Breast cancer (BC), the most prevalent cancer type in women worldwide, exhibits significant heterogeneity across individual cases. The existing subtyping schemes for BC are all based on the expression of a few marker genes or proteins, which could not well capture samples sharing common or similar biology, which has substantially limited their applications.1 We present a new classification scheme for BC samples based on predominantly co-expression patterns of metabolic genes at the genome scale using a suite of machine-learning methods, resulting in four classes. Our consideration is: i) Extensive metabolic reprogramming is observed in cancer and considered as a hallmark of cancer2; and ii) all abnormal behaviors exhibited by cancer tissue cells are the direct results of these reprogrammed metabolisms, which are induced to adapt to specific stressors in the cancer-promoting microenvironments.3 In addition, the classification of BC tissues based on genome-scale metabolic genes expressions can ensure the stability of the classification results. Compared with the current PAM50-based BC subtypes, samples in each of our four classes have more outstanding shared metabolic characteristics, clinical features, genomic alterations, tumor microenvironments, immune-cell infiltrations, immunotherapy responses, and chemotherapy sensitivities, making our classification results more informative and potentially more useful.