Machine learning developed an intratumor heterogeneity signature for predicting prognosis and immunotherapy benefits in cholangiocarcinoma

Background: Cholangiocarcinoma is a kind of epithelial cell malignancy with high mortality. Intratumor heterogeneity (ITH) is involved in tumor progression, aggressiveness, treatment resistance, and disease recurrence. Methods: Integrative machine learning procedure including 10 methods (random survival forest, elastic network, Lasso, Ridge, stepwise Cox, CoxBoost, partial least squares regression for Cox, supervised principal components, generalized boosted regression modeling, and survival support vector machine) was performed to construct an ITH-related signature (IRS) for cholangiocarcinoma. Single cell analysis was performed to clarify the communication between immune cell subtypes. Cellular experiment was used to verify the biological function of hub gene. Results: The optimal prognostic IRS developed by Lasso method served as an independent risk factor and had a stable and powerful performance in predicting the overall survival rate in cholangiocarcinoma, with the AUC of 2-, 3-, and 4-year ROC curve being 0.955, 0.950 and 1.000 in TCGA cohort. low IRS score indicated with a lower tumor immune dysfunction and exclusion score, lower tumor microsatellite instability, lower immune escape score, lower MATH score, and higher mutation burden score in cholangiocarcinoma. Single cell analysis revealed a strong communication between fibroblasts, microphage and epithelial cells by specific ligand-receptor pairs, including COL4A1-(ITGAV+ITGB8) and COL1A2-(ITGAV+ITGB8). Down-regulation of BET1L inhibited the proliferation, migration and invasion as well as promoted apoptosis of cholangiocarcinoma cell. Conclusion: Integrative machine learning analysis was performed to construct a novel IRS in cholangiocarcinoma. This IRS acted as an indicator for predicting the prognosis and immunotherapy benefits of cholangiocarcinoma patients.