Singapore researchers have developed a machine-learning tool that predicts which patients with hepatocellular carcinoma (HCC) are most likely to experience cancer recurrence after surgery. The system combines genomic, clinical and laboratory data to assess biological features that are not captured by conventional tumour staging. In testing across independent patient groups, including a large publicly available Western dataset, the tool substantially outperformed the commonly used TNM staging system, which primarily measures tumour burden and anatomical spread.
The study was led by investigators from the National Cancer Centre Singapore (NCCS), Duke-NUS Medical School and the ASTAR Genome Institute of Singapore (ASTAR GIS) through the National Medical Research Council-funded PLANet programme, or Precision Medicine in Liver Cancer across an Asia-Pacific Network. The findings, published in Gut on 21 July 2026, also reveal that HCC can return through two biologically distinct routes. These insights could help clinicians identify patients who need closer surveillance, select individuals for additional treatment after surgery and design more focused clinical trials.
HCC is the most common form of primary liver cancer and remains a major cause of cancer mortality worldwide. It is the third leading cause of cancer-related death globally and disproportionately affects Asian populations, where approximately 72 per cent of cases occur. In Singapore, liver cancer is the third most common cause of cancer death among men and the fifth among women. The region’s disease burden is strongly linked to chronic hepatitis B virus infection, which can drive long-term inflammation, liver damage and malignant transformation.
Surgical removal of the tumour can offer patients a chance of long-term survival, particularly when the disease is detected at an earlier stage. However, recurrence remains common even after apparently successful resection. Around 70 to 80 per cent of recurrences occur within the liver, while the remainder involve metastasis to distant organs. Current clinical staging systems can estimate risk based on tumour size, number, vascular invasion and spread, but they provide limited information about the genetic diversity and evolutionary behaviour of individual cancers.
To investigate why some tumours return, the researchers analysed clinical information and performed comprehensive matched genetic studies on tumour samples collected through the PLANet cohort. The analysis included samples from recurrent tumours in a subset of patients, enabling the team to compare the original cancer with the disease that later re-emerged. Among 106 patients included in the study, 68, or 64.2 per cent, developed recurrence. Forty-eight patients experienced recurrence inside the liver, 11 developed disease elsewhere in the body and nine had both intrahepatic and distant recurrence.
The genetic comparisons identified two principal patterns of tumour spread. In the first, known as polyclonal seeding, several genetically distinct groups of cancer cells appear to leave the original tumour and establish new growths at the same time. This mechanism was more frequently associated with recurrence within the liver. The findings suggest that these tumours contain multiple malignant subpopulations, some of which may interact differently with the immune system and could be more responsive to selected immunotherapies.
The second pattern, monoclonal seeding, occurs when recurrence is driven primarily by a single cancer-cell clone from the original tumour. This route was more commonly linked to later recurrence and the spread of cancer beyond the liver. A surviving clone may possess biological properties that allow it to remain dormant, resist treatment or adapt to a new tissue environment before forming a detectable metastasis. Distinguishing between these mechanisms may therefore provide information about both the timing and likely location of recurrence.
Building on the genetic findings, the investigators created a multi-omics prediction model that integrates information from several biological layers. The tool considers clinical features such as tumour size and cancer stage, blood markers and genomic measurements, together with the activity of a 15-gene signature strongly associated with recurrence. Machine-learning methods allow these variables to be combined into a single risk estimate, potentially revealing patterns that would be difficult to recognise from any individual measurement.
The model was evaluated in three independent cohorts, including data from The Cancer Genome Atlas Liver Hepatocellular Carcinoma dataset. This validation was important because the TCGA group is largely Western, while the PLANet programme focuses on Asia-Pacific populations. Across the evaluations, the tool achieved an area under the receiver operating characteristic curve of approximately 86 per cent, compared with reported values of 56 to 68 per cent for TNM staging. A higher area under the curve indicates that a model is better able to distinguish patients at higher risk from those at lower risk, although further prospective testing will be needed before routine clinical use.
The researchers say the tool could eventually support more personalised surveillance and treatment after surgery. Patients predicted to be at high risk might be prioritised for intensive monitoring or clinical trials of adjuvant systemic therapies, while those at lower risk could avoid unnecessary treatment and its associated toxicities. The team is also using spatial sequencing to examine the tumour microenvironment and identify potential drug targets. Additional data, including CT imaging, may be incorporated into future versions of the model. Together, these efforts could improve understanding of how HCC evolves and help match patients with therapies designed for the specific biology of their disease.
Subject of Research: Human tissue samples
Article Title: Clonal diversity underpins distinct modes of recurrence in hepatocellular carcinoma: the PLANet cohort study
News Publication Date: 3 August 2026
Web References: https://doi.org/10.1136/gutjnl-2026-338227; https://www.nccs.com.sg; https://www.a-star.edu.sg/gis; https://www.duke-nus.edu.sg/
References: National Registry of Diseases Office. (2026). Singapore Cancer Registry Annual Report 2023. Ministry of Health Singapore. Chan S, Sun H, Xu Y et al. “The Lancet Commission on addressing the global hepatocellular carcinoma burden: comprehensive strategies from prevention to treatment.” The Lancet. 2025;406:731–778.
Keywords: hepatocellular carcinoma, liver cancer, cancer recurrence, machine learning, multi-omics, genomics, precision medicine, tumour evolution, polyclonal seeding, monoclonal seeding, hepatitis B virus, PLANet cohort
Tags: advanced liver cancer staging toolsAI-driven cancer prognosis modelsbiological pathways of liver cancer recurrencegenomic and clinical data for cancer prognosishepatocellular carcinoma recurrence predictionimproving liver cancer clinical trialsliver cancer recurrence predictionliver cancer recurrence risk assessmentliver cancer surgical outcomesmachine learning in liver cancerpersonalized liver cancer treatmentSingapore liver cancer research
