Résumé
Abstract Background: Long-term treatment-related toxicities, such as neurological and metabolic toxicities, are major issues in breast cancer. We investigated the interest of metabolomic profiling to predict toxicities. Methods: Untargeted high-resolution metabolomic profiles of 992 patients with ER+/HER2- breast cancer from the prospective CANTO cohort were acquired (n=1935 metabolites). A residual-based modeling strategy with a discovery and validation cohort was used to benchmark machine learning algorithms, taking into account confounding variables. Results: Adaptive LASSO has a good predictive performance, has limited optimism bias, and allows the selection of metabolites of interest for future translational research. The addition of low-frequency metabolites and non-annotated metabolites increases the predictive power. Metabolomics adds extra performance to clinical variables to predict various neurological and metabolic toxicity profiles. Conclusions: Untargeted high-resolution metabolomics allows better toxicity prediction by considering environmental exposure, metabolites linked to microbiota, and low-frequency metabolites.