Breast cancer is the cancer with the highest incidence in women worldwide, and is the leading cause of cancer-related death, mainly due to treatment resistance. Recently, tumor heterogeneity has been described as one of the key driver in treatment failure. Indeed, tumor is not a homogeneous entity to treat, but a complex association of subclonal populations driven by their own genetic alterations, and immune and stromal cells from microenvironment. Breast cancer subtypes and tumor heterogeneity advocate for the development of tailored, personalized treatments, but so far, the discovery of efficient predictive markers has been compromised by the lack of adapted biological models and methodological tools. The recent developments of high-throughput methods for bulk and single-cell analyses has generated large ‘omics’ datasets from patients, stored in open access databases (ArrayExpress, GEO). Combining these numerous datasets will grant a sufficient statistical power to reveal a comprehensive overview of tumor complexity. However, this data mining is currently limited by methodological challenges like cross-platform normalization and the difficulty to analyze complex data structure with high dimension observations. To overcome these issues, I propose to implement a multidisciplinary project at the interface between mathematics, biology, and information technologies. With the support of the mathematicians and bioinformaticians from the Bioinfomics unit of the regional comprehensive cancer center (ICO), I will develop and implement machine-learning algorithms in the search of predictive biomarkers for breast cancer treatment. This innovative strategy will lead to personalized medicine in breast cancer by guiding clinicians in the selection of the optimal therapeutic option. Moreover, this generated pipeline for predictive marker discovery could be further adapted for the treatment of other cancer types.