Integrated multi-omics and machine learning analyses to improve and explore molecular subtypes and immune profiles of primary hepatocellular carcinoma

Hepatocellular carcinoma (HCC) is a highly heterogeneous malignancy with poor prognosis and limited treatment options. A comprehensive understanding of its molecular subtypes and immune context is urgently needed to guide precision therapy. We performed an integrative analysis of transcriptomic, epigenetic, and genomic data from TCGA-LIHC and 10 external HCC cohorts. Using MOVICS-based consensus clustering and a predefined multi-algorithm survival-learning framework, we identified molecular subtypes and constructed a prognostic model (Consensus Machine Learning Signature, CMLS) based on 14 survival-associated genes. Subtype validation and immune characterization were conducted using NTP/PAM, GSVA/ssGSEA, ESTIMATE, RTN, and external immunotherapy cohorts. Two molecular subtypes (CS1 and CS2) with distinct prognostic and immune profiles were defined. CS1 exhibited an aggressive, immunosuppressive landscape characterized by T-cell exhaustion and high regulatory T-cell activity, whereas CS2 featured a more immunologically active microenvironment with enriched NK cells and M2 macrophages. Furthermore, the CMLS model effectively stratified patients into high- and low-risk groups with significant differences in survival across cohorts ( p < 0.001), with an average C-index of 0.751 across the TCGA training cohort and the two external validation cohorts. High-risk groups represented “cold” tumors with limited immune surveillance and high infiltration of myeloid-derived suppressor cells (MDSCs), while low-risk groups displayed “hot” tumor characteristics with robust lymphocyte infiltration and heightened immunogenicity. Notably , CDC20 was identified as a key gene with strong prognostic value and involvement in tumor progression. This study proposes a multi-omics-based molecular classification and prognostic scoring system for HCC. The CMLS model showed reproducible prognostic performance across retrospective cohorts and may provide a useful computational framework for risk stratification and hypothesis generation for future therapeutic studies.