ML-BUSMetab: Machine Learning-Based Metabolomic Profiling for Predicting Aspirin Response in Colorectal Cancer Chemoprevention: A Multi-Model Explainable Artificial Intelligence Approach with External Validation

Background/Objectives: Aspirin-based colorectal cancer (CRC) chemoprevention remains a promising yet individually variable strategy. As a proof-of-concept toward future personalized chemoprevention frameworks, we aimed to develop and validate machine learning (ML) models capable of distinguishing aspirin-exposed from placebo-exposed participants based on their plasma metabolomic signatures, thereby characterizing the metabolomic footprint of aspirin administration rather than directly predicting clinical chemoprevention benefit. Methods: Training was performed on the Aspirin/Folate Polyp Prevention Study (AFPPS) dataset ST001422 (n = 300) and external validation on ST001423 (n = 223). After multi-method consensus feature selection, reducing 19,433 features to 300, sixteen ML and deep learning (DL) architectures were benchmarked under nested cross-validation. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) analyses. Results: GBM_sklearn achieved the highest cross-validation Precision-Recall AUC (PR-AUC) of 0.945, while ensemble stacking (Stack_LGB) offered superior calibration (Brier = 0.117). DL models consistently underperformed traditional ML (PR-AUC: 0.673-0.843 vs. 0.881-0.945), attributable to limited sample size. SHAP and LIME analyses independently identified m/z 196.0604 (C18, RT 89.4 s) as the top metabolic biomarker, consistent with aspirin-induced glycerophospholipid pathway alterations. External validation performance degraded substantially (PR-AUC: 0.945 → 0.711), attributable to inter-study analytical batch effects. Conclusions: This framework demonstrates the feasibility of metabolomics-driven personalized chemoprevention. Although the high feature-to-sample ratio (300:300) and the substantial drop between internal and external performance indicate that the cross-validation estimates likely include dataset-specific noise in addition to the true biological signal. While highlighting batch harmonization and aggressive feature reduction (e.g., LASSO/RFE-based selection of 10-20 high-impact metabolites) as a prerequisite for clinical translation.