Deciphering ENTPD5-driven radioresistance in hepatocellular carcinoma via multi-algorithm machine learning and single-cell virtual knockout

Abstract Background Radioresistance limits durable local control in hepatocellular carcinoma (HCC), but the relationship between cuproptosis-related metabolism and radiation response remains incompletely characterized. Methods We integrated 26 cuproptosis- and radiation-related genes across LIRI-JP ( n = 231), GSE14520 ( n = 225), and TCGA-LIHC ( n = 325). Consensus clustering and 127 machine-learning configurations were used to develop and validate a prognostic score. GSE149614 single-cell data, scTenifoldKnk virtual knockout, an irradiated MHCC97H model, and postoperative immunohistochemistry were used for biological evaluation. Results A gradient boosting machine score consistently stratified overall survival across the three cohorts. In bulk cohorts, ENTPD5 expression was lower in the high-risk group and inversely correlated with the risk score. By contrast, ENTPD5 showed a weak positive association with the score in 6,033 malignant cells and was upregulated in radiation-selected MHCC97H cells ( p = 0.036; log2 fold change = 0.316). Virtual knockout predicted attenuation of unfolded-protein-response and radioresistance-associated programs. Among 42 patients with evaluable clinical outcomes, recurrence occurred in 11/39 ENTPD5-positive and 3/3 ENTPD5-negative cases (28.2% vs. 100%; Fisher’s exact p = 0.032). ENTPD5-positive cases had longer progression-free survival (log-rank p = 0.019; univariable Cox hazard ratio = 0.22, 95% confidence interval 0.05–0.89), although the ENTPD5-negative group was small. Conclusions The integrated score is prognostic in retrospective HCC cohorts, and ENTPD5 shows context-dependent associations across bulk tumors, malignant-cell states, radiation-selected cells, and postoperative tissues. These findings support ENTPD5 as a candidate marker for further study rather than establish a causal or treatment-predictive role.