This study included 99 endometrial samples, comprising 60 EC tissues and 39 matched normal endometrial tissues (non-malignant, collected within 2 cm of the tumor). Samples were obtained at Fujian Maternity and Child Health Hospital during routine clinical procedures. All patients provided written informed consent for their tissues to be used in this study.

Specifically, 35 EC samples and 15 normal endometrial samples were used for metabolomics analysis. A further 24 EC samples and 23 matched adjacent normal endometrial tissues were used for proteomics analysis. The final EC sample and paired adjacent normal endometrial tissue were used for single-cell RNA sequencing. The datasets were generated from independent sample cohorts, with no overlap among different omics analyses.

After collection, EC and adjacent normal tissues (Supplementary Table 1) were numbered, weighed, and placed in EP tubes. Extraction buffer (methanol: acetonitrile: water = 2:2:1, with isotope-labeled standards) was added to the samples, which were then homogenized and sonicated in cycles. They were subsequently incubated at –40 °C for 1 h and then centrifuged at 12,000 rpm for 15 minutes at 4 °C. The supernatant was transferred to glass vials for liquid chromatography-mass spectrometry/mass spectrometry analysis (Thermo Fisher Vanquish system, Waltham, MA, USA). Metabolomics analysis was performed by BIOTREE BIOTECH (Shanghai, China), and the metabolites were identified using an in-house database.

Differential metabolites were identified using two criteria: p 0.05 and a variable importance in projection score $>$1 for the first principal component in the OPLS-DA model. Enrichment analysis of these metabolites was then conducted with reference to Kyoto Encyclopedia of Genes and Genomes (KEGG) and PubChem databases [9, 10, 11].

The study workflow is shown in Supplementary Fig. 1. TCGA-Uterine Corpus Endometrial Carcinoma (UCEC) data were obtained from The Cancer Genome Atlas, GSE17025 data from the Gene Expression Omnibus (GEO), and PDC000125 data from the CPTAC database.

Using the WGCNA package (Version 1.73, Bioconductor, USA) [12], 539 tumor samples and 35 normal samples from TCGA-UCEC were analyzed to find gene modules linked to EC development. We started by selecting the top 25% most variable genes and calculating a Pearson correlation matrix based on mRNA data. After clustering samples to identify outliers, the optimal soft threshold ($\beta$) was determined for network construction. The correlation matrix was converted into a topological overlap matrix, and dynamic tree cutting was used for hierarchical clustering to group co-expressed genes into modules.

The limma package (Version 3.58.1, Bioconductor, USA) [13] was used for differential expression analysis of TCGA-UCEC and GEO data, with screening criteria of $|$log2-fold change$|$ $\ge$1 and adjusted p 0.05.

The ConsensusClusterPlus package (Version 1.66.0, Bioconductor, USA) [14] was used for consensus clustering analysis, with 1000 iterations for stability and reproducibility. The principal component analysis (PCA) method determined the optimal cluster number (k = 2–10), and the internal consistency index assessed the clustering stability. PCA was used to evaluate subtype separation based on expression patterns. Survival analysis was performed using the survival (Version 3.5.7, CRAN, Vienna, Austria) and survminer (Version 0.4.9, CRAN, Vienna, Austria) packages to compare prognostic outcomes across EC subtypes.

A total of 537 patients with survival data were selected from the TCGA database and randomly split into training and test sets at a 7:3 ratio. Univariate Cox regression was used to analyze 90 key metabolic genes for survival associations, identifying significant prognostic genes. Least absolute shrinkage and selection operator (LASSO) regression further filtered prognosis-related genes, and common genes from both methods were analyzed with multivariate Cox regression. The prognostic model was built using LASSO-Cox regression. The risk score was calculated using the following formula: $\sum$ (Xi * Yi) (X: coefficient, Y: gene expression level). Kaplan-Meier (K-M) curves were used to compare survival across risk groups, and risk-survival curves for mortality rates. To validate the model, receiver operating characteristic (ROC) analysis at 1, 3, and 5 years was performed using the survival (Version 3.5.7, CRAN, Vienna, Austria), survminer (Version 0.4.9, CRAN, Vienna, Austria), and timeROC packages (Version 0.4, CRAN, Vienna, Austria) [15].

The maftools package (Version 2.18.0, Bioconductor, USA) was used to retrieve tumor mutation data for UCEC patients from the TCGA database. The mutation status of different risk populations was analyzed, and the combined effects of risk scores and TMB on patient survival outcomes were examined.

Tumor immune dysfunction and exclusion (TIDE) scores for UCEC patients were obtained from the TIDE database. The chi-square test was used to assess differences in immunotherapy responses between different risk groups of EC patients. The CIBERSORT algorithm [16] was employed to assess differences in the levels of infiltrating immune cells in the tumor microenvironment (TME). The limma package was used for differential expression analysis of several classical immune checkpoint genes, while the estimate package (Version 1.0.13, CRAN, Vienna, Austria) was used to evaluate the TME.

Formalin-fixed, paraffin-embedded tissue samples were deparaffinized with xylene and ethanol, then homogenized in lysis buffer containing Tris-hydrochloride, dithiothreitol, and sodium dodecyl sulfate. After incubation at 99 °C for 60 minutes, the samples underwent ultrasonication, centrifugation, iodoacetamide alkylation, and acetone precipitation. Proteins were resuspended in 8M urea and analyzed by mass spectrometry using DIA-NN (Version 1.8, Berlin, Germany) with a false discovery rate set at 1%. Fisher’s exact test and Pearson’s test identified differentially expressed proteins (p 0.05), with fold-changes $\ge$1.2 or $\le$1/1.2 considered significant.

Fresh tumor tissues were cut into 1–3 mm³ fragments and single-cell suspensions subsequently prepared using a single-cell 3^′ Library and Gel Bead Kit V3.1 (10x Genomics, Pleasanton, CA, USA). Following enzymatic digestion and cell purification, the cell viability exceeded 90%, with a final cell concentration of 700–1200 cells/µL. Gel bead-in-emulsions (GEMs) were generated using a Chromium Single-Cell G Chip Kit (10x Genomics, Pleasanton, CA, USA), after which cell lysis, reverse transcription with unique molecular identifiers (UMIs), cDNA amplification, fragmentation, end repair, and adapter ligation were sequentially performed according to the manufacturer’s protocol. The resulting libraries were sequenced on the Illumina platform. Raw sequencing data were processed using Cell Ranger (10x Genomics, Pleasanton, CA, USA) for read alignment, barcode processing, and UMI counting. Downstream analyses were performed in Seurat (Version 4.3.0, CRAN, Vienna, Austria), including quality control (cells with $>$200 detected genes and $\le$25% mitochondrial gene expression), normalization, PCA, and uniform manifold approximation and projection (UMAP). This enabled cell type identification and characterization of subpopulations.

GSVA was conducted to evaluate enrichment with hallmark pathways using the GSVA package (Version 1.50.0, Bioconductor, USA). Hallmark gene sets were obtained from MSigDB through the msigdbr package (Version 7.5.1, CRAN, Vienna, Austria) and converted to Entrez IDs to match the RNA-seq expression matrix. Differential pathway activity between groups was analyzed with limma using empirical Bayes moderation, with significance defined as p 0.05 and $|$T$|$ $>$2.

All cell lines were maintained in culture media containing 10% fetal bovine serum (FBS; Gibco, MT, USA) and 1% antibiotic–antifungal mixture (Basal Media, Shanghai, China). HEC-1A cells (KeyGEN BioTECH, Nanjing, China) were grown in McCoy’s 5A medium without sugar and supplemented with glucose (Procell Life Science & Technology Co., Ltd., Wuhan, China). Ishikawa cells (KeyGEN BioTECH, Nanjing, China) were maintained in RPMI-1640 medium containing glucose (Procell Life Science & Technology Co., Ltd., Wuhan, China), whereas endometrial epithelial cells (EECs) (Whelab, Shanghai, China) were cultured in MEM supplemented with non-essential amino acids (NEAA; Procell Life Science & Technology Co., Ltd., Wuhan, China). All cell lines were maintained at 37 ℃ in a humidified atmosphere, validated by STR profiling, and tested negative for mycoplasma (Biowing Biotechnology Co., Ltd., Shanghai, China).

Cells were lysed directly in the culture dishes using RNA extraction reagent, and the lysates were thoroughly vortexed. After centrifugation, the supernatant was collected into a new tube. For phase separation, 0.25 mL of chloroform was added to every 1 mL of the extraction reagent. Following another centrifugation, the upper aqueous phase was carefully transferred to a clean tube. RNA was precipitated with isopropanol and subsequently washed with 75% ethanol. First-strand cDNA synthesis was carried out according to the manufacturer’s instructions. Following preparation of the PCR master mix (Promega, Madison, WI, USA), quantitative PCR was conducted using the Eastep qPCR Master Mix (Promega, USA). The primer sequences are shown in Table 1. GAPDH was used as the reference gene to determine relative gene expression levels with the 2^{-Δ⁢Δ⁢CT} method.

Statistical analyses were conducted in R (Version 4.3.1, R Foundation for Statistical Computing, Vienna, Austria) and GraphPad Prism 9.5 (GraphPad Software, San Diego, CA, USA), with data visualized using ggplot2 (Version 3.5.2, CRAN, Vienna, Austria) and ggpubr (Version 0.6.0, CRAN, Vienna, Austria). Group comparisons were carried out with Student’s t-test (normal data), Wilcoxon rank-sum test (non-normal data), and Kruskal-Wallis test (multiple groups). Chi-square tests were used for categorical variables. Survival analyses were performed with the K-M method. Results with p 0.05 were considered statistically significant.

We performed metabolomic analysis to investigate whether metabolic reprogramming occurs in EC (Fig. 1A). PCA revealed distinct metabolic profiles between EC and normal endometrial tissues (Fig. 1B). Subsequent differential analysis identified 4529 upregulated and 3222 downregulated metabolites (Fig. 1C). Pathway enrichment analysis revealed significant enrichment of the Gly/Ser/Thr metabolism pathways (Fig. 1D). We then screened for genes associated with this pathway, and identified and matched 1741 genes with relevance scores $>$10 through the GeneCards database (Supplementary Table 2). WGCNA was performed on the TCGA dataset, with a soft threshold of $\beta$ = 7 (R² = 0.85) to construct a scale-free network (Fig. 1E,F). Six distinct modules were identified (Fig. 1G), among which the green module exhibited the strongest correlation with tumor tissues and contained 277 genes. Differential analysis of tumor and normal tissues identified 1095 up-regulated genes (log₂$|$fold change$|$ $>$1, adjusted p 0.05) (Fig. 1H). Intersection analysis of Gly/Ser/Thr metabolism–related WGCNA module genes and up-regulated differentially expressed genes (DEGs) in EC yielded 90 candidate genes (Fig. 1I).

Fig. 1.

Metabolomics and WGCNA analysis for the screening of Gly/Ser/Thr metabolism-related genes. (A) Metabolomics workflow. This figure was created in BioRender (https://BioRender.com/tkklo3h). (B) 3D PCA score scatter plot. (C) Volcano plot of differential metabolites. (D) Enrichment pathway map of differential metabolites. (E) Scale independence and mean connectivity analysis for soft-thresholding selection. (F) Gene dendrogram and module color assignment. (G) Module-trait relationship heatmap. (H) Volcano plot of DEGs. (I) Venn diagram of overlapping genes. LC-MS, liquid chromatography-mass spectrometry; MS2, secondary mass spectrometry; PCA, principal component analysis; WGCNA, weighted gene co-expression network analysis; Gly/Ser/Thr, glycine, serine, and threonine; DEGs, differentially expressed genes.

Unsupervised clustering analysis was performed on 539 EC tissues from the TCGA database using expression profiles of Gly/Ser/Thr metabolism-related genes. Optimal clustering stability was achieved at k = 3 (Fig. 2A–D), dividing patients into three distinct subgroups (C1: n = 266; C2: n = 196; C3: n = 77). PCA revealed significant transcriptomic heterogeneity among clusters (Fig. 2E). Survival analysis demonstrated pronounced prognostic heterogeneity, with C1 exhibiting the shortest overall survival (OS) and disease-specific survival (DSS) (Fig. 2F,G). Notably, C1 contained significantly higher proportions of advanced International Federation of Gynecology and Obstetrics (FIGO) stage (III/IV) and high-grade (G3) tumors (Fig. 2H,I), indicating greater clinical aggressiveness.

Fig. 2.

Identification of Gly/Ser/Thr metabolism-related subtypes. (A) Consensus clustering matrices for k = 2–5. (B) CDF curves for k = 2–10. (C) Relative change in the area under the CDF delta curve. (D) Tracking plot for k = 2–10. (E) PCA demonstrating three distinct subgroups. (F,G) Differences in OS and DSS among the three subtypes. (H,I) Bar plots illustrating the distribution of different pathological grades and FIGO stages across subtypes. CDF, cumulative distribution function; PCA, principal component analysis; OS, overall survival; DSS, disease-specific survival; FIGO, International Federation of Gynecology and Obstetrics.

Consistent with established literature demonstrating metabolic-driven immune evasion in the TME [17, 18], analysis of immune cell infiltration revealed distinct imbalances in the C1 phenotype, with increased Th2 cells and decreased activated CD8⁺ T cells, DCs, NK cells, and myeloid-derived suppressor cells (MDSCs) (Fig. 3A). This phenotype likely reflects STAT6 pathway activation via Th2-secreted IL-4/IL-13, promoting M2 macrophage polarization and immunosuppression [19]. Concurrent DC dysfunction may impair antigen presentation and CD8⁺ T cell activation [20], while NK cell exhaustion compromises innate immune surveillance, collectively facilitating metastatic dissemination. In contrast, C2 showed enhanced anti-tumor immunity with activated CD4⁺ T cells and DCs, while C3 had an immunomodulatory environment dominated by NK cells, mast cells, and Tregs. Immunocheckpoint analysis showed significant upregulation of CD274 (PD-L1), lymphocyte activation gene 3 (LAG3), T-cell immunoreceptor with Ig and ITIM domains (TIGIT), and interferon gamma receptor 1 (IFNGR1) in C1 (Fig. 3B). Moreover, C1 demonstrated the lowest immune score, the highest TIDE score (Fig. 3C,D), elevated tumor purity, and a lower stromal score (Fig. 3E,F), suggesting an immunosuppressive environment with poor response to immunotherapy.

Fig. 3.

Immune analysis of the different metabolic subtypes. (A) Comparison of infiltration levels of immune cell types among the three metabolic subtypes. (B) Expression levels of immune checkpoint-related genes in the three metabolic subtypes. (C–F) Evaluation of TIDE score (C), immune score (D), stromal score (E), and tumor purity score (F) in the three metabolic subtypes. ns, not significant, * p 0.05, ** p 0.01, *** p 0.001, **** p 0.0001. TIDE, tumor immune dysfunction, and exclusion; LAG3, lymphocyte activation gene 3; TIGIT, T-cell immunoreceptor with Ig and ITIM domains; IFNGR1, interferon gamma receptor 1; CTLA-4, cytotoxic T-lymphocyte-associated protein 4; PDCD1, programmed cell death 1; SIGLEC15, sialic acid-binding Ig-like lectin 15; HAVCR2, hepatitis A virus cellular receptor 2.

To evaluate the prognostic significance of metabolic genes, we developed a prognostic model using data from 537 TCGA-UCEC cases. These were randomly divided into training (n = 429) and test (n = 108) cohorts at a ratio of 7:3. Univariate Cox regression identified 29 OS-associated metabolic genes (Fig. 4A, Supplementary Table 3). These were narrowed to 6 via LASSO-Cox regression ($\lambda$ = 0.0351; Fig. 4B, Supplementary Table 4), and then further refined to three core genes (methylenetetrahydrofolate dehydrogenase 2 (MTHFD2), ribosomal protein S6 kinase A1 (RPS6KA1), cyclin-dependent kinase inhibitor 2A (CDKN2A)) using multivariate Cox analysis (Fig. 4C). The risk score was calculated as: (0.275 $\times$ MTHFD2 expression) + (–0.249 $\times$ RPS6KA1 expression) + (0.28 $\times$ CDKN2A expression). Patients were stratified into high- and low-risk groups based on their median risk score (Supplementary Table 5). K-M analysis revealed significantly worse OS in the high-risk group (p 0.001, Fig. 4D). ROC analysis confirmed high predictive accuracy, with the area under the curve (AUC) $>$0.7 for 1-, 3-, and 5-year OS (Fig. 4E). In both the test cohort and the overall dataset, high-risk scores correlated with poorer survival (p 0.05), with AUC $>$0.6 (Fig. 4F,G,J,K). Scatter plots further showed increased EC patient mortality with higher risk scores (Fig. 4H,I,L,M). Together, these results indicate the three-gene model (MTHFD2, RPS6KA1, CDKN2A) is a reliable predictor of EC prognosis.

Fig. 4.

Construction and validation of the risk prediction model. (A) Identification of genes associated with OS. (B) Variable selection in the LASSO model. (C) Determination of the most relevant genes. (D,F,J) K-M survival analysis comparing high- and low-risk groups. (E,G,K) Time-dependent ROC curves for predicting 1-, 3-, and 5-year survival rates. (H,I,L,M) Distribution of risk scores and survival status for each patient. OS, overall survival; LASSO, least absolute shrinkage and selection operator; K–M, Kaplan–Meier; ROC, receiver operating characteristic.

Tumor samples from 24 EC patients with different prognoses were analyzed by proteomic profiling (Fig. 5A). Stratifying patients by risk scores revealed the high-risk group had significantly worse survival (p 0.05, Fig. 5B). AUC values for predicting 1-, 3-, and 5-year survival were all $>$0.7 (Fig. 5C–E), confirming the model’s robustness. Scatter plots also showed a positive correlation between mortality and risk scores, further validating the prognostic model (Fig. 5F,G).

Fig. 5.

Validation of the risk model through proteomics. (A) Workflow for the proteomics study. This figure was created in BioRender (https://BioRender.com/tkklo3h). (B) K-M survival analysis comparing high- and low-risk groups. (C–E) Time-dependent ROC curves for predicting 1-, 3- and 5-year survival rates. (F,G) Distribution of risk scores and survival status for each patient. K–M, Kaplan–Meier; ROC, receiver operating characteristic.

Analysis of mutational profiles showed that missense mutations were the main variant type in both risk groups, with higher frequencies of PIK3CA and TP53 mutations in high-risk patients (TP53: 48% vs. 11%) (Fig. 6A,B). High-risk tumors had a lower TMB (Fig. 6C,D), suggesting reduced neoantigen production and poorer treatment response. High-risk scores were consistently observed across microsatellite stability (MSS), microsatellite instability low (MSI-L), and MSI-high (MSI-H) subtypes. MSS tumors showed particularly high scores, whereas MSI-H tumors with high mutational load and typically better immunotherapy response were predominantly clustered in the low-risk group (Fig. 6E).

Fig. 6.

Analysis of tumor mutations and immune microenvironment infiltration. (A,B) Comparison of mutation frequency and type based on risk scores. (C) Violin plot comparing tumor mutation burden (log2-transformed) between low- and high-risk groups. (D) Bar plot showing the proportion of tumor mutation burden in low- and high-risk groups. (E) Distribution of microsatellite stability subtypes in low- and high-risk groups. (F) Levels of infiltration by different immune cell types. (G) Differential distribution of immune risk scores between responders and non-responders to immunotherapy. (H–K) Correlation analysis between immune-related scores and risk scores. * p 0.05, ** p 0.01, *** p 0.001, **** p 0.0001. MSS, microsatellite stable; MSI-L, microsatellite instability-low; MSI-H, microsatellite instability-high.

CIBERSORT analysis of immune cell infiltration patterns revealed distinct microenvironment profiles between the two risk groups. Low-risk tumors were characterized by higher proportions of CD8⁺ T cells, Tregs, and resting DCs, whereas high-risk tumors showed enrichment of naïve B cells, M1 macrophages, activated DCs, and neutrophils (Fig. 6F). TIDE score analysis predicted poorer response to immune checkpoint blockade (ICB) in high-risk patients (p 0.05), indicating greater immune escape potential (Fig. 6G). Analysis with the ESTIMATE algorithm revealed an inverse correlation between risk scores and immune scores (cor –0.1, p 0.05) (Fig. 6H). The high-risk group showed lower stromal scores and ESTIMATE scores, together with higher tumor purity (Fig. 6I–K), collectively suggesting an immunologically suppressed TME.

Single-cell RNA sequencing of paired tumor and adjacent normal endometrial tissues from one EC patient identified five major cell populations: epithelial cells (epithelial cell adhesion molecule (EPCAM), keratin 8 (KRT8)), stromal fibroblasts (regulator of G-protein signaling 5 (RGS5)), endothelial cells (adhesion G protein-coupled receptor L4 (ADGRL4); plasmalemma vesicle associated protein (PLVAP)), smooth muscle cells (decorin (DCN)), and immune cells (protein tyrosine phosphatase receptor type C (PTPRC)) (Fig. 7A,B). UMAP analysis revealed widespread MTHFD2 expression in non-immune cells, RPS6KA1 expression in both epithelial and endothelial cells, and scattered CDKN2A expression across epithelial, immune, and smooth muscle cells (Fig. 7C). Validation with TCGA-UCEC, GEO, and CPTAC confirmed tumor-upregulated expression of all three genes (Fig. 7D,E). MTHFD2 and CDKN2A were highly expressed in higher-grade tumors and TP53-mutated samples, while RPS6KA1 expression was lower in cases with poor prognosis (Fig. 7F,G). Survival analysis indicated that high MTHFD2 expression and CDKN2A expression were adverse prognostic factors, while downregulation of RPS6KA1 also predicted poor outcome (Fig. 7H).

Fig. 7.

Validation of prognostic genes in external datasets. (A) UMAP plot displaying five major cell type clusters. (B) UMAP visualization of canonical marker genes across five cell subpopulations. (C) Expression profiles of the three prognostic genes across cell subpopulations. (D,E) Differential expression of the three prognostic genes in TCGA-UCEC and GSE17025 datasets. (F) Expression patterns of the three prognostic genes across different pathological grades in the CPTAC database. (G) Expression levels of the three prognostic genes: comparison between TP53-mutated and wild-type samples. (H) Prediction of disease outcome with the three prognostic genes in the TCGA-UCEC cohort. ns, not significant, * p 0.05, **** p 0.0001. UMAP, uniform manifold approximation and projection; EPCAM, epithelial cell adhesion molecule; KRT8, keratin 8; RGS5, regulator of G-protein signaling 5; ADGRL4, adhesion G protein-coupled receptor L4; PLVAP, plasmalemma vesicle associated protein; DCN, decorin; PTPRC, protein tyrosine phosphatase receptor type C; MTHFD2, methylenetetrahydrofolate dehydrogenase 2; RPS6KA1, ribosomal protein S6 kinase A1; CDKN2A, cyclin-dependent kinase inhibitor 2A.

To clarify the roles of prognostic genes in the EC subgroups, we further analyzed their expression and pathway associations. MTHFD2 and CDKN2A were highly expressed in C1, while RPS6KA1 was highly expressed in C2 (Fig. 8A,B), consistent with the subgroup outcomes. Risk score analysis showed that C1 had the highest scores (Fig. 8C), matching its poor prognosis. GSVA revealed distinct signatures: C1 showed global metabolic suppression (fatty acid, bile acid, cholesterol, hormone signaling) and inhibition of P53, apoptosis, and ROS response, indicating dedifferentiation and immune-metabolic suppression (Fig. 8D). The C1 subgroup also exhibited low CD8⁺ T/NK infiltration and high CD274/PD-L1 and LAG3, defining an “immune-metabolic suppressed type” with the worst prognosis. C2 featured active proliferation (E2F transcription factor/MYC proto-oncogene, BHLH transcription factor (E2F/MYC) targets, G2/M checkpoint), partial interferon-$\alpha$/$\gamma$ activation, and reduced Wnt/$\beta$-catenin signaling (Fig. 8E). Despite increased MDSCs, the C2 subgroup had high CD4⁺ memory T cells, low LAG3/TIGIT, and an immune-balanced state, defining a “proliferation-immune balanced type” with intermediate prognosis. C3 showed upregulated DNA repair and oxidative phosphorylation, but reduced epithelial-mesenchymal transition (EMT), angiogenesis, and KRAS signaling (Fig. 8F). The C3 subtype also had abundant activated CD8⁺ T and CD56dim NK cells and low IFNGR1/SIGLEC15, indicating an immune-active microenvironment. C3 was therefore defined as a “mitochondrial hypermetabolic-immune activated type” and was linked to better prognosis.

Fig. 8.

Redefining the three Gly/Ser/Thr metabolism-related subtypes in EC. (A,B) Expression patterns for the three prognostic genes across the metabolic subtypes are shown as bar plots (A) and heatmaps (B). (C) Distribution of risk scores among the three metabolic subtypes. (D–F) Bar plots showing hallmark pathway enrichment for C1 (D), C2 (E), and C3 (F) subtypes. ns, not significant, ** p 0.01, **** p 0.0001. C, cluster; MTHFD2, methylenetetrahydrofolate dehydrogenase 2; RPS6KA1, ribosomal protein S6 kinase A1; CDKN2A, cyclin-dependent kinase inhibitor 2A.

We next examined differential expression of the key genes (CDKN2A, MTHFD2, and RPS6KA1) in vitro by comparing their levels in EECs with those found in EC cell lines (HEC-1A and Ishikawa). Compared with EECs, EC cells showed significantly higher expression of CDKN2A and MTHFD2, but lower expression of RPS6KA1 (Fig. 9). These findings provide valuable insights for subsequent investigations.

Fig. 9.

Validation of the differential expression of key genes in vitro. (A) Differential expression of CDKN2A, MTHFD2, and RPS6KA1 between EECs and HEC-1A cells. (B) Differential expression of CDKN2A, MTHFD2, and RPS6KA1 between EECs and Ishikawa cells. Data are the mean $\pm$ SD, with n = 3. Statistical method: paired Student’s t-test. * p 0.05, ** p 0.01. EECs, endometrial epithelial cells; MTHFD2, methylenetetrahydrofolate dehydrogenase 2; RPS6KA1, ribosomal protein S6 kinase A1; CDKN2A, cyclin-dependent kinase inhibitor 2A.