B1767
Title: Bayesian cross-study models: From epidemiological to genomics applications
Authors: Roberta De Vito - Brown University (United States) [presenting]
Abstract: One of the most important challenges in biological sciences today is to elucidate how complex experiments, which measure hundreds of thousands of variables, generate consistent answers when repeated; and how these answers can be learned by analyzing several studies together. We start from the premise that genuine biological patterns are more likely than spurious patterns to be consistently present in multiple studies. Our challenge is to systematically and reliably identify the consistent biological patterns shared among studies and remove variation that lacks such reproducibility. To meet this challenge, we propose a novel analytical concept by upgrading a very widely used statistical technique known as factor analysis. Our Bayesian Multi-study Factor model is able to handle multiple studies simultaneously in a Bayesian shrinkage prior approach by using a Gibbs Sampling algorithm for parameter estimates. We present several different biological: nutritional epidemiological data in seven different countries, microarray gene expression in 4 different studies, and 12 brain regions in tissue studies.
