Bayesian multimodal modeling with explainable artificial intelligence to predict response to GBE1-targeted metabolic therapy in colorectal cancer.

9

Background: Colorectal cancer (CRC) remains a leading cause of cancer-related mortality worldwide, with therapeutic resistance limiting durable clinical benefit. Metabolic reprogramming is an important yet incompletely understood driver of tumor progression, enabling CRC cells to adapt to metabolic stress and shape an immunosuppressive tumor microenvironment. Our prior work identified glycogen metabolism, including GBE1-driven pathways, as a key regulator of CRC metabolic and treatment response. We integrate multimodal data with explainable artificial intelligence to model metabolic heterogeneity and improve prediction of therapeutic effectiveness while preserving biological interpretability. Methods: Multimodal datasets were integrated, including genomics, transcriptomics, CMS subtypes, histopathology-derived features, and clinical outcomes from TCGA-COAD, CRC-SG, AACR BPC CRC v2.0, and institutional cohorts. Genotyping data were processed through the variant-calling pipeline, and pseudo-bulk expression profiles were performed and visualized across 629 available samples. Clinical and molecular variables were harmonized to construct a framework for estimation. NetraAI, an explainable artificial intelligence platform employing an iterative machine learning strategy, was applied to identify treatment-responsive patient subpopulations. Predictive modeling combined ensemble learning and penalized regression approaches to generate interpretable feature attribution metrics (SHAP values). Model outputs were further incorporated into a Bayesian inference framework to enable probabilistic outcome prediction while accounting for prior data. Model performance was evaluated using stratified cross-validation, ROC analysis, and Bayesian credible intervals. Results: Multimodal modeling identified six key features associated with response to GBE1 inhibition, including glycogen content, GBE1 expression, CMS subtype, and lipid metabolic gene signatures. SHAP analyses ranked glycogen content (6.2), GBE1 expression (3.92), and CMS4 subtype (0.35) as dominant predictors. Bayesian integration improved predictive performance compared by validation cohorts (AUC 0.81 vs 0.61). Explainability analyses highlighted glycogen and lipid metabolism, hypoxia, and immune features as key drivers of response. Synthetic multimodal data preserved biologically plausible relationships and enhanced model generalizability across molecular subtypes and treatment contexts. Conclusions: Integration of multimodal CRC data with explainable AI enables accurate and interpretable prediction of response to GBE1-targeted metabolic therapy. This framework supports the high dimensional analysis and metabolic reprogramming as a clinically actionable vulnerability and advancing precision metabolic therapy in CRC.