Cancer heterogeneity represents a major hurdle in the development of effective theranostic strategies, as it prevents to devise unique and maximally efficient diagnostic, prognostic and therapeutic procedures even for patients affected by the same tumor type. Computational techniques can nowadays leverage the huge and ever increasing amount of (epi)genomic data to tackle this problem, therefore providing new and valuable instruments for decision support to biologists and pathologists, in the broad sphere of precision medicine. In this context, we here introduce a novel cancer subtype classifier from gene expression data and we apply it to two different Breast Cancer datasets, from TCGA and GEO repositories. The classifier is based on Support Vector Machines and relies on the information about the relevant pathways involved in breast cancer development to reduce the huge variable space. Among the main results, we show that the classifier accuracy is preserved at excellent values even when the variable space is reduced by a 20-fold, hence providing a precious tool for cancer patient profiling even in case of limited experimental resources.