Survival Reinforced Transfer Learning for Multicentric Proteomic Subtyping and Biomarker Discovery

Omics-based molecular subtyping in large-scale and multicentric cohort studies is a prerequisite for proteomics-driven precision medicine (PDPM). However, keeping the subtypes with robust molecular features and significant associations with prognosis among different cohorts is challenging due to the biological heterogeneity and technical inconsistency. Herein, we propose a subtyping algorithm, named Survival Reinforced Patient Stratification (SRPS), to adapt the known subtypes from the discovery cohort to another by simultaneously preserving the distinct prognosis and molecular characteristics of each subtype. SRPS has been benchmarked on simulated and real-world datasets, where it shows a 12% increase in classification accuracy and possesses the best prognostic discriminations. Moreover, based on the calculated subtype significance score, an unpopular protein, Peptidylprolyl Isomerase C (PPIC), was identified as the top-1 remarkable protein for subtyping the hepatocellular carcinoma (HCC) patients with the worst prognosis. Eventually, PPIC was experimentally proved to be a pro-cancer protein in HCC, confirming our work as a practice of interpretable machine learning guided biological discovery in PDPM research.