Single-nuclei RNA sequencing reveals obesity-associated rewiring of estrogen signaling in endometrioid adenocarcinoma

Summary

Endometrioid adenocarcinoma has one of the strongest body mass index associations of any solid tumor. This is widely attributed to amplification of estrogen-driven signaling via increased production of unopposed estrogen in peripheral adipose tissue. However, there is little evidence as to whether this is a pure dosage effect or if persistent signaling leads to feedback on the execution of that signaling. Using single-nucleus RNA sequencing of primary tumors from postmenopausal patients with normal and obese body mass indices, we identified coordinated transcriptional remodeling across tumor epithelial, immune, and stromal compartments. Network-level analysis of tumor epithelial cells revealed that ESR1-associated co-expression modules in tumors from patients with obesity were distinct across two independent patient cohorts and transcriptomic modalities. These network constructions were then shown to be enriched for estrogen receptor targets using existing chromatin immunoprecipitation data from endometrioid adenocarcinoma cell lines and normal endometrial tissue, as well as RNA-seq following exogenous exposure to estrogen for the cell lines. Thus, this study provides a single-nuclei atlas of endometrioid adenocarcinoma and suggests that obesity likely both amplifies estrogenic signaling and qualitatively reshapes its regulatory context.

Graphical abstract

Highlights

• •Single-nucleus RNA-seq reveals obesity-associated changes across cell populations
• •Obesity reshapes ESR1 co-expression networks in endometrial tumor cells
• •Network rewiring is reproducible in an independent bulk RNA-seq cohort
• •ESR1-centered modules are enriched for ESR1 binding and estradiol-responsive genes

Microenvironment; Bioinformatics

Introduction

Endometrial cancer is the most common gynecologic malignancy in the United States, with over 69,000 new cases projected in 2026.¹ Its incidence has risen in parallel with the obesity epidemic, particularly for endometrioid adenocarcinoma—the most common histological subtype, which is largely considered hormonally driven by estrogen.^2,3 The most frequently cited risk factors disrupt the balance between estrogen’s proliferative effects and progesterone’s inhibitory effects on the endometrium, including obesity, tamoxifen use, anovulation, and estrogen-only hormone replacement therapy.^4,5,6 Notably, obesity perturbs this hormonal balance through aromatase activity in peripheral adipose tissue, leading to elevated circulating estrogen levels that are not counterbalanced.⁷

In postmenopausal women, who comprise the majority of patients, peripheral adipose tissue is the dominant source of estrogen.^8,9,10 Though early abnormal uterine bleeding often leads to timely diagnosis and curative hysterectomy, rising obesity rates have coincided with increased diagnoses of aggressive, high-grade tumors and an annual 2.2% incidence growth among women under 50.^9,11,12 This presents a growing clinical challenge given the reliance on early-stage detection due to early presentation and non-fertility-sparing surgery.¹³ Despite this, the consequences of obesity-associated estrogen excess for ESR1-mediated transcriptional coordination remain poorly understood.

Thus, we systematically evaluated the tumor transcriptome across cell populations using single-nucleus RNA sequencing in tumors from patients with normal and obese body mass indices. Given the strong biological connection between estrogen signaling and BMI in endometrioid adenocarcinoma, we used co-expression network analysis to compare ESR1-centered gene networks across BMI groups to determine whether obesity is associated with coordinated rewiring of estrogen-linked gene networks in tumor epithelial (TE) cells. We further validated these networks by intersecting co-expression modules with ESR1 chromatin binding and estrogen-responsive gene expression profiles, supporting these genes as bona fide ESR1 targets.

Results

Single-nucleus analysis reveals distinct cell populations within the endometrioid adenocarcinoma tumor microenvironment

To investigate the influence of obesity on the tumor immune microenvironment and transcriptomic landscape of endometrioid adenocarcinoma, we conducted single-nucleus RNA sequencing on snap-frozen primary tumors from eight obese and seven normal-weight, postmenopausal patients. The clinical characteristics of the patients and the clinicopathological features of their tumors are detailed in Table S1. Following quality control, one obese and one normal-weight sample were excluded due to insufficient nuclei counts based on quality control metrics (Figures S1, S2, S3, and S4), resulting in a final dataset comprising 42,574 nuclei across the remaining 13 tumors.

Data were analyzed using Seurat and Harmony, with integration performed to account for batch effects and enable a unified representation of nuclei transcriptomes. Clustering followed by UMAP visualization identified six major nuclei populations: ciliated epithelial, endothelial, lymphoid, myeloid, stromal/myometrial, and TE (Figure 1A). Cell identities were assigned based on canonical marker gene expression (Figure 1B) and independently validated using SingleR against reference transcriptomic profiles (Figure S5).

Figure 1.

Single-nuclei characterization of endometrioid adenocarcinoma

(A) UMAP of 42,574 nuclei from 13 primary tumors, colored and labeled by six major cell types: ciliated epithelial, endothelial, lymphoid, myeloid, stromal/myometrial, and tumor epithelial.

(B) Expression of canonical marker genes for cluster annotation: DNAH12 (ciliated epithelium), VWF (endothelium), IKZF1 (lymphoid), MS4A6A (myeloid), COL3A1 (stromal/myometrial), and GRIA2 (tumor epithelial).

(C) UMAPs visualizing the integrated dataset colored by clinical and pathological variables (race, patient of origin, tumor grade, BMI) to assess potential batch effects.

(D) Composition of each tumor sample by cluster.

(E) Distribution of body mass index group across each identified cluster. n = 13 tumors (7 obese, 6 normal-weight), 42,574 nuclei total. No formal statistical tests applied. See also Figures S1, S2, S3, S4, and S5.

We next assessed whether clinical covariates, including self-identified race, patient of origin, tumor grade, and BMI, influenced cluster structure. No clusters were exclusively driven by any single variable, supporting effective data integration and indicating that cluster identities reflect underlying biological heterogeneity rather than technical artifacts (Figure 1C).

Analysis of cell type composition showed that TE cells predominated in most samples (Figure 1D), while the relative abundance of immune and stromal populations varied across tumors, indicating substantial inter-patient heterogeneity in the tumor microenvironment. Both obese and normal-weight groups contained sufficient nuclei from each major cell type to support downstream differential expression analyses (Figure 1E). Together, these analyses establish a single-nuclei atlas of endometrioid adenocarcinoma that captures inter-patient heterogeneity while minimizing technical and clinical confounding.

Obesity leads to gene expression and immune-metabolic signaling changes across tumor microenvironment cell types

To assess the broader impact of obesity on the tumor microenvironment, we examined gene expression changes across five additional cell types beyond TE cells. Differential expression analysis revealed significant obesity-associated transcriptional alterations in every examined cluster, indicating a widespread effect of obesity on the tumor microenvironment (Figures S6A–S6E). Genes were considered differentially expressed based on a threshold of log₂ fold change >1 and adjusted p value <0.05.

Pathway-level enrichment analysis using single-cell pathway analysis (SCPA; Wilcoxon-based) of Hallmark gene sets further highlighted cell type-specific effects of obesity. Across multiple cell types, we observed signatures of immune dysregulation and metabolic shifts (Figure S6F). These findings suggest that obesity, and obesity-related factors, lead to shifts in immune and metabolic gene expression programs, potentially contributing to a tumor-permissive microenvironment.

To determine whether obesity reprograms intercellular communication within the tumor microenvironment, we applied CellChat to infer ligand-receptor signaling networks in tumors from obese and normal-weight patients. Differential interaction heatmaps (Figure S6G) revealed modest but specific changes in both the number and strength of signaling interactions in obese tumors compared to normal-weight controls. These findings indicate that obesity may modestly impact specific cell-cell communication pathways in the tumor microenvironment, promoting a more interconnected and potentially pro-tumorigenic microenvironment. Thus, obesity is associated with widespread but cell type-specific transcriptional and signaling alterations across the tumor microenvironment, extending beyond tumor epithelial cells.

Tumor epithelial subclusters include proliferative and stem-like populations

We next focused on the TE compartment, which comprised seven transcriptionally distinct subclusters that were merged into a single epithelial cluster for high-level annotation (Figure 2A), while retaining subcluster structure to examine functional heterogeneity. Marker gene analysis identified subcluster-defining transcripts (Table S2), and Seurat’s cell cycle scoring revealed that subcluster 4 was highly enriched for nuclei in the S and G2/M phases, consistent with active proliferation (Figure 2B). SingleR reference-based labeling further suggested that many nuclei in subcluster 4 resembled neuroepithelial or neuronal lineages, as previously reported in aggressive uterine tumors such as uterine serous carcinoma.¹⁴

Figure 2.

Characterization of the tumor epithelial cluster

(A) UMAP displays seven subclusters within tumor epithelial nuclei prior to merging for downstream analysis.

(B) Cell cycle phase scoring (G1, S, G2/M) using Seurat’s cell cycle workflow.

(C) Predicted pseudotime trajectory from Monocle 3, with subcluster 4 positioned at the root using get_earliest_principal_node().

(D) Enrichment of Hallmark pathways G2M checkpoint and E2F targets in subcluster 4 based on Seurat FeaturePlot visualization.

(E) Distribution of classical cancer stem cell markers (BMI1, CD44, POU5F1 [OCT4], PROM1 [CD133]) across tumor epithelial nuclei. Trajectory and pathway analyses were conducted without formal statistical testing. See also Figure S2 and Table S2.

To investigate differentiation hierarchies, we used Monocle 3 pseudotime trajectory analysis, which positioned subcluster 4 at the root of the trajectory based on the selection of the earliest principal node (get_earliest_principal_node()), consistent with a less differentiated or stem-like state (Figure 2C). Enrichment of Hallmark gene sets related to proliferation, including G2M checkpoint and mitotic spindle programs, supported high mitotic activity within this population (Figure 2D). However, classical stemness markers,¹⁵ including CD44, PROM1 (CD133), POU5F1 (OCT4), and BMI1, were not specifically enriched in subcluster 4, despite its strong proliferative state (Figure 2E).

These findings suggest that while subcluster 4 is highly proliferative and occupies an early pseudotemporal position, it does not represent a transcriptionally distinct cancer stem cell population based on canonical markers, but instead reflects a transit-amplifying state.¹⁶ The absence of clear stemness marker enrichment suggests that proliferative epithelial states reflect transcriptional plasticity rather than a discrete stem cell compartment. Alternatively, this may reflect tumor-to-tumor variability despite integration minimizing technical batch effects. These observations underscore the need for functional validation to define stem-like populations and provide context for subsequent analyses of coordinated gene expression programs.

Obesity is associated with coordinated estrogen and metabolic transcriptional changes in tumor epithelial cells

To determine whether obesity-associated differences extend beyond composition to epithelial state programs, we performed differential expression analysis on the TE cluster, identifying genes differentially expressed in TE nuclei between obese and normal-weight patients. Differential expression was performed at the single-nucleus level using Seurat’s FindMarkers() function (Wilcoxon rank-sum test); to assess robustness to patient-level aggregation, we additionally performed pseudobulk analysis, which showed concordant directionality of the major differential expression patterns. A volcano plot highlights significant changes (Figure 3A), and subsequent pathway analysis revealed upregulation of hallmark pathways including estrogen response (early and late), epithelial-to-mesenchymal transition (EMT), fatty acid metabolism, TNFα signaling, hypoxia, and KRAS signaling (Figure 3B). Proliferation-associated pathways identified in subcluster analysis were not among the most differentially expressed programs between BMI groups, suggesting that proliferative epithelial states are present across tumors regardless of BMI rather than being obesity-specific.

Figure 3.

Obesity leads to gene expression changes in tumor epithelial nuclei

(A) Volcano plot of differentially expressed genes between tumor epithelial nuclei from obese vs. normal-body mass index patients. Green dots denote genes with log₂ fold change >1 and adjusted p < 0.05 (Wilcoxon rank-sum test, Bonferroni-corrected).

(B) Heatmap shows differentially enriched Hallmark pathways identified using single-cell pathway analysis (SCPA), using Wilcoxon-based scoring.

(C) Comparison of the top 50 hub genes in the estrogen receptor alpha (ESR1) modules identified in obesity- and normal-weight-specific WGCNA. Gene-level differential expression overlaid (blue = obese > normal-weight, red = normal-weight > obese).

(D) ESR1-containing module hub-rank reshuffling across BMI groups. Comparison of intramodular connectivity ranks (kME) between normal-BMI and obese tumors in TCGA bulk RNA-seq.

(E) Cross-platform concordance of obesity-associated ESR1 network rewiring. Bulk-defined ESR1-network rewiring genes exhibit elevated rewiring percentiles in epithelial snRNA-seq compared to size-matched random gene sets (p = 0.0044 across 10,000 permutations).

(F) Observed rewired-edge network formed by filtering identified ESR1-associated genes retained across tumor epithelial snRNA-seq and bulk RNA-seq (top) compared to a degree-matched random network (bottom).

Notably, VIM, a key marker of EMT, was upregulated in TE nuclei from obese compared to normal-weight patients, supporting enhanced mesenchymal features at the epithelial compartment level.¹⁷ HIF1A, a central regulator of the hypoxic response, was also increased, consistent with pathway-level enrichment.¹⁸ Additionally, OVGP1, involved in epithelial differentiation and known to interact with ERα signaling, was elevated.¹⁹ Among differentially expressed transcription factors linked to ESR1-associated signaling, FOSL2 was upregulated while JUN was downregulated, suggesting complex modulation of ESR1-associated gene networks in obesity.^20,21 Together, these results show that obesity is associated with coordinated shifts in estrogen-linked, metabolic, and stress-related transcriptional programs within TE cells.

ESR1-centered co-expression shows partial preservation with BMI-associated rewiring in tumor epithelial cells

Given the central role of estrogen signaling in endometrioid adenocarcinoma and its strong association with BMI, we examined differences in estrogen-related gene co-expression networks in TE cells. Using hdWGCNA, we independently constructed gene co-expression networks from single-nucleus RNA-seq data by grouping genes with correlated expression across TE nuclei from obese and normal-BMI patients (Figure S7).

In each co-expression network, we identified a module, or a group of genes whose expression levels are more strongly correlated with each other than with genes outside the group, that contained ESR1 as a highly connected gene. The stronger eigengene-based connectivity (kME) of ESR1 in tumors from patients with obesity (normal BMI kME = 0.25; obese BMI kME = 0.48) suggests a more central and coordinated role in obese tumors. While kME is traditionally defined in bulk co-expression networks, here it is used comparatively to assess relative differences in network organization between BMI groups.

To examine this, we used rank-based hub overlap and hypergeometric tests as descriptive and heuristic measures to compare the ESR1-containing modules across BMI cohorts since our snRNA-seq cohort contained few biologically independent tumors and gene-gene correlations within modules violate independence assumptions. Looking at the genes most central to the ESR1-containing modules in each network, only ESR1 itself was shared among the top 25 kME-ranked genes (9.0-fold enrichment, hypergeometric p = 0.11). Expanding to the top 50 kME-ranked genes revealed similarly limited overlap, although more than expected by chance (6.6-fold enrichment, hypergeometric p = 0.01). Even amongst the top 100 kME-ranked genes, only six genes overlapped (3.4-fold enrichment, hypergeometric p = 0.009).

To complement this approach, we also examined module preservation statistics (1,000 permutations). Using the obese network as the reference, given its larger number of nuclei, the composite Zsummary statistic, which integrates standardized measures of module density and intramodular connectivity preservation, revealed that the module containing ESR1 in the normal BMI network exhibited only moderate preservation (Zsummary = 5.31). The statistic was largely driven by preservation of module density (Zdensity = 8.91), as there was weak preservation of connectivity patterns (Zconnectivity = 1.7). Together, these results, including differences in kME, limited overlap of highly connected genes, and weak preservation of intramodular connectivity, suggest that while the ESR1 signaling program remains broadly intact across BMI groups, shifts in BMI substantially reconfigure the genes that regulate and mediate its downstream effects.

To determine whether the observed differences reflected true network rewiring rather than differential expression artifacts, we examined the direction and significance of differential expression for each of the top 50 kME-ranked genes in the ESR1-containing modules (Figure 3C). Each module included genes upregulated in patients with obesity, genes upregulated in normal-BMI patients, and genes without significant differential expression, suggesting that differences in ESR1-associated co-expression are not explained by uniform shifts in gene expression alone.

ESR1-linked co-expression structure is conserved yet reconfigured across BMI groups in bulk tumors

Preservation testing, using the more robust obese cohort as reference, showed strong module preservation (Zsummary = 11.0). However, because bulk RNA-seq aggregates cell-type heterogeneity and tumor purity into shared correlation structure, it can artificially inflate preservation statistics,²² potentially obscuring underlying differences in network organization between BMI groups. We therefore next examined ESR1-associated network properties that are more robust to such averaging.

There was little gene-gene overlap between the ESR1-containing modules identified in obese and normal-BMI bulk tumors (Jaccard index ≈0.1), consistent with our snRNA-seq results. Consistent with that, hub-rank preservation analyses further clarified how ESR1-associated networks differ across BMI strata. Comparing intramodular hub rankings, which evaluate how central a gene is within a specific co-expression module, between obese and normal-BMI networks showed little preservation across the complete set of ESR1-associated genes identified in either network (Spearman ρ ≈ 0.08; Kendall τ ≈ 0.06). In contrast, restricting the comparison to genes shared between both BMI-specific networks revealed modest but non-random preservation of hub ordering (Spearman ρ = 0.33; Kendall τ = 0.24; n = 74), supporting partial conservation of an estrogen signaling core with BMI-dependent network reconfiguration (Figure 3D).

To validate that BMI-associated ESR1 network differences were not an artifact of module definitions or hub thresholds, we compared the full ESR1-associated gene-gene correlation structure between obese and normal-BMI tumors. Overall matrix similarity was modest (Spearman ρ = 0.25) and significantly lower than expected under permutation of BMI labels (p = 5 × 10⁻⁴), with increased matrix deviation as measured by the Frobenius norm (p = 0.023), which captures absolute differences in correlation strength between matrices (Figure S8). Nonetheless, the majority of ESR1–gene correlations (∼83%) retained the same direction across BMI groups, indicating that estrogen signaling remains intact while the pattern of ESR1 coordination is selectively reorganized with BMI (Figure S9).

Obesity-associated ESR1 network rewiring generalizes across cellular and bulk contexts

We next asked whether obesity-associated ESR1-network rewiring, defined by BMI-dependent changes in the pattern or strength of co-expression between ESR1 and other ESR1-associated genes, was concordant between bulk tumors and TE cells profiled by snRNA-seq. Because bulk RNA-seq aggregates signals across multiple cell types and tumor purity, whereas snRNA-seq captures epithelial cell-intrinsic expression, gene-gene correlation structure reflects different sources of variation across platforms. In addition, the sets of genes included in each network differ due to platform-specific detection and feature selection, further limiting direct correspondence. Consequently, direct gene-by-gene comparison of network features is not expected. We therefore tested whether genes identified as rewired in the bulk network were non-randomly enriched in epithelial regions of high rewiring in the snRNA-seq network, anchoring the analysis in the bulk-defined signal rather than performing a reciprocal comparison across non-equivalent feature spaces.

Bulk RNA-seq-defined rewiring genes were distributed toward higher rewiring percentiles in the snRNA-seq epithelial network (median = 0.60), indicating that genes identified as rewired in bulk tend to occupy regions of elevated rewiring in the epithelial network. This pattern was robust to permutation testing, with the observed mean rewiring percentile exceeding that of most size-matched random gene sets (p = 0.0044), and was further supported by a rank-based Wilcoxon test (two-sided p = 0.007; one-sided p = 0.003). Together, these results indicate partial, statistical concordance in rewiring signal across platforms, rather than direct gene-by-gene or network-level correspondence, despite differences in cellular resolution and measurement modality (Figure 3E).

To derive a gene-level summary of cross-platform ESR1 network rewiring, we examined BioGRID-annotated ESR1 interaction partners and known cofactors.²³ Within this context, we restricted the analysis to genes exhibiting at least one significant rewired co-expression relationship (i.e., rewired edge) in both TE snRNA-seq and bulk RNA-seq analyses.

As expected, the number of detectable rewired co-expression relationships per gene was substantially lower in bulk tumors than in epithelial single-nucleus profiles, reflecting both cell-type averaging and the conservative bias of bulk RNA-seq toward network preservation. Accordingly, we summarized rewiring at the gene level based on the number of incident significant relationships. Rather than treating the inferred network as a fixed biological entity, this approach focuses on identifying a small set of high-confidence genes that retain detectable rewiring across platforms (Table 1). These genes therefore represent robust candidates associated with ESR1 network reorganization in obesity.

Table 1.

Cross-platform filtered set of ESR1-associated genes show significant co-expression rewiring in both epithelial single-nucleus and bulk tumor RNA-seq data

Gene	# Rewired edges Single nuclei RNAseq	# Rewired edges Bulk RNAseq
ALDH3A2	925	2
ARFGAP3	766	2
DNMBP	319	1
ESR1	890	10
MYOF	517	1
NFIB	1013	2
PBX1	515	7
PGR	913	2
PIK3R1	959	3

We next asked whether genes retained across platforms reflected coordinated ESR1 network reorganization or merely isolated rewired gene pairs. Compared to degree-matched random gene sets, these genes exhibited significantly greater internal connectivity and higher overall degree within the rewired-edge network (Figure 3F), indicating enrichment for centrally positioned nodes and supporting coordinated, rather than purely pairwise, network remodeling. As a result, these genes are more likely to reflect biologically meaningful ESR1-associated processes perturbed by obesity and thus represent higher-priority candidates for downstream functional investigation.

BMI stratifies coordinated transcriptional programs across tumor epithelial cells

To examine BMI-associated differences in coordinated gene expression programs, we pooled the BMI cohorts from our snRNA-seq dataset and constructed a new hdWGCNA network. In contrast to the BMI-stratified networks used above to assess network rewiring, this combined network enables direct comparison of module activity across BMI groups. Modules were defined by constructing a signed WGCNA network on pooled samples, clustering genes based on topological overlap, and identifying modules via dynamic tree cutting with eigengene-based merging, after filtering for variable genes and retaining key genes such as ESR1. This enabled direct comparison of module eigengenes, which represent the aggregate expression behavior of co-expressed gene sets and provide a quantitative measure of pathway-level activity, across BMI groups. This analysis identified multiple co-expression modules that were either upregulated or downregulated in obesity, including a TE module (TE7; module 7 identified by hdWGCNA) in which ESR1 is a top hub gene and whose eigengene differed significantly between BMI groups (Figure 4A). An analogous eigengene shift for this module was observed in the TCGA bulk RNA-seq dataset (Figure 4B).

Figure 4.

BMI stratifies coordinated transcriptional programs across tumor epithelial cells

(A) Each point represents a tumor epithelial hdWGCNA co-expression module summarized by its eigengene in the snRNA-seq dataset. The x axis shows the average eigengene difference between obese and normal-BMI groups (Obese − Normal), and the y axis shows statistical significance (−log₁₀ FDR) from a Wilcoxon rank-sum test with Benjamini-Hochberg correction. Points are colored by module assignment; the vertical dashed line indicates no eigengene difference, and the horizontal dashed line marks the −log₁₀(FDR) = 15 threshold.

(B) Each point represents a WGCNA co-expression module from the bulk RNA-seq summarized by its eigengene. The x axis shows the average eigengene difference between obese and normal-BMI samples (obese − normal), and the y axis shows statistical significance (−log₁₀ FDR) from a Wilcoxon test with Benjamini-Hochberg correction. Points are colored by module assignment. The vertical dashed line indicates no eigengene difference, and the horizontal dashed line marks the FDR = 0.05 threshold.

(C) Euler diagram shows the overlap between DNA replication/repair-associated genes identified in tumor epithelial snRNA-seq hdWGCNA module TE10 and bulk RNA-seq WGCNA module ME4. Counts indicate exclusive and shared genes within the shared cross-platform universe; overlap significance was assessed by hypergeometric testing using the restricted universe.

(D) Euler diagram shows the overlap between ESR1-containing modules from tumor epithelial snRNA-seq hdWGCNA module TE7 and the corresponding ESR1-associated bulk RNA-seq WGCNA module ME7 (module numbering is dataset-specific and does not imply direct correspondence). Numbers denote exclusive and shared genes within the shared gene universe used for enrichment testing.

We next assessed whether the differences in module eigengene expression between obese and normal-BMI groups for modules exhibiting BMI-associated eigengene shifts were concordant across datasets. A DNA replication/repair module was consistently higher in normal-BMI tumors across platforms, with the bulk module significantly overlapping the snRNA-seq module, sharing 67 genes—nearly 2-fold more than expected by chance (fold enrichment = 1.95; hypergeometric p = 1.16 × 10⁻¹⁰) (Figure 4C). In contrast, an ESR1-related module was consistently higher in obese-BMI tumors across platforms. The snRNA-seq ESR1 module (284 genes) showed strong enrichment within the bulk ESR1 module (172 genes), sharing 60 genes, over 3-fold more than expected by chance (fold enrichment = 3.30; hypergeometric p = 3.7 × 10⁻¹⁹) (Figure 4D). Thus, BMI-associated differences are reflected in coordinated transcriptional programs, including estrogen receptor-associated modules, and are reproducibly observed across datasets.

Functionally distinct ESR1 modules differ in ERα binding profiles

To determine whether these ESR1-associated genes exhibited evidence of ERα binding, we examined ERα ChIP-seq data from both estrogen-treated endometrial cancer cell lines and human endometrial biopsies collected across menstrual cycle phases (GSE200802).²⁴ This analysis allowed us to assess whether genes prioritized by integrative co-expression network analyses were bound by ERα in the intact endometrium under physiologic hormonal conditions, independent of experimental perturbation. ESR1-associated genes identified from both snRNA-seq and bulk RNA-seq networks showed widespread and significant enrichment for ERα binding across menstrual cycle states, including enrichment for estrogen-inducible ERα binding sites (Figures S10 and S11), indicating that many prioritized genes are likely direct ERα targets in vivo. Establishing ERα binding in patient tissue provided a physiologic foundation for subsequent analyses in model systems, motivating our examination of ERα chromatin-binding dynamics in estrogen-treated endometrioid adenocarcinoma cell lines.

We next sought to functionally ground WGCNA-defined modules in estrogen receptor biology by examining ERα chromatin binding in endometrioid adenocarcinoma models. To this end, we integrated previously published ERα ChIP-seq data from estrogen- and vehicle-treated Ishikawa and HCI-EC-23 cell lines (GSE210124).²⁵ Across snRNA-seq modules, several showed significant enrichment for ERα binding, including TE1, TE5, TE6, and our a priori module of interest, TE7. Notably, TE7 exhibited particularly strong enrichment for estrogen-inducible ERα binding (E2-gained peaks; 1.8-fold enrichment, hypergeometric p = 1.58 × 10⁻¹³) (Figure 5A).

Figure 5.

Validation of estrogen association for genes in the ESR1-containing modules

(A) Fold enrichment of ESR1 cell line ChIP-seq-associated genes within tumor epithelial (TE) modules is shown for all estrogen-bound sites (E2 union; triangles) and estrogen-inducible sites (E2-gained; circles). Enrichment significance was assessed using a one-sided hypergeometric test with Benjamini-Hochberg correction across modules. Point size reflects the number of overlapping genes, and color encodes –log₁₀(FDR). The dashed vertical line indicates no enrichment (fold enrichment = 1).

(B) Fold enrichment of ESR1 cell-line ChIP-seq-associated genes within tumor epithelial (ME) modules is shown for all estrogen-bound sites (E2 union; triangles) and estrogen-inducible sites (E2-gained; circles). Enrichment significance was assessed using a one-sided hypergeometric test with Benjamini-Hochberg correction across modules. Point size reflects the number of overlapping genes, and color encodes –log₁₀(FDR). The dashed vertical line indicates no enrichment (fold enrichment = 1). Modules at or below this threshold (left of the dashed line) show little or no enrichment, and interpretation prioritizes modules with both statistically significant enrichment and substantial gene overlap.

(C) Dot plot shows fold enrichment (observed/expected) of MSigDB Hallmark gene sets among genes in bulk RNA-seq ESR1-associated modules ME7 (circles) and ME10 (triangles). Enrichment significance was assessed using over-representation analysis with hypergeometric testing and Benjamini-Hochberg correction. Point size reflects the number of overlapping genes (k), and color encodes −log₁₀(FDR). The dashed vertical line indicates no enrichment (fold enrichment = 1).

(D) Gene set enrichment analysis (fgsea) shows the enrichment of bulk RNA-seq modules ME7 (black) and ME10 (purple) along a ranked list of genes ordered by estrogen-induced differential expression. The running enrichment score is shown in green, with vertical tick marks indicating positions of module genes in the ranked list. Normalized enrichment score (NES) and false discovery rate (FDR) are indicated for each module, demonstrating significant estrogen-responsive transcription for ME7 but not ME10.

Analysis of bulk RNA-seq modules revealed a corresponding enrichment pattern in the ESR1-associated module, ME7. In contrast, another module, ME10, exhibited greater overall ERα binding enrichment relative to module size and a comparable number of genes with associated ERα peaks (Figure 5B), but did not display a significant BMI-associated eigengene shift (Figure 4B). To further characterize differences between these modules, we compared their functional pathway composition. ME10 contained genes spanning a broad range of hallmark pathways, whereas ME7 was more narrowly enriched for genes associated with hormonal response pathways (Figure 5C).

Only a subset of ERα-bound modules exhibits coordinated estrogen-induced transcription

To assess whether ESR1-associated modules participate in coordinated estrogen-responsive transcription, we examined transcriptomic data from Ishikawa and HCI-EC-23 endometrioid adenocarcinoma cell lines treated with estrogen or vehicle (GSE210123).²⁵ Genes were ranked by estrogen-induced expression changes, and module-level enrichment was evaluated using a rank-based gene set enrichment approach.

The ESR1-containing module, ME7, showed significant positive enrichment among estrogen-induced genes (NES = 1.56, FDR = 0.03), indicating that genes within this module tended to be coordinately upregulated following estrogen stimulation (Figure 5D). This enrichment was driven by a subset of module members that exhibited concordant estrogen-responsive expression changes rather than by a small number of highly responsive genes.

In contrast, ME10 showed no significant enrichment for estrogen-induced transcriptional changes (NES = −1.33, FDR = 0.16), despite containing a comparable number of ERα-bound genes identified by ChIP-seq. Thus, modules differed in their transcriptional response to estrogen stimulation, with coordinated estrogen-induced expression observed for ME7 but not for ME10. Furthermore, ERα binding alone is insufficient to define a functional estrogen-responsive network, with coordinated estrogen-induced transcription restricted to a specific co-expression context.

Convergent multi-modal evidence defines a core ESR1 regulatory network

We next integrated ESR1-associated co-expression modules identified from snRNA-seq and bulk RNA-seq with ERα ChIP-seq binding data and estrogen-induced transcriptional response profiles to identify genes supported by multiple independent lines of evidence. By requiring concordance across expression-based network membership and at least one functional readout of ESR1 activity, this approach prioritized genes most consistently associated with ESR1-linked regulation in endometrioid adenocarcinoma.

This integrative analysis identified a core set of genes with convergent support for ESR1 association, including BMPR1B, CPM, EPB41L2, ESR1, FBLN1, MECOM, PGR, and PLCB1. These genes were present in ESR1-associated modules in both snRNA-seq and bulk RNA-seq analyses and were additionally supported by either ERα chromatin binding or estrogen-responsive transcriptional activity. A complete list of genes meeting these criteria is provided in Table 2.

Table 2.

Datasets shared module genes were identified in

Gene	Normal mid Secretory ChIP	Normal proliferative ChIP	Ishikawa ChIP	HCI-EC-23 ChIP	E2 upregulated RNA-Seq
ADCY1	✓	✓	✓	–	–
ADAMTS6	✓	✓	–	✓	–
ANAPC4	✓	✓	✓	–	–
ARHGAP26	✓	✓	✓	–	✓
BMPR1B	✓	✓	✓	✓	–
CPM	–	✓	✓	✓	✓
DGKH	✓	✓	✓	–	✓
DNAJC10	–	✓	✓	–	–
EHF	✓	✓	✓	–	✓
EPB41L2	✓	✓	✓	✓	✓
ESR1	✓	✓	✓	✓	✓
EYA2	✓	✓	–	–	✓
FBLN1	✓	✓	✓	✓	✓
FOSL2	–	✓	–	✓	–
MAST4	✓	✓	✓	–	✓
MECOM	✓	✓	✓	✓	–
NPAS3	–	✓	–	–	✓
PAM	✓	✓	✓	–	–
PGR	✓	✓	✓	✓	✓
PLCB1	✓	✓	✓	✓	✓
PODXL	✓	✓	✓	–	✓
PPP2R2C	✓	✓	–	✓	–
PTPRM	–	✓	–	✓	–
RHEX	✓	✓	✓	–	✓
SLC25A35	–	✓	–	–	✓
SLC40A1	✓	✓	✓	–	✓
SOX17	✓	✓	–	–	✓
SPATA13	–	✓	–	✓	–
STX18	✓	✓	✓	–	✓
WDR77	–	✓	–	–	✓
ZNF516	✓	✓	–	–	✓

Discussion

Obesity is a well-established risk factor for endometrioid adenocarcinoma, which is most commonly attributed to increased estrogen bioavailability.^4,5,6 In this study, we sought to move beyond this quantitative framing by examining how obesity is associated with changes in the organization of estrogen receptor-linked transcriptional networks within tumors. Using single-nucleus RNA sequencing, complemented by bulk RNA-seq, we show that BMI-associated obesity is associated with coordinated alterations in gene expression across TE, immune, and stromal compartments. These changes converge on pathways related to estrogen signaling, metabolism, inflammation, and angiogenesis—processes with well-established roles in endometrial cancer progression.

These findings are consistent with epidemiological evidence linking higher BMI to increased endometrial cancer risk and poorer outcomes, likely mediated in part by chronic estrogen excess.^26,27 However, our results suggest that obesity-associated estrogen exposure does not simply amplify a fixed ESR1 transcriptional program. Instead, across both snRNA-seq and bulk RNA-seq data, we observed preservation of an ESR1-centered co-expression framework alongside extensive redistribution of module membership, hub ranking, and intramodular connectivity. This pattern suggests that obesity is associated with the reorganization of estrogen-related gene networks, rather than wholesale disruption or sole amplification.

Multiple orthogonal analyses support this network-level reorganization. ESR1-associated modules were enriched for estrogen-inducible ESR1 chromatin binding and transcriptional responses in endometrioid adenocarcinoma cell lines, whereas an ESR1-bound but transcriptionally unresponsive module, ME10, provided an informative internal contrast. Despite exhibiting substantial ESR1 binding enrichment, ME10 showed neither a BMI-associated eigengene shift nor coordinated estrogen-induced transcriptional activation, indicating that ESR1 binding alone is insufficient to define a functional estrogen-responsive network.^28,29 Notably, no direct analog of ME10 was identified in the epithelial-only snRNA-seq data, consistent with bulk RNA-seq capturing broader mixed-state covariance that may support ESR1-bound but functionally heterogeneous modules. Together, these observations argue against a simple global, dose-dependent amplification of ESR1-associated biology in obesity and instead suggest that functional estrogen signaling is selectively engaged within specific co-expression contexts.

This dissociation between ESR1 chromatin occupancy and coordinated transcriptional output is consistent with context-dependent engagement of estrogen receptor cofactors and partner transcription factors, which govern enhancer selection and transcriptional competence downstream of ESR1 binding.^30,31,32 Such cofactor-dependent regulation provides a plausible mechanism by which ESR1-bound networks can be transcriptionally active or inert depending on cellular and metabolic context.

Integration of network analyses with ChIP-seq and transcriptional response data further identified a set of genes consistently associated with ESR1-linked regulation in endometrioid adenocarcinoma. Validation using ESR1 ChIP-seq data from human endometrial biopsies demonstrated that nearly all prioritized genes were bound by ESR1 in vivo under physiologic estrogen exposure, indicating that these networks reflect endogenous estrogen regulatory circuitry rather than tumor-specific inventions.²⁴ Although co-expression networks were not constructed from non-tumor endometrium, integration with ESR1 ChIP-seq data from normal endometrial tissue supports that these modules reflect endogenous estrogen regulatory programs, with obesity-associated factors reshaping their organization within tumors. This supports a model in which endometrial tumors, regardless of BMI, reutilize normal estrogen regulatory programs, with obesity-associated factors reshaping how these programs are deployed.

Conceptually, these findings suggest that sustained remodeling of estrogen receptor signaling networks can arise either from obesity-associated estrogen excess or from metabolic and inflammatory states that modulate estrogen receptor activity and transcriptional competence, leading to adaptive reorganization of ESR1-linked regulation over time. Such remodeling is consistent with observations from other chronically stimulated receptor systems, where sustained ligand exposure leads to feedback regulation, altered cofactor engagement, and redistribution of regulatory influence rather than linear increases in signaling output.^33,34,35,36 Notably, ER signaling is known to be plastic in breast cancer under prolonged SERM exposure, a phenomenon loosely analogous to the chronic, non-physiologic estrogenic environment associated with obesity, in that sustained ER perturbation leads to adaptive reorganization of downstream regulatory networks.^37,38 Thus, in this context, obesity-associated estrogen excess may act not merely as a stronger signal, but as a persistent perturbation that reshapes the ESR1 network over time.

Clinically, this framework may help explain heterogeneity in hormonal therapy responses among patients with endometrioid adenocarcinoma.^39,40 If obesity alters the organization of functional estrogen networks rather than simply increasing estrogen levels, this could influence responsiveness to progestins and other endocrine therapies. Although not directly tested here, emerging evidence that metabolic interventions—such as GLP-1 receptor agonists—can modulate hormone receptor signaling raises the possibility that targeting obesity-related pathways could indirectly reprogram tumor hormone responsiveness.⁴¹ Whether obesity-associated network adaptations are reversible, however, remains unknown.

Overall, our findings support a model in which obesity contributes to endometrial tumor progression by inducing adaptive reorganization of ESR1-associated gene networks under chronic estrogenic pressure. While our conclusions are necessarily network-level and primarily computational, the consistency of these patterns across platforms, cohorts, and orthogonal data types meaningfully constrains plausible mechanisms and provides a focused framework for future experimental studies.

Limitations of the study

The findings presented here should be interpreted in the context of several constraints that define the scope of inference. BMI was used as a pragmatic measure of obesity, but it does not capture adipose distribution or distinguish adiposity from lean mass—features that influence inflammation and metabolic signaling.^42,43,44,45 More refined measures of adiposity were unavailable in the analyzed cohorts, and future studies incorporating metabolic phenotyping will be needed to determine how specific aspects of obesity shape estrogen-associated transcriptional networks. Future studies should incorporate more refined adiposity metrics and larger cohorts to enable investigation of obesity’s impact across different classes, especially given the known 50% increased risk of endometrioid adenocarcinoma per 5-unit BMI increase.²⁶

Although all tumors analyzed were classified as endometrioid adenocarcinomas, residual heterogeneity within this histologic subtype—including TCGA/ProMisE molecular subtypes, tumor grade, and stage—could not be explicitly modeled.^46,47 This reflects patient availability rather than study design. In particular, the number of normal-BMI tumors was small (n = 23 in the TCGA bulk RNA-seq cohort, with similarly limited representation in the snRNA-seq cohort), making stratified analyses statistically impractical and, in some cases, infeasible due to incomplete annotation. Consequently, BMI-associated network patterns identified here should not be interpreted as independent of subtype, grade, or stage. Rather, they may reflect direct effects of BMI, systematic coupling between BMI and tumor characteristics, or interactions among these factors. Nevertheless, genes prioritized by cross-platform network centrality and rewiring are expected to represent core regulatory nodes in endometrioid adenocarcinoma, irrespective of whether BMI acts as a direct driver or a correlated modifier of tumor state.

Functional integration of ESR1 binding relied on ChIP-seq data from two estrogen-responsive endometrioid adenocarcinoma cell lines, both of which are mismatch repair-deficient.²⁵ This reflects a broader constraint in the field, as relatively few endometrial cancer cell lines robustly express ESR1. In addition, while Ishikawa cells are widely used as a canonical estrogen-responsive model, they harbor TP53 loss, which may influence transcriptional responses to estrogen.⁴⁸ These factors limit the extent to which cell line-based ChIP data capture the full diversity of ESR1 regulatory behavior across endometrial tumors. However, convergence of ChIP-seq, transcriptional response data, and network analyses across snRNA-seq and bulk RNA-seq mitigates reliance on any single model system.

Several technical considerations further shape interpretation. Endothelial cell populations were derived primarily from a single patient per BMI group, limiting generalization of endothelial-specific findings. Stromal and myometrial cells clustered together due to transcriptional similarity, restricting resolution of their distinct roles. While snRNA-seq minimizes dissociation-induced artifacts, it captures nuclear transcripts rather than whole-cell expression, potentially underrepresenting cytoplasmic and mitochondrial programs relevant to metabolic remodeling. Immune cell analyses were also constrained by low nuclei counts, consistent with the immunologically cold nature of endometrioid adenocarcinoma.

Finally, the network-level alterations described here are observed within established tumors. It therefore remains unclear whether these ESR1 network adaptations contribute to increased risk of developing endometrial cancer or instead reflect obesity-associated modulation of tumor behavior after tumor initiation. In addition, correlation-based network approaches such as WGCNA capture coordinated transcriptional structure rather than direct regulatory interactions; inferred rewiring reflects changes in coordinated gene behavior and network organization, not explicit gains or losses of direct ESR1 regulation.⁴⁹ Because bulk RNA-seq data are inherently biased toward network conservation due to stromal and immune admixture and tumor purity variability, detection of BMI-associated shifts in ESR1 module activity and hub prioritization under this conservative bias likely underestimates the magnitude of epithelial-specific network differences.²² Experimental interrogation of inferred network rewiring, through chronic estrogen exposure models, cofactor manipulation, or obesity-mimetic systems, will be required to determine functional consequences.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Mike R Wilson (wilsonmr@wayne.edu).

Materials availability

The raw single-nucleus sequencing data generated in this study have been deposited in GEO: GSE300407. ChIP-seq data generated from human endometrial biopsies collected across menstrual cycle phases can be downloaded through GEO: GSE200802. ESR1 ChIP-seq data from estrogen- and vehicle-treated Ishikawa and HCI-EC-23 cell lines is publicly available on GEO: GSE210124. Transcriptomic data from Ishikawa and HCI-EC-23 endometrioid adenocarcinoma cell lines treated with estrogen or vehicle can be downloaded via GEO: GSE210123.

Data and code availability

• •Single-nucleus RNA-seq data have been deposited at GEO: GSE300407 and are publicly available as of the date of publication.
• •Previously published data analyzed in this study are available at GEO: GSE200802, GSE210124 and GSE210123.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work was funded by the National Cancer Institute Pathway to Independence Award R00 CA252152 (M.R.W.), Detroit Medical Center Foundation grant DMC-G-202323515 (M.R.W.), the Ruth L. Kirschstein National Research Service Award T32 CA009531 (J.L.L.), and funds from the Karmanos Cancer Institute Molecular Therapeutics Program. The Biobanking and Correlative Sciences Core is supported, in part, by NIH Center grant P30 CA022453 to the Karmanos Cancer Institute at Wayne State University. The Biostatistics and Bioinformatics Core is supported, in part, by NIH Center grant P30 CA022453 to the Karmanos Cancer Institute at Wayne State University. We thank Drs. Asfar Azmi, Victoria Bae-Jump, Hasan Korkaya, Larry Matherly, Izabela Podgorski, Adi Tarca, Kay-Uwe Wagner, Gen Sheng Wu, and Zeng-Quan Yang for their advice and support.

Author contributions

J.L.L., K.G., J.L.B., and M.R.W. designed the project. J.L.L., S.A., S.P., S.G., A.Q., M.F., K.G., J.L.B., and M.R.W. performed research. A.A. and C.F. contributed new reagents and methods. J.L.L. and M.W. analyzed data. M.G., M.F., K.G., A.G., G.D., M.J.C., and H.M.G. provided guidance on data analysis. R.T.M., J.J.W., and J.L.B. supplied clinical samples and information. J.L.L. wrote the manuscript. J.L.L., S.A., A.G., A.A., C.F., J.J.W., H.M.G., K.G., G.D., M.J.C., J.L.B., and M.R.W. edited the manuscript. M.R.W. supervised the project.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work, the authors used ChatGPT (OpenAI) for proofreading and occasional sentence rewording. The authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
	Karmanos Cancer Institute Biobank	Primary frozen endometrial tumors, IRB #080215M1F
Critical commercial assays
Chromium Nuclei Isolation Kit	10x Genomics	PN-1000493
Chromium Dissociation Tubes	10x Genomics	PN-2000564
Nuclei Isolation Columns	10x Genomics	PN-2000562
Deposited data
Single-nuclei RNA-seq dataset generated in this study	GEO	GSE300407
Previously published human endometrial ERα ChIP-seq dataset	GEO	GSE200802
Previously published Ishikawa/HCI-EC-23 RNA-seq dataset	GEO	GSE210123
Previously published Ishikawa/HCI-EC-23 ERα ChIP-seq dataset	GEO	GSE210124
Software and algorithms
R	R Foundation for Statistical Computing	v4.4.1
Seurat	Satija Lab	v5.3.0
Harmony	Raychaudhuri Lab	v1.2.3
Monocle 3	Trapnell Lab	v.1.3.7
clusterProfiler	Bioconductor	v4.13.0
SCPA	Bioconductor	v1.6.2
hdWGCNA	Swarup Lab	v0.3.03
CellChat	Nie Lab	v1.6.1
SingleR (celldex)	Bioconductor	v2.8.0
WGCNA	CRAN	V1.73
Biorender	Biorender.com	N/A

Experimental model and study participant details

Cell lines

Previously published ChIP-seq and RNA-seq datasets generated using Ishikawa and HCI-EC-23 cells (GSE210123 and GSE210124) were analyzed. HCI-EC-23 was reported by the original authors to be mycoplasma negative following IDEXX BioAnalytics testing; no mycoplasma testing information was reported for Ishikawa.

Human tumors

Primary snap-frozen endometrioid endometrial adenocarcinoma specimens were obtained from the Karmanos Cancer Institute Biobanking and Correlative Sciences Core. Tumors were collected from postmenopausal female patients and snap frozen at the time of collection, then stored in liquid nitrogen prior to nuclei isolation. Patients were selected to generate normal-weight (BMI <25 kg/m²) and obese (BMI ≥30 kg/m²) cohorts. Because normal-weight tumor specimens were relatively limited, cases were selected to provide the closest available matching between BMI groups based on available clinicopathologic characteristics. Treatment history was not available for all patients and was therefore not considered during sample selection.

The initial cohort consisted of 15 tumors from 8 obese and 7 normal-weight patients. Following quality-control filtering, one tumor from each BMI group was excluded due to insufficient nuclei counts, resulting in a final cohort of 13 tumors (7 obese and 6 normal-weight patients) comprising 42,574 nuclei. Patient ages ranged from 49 to 89 years. All tumors were classified as endometrioid adenocarcinoma and represented a range of FIGO grades and clinical stages. The cohort included patients self-identifying as White (n = 8), African American (n = 4), or unknown race (n = 1). Microsatellite instability (MSI) and mismatch repair (MMR) status were available for a subset of tumors. Additional clinicopathologic characteristics that were available are provided in Table S1.

All human tumor specimens were received de-identified from the KCI Biobank through IRB approved protocol #080215M1F. Written informed consent and HIPAA authorization were obtained from participants prior to specimen collection. All procedures were conducted in accordance with relevant institutional guidelines and applicable ethical regulations, including the principles outlined in the Declaration of Helsinki.

Because all participants were postmenopausal women with endometrial cancer, the influence of sex as a biological variable could not be evaluated in this study.

Method details

TCGA data

Data for endometrioid adenocarcinoma patients were retrieved from the TCGA cohort. Tumors were categorized based on patient BMI at the time of diagnosis. A total of 177 cases met the inclusion criteria, consisting of an obese group (n = 154) and a normal-BMI control group (n = 23). Patients with overweight BMI were excluded.

Tissue samples

Patient samples were provided by the Karmanos Cancer Institute Biobanking and Correlative Sciences Core. Tissues were snap frozen and stored in liquid nitrogen prior to nuclei isolation.

Single nuclei isolation

The Chromium Nuclei Isolation Kit (10x Genomics, PN-1000493) was used to isolate single nuclei. Tissue samples were thawed and dissociation tubes (PN-2000564) pre-chilled on dry ice. Samples were dissociated on wet ice using lysis buffer and pestle agitation until samples reached homogenization. Dissociated tissues were centrifuged in nuclei isolation columns (PN-2000562) for initial removal of cell debris. To remove remaining cell debris, pelleted nuclei were washed by resuspending nuclei pellet once in Debris Removal Buffer and three times in Wash and Resuspension Buffer, respectively, with centrifugation between washes. Nuclei were resuspended in Wash and Resuspension buffer on ice for further analysis. RNA-seq was performed by the Wayne State University Genome Sciences Core.

Single-nuclei RNA-seq data processing

After sequencing, the snRNA-seq library was demultiplexed using CellRanger mkfastq. Then, the CellRanger count was used to align the reference genome and tabulate counts by gene region. The cell-gene count matrix was converted to a Seurat object using the R Seurat package (v.5.3.0) for downstream analysis.⁵⁰ Cells with >5% mitochondrial gene expression and/or the number of genes identified in individual cells <200 or >2500 were filtered out. Normalization of the expression matrix to account for sequencing depth differences was performed using the “NormalizeData” function of Seurat. To generate a list of 2,000 highly variable genes, we used Seurat’s “FindVariableFeatures” function. Data was scaled using the “ScaleData” function of Seurat. After which, the data was subjected to linear dimensional reduction using Seurat’s “RunPCA” function. We used Harmony (v.1.2.3) and the top 20 principal components with a resolution of 0.5 to perform integration and clustering. The final dataset comprised 42,574 nuclei from 13 primary tumors (7 obese and 6 normal-weight). One obese and one normal-weight sample were excluded because they did not meet nuclei count or sequencing quality criteria.

Cluster classification

Differentially expressed genes between conditions in each cluster were identified using the “FindAllMarkers” function of Seurat, using default parameters. The annotation of cell types and subtypes was determined based on the expression of known canonical marker genes associated with respective cell types. Immune cells were subclustered using Seurat to identify more specific subtypes. To validate classifications, we used singleR with celldex “Human Primary Cell Atlas Data” and “Database Immune Cell Expression Data” to conduct an independent reference based annotation.⁵¹

Identification of differentially expressed genes

Differentially expressed genes between conditions in each cluster were identified using the “FindMarkers” function of Seurat, using default parameters. Pseudobulk aggregation was performed using Seurat to assess the concordance of differential expression results at the patient level.

Pathway analysis

Gene ontology (GO) analysis on differentially expressed genes was performed using clusterProfiler (v.4.13.0).⁵² Single Cell Pathway Analysis on differentially expressed genes was performed using the SCPA package (v.1.6.2).⁵³

Weighted gene correlation network analysis

Signed weighted gene co-expression networks were constructed separately for each BMI group using hdWGCNA (v0.3.03) to identify condition-specific gene modules.⁵⁴ Modules were compared between obese and normal-BMI groups. A pooled snRNA-seq network was subsequently generated and used for differential module eigengene analysis. WGCNA (v1.73) was then applied to the TCGA dataset following batch correction with limma (v3.62.2).^49,55 After variance-stabilizing transformation (VST), genes were filtered using a Median Absolute Deviation threshold (35%), while a curated list of biologically relevant genes (e.g., ESR1 and PGR) was retained. A signed co-expression network was constructed using Pearson correlation and a soft-thresholding power selected to satisfy the scale-free topology criterion (R² > 0.8). Modules were identified by hierarchical clustering of the Topological Overlap Matrix with a minimum module size of 30 and merged using a height cut-off of 0.25. Finally, snRNA-seq and TCGA co-expression modules were integrated with ESR1 ChIP-seq binding data and estrogen-induced transcriptional response datasets to identify a core set of genes supported by multiple independent lines of evidence.

CellChat

Cell-cell communication analysis was performed using the CellChat R package (v.1.6.1) to compare signaling interactions between obese and normal BMI conditions.⁵⁶ Single-cell RNA-seq data were preprocessed, and CellChat objects were constructed for each condition using the CellChatDB.human ligand-receptor database. Interactions were inferred by identifying overexpressed signaling genes, computing communication probabilities, and aggregating networks.

Cellular trajectories

Cellular trajectories and differentiation pathways in single-cell RNA sequencing data were analyzed using Monocle 3, a computational framework designed for pseudotime ordering and trajectory analysis.⁵⁷ Preprocessed and normalized data were used to construct a single-cell trajectory graph, employing Monocle 3’s reverse graph embedding method to identify developmental states and branch points. Differential gene expression analysis along identified trajectories was conducted to reveal key regulatory genes driving cellular differentiation.

Peak processing and annotation

ChIP-seq data for HCI-EC-23 and Ishikawa cell lines were analyzed using the ChIPseeker and rtracklayer R packages. Genomic coordinates were mapped to the hg38 reference genome. Treatment-specific peaks were identified by generating a union set of all detected peaks and constructing a presence/absence matrix. E2-gained peaks were defined as regions present in 1h E2-treated samples but absent in 8h DMSO controls. All peaks were annotated to the nearest gene using a promoter window of 3 kb around the Transcription Start Site.

ESR1 ChIP-seq and input control bigWig files for normal endometrium (Prolif, MS, d1, and d2 stages) were obtained from GSE200802. Files were converted to bedGraph format via bigWigToBedGraph and sorted by genomic coordinates. Peak calling was performed using SEACR (Sparse Enrichment Analysis for CUT&RUN/ChIP-seq) using a non-stringent threshold against the top 1% of enriched regions. The resulting BED files were imported into R as GRanges objects and annotated using ChIPseeker and the TxDb.Hsapiens.UCSC.hg38.knownGene database. Promoter regions were defined as the interval within 3 kb of the Transcription Start Site.

Quantification and statistical analysis

Software and computing environment

All statistical analyses were conducted in R (v4.4.1) using Seurat (v5.3.0), hdWGCNA (v0.3.03), WGCNA (v1.73), Monocle 3 (v1.3.7), clusterProfiler (v4.13.0), SCPA (v1.6.2), CellChat (v1.6.1), and SingleR (v2.8.0). Statistical tests, sample sizes, and measures of significance are reported throughout the Results and corresponding figure legends.

Differential gene expression and pathway analysis

Differential gene expression analyses were performed using Seurat’s FindMarkers() and FindAllMarkers() functions, which implement the Wilcoxon rank-sum test. Adjusted p values were calculated using the Bonferroni correction. Genes were considered significantly differentially expressed if they met both an adjusted p value <0.05 and an absolute log2 fold change >1. Gene ontology enrichment analyses were performed using clusterProfiler, which employs a hypergeometric test with Benjamini-Hochberg correction for multiple comparisons. Single-cell pathway enrichment was performed using SCPA, which uses Wilcoxon rank-sum testing to compare pathway activity between groups.

Co-expression network and signaling analysis

Module eigengene differences between BMI groups were assessed using the two-sided Wilcoxon rank-sum test. Enrichment of ESR1-bound genes within co-expression modules was evaluated using a hypergeometric test against a gene universe defined by the unique genes present in both the module dataset and the ChIP-seq dataset. Cell-cell communication significance was assessed using the permutation framework implemented in CellChat. Differential interaction strength and frequency were evaluated using chi-squared and permutation-based approaches.

Validation

Network rewiring intensity was evaluated using rank-based Wilcoxon tests and permutation testing (n = 10,000) to assess cross-platform concordance.

Additional statistical parameters and thresholds are provided in the figure legends and results section.

Additional resources

Biorender was used to generate the Graphical Abstract.