Integrative machine learning and network toxicology identifies EZH2, MMP9, and PPIA as glyphosate-associated hub genes in uterine corpus endometrial carcinoma

Endometrial cancer, particularly uterine corpus endometrial carcinoma (UCEC), is a common malignancy with substantial molecular and clinical heterogeneity. Glyphosate (GLP), a widely used herbicide, has been linked to oxidative stress, endocrine disruption, immune modulation, and carcinogenic processes; however, its molecular relationship with UCEC remains poorly understood. We developed an integrative systems biology framework combining GLP target prediction (from ChEMBL, STITCH, and SwissTargetPrediction) with transcriptomic profiling from the GEO dataset GSE17025 and machine learning algorithms (LASSO, SVM-RFE, and Random Forest) to identify GLP-related hub genes in UCEC. GSE17025 comprised 103 samples (91 UCEC tumors and 12 normal endometrial controls) and served as the training cohort for differential expression analysis and diagnostic model construction. UCEC-associated genes were retrieved from GeneCards (top 2,000 by relevance score) and OMIM (all 170 genes), then merged and deduplicated. Protein-protein interaction (PPI) networks and CytoHubba algorithms were used for network‑based screening. Predictive performance was assessed using SHAP analysis, receiver operating characteristic (ROC) curves, decision curve analysis, calibration assessment, and nomogram construction, with external validation in the TCGA-UCEC cohort (552 tumors, 35 normal samples). Importantly, because our analysis is based on public transcriptomic data comparing tumor versus normal tissues rather than GLP‑exposed versus unexposed samples, all identified associations should be interpreted as correlative and hypothesis‑generating, not as evidence of causation. A total of 350 potential GLP targets and 4,253 differentially expressed genes (DEGs) were identified from GSE17025. Overlap between GLP targets and DEGs yielded 78 GLP‑related DEGs; further intersection with UCEC‑associated genes produced 24 candidate genes. PPI network analysis and CytoHubba screening identified 11 hub genes, and machine learning consistently selected EZH2, MMP9, and PPIA as the final machine learning‑derived hub genes. The three‑gene signature showed strong diagnostic performance in both the training set (combined model AUC = 0.995) and the TCGA‑UCEC validation cohort (combined model AUC = 0.964; individual gene AUCs: EZH2 = 0.964, MMP9 = 0.909, PPIA = 0.931). The optimism-corrected bootstrap AUC was 0.975, and the nested cross-validation AUC was 0.978, suggesting robust model performance. SHAP analysis ranked MMP9 as the top contributor, followed by PPIA and EZH2. Immune infiltration analysis revealed significant correlations between these genes and multiple immune cell subsets. Molecular docking computationally predicted potential interactions of GLP with EZH2, MMP9, and PPIA, with the most favorable binding energy for PPIA (-6.5 kcal/mol). These computational findings are preliminary and warrant experimental validation. The computational intersection of GLP targets with UCEC‑associated transcriptional changes may be associated with EZH2, MMP9, and PPIA‑mediated transcriptional regulation, immune microenvironment remodeling, and potential chemical‑protein interactions. Our results provide candidate biomarkers and a computational framework for investigating environmental chemical‑linked molecular mechanisms in endometrial cancer, supporting future experimental studies. However, causal inference and biological functionality require rigorous in vitro and in vivo validation.