Gene expression profiles separate endometriosis lesion subtypes and indicate a sensitivity of endometrioma to estrogen suppressive treatments through elevated ESR2 expression

Normalisation techniques for gene expression in both datasets were performed as previously described [18, 22]. Data were pre-processed using Illumina GenomeStudio software (Illumina Inc., San Diego) where data was background corrected and any probes with a detection p-value less than 0.05 was considered expressed in that sample. Data was subsequently quantile normalised and log transformed. In total, 48,701 probes were expressed in dataset A and 47,235 probes were expressed in dataset B.

Principal component analysis (PCA) was used to evaluate variation in gene expression between lesion samples in the dataset A. PCA was performed on normalised log-transformed expression values using the prcomp function in R (V4.1.3). ANOVA was used to test the association between the top five PCs as well as the menstrual cycle stage and lesion subtype (SUP, OMA, DIE).

To avoid introducing expression bias with genes not expressed in some samples, the analysis was restricted to genes expressed in > 90% of all samples, leaving 14,747 probes in dataset A and 12,247 probes in dataset B [18].

In endometrial samples, differential gene expression was performed using the eBayes method in the limma package (R version-4.1.3). A comparison was conducted between the proliferative and secretory stages with endometriosis status as a covariate in each dataset. Within dataset A differential expression between samples collected from those taking and not taking hormonal medication was also tested with menstrual stage as a covariate. The Pearson correlation coefficient was used to determine the correlation between log-transformed gene expression in the two datasets.

In endometriotic lesions, normalised and batch-corrected gene expression levels were used to conduct differential gene expression analysis between samples collected at different menstrual cycle stages. This was conducted in lesions of all subtypes, as well as in each subtype separately (SUP v DIE; SUP v OMA; DIE v OMA). P-values were corrected for multiple testing using the Bonferroni method and Benjamini–Hochberg method and an adjusted p-value less than 0.05 was considered statistically significant. The top 50 differentially expressed genes (DEGs) were visualised and clustered using the heatmap function in R.

Using the normalised gene expression dataset, gene expression across lesion subtypes for nine hormone receptors (oestrogen receptor 1 (ESR1), oestrogen receptor 2 (ESR2), androgen receptor (AR), PGR, progesterone receptor membrane component 1 (PGRMC1), progesterone receptor membrane component 2 (PGRMC2), gonadotropin-releasing hormone 1 (GnRH1), gonadotropin-releasing hormone 2 (GnRH2), mineralocorticoid receptor (MCR)) was investigated using a two-way ANOVA. Some genes were not expressed in more than 90% of samples; hence, this analysis was not restricted. Covariates included were batch number, menstrual stage and medicated status. A p-value of 0.05 or less was deemed statistically significant.

A genome-wide correlation analysis was conducted on the medicated and un-medicated samples separately for ESR1 across all lesion subtypes and for ESR2 in OMA subtype only. Pair-wise correlations were performed controlling for the false discovery rate (FDR) using the Benjamini–Hochberg method. An FDR < 0.05 was considered significant.

Using the xCell pipeline in R [23], cell-type enrichment analysis was performed on dataset A to estimate the cell-type composition in each sample from lesion data. xCell is a bioinformatics application that generates cell type enrichment scores by comparing expression profiles to reference data for 64 cell types.

ClusterProfiler [24], which is capable of analysing and visualising data for enrichment analysis, was utilised to acquire better biological insight into differentially expressed genes identified as significant from the differential gene expression analysis. Enrichment was considered significant if Benjamini–Hochberg adjusted p-values were less than 0.05.

The correlation between expressed genes was assessed to establish consistency across independent datasets. A direct comparison of constitutively expressed endometrial genes in the endometrial samples from all patients revealed a significant correlation between dataset A and dataset B (r² = 0.70) (Fig. 2a). Comparing the gene expression in the endometrial samples from the proliferative and secretory stages independently identified 3202 and 8361 FDR significant DEG in dataset A and dataset B, respectively. The correlation of these DEG between the two datasets was also significant (r² = 0.9) (Fig. 2b), indicating the consistency of endometrial gene expression across datasets generated at different locations, using different microarray chips, and from individuals of different European ancestry.

As endometriosis status may have a confounding effect on endometrial gene expression [18], analysis was limited to cases exclusively. A comparison of endometrial samples from the proliferative (n = 16) and secretory (n = 21) stages revealed 2835 DEGs following FDR correction and 142 DEGs following Bonferroni correction (Additional file 1: Table S2; Fig. 3a), confirming the influence of cycle stage in cases only. A pair-wise comparison of hormone-medicated (n = 21) endometrium samples and endometrium from the un-medicated proliferative stage identified 1012 DEGs following FDR correction and six DEGs following Bonferroni correction (Additional file 1: Table S3; Fig. 3b), and with endometrium from the un-medicated secretory stage samples identified 77 DEGs following FDR correction (Additional file 1: Table S4; Fig. 3c). No genes were significantly different between hormone-medicated samples and the secretory stage following Bonferroni correction.

Of the 1012 DEG in the medicated and proliferative stage endometrial samples, a total of 562 genes were upregulated and 450 were downregulated in medicated samples. In contrast, 32 were upregulated in medicated samples and 45 downregulated compared to the secretory stage. The influence of the menstrual stage and hormones created distinct profiles for each gene. FAM125B was significantly upregulated in medicated compared to secretory stage samples (p_prol = 0.10; p_secr = 2.58 × 10⁻⁵). WNT7A displayed similar concentrations in medicated and proliferative stage samples but was significantly upregulated compared to secretory stage samples (p_prol = 0.77; p_secr = 0.00024). FKBP5 expression was similar between medicated and secretory stage samples, but significantly upregulated compared to proliferative stage samples (p_prol = 9.51 × 10⁻⁵; p_secr = 0.26) (Fig. 3d). A gene ontology analysis suggested that the major distinction between proliferative and secretory stages is cellular processes (Fig. 3e). In contrast, medicated samples demonstrated change in MHC class II protein activity relative to the proliferative stage and morphogenesis variation compared to the secretory stage. Combined together this data confirms hormonal treatment significantly influenced gene expression in endometrium. It influenced each gene differently, although the gene expression profiles of medicated samples appear more similar to the secretory stage, as evidenced by the smaller number of DEGs, compared to the proliferative endometrium.

To explore gene expression differences in lesions, we first assessed the menstrual cycle given its significant influence on the endometrium. In contrast to the endometrium, no significant association was observed between variation in gene expression and different cycle stages in the lesions (Fig. 4a). Of these lesions, 29 were SUP (11 proliferative and 18 secretory), 19 were OMA (10 proliferative and 9 secretory), and 26 were DIE (10 proliferative and 16 secretory). In lesions stratified by subtype, the influence of menstrual stage on gene expression for SUP, OMA, or DIE was similarly not significant.

PCA revealed clustering of samples with PC2 significantly associated with subtype (p = 1.37 × 10⁻¹³) suggesting subtype could explain a significant amount of variability in gene expression (Fig. 4b). Analyses of differential gene expression between SUP and DIE found 1108 and 27 DEGs following FDR and Bonferroni correction, respectively (Additional file 1: Table S5). Differential expression analysis between SUP and OMA found 1677 and 202 DEGs following FDR and Bonferroni correction, respectively (Additional file 1: Table S6). Between OMA and DIE lesions; 2663 and 334 DEGs following FDR and Bonferroni correction, respectively (Additional file 1: Table S7). A heatmap analysis based on top 50 genes from each combination confirmed OMA clustered separately from the DIE and SUP lesions. Additionally, the DIE and SUP lesions also formed separate clusters (Fig. 4c). Gene ontology analysis revealed that differences in gene expression drove variation in pathways relevant to antigen processing presentation in peritoneal lesions, while DIE lesions showed variation in growth regulation pathways (Fig. 4d).

Lastly, to estimate cell types within lesion tissue and to determine whether clustering could be driven by differences in cell type composition, gene expression profiles of each sample were compared to the Human cell atlas using xCell. A comparison of the cell type enrichment indicated the presence of a considerable stem cell component consisting of mesenchymal stem cells (MSC), multipotent progenitor cells (MPP) and hematopoietic stem cells (HSC) (Fig. 5). Samples were also enriched for immune cells, in particular NKT cells and eosinophils. The expression level of epithelial and fibroblast cells was lower but consistent across all sample types. Comparison to the human cell atlas however did not indicate that there were sufficient differences between cell type composition to drive sample clustering. The findings of this study indicate that the subtype of the lesion, rather than the menstrual stage or cell composition, may be responsible for the observed differences in gene expression profiles amongst individual lesions. These differences in gene expression profiles could potentially contribute to variations in the functional characteristics of specific lesions. However, it is important to note that the validity of identifying the cell composition of mixed tissue using will be limited by the genes that are included on the array.

In the absence of menstrual cycle stage effects on lesions, samples from all cycle stages were combined to test the effect of hormonal treatment (Fig. 6). Unsupervised heatmap clustering based on the top 100 genes from the DE analysis revealed that hormonally medicated OMA samples cluster separately from un-medicated OMA samples (Additional file 1: Table S8, Fig. 6a). In contrast, no clustering was observed in SUP (Additional file 1: Table S9, Fig. 6b) or DIE lesions (Additional file 1: Table S10, Fig. 6c) that could be attributed to the medication. Genome-wide differential expression analysis between medicated and non-medicated samples in each lesion subtype resulted in 569 (DIE), 836 (OMA) and 638 (SUP) nominally significant genes (p < 0.05).

The expression of target receptors for steroid hormones was evaluated by plotting their expression in both medicated and non-medicated samples and in each lesion subtype. The expression of ESR2 (Fig. 7a), ESR1 (Fig. 7b), AR (Fig. 7c), PGR (Fig. 7d), the progesterone membrane receptors PGRMC1 (Fig. 7e) and PGRMC2 (Fig. 7f), GnRH1 (Fig. 7g), GnRH2 (Fig. 7h) and MCR (Fig. 7i) was observed in each lesion subtype. A significant increase in the expression of ESR2 (p < 2 × 10⁻¹⁶) was observed in the OMA lesions compared to SUP, or DIE. A significant decrease in the expression of ESR1 (p = 4.42 × 10⁻⁵) and PGR (p = 4.85 × 10⁻⁵) was observed in OMA lesions compared to DIE and SUP lesions. Additional analysis of ESR2 expression revealed that only 52 SUP and 53 DIE samples expressed ESR2. Together this data suggests a coordinated response to exogenous hormones in OMA and that an elevated ESR2 expression may be important in this process. How this influence is mediated is not yet clear as changes stimulated by hormonal treatment to endogenous hormones may also impact the gene expression profiles.

Steroid receptors are transcription factors that modulate gene transcription by interacting with co-regulators. A pair-wise correlation between steroid receptors in SUP lesions without (Fig. 8a) and with (Fig. 8b) treatment, OMA lesions without (Fig. 8c) and with (Fig. 8d) and DIE lesions without (Fig. 8e) and with (Fig. 8f) treatment was assessed. This analysis revealed a similar correlation pattern between the SUP and DIE lesions in samples derived from patients with and without treatment, but distinct patterns in OMA samples.

In SUP lesions (Fig. 8a, b), the direction of correlation between the hormonal receptors was consistent between medicated and non-medicated samples. There was a significant positive correlation between ESR1 and PGR and the membrane PGR receptors PGRMC1 and PGRMC2. DIE lesions (Fig. 8e) revealed a significant positive correlation between PGR and ESR1 as well as PGRMC1. ESR1 was also correlated significantly with AR, but negatively with NR3C2. Additionally, NR3C2 exhibited a significant positive correlation with PGRMC2. In hormone-medicated patients (Fig. 8f), the direction of these correlations remained consistent, with ESR1 exhibiting a significant positive correlation with PGR and a negative association with NC3R2, and PGR exhibiting a positive correlation with AR.

In contrast, both the direction and strength of the correlation between receptors was significantly altered in OMA samples taken from medicated patients compared to non-medicated. The correlation between ESR1 and PGR was positive in non-medicated samples (Fig. 8c), with no expression of PGR in medicated samples (Fig. 8d). A negative association between ESR1 and NR3C2 was observed in non-medicated (positive) and with no expression of NR3C2 in medicated patients. There was a positive correlation between ESR1 and PGRMC1. Correlations with AR were also impacted after treatment with a stronger correlation with ESR2 (negative) and ESR1 (positive) that were not evident in non-medicated patients. According to this study, OMA appears more sensitive to exogenous hormonal treatment than SUP or DIE, which generated distinct hormone receptor profiles after exposure.

To determine whether the increased expression of ESR2 in OMA provided the potential for exogenous hormones to initiate a change in gene expression profiles, the top 20 genes correlated with ESR2 expression were evaluated (Fig. 9), and the direction and strength was compared between medicated and non-medicated patients. This analysis revealed a significant alteration in the genes correlated with ESR2 in OMA samples from medicated and non-medicated patients. At least 10 of these 20 genes, including MEGF8, ENG, CCDC95, HERC2, and GLI4, exhibited an opposite direction of action in medicated and non-medicated samples (Fig. 9), indicating that genes associated with ESR2 are significantly altered by medication.