Uterine microbiome signatures associated with endometriosis

Abstract

Background

Endometriosis is a chronic inflammatory disorder affecting ~ 10% of reproductive-age women, often causing pelvic pain and infertility. Despite its prevalence, diagnosis remains delayed due to non-specific symptoms and lack of reliable non-invasive biomarkers. Emerging evidence implicates the microbiome in disease pathogenesis.

Results

We analyzed uterine microbiomes from 266 tissue samples collected during either the proliferative or secretory phase, using 16S rRNA gene sequencing. Genus-level analysis revealed variable Lactobacillus abundance among all individuals. Prevotella showed borderline enrichment in proliferative-phase patients. Sub-genus analyses identified a small number of differentially abundant taxa, though none remained significant after FDR correction. To capture subtle microbial shifts, we developed a feature set combining weakly differential taxa, algorithmically selected taxa via machine learning, and a functional dysbiosis score. A supervised classifier trained on proliferative-phase data achieved moderate predictive performance (AUC = 0.70), while secretory-phase models performed more poorly (AUC = 0.58).

Conclusions

The uterine microbiome shows phase-dependent differences in its potential to inform endometriosis status. Although no robust individual microbial biomarkers were identified, machine learning models incorporating subtle community features from the proliferative phase yielded modest diagnostic potential. These results highlight the importance of menstrual cycle-aware sampling and support further development of microbiome-informed diagnostic tools for endometriosis.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12915-026-02659-8.

Background

Endometriosis is a chronic inflammatory condition where endometrial-like tissue grows outside the uterus, often affecting the ovaries, fallopian tubes, and peritoneum [1, 2]. This ectopic tissue responds to hormonal changes, leading to symptoms like chronic pelvic pain, dysmenorrhea, dyspareunia, and infertility [1, 2]. Affecting an estimated 10% of reproductive-age women, the disease is frequently underdiagnosed, with an average delay of 6–11 years [3]. This is due to its non-specific symptoms, overlap with other conditions, and the lack of reliable diagnostics.

Multiple theories exist about the origin of endometriosis—including retrograde menstruation, coelomic metaplasia, and stem cell involvement [4]. Immune and inflammatory responses are also thought to play a key role, and recent research has begun exploring the gut and vaginal microbiomes as potential contributors [5]. Studies show that women with endometriosis often have higher levels of Escherichia coli in the gut, linked to elevated serum lipopolysaccharide, which activates inflammatory pathways and promotes pro-inflammatory cytokine production. These cytokines may aid the survival and implantation of ectopic endometrial cells [6]. Additionally, Streptococcus species have been linked to advanced disease stages, possibly by stimulating prostaglandin E2 expression, a contributor to pelvic pain [7]. However, findings across studies are inconsistent; notably, one large study involving around 1000 participants found no significant association between gut microbiome and endometriosis [8]. The vaginal microbiome, crucial for reproductive health, remains relatively understudied in the context of endometriosis [9]. Shifts in vaginal microbiome, particularly the reduction of Lactobacillus and overgrowth of bacteria such as Gardnerella, Prevotella, and Mobiluncus are seen in bacterial vaginosis, which is associated with inflammation and may contribute to endometriosis pathogenesis [10]. Moreover, menstrual cycle phases influence vaginal microbial composition, with reduced Lactobacillus and increased diversity during menstruation, complicating research and emphasizing the importance of cycle-aware sampling [11].

While the majority of microbiome studies in endometriosis have focused on the vaginal or gut compartments, presumably due to the relative ease of sample collection, only a handful of investigations have directly examined the uterine microbiome. These studies often involved relatively small cohorts with just a few dozens of individuals [12, 13] or were conducted within broader clinical settings such as repeated implantation failure or infertility [14, 15], where endometriosis was one of several coexisting clinical factors rather than the primary focus of analysis. Despite heterogeneity in study designs, these reports collectively support the biological relevance of the uterine microbial environment to endometrial function.

To investigate this further, we analyzed uterine microbiomes in 266 samples from women in either the proliferative or secretory phases. A total of 138 uterine tissue samples were collected from women in the proliferative phase of their menstrual cycle. Among these, 78 samples were obtained from women with a laparoscopic diagnosis of endometriosis, while the remaining 60 were from women without the disease. An additional 128 uterine tissue samples were collected during the secretory phase, including 88 from women diagnosed with endometriosis and 40 from unaffected individuals (Table 1; Fig. 1). Total genomic DNA was extracted from all tissue samples and used to prepare targeted bacterial 16S rRNA gene libraries for sequencing, as detailed in the “Methods” section. Raw sequence data were processed to remove technical artifacts, host DNA, and environmental contaminants. High-confidence bacterial reads were then taxonomically annotated using an internally curated version of the Greengenes2 database. Downstream analyses focused on identifying differentially abundant microbial taxa associated with endometriosis and determining taxa with potential predictive value for disease diagnosis.

Table 1.

Summary of study samples. Median values are presented for each demographic parameter, with ranges shown in brackets. Statistical comparisons were conducted using T-tests

Characteristics	Endometriosis	Control	p-value
All samples (proliferative and secretory phases)
Sample size	166	100	-
BMI	21.48 (15.62–36.85)	22.50 (17.1–31.22)	0.83
Age	35.5 (20–51)	38.5 (21–50)	0.0002
Proliferative phase samples
Sample size	78	60	-
BMI	21.51 (16.21–36.85)	22.50 (17.1–30.42)	0.76
Age	36 (20–51)	40.5 (24–50)	0.026
Secretory phase samples
Sample size	88	40	-
BMI	21.45 (15.62–34.22)	22.49 (18.22–31.22)	0.94
Age	35 (21–51)	37.5 (21–49)	0.005

Fig. 1.

Overview of the study design. A total of 266 samples were analyzed, including 138 from participants in the proliferative phase (78 patients and 60 controls) and 128 from the secretory phase (88 patients and 40 controls). Targeted sequencing of the 16S rRNA V4 hypervariable region was performed. Bacterial reads were annotated using the Greengenes2 database. Differential and informative taxa were identified for machine learning-based prediction of endometriosis

Results

Uterine microbiome landscape in the study cohort

The initial objective of our analysis was to evaluate both alpha and beta diversity of the uterine microbiome in women diagnosed with endometriosis compared to those without the disease, stratified by the proliferative and secretory phases of the menstrual cycle. Alpha diversity, which reflects the richness and evenness of microbial species within individual samples, was assessed using the Shannon index. No statistically significant differences in alpha diversity were observed between endometriosis and control groups in either the proliferative or secretory phase (Fig. 2a). Similarly, beta diversity, which measures compositional differences in microbial communities between groups, was evaluated using the Bray–Curtis dissimilarity metric. This analysis also revealed no significant differences between women with and without endometriosis across both menstrual phases (Fig. 2b). These findings suggest that the overall diversity, including both the number of microbial taxa and their relative abundance distribution, is comparable between affected and unaffected individuals, irrespective of the menstrual cycle phase.

Fig. 2.

a Alpha diversity comparisons between patients and controls within each menstrual phase. b Beta diversity comparisons between patients and controls in both proliferative and secretory phases

Genus-level analysis revealed substantial variability in the relative abundance of Lactobacillus among individuals, both in patients and controls, across both menstrual phases. Although Lactobacillus is typically considered a hallmark of a healthy vaginal microbiome, its levels varied considerably, especially among patients with endometriosis in the proliferative phase compared to the controls (Fig. 3a, b), recapitulating a previous study which showed considerable variations in Lactobacillus abundance even among healthy females [11]. An analysis of the top 20 most abundant genera in both proliferative and secretory phase samples showed that the overall distribution of relative abundance was similar between patients and controls, with most of these dominant taxa not differentially abundant. One notable observation was made: the genus Prevotella showed a trend toward enrichment in patient samples from the proliferative phase, with a borderline significant p-value (0.0509). Bacteria species in the genus Prevotella are commonly associated with vaginal dysbiosis and pro-inflammatory states [10].

Fig. 3.

a Genus-level relative abundance and microbial community profiles in proliferative phase samples. b Genus-level relative abundance and microbial community profiles in secretory phase samples

Differentially abundant and machine learning informative taxa

A more in-depth analysis of sub-genus level taxonomic units revealed eight taxa that were differentially abundant (p-value ≤ 0.05) between patients and controls in the proliferative phase, and three differential taxa in the secretory phase (Fig. 4a), after adjusting for potential confounding effects of body mass index (BMI) and age using MaAsLin2. A supplementary analysis using ALDEx2 showed similar results (Additional file 1: Table S1). There is no overlap in differential taxa between proliferative and secretory phases. Although the number of differentially abundant taxa was modest, this finding aligns with expectations based on a prior large-scale study where it reported no statistically significant differences in gut microbial composition between women with and without endometriosis [8], despite earlier, smaller studies suggesting such associations. This discrepancy highlights the challenges in identifying consistent microbial biomarkers of gynecological disease, particularly in extraintestinal sites. Motivated by the hypothesis that the uterine microbiome may more directly reflect gynecological pathophysiology than the gut microbiome, we undertook this study to examine microbial community profiles in uterine tissue. In fact, after correcting for multiple comparisons using false discovery rate (FDR) adjustment, the initially observed differential taxa no longer reached statistical significance (i.e., FDR > 0.05). Nonetheless, we recognize that subtle yet consistent shifts in microbial composition across multiple taxa may carry predictive value [16]. Therefore, we aim to employ machine learning approaches to integrate these signals, under the premise that the cumulative effect of multiple weakly informative taxa would enable and/or enhance predictive performance in distinguishing disease states [17]. For instance, taxa such as Prevotella sp.1 and Ureaplasma sp.1—both belonging to genera frequently linked to bacterial vaginosis [10], while not individually conclusive after FDR correction, may collectively contribute to distinguishing disease states when incorporated as features in a supervised machine learning model.

Fig. 4.

a Regression coefficients of differential taxa identified by MaAsLin2 (p ≤ 0.05) that distinguish endometriosis from controls in proliferative and secretory phase samples. b Boxplots illustrating the relative abundance distribution of differential taxa in proliferative and secretory phases. c Bacterial taxa identified as informative features through machine learning-based selection using the Random Forest algorithm

To develop the feature set for supervised machine learning classification, we implemented a three-step selection strategy combining statistical and algorithmic criteria. First, we identified weakly differential taxa (i.e., nominal p-values ≤ 0.05; Fig. 4a, b), indicating potential biological relevance despite not meeting strict multiple-testing thresholds. These taxa were initially included to ensure that subtle, non-random differences were not overlooked. In the second step, we applied a machine learning-based feature selection process by systematically evaluating the importance of each taxon detected in our profiling pipeline. This involved training preliminary models to score each taxon’s contribution to classification performance, using a predefined threshold of feature importance score ≥ 0.015 as a cutoff for inclusion. Taxa meeting this threshold were selected as additional candidates for the final feature set (Fig. 4c; Additional file 1: Table S2–S3). A subset of taxa was identified exclusively through the feature importance criterion. Specifically, 14 additional taxa including two other Prevotella spp. were added in the proliferative phase. In the secretory phase, 11 additional taxa were incorporated. Notably, Gardnerella sp.1, a taxon traditionally linked to bacterial vaginosis was included in both proliferative and secretory cohorts. In the third step, a functional dysbiosis score (FDS) was calculated for each sample (Additional file 1: Table S4), representing an aggregate measure of microbial dysbiosis within the uterine tissue. The final feature set used for model training therefore comprised three components: the weakly differential taxa, taxa selected by feature importance scoring, and the FDS (Additional file 1: Table S5–S6).

Diagnostic performance of machine learning models in endometriosis detection

Using the microbial profiles from the proliferative phase, we achieved a modest predictive performance in distinguishing endometriosis patients from controls. Specifically, across 50 rounds of repeated random subsampling cross-validation, the average area under the curve (AUC) reached 0.70, indicating reasonable discriminative capability. The model demonstrated a sensitivity of 0.71, and a specificity of 0.54 (Fig. 5a). While not optimal, the overall performance suggests that the microbiome during the proliferative phase carries meaningful signals that could aid in endometriosis diagnosis, especially when combined with other tools such as laparoscopy and blood-based biomarker tests. In contrast, models trained on the microbial profiles from secretory phase showed weaker overall performance, with an average AUC of only 0.58 (Fig. 5b). Overall, these findings suggest that microbial signatures differ in both menstrual phases, and that the proliferative phase carries more informative profiles. A supplementary analysis stratifying samples by disease stage indicated that early-stage cases exhibit modestly distinct microbiome profiles compared with late-stage cases (Additional file 2: Fig. S1). Given the limited number of early-stage samples, a three-class classification model was not pursued, and future studies with larger cohorts are warranted to address this question more robustly. Additionally, within-group comparisons demonstrated that there are phase-associated taxa in both the control and patient groups (Additional file 2: Fig. S2). Importantly, several of the taxa that were differentially abundant between proliferative and secretory phases were also identified as differentially abundant between disease and control within a given menstrual phase. From a clinical and biological perspective, these results underscore the importance of menstrual cycle timing when considering the microbiome as a diagnostic aid for endometriosis. The result also suggests that cycle-phase-specific sampling may be crucial for optimizing microbiome-based diagnostics, and that future models may benefit from integrating hormonal phase information, or adjusting for it explicitly.

Fig. 5.

a Predictive performance of the differential microbial profile from the proliferative phase in classifying endometriosis. b Predictive performance of the differential microbial profile from the secretory phase in classifying endometriosis

Discussion

This study investigated the uterine microbiome in women with and without endometriosis, with a focus on menstrual cycle phase-specific microbial signatures and their predictive value for disease status. The differentially abundant taxa, coupled with the use of supervised machine learning allowed us to uncover subtle patterns of microbial variation that may hold diagnostic potential.

Our findings add to a growing body of literature suggesting that microbial dysbiosis may contribute to endometriosis pathogenesis, though not through overt community-level disruption. Previous studies have implicated the gut and vaginal microbiomes, such as the elevated abundance of Gardnerella and Streptococcus being associated with advanced disease. However, large-scale studies such as the one by Pérez-Prieto et al. [8], have failed to find significant associations between gut microbiome and endometriosis despite other studies suggesting so, indicating that location-specific microbial assessments may be more informative. It is important to note that characterization of the uterine microbiome in endometriosis remains relatively limited, as outlined in the Introduction. Existing studies have often involved relatively small cohorts or have been conducted within broader clinical contexts such as infertility, rather than focusing specifically on endometriosis [12–15]. In this regard, our results demonstrate that the uterine microbiome may contain weak but biologically meaningful signals in distinguishing endometriosis, especially during the proliferative phase, where the tissue is more hormonally responsive and immunologically active. The modest predictive power (AUC = 0.7) achieved by our model trained on proliferative phase cohort is consistent with recent studies showing that cumulative patterns of multiple taxa can outperform single-biomarker approaches. Furthermore, the menstrual cycle phase significantly influences reproductive tract microbiota, with menstruation being associated with reduced Lactobacillus and increased microbial diversity. These fluctuations likely contribute to inconsistencies across studies in the literature and emphasize the importance of phase-aware sampling. Our observation that microbial profiles from the secretory phase were less predictive supports this notion and aligns with reports that progesterone-dominant conditions may dampen inflammatory signals or microbiome-host interactions. As such, we emphasize that the microbiome-based classifier presented in this study is not intended to supplant established diagnostic modalities currently used in clinical practice. Rather, it should be viewed as a complementary tool within a broader, multi-model diagnostic framework. In this context, microbiome-derived signatures could add orthogonal biological information that augments existing approaches, rather than functioning as a standalone diagnostic test. Notably, blood- and saliva-based molecular biomarkers have demonstrated slightly superior discriminative performance when used in isolation [18–20], underscoring the importance of integrating multiple sources of evidence to achieve robust and clinically meaningful diagnostic accuracy. Furthermore, recent advances in artificial intelligence (AI) driven medical imaging analysis have opened new avenues for the systematic identification of subtle and complex imaging features that may not be readily discernible through conventional radiological assessment [21, 22]. The incorporation of such AI-derived imaging features, alongside molecular and microbiome data integrated through a machine learning framework, holds considerable promise for further enhancing diagnostic performance in highly heterogenous diseases such as endometriosis.

Notwithstanding, there are several limitations in our study. First, uterine microbiome studies are inherently challenged by low microbial biomass and the potential for cross-contamination introduced during transcervical sampling. In the present study, uterine tissue samples were obtained using Pipelle endometrial sampler, with specific procedural steps implemented to minimize contact with vaginal and cervical microbiota, as described in the “Methods” section. We acknowledge that, as with all transcervical and transvaginal sampling approaches, complete elimination of lower tract microbial carryover is not technically feasible. We also acknowledge that this study did not include technical replicates and dual-sampling validation approaches, which represents a limitation. However, such constraints are common in clinically obtained uterine microbiome datasets due to practical considerations (i.e., low biomass), and the methodological framework employed here is consistent with prevailing standards in the field. Notably, several high-profile microbiome studies [23, 24], similarly relied on rigorous aseptic sampling combined with negative controls and bioinformatic decontamination consistent with our workflow. Taken together, while we recognize the inherent limitations associated with transcervical uterine sampling, we believe that the combination of controlled sampling procedures, inclusion of negative controls, and conservative computational decontamination supports the robustness of the observed microbial patterns and the biological conclusions drawn from these samples.

Another limitation of this study is the presence of significant age differences between endometriosis patients and control subjects. Age is known to influence both host physiology and the composition of microbiome and therefore could partially contribute to the observed microbiome differences between groups. We have adjusted age explicitly for the statistical analyses to mitigate its potential confounding effects; however, residual confounding might not be entirely excluded. Future studies with a larger sample size that enables age-matched stratification will be important to further disentangle age-related effects from disease-specific signals.

Although this study was designed specifically to investigate the uterine microbiome in endometriosis, it is important to acknowledge the clinical complexity of this patient population. Gynecological conditions such as adenomyosis and uterine fibroids frequently co-exist with endometriosis and share overlapping symptoms and pathological features, making it challenging to fully disentangle their individual contributions. Therefore, the presence of these co-morbidities could not be entirely excluded, and they may have contributed, to some extent, to the observed uterine microbiome profiles. This potential confounding factor should be considered when interpreting the results. Future studies incorporating larger cohorts, condition-matched controls, and stratified analyses will be necessary to disentangle the individual and combined effects of these gynecological conditions on the uterine microbiome.

Conclusions

In conclusion, while the uterine microbiome alone is unlikely to serve as a stand-alone biomarker for endometriosis, our findings suggest that when analyzed in a cycle-phase-aware and integrative manner, it may contribute to a broader diagnostic framework. Future studies should explore longitudinal sampling, integrate host transcriptomic and immunologic data, assess signals in vaginal mucus, and validate predictive models in larger, independent cohorts. Collectively, these efforts may pave the way toward a microbiome-informed, minimally invasive diagnostic toolkit for endometriosis.

Methods

Specimen collection

This study was approved by the institutional review board of the Women’s Hospital, Zhejiang University School of Medicine (IRB-20240110-R). Endometrial tissue samples were collected from 266 individuals, all of whom were clinically suspected of having a gynecologic condition and scheduled for laparoscopy with histopathological evaluation (Additional file 1: Table S7). To explore menstrual cycle-related differences, samples were obtained from women in either the proliferative or secretory phases of their cycle. The menstrual phase was initially assessed by physicians or surgeons based on self-reported cycle days and clinical evaluations. To confirm this classification, serum progesterone levels were measured using a protein assay from Kangrun Biotech Co. Ltd. (Guangdong, China), with levels above 1.08 ng/mL indicating the secretory phase, as per the manufacturer’s guidelines. Uterine tissue samples were obtained using Pipelle endometrial sampler inserted transcervically [23–25]. Following insertion of a sterile speculum, the cervix was visualized, cleansed and dried using standard sterile procedures. The Pipelle device was then advanced through the cervical canal into the uterine cavity, with deliberate avoidance of contact with the vaginal walls. Upon reaching the uterine cavity, the uterine tissue samples were aspirated into the Pipelle sampling loop. Throughout passage of the device through the vaginal and cervical canals, the sampling loop was kept retracted within the cannula to avoid contact with the vaginal and cervical microbiota in order to minimize cross-contamination between uterine and lower tract microbiomes. All procedures were performed using single-use sterile instruments under uniform aseptic conditions. Among the samples, 138 were from the proliferative phase (78 from individuals with endometriosis and 60 from controls), and 128 were from the secretory phase (88 with endometriosis and 40 controls). Endometriosis was diagnosed and confirmed via gold-standard laparoscopic surgery. The disease stage was determined according to the revised American Society for Reproductive Medicine (rASRM) classification system, with stages I and II categorized as early-stage disease, and stages III and IV categorized as late-stage disease. Among the endometriosis patients, 22/78 from proliferative phase and 24/88 from secretory phase were classified as early stage. All participants provided written informed consent. The collected tissue samples were immediately transported at 4 °C to the HerAnova Lifesciences laboratory and stored at − 20 °C upon arrival.

Uterine tissue processing and targeted 16S library preparations

A 3–5-mm fragment of endometrial tissue from each sample was placed into an individual centrifuge tube with 20 µL of Proteinase K and 180 µL of Buffer ATL. Prior to DNA extraction, defined spike-in taxa were introduced into all samples including negative controls. This spike-in strategy was deliberately employed to ensure the presence of a measurable and internally consistent microbial signal under low-biomass conditions, where background contamination and stochastic amplification effects can otherwise dominate sequencing outputs. Specifically, ZymoBIOMICS™ Spike-in Control I (Zymo Research) was used as an internal reference to facilitate downstream quality control and contamination assessment (Additional file 1: Table S8). The mixture was vortexed thoroughly and incubated at 58 °C with shaking at 1200 rpm for 3 h. After incubation, add 210 µL of Buffer ATL to each sample and homogenize them using the TissueLyser II (2 min at 30 Hz, 1-min pause, repeated for 15 cycles), and DNA was extracted using the QIAsymphony SP instrument (QIAGEN, 35459). DNA concentration and purity were assessed using the MultiSkan GO spectrophotometer (Thermo, 1510). The V4 variable region of the 16S rRNA was amplified using Invitrogen Platinum SuperFi II DNA Polymerase with the following PCR conditions: initial denaturation at 98 °C for 30 s (1 cycle), followed by 30 cycles of 98 °C for 10 s, 60 °C for 10 s, and 72 °C for 30 s, and a final extension at 72 °C for 5 min before holding at 4 °C. The forward primer is 5′- TAATTGTGTGCCAGCMGCCGCGGTAA-3′ while the reverse primer is 5′- TCAGCCGGACTACHVGGGTWTCTAAT-3′. The PCR products were purified using VAHTS DNA Clean Beads. Adapter ligation was carried out using the UltraClean Universal DNA Library Prep Kit for Illumina V3 (Vazyme, UND607-02). First, 45 µL of End Repair reaction mix was added to the purified PCR product, followed by incubation at 20 °C for 15 min, 65 °C for 15 min, and held at 4 °C. The ligation reaction mix was prepared on ice, added to the end-repaired DNA, and incubated at 20 °C for 15 min, then held at 4 °C. The ligated products were purified again with VAHTS DNA Clean Beads. Library amplification was performed under the following thermal conditions: 95 °C for 3 min (1 cycle), then 5 cycles of 98 °C for 20 s, 60 °C for 15 s, and 72 °C for 30 s, with a final extension at 72 °C for 5 min and a hold at 4 °C. The final libraries were purified using VAHTS DNA Clean Beads. Library concentrations were quantified using the KAPA Library Quantification Kit (KAPA, KK4824), and fragment sizes were evaluated with the Agilent 4200 TapeStation (Agilent, G2991A). Sequencing was performed on the Illumina MiSeq platform using the MiSeq Reagent Kit v2 (300 cycles).

Bioinformatic processing of targeted 16S sequencing data

The demultiplexed FASTQ files from Illumina MiSeq sequencing were processed to extract the forward reads. To improve data quality, a two-step trimming and filtering process was employed. First, fastp [26] was used to identify and remove polyX artifacts—artificial stretches of a single nucleotide—commonly introduced during sequencing. Next, cutadapt [27] was used to trim any residual adapter sequences from the reads. To eliminate host-derived contamination, the filtered reads were aligned to the human reference genome (hg38) using Bowtie2 [28]. Alignment results were processed with SAMtools [29], and reads mapping to the human genome were removed, ensuring that only non-host (primarily bacterial) sequences were retained for microbiome analysis.

The resulting high-quality, non-human reads were then imported into the QIIME2 platform [30] for microbial community analysis. Within QIIME2, chimeric sequences were identified and removed using the vsearch uchime-denovo method. Subsequently, redundant sequences were collapsed using vsearch dereplicate-sequences, enhancing computational efficiency and reducing noise. The remaining high-confidence bacterial reads were annotated using an internally curated version of the Greengenes2 reference database [31]. Taxonomic assignments were made using the Greengenes2 taxonomy-from-table classifier, providing genus-level and species-level annotations where possible. To ensure the validity and accuracy of the microbiome profiles, decontamination was performed using SCRuB [32], a statistical tool designed to identify and remove background contaminants. A blank negative control, which underwent the entire experimental workflow alongside the tissue samples, was included in the analysis to model and subtract any environmental or reagent-based contaminants. This approach ensured that the final dataset reflected true biological signals and minimized the risk of false microbial detection. Microbiome Shannon index and beta diversity were calculated and visualized using vegan [33] in R (v.4.4.2). A functional dysbiosis score was computed for each sample using the following formula (0.5*(1 − relative abundance of Lactobacillus) + 10*(sum of relative abundance of pathogenic genera)) where pathogenic taxa consist of genus commonly associated with bacterial vaginosis including Gardnerella, Prevotella, Anaerococcus, Streptococcus, Megasphaera, Mobiluncus, Sneathia, Atopobium, Peptoniphilus, Mycoplasmoides, Ureaplasma, Bacteroides, Peptostreptococcus, and Dialister.

Disease prediction model construction using a random forest classifier

Samples were divided into proliferative (n = 138) and secretory (n = 128) groups for the following analysis, and the subsequent analysis was based on bacterial species relative abundances. MaAsLin2 [34] was performed to determine the multivariable association between bacterial species and endometriosis/non-endometriosis groups (P ≤ 0.05), with age and BMI controlled. A complementary analysis using ALDEx2 [35, 36] was also performed (Additional file 1: Table S1). Features with importance scores ≥ 0.015 to the endometriosis/non-endometriosis groups among bacterial species were selected via random forest implemented in Python sklearn package. Models to predict endometriosis/non-endometriosis were built based on features that were selected by MaAsLin2, by random forest feature scoring and by the addition of the functional dysbiosis score. Model performance was assessed through 50 iterations of repeated random subsampling cross-validation, in which the data was randomly split into 80% for training and 20% for testing in each iteration. This strategy helped account for variability arising from random data splits and yielded a more robust estimate of the predictive accuracy.

Abbreviations

BMI

Body mass index

FDR

False discovery rate

FDS

Functional dysbiosis score

AUC

Area under the curve

AI

Artificial intelligence

rASRM

Revised American Society for Reproductive Medicine

Authors’ contributions

LZ and JH analyzed the data. LZ, JH, WHW, FZB and XZ conceptualized the study. LZ, JH and WHW wrote the manuscript with input from all authors. XZ and LZ provided the samples for sequencing. SL coordinated sample collection and shipment to the laboratory. YY and XX performed molecular experiments. WHW, FZB and XZ supervised the study. FZB and XZ secured R&D funding for the study. All authors read and approved the final manuscript.

Funding

The research was funded by the internal R&D budget at HerAnova Lifesciences Inc, and by the National Key R&D Program of China (2022YFC2704003).

Data availability

Sequencing data were deposited into Genome Sequence Archive (accession number PRJCA041172).

Declarations

Ethics approval and consent to participate

This study was approved by the institutional review board of the Women’s Hospital, Zhejiang University School of Medicine (IRB-20240110-R), and consent was given by enrolled participants.

Consent for publication

The enrolled participants provided informed consent for the use of their biological materials in research and publication. All authors have reviewed and approved this manuscript for publication.

Competing interests

All authors, except Zhang Xinmei and Zhu Libo, are employees of HerAnova Lifesciences, a company engaged in the commercial development of a non-invasive test for endometriosis. Farideh Z Bischoff, Yanqin Yu and Wing Hing Wong hold stock options of HerAnova Lifesciences. The authors declare no other conflicts of interest, financial or otherwise.