Machine Learning Study Links ENTPD5 to Radioresistance in Liver Cancer
Published on 26 October 2026, this study combines analysis of tumor gene activity, single-cell analysis, machine learning, and a laboratory model of radiation exposure to study ENTPD5. The researchers focused on how copper-dependent cell death pathways may be connected to how cancer cells respond to radiation. They combined data from three separate groups of patients with hepatocellular carcinoma, a type of liver cancer: LIRI-JP, with 231 patients; GSE14520, with 225 patients; and TCGA-LIHC, with 325 patients. They used a method called consensus clustering and tested 127 different machine-learning setups to build a score that could predict patient outcomes.
A gradient boosting machine model performed consistently well. It separated patients into groups with different overall survival outcomes in all three datasets, suggesting that the results were not caused by something unique to just one dataset. However, ENTPD5 showed a complicated pattern. In the analyses of whole tumor samples, ENTPD5 activity was lower in the high-risk group and was negatively linked to the risk score. At first glance, this might suggest that ENTPD5 helps protect patients from poor outcomes.
But when the researchers examined 6,033 cancer cells using single-cell RNA sequencing data from GSE149614, ENTPD5 showed a weak positive relationship with the same risk score. The results changed again in the laboratory. In a model using MHCC97H liver cancer cells exposed to radiation, ENTPD5 activity increased in cells selected after radiation exposure. The log2 fold change was 0.316, and the p-value was 0.036. To investigate what ENTPD5 might do, the team used scTenifoldKnk, a computer method that simulates removing a gene from the regulatory networks of individual cells.
The simulated gene removal predicted a weaker unfolded protein response. This is a cell stress pathway that can help tumor cells survive the damage that ionizing radiation causes to proteins. The simulation also predicted lower activity in gene programs linked to resistance to radiation. These computer results suggest a possible explanation: ENTPD5 may help irradiated tumor cells cope with stress in the endoplasmic reticulum, a part of the cell that makes and processes proteins. Removing ENTPD5 might therefore make the cells more sensitive to radiation. However, the clinical findings have important limitations.
The researchers examined tissue collected after surgery from 42 patients whose outcomes could be assessed, using a method called immunohistochemistry to detect ENTPD5. Cancer returned in 11 of the 39 patients whose samples tested positive for ENTPD5. It returned in all 3 patients whose samples tested negative. These rates were 28.2 percent and 100 percent, respectively, with a Fisher's exact test p-value of 0.032. Patients with ENTPD5-positive samples also had longer progression-free survival, meaning they lived longer without their disease worsening. The log-rank test gave a p-value of 0.019. A Cox model that considered ENTPD5 alone gave a hazard ratio of 0.22, with a 95 percent confidence interval from 0.05 to 0.89.
The confidence interval is wide, and the researchers clearly explain the reason: only three patients were in the ENTPD5-negative group. A hazard ratio calculated from such a small group is unreliable and can change greatly with only a few cases. The Fisher's exact test result is statistically significant at the usual threshold, but it is also based on very few patients in the negative group. The study was retrospective, meaning the researchers used existing datasets and stored tissue samples rather than planning a study in advance to test whether ENTPD5 could predict treatment outcomes.
The authors are careful about what they conclude. They report that the combined risk score predicted patient outcomes in retrospective groups and that ENTPD5 showed different relationships depending on the type of sample or analysis. These included whole tumor samples, individual cancer-cell states, cells selected after radiation exposure, and tissue collected after surgery. The authors say these results make ENTPD5 a possible marker worth studying further. They do not prove that ENTPD5 directly causes radiation resistance or can predict which treatments will work.
This difference is important for doctors and patients. The study does not yet show that measuring ENTPD5 can help doctors decide how to use radiation treatment, or that targeting this enzyme can overcome radiation resistance in patients. Instead, the study provides an example of how researchers can combine several methods: gene analysis across multiple patient groups, detailed study of individual cells, computer simulations of gene removal, laboratory experiments with radiation exposure, and tests on patient tissue. It also shows how researchers can report conflicting results from different types of data without forcing them into one simple explanation.
Subject of Research: ENTPD5-related radiation resistance and the use of copper-dependent cell death-related information to build a model for predicting outcomes in hepatocellular carcinoma
Article Title: Deciphering ENTPD5-driven radioresistance in hepatocellular carcinoma via multi-algorithm machine learning and single-cell virtual knockout
Article References: Wu, Q.-Q., Zhang, X.-F., Zhang, Y., Wang, H.-Y., Wang, L.-F., Xu, Y., Chen, X.-L., Yi, X., Liu, K., Wu, Z.-F., & Zeng, Z.-C. (2026). Deciphering ENTPD5-driven radioresistance in hepatocellular carcinoma via multi-algorithm machine learning and single-cell virtual knockout. BMC Cancer. https://doi.org/10.1186/s12885-026-17093-w
Image Credits: AI Generated
DOI: 10.1186/s12885-026-17093-w
Keywords: hepatocellular carcinoma, resistance to radiation, ENTPD5, copper-dependent cell death, machine learning, single-cell RNA sequencing, virtual gene knockout, scTenifoldKnk, unfolded protein response, outcome prediction score, BMC Cancer, biomarker