Patient-specific metabolic fluxes reflect molecular subtype and heterogeneity in breast cancer

Tumor metabolism is dynamic and context-dependent, yet gene expression provides only an indirect view of metabolic activity. We combined patient-specific genome-scale metabolic models with supervised and unsupervised machine learning to infer metabolic fluxes for 1,223 TCGA breast tumor and normal samples under a uniform physiological medium. Flux profiles distinguished tumor from normal tissue under patient-grouped nested cross-validation (ROC-AUC up to 0.989) and predicted receptor status and molecular subtype (e.g., Basal-like versus other tumors, AUC = 0.966), but not stage, nodal status or age. Within tumors, metabolic variation was continuous rather than organized into discrete groups and aligned only partially with clinical classifications. These findings support a view of breast cancer metabolism as a structured but continuous landscape that reflects molecular subtype and receptor status, and provide a reproducible framework to characterize tumor heterogeneity.