Abstract
Hepatocellular carcinoma (HCC) is highly heterogeneous, making prognosis and treatment prediction challenging. Using multi-omics data from multiple HCC cohorts, we identified five biomarkers (AKR1B10, ANXA2, COL15A1, SPARCL1, and SPINK1) and developed dual serum and tissue signatures by machine learning. The tissue mRNA signature could stratify prognostic risk and reflect alterations in the tumor's genome, metabolism, and immune microenvironment. High-risk HCC responded poorly to sorafenib and transarterial chemoembolization (TACE) but sensitively to agent ABT-263 in silico, in vitro, and in vivo experiments. The serum protein signature outperformed the clinical tumor staging systems in predicting 24-month disease-free survival, with median time-dependent areas under the receiver operating characteristic curve (AUC(t)) of 0.79 and 0.75 in two postoperative cohorts, and the AUC was 0.90 for predicting treatment benefit in a TACE-treated cohort. Interpretability analysis revealed consistent biomarker contributions in both signatures. Conclusively, the dual signatures show promise for HCC risk stratification, pending external validation.