performance 1134 \nof each of the different gene sets to stratify lesions by tissue of origin as shown via a ROC curve. 1135 \nE, Log normalized expression of PDLIM3 and SERPINE2 across all groups. Statistics shown are 1136 \nfrom a Wilcoxon rank sum test with a Benjamini-Hochberg correction across all possible 1137 \ncomparisons for a given gene. * is for a corrected p-value below 0.05, ** is for a corrected p-1138 \nvalue below 0.01, and *** is for a corrected p-value below 0.001. All AUC curves show the 1139 \ncalculated curve and a 95% confidence interval.  1140 \n1141 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n47 \n \nSupplementary Figures 1142 \n 1143 \n 1144 \n 1145 \nSupplementary Figure 1: Semi-automated cell typing recovers previously identified cell 1146 \npopulations 1147 \nA, Number of transcripts per cell, line on the violin plot indicates the median value. B, Number 1148 \nof genes per cell, line on the violin plot indicates the median value. C, Cluster map showing the 1149 \nrelative percentage of HECA labels for each of our semi-automated pipeline created labels. D, 1150 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n48 \n \nUMAP representation of all cells colored by tissue. E, UMAP representation of all cells colored 1151 \nby Dataset. F, UMAP representation of all cells colored by endometriosis stage. 1152 \n 1153 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n49 \n \n 1154 \nSupplementary Figure 2: Somatic mutation accumulation in eutopic and ectopic 1155 \nendometrium 1156 \nA-B, 6-class (A) and 96 class (B) mutation spectra between the eutopic endometrium and the 1157 \nectopic endometrium. C, Permutation test results showing expected against null dN/dS for the 1158 \nobserved mutation spectra and mutation sites. D, Number of mutations shared between the 1159 \neutopic and ectopic endometrium in a given patient. E, Percentage of cells, bulked across the two 1160 \npatients with shared mutations, that display the shared mutation in the eutopic (left) and ectopic 1161 \n(right) endometrium. F, cumulative density functions of the variant allele frequencies across the 1162 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n50 \n \ndifferent patients eutopic and ectopic samples. G, Total number of mutations attributable to each 1163 \ncell type, colored by sample location.  1164 \n 1165 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n51 \n \n 1166 \nSupplementary Figure 3: Validation of cell type-level endometrial origin classifiers 1167 \nA, Confusion matrices of classifier predictions across populous cell types B, Bar plot showing 1168 \nthe number of endometrium tissue-enhanced or tissue enriched HPA genes which were used in 1169 \nthe classifier for each cell type. C, UMAP showing value of applied classifier across the whole 1170 \nintegrated dataset. D-E, Stacked bar plots showing classifier values across different cell lineages 1171 \nin the (D) endometrioma and the (E) peritoneal endometriosis samples. 1172 \n 1173 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n52 \n \n 1174 \nSupplementary Figure 4: Differential cell-cell interaction between the eutopic 1175 \nendometrium and core or non-core cell populations 1176 \nA, Heatmap showing the difference in odds ratio scores of preserved cell-cell interactions from 1177 \nthe eutopic endometrium to core and non-core cells in endometrioma. B, Heatmap showing the 1178 \ndifference in odds ratio scores of preserved cell-cell interactions from the eutopic endometrium 1179 \nto core and non-core cells in peritoneal endometriosis. 1180 \n 1181 \n 1182 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n53 \n \n 1183 \n 1184 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n54 \n \nSupplementary Figure 5: Core- and non-core-specific differential gene expression between 1185 \nthe eutopic endometrium and different lesion sites. 1186 \nA, Summary of differentially expressed genes as core-specific, non-core-specific, or shared for 1187 \nendometrioma samples across all cell types. B, Summary of differentially expressed genes as 1188 \ncore-specific, non-core-specific, or shared across peritoneal endometriosis samples across all cell 1189 \ntypes. C-E, Gene Ontology Biological Processes for endometrioma macrophage (C) non-core 1190 \nspecific, (D) shared, and (E) core-specific differentially expressed genes. F-H Gene Ontology 1191 \nBiological Processes for peritoneal endometriosis macrophage (F) non-core specific, (G) shared, 1192 \nand (H) core-specific differentially expressed genes. 1193 \n 1194 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n55 \n \n1195 \nSupplementary Figure 6: Hormone-responsive gene modules explain endometriosis pain. 1196 \nA, Barplot showing the number of cycle-dependent differentially expressed genes for each cell 1197 \ntype across the ectopic ovary (EcO), ectopic peritoneum (EcP), and eutopic endometrium 1198 \n(Eutopic) tissues. B, Hormone responsive cluster profiles identified through k-means clustering 1199 \nof the mean expression for a tissue, cell type, and gene triplet at each hormonal state. Clusters are 1200 \nlabeled with the top 3 gene ontology terms for the union of all genes present in the cluster. C, 1201 \nStacked bar plot representing the number of cycle-responsive genes in Cluster 3 (C3) for EcO, 1202 \nEcP, Eutopic, and shared between two or more. D, Stacked bar plot representing the number of 1203 \ncycle-responsive genes in Cluster 1 or Cluster 6 (C1 | C6) for EcO, EcP, Eutopic, and shared 1204 \nbetween two or more. E, Bubble plot showing gene ontology biological processes enriched in the 1205 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n56 \n \nectopic only (EcO or EcP) genes for Cluster 3 across each cell type. F, Bubble plot showing gene 1206 \nontology biological processes enriched in the ectopic only (EcO or EcP) genes for Cluster 1 or 1207 \nCluster 6 across each cell type. 1208 \n 1209 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n57 \n \n 1210 \nSupplementary Figure 7: Site-dependent cell type signatures are conserved across the full 1211 \nspectrum of hormonal conditions. 1212 \nA, Boxplot showing the euclidean distance between the usage space for each individual cell type 1213 \nacross sites for each hormone state. n.s represents no statistically significant deviation. B, 1214 \nClustered heatmap showing z-scored expression of top 15 differentially expressed genes across 1215 \nmesenchymal cell types in a hormonally controlled dataset as applied to a non-hormonally 1216 \ncontrolled dataset. 1217 \n 1218 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n58 \n \n 1219 \n 1220 \nSupplementary Figure 8: Severity associated differentially expressed genes contain cell 1221 \ntype- and tissue-specific genes.  1222 \nA, Boxplot representation of the top 250 genes associated with each lineage across different 1223 \nseverity classifications of endometriosis. B, Venn diagram representation of the differentially 1224 \nexpressed genes split by severity across endometrioma (left) and percentage of severity shared 1225 \ngenes that are non-specific, cell type-specific, tissue specific, or both (right). C, Venn diagram 1226 \nrepresentation of the differentially expressed genes split by severity across peritoneal 1227 \nendometriosis (left) and percentage of severity shared genes that are non-specific, cell type-1228 \nspecific, tissue specific, or both (right). 1229 \n 1230 \n 1231 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n59 \n \n 1232 \n 1233 \nSupplementary Figure 9: Bulk tissue microarray data can be separated by cell type- and 1234 \ntissue-specific gene signatures 1235 \nA-B, PCA projections of bulk tissue microarray data using cell type-specific (A) genes and 1236 \ntissue-specific (B) genes. 1237 \n 1238 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n60 \n \n 1239 \n 1240 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint \n\n \n61 \n \nSupplementary Figure 10: Combined cell type- and tissue-specific gene signature 1241 \nimportances across diverse classifiers 1242 \nA, Combined cell type- and tissue-specific gene signature classifier importances for predicting 1243 \ncontrol against diseased eutopic endometrium. B, Combined cell type- and tissue-specific gene 1244 \nsignature classifier importances for predicting control against diseased peritoneum. C, Combined 1245 \ncell type- and tissue-specific gene signature classifier importances for multi-class prediction 1246 \nacross three separate lesion types (deep infiltrating endometriosis, peritoneal endometriosis, and 1247 \nendometrioma). 1248 \n  1249 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted July 16, 2026. ; https://doi.org/10.64898/2026.07.10.737856doi: bioRxiv preprint","source_license":"CC0","license_restricted":false}