Molecular subtyping and therapeutic prediction of hepatocellular carcinoma based on exosome proteomics

Precision medicine increasingly relies on molecular subtyping to guide individualized therapy for hepatocellular carcinoma (HCC), yet conventional tissue biopsies have limited capacity to capture tumor heterogeneity and temporal changes. This study utilized patient-derived xenograft (PDX) models to isolate authentic tumor-derived exosomal proteins, which were profiled to classify HCC patients into molecular subtypes. Differential expression analysis combined with survival association screening and a bootstrap strategy was used to identify robust prognostic candidates, followed by Boruta feature selection to refine the list to 14 key proteins. These were integrated into a machine learning model (TSExs) to predict patient survival and treatment response. Immune-related pathway enrichment was performed to characterize the immune microenvironment of each subtype. Two robust subtypes, S1 and S2, were identified. S1 tumors demonstrated heightened antigen presentation and T-cell activation but impaired immune clearance, whereas S2 tumors exhibited lower immune infiltration yet more efficient immune-mediated tumor elimination. The TSExs model achieved high prognostic accuracy and reliably predicted responses to chemotherapy and immunotherapy across independent cohorts. By integrating exosomal proteomics with machine learning, this study establishes a non-invasive, biologically informed framework for immune profiling and therapeutic prediction in HCC, supporting the clinical potential of exosomal proteins as precision oncology biomarkers.