MOTL: enhancing multi-omics matrix factorization with transfer learning

Abstract

AbstractJoint matrix factorization is a popular method for extracting lower dimensional representations of multi-omics data. It disentangles underlying mixtures of biological signals, facilitating efficient sample clustering, disease subtyping, or biomarker identification, for instance. However, when a multi-omics dataset is generated from only a limited number of samples, the effectiveness of matrix factorization is reduced. Addressing this limitation, we introduce MOTL (Multi-Omics Transfer Learning), a novel framework for multi-omics matrix factorization with transfer learning based on MOFA (Multi-Omics Factor Analysis). MOTL infers latent factors for a small multi-omics dataset, with respect to those inferred from a large heterogeneous learning dataset. We designed two protocols to evaluate transfer learning approaches, based on simulated and real multi-omics data. Using these protocols, we observed that MOTL improves the factorization of multi-omics datasets, comprised of a limited number of samples, when compared to factorization without transfer learning. We showcase the usefulness of MOTL on a glioblastoma dataset comprised of a small number of samples, revealing an enhanced delineation of cancer status and subtype thanks to transfer learning.