Machine Learning Tool Improves Prediction of Liver Cancer Recurrence

The burden is particularly high in Asia, and clinicians still lack reliable biomarkers to stratify recurrence risk. Most relapses occur within the liver, with a smaller proportion presenting as distant metastasis. A new study shows that a machine learning-based multi-omics tool can more accurately predict post-resection recurrence and delineate distinct biological routes of relapse.

Researchers at the National Cancer Center Singapore (NCCS), Duke-NUS Medical School, and the A*STAR Genome Institute of Singapore (A*STAR GIS) developed a machine learning-based multi-omics tool to estimate an individual patient’s risk of HCC recurrence after surgical resection. The model combines multiple biological and clinical indicators, including tumor size and cancer stage, with genetic information and blood-based biomarkers. It also incorporates a 15-gene panel strongly linked to recurrence to refine risk estimates, integrating tumor burden with molecular features not captured by conventional staging.

Using clinical information and comprehensive genetic analyses from the Precision Medicine in Liver Cancer across an Asia-Pacific Network (PLANet) cohort, the team also defined two distinct recurrence pathways: polyclonal seeding and monoclonal seeding. Polyclonal seeding, in which multiple groups of cancer cells disseminate from the primary tumor, was more often associated with intrahepatic relapse and displayed characteristics suggesting greater susceptibility to certain immunotherapies. Monoclonal seeding, in which recurrence arises from a single clone, was more commonly linked to later relapse and spread beyond the liver. Among 106 patients studied, 68 (64.2%) relapsed: 48 within the liver, 11 at distant sites, and nine at both.

The prediction tool was validated across three independent cohorts, including The Cancer Genome Atlas (TCGA) Liver Hepatocellular Carcinoma dataset. Successful validation in TCGA, a largely Western cohort complementing the PLANet Asia-Pacific population, supported reproducibility and broader generalizability across populations. Performance measured by area under the curve (AUC) reached 86%, markedly exceeding the 56%–68% achieved by tumor-node-metastasis (TNM) staging. Overall, the model outperformed TNM for resected liver cancer, which does not take cancer genetics into account.

Published in Gut, the findings indicate that improved risk stratification could help tailor follow-up, identify patients likely to benefit from additional therapy after resection, and inform more targeted clinical trial design. Building on this work, the team is applying spatial sequencing to identify tumor microenvironment biomarkers that could serve as drug targets for HCC. They are also working to enhance the tool by adding computed tomography (CT) imaging and additional machine learning inputs to further improve accuracy and enable focused surveillance of high-risk patients.

“By uncovering two distinct mechanisms of liver cancer recurrence, we have gained new insights into the biological processes that drive disease progression after surgery. Understanding these differences is an important step toward more personalized approaches to risk prediction, and the development of more precise biomarkers and targeted therapies for HCC patients,” said Dr. Zhang Ying, co-first author of the study and senior scientist at A*STAR GIS.

Related Links
National Cancer Center Singapore 
Duke-NUS Medical School
A*STAR GIS