Machine learning-driven prediction of immune checkpoint inhibitor responses against cholangiocarcinoma: a bile biopsy perspective.

BACKGROUND: The treatment of cholangiocarcinoma (CCA) continues to face numerous clinical challenges, including the prediction of sensitivity to immunotherapy and the development of preoperative diagnostic models. METHODS: In this study, we aimed to address these challenges by collecting bile samples from CCA patients for metabolomic and microbiomic analyses. We also performed immunofluorescence (IF) staining on tissue formalin-fixed, paraffin-embedded (FFPE) blocks to assess the expression of relevant biomarkers. Additionally, we followed up with patients to analyze prognostic indicators based on their survival times. Using advanced machine learning techniques, specifically LASSO regression, we constructed a predictive model to determine the effectiveness of programmed cell death protein 1 (PD-1) inhibitors in treating CCA. The model integrates bile metabolomic data with an Immune Hot-Cold Index (IHC Index) derived from IF results, providing a comprehensive metric of the patient's immune environment. RESULTS: Our findings revealed significant differences in metabolomic profiles between CCA patients and those with non-malignant liver diseases, as well as between patients with different genetic mutations. The IHC Index successfully differentiated between immune "hot" and "cold" states, correlating strongly with patient responses to immunotherapy. Furthermore, in one CCA patient, the model's predictions were validated, demonstrating high accuracy and clinical relevance. CONCLUSION: Our predictive model offers a robust tool for assessing the sensitivity of CCA patients to PD-1 inhibitors, potentially guiding personalized treatment strategies. Additionally, the integration of bile metabolomics with IF data provides a promising approach for developing preoperative diagnostic models, enhancing early detection and treatment planning for CCA.