Identification of Key Genes via Integrated Multi-Omics and Machine Learning Uncovers Tumor Biological Features and Prognostic Biomarkers in Uterine Leiomyosarcoma.

Abstract

Background: Uterine leiomyosarcoma (ULMS) is a rare, aggressive uterine malignancy with high misdiagnosis rates, poor prognosis, and limited molecular biomarkers. Its pathogenesis, links between specific genes and the tumor immune microenvironment (TIME), and applications of machine learning (ML) and Mendelian randomization (MR) remain understudied.

Methods: Multi-cohort data (4 GEO datasets, TCGA-SARC, single-cell sequencing) were integrated. Differentially expressed genes (DEGs) and WGCNA-derived key modules identified "InteGenes". 113 ML algorithms were compared to build a diagnostic model (top: GBM, core genes = "Mgenes"). CIBERSORT analyzed TIME; MR explored Mgenes-ULMS causal links.

Results: 96 InteGenes enriched in cell cycle/p53/DNA repair pathways. The GBM model had training AUC = 1 and validation accuracy 92.3-100%; 36 Mgenes (e.g., TRIP13, AUC = 0.972) showed diagnostic value. Mgenes correlated with TIME (upregulated Mgenes ↔ M2 TAMs/Tregs; downregulated ↔ effector cells). MR found no genetic causality between Mgenes and ULMS.

Conclusion: InteGenes reflect ULMS pathogenesis; the GBM model and Mgenes are promising diagnostic tools. Mgenes modulate ULMS's TIME, offering immunotherapeutic targets. This study advances ULMS molecular/immune understanding for translational research.