Whole genome methylation profiling of menstrual stem cells identifies novel biomarkers for endometriosis

Abstract

Background

Endometriosis, despite its high prevalence, is underdiagnosed and poorly managed due to lack of clinically validated biomarkers and pathophysiological insight. Menstrual blood-derived stem cells have been implicated in disease pathogenesis, but their diagnostic potential remains unexplored.

Methods

We conducted a case-control clinical study in women (n = 42; 19 endometriosis, 23 controls). Menstrual blood samples were collected, and menstrual blood-derived stem cells were isolated for whole-genome DNA methylation sequencing. Differential methylation analysis was performed to identify disease-specific epigenetic biomarkers, and machine learning models were applied to evaluate the diagnostic performance of candidate markers. An external endometrial single-cell RNA sequencing atlas including endometriosis samples was employed to correlate RNA expression with the identified disease-specific methylation signature.

Results

Here we identify differentially methylated regions enriched in genes linked to hallmarks of endometriosis such as inflammation, tissue remodelling and development. These differentially methylated regions robustly distinguish cases from controls, independent of technical and clinical variables. Machine learning models trained and validated on these differentially methylated regions achieve high diagnostic performance (specificity 83%, sensitivity 79%). Integration with an independent single-cell RNA sequencing dataset shows that the differentially methylated regions may modulate gene expression, further supporting their biological relevance.

Conclusions

These findings position menstrual blood-derived stem cell DNA methylation profiling as a promising, non-invasive approach for early endometriosis diagnosis and personalised care.

Plain language summary

Endometriosis affects about 1 in 10 women of reproductive age and can cause long-term pain and fertility problems. Yet many people wait 7 to 10 years for a diagnosis. This study explored whether menstrual-derived stem cells, which are found in menstrual blood and are linked to endometriosis, could help identify the condition earlier by providing biological markers. We analysed DNA methylation, patterns of genetic regulation in these cells. These DNA methylation patterns differed between patients with endometriosis and healthy individuals and were connected to biological processes involved in the disease. These findings suggest that menstrual blood could offer a simple, non-invasive way to detect endometriosis earlier, improve understanding of the condition and guide personalised care.

Tiniakou et al perform DNA methylation profiling on menstrual blood stem cells from healthy and endometriosis participants to identify epigenetic biomarkers. They find differentially methylated regions that distinguish cases from controls with high accuracy in machine learning models and influence gene expression, supporting biological relevance.

Introduction

Endometriosis is a common gynaecological disorder characterised by extra-uterine growth of endometrial-like tissue and chronic low-grade inflammation¹. Despite an estimated prevalence of 10% among women of reproductive age, diagnosis is typically delayed by an average of 7 to 10 years². This stems from a limited understanding of disease pathogenesis and a lack of reliable non-invasive diagnostics, often leading to misdiagnosis³. Endometriosis presents with diverse non-specific symptoms that may not reflect disease severity, rendering clinical assessment and symptom management challenging. It often causes debilitating pelvic pain, fatigue, and infertility, significantly impairing quality of life and mental health. Historically, the gold standard for diagnosis has been laparoscopic surgery, during which endometriotic tissue biopsies are obtained for histological confirmation. To reduce diagnostic delay and facilitate early intervention, recent guidelines from the European Society of Human Reproduction and Embryology recommend a combined assessment of clinical and imaging findings to establish a diagnosis⁴. These guidelines strongly advise against the use of biomarkers for diagnostic purposes, citing insufficient clinical validation and reproducibility. Post-diagnosis, endometriosis management remains complex as existing staging systems offer limited prognostic value and poorly correlate with symptom severity or treatment response. Molecular biomarkers may address these gaps by advancing the understanding of disease mechanisms, enabling earlier detection, and identifying disease subtypes linked to symptomatology and therapeutic response.

Menstrual blood has emerged as a valuable, non-invasive source of stem cells with mesenchymal properties, commonly referred to as menstrual blood-derived stem cells (MenSCs)⁵. When localised in the perivascular regions of the basalis and functionalis layers of the endometrium, these cells are termed as endometrial mesenchymal stem cells (eMSCs) and contribute to endometrial regeneration and repair during the menstrual cycle⁶. MenSCs exhibit high proliferative and migratory capacity, robust colony-forming ability and the potential for multilineage differentiation⁷. These features, combined with their endometrial origin, have positioned MenSCs at the centre of several theories on endometriosis pathogenesis.

Retrograde menstruation, the leading proposed mechanism underlying endometriosis, suggests that viable endometrial cells, including MenSCs, reflux through the fallopian tubes into the pelvic cavity where they can implant, grow and form endometriotic lesions⁸. Although retrograde menstruation is common (approximately 90% of women)⁹, only a subset of individuals develop endometriosis, suggesting differences in MenSCs may contribute to disease susceptibility¹⁰. The stem cell theory further proposes that MenSCs can disseminate through lymphatic or haematogenous routes, resulting in lesions at both pelvic and extra-pelvic sites¹¹. The high proliferative and migratory potential of MenSCs, demonstrated both in vitro and in vivo, lends support to this mechanism^12,13.

MenSCs express mesenchymal stem cell markers such as CD105, CD90, CD73, CD44 and CD29^14,15 that are also preserved in eMSCs found in endometriotic lesions¹⁶. In endometriosis patients, MenSCs show dysregulation of genes associated with proliferation, immune modulation, and hormone responsiveness¹⁷. These findings suggest that MenSCs contribute to both lesion initiation and disease maintenance, underscoring their diagnostic and therapeutic potential.

To date, studies attempting to characterise MenSCs have predominantly utilised cultured cells. However, even short-term in vitro culture is known to alter the epigenetic and transcriptional landscapes of cells, thus limiting the relevance of such findings¹⁸. In contrast, direct isolation from menstrual blood allows for non-invasive, phase-standardised sampling of MenSCs, circumventing the need for cell culture. This approach ensures high specificity by targeting a single disease-relevant population, thereby enhancing molecular resolution and enabling precise analysis of native omic states in both health and disease. Among these omic features, DNA methylation stands out for its chemical stability, resistance to short-term physiological fluctuations, and growing application in in vitro diagnostic devices¹⁹. In endometriosis, altered methylation has been linked to modified gene expression, leading to increased cellular proliferation, invasion, and progesterone resistance²⁰.

Here, we investigate whether methylomic profiling of MenSCs can identify biomarkers for endometriosis. We perform enzymatic methylation sequencing (EM-seq) on fluorescence-activated cell sorting (FACS)-isolated MenSCs from women with or without endometriosis and identify differentially methylated regions (DMRs) linked to endometriosis-related pathways. We proceed to use machine learning models trained on MenSC-derived methylation data to evaluate the potential use of the identified DMRs as non-invasive biomarkers. Finally, we explore the functional relevance of these methylation patterns by integrating them with publicly available single-cell RNA-seq (scRNA-seq) data from eutopic endometrium, revealing transcriptional changes associated with the disease. Taken together, our findings provide insights into the epigenetic landscape of endometriosis and highlight MenSCs as a promising source of disease-specific biomarkers.

Materials and methods

Ethics statement

The study was approved by the Ethics Committee for Investigation with medicinal products (CEIm) of both hospitals: the Ethics Committee for Clinical Research of Hospital Clínic de Barcelona (reference number: HCB/2023/0050) and the Ethics Committee of Las Palmas H.U.G.C. Dr. Negrín (reference number: 2022-435-1). Written informed consent was obtained from all participants. All experiments were conducted in compliance with the ICH E6 Good Clinical Practice guidelines and the Declaration of Helsinki.

Participant recruitment

Participants were recruited through the Hospital Clínic de Barcelona (Barcelona, Spain) and the Complejo Hospitalario Universitario Insular Materno Infantil (CHUIMI; Gran Canaria, Spain) according to the following inclusion criteria: (i) premenopausal women aged 18 to 45 years; (ii) BMI of 18–27 kg/m²; (iii) history of spontaneous regular menstrual cycles (21–35 days); (iv) no menstrual cycle-altering drug treatment (e.g. oral contraceptive pills) for at least 3 months prior to participation. The number of patients recruited was similar across the two centres. Participants in the endometriosis group had a confirmed diagnosis of ovarian endometriosis, identified by laparoscopy and/or imaging within the previous year, in accordance with the ESHRE endometriosis diagnosis guidelines⁴. While ovarian endometriosis served as a minimum inclusion criterion, many participants presented with additional lesion types (see Table 1). Eligible endometriosis participants reported pelvic pain symptoms commonly associated with the condition, such as dysmenorrhea, dyspareunia, and dyschezia. Non-endometriosis participants had no history or suspicion of endometriosis, had endometriosis ruled out by the same diagnostic methods, and reported no pelvic pain or infertility issues. All participants were further excluded if (i) pregnant or breastfeeding; (ii) used an intrauterine birth control device; (iii) experienced abnormal gynaecological bleeding without a known cause; (iv) had history of severe cardiac, respiratory, renal, endocrine or haematological diseases; or (v) were diagnosed with any type of cancer and/or autoimmune disease. The clinical pilot study was conducted from March 2023 to September 2024 and included participant recruitment, clinical data and sample collection.

Table 1.

Demographic and clinical characteristics of study participants

	Control	Endometriosis	p
n	23	19
Ethnicity (n (%))			0.07#
Black/African	0 (0.0)	1 (5.3)
Hispanic/Latina	0 (0.0)	3 (15.8)
White/European	23 (100.0)	15 (78.9)
Age (years)	33.04 (6.34)	37.11 (3.93)	0.02
Parity (n)	1.50 (0.53)	1.00 (1.10)	0.34
Age of menarche (years)	12.57 (1.85)	12.05 (1.22)	0.29
Duration of menstrual bleeding (days)	5.00 (1.28)	4.79 (1.23)	0.59
BMI (kg/m²)	21.55 (2.32)	21.94 (2.62)	0.62
Other type of endometriosis (%)
DE	NA	5 (26.3)
DE, Adenomyosis	NA	4 (21.1)
DE, Adenomyosis, Peritoneal, Other	NA	1 (5.3)
Peritoneal	NA	1 (5.3)
Chronic pelvic pain (n (%))	0 (0.0)	8 (42.1)	0.002 #
	Control	Endometriosis	p
Family history of endometriosis (n (%))	0 (0.0)	3 (15.8)	0.17 #
Family history of cancer (n (%))	10 (43.5)	11 (57.9)	0.54 #

The results are presented as mean (SD) and p-value refers to two-tailed t-test, unless stated otherwise.

BMI body mass index, DE deep endometriosis.

^#chi-squared test.

Menstrual blood sample collection and initial processing

Participants collected menstrual blood overnight on the first or second day of their cycle using menstrual cups. Upon collection, menstrual blood was immediately transferred into sterile 50 mL conical tubes containing 30 mL collection medium (DMEM/F12, 1% Pen/Strep 100 U/mL, 1% Amphotericin B 100×, 1% L-Glutamine 2 mM, EDTA 2 mM) and maintained at 4 °C until further processing. Menstrual blood samples collected in Hospital Clínic de Barcelona were shipped at 4 °C and processed at endogene.bio’s laboratories within 48 h. Upon receipt, sample temperature and volume were measured, and a rapid HIV test (ReLab) was performed. Samples collected in CHUIMI were processed in the hospital’s laboratories using the same standardised protocol. The isolated mononuclear cells were frozen and shipped to endogene.bio’s laboratories for further processing.

Mononuclear cell isolation by density gradient centrifugation

Mononuclear cells were isolated from collected samples by centrifugation on a Ficoll gradient. Briefly, menstrual blood in the collection medium was washed with 1× PBS (Gibco) and filtered to remove blood clots and tissue. The filtered suspension was slowly overlaid on Ficoll-Paque PREMIUM (Cytiva) and mononuclear cells were isolated by density gradient centrifugation²¹.

Flow cytometry and MenSC isolation by FACS

Menstrual blood-derived mononuclear cells prepared as described above were washed and counted on a MACSQuant Analyzer 10 (Miltenyi Biotec). Cells were subsequently stained with surface antibodies (Supplementary Table 1) for multicolour analysis and sorting. Lineage (Lin) cocktail included antibodies to human CD45, CD34, CD14, CD19 and HLA-DR. Fluorochrome-conjugated antibodies against human CD90, CD73, CD105 and CD44 were used to identify the MenSC population. Viobility 405/520 Fixable dye (Miltenyi Biotec; catalogue #130-130-404; lot #5231110365 and 5240501993) was used to assess cell viability, and staining was performed at a 1:100 dilution in 1× PBS for 15 min at RT in the dark. Staining for extracellular markers was performed in FACS buffer (1× PBS, 0.5% BSA, 2 mM EDTA) for 20 min at 4 °C in the dark. Samples were acquired on MACSQuant Analyzer 10 (Miltenyi Biotec) and further analysed with FlowJo v10.10.0 software (Tree Star). Cell sorting was performed on the MACSQuant Tyto Cell Sorter (Miltenyi Biotec). Statistical analysis of flow cytometry data was performed using Prism (GraphPad). Normal distribution of data was not assumed, and statistical significance of differences between experimental groups was determined by the non-parametric Mann–Whitney test.

Sample preparation for EM-seq

Genomic DNA was isolated using NucleoSpin Tissue XS or NucleoSpin Tissue Mini kit (Macherey-Nagel) and quantified on a Qubit 4 Fluorometer using the Qubit dsDNA High Sensitivity Assay Kit. DNA quantities ranging from 0.5 to 10 ng were sheared to ~350–550 bp length by Covaris LE220 Plus and used as input to generate sequencing libraries. The libraries were prepared and amplified using the NEBNext Enzymatic Methyl-seq Kit (New England Biolabs) according to the manufacturer’s instructions, adjusting the PCR cycles based on the starting DNA quantity. The sequencing libraries passing technical QC were sequenced in 150 bp paired-end mode on Illumina NovaSeq X Plus using 6–8% unmethylated PhiX Control v3 (Illumina).

EM-seq data quality control

Fastq files were pre-processed to assess data quality and identify potential sequencing issues. Briefly, the quality of reads was assessed using FastQC²², followed by adapter, 10 bp low-quality ends and polyA tail trimming with TrimGalore²³. Reads with less than 20 bp after trimming were discarded. To investigate cross-species contamination, FastQ Screen was used to perform mapping of samples with a high percentage of duplicates (> 60%)²⁴.

To maintain data integrity, samples exhibiting both low MenSC counts post-sorting (< 250 cells, corresponding to approximately 1.6 ng gDNA if fully recovered) and a high proportion of duplicate reads post-sequencing (> 80%) were excluded (n = 9). This dual criterion strategy minimised technical artifacts and ensured that only high-confidence samples were retained for downstream analyses. Additionally, quality control metrics were calculated for each sample, including the percent covered fraction, defined as the proportion of CpG sites with at least one read, and coverage, representing the number of sequencing reads covering each CpG site. Samples with a covered genome fraction of less than 60% and with a mean coverage below 1× were excluded from further analysis. After applying these quality control filters, 42 high-confidence samples (19 endometriosis and 23 controls) were retained for downstream analyses (Table 1).

EM-seq data read alignment and methylation calling

Trimmed reads were mapped to the human GRCh38 reference genome using the default parameters of BSBolt (v1.6.0), which integrates a modified BWA-MEM algorithm optimised for both bisulfite and enzymatic methylation sequencing data²⁵. Post-alignment, duplicate reads were removed using samtools (v1.21)²⁶. Methylation calling was performed using BSBolt with default settings, generating output in CGmap format. These files were then converted to Bismark-compatible CpG report format using a custom Python script to facilitate downstream analysis.

EM-seq data downstream analysis

The total number of CpGs analysed was assessed using the comethyl (v1.3.0) package²⁷ to explore various coverage thresholds and shared CpG positions across samples. Coverage thresholds ranging from 1× to 10× per CpG, and the proportion of samples required for a CpG to be considered valid, were evaluated. Final cutoffs were set at a minimum coverage of 3× in at least 80% of samples. The distribution of methylation positions with a minimum coverage of 3× was plotted to assess conformity of the methylation profiles to a binomial distribution (Supplementary Fig. 1). After applying these cutoffs, dmrseq²⁸ and bsseq²⁹ (implemented via DMRichR (v1.7.8)³⁰) were used to identify and calculate DMRs in the dataset. The discovery process utilised a smoothing and weighting algorithm to adjust CpG counts based on coverage. Candidate background regions were defined by grouping CpGs with similar genomic proximity and methylation levels. Statistical significance of the regions was determined through permutation testing, with empirical p-values calculated by comparing observed test statistics against a null distribution generated from 41 permutations. Analysis parameters included a minimum CpG coverage of 3×, at least five CpGs per DMR present in 80% of samples, and a single-CpG coefficient cutoff corresponding to a 6.5% methylation difference, with adjustment for age.

Finally, DMRs were annotated using a combination of the annotatr (v1.32.0)³¹ and ChIPseeker (v1.42.1)³² R packages. Promoter regions were first defined using the build_annotations function from annotatr with the hg38 genome build. DMRs were then annotated to genomic features (e.g. promoters, introns, exons) using the annotate_regions function from the DMRichR pipeline, incorporating genome-wide annotation tracks. To assign each DMR to the nearest gene, we used annotatePeak from ChIPseeker with the TxDb.Hsapiens.UCSC.hg38.knownGene and org.Hs.eg.db databases. The resulting annotations included gene symbols, gene names, and functional context relative to transcription start sites.

Selection of informative DMRs

From the identified DMRs, the methylLearn function was used to select the most informative regions distinguishing between endometriosis and control samples. This function applies Random Forest (RF) and Support Vector Machine (SVM) algorithms to identify key DMRs. We selected the top 5% of DMRs that were consistently prioritised by both algorithms. To evaluate these DMRs, Principal Component Analysis (PCA) was performed using the PCAtools (v2.20.0) package³³. Data were scaled and centred with the pca function, and correlations between DMRs and various clinical or experimental variables were assessed using eigencorplot.

Enrichment analysis

Overrepresentation analysis (ORA) was performed using the WebGestaltR (v0.4.6) package³⁴, considering only genes corresponding to hypermethylated DMRs. GO Biological Process terms were selected as the ontological framework, with minimum and maximum category sizes set to 10 and 500, respectively. All genes annotated in the hg38 genome assembly were used as the reference set, and genes identified in the DMR signature served as the input list. Statistical significance was determined using false discovery rate (FDR) adjusted p-values, with a cutoff of 0.05.

Machine learning

To prevent potential data leakage into the training set and ensure reliable test results, given the limited dataset size, an leave-one-out (LOO) data split strategy was implemented. The dataset was divided into 42 folds, with each fold consisting of 41 samples for training and 1 independent sample for testing. DMR detection was performed on the training data using the same settings described in the ‘EM-seq Data Downstream Analysis’ section. The resulting CpG matrices were smoothed, and features were derived from the top 2 to 20 DMRs, ranked by imputed p-value. These reduced matrices were then used to train machine learning models using the Python library AutoGluon³⁵ with the ‘medium_quality’ setting. The following models were trained: LightGBM, CatBoost, XGBoost, Neural Networks (Torch and FastAI), Random Forest (Gini and Entropy), Extra Trees (Gini and Entropy), Weighted Ensemble, and k-Nearest Neighbors (kNN). Accuracy was used as the primary metric during training. For each fold and each number of DMRs used, predictions on the test sample (excluded from training) were collected to evaluate overall model performance in the test phase. Final performance metrics included sensitivity, specificity, and accuracy.

Methylation and scRNA-seq validation data

The impact of the DMRs identified in this study on the transcriptome was investigated using the state-of-the-art scRNA-seq dataset of eutopic endometrium from the Human Endometrial Cell Atlas (HECA) dataset^36,37. This dataset comprises 313,527 cells from 63 women, both with and without endometriosis, integrated from multiple studies. First, the mesenchymal lineage was extracted and re-clustered using the Seurat R (v5.1.0) package. The count matrix was normalised by dividing each cell’s feature counts by the total counts for that cell, multiplying by a scale factor, and then applying a natural log transformation using the log1p function. Next, the most variable RNA features were identified, and the top 2000 were selected for integration using the SCTransform function. Data integration was performed using the harmony (v1.2.3) package and integrated by ‘dataset’ variable to remove batch effects related to dataset origin. PCA was performed and the top 30 PCs were used to compute the Uniform Manifold Approximation and Projection (UMAP). The FindNeighbors and FindClusters functions were applied for graph-based clustering by constructing a kNN graph using Euclidean distance in the PCA space. Clusters were subsequently defined using the Louvain algorithm to optimise the standard modularity function.

To detect the stromal stem cell population within the mesenchymal lineage, the AddModuleScore was used with previously established MenSC marker genes (THY1, NT5E, CD44). This function calculates the average expression of a specified gene set per cell, subtracting the aggregated expression of control feature sets. All analysed features are binned based on average expression, and control features are randomly selected from each bin. Log-normalised counts were used for this analysis. A cut-off of  ≥ 0.8 was used to select highly enriched cells expressing MenSC-associated markers³⁸. Finally, the stromal cells from the menstrual/early proliferative phase were exclusively selected to perform downstream analysis. Differential gene expression (DGE) analysis was performed on the selected cells using the Wilcoxon test, with p-values corrected by the FDR. Finally, the DMR and DGE results were integrated to assess the correlation between the DMR signature identified in MenSCs and the DEGs identified in the stromal stem cell population from an external eutopic endometrium dataset.

Statistics and reproducibility

Statistical analyses of sample-associated and flow cytometry data were conducted using Prism (GraphPad). Differences between experimental groups were analysed using two-tailed non-parametric Mann–Whitney test, as normal distribution of data was not assumed, and p < 0.05 was considered statistically significant.

Statistical analyses of bioinformatic and clinical data were performed using R (v4.4.0) and Python (v3.10.14). Unless otherwise specified, statistical tests were two-sided and a p-value < 0.05 was considered statistically significant.

Comparisons of clinical and demographic variables between groups were performed using two-tailed Student’s t-tests for continuous variables and chi-squared tests for categorical variables, as indicated. Summary statistics are reported as mean and standard deviation (SD) for continuous variables and as counts and percentages for categorical variables.

Statistical inference for differential methylation analyses was conducted using region-based permutation testing as implemented in dmrseq, with empirical p-values derived from the null distributions generated during permutations. For machine learning analyses, each biological sample represented one independent observation. Model performance was assessed using LOO cross-validation, and evaluation metrics were calculated on held-out samples only. We used the following metrics to measure the model performance:

$$\mathrm{Sensitivity}=\mathrm{TP}/\left(\mathrm{TP}+\mathrm{FN}\right)$$

$$\mathrm{Specificity}=\mathrm{TN}/\left(\mathrm{TN}+\mathrm{FP}\right)$$

$$\mathrm{Accuracy}=\left(\mathrm{TP}+\mathrm{TN}\right)/\left(\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}\right),$$

where TP, TN, FP and FN refer to true positive, true negative, false positive and false negative.

We performed ORA and considered results significant at FDR < 0.05. For single-cell RNA-seq validation analyses, statistical testing for DGE was performed using non-parametric tests, Wilcoxon test, as specified.

Sample sizes are indicated in the figure legends and reflect the number of unique biological samples included in each analysis. No technical replicates were included. Missing data were not imputed unless stated otherwise.

Results

Endometriosis and control samples exhibit similar metrics and processing outcomes

Participant recruitment was conducted across two clinical centres, and sample processing was performed as outlined in Fig. 1a (for details see ‘Methods’ section). A total of 42 participants, including 19 diagnosed with endometriosis and 23 non-endometriosis controls, were included in the final analysis (Supplementary Fig. 2). The cohort was predominantly composed of individuals of White European ancestry (approximately 90%), with three participants of Hispanic and one of African descent, all of whom belonged to the endometriosis group (Table 1). Participants in the control and endometriosis groups had mean BMIs of 21.55 and 21.94, mean ages of 33.04 and 37.11 years, mean ages of menarche of 12.57 and 12.05 years, and mean durations of menstrual bleeding of 5.00 and 4.79 days, respectively (Table 1). Menstrual blood volumes and flow rates were similar between the two study groups (median volume: 5.0 vs 5.0 mL; median flow rate: 0.50 vs 0.54 mL/h, control vs endometriosis), with no significant differences detected (Fig. 1b, and Supplementary Data 1). To isolate MenSCs from menstrual blood we performed a two-step procedure involving mononuclear cell isolation followed by sorting of the target cells. Quantitative analysis revealed no significant differences in mononuclear cell yield (median total cell count: 15 × 10⁶ vs 20 × 10⁶; median cell count per mL of MB: 4 × 10⁶ vs 4 × 10⁶, control vs endometriosis) or in the number of sorted MenSCs (median total cell count: 4293 vs 4904; median cell count per mL of MB: 1112 vs 1639, control vs endometriosis) between the two groups (Fig. 1c, d, and Supplementary Data 1). To achieve selective isolation, we characterised MenSCs using a combination of multiple negative selection markers along with four established positive markers specific to this cell type (CD90, CD73, CD105, CD44; see details in ‘Methods’; Fig. 1e). No significant differences were observed in the frequency of MenSCs among live cells between the two groups (median frequency: 0.37 vs 0.32%, control vs endometriosis; Fig. 1f and Supplementary Data 1). EM-seq of sorted MenSCs from each sample yielded methylation profiles, which were subsequently analysed using several downstream approaches (Fig. 1g). These included exploratory data analysis, identification of an endometriosis-associated DMR signature, development of a machine learning-based diagnostic model, and independent validation of the DMR signature’s impact on the eutopic endometrium transcriptome.

Fig. 1. Experimental workflow and quantitative profiling of MenSC isolation from menstrual blood.

a Schematic of participant recruitment and sample processing workflow. b Menstrual blood volume and flow rate collected overnight among participants of the different collection centres. Shown are box plots of menstrual blood volume or volume per hour. Absolute numbers of c Ficoll-isolated mononuclear cells and d sorted MenSCs. Shown are box plots of total cell numbers and cell numbers per mL of collected menstrual blood. e Phenotypic profile of MenSCs (Lin- CD90 + CD73 + CD105 + CD44 + ) within the mononuclear cell population in menstrual blood. Shown is representative staining plot corresponding to the gating strategy used to isolate the cells by FACS. Cells were gated on viable cells. Numbers represent frequencies among total live cells. Lineage was defined as CD45 + CD34 + CD14 + CD19 + HLA-DR + . f Frequency of sorted MenSCs among total live cells. g Schematic of bioinformatic analysis workflow. Box plots include values from n = 23 control and n = 19 endometriosis samples, except in c (n = 14 control, n = 17 endometriosis). Centre lines in box plots indicate the median; whiskers denote the minimum and maximum values. Statistical significance was determined using a two-tailed Mann–Whitney test. Exact p-values are as indicated in the respective graphs. BMI body mass index, endo endometriosis group, CHUIMI Complejo Hospitalario Universitario Insular Materno Infantil, MB menstrual blood, MenSC menstrual-blood derived stem cell, QC quality control, PCA principal component analysis, DMR differentially methylated region, ML machine learning, scRNA-seq single-cell RNA sequencing.

DMR analysis reveals predominant hypermethylation in MenSCs in endometriosis and distinct separation of clinical groups

To account for potential confounding effects in downstream analysis, we first assessed baseline clinical variables. Age was the only variable to differ significantly between control and endometriosis groups (p = 0.02; see Table 1). Therefore, age was included as a covariate in the generalised least squares model used for identifying DMRs. Differential methylation analysis at the region level identified a total of 466 DMRs, with 458 regions being hypermethylated and 8 regions hypomethylated (imputed p < 0.05), indicating a strong MenSC hypermethylation profile in endometriosis patients. Notably, the hypermethylation patterns observed in endometriosis exhibit parallels with those documented in other conditions such as ovarian cancer^39,40, uterine fibroids⁴¹, and other cancers⁴².

In terms of genomic context, the majority of identified DMRs (362 DMRs) were located in open sea CpG regions, with the remainder distributed across CpG shores (42 DMRs), CpG shelves (38 DMRs), and CpG islands (24 DMRs) (Fig. 2a). Genomic annotation indicated that most DMRs were located within gene bodies (238 intronic and 56 exonic regions), while a substantial proportion (70 DMRs) was associated with transcriptional regulatory elements, including promoter regions.

Fig. 2. Exploratory analysis of EM-seq data and DMR detection.

a Genomic context of the detected DMRs, showing CpG annotations and genomic region types. b PCA plot of the 19 DMRs selected by the feature-selection procedure common to both Random Forest (RF) and Support Vector Machine (SVM) algorithms (n = 23 control, n = 19 endometriosis). c Correlation plot depicting relationships between the selected DMRs and various clinical variables (Pearson correlation p-value thresholds: 0.05 < p < 0.01 = *, 0.01 < p < 0.001 = **, p < 0.001 = ***). PCA principal component analysis, DMR differentially methylated region.

To identify the most critical DMRs that effectively distinguish between endometriosis and control samples, a dual feature selection approach was employed, combining SVM and RF algorithms. This approach yielded a consensus list of 19 key DMRs identified as critical for distinguishing endometriosis from controls (Supplementary Data 2). To assess their discriminative capacity, an unsupervised PCA was performed, revealing distinct methylation pattern differences between the two groups (Fig. 2b). PCA results showed that PC1 accounted for 48.63% of the total variance and distinguished endometriosis and control samples, supporting further exploration to refine candidate biomarkers.

To further evaluate the relevance of the selected DMRs, we examined the correlation between PCs and the clinical and technical variables included in the study. The top 10 PCs were analysed, as they collectively accounted for 93.5% of the total variance in the dataset. Figure 2c demonstrates that PC1 showed a strong correlation with the diagnosis status (correlation coefficient = 0.68, p < 0.001) and a weak correlation with body mass index (BMI) (correlation coefficient = -0.35, p < 0.05). No significant correlations were observed with the remaining clinical variables analysed, including processing time and collection centre, highlighting the robustness of the protocol and resilience to pre-analytical variability. Taken together, these results indicate that PC1 captures critical endometriosis-related biological variability, highlighting the promise of the selected DMRs for further evaluation.

DMR signature highlights genes with disease-relevant functions and pathways

To assess the discriminative power of the identified 19-DMR signature, we performed hierarchical clustering using Euclidean distance. The DMR-based clustering accurately classified the majority of endometriosis and control samples (Fig. 3a), supporting the robustness of the signature in distinguishing endometriosis samples from non-endometriosis controls.

Fig. 3. Comprehensive analysis of DMRs and their functional annotation.

a Heatmap displaying the 19-DMR signature identified through the dual feature selection algorithm. Red tiles represent hypermethylated DMRs, while blue tiles indicate hypomethylated DMRs. b Summary of the ORA analysis for GO biological processes. Bar labels represent biological processes associated with the respective GO terms relevant to endometriosis (FDR < 0.05). CHUIMI Complejo Hospitalario Universitario Insular Materno Infantil, HCB Hospital Clínic de Barcelona, ORA over-representation analysis, GO gene ontology.

Among the genes captured in the DMR signature were ZNF516, GPM6B, PSMD1 and CLMP. Notably, the dual feature selection algorithm identified multiple DMRs within intronic and exonic regions of ZNF516, a DNA-binding protein which regulates transcriptional activity of downstream target genes. Furthermore, a hypermethylated intronic region within GPM6B, a membrane glycoprotein implicated in cell-cell communication, was detected. A hypermethylated region was also identified within the 3′ UTR of PSMD1, a regulatory subunit of the 26S proteasome that plays a key role in the maintenance of protein homeostasis. Lastly, a DMR was mapped to the first intron of CLMP, a transmembrane protein localised at junctions between endothelial and epithelial cells with a role in cell-cell adhesion.

Additional genes within the DMR signature included MAP3K2, SPRY1 and PRDM1. Specifically, a hypermethylated intronic region was detected within MAP3K2, a serine/threonine protein kinase regulating multiple signalling pathways. Another hypermethylated region was mapped within the 5′ UTR of SPRY1, a negative regulator of growth factor signalling pathways. Finally, a DMR in the exonic region of PRDM1, a transcription factor important for immune cell development and differentiation, was hypermethylated.

To characterise the identified DMR signature and its potential impact at the functional level, Gene Ontology (GO) over-representation analysis was performed using the 458 hypermethylated DMRs described above (Supplementary Data 3). As shown in Fig. 3b, several endometriosis-relevant biological processes were found to be significantly enriched. Notably, multiple terms related to the WNT signalling pathway emerged among the most overrepresented categories (Supplementary Fig. 3) alongside an enrichment of terms related to extracellular matrix (ECM) organisation. Moreover, the angiogenesis-related processes were prominent, as reflected by terms such as ‘Regulation of cellular response to vascular endothelial growth factor stimulus’. Finally, several processes associated with epithelial and tissue morphogenesis were listed among the enriched terms. Collectively, these results reveal the biological processes and functional relevance of the genes within the MenSC-derived DMR signature, linking significant DNA methylation changes to key pathways involved in endometriosis pathogenesis.

DMR signature-based machine learning model accurately diagnoses endometriosis and outperforms prior approaches

To evaluate the predictive capacity of methylation profiling of MenSCs, we trained and tested multiple machine learning algorithms to distinguish between endometriosis and control samples. Given the limited sample size, a LOO validation pipeline was implemented to minimise overfitting and provide a realistic estimate of diagnostic performance (Supplementary Fig. 4, and Supplementary Data 4).

We trained 14 different models with varying numbers of DMRs (ranging from 2 to 20), ranked by their imputed p-values in each training iteration. Most models achieved a mean accuracy exceeding 0.90 in the training phase, except for LightGBM and LightGBMXT models, which maintained accuracies of ~0.50 (Fig. 4a). During the testing phase, the top three performers Weighted Ensemble L2, ExtraTreesGini, and NeuralNetTorch, achieved accuracies of 0.81, 0.76, and 0.76, respectively, using four DMRs (Fig. 4b, and Supplementary Fig. 5). Focusing on the best-performing Weighted Ensemble L2 model, we calculated an accuracy of 0.81, a specificity of 0.83, and a sensitivity of 0.79 (Fig. 4c). Consistent with our previous exploratory analysis, the most influential DMRs driving classification (Fig. 4d) were those located in the exonic and intronic regions of ZNF516, as well as in distal intergenic and promoter regions of LOC101929614 and LOC112267908, respectively. These findings reinforce the diagnostic potential of leveraging methylome dysregulation in MenSCs to accurately detect endometriosis in a non-invasive manner.

Fig. 4. Machine learning model performance using a leave-one-out training and testing pipeline.

a Line plot showing the mean accuracy during the training stage, evaluated using between 2 to 20 DMRs. b Line plot illustrating the accuracy, sensitivity, and specificity of the top 3 best-performing models in the test stage. The dotted line indicates the optimal number of DMRs to use, which is 4. c Confusion matrix displaying the performance of the best model, WeightedEnsemble L2. d Methylation levels of the top 4 DMRs utilised by WeightedEnsemble L2. Box plots include values from n = 23 control and n = 19 endometriosis samples. Whiskers denote the minimum and maximum values and the black horizontal line within each boxplot represents the median. Outliers are shown as individual values beyond the whiskers.

MenSC-derived methylation signature is associated with suppressed target gene expression in an independent eutopic endometrium scRNA-seq dataset

To independently validate the effect of the identified DMR signature on the endometrium, we used the HECA dataset^36,37, the most comprehensive scRNA-seq dataset of eutopic endometrium from individuals with and without endometriosis. Figure 5a depicts the UMAP plot of 159,052 mesenchymal lineage cells, comprising perivascular, smooth muscle, and stromal cell populations. These cells were sampled across various stages of the menstrual cycle, from the early proliferative phase to the late decidual stage, thereby capturing the complete transcriptomic landscape of stromal cells throughout the cycle in both experimental groups (Supplementary Fig. 6). To detect the possible presence of stem cell-like cells, we calculated a module score using MenSC markers to identify potential stromal stem cells within the eutopic endometrium. By applying a module score cutoff of ≥ 0.8, we detected 1198 cells (Fig. 5b), which were locally enriched in early-proliferative and proliferative stromal (eStromal) cell clusters. Most of these cells mapped to the eStromal (stromal cells in early-proliferative and proliferative phases) and eStromal_MMP (stromal cells in the menstrual phase) clusters (Fig. 5c). To investigate the transcriptomic differences in stromal stem cells between endometriosis and control samples, we selected cells annotated as eStromal or eStromal_MMP. This yielded a total of 826 cells (403 endometriosis and 423 control cells).

Fig. 5. Integrative analysis of the DMR signature detected in the HECA scRNA-seq external dataset.

a UMAP plot displaying the module score for MenSC markers. b Bar plot showing the distribution of cell types between cells with a MenSC module score > 0.8 and all remaining cells. This reveals an enrichment of stromal compartment cells in proliferative phase. dHormones, hormone-induced decidualised stromal cells; dStromal_early, decidualised stromal cells (early stage); dStromal_late, decidualised stromal cells (late stage); dStromal_mid, decidualised stromal cells (mid stage); ePV_1a, endometrial perivascular cells; ePV_1b, endometrial perivascular cells; ePV_2, endometrial perivascular cells; eStromal, endometrial stromal cells (proliferative phase); eStromal_cycling, endometrial stromal cells in cycling stage; eStromal_MMPs, endometrial stromal cells (menstrual/early proliferative phase); Fibroblast_basalis, fibroblast cells from the basalis part; HOXA13, HOXA13 positive cells; mPV, myometrial perivascular cells; sHormones, hormone-exposed stromal cells; uSMCs, uterine smooth muscle cells. c Dot plot displaying the markers of the cellular subpopulation. d Scatter plot of the beta coefficient versus the average log2 fold change. Each point represents a DMR associated with a gene present in the scRNA-seq dataset and identified as differentially expressed (FDR < 0.05 using two-tailed Wilcoxon test) in stromal cells labelled as eStromal and eStromal_MMPs. The highlighted points are genes detected using the dual machine learning feature selection strategy from the previous analysis (n = 423 control, n = 403 endometriosis cells). e Ridge plot illustrating the expression of these selected genes in the aforementioned stromal cells (normalised log-transformed expression values). FDR false discovery rate, UMAP Uniform Manifold Approximation and Projection.

We next performed DGE analysis comparing endometriosis to control cells and correlated these results with our previously characterised DMR signature (Supplementary Data 5). A substantial proportion (~90%) of genes that overlap both hypermethylated DMRs and differentially expressed genes showed reduced expression in endometrial stromal stem cell populations from endometriosis-affected individuals (Fig. 5d, and Supplementary Data 6). These findings support a strong correlation between the MenSC methylation signature and the transcriptional profiles observed in the MenSC marker-positive cells in the HECA dataset.

Interestingly, several genes identified through our dual feature-selection algorithm (ZNF516, TPGS1, GPM6B, CLMP, and PPFIA1) were downregulated at the transcriptional level (Fig. 5e). The results shown here suggest that the identified methylation signature has a functional impact at the transcriptome level of MenSC marker-positive cells within the eutopic endometrium of patients.

Discussion

Menstrual blood provides a non-invasive window into the uterine environment, with low intra-cycle biological variability, making it suitable for molecular profiling of endometriosis and other gynaecological diseases. In this study, we validated the robustness of our protocol for DNA methylation analysis in MenSCs and identified disease-specific methylation signatures with strong diagnostic performance, achieving 81% accuracy in testing. We further demonstrated the biological relevance of the identified DMRs through external transcriptomic correlation and pathway analysis, which implicated processes involved in disease development and persistence. Altogether, these findings support the potential of menstrual blood-derived epigenetic biomarkers to advance our understanding of endometriosis and improve its detection.

Previous studies on methylation changes in endometriosis have largely used array-based methods which are prone to batch effects⁴³ and cover only about 1.5–3% of all CpG sites⁴⁴. These studies were performed on bulk tissue, potentially obscuring cell-type-specific methylation patterns in heterogeneous tissues such as the endometrium. Notably, changes in rare cell populations such as eMSCs, which comprise only 0.02% to 1.23% of endometrial cells⁴⁵, may be masked in such analyses. This challenge is further compounded by the fact that the endometrium undergoes substantial methylation changes throughout the menstrual cycle, an effect not observed in peripheral blood⁴⁶. A recent study by Mortlock et al., analysing samples from 984 participants, provided foundational insights into the impact of menstrual cycle phase on endometrial methylation. The study showed that methylation changes were predominantly driven by cycle phase, with disease-specific signals emerging only in moderate-to-severe cases (i.e. stage III and IV combined)⁴⁷. To the best of our knowledge, this is the first study designed to overcome previous limitations by isolating MenSCs directly from menstrual blood and performing cell type-specific whole genome methylome profiling on primary, freshly isolated cells. The collection of menstrual blood standardised the sample timing across participants, minimising phase-dependent variability. Profiling freshly isolated MenSCs allowed an unbiased, comprehensive analysis of the epigenetic landscape in this disease-relevant key cell population, facilitating direct comparisons between patients and controls. Finally, the use of a recently developed whole genome methylation sequencing protocol that avoids harsh bisulfite treatment, better preserves DNA integrity and improves coverage uniformity, thereby enabling accurate, high-resolution CpG methylation profiling of MenSCs.

MenSCs have been previously implicated in the aetiopathogenesis of endometriosis, as discussed earlier⁴⁸. Key genes captured in the MenSC-derived DMR signature included ZNF516, GPM6B, PSMD1 and CLMP. Previous studies have reported hypermethylation and transcriptional downregulation of ZNF516 in ovarian endometriosis⁴⁹. Further supporting its epigenetic regulation in endometriotic tissue, a DMR within GPM6B has also been identified in healthy endometrial stromal cells following progesterone and oestradiol treatment, suggesting a possible role in hormone-responsive endometrial function⁵⁰. In addition, the downregulation of GPM6B in eutopic endometrial samples⁵¹, together with evidence that loss of expression is linked to the development of a mesenchymal subtype in glioblastoma⁵², suggests that GPM6B may actively contribute to endometriosis progression. Furthermore, dysregulation of PSMD1, a key proteasome component, may impair proteasomal function and thereby contribute to the abnormal tissue growth and inflammation observed in endometriosis⁵³. Consistent with this, the dysregulation of Psmd1 reported in mouse xenograft models of endometriotic lesions supports a direct link between its epigenetic regulation and disease pathogenesis. Finally, DMRs identified in the first intron of CLMP could disrupt cell-cell adhesion and endometrial stromal cell integrity⁵⁴. As a tumour suppressor in colorectal cancer, CLMP loss has been associated with WNT/β-catenin pathway activation and enhanced cellular proliferation⁵⁵, further emphasizing the potential role of CLMP dysregulation in endometriosis.

Other genes within the DMR signature with potential functional relevance to endometriosis pathogenesis include MAP3K2, SPRY1 and PRDM1. In vitro targeting of MAP3K2 in endometrial stromal cells from endometriosis patients demonstrated that reduced MAP3K2 expression altered MAPK signalling, a pathway central to inflammation, hormone responsiveness and lesion survival⁵⁶. SPRY1, a negative feedback inhibitor of FGF and EGFR signalling that regulates cell proliferation and differentiation, has been shown to exhibit context-dependent regulation in endometriotic epithelial cells, with its expression notably increased under progesterone resistance^57,58. Lastly, transcriptional profiling of patient-derived samples has shown that PRDM1, a transcriptional repressor of immune cell differentiation and function, is dysregulated in endometriosis^59,60. Overall, among the DMR-associated genes identified in the MenSC-derived signature PSMD1, CLMP, MAP3K2, SPRY1 and PRDM1 represent previously unreported methylation loci in endometriosis, underscoring the significance of these findings.

Building on these observations, our findings demonstrate that the genome-wide methylation profile of these cells reflects disease-specific alterations, providing molecular evidence for their role in ectopic lesion development. In support, GO analysis revealed methylation changes in genes associated with mechanisms of endometriosis pathogenesis and lesion persistence. Enriched terms included regulation of the WNT signalling pathway, as well as processes related to ECM organisation and angiogenesis^61,62, indicating epigenetic alteration of pathways critical to lesion establishment. Dysregulation of WNT signalling has been linked to aberrant cellular proliferation, invasion, and tissue remodelling in endometriosis⁶³, while genes within the DMR signature may influence ECM dynamics and adhesion properties, potentially promoting the invasive behaviour of MenSCs^64,65. Enriched processes associated with epithelial and tissue morphogenesis further underscore the relevance of these epigenetic changes⁶⁶. Collectively, these results suggest that the epigenetic alterations identified in MenSCs are consistent with dysregulation of cellular behaviours relevant to lesion biology, including tissue remodelling, cellular invasion and adhesion, and neovascularisation. Thus, our findings highlight the biological relevance of the MenSC-derived DMR signature by linking DNA methylation changes to key pathways involved in endometriosis pathogenesis and ectopic lesion progression.

Our study demonstrates that a machine learning-based approach leveraging DMRs from isolated MenSCs achieves a diagnostic accuracy of 0.81, with a specificity of 0.83 and a sensitivity of 0.79. This performance is particularly notable given the current state of biomarker-based diagnostic tools for endometriosis, which, despite extensive research, have yet to produce a clinically validated, non-invasive diagnostic method^67,68. To date, over 1100 potential biomarkers, primarily proteins and RNAs, have been identified across various biological compartments, including peripheral blood, peritoneal fluid, follicular fluid and urine. However, only four of these candidates (TNF-α, MMP-9, TIMP-1, and miR-451) have been consistently validated across multiple studies⁶⁸. Even among these, sensitivity and specificity vary substantially depending on patient phenotype, menstrual cycle phase, and comorbidities, underscoring the complexity and heterogeneity of endometriosis^68,69. In contrast, our model exhibited robust predictive performance using four DMRs, highlighting the potential of methylation-based classification as a more reliable approach to non-invasive endometriosis diagnosis. Moreover, our findings support the use of menstrual blood as a stable diagnostic sample, circumventing the biological variability associated with menstrual phase-sensitive biomarkers and leveraging epigenomic signatures, which are inherently more stable than miRNAs or proteins⁷⁰. Nevertheless, to fully capture the clinical and molecular heterogeneity of endometriosis and enhance the clinical applicability of this approach, validation in larger, more diverse patient cohorts is essential.

The identified MenSC-derived DMR signature not only distinguishes endometriosis samples at the epigenetic level but also demonstrates potential functional relevance through its influence on gene expression, as supported by analysis of an independent transcriptomics dataset. Specifically, analysis of HECA scRNA-seq dataset revealed that genes marked by hypermethylation in our DMR signature were significantly downregulated in eutopic stromal stem cell populations from endometriosis patients. This result indicates that the detected epigenomic signature may influence transcriptional changes.

This study has a few limitations that should be considered. First, the low yield of sorted MenSCs resulted in limited DNA quantities obtained, which constrained sequencing depth and reduced the extent of methylation changes that could be detected across the genome. This limitation reflects the biological challenge of working directly with a rare cell population without prior in vitro expansion. Second, the relatively small sample size reduces the statistical power of the study and limits the ability to perform stratified analysis. Nevertheless, the use of whole-genome methylation sequencing provides high-resolution, unbiased coverage across the genome, enhancing our capacity to detect meaningful disease-associated methylation changes despite sample size constraints. To further mitigate the impact of limited sample size, we adjusted the statistical cutoff parameters accordingly and emphasise the importance of independent validation in larger, multi-ancestry cohorts. Third, although our study controls were selected based on the absence of clinical suspicion or history of endometriosis, and confirmed by negative imaging or surgical findings, the inherent limitations of current diagnostic methods mean that undiagnosed disease cannot be entirely ruled out. This common challenge in endometriosis research may lead to conservative estimates of test specificity. Fourth, while we assessed the transcriptional impact of the DMR signature using scRNA-seq data from eutopic endometrium, these data were not derived from the same MenSC population. In the absence of transcriptomic data from freshly isolated MenSCs, we focused on stromal cells expressing canonical MenSC markers to approximate the in vivo population following menstrual shedding. The observed downregulation of hypermethylated genes in these cells supports the functional relevance of our methylomic signature across molecular layers and biological contexts, and underscores the need for direct transcriptomic profiling of freshly isolated MenSCs in future studies. Lastly, participants of different ethnic backgrounds were not equally represented in this study, primarily due to the narrow geographic scope of recruitment. The majority of participants were of White European ancestry, while the limited size of the Hispanic and African subgroups precluded reliable ancestry-stratified analyses or the inclusion of ethnicity as a covariate in the DMR model. Therefore, further validation in larger, ethnically diverse cohorts is required.

In conclusion, our findings highlight the potential of MenSCs to reveal endometriosis-associated molecular alterations. The identified methylation signatures were not only predictive but also mapped to pathways implicated in lesion formation, including WNT signalling, stem cell regulation, and ECM remodelling. Their transcriptional suppression in MenSC-like stromal populations further supports a functional role in disease pathogenesis. By providing a stable, cell-type-specific molecular readout, this approach offers a promising avenue for non-invasive diagnosis and, with further validation, may inform future efforts toward more individualised patient care. Larger and more diverse studies will be essential to assess clinical utility across varying disease presentations.

Author contributions

I.T., S.H. and S.R.V. performed the experiments. I.T., C.B. and R.P.M. analysed the data and interpreted the results. R.N.M. conducted the literature review and provided scientific support for the project’s development. M.T.P.Z. conceived the project and secured funding. C.F.M. led the overall project design and supervision. A.M.M. and F.C. acted as clinical principal investigators. A.S.R., M.A.S.S., M.Á.M.Z. and M.T.V. collected and processed clinical samples. I.T., C.B., R.P.M. and C.F.M. wrote the manuscript with input from all authors.

Peer review

Peer review information

Communications Medicine thanks the anonymous reviewers for their contribution to the peer review of this work.

Data availability

Source data for the main figures for this manuscript are provided in the Supplementary Data files as follows: Source Data for Fig. 1b-d, f are included in Supplementary Data 1; Source Data for Fig. 2b are provided in Supplementary Data 2, Source Data for Fig. 3b are provided in Supplementary Data 3, Source Data for Fig. 4a, b, d are provided in Supplementary Data 4, and Source Data for Fig. 5d are provided in Supplementary Data 6.

Methylation data generated from this study is available through the European Genome-Phenome Archive (EGA; EGA study EGAS50000001640 and EGA dataset EGAD50000002353). Access will be restricted due to privacy protections under the EU General Data Protection Regulation (GDPR), and because participants did not consent to public data sharing. Qualified researchers may request access through the relevant Data Access Committee (DAC; EGA DAC EGAC50000000786). Approval will require a data use agreement and evidence of compliance with GDPR requirements.

Code availability

All source code used to generate the figures in this study is available via Zenodo at 10.5281/zenodo.15878286⁷¹.

Competing interests

C.B., R.P.M., I.T., S.H., S.R.V., R.N.M. and C.F.M. are employees of endogene.bio. M.T.P.Z. is the Chief Executive Officer of endogene.bio. All other authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-026-01641-3.