A Bayesian Methodology for High-Dimensional Discrete Graphical Models with an Application to Protein Structure Prediction

• Activity Number: 581 - High-Throughput Biological Data Analyzed with Bayesian Methods
• Type: Contributed
• Date/Time: Wednesday, August 2, 2017 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Anwesha Bhattacharyya* and Yves Atchade
• Companies: and University of Michigan
• Keywords: Bayesian ; High Dimension ; Variable Selection ; Moreau-Yosida approximation ; Protein Structure Prediction

Abstract

This work introduces a bayesian methodology for fitting large discrete graphical models (Potts models) with spike-and-slab priors with point-mass at the origin to encode sparsity. Resulting posterior distributions are well-known to be difficult to handle by standard MCMC algorithms. A Moreau-Yosida approximation of the posterior distribution is proposed that is tractable and scales well with the size of the problem. The motivation of this problem comes from protein contact prediction, where Potts models have been used to infer amino acid residue contacts and have played an important role in predicting folded three dimensional structures of the proteins. The use of a Bayesian approach for this problem allows us to leverage existing prior information, and evaluate posterior uncertainty. We present simulation results to illustrate the performance.