Fig. 1. Study summary.

All protein-coding genes were examined for their potency as prognosis-related genes with the utility of the data from the TCGA-LIHC cohort and five international multicenter datasets (Step 1). The number of validated genes fell off to 23 with the random forest machine-learning approach (Step 2). A universal prognostic score, appointed as mHPS, was established through a neural network approach (Step 3). Finally, the usage of mHPS was validated in different circumstances (Step 4). LIHC, Liver Cancer Cohort; TCGA, The Cancer Genome Atlas; mHPS, molecular hepatocellular carcinoma prognostic score.