Driver gene-based prognostic model benchmarked against machine learning survival models is associated with a CAF–TGFβ–EMT-like phenotype in hepatocellular carcinoma

Abstract

Hepatocellular carcinoma (HCC) is molecularly heterogeneous, limiting conventional prognostic tools. Whole-genome resequencing of six HCC cases was used only to filter candidates. Somatically mutated genes were intersected with HCC-annotated entries from five predefined driver resources, yielding 154 genes. DGPM-HCC, a 13-gene model summarized by TotalScore, was developed in 170 bulk transcriptome samples. In 20 repetitions of event-stratified fivefold cross-validation, DGPM-HCC achieved a mean overall-survival C-index of 0.761; versus continuous Cox13, the median paired ΔC-index was 0.125 (bootstrap 95% confidence interval, 0.111–0.142; P = 9.57 × 10 −5 ). The combined nomogram achieved a training-cohort 3-year disease-free-survival AUC of 0.924. External discrimination in TCGA-LIHC was modest (C-index, 0.533; 1-, 3-, and 5-year AUCs, 0.503, 0.592, and 0.590), whereas GSE14520 lacked three model genes. Exploratory single-cell RNA sequencing of one tumor–adjacent pair, public single-cell integration, and multiplex immunohistochemistry on 16 specimens associated high TotalScore with a CAF–TGFβ–EMT-like state. Drug-sensitivity findings were computational and exploratory.