Constructing a bladder cancer prognostic model related to exosome using machine learning and identifying THBS1 as a potential target.

BACKGROUND: Exosome-mediated molecular processes significantly influence bladder cancer (BCa) development, but clinically applicable exosome-based prognostic systems are still lacking. This research aims to develop an exosome-related prognostic model using machine learning approaches and discover potential therapeutic targets. METHODS: We implemented ten machine learning algorithms with eighty-one combinatorial configurations to analyze BCa transcriptomic data. Model validation incorporated time-dependent receiver operating characteristic analysis, Kaplan-Meier survival curves, nomogram assessment, and Cox regression. Biological mechanisms were explored through immune microenvironment evaluation (CIBERSORT) and functional enrichment analysis (GSEA). Molecular docking using PubChem, PDB structures, and CB-DOCK2 identified potential targets, followed by experimental validation including quantitative reverse transcription PCR (qRT-PCR), Western Blotting (WB), cell proliferation (CCK-8, colony formation), migration (Transwell, wound healing), EMT marker detection, and tail vein injection-based lung metastasis assays in nude mice. RESULTS: Analysis revealed 132 differentially expressed genes specific to BCa, which were subsequently refined to 15 prognosis-associated genes through univariate Cox regression. The machine learning-derived prognostic model (MLDPM) outperformed existing clinical indicators, effectively stratifying patients into distinct risk categories. High-risk patients demonstrated characteristics of immune evasion and poorer survival outcomes. Computational docking analysis identified strong molecular interactions between THBS1/MMP9/CXCL12 and standard chemotherapeutic compounds. Importantly, suppression of THBS1 expression significantly inhibited BCa cell migration and invasion. CONCLUSION: This study establishes the first machine learning-based exosomal prognostic system for BCa and identifies THBS1 as both a potential biomarker and therapeutic target. The combined computational and experimental methodology offers a new approach for personalized BCa treatment strategies.