Cell-resolved genetic risk reveals hyperandrogenism-associated immune and metabolic rewiring in the PMOS follicular microenvironment

Abstract

Background

Polyendocrine metabolic ovarian syndrome (PMOS) is a leading cause of anovulatory infertility. Although follicular developmental arrest is a defining feature of PMOS, the cell-type-specific mechanisms linking genetic susceptibility, endocrine imbalance and follicular microenvironmental dysfunction remain incompletely understood, particularly across clinical subtypes.

Methods

We performed single-cell RNA sequencing of preovulatory follicular-fluid cells from healthy controls and women with uncomplicated PMOS, hyperandrogenic PMOS and metformin-treated insulin-resistant PMOS. We integrated these data with GWAS-derived disease relevance scores and ovarian chromatin-accessibility annotations to map genetic risk to follicular cell populations and regulatory programmes. Ligand–receptor analyses were used to infer intercellular communication networks. Transcriptomic metabolic findings were further assessed by follicular-fluid metabolic and steroid-hormone measurements, DHT-treated granulosa-cell functional assays, and reanalysis of a DHEA-induced mouse PMOS-like ovarian dataset.

Results

We generated a single-cell atlas of the PMOS follicular microenvironment comprising 220,983 high-quality cells. Granulosa cells showed the strongest enrichment of PMOS-related genetic risk, with regulatory-genetic support from ovarian snATAC-seq integration. Among granulosa cell states, a GC2_CYP19A1⁺ population was preferentially associated with hyperandrogenic PMOS and displayed a transcriptional programme characterized by steroidogenic remodelling coupled to reduced energy-metabolic activity. Biochemical profiling supported this endocrine–metabolic imbalance, showing increased follicular-fluid androstenedione (ASD) and estrone, reduced estradiol and E2/ASD ratio, decreased ATP, citrate and fumarate hydratase activity, and increased lactate and β-hydroxybutyrate. Prolonged DHT exposure impaired granulosa-cell mitochondrial respiration, ATP production and mitochondrial membrane potential. In parallel, PMOS subtypes showed immune remodelling and altered ligand–receptor signalling, with hyperandrogenic PMOS associated with increased inflammatory signalling and reduced homeostatic or tissue-repair pathways. Reanalysis of a DHEA-induced mouse ovarian dataset provided supportive evidence for androgen-associated steroidogenic pathway perturbation.

Conclusions

This study provides a cell-resolved framework linking PMOS genetic susceptibility to endocrine–metabolic and immune remodelling in the preovulatory follicular microenvironment. The identification of a hyperandrogen-associated GC2_CYP19A1⁺ state highlights a candidate cellular programme in which steroidogenic imbalance is coupled to impaired granulosa-cell bioenergetics. These findings provide a resource for future validation and may inform microenvironment-guided stratification of PMOS and the development of strategies aimed at restoring follicular function.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13073-026-01734-7.

Background

Polyendocrine metabolic ovarian syndrome (PMOS), formerly known as polycystic ovary syndrome (PCOS), is one of the most prevalent endocrine–metabolic disorders affecting women of reproductive age, with a global prevalence of approximately 8–12%, and accounting for up to 50–70% of cases of anovulatory infertility [1, 2]. The transition from PCOS to PMOS reflects the progressive evolution of scientific understanding of this condition. The new term PMOS more explicitly captures the interconnected polyendocrine, metabolic, and ovarian features that have long been recognized as central to the syndrome [3]. PMOS is frequently accompanied by metabolic disturbances such as insulin resistance, obesity, and type 2 diabetes [4]. Beyond its role as a leading cause of infertility [5], PMOS substantially increases the long-term risk of cardiovascular disease [6], endometrial cancer [7], and other chronic complications, thereby exerting a profound impact on women’s physical and mental health.

Despite decades of intensive research, the aetiology of PMOS remains incompletely understood. Current evidence suggests that its pathogenesis arises from complex interactions among genetic susceptibility, endocrine dysregulation, metabolic abnormalities, and immune imbalance [8]. Among these, follicular developmental arrest represents a central pathological hallmark underlying ovulatory dysfunction in PMOS [9].

Follicular fluid constitutes the immediate microenvironment for oocyte and granulosa cell development and contains bioactive molecules such as steroid hormones, cytokines, and metabolic intermediates. Together, these elements orchestrate follicular maturation and ovulation [10]. PMOS is a highly heterogeneous disorder, encompassing distinct clinical subtypes such as hyperandrogenic and insulin-resistant phenotypes [11]. However, the molecular mechanisms underlying these subtypes remains poorly understood.

Previous studies of PMOS follicular fluid have largely relied on bulk molecular profiling or analyses restricted to limited cell populations, precluding the resolution of cellular heterogeneity and dynamic intercellular interactions at single-cell resolution. For instance, although immune cell infiltration has been reported in PMOS ovaries [12], the specific immune subsets involved, their functional states, and their regulatory interactions with granulosa cells remain unclear. Granulosa cells, the principal somatic support cells of follicular development, are known to exhibit functional impairments in PMOS [13]; however, their intrinsic metabolic defects, genetic risk signatures, and links to endocrine disturbances have not been fully elucidated at the single-cell level. Notably, genome-wide association studies (GWAS) have identified numerous genetic risk loci associated with PMOS [14], yet how these susceptibility signals are manifested within specific follicular cell types remains a critical unanswered question.

To decode the cellular and molecular landscape underlying follicular developmental arrest in PMOS at single-cell resolution, we performed single-cell RNA sequencing (scRNA-seq) on preovulatory follicular fluid obtained from patients with PMOS multiple subtypes. Here, we delineate the immune and metabolic reprogramming of the PMOS follicular microenvironment. We observed elevated infiltration of immune cells, particularly T lymphocytes, in PMOS, with trends suggesting that a hyperandrogenic milieu may influence the differentiation and functional states of lymphocytes and macrophages. Importantly, we identify a granulosa cell subpopulation enriched for genetic risk signals that exhibits a distinctive form of metabolic reprogramming under hyperandrogenic conditions, characterized by steroidogenic remodeling concomitant with suppression of energy metabolism. Through cell–cell communication network analyses, we further uncover disrupted signaling crosstalk between granulosa cells and immune cells. Collectively, our study provides a single-cell atlas of the PMOS follicular microenvironment, integrating genome-wide association study (GWAS) signals to map genetic risk to specific cell populations. These findings offer potential targets for precision stratification and therapeutic intervention aimed at restoring follicular development and ovulatory function in PMOS.

Methods

Study design and patient sample collection

PMOS was diagnosed according to the Rotterdam criteria. Throughout this manuscript, PMOS is used as the updated terminology for polycystic ovary syndrome (PCOS). For consistency with original figure annotations and previously generated datasets, the term PCOS is retained in certain figures and tables, referring to the same condition now termed PMOS. Adult patients were diagnosed with PMOS when at least two of the following three features were present after exclusion of related disorders: oligo- or anovulation, clinical and/or biochemical hyperandrogenism, and polycystic ovarian morphology on ultrasonography. PMOS group: patients with PMOS characterized by oligo-/anovulation and polycystic ovarian morphology, but without biochemical or clinical hyperandrogenism and without insulin resistance. Hyper group: patients with PMOS and hyperandrogenism. Hyperandrogenism was defined by biochemical evidence of androgen excess, including serum total testosterone above the upper limit of the range, and/or elevated free androgen index, or by clinically evident hyperandrogenism, such as hirsutism assessed by the modified Ferriman–Gallwey score (> 4). Met-IR group: patients with PMOS and insulin resistance who were receiving metformin treatment at the time of follicular-fluid collection. Insulin resistance was defined using HOMA-IR.

All follicular fluid samples were collected after obtaining written informed consent from all participants, in accordance with the principles of the Declaration of Helsinki (2000 revision). Follicular fluid was obtained from 20 patients diagnosed with PMOS presenting with ovulatory dysfunction who underwent in vitro fertilization or intracytoplasmic sperm injection (IVF/ICSI) at the Reproductive Medicine Center of the First Affiliated Hospital of Naval Medical University. Among the PMOS cohort, seven patients met the diagnostic criteria for PMOS based on polycystic ovarian morphology on ultrasonography together with oligo- or anovulation (PMOS group), six patients were diagnosed with PMOS accompanied by hyperandrogenism (Hyper group), and seven patients were diagnosed with PMOS with concomitant insulin resistance and were receiving metformin treatment (Met-IR group). In addition, single-cell transcriptomic data from three healthy control individuals were integrated from our previously published dataset [15] (https://www.ncbi.nlm.nih.gov/bioproject/PRJNA1155666). These samples were previously collected and generated by our group as healthy control samples and were processed using the same standardized workflow as the PMOS samples. Specifically, the control and PMOS samples were collected during comparable IVF/oocyte retrieval procedures, processed using the same follicular fluid handling and cell preparation pipeline, captured using the BD Rhapsody single-cell platform, and sequenced using the BD Rhapsody WTA workflow.

Patients underwent ovarian stimulation according to individualized protocols determined by their attending clinicians, including mild stimulation and gonadotropin-releasing hormone antagonist regimens. When transvaginal ultrasonography indicated that the leading follicle had reached a diameter of at least 18 mm, a single dose of 0.25 mg human chorionic gonadotropin (hCG) was administered to trigger final oocyte maturation. For downstream single-cell analyses, only follicles with a diameter greater than 8 mm were aspirated and included. These follicles are generally classified as antral follicles and are considered to be at comparable stages of maturation. By excluding smaller or atretic follicles, we aimed to minimize transcriptional variability attributable to differences in follicular developmental stage. Oocyte retrieval was performed 34–38 h after hCG administration under transvaginal ultrasound guidance. Retrieved oocytes were used for IVF/ICSI procedures, while the corresponding follicular fluid was collected for subsequent analyses. This study was approved by the Medical Ethics Committee of the First Affiliated Hospital of Naval Medical University (approval number: CHEC2025-132).

Follicular fluid sample processing

Following oocyte retrieval from patients undergoing controlled ovarian hyperstimulation for IVF/ICSI, oocyte–cumulus complexes were mechanically dissociated. The cumulus cell complexes were washed twice in DMEM/F12 medium supplemented with 10% fetal bovine serum (FBS) to remove residual MOPS buffer and mineral oil remaining from the oocyte retrieval procedure. Follicular fluid samples were collected separately from IVF/ICSI patients, excluding oocytes and oocyte–cumulus complexes. Red blood cell lysis buffer was added directly to the follicular fluid, followed by gentle pipetting to lyse erythrocytes. The mixture was then centrifuged, and the resulting cell pellet was washed twice with phosphate-buffered saline (PBS) before being resuspended in DMEM/F12 medium containing 10% FBS. The resuspended follicular fluid–derived cells were subsequently combined with the mechanically isolated cumulus cells and maintained at 4 °C prior to downstream scRNA-seq.

Singe-cel library preparation and sequencing

scRNA-seq was performed using the BD Rhapsody system in accordance with the manufacturer’s instructions. Briefly, single-cell capture was achieved by randomly distributing a single-cell suspension into more than 20,000 microwells using a limiting dilution approach. Barcoded magnetic beads were then loaded into the microwells to pair with individual cells. Following cell lysis, mRNA molecules hybridised to bead-bound barcoded oligonucleotides. The beads were subsequently collected for reverse transcription and Exonuclease I (ExoI) treatment. Each cDNA molecule was labelled at its 5′ end with a unique molecular identifier (UMI) and a cell-specific barcode. Whole-transcriptome libraries were generated using the BD Rhapsody Single-Cell Whole Transcriptome Amplification (WTA) workflow, which includes random primer extension (RPE), RPE amplification PCR, and WTA indexing PCR. Library quality and concentration were assessed using an Agilent 2200 Bioanalyzer with High Sensitivity DNA chips and the Qubit High Sensitivity DNA Assay Kit (Thermo Fisher Scientific). Sequencing was performed on an Illumina NovaSeq 6000 platform using 150-bp paired-end reads. For each scRNA-seq library, raw sequencing data were processed using the BD Genomics Rhapsody analysis pipeline (version 2.2.1) with the RhapRef Human WTA reference genome (available at bd-rhapsody-public.s3-website-us-east-1.amazonaws.com/Rhapsody-WTA/). Read alignment, transcript quantification, and initial quality control were conducted using default parameters. Filtered cell-by-gene count matrices were used for all downstream analyses.

Quality control and doublet detection

Quality control was performed using the scAutoQC workflow implemented using SCTK (v0.1.1) [16]. Cells were excluded if they had fewer than 500 unique molecular identifiers (UMIs), fewer than 200 detected genes, or if mitochondrial transcripts accounted for more than 25% of total expression. For each sample, three quality control metrics were calculated: log-transformed UMI counts per cell (log1p_n_counts), log-transformed number of detected genes per cell (log1p_n_genes), and the percentage of mitochondrial transcripts (percent_mito). These metrics were used for dimensionality reduction, neighbourhood graph construction, and low-resolution clustering to generate quality control–based clusters (QC clusters).

The multi_resolution_cluster_qc procedure was then applied to evaluate multiple clustering resolutions ranging from 0.1 to 1.0, yielding two alternative quality control decision schemes. The final quality control configuration was determined by maximising the Jaccard index between cell-level and cluster-level quality control outcomes, while also considering resolution-dependent pass rates. Following the removal of low-quality cells, doublets were quantified using Scrublet (v0.2.1) [17] with a two-step diffusion-based approach and Bonferroni correction. At this stage, predicted doublets were not excluded but were retained to assess enrichment at the cluster level during subsequent analyses.

For the PMOS mouse ovary single-cell datasets, the same quality control strategy was applied with adjusted thresholds reflecting dataset-specific characteristics (n_counts ≥ 2,000, n_genes ≥ 1,000, percent_mito ≤ 15%).

Batch correction, clustering, and cell type annotation

After merging all samples, data integration was performed using scVI (v1.2) [18] with the top 2,000 highly variable genes (HVGs). Sample identity was specified as the batch covariate, and percent_mito was included as a continuous covariate. The resulting latent representations were used to construct neighbourhood graphs, generate UMAP visualisations, and perform Leiden clustering. Cells were initially assigned to seven major lineages based on canonical marker gene expression. Prior to subclustering, genes related to the cell cycle, ribosomal proteins, and haemoglobin were excluded, and each lineage was re-integrated using scVI with the top 2,000 HVGs. Subtype annotation was performed using Celltypist (v1.6.3) [19] prediction models, complemented by manual curation based on marker gene co-expression, Scrublet scores, and predicted labels to ensure the removal of residual doublets. Immune cell subtypes were annotated using predictions from the Immune_All_Low v2 model, further refined by Leiden marker-based analysis. To ensure high-quality granulosa cell analyses, cells with fewer than 2,000 detected genes were excluded prior to subclustering.

To quantitatively evaluate batch correction and biological conservation after integration, we benchmarked the uncorrected PCA embedding and the scVI latent representation using scIB metrics [20]. Sample identity was used as the batch variable, and major cell-type or subcluster annotations were used as biological labels. Batch-mixing performance was assessed using iLISI, kBET, graph connectivity, BRAS and PCR comparison, while biological conservation was evaluated using isolated-label score, KMeans NMI, KMeans ARI, silhouette label score and cLISI. Aggregate batch-correction, bio-conservation and total integration scores were then calculated to compare the uncorrected and scVI-corrected embeddings.

Integration, dimensionality reduction, and annotation of mouse ovary datasets followed a similar strategy. Previously published single-cell atlases of the mouse ovary were used as reference models for Celltypist predictions [21]. After the initial round of dimensionality reduction and clustering, two low-quality clusters were removed: one characterised by the concurrent expression of markers from multiple cell types, and another detected exclusively in a single sample.

scDRS analysis

GWAS summary statistics for PMOS and PMOS-related traits were obtained from published studies [14, 22–26]. Given that our single-cell dataset consisted exclusively of individuals of East Asian ancestry, GWAS summary statistics from East Asian populations were prioritized whenever available; when unavailable, data from European populations were used. Gene-level disease association scores were computed using MAGMA [27] based on SNP identifiers. These gene-level scores were then used as input for scDRS [28], which calculates disease relevance scores using the top 1,000 genes with the strongest associations and outputs a disease score and statistical significance for each individual cell. Statistical significance was assessed using the default scDRS control-score framework. Briefly, 1,000 matched control gene sets were generated using mean–variance matching of gene expression properties. The observed disease score of each cell was normalized against scores from these matched control gene sets, and empirical Monte Carlo P values were calculated from the resulting null distribution. Observed group-level association statistics were compared with the corresponding statistics derived from matched control scores. Cell group-level association was assessed using the default 5% quantile-based test, and within-group heterogeneity was evaluated using Geary’s C. P values were adjusted using the Benjamini–Hochberg procedure, with adjusted P < 0.05 considered significant.

For downstream analyses, disease enrichment across annotated cell types or clinical subgroups was evaluated using the default scDRS group-level testing framework. Observed group-level association statistics were compared with the corresponding statistics derived from matched control scores. Cell group-level association was assessed using the default 5% quantile-based test, and within-group heterogeneity was evaluated using Geary’s C. P values were adjusted using the Benjamini–Hochberg procedure, with adjusted P < 0.05 considered significant.

Integration of GWAS summary statistics with ovarian snATAC-seq data

To assess whether PMOS-related genetic associations converge on ovarian cell type-specific regulatory elements, we integrated GWAS summary statistics with the public ovarian snATAC-seq dataset GSE202601 [29] (https://www.ncbi.nlm.nih.gov/bioproject/PRJNA836755). Published cell type-associated differentially accessible regions (DARs) from GSE202601 were used, including DARs assigned to granulosa cells (GC), theca cells (TC), epithelial cells (EpiC), endothelial cells (EC), immune cells (IC) and smooth muscle/stromal cells (SMC). DAR coordinates were harmonized according to the downstream analysis: hg38 coordinates were used for direct SNP–DAR overlap analysis, whereas hg19/GRCh37 coordinates were used for stratified LD score regression (S-LDSC) to match the 1000 Genomes Phase 3 LDSC reference panels.

Variants with invalid alleles, non-SNV alleles, invalid P values or missing key statistics were removed. When rsIDs were unavailable, variants were matched to the corresponding 1000 Genomes reference panel by chromosome, position and allele information. Signed Z statistics were calculated as β/SE for quantitative traits or log(OR)/SE for case–control traits. For case–control PMOS GWAS, effective sample size was calculated as 4/(1/N_case + 1/N_control). European and East Asian GWAS traits were analyzed using ancestry-matched LDSC reference panels.

SNP–DAR–gene overlap analysis

Genome-wide significant GWAS variants (P < 5 × 10⁻⁸) were converted to BED format and intersected with GSE202601 cell type-associated DARs. Overlapping SNP–DAR pairs were linked to candidate genes using promoter annotation, defined as ± 2 kb from the transcription start site, and nearest-gene annotation for distal regions. Promoter-based annotations were prioritized when both annotations were available. For each trait and cell type, we summarized the number of overlapping GWAS variants, DARs and candidate genes. Candidate genes were prioritized using GWAS significance, DAR accessibility effect size and the number of SNP–DAR links.

Stratified LD score regression

S-LDSC was performed to test whether ovarian cell type-associated DARs were enriched for trait heritability. GWAS summary statistics were formatted using munge_sumstats.py and restricted to HapMap3 SNPs. Cell type-specific DAR annotations were lifted to hg19/GRCh37 and converted into LDSC annotation files using ancestry-matched 1000 Genomes Phase 3 PLINK reference panels. Annotation-specific LD scores were calculated using ldsc.py –l2 with a 1-cM LD window. For each trait and cell type, the cell type-specific DAR annotation was jointly modeled with baselineLD v2.2 annotations using –overlap-annot, ancestry-matched allele frequency files and HapMap3 no-MHC regression weights. The S-LDSC coefficient z-score was used as the primary statistic for cell type prioritization. Nominal P values were derived from coefficient z-scores, and Benjamini–Hochberg correction was applied across cell types within each trait.

Metabolic analysis

Cell-type–specific metabolic activity was inferred using a task-based metabolic modelling framework implemented in scCellFie [30]. For each metabolic reaction, scCellFie scores were compared across conditions using Wilcoxon rank-sum tests. scCellFie visualizations include both positive and negative reactions grouped according to their predefined metabolic subsystems, allowing systematic comparison of metabolic pathway activity across cell populations and clinical conditions.

Differential abundance analysis

To assess enrichment or depletion of specific cell populations across clinical subgroups, we quantified the distribution of all second-level cell types in subgroups. For each subgroup, the frequency of a given cell type was compared with its frequency across all other subgroups combined. Two-sided Fisher’s exact tests were performed for each subgroup–cell type combination to evaluate whether the observed proportions differed significantly from background expectations. The magnitude and direction of enrichment were summarized using odds ratios (ORs). All P values were corrected for multiple testing using the Benjamini–Hochberg false discovery rate (FDR) procedure, with FDR < 0.05 considered statistically significant. To complement this analysis, differential abundance of cells within local neighborhoods between conditions (Hyper versus PMOS or Hyper versus Met-IR) was further assessed using miloR [31], providing an independent neighborhood-based framework for compositional comparison.

Gene set and pathway activity analysis

Pathway activity inference was performed using PROGENy, with the top 500 genes selected for each pathway. Pathway activity scores were calculated using the univariate linear model (ULM) implemented in decoupleR [32] and applied to normalized expression matrices. Mean pathway activity scores were subsequently computed and visualized for each cluster. In addition, gene set scoring was performed using the score_genes function in Scanpy, applied to normalized expression values.

Differential expression analysis

Differential gene expression between cell types was assessed using the rank_genes_groups function in Scanpy with the Wilcoxon rank-sum test. For comparisons between clinical conditions, differential expression analyses were performed using MAST [33], which is well suited for single-cell data characterized by zero inflation and provides more stable inference than non-parametric methods in the presence of sparse expression patterns.

To account for donor-level variability and avoid treating individual cells as independent biological replicates, differential expression analysis was additionally performed using a pseudo-bulk framework. Genes with low expression across pseudo-bulk samples were removed before model fitting. Differential expression was then performed on donor-level pseudo-bulk count matrices using DESeq2 [34], with subgroup specified as the main variable. Multiple testing correction was performed using the Benjamini–Hochberg procedure, and genes with FDR-adjusted P values below the indicated threshold were considered significant.

Trajectory inference analysis

Cellular differentiation trajectories were reconstructed using diffusion map–based approaches [35]. For each lineage of interest, including CD4⁺ T cells and monocyte/macrophage populations, cells were first embedded using diffusion maps. DPT was then computed to estimate each cell’s relative progression along the inferred manifold. For CD4⁺ T cell analyses, naïve T cells identified based on established markers and prior literature were selected as the root population. For the monocyte–macrophage lineage, CD14⁺ monocytes were specified as the root cells.

SCENIC regulatory network analysis

Regulon activity inference was performed using the Python implementation of the SCENIC pipeline [36]. The input expression matrix comprised barcodes from all four conditions. Genes expressed in fewer than 5% of nuclei within any given cell state were excluded, as were mitochondrial, ribosomal, haemoglobin, non-coding, antisense, and microRNA genes. Cell-level regulon activity scores were compared between conditions using non-parametric Wilcoxon rank-sum tests to identify condition-specific regulatory programs across cell states and cell types.

ELISA

Steroid hormone concentrations and metabolic parameters were measured according to the manufacturers’ instructions. Estrone (Mkbio, MK0125ZA), estradiol (Mkbio, MK0376OA), and androstenedione (Mkbio, MK0121ZA) were quantified using competitive ELISA kits. Lactate (Adsbio, ADS-W-T009-96), pyruvate (Adsbio, ADS-W-S014), ATP (Adsbio, ADS-W-A001-96), citrate (Adsbio, ADS-W-S002-96), β-hydroxybutyrate (Adsbio, ADS-W-D015-48), and fumarase activity (Adsbio, ADS-W-FM019) were measured using microplate-based assay kits..

Cell culture

Granulosa KGN cell lines were cultured in DMEM/F12 medium (GIBCO) supplemented with 10% fetal bovine serum (GIBCO) and 1% penicillin–streptomycin (GIBCO) at 37 °C in a humidified incubator with 5% CO₂. Cells were treated with different concentrations of dihydrotestosterone (DHT, TargetMol, T84319). The experimental groups included NC, DHT-500 nM, DHT-1 μM and DHT-2 μM. Cells were collected after 2, 4 or 7 days of treatment for downstream assays.

KGN cells were seeded in six-well plates at a density of 40,000 cells per well. Cells from each treatment group were collected at the indicated time points, and ATP and lactate levels were measured using ATP Assay Kit (Beyotime, S0026) and Lactate Assay Kit (APPLYGEN, E2063) according to the manufacturer’s instructions. Total protein concentration was determined using a BCA Protein Quantification Kit (Epizyme, ZJ102).

Seahorse

Mitochondrial oxygen consumption was measured using a mitochondrial oxidative phosphorylation assay kit (Alicelligent, ALS22012) and an Agilent Seahorse XFe96 Analyzer according to the manufacturers’ instructions. KGN cells were seeded into Seahorse XF96 cell culture microplates at a density of 5,000 cells per well on days 1, 3 and 6, and were cultured for 24 h at 37 °C with 5% CO₂ before analysis. The XFe96 extracellular flux assay cartridge was hydrated overnight at 37 °C in a non-CO₂ incubator.

On the day of the assay, Seahorse XF DMEM assay medium (Agilent, 103,680–100) was supplemented with glucose, pyruvate and glutamine to final concentrations of 10 mM glucose, 1 mM pyruvate and 2 mM glutamine, and pre-warmed at 37 °C in a non-CO₂ incubator. Culture medium was removed from the cell plate, and cells were washed three times with assay medium. A total of 180 μl assay medium was then added to each well, and the plate was incubated at 37 °C in a non-CO₂ incubator for 1 h. Oxygen consumption rate (OCR) was measured under basal conditions and after sequential injection of oligomycin (1.5 μM), FCCP (0.5 μM) and rotenone/antimycin A (1 μM).

JC-1 staining probes for live cell imaging

50,000 cells from each group were seeded into 20-mm confocal dishes on day 2 and cultured until day 4. Mitochondrial membrane potential and nuclei were stained using a JC-1 MitoMP Detection Kit (DOJINDO, MT09) and Cellstain-Hoechst 33,342 solution reagent (DOJINDO, H342), respectively, according to the manufacturers’ instructions. Fluorescence images were acquired using a laser scanning confocal microscope (Nikon, AXR).

Cell–cell interaction analyses

To characterize heterogeneity in granulosa cell communication states within the follicular fluid microenvironment, ligand–receptor interactions were inferred using CellPhoneDB (v5.0.0) [37] with its differentially expressed gene–based approach. Ligand–receptor pairs were retained if both interaction partners were expressed in at least 10% of cells in the cell types and if at least one partner (ligand or receptor) was significantly upregulated in the corresponding cell state compared with other states within the same lineage (Wilcoxon test; adjusted P < 0.05 and log₂ fold change > 0.25). Differential expression analyses were conducted within each lineage to identify genes specifically upregulated in individual cell states relative to other states in the same lineage, enabling the identification of hormone-related signaling pathways.

For broader comparative analyses of intercellular communication, CellChat (v2) [38] was used to infer cell–cell communication networks. Communication networks were inferred separately for each condition. Each condition was downsampled to 20,000 barcodes to avoid any confounding effects arising from higher cell numbers in each group, and cell types with very low numbers were removed because these cell types often have higher mean gene expression owing to low cluster background. Differential communication was then assessed using CellChat rankNet, including both overall pathway-level changes and sender–receiver-specific comparisons.

Results

Mapping PMOS follicular fluid remodeling dynamics

To elucidate the cellular and molecular mechanisms underlying preovulatory follicular developmental arrest in PMOS, we performed scRNA-seq from preovulatory follicular fluid obtained from patients with uncomplicated PCOS referring to the new term PMOS (PMOS, n = 7), PMOS with hyperandrogenism (Hyper, n = 6), and PMOS with insulin resistance receiving metformin treatment (Met-IR, n = 7). These data were integrated with our previously published scRNA-seq dataset from healthy controls [15] (NC, n = 3) (Fig. 1A and Additonal file 2: Additonal file 2: Table S1). All PMOS patients and control individuals underwent standardized clinical evaluation and ovarian stimulation protocols prior to in vitro fertilization (IVF). Consequently, key baseline clinical characteristics at the time of oocyte retrieval were comparable across groups, ensuring physiological consistency of follicular fluid samples used for downstream analyses (Fig. 1B).

Fig. 1.

Single-cell transcriptomic landscape of human follicular fluid across PMOS subtypes and its genetic and metabolic associations (For consistency with the original figure annotations and datasets generated before the updated nomenclature, the term PCOS is retained in the figures while the term PMOS in the figure legends corresponds to PCOS labels retained in the original figures, reflecting the updated nomenclature. This terminology convention applies consistently to all figures and tables throughout the manuscript.). A Schematic illustration of the included participants, follicular fluid samples collected during in vitro fertilization, and single-cell RNA sequencing performed using BD Rhapsody. NC, normal control; PMOS, polyendocrine metabolic ovarian syndrome; Hyper, PMOS combined with hyperandrogenism; Met-IR, metformin-treated PMOS with insulin resistance. B Clinical characteristics of the primary cohort. Boxplot, median IQR min/max. BMI, body mass index (kg m^–2); FBG, fasting blood glucose (mmol L.⁻¹). Two-tailed Wilcoxon Unpaired test. C Uniform manifold approximation and projection (UMAP) plot of 220, 983 cells and colored according to cell type in NC (n = 3), PMOS (n = 7), Hyper (n = 6) and Met-IR (n = 7). D Cell-type proportions in global dataset (excluding neutrophils), mean per group, and for each sample. E Proportional changes in immune cells(upper), CD4 +, CD8 + and B cells (below) between NC (n = 3), PMOS (n = 7), Hyper (n = 6) and Met-IR (n = 7). Boxplot, median IQR min/max. Two-tailed Wilcoxon Unpaired test. F Dot plot showing the associations between PMOS-related genetic traits and cell-types. Each column represents a disease or genetic trait, and each row corresponds to a specific cell type. Dot size reflects the magnitude of the association, and dots outlined with a border indicate statistically significant associations (P < 0.05). Significance was assessed using Monte Carlo permutation testing (see Methods). Bio. T., bioavailable testosterone; Insulin, fasting insulin; T2D, Type 2 diabetes. G Alterations in pathway-wide metabolic flux between groups in major cell types. Cohen’ s D, colored at absolute value > 0.5 and FDR < 0.01

To uniformly process the data, we remapped raw sequencing data and processed gene counts.

through quality control pipeline (scAutoQC), removing low-quality cells in an unbiased and automated way (Methods; Additional file 1: Fig. S1). We used single-cell variational inference (scVI) to integrate the data (Additional file 1: Fig. S2). After stringent quality control, we retained 220,983 high-quality single-cell transcriptomes derived from 23 samples for subsequent analyses (Fig. 1C, Additional file 1: Fig. S1-S2 and Additonal file 2: Table S2). Unsupervised clustering identified 12 major cell types within the follicular fluid (Additional file 1: Fig. S3A). Notably, neutrophils constituted the largest immune cell population, representing a substantial fraction of the total cellular landscape (Additional file 1: Fig. S3B and Additonal file 2: Table S3). Given that follicular fluid is an exudative compartment containing circulating immune cells, together with the high capture efficiency of the BD platform for cells with high granularity and surface antigen expression, the enrichment of neutrophils was not unexpected and did not interfere with cell type annotation or downstream analyses. Following exclusion of neutrophils, we observed an increase in immune cells, including CD4⁺ T cells, CD8⁺ T cells, and B cells in PMOS subgroups compared with healthy controls (Fig. 1D-E and Additional file 1: Fig. S3C-D). In contrast, granulosa cells exhibited a modest but reduction in relative abundance in PMOS samples (Additional file 1: Fig. S3E-F).

To further interrogate the potential genetic drivers shaping follicular fluid cell populations in PMOS, we applied single-cell disease relevance score (scDRS) analysis [28], integrating GWAS data [14, 22–24, 29] for PMOS and related metabolic and endocrine traits with our scRNA-seq dataset (Additonal file 2: Table S4). This approach enabled the identification of follicular fluid cell types enriched for PMOS genetic susceptibility. Granulosa cells exhibited the strongest genetic risk enrichment, showing significant associations not only with PMOS itself but also with bioavailable testosterone and fasting insulin levels (Fig. 1F). Stratified analyses further revealed that granulosa cells in the Hyper group displayed particularly strong associations with multiple PMOS-related diseases and traits (Additional file 1: Fig. S3G). Intriguingly, body mass index (BMI) and fasting insulin were preferentially associated with B cells and natural killer (NK) cells, respectively, suggesting that PMOS risk loci may influence ovarian endocrine function while indirectly modulating the local immune–metabolic microenvironment to drive disease pathogenesis. To provide an regulatory-genetic validation, we integrated these GWAS summary statistics with the public ovarian snATAC-seq dataset [29] (Methods). Stratified LD score regression (S-LDSC) using cell type-associated differentially accessible regions (DARs) revealed trait-dependent enrichment of GWAS heritability in ovarian regulatory elements, with positive enrichment signals involving granulosa cells, theca cells and immune/stromal-related compartments across PMOS-related traits (Additional file 1: Fig. S3H-I and Additonal file 2: Table S5). Coordinate-based GWAS–DAR overlap analysis further nominated cell type-specific regulatory candidate genes (Methods, Additonal file 2: Table S6). Granulosa cell DARs contained the largest number of candidate genes for PMOS and bioavailable testosterone, highlighting genes such as PPARG, FTO, YAP1, CDKAL1, AUTS2, ATP1B2, CYP19A1, GREB1, TCF12, EPHA2, COBLL1 and FAM13A (Additional file 1: Fig. S4A-B). These genes showed corresponding expression patterns in our follicular fluid single-cell RNA-seq dataset, with many displaying detectable or relatively enriched expression in granulosa cells (Additional file 1: Fig. S4C). These findings provide regulatory and transcriptional support for the convergence of PMOS-related genetic risk on granulosa cell endocrine–metabolic programs.

Given that PMOS is fundamentally a metabolic disorder [39], we next sought to determine whether altered energy utilization and metabolic reprogramming within the follicular microenvironment contribute to cellular dysfunction. Using gene expression–based metabolic flux modelling [30], we systematically inferred pathway-level metabolic activities across cell types. This analysis revealed a suppression of metabolic flux in PMOS samples, relative to healthy controls (Fig. 1G). Collectively, these complementary analyses delineate an initial cellular and molecular framework of follicular developmental arrest in PMOS. Together with the genetic-risk enrichment, these findings provide an initial cellular and molecular framework linking PMOS-associated genetic susceptibility, and metabolic remodeling.

T cell heterogeneity and signatures in PMOS

Given the alterations in immune cell composition within PMOS follicular fluid, we next focused on the functional heterogeneity of lymphocyte populations. Using a combination of semi-automated annotation and manual curation, lymphocytes were classified into 11 distinct subtypes (Methods; Fig. 2A-B and Additional file 1: Fig. S5A-B). CD4⁺ T cell subtypes were consistently enriched in both the PMOS and Hyper groups, whereas their abundance was reduced in the Met-IR group (Fig. 2C-D and Additional file 1: Fig. S5C). Within the CD8⁺ T cell compartment, uncomplicated PMOS was characterized by a significant expansion of proliferative naïve/central memory T cells, whereas the Hyper group exhibited marked enrichment of effector memory and tissue-resident memory T cells (Tem/Trm) with attenuated cytotoxic signatures and elevated exhaustion scores (Fig. 2C and Additonal file 2: Table S7). In contrast, the Met-IR group displayed pronounced enrichment of effector memory/recently activated effector memory T cells (Tem/Temra) as well as γδ T cells, which further underscores the subgroup-specific immune heterogeneity of the PMOS follicular microenvironment.

Fig. 2.

T and NK cell heterogeneity and subgroup-specific transcriptional programs across PMOS subtypes. A UMAP plots of T and NK cells, colored according to cell type. Tem, Effector Memory T cells; Treg, Regulatory T cells; Tcm/Tn, Central Memory/Naive T cells; Tem/Temra, Effector Memory/recently activated effector memory; Trm, Tissue-resident Memory T cells; MAIT, Mucosal-associated invariant T cells; ILC2, Group 2 Innate Lymphoid Cells. B UMAP plot of T/NK cell specific subclusters. C Left, dot plot showing the relative representation of each cell type across the four groups. For each subgroup–cell-type combination, the odds ratio (OR) was calculated by Fisher’ s exact test comparing the proportion of that cell type within the indicated subgroup versus all other subgroups combined. The color of each dot encodes the log2(OR), with more extreme colors indicating stronger enrichment or depletion relative to the other groups. Dot outline denotes statistical significance after Benjamini–Hochberg correction (FDR < 0.05). Middle, heatmap of signaling pathway activity scores across T/NK cell subclusters. Right, heatmap of average T cell state scores across T/NK cell subclusters. D UMAP of embedding density represented as a Gaussian density estimation, of T and NK cells by groups. E DiffMap plot colored by pseudotime and cell types across subclusters of CD4⁺ T cells. Scaled module scores within these subclusters of CD4⁺ T cells with respect to two pseudotime trajectories. F Metabolic and Progeny pathway activities along two pseudotime trajectories. G Volcano plots showing pseudo-bulk differential expression results for CD4⁺ and CD8⁺ T cells in pairwise comparisons of the Hyper group versus the PMOS and Met-IR groups. H Heatmaps showing Hyper group-specific genes identified in CD4⁺ and CD8⁺ T cells. Colors represent log₂ fold changes, with red indicating higher expression and blue indicating lower expression in the Hyper group

To further elucidate the relationship between metabolic programs, functional states, and CD4⁺ T cell subtypes, we constructed differentiation trajectories of CD4⁺ T cells. Trajectory inference projected CD4⁺ T cells along two trajectories, terminating in effector memory T cell (Tem) and regulatory T cell (Treg) phenotypes (Fig. 2E). Along these trajectories, dynamic profiling of signaling pathways and metabolic features revealed functional divergence during differentiation and maturation (Fig. 2F and Additional file 1: Fig. S5D). Cells along both trajectories exhibited increased activity of ketone body synthesis (acetoacetate synthesis) and glycolysis, reflecting the metabolic demands associated with T cell activation and transition from quiescence to effector states [40]. Notably, glycolytic scores increased more sharply during late-stage differentiation along the Tem trajectory than along the Treg trajectory, consistent with the heightened glucose utilization and energetic reprogramming required for effector T cell function. Canonical pathways involved in T cell activation and proliferation, including PI3K and EGFR signaling, were persistently and robustly upregulated in the late stages of the Tem trajectory, far exceeding the levels observed along the Treg trajectory [41]. We also observed a coupling between T cell differentiation and the local endocrine milieu. Estrogen receptor signaling peaked early along the Treg trajectory before rapidly declining, whereas it remained relatively stable along the Tem trajectory. In contrast, androgen receptor signaling was markedly upregulated during the late stages of the Tem trajectory, reaching levels substantially higher than those observed in Treg cells. These findings suggest that a hyperandrogenic environment may selectively promote Tem cell maturation and effector function. Consistent with this notion, apoptotic signaling mediated by TRAIL [42] was elevated at the onset of the Treg trajectory but declined to minimal levels at the terminal stages of the Tem trajectory.

To systematically compare transcriptional alterations in Hyper group, we performed pseudobulk- differential gene expression (DEG) analyses for CD8⁺ and CD4⁺ T cells (Additional file 1: Fig. S5E and Additonal file 2: Table S8-9). In CD4⁺ T cells, the Hyper group showed 33 DEGs relative to the PMOS group and 59 DEGs relative to the Met-IR group. By intersecting these two comparisons, we identified 21 CD4⁺ T-cell Hyper-specific genes, including 1 upregulated gene and 20 downregulated genes (Fig. 2G). The downregulated genes included FOS, JUN, NFKBIA, GADD45B, LTB and PIK3CD, suggesting altered stress-response programs in CD4⁺ T cells from the Hyper group. In CD8⁺ T cells, the Hyper group showed 61 DEGs relative to the PMOS group and 59 DEGs relative to the Met-IR group. Intersection of these two comparisons identified 28 CD8⁺ T-cell Hyper-specific genes, including 8 upregulated and 20 downregulated genes (Fig. 2G). The upregulated genes included MAF, CREBBP, KAT6A, MECP2, IKZF3 and ARHGAP25, indicating subtype-associated transcriptional and regulatory remodeling in CD8⁺ T cells. Notably, 11 Hyper-specific downregulated genes were shared between CD4⁺ and CD8⁺ T cells, (Additional file 1: Fig. S5F). Gene set enrichment analysis further indicated that Hyper-associated T-cell transcriptional changes were linked to translation, peptide biosynthesis, amide metabolism and RNA splicing-related biological processes (Fig. 2H). Together, these results indicate that the Hyper group is associated with distinct transcriptional features in follicular fluid T cells.

Myeloid alterations and metabolic rewiring in PMOS

We next expanded our analysis to myeloid cells to comprehensively assess the contribution of innate immunity to PMOS pathogenesis. Through refined clustering of monocytes/macrophages and dendritic cells, we identified 12 distinct myeloid cell clusters, including CD14⁺ and CD16⁺ monocytes, multiple macrophage subsets, and both classical and non-classical dendritic cells (Fig. 3A, Additional file 1: Fig. S6A and Additonal file 2: Table S10). Density mapping revealed that macrophage-enriched regions were more prominently concentrated in the PMOS and Hyper groups, whereas the Met-IR group exhibited a distribution pattern more similar to that of healthy controls, clustering predominantly within monocyte-rich regions (Fig. 3B and Additional file 1: Fig. S6B). Further analysis demonstrated that the PMOS group was selectively enriched for a LYVE1⁺ macrophage subpopulation characterized by elevated M2-like features and signatures of lipid-associated macrophages [43]. Notably, this enrichment was reversed in the Met-IR group, which closely resembled the NC group (Fig. 3C), may reflecting the anti-inflammatory and immunometabolic effects of metformin.

Fig. 3.

Transcriptional, functional and metabolic diversification of myeloid populations in PMOS follicular fluid. A UMAP plots of myeloid cells (excluding neutrophils), colored according to cell type. B UMAP of embedding density represented as a Gaussian density estimation, of myeloid cells (excluding neutrophils) by subgroups. C Left, dot plot showing the relative representation of each cell type across the four groups. For each subgroup–cell-type combination, the odds ratio (OR) was calculated by Fisher’ s exact test comparing the proportion of that cell type within the indicated subgroup versus all other subgroups combined. The color of each dot encodes the log2(OR). Dot outline denotes statistical significance after Benjamini–Hochberg correction (FDR < 0.05). Middle, boxplot of M1/M2 gene characterization across Myeloid cell subclusters. Right, heatmap of average myeloid cell state scores. Inflam, inflammation; LA, lipid-associated; Angio, angiogenesis; Reg. regulation; Prolif, proliferation. D Violin plots showing the distribution of DPT values across cell subsets along the two inferred branches. The upper panel shows the CD14⁺ monocyte-to-CD16⁺ monocyte trajectory, and the lower panel shows the CD14⁺ monocyte-to-macrophage trajectory. E Metabolic pathway activities along Monocyte to Macrophage pseudotime trajectories. F Alterations in pathway-wide metabolic flux between groups in Monocytes/Macrophages. Cohen’ s D, colored at absolute value > 0.5 and FDR < 0.01. G Heat map showing metabolic pathways that are ameliorated in macrophages from Met-IR group. Pathways are annotated according to their corresponding metabolic systems. H Dot plot showing Gene Ontology (GO) terms enriched from marker genes identified for each neutrophil subcluster (Wilcoxon test, adjusted P < 0.05, log₂ fold change > 0.25). Enrichment P values were calculated using two-sided Fisher’ s exact tests and adjusted for multiple testing using the Benjamini–Hochberg. I Dot plot showing the relative representation of each neutrophil subcluster across the four groups

To delineate the developmental dynamics of monocytes and macrophages, we reconstructed trajectories using diffusion maps and diffusion pseudotime (DPT) analysis (Methods). CD14⁺ monocytes differentiated along two principal trajectories: one leading toward CD16⁺ monocytes and the other toward macrophage lineages [44] (Fig. 3D and Additional file 1: Fig. S6C). To characterize metabolic adaptations during monocyte-to-macrophage differentiation, we next assessed the dynamic activity of key metabolic pathways along the CD14⁺ monocyte–macrophage pseudotime axis (Fig. 3E and Additional file 1: Fig. S6E). Multiple pathways related to protein processing and amino acid metabolism were progressively upregulated along the macrophage differentiation trajectory. Among these, the pathway associated with recognition of misfolded proteins exhibited the highest activity scores, may reflecting an increase in endoplasmic reticulum (ER) stress responses and protein synthesis demands as monocytes transition into fully differentiated macrophages [45]. In contrast, several steroid hormone biosynthesis pathways displayed a consistent decline along pseudotime, including pathways mediating the conversion of androstenedione to estradiol and estrone. This progressive attenuation of steroidogenic capacity during differentiation may support the that under conditions of reduced estrogen availability, macrophages may preferentially adopt mature inflammatory or lipid-associated phenotypes [46].

Metabolic flux analysis revealed suppression of metabolic pathway activity in macrophages from the PMOS subgroup compared with healthy controls (Fig. 3F). Specifically, macrophages from the NC group exhibited the highest flux scores across multiple amino acid biosynthesis pathways, as well as gluconeogenesis pathways driven by pyruvate, alanine, or lactate (Fig. 3G). These pathways were globally attenuated in the PMOS and Hyper groups. Similarly, ketone body–related metabolism and hypoxia-associated glycolytic ATP production remained robust in NC macrophages but were markedly reduced in the PMOS and Hyper groups. In addition, lipid metabolic pathways maintained relatively high activity in NC group but were modestly reduced in PMOS and Hyper samples. Macrophages from the Met-IR group displayed intermediate flux levels between NC and PMOS, suggesting that metformin-treated insulin-resistant PMOS is associated with a distinct macrophage metabolic state.

To systematically resolve neutrophil heterogeneity in PMOS follicular fluid, we performed unsupervised clustering and identified seven major neutrophil subpopulations (Additional file 1: Fig. S6F). Marker gene expression and Gene Ontology enrichment analyses revealed pronounced functional specialization among these subsets (Fig. 3H, Additional file 1: Fig. S6G and Tables S11-12). Inflammation-associated pathways—including cytokine-mediated signaling, positive regulation of chemokine production, and neutrophil degranulation—were predominantly driven by the Neu 3–5 subsets. Type I interferon signaling, indicative of antiviral responses, was most prominent in the Neu4 subset. Notably, Neu7 were significantly enriched for regulatory pathways, including regulation of TGF-β activation and negative regulation of cell death. Collectively, these findings indicate that neutrophils may play critical and subtype-specific roles in shaping the immune and inflammatory microenvironment of PMOS follicles (Fig. 3I and Additional file 1: Fig. S6H).

Metabolic impairment of GCs in hyperandrogenism

GCs dysfunction is widely regarded as a central cellular basis for follicular developmental arrest and ovulatory failure in PMOS [47]. In this study, we identified granulosa cells as one of the follicular fluid cell types exhibiting the highest PMOS genetic risk scores (Fig. 1F and Additional file 1: Fig. S3H-I), with particularly strong genetic associations observed in Hyper group, where GCs showed significant correlations with multiple PMOS-related diseases and metabolic traits (Additional file 1: Fig. S1E).

Subclustering analysis resolved nine distinct GC subpopulations (Fig. 4A-B, Additional file 1: Fig. S7A-C and Additonal file 2: Table S13), corresponding to discrete functional states along follicular maturation. GC1 and GC2 were characterized by high expression of FSHR, may indicating early-to-mid follicular developmental stages [48]. GC3 and GC4 exhibited elevated expression of SLC40A1 and SLC7A11, respectively, suggestive of differential capacities for iron homeostasis and enhanced antioxidant defence [49]. GC5, marked by high WNT5A expression, likely antagonises gonadotropin signaling through suppression of canonical WNT pathways [50]. GC6 and GC7 showed strong expression of EREG and AREG, consistent with luteinising hormone (LH)–responsive states that drive cumulus expansion, promote oocyte maturation, and initiate luteinization and follicular wall remodeling [51]. GC8, defined by PTX3 expression, likely represents a core LH–EGFR axis–induced cumulus expansion program, stabilising the extracellular matrix and supporting oocyte maturation while participating in peri-ovulatory inflammatory and remodelling processes [52]. In contrast, GC9 exhibited high expression of the pro-apoptotic gene BAX, indicative of post-ovulatory apoptotic initiation [53]. To further define the anatomical and functional identity of these GC subsets, we examined canonical cumulus and mural GC markers across subclusters (Additional file 1: Fig. S7D-E). Cumulus-associated genes were preferentially enriched in GC8, supporting its annotation as a cumulus expansion-associated population. In contrast, mural GC markers were more broadly expressed across GC1, GC2 and several steroidogenic or gonadotropin-responsive subsets, consistent with mural-like functional properties. Across clinical subgroups, the predicted proportions of cumulus-like, mural-like and intermediate GC states were broadly comparable, with no marked global shift in GC source composition (Additional file 1: Fig. S7F-G). These findings suggest that PMOS-associated GC alterations are more likely to reflect changes in transcriptional and functional states within existing GC lineages rather than a wholesale redistribution of cumulus-like versus mural-like compartments.

Fig. 4.

Pathogenic specialization of GC2 links PMOS genetic burden to disrupted steroid metabolism and metabolic–transcriptional rewiring. A UMAP plots of granulosa cells, colored according to cell type. B Heatmap showing the expression of top 50 significantly highly expressed genes for each granulosa subclusters, with Specific marker genes highlighted. Z score normalized mean expression. C UMAP projections showing cell-wise relevance scores for PMOS genetic traits. Higher (red) and lower (blue) scores indicate stronger or weaker trait relevance, respectively. D Violin plots (bottom) show scDRS-derived relevance scores for PMOS genetic traits across granulosa cell (GC) subclusters. The top scatter plot depicts the proportion of cells within each GC subcluster identified by scDRS as exhibiting significant trait-associated heterogeneity (P < 0.05). Point size denotes the odds ratio of heterogeneous cells within the subcluster, and colour indicates the FDR-adjusted P value. E Spearman correlation analysis showing genes positively or negatively associated with PMOS trait relevance scores in GC2 subcluster. Genes with significant positive or negative correlations are highlighted. F The Ro/e value of group in each granulosa subcluster (upper). Beeswarm plots showing changes in neighborhood abundance in Hyper-PMOS and Hyper-IR comparisons in GC2 subcluster. Log2FC, colored by spatial FDR < 0.1: red, Hyper high; blue, Hyper low. G Heat map showing the Z-score–normalized mean activity of hormone-associated metabolic flux pathways in each granulosa cell (GC) subcluster. P4 from Chol, synthesis of progesterone from cholesterol; E1 to E2, Conversion of estrone (E1) to estradiol-17beta (E2); E1 to E1S, Conversion of estrone (E1) to estrone sulfate (E1S); E1 from A4, Synthesis of estrone (E1) from androstenedione; E2 from A4, Synthesis of estradiol-17beta (E2) from androstenedione; A4 from Chol, Synthesis of androstenedione from cholesterol; T from A4, Synthesis of testosterone from androstenedione. H Dot plot showing differential expression of key enzymatic genes involved in hormone-associated metabolic pathways when comparing the Hyper group with the Met-IR, NC, and PMOS groups. Dot colour reflects log₂ fold change (red, higher in Hyper; blue, lower in Hyper), and dot size represents the − log₁₀ (adjusted P value) (MAST test). Only comparisons with |log₂ fold change|> 0.25 and adjusted P < 0.05 are shown. I Dot plot showing differential expression of key enzymatic genes involved in energy metabolic pathways when comparing the Hyper group with the Met-IR, NC, and PMOS groups. Dot color reflects log₂ fold change (red, higher in Hyper; blue, lower in Hyper), and dot size represents the − log₁₀ (adjusted P value). Only comparisons with |log₂ fold change|> 0.25 and adjusted P < 0.05 are shown. FA meta., fatty acid metabolism; Glycol., glycolysis; OXPHOS, Oxidative phosphorylation; Redox., redox metabolism; TCA&Amino., TCA cycle and amino acid metabolism. J Volcano plot of differentially expressed genes in Hyper group and others. MAST test, adjusted P < 0.05. Red, Hyper high, Blue, Hyper low. K Bar plots showing Gene Ontology (GO) biological processes enriched among genes downregulated (top) or upregulated (bottom) in the Hyper group compared with all other groups. L Left, heat map showing transcription factors up- or downregulated in each granulosa cell subcluster relative to all other subtypes (Wilcoxon test, FDR < 0.05). Right, differential expression of transcription factors in Hyper versus the NC, Met-IR, and PMOS groups across all granulosa cells, displayed as log₂ fold changes (adjusted P < 0.05, red indicates higher activity in Hyper)

To determine which GC subpopulations are most susceptible to PMOS genetic predisposition, we integrated scRNA-seq profiles with GWAS data using scDRS analysis. Multiple GC subtypes displayed significant associations with PMOS and related metabolic traits (Fig. 4C and Additional file 1: Fig. S8A). Quantitative comparison revealed that GC1_SOX4⁺, GC2_CYP19A1⁺, and GC3_SLC40A1⁺ populations exhibited markedly elevated scDRS risk scores, with GC2 contributing most prominently (Fig. 4D). Consistently, expression profiling of GWAS-linked candidate genes showed that PMOS-associated genes were distributed heterogeneously across GC subclusters (Additional file 1: Fig. S8B), further supporting the convergence of metabolic genetic susceptibility within the GC compartment. Dissection of molecular features underlying GC-specific genetic risk revealed a highly concordant regulatory architecture between GC1 and GC2 (Fig. 4E and Additional file 1: Fig. S8C). In these populations, genes involved in follicular growth (FOXL2), cell cycle regulation (H19, UBE2S, TUBB, H2AX, FOXM1), and steroidogenesis (INHA, HSD11B2) were positively correlated with genetic risk scores, collectively indicating heightened vulnerability to genetic risk during proliferative and early differentiation stages of granulosa cells [54–56]. In contrast, elevated genetic risk in GC3 was preferentially associated with increased expression of genes related to antigen presentation (HLA-A, HLA-C), iron storage (FTL), mitochondrial energy metabolism (ATP5MC2, NDUFB11), and oxidative stress defence (PRDX5) (Additional file 1: Fig. S8D).

Pseudotime trajectory analysis suggested that GC2 may represent an aberrantly differentiated preovulatory GC population (Additional file 1: Fig. S8E). Consistently, GC2 was most strongly enriched in the Hyper group, exceeding levels observed in uncomplicated PMOS and Met-IR groups (Fig. 4F and Additional file 1: Fig. S7C). Notably, GC-specific PMOS genetic risk was also significantly elevated in Hyper group (Additional file 1: Fig. S8F), underscoring the profound impact of preovulatory hyperandrogenism on granulosa cell function.

Given the endocrine–metabolic nature of PMOS [57], we next systematically evaluated hormone-related metabolic fluxes across GC subpopulations (Fig. 4G). GC2 exhibited the highest metabolic flux across multiple steroidogenic reactions, particularly in the conversion of androstenedione to estrone and estradiol. Differential flux analysis revealed that, relative to other groups, Hyper samples showed the most pronounced downregulation of metabolic fluxes in GC2, predominantly affecting fatty acid metabolism and gluconeogenesis, while selectively upregulating inositol synthesis and androstenedione-to-estrone conversion pathways (Additional file 1: Fig. S9A-C). Moreover, in GC2, the Hyper group displayed a distinctive pattern characterized by upregulation of androstenedione conversion to estrone and estradiol alongside suppression of other steroidogenic pathways (Additional file 1: Fig. S9D). Correspondingly, hyperandrogen-induced gene perturbation analysis revealed upregulation of CYP19A1 and HSD17B12, coupled with downregulation of CYP11A1 (Fig. 4H). This transcriptional signature aligns with previous reports describing selective enhancement of aromatisation-related pathways under hyperandrogenic conditions [58] (Additional file 1: Fig. S9E).

Beyond steroid metabolism, the Hyper group exhibited broad suppression across energy metabolic pathways. The overall activity of cholesterol homeostasis, fatty acid metabolism, glycolysis, and oxidative phosphorylation (OXPHOS) modules was significantly reduced (Additional file 1: Fig. S9F). Differential expression analysis revealed coordinated downregulation of genes involved in glycolysis (ALDOC, GPI, HK1, PKM) and fatty acid metabolism (FADS2, ACADM). In parallel, key subunits of mitochondrial respiratory complexes I–V (NDUFS6, NDUFB10, COX6A1, ATP5MC2, among others) were consistently decreased in Hyper samples (Fig. 4I and Additional file 1: Fig. S9G).

Gene set enrichment analysis further confirmed significant suppression of ATP metabolic processes, oxidative phosphorylation, aerobic respiratory chain activity, and mitochondrial electron transport in the Hyper group (Fig. 4J-K), indicating a global impairment of energy production. Network analysis identified hyperandrogen-specific upregulation of transcription factors such as PEG3, associated with cellular stress responses and apoptosis [59], while downregulated transcription factors, including STAT3, STAT5B and CEBPB, have been previously implicated in preovulatory GC regulatory networks governing steroidogenesis, endoplasmic reticulum homeostasis, and rapid LH-induced transcriptional responses [60, 61] (Fig. 4L and Additional file 1: Fig. S9H).

To substantiate these transcriptomic signatures, we further quantified follicular fluid metabolites and steroid hormones across subgroups. Consistent with impaired mitochondrial energy metabolism, the Hyper group showed reduced levels of citrate, fumarate hydratase, pyruvate and ATP compared with NC or PMOS samples, whereas β-hydroxybutyrate and lactate were increased (Fig. 5A-F). In parallel, steroid profiling revealed a marked elevation of androstenedione (ASD) and estrone (E1), accompanied by a reduction in estradiol (E2) and a pronounced decrease in the E2/ASD ratio in the Hyper group (Fig. 5G-I and Additional file 1: Fig. S10A). These changes indicate a metabolic–endocrine imbalance characterized by androgen accumulation, impaired conversion toward E2, and disrupted energy substrate utilization. Integrated analysis of metabolic and hormonal indicators further showed that ASD levels were positively correlated with lactate, β-hydroxybutyrate and the lactate/pyruvate ratio, but negatively correlated with ATP, citrate and fumarate hydratase (Fig. 5J-K and Additional file 1: Fig. S10B), linking hyperandrogenic status to altered redox balance and mitochondrial metabolic dysfunction within the follicular microenvironment.

Fig. 5.

Follicular-fluid biochemical profiling and 5α-dihydrotestosterone (DHT)-treated granulosa cell functional assays. A-I Box plots showing follicular-fluid levels of citrate (A), β-hydroxybutyrate (B), fumarate hydratase (C), pyruvate (D), lactate (E), ATP (F), estradiol (E2; G), estrone (E1; H), and androstenedione (ASD; I) across groups: NC (n = 9), PMOS (n = 12), Hyper (n = 18) and Met-IR (n = 8). Two-way ANOVA followed by Tukey’s multiple comparisons test was applied. J Heat map showing subgroup median Z scores for metabolic and hormone-related indicators. K Summary bar plot of Spearman correlations between ASD and the indicated energy-metabolic readouts. Positive correlations are shown in red and negative correlations in blue. L-N Seahorse mitochondrial stress test analysis of oxygen consumption rate (OCR) in granulosa cells treated with vehicle control (NC) or DHT at 500 nM, 1 μM or 2 μM for 2 days (L), 4 days (M) or 7 days (N). For each time point, representative OCR traces are shown together with matched lactate and ATP measurements. Two-way ANOVA followed by Tukey’s multiple comparisons test was applied. O Representative JC-1 fluorescence images and quantification of the JC-1 aggregate/monomer ratio in granulosa cells treated with vehicle or DHT for 4 days. Scale bars, 50 µm. Two-way ANOVA followed by Tukey’s multiple comparisons test was applied

To experimentally examine whether androgen exposure could contribute to granulosa-cell metabolic impairment, we treated KGN cells with increasing concentrations of dihydrotestosterone (DHT) for 2, 4 and 7 days. Short-term DHT exposure induced only modest changes in mitochondrial respiration, whereas prolonged or higher-dose exposure progressively suppressed basal respiration, maximal respiration and ATP-linked oxygen consumption, with the strongest effects observed after 4–7 days of treatment (Fig. 5L-N and Additional file 1: Fig. S10C-E). Consistently, DHT treatment reduced cellular ATP production and impaired mitochondrial membrane potential, as indicated by decreased JC-1 aggregate/monomer ratios (Fig. 5O). Although lactate levels varied across exposure duration and dosage, sustained androgenic stress compromises mitochondrial oxidative metabolism and energy production. Together, these results support the transcriptomic inference that hyperandrogenism is associated with coordinated disruption of steroidogenesis and energy metabolism in PMOS granulosa cells.

Cross-species assessment of steroidogenic pathway perturbation

To examine whether steroidogenic alterations observed in human PMOS granulosa cells could be recapitulated in an independent experimental model, we next analyzed publicly available single-cell transcriptomic data from ovaries of a dehydroepiandrosterone-induced mouse model of PMOS [62]. Reanalysis of the dataset revealed that, after excluding two low-quality clusters—one characterized by concurrent expression of markers from multiple cell types, likely reflecting doublet contamination (LowQC1), and another detected in only a single sample (LowQC2)—the overall cellular composition was largely preserved between control and PMOS model mice (Fig. 6A-C and Additional file 1: Fig. S11A-D).

Fig. 6.

Mouse ovarian analysis supports steroidogenic pathway remodeling in an androgen-excess PMOS-like model. A UMAP plot of mouse ovarian tissues and colored according to cell type in NC (normal control, n = 3), PMOS model (n = 3). B Dot plot of Cell type marker genes in mouse global dataset. C Cell-type proportions in mouse dataset, mean per group, and for each sample. D Transcriptomic flux-based analyses of metabolic pathways in PMOS model compared NC in each cell type. Task level, Cohen’s D, colored by FDR < 0.05 (Wilcoxon): red, model high; blue, model low; grey, nonsignificant. E Volcano plots showing metabolic flux differences in luteal cell from PMOS model compared with NC. Pathways meeting significance thresholds (Cohen’ s D > 0.3 and FDR < 0.05) are colored, with labelled upregulated hormone-associated metabolic flux pathways. F UMAP plot of granulosa cells in mouse ovarian dataset and colored according to cell type. G Dot plot of cell type marker genes in granulosa cells in mouse dataset. H Transcriptomic flux-based analyses of metabolic pathways in PMOS model compared NC in each granulosa cell. Task level, Cohen’s D, colored by FDR < 0.05 (Wilcoxon): red, model high; blue, model low; grey, nonsignificant. I Volcano plots showing metabolic flux differences in each granulosa subcluster from PMOS model compared with NC. Pathways meeting significance thresholds (Cohen’ s D > 0.3 and FDR < 0.05) are colored, with labelled hormone-associated metabolic flux pathways

Gene expression–based metabolic flux inference revealed that luteal cells from PMOS model mice exhibited increased activity of several steroid hormone biosynthesis-related pathways compared with controls, including reactions linked to the production of androstenedione, estrone and estradiol (Fig. 6D-E). Further subclustering of granulosa cells showed that increased inferred estradiol synthesis activity was detectable across multiple granulosa cell states rather than being confined to a single subpopulation (Fig. 6F-I and Additional file 1: Fig. S11E). These findings suggest that the DHEA-induced PMOS mouse model captures a component of ovarian steroidogenic remodeling, independent supporting that perturbation of steroidogenic pathway activity is a recurrent feature of PMOS-associated ovarian dysfunction.

Disrupted signaling networks in PMOS follicular microenvironment

Finally, we systematically interrogated ligand–receptor interactions among distinct cell types within the follicular fluid microenvironment to explore potential links between aberrant steroidogenesis and dysregulated local signaling in PMOS. Using CellPhoneDB [37], we observed heterogeneity among granulosa cell subpopulations in their capacity to sense and respond to endocrine cues. GC2 exhibited high expression of aromatisation-related genes, including CYP19A1, HSD17B1, and HSD17B12, together with steroid hormone receptors (ESR1, ESR2, and AR) and AMHR2 (Fig. 7A). This expression profile suggests that GC2 may be particularly responsive to AMH signaling, a pathway known to suppress FSH responsiveness and constrain follicular progression at immature stages [63]. In parallel, elevated expression of INHA in GC2 contrasted with activation of ACVR1B in other granulosa cell subpopulations, indicative of an imbalanced Activin–Inhibin signaling axis. Such a configuration is consistent with attenuation of pro-maturational TGF-β signaling and may exacerbate granulosa cell developmental dysregulation within the PMOS follicular environment [64].

Fig. 7.

Altered signaling interactions among granulosa cell subtypes in PMOS. A Dot plots showing expression of genes coding for hormone-related interacting ligand–receptor proteins in granulosa cells. Specific interacting partners are linked with a matching symbol. B Chord diagram illustrating WNT pathway–mediated cell–cell communication across granulosa and immune cell subsets. Ribbon thickness reflects the strength of predicted ligand–receptor interactions between source and target populations. C Scatter plot showing outgoing and incoming interaction strengths for WNT pathway signaling across granulosa and non-granulosa cell subsets; each point represents a cell subset. D Dot plots showing expression of genes coding for WNT-related interacting ligand–receptor proteins in granulosa cells. E Left, heat map showing predicted ligand-mediated outgoing interaction strengths for each granulosa cell subcluster; ligand genes are highlighted in red. Right, corresponding heat map showing predicted receptor-mediated incoming interaction strengths. F Left, chord diagram illustrating CXCL pathway–mediated cell–cell communication in all subsets. Ribbon thickness reflects the strength of predicted ligand–receptor interactions between source and target populations. Right, CXCL ligand expression level in granulosa subsets. G Ligand–receptor pathways with significant differential interactions in Hyper-NC, Hyper-PMOS and Hyper-IR comparisons. Colored by relative flow: red, Hyper high; blue, Hyper low; *FDR < 0.05. Infinity (Inf) represents pathways that were present in only one of the conditions. H Cell–cell interactions mediated by GAS signaling pathway in granulosa cells and immune cells

The intercellular communication network further highlighted a highly active WNT signaling module within granulosa cells. Global ligand–receptor flow analyses revealed that granulosa cells represent one of the principal sources and targets of WNT signaling in follicular fluid. Among these, GC1–GC3 displayed relatively stronger signal-receiving capacities, whereas GC4–GC9 acted as more active signal-sending populations (Fig. 7B-C). Gene-level inspection revealed pronounced heterogeneity at the ligand level. GC5 preferentially expressed the non-canonical WNT ligand WNT5A, whereas GC7 and GC8 were enriched for canonical ligands WNT3 and WNT2B (Fig. 7D). At the receptor level, GC2 expressed a broad repertoire of Frizzled family receptors (FZD3/4/5) along with the co-receptors LRP5/6, suggesting that GC2 may function as a dominant responder to WNT cues within the follicle (Fig. 7C). Collectively, these findings reveal a hierarchical and asymmetric organization of WNT signaling among granulosa cell subtypes. Beyond WNT pathways, multiple additional signaling modules displayed strong subpopulation specificity, pointing to multilayered and finely tuned communication both among granulosa cells and between granulosa cells and other follicular cell types (Fig. 7E and Additional file 1: Fig. S12A). Notably, all granulosa cell subpopulations except GC2 expressed a range of cell adhesion molecules, including CADM1, PTPRM, and CLDN11, implicating these populations in intercellular connectivity, and local barrier regulation. In contrast, GC2 preferentially contributed as a ligand source to the pro-inflammatory SEMA7A signaling axis [65]. In addition, GC5 and GC9 exhibited high expression of the chemokines CXCL1 and CXCL6 (Fig. 7F), suggesting a potential role in coordinating neutrophil recruitment during the peri-ovulatory inflammatory cascade [66].

To integrate signaling alterations across conditions, we next applied parallel communication analysis using CellChat [38], linking signaling strength to both source and target cell populations (Fig. 7G and Additional file 1: Fig. S12C). In communication patterns involving immune cells, the Hyper group exhibited significantly enhanced signaling through inflammation- and immune-associated ligand–receptor pairs, including IL10 and IL16. In contrast, several pathways associated with tissue repair, homeostatic maintenance, or anti-inflammatory regulation, such as ANXA1, NRG, OCLN, TGFβ, GAS, and PVR signaling, were markedly attenuated in the Hyper group. Together, these shifts reflect a global communication landscape characterized by heightened inflammatory tone coupled with weakened homeostatic restraint.

Consistent with this pattern, GC2 showed relatively low detectable expression of GAS6 and its cognate receptors AXL and MERTK compared with other granulosa cell states (Fig. 7H and Additional file 1: Fig. S12A), suggesting that this subpopulation may be less receptive to GAS–AXL/MERTK–mediated anti-inflammatory signaling. Taken together, these results indicate that inter-subpopulation communication among granulosa cells is imbalanced in PMOS follicular fluid. Under hyperandrogenic conditions, pro-inflammatory and maturation-inhibitory signals appear to be amplified, whereas pathways supporting follicular development, tissue homeostasis, and inflammatory resolution are attenuated.

Discussion

PMOS is a complex endocrine–metabolic disorder in which ovulatory dysfunction emerges from the convergence of genetic susceptibility, hormonal imbalance, metabolic stress, and immune dysregulation [57]. Our study provides a comprehensive single-cell atlas of PMOS follicular microenvironment, uncovering the multi-lineage coordination that drives ovulation disorders. By integrating large-scale scRNA-seq data from clinically distinct PMOS phenotypes with GWAS, we provide a mechanistic understanding of how genetic susceptibility translates into cellular dysfunction. While ovarian tissue–based single-cell studies have yielded valuable insights [67], they often capture static or mixed-stage signals. In contrast, our approach focuses on a precisely defined physiological window, just prior to ovulation, allowing us to resolve dynamic cellular states and interactions that are most relevant to follicular maturation and ovulatory failure. The inclusion of multiple PMOS subtypes, further enabled us to distinguish disease-associated alterations from treatment-responsive or potentially reversible features.

Granulosa cells constitute the central regulatory unit of the oocyte developmental microenvironment. Their metabolic homeostasis and signaling competence critically determine oocyte growth, cytoplasmic and nuclear maturation, fertilization potential, and subsequent pregnancy outcomes. Conversely, oocyte quality can be accurately inferred from the functional state of surrounding granulosa cells, reflecting a finely tuned bidirectional regulatory circuit between these two compartments [68, 69]. Granulosa and cumulus cells obtained during oocyte retrieval are clinically accessible yet remain underused for routine decision-making. Single-cell profiling of follicular-fluid-derived granulosa cells may therefore reveal disease-relevant cellular states and molecular signatures linked to impaired follicular development in PMOS. In this study, we resolved nine functionally distinct granulosa cell subpopulations. Importantly, our analyses highlight the GC2_CYP19A1⁺ subset as a focal cellular state in PMOS-associated follicular dysfunction. This population exhibited the highest enrichment of PMOS genetic risk and displayed a distinctive metabolic profile characterized by enhanced steroidogenic activity concomitant with impaired energy metabolism, which were strongly associated with hyperandrogenic conditions [70]. This discovery supplements the molecular details of how hyperandrogenism precisely regulates granulosa cell function, providing a new explanatory framework of "genetic susceptibility—metabolic reprogramming" for granulosa cell dysfunction [47].

By integrating GWAS-derived polygenic signals with follicular-fluid scRNA-seq and ovarian chromatin accessibility, our study further places granulosa-cell dysfunction within a regulatory-genetic framework. Recent large-scale and multi-ancestry GWAS have expanded the catalogue of PMOS-associated loci, but most risk variants reside in non-coding regions, making it essential to define the cellular and regulatory contexts in which these variants act [14]. scDRS analysis localized the strongest PMOS-related genetic burden to granulosa cells, with additional associations with bioavailable testosterone and fasting insulin, suggesting that the same follicular compartment may integrate reproductive, endocrine and metabolic components of PMOS risk [28]. The incorporation of ovarian snATAC-seq data extends this inference beyond transcriptional enrichment. S-LDSC and GWAS–DAR overlap analyses showed that PMOS-related heritability is not randomly distributed across the ovarian genome, but is preferentially embedded within cell type-associated regulatory elements, particularly those assigned to granulosa cells, theca cells and immune or stromal compartments. This pattern is biologically coherent: theca cells provide androgenic substrates, granulosa cells mediate FSH-dependent aromatisation and metabolic support, and immune–stromal cells shape the inflammatory and paracrine niche required for peri-ovulatory remodelling.

Although transcriptome-based metabolic modelling cannot directly measure enzymatic flux, metabolite turnover or ATP-generating capacity, it provided a useful framework for prioritizing metabolic programs that may be altered in the hyperandrogenic follicular microenvironment. In this context, the GC2_CYP19A1⁺ state should be interpreted not as definitive evidence of increased biochemical flux, but as a transcriptionally defined steroidogenic state in which aromatisation-related potential is uncoupled from broader energy metabolism. This interpretation is supported by complementary biochemical and functional data showing that hyperandrogenic follicles accumulated androstenedione and estrone while exhibiting reduced estradiol levels and a lower E2/ASD ratio, consistent with steroidogenic imbalance and inefficient conversion of androgenic substrates into bioactive oestrogens [54]. Concomitantly, follicular fluid displayed reduced ATP, citrate and pyruvate, together with elevated lactate, and androstenedione levels were negatively associated with mitochondrial metabolic indices but positively associated with lactate and the lactate/pyruvate ratio [71, 72]. These findings link androgen accumulation to impaired mitochondrial function, altered redox homeostasis and compensatory reliance on glycolytic or ketone-related metabolism. Consistently, prolonged DHT exposure progressively impaired OCR, intracellular ATP production and mitochondrial membrane potential, whereas short-term exposure exerted only modest effects, may indicating that sustained androgenic stress drives the transition from an adaptive steroidogenic response to mitochondrial insufficiency [73]. Given that follicular fluid reflects the integrated metabolic activity of granulosa cells, cumulus cells, immune cells and the peri-ovulatory inflammatory milieu, increased lactate likely represents both physiological metabolic coupling and pathological metabolic stress; in hyperandrogenic PMOS, this balance appears shifted towards lactate accumulation in the context of reduced mitochondrial bioenergetic capacity. Together, these data support hyperandrogenic PMOS follicles granulosa cells in a incomplete oestrogenic conversion and impaired ATP production microenvironment, ultimately compromising the metabolic and paracrine support required for oocyte maturation and contributing to follicular arrest and ovulatory dysfunction.

Recent single-cell profiling of PMOS follicular fluid supports the notion that ovulatory immune programs are compartment-specific and persistently dysregulated in PMOS, even under ovarian stimulation [10, 74]. Our data extend this framework by showing that hyperandrogenism imposes a network-level bias—amplifying inflammatory and maturation-inhibitory signaling while dampening tissue-repair and resolution pathways—thereby embedding granulosa cell endocrine–metabolic dysfunction into a broader immune–communication cascade. Mechanistically, emerging studies linking androgen excess to mitochondrial stress–innate immune activation and to inflammatory microenvironment rewiring offer plausible routes through which endocrine stress can be converted into persistent inflammatory tone within the follicle [75–77].

Several limitations should be considered. First, although this study profiled a large number of single cells, the number of healthy control donors was limited, which reduces statistical power and warrants validation in larger, independently collected cohorts. Second, the Met-IR subgroup represents insulin-resistant PMOS under metformin treatment rather than untreated insulin resistance; thus, cellular and metabolic features in this group may reflect the combined effects of insulin resistance, metformin exposure and inter-individual heterogeneity. Third, although follicular-fluid biochemical profiling and DHT-treated granulosa-cell assays support the transcriptomic findings, transcriptome-based metabolic modelling and ligand–receptor inference remain indirect measures of biochemical flux and intercellular signalling, and low-abundance transcripts may be affected by dropout. Finally, cumulus-associated and follicular-fluid-derived cells were combined before single-cell capture; marker-based inference suggests that disease-associated granulosa-cell states are not solely explained by cumulus–mural origin, but definitive anatomical assignment will require physically separated or spatially resolved profiling. Future studies integrating longitudinal follicular sampling, primary-cell perturbation, spatial proteomic or metabolomic validation, and clinical outcome linkage will be required to define the reversibility and functional consequences of these cellular states.

Conclusions

In summary, our study provides a cell-resolved framework linking PMOS (historically studied as PCOS) genetic susceptibility to endocrine–metabolic and immune remodelling within the preovulatory follicular microenvironment. We identify granulosa cells, particularly a hyperandrogen-associated GC2_CYP19A1⁺ state, as a candidate cellular node in which steroidogenic imbalance is coupled to impaired mitochondrial bioenergetics and altered follicular-fluid metabolism. Together with the observed shifts in immune composition and ligand–receptor signalling, these findings suggest that follicular arrest in PMOS reflects a coordinated disruption of granulosa-cell metabolism, endocrine signalling and immune homeostasis, rather than altered hormone levels alone. By defining clinically accessible cellular states and signalling programmes associated with PMOS subtypes, this atlas provides a resource for future validation studies and may inform microenvironment-guided stratification and therapeutic development aimed at restoring follicular function.

Abbreviations

PMOS

Polyendocrine metabolic ovarian syndrome

PCOS

Polycystic ovary syndrome

DHT

Dihydrotestosteron

DHEA

Dehydroepiandrosterone

ASD

Androstenedione

GWAS

Genome-wide association studies

scRNA-seq

Single-cell RNA sequencing

IVF

In vitro fertilization

hCG

Human chorionic gonadotropin

FBS

Fetal bovine serum

PBS

Phosphate-buffered saline

ExoI

Exonuclease I

UMI

Unique molecular identifier

RPE

Random primer extension

DARs

Differentially accessible regions

GC

Granulosa cells

TC

Theca cells

EpiC

Epithelial cells

EC

Endothelial cells

IC

Immune cells

SMC

Smooth muscle/stromal cells

S-LDSC

Stratified LD score regression

OR

Odds ratios

ULM

Univariate linear model

scDRS

Single-cell disease relevance score

BMI

Body mass index

NK

Natural killer

Tem

Effector memory T cells

Trm

Tissue resident memory T cells

Temra

Recently activated effector memory T cells

Treg

Regulatory T cell

LH

Luteinising hormone

OXPHOS

Oxidative phosphorylation

Authors’ contributions

C.Y., J.Y and Z.N. conceived and designed the project. D. S, X.Li. and J.D collected the human samples and clinical information. Y-S.L. performed the computational analysis help form Q.Z. and X.Liang. Y-S.L., D.Z. and Y.L. drafted the manuscript. M.L provided relevant advice and reviewed the manuscript. Z.N. and C.Y. supervised the work. All authors read and approved the final manuscript.

Funding

This study was supported by the National Natural Science Foundation of China (82474281, 82405469), Shanghai “Science and Technology Innovation Action Plan” by Science and Technology Commission of Shanghai Municipality (23ZR1478600), Medical New Technology Innovation Cultivation Project of the First Affiliated Hospital of Naval Medical University (ZXJS2024B20), National Outstanding Young Scientist Training Program by Chinese Academy of Sciences (ZZ18-YQ-026) and Subordinate Project of Guang’anmen Hospital by China Academy of Chinese Medical Sciences (2022s477).

Data availability

The sequencing dataset supporting the conclusions of this article is available in the Genome Sequence Archive for Human (GSA-Human), National Genomics Data Center, under accession number HRA015615 [78]. And ovary scATAC-seq dataset is available from the National Institutes of Health (NIH) BioProject, accession no. PRJNA836755 (Gene Expression Omnibus (GEO) ID: GSE202601) [29]. https://ngdc.cncb.ac.cn/gsa-human.

Declarations

Ethics approval and consent to participate

All study participants provided informed written consent, and the study was approved by the local ethics committees and conducted in accordance with the principles of the Helsinki Declaration. This study was approved by the Medical Ethics Committee of the First Affiliated Hospital of Naval Medical University (approval number: CHEC2025-132).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.