Prognostic Classification of Hepatocellular Carcinoma Under Incomplete Data Conditions Using a Public Reference Cohort: A Weakly Supervised Multi-Omics Study

Abstract

Background: Prognostic stratification of hepatocellular carcinoma (HCC) remains difficult because the disease is highly heterogeneous and complete matched multi-omics data are not always available in clinical cohorts.  Objective: To develop a weakly supervised multi-omics framework that derives prognostic subtype labels from a public reference cohort and transfers them to cohorts with incomplete data. Methods: This study has analyzed 363 patients in TCGA-LIHC with matched mRNA, miRNA, DNA methylation, and clinical data. Overall-survival-related features were selected by univariate Cox regression, integrated by similarity network fusion (SNF), and clustered by spectral clustering to generate pseudo-labels. TCGA-LIHC was then split 6:4 into training and test sets for supervised modeling. External validation used LIRI-JP, GSE14520, GSE54236, and GSE31384, with models rebuilt on features shared with each cohort. Prognostic performance was evaluated by Kaplan-Meier analysis, log-rank testing, and the concordance index (C-index). Results: A total of 3,890 mRNA features, 150 miRNA features, and 1,889 methylation features were retained. SNF plus spectral clustering identified two subtypes: S1 (n = 257) and S2 (n = 106). S2 had significantly worse overall survival than S1 (log-rank P = 3.891 × 10^-9; C-index = 0.866). In internal validation, XGBoost showed the highest AUC (0.983). In external validation, the predicted subtypes remained prognostically informative, with C-index values of 0.857 in LIRI-JP, 0.875 in GSE14520, 0.930 in GSE54236, and 0.883 in GSE31384. Conclusions: In the public datasets included in this study, this weakly supervised framework identified two prognostically distinct HCC subtypes and retained prognostic discrimination after transfer to external cohorts with incomplete data.