Dimensional reduction approaches are crucial to extract relevant information from high-dimensional datasets, such as clinical or omics data. We have developed a toolbox based onnon-negative matrix factorization, in order to extract relevant biological and clinicalsignatures from high-dimensional datasets and combine various modalities into molecularsignatures. I will present some recent applications of these approaches, fromneuroblastoma epigenomics to the study of comorbidities between mental and somaticdisorders and the application to transfer learning from single-cell to bulk datasets.