Bayesian multistudy factor analysis for high-throughput biological data

This paper analyzes breast cancer gene expression across seven studies to identify genuine and thus replicable patterns shared among these studies. Our premise is that biological signal more likely be reproducibly present in multiple than spurious signal. analysis uses a new modeling strategy for the joint of high-throughput which simultaneously identifies as well study-specific To this end, we generalize multi-study factor model handle high-dimensional data sparse Bayesian infinite context. We provide strategies identification loading matrices, common study-specific. Through extensive simulation analysis, characterize performance proposed approach various scenarios show it outperforms standard identifying all considered. The clear patterns. These are related well-known pathways involved cancer, such ER, cell cycle, immune system, collagen, metabolic pathways. Some also associated with existing subtypes, LumA, Her2, basal while other novel active subtypes missed by hierarchical clustering approaches. R package MSFA implementing method available on GitHub.