Hierarchical Bayesian Graphical Models in Genomics

• Activity Number: 102
• Type: Invited
• Date/Time: Monday, August 4, 2014 : 8:30 AM to 10:20 AM
• Sponsor: General Methodology
• Author(s): Riten Mitra*+ and Yuan Ji and Peter Mueller
• Companies: University of Louisville and NorthShore University HealthSystem and University of Texas at Austin
• Keywords: Bayesian ; graphical priors ; posterior inference ; epigenetics ; latent variables ; multiple graphs

Abstract

We describe a class of hierarchical Bayesian graphical models having special relevance to next generation genomics and proteomics. Compared to traditional graphical models, these models induce greater flexibility by means of latent variables at different levels of the model hierarchy. This is critical, for example, in modeling interactions among latent presence of epigenetic markers going after a biologically more meaningful signal than raw correlations. The leaves of the hierarchy allow the choice of a wide range different sampling distributions- a desirable property in the context of sequencing data which deviate markedly from standard distributions. Lastly, the graphical priors at the roots allow easy extension to model dependent families of graphs. The latter combines networks across different biological platforms and facilitates differential network inference. Our description , though mostly limited to undirected graphs, are easily applicable to time directed networks. We shall highlight some novel posterior inference strategies invoked by such models. We conclude with some interesting future directions in modeling multiple dependent graphs.