A0344
Title: A supervised NMF extension for integrating omics data
Authors: Aurelie Mercadie - INRAE (France) [presenting]
Eleonore Gravier - Laboratoires Pierre Fabre (France)
Gwendal Josse - Laboratoires Pierre Fabre (France)
Nathalie Vialaneix - MIAT INRAE (France)
Celine Brouard - MIAT INRAE (France)
Abstract: The motivation is to tackle a frequent problem in clinical research: patients stratified into K groups of interest (typically healthy/sick or control/treated patients) are described by biological measurements corresponding to different omics (metabolomics, proteomics, etc.). The aim is then to discover molecular signatures characterizing the groups. A non-negative matrix factorization (NMF) approach is extended to this framework. More specifically, this proposal is related to a supervised NMF, the FR-lda, which is based on an objective function that includes a supervised term to explain the groups of individuals. The proposal adapts this method to a multi-table framework by integrating information through a contribution matrix common to all omics. Compared to FR-lda, the supervised term is reworked to account for the non-negativity of the solution and resumes a criterion equivalent to K independent linear regressions, one for each group. Finally, two optimization methods are proposed to solve the induced optimization problem: the classical multiplicative approach (MU) and a novel proximal approach that achieves exact sparsity in molecular signatures, easing result interpretation. The method has been successfully tested on simulated as well as real data and compared with state-of-the-art methods for omics integration.
