A1688
Title: Joint Bayesian additive regression trees for inference of multiple non-linear dependence networks
Authors: Christine Peterson - The University of Texas MD Anderson Cancer Center (United States) [presenting]
Abstract: Many diseases are heterogeneous, with subgroups of patients that have differing biological mechanisms. Estimating a single graph using the entire dataset may miss subtype-specific relations, while analyzing the data separately for each subgroup reduces statistical power to identify shared mechanisms. Importantly, the dependence relations among features may be non-linear. To address this challenge, we propose a hierarchical Bayesian model that encourages shared edge selection while allowing each group to have its own subtype-specific network. We formulate a flexible model based on the Bayesian additive regression tree framework to allow for non-linear dependencies within each subgroup, and encourage the joint selection of edges across groups through a hierarchical prior. The proposed method will be illustrated through both simulation studies and a real data application to protein-protein interaction networks across colorectal cancer subtypes.
