Deep learning facilitated discovery of prognosis biomarkers and their ligands to improve liver cancer treatment.

[BACKGROUND] Multi-kinase and immune checkpoint inhibitors have been developed for the treatment of hepatocellular carcinoma (HCC). However, the improvement in cancer survival remains limited due to their similar structures and targets. It is very interesting to identify novel prognosis biomarkers and corresponding ligands to improve patient survival.

[MATERIALS AND METHODS] Herein, we propose DLCP, a deep learning (DL)-based framework for improving Clinical Prognosis, to stratify patients with HCC, identify prognostic biomarkers, and recommend potential anti-HCC agents. Genomics, transcriptomics, epigenetics, and survival outcomes were integrated into a deep neural network to extract survival-related signatures, enabling stratification of patients accordingly. Prognostic biomarkers were identified by comparing molecular profiles across patient subgroups, and their ligands were screened from chemical compounds and natural products (NPs) using molecular docking and a machine learning-based predictive model. The protein-ligand interactions were validated by molecular dynamics simulations, surface plasmon resonance (SPR), and cellular thermal shift assay.

[RESULTS] The DLCP model stratified The Cancer Genome Atlas HCC patients into two subgroups with distinct survival outcomes, validated in the LIRI-JP cohort. Molecular traits of high- and low-risk patients aligned well with previous HCC findings. DLCP pinpointed RAC1 that exhibited significantly elevated expression in mice with primary liver cancer and patients with HCC and strong associations with cancer survival as a key biomarker. Through screening chemical small molecules, NPs, and NP derivatives from databases and literatures, a germacrane-guaiane dimer derivative (KGA-1083b) emerged as a promising anti-HCC agent. The complex reached equilibrium with low variations and achieved a K D value of 17.3 by SPR. RAC1 exhibited increased thermal stability in the presence of KGA-1083b, indicating direct binding and stabilization of the protein structure.

[CONCLUSIONS] DLCP provides a successful example of introducing DL into cancer prognosis by incorporating heterogeneous multi-omics and clinical phenotype, highlights the molecular mechanism of HCC progression, and accelerates the discovery of drug candidate molecules.