A Non-Invasive Serum Metabolite-Based Machine Learning Model Predicts Response to Neoadjuvant Immunotherapy in Mismatch Repair-Deficient Colorectal Cancer

Colorectal cancer (CRC) with microsatellite instability-high (MSI-H) or mismatch repair-deficient (dMMR) shows significant sensitivity to immune checkpoint inhibitors (ICIs). However, a considerable proportion of patients still exhibit primary or acquired resistance to ICIs. Until now, efficient and non-invasive biomarkers for accurately predicting immunotherapy efficacy remain unavailable. In this multicentre study, liquid chromatography–mass spectrometry (LC–MS) and enzyme-linked immunosorbent assay (ELISA) were applied to identify and validate serum metabolites linked to immunotherapy response. Using machine learning algorithms, a random forest predictive model (5-MPM) was constructed based on five metabolites-PGE2, tryptophan, arginine, citrulline, and histidine. The 5-MPM model demonstrated robust predictive performance in both training and independent validation cohorts, with AUC values of 0.85 and 0.88, respectively. The SHAP analysis elucidated the contribution of each metabolite to model predictions. Integrating above five metabolites with metastasis stage did not further improve the predictive performance of this model. This study provides the first systematic characterization of metabolic reprogramming in dMMR colorectal cancer with different response to immunotherapy, and establishes a non-invasive, high-precision predictive tool that offers a new basis for individualized therapeutic decision-making.