|
Abstract:
|
Multi-platform genomics data contain extensive information, but are challenging to integrate in a coherent fashion. We describe two integration efforts: (1) Gene-specific epigenetic profiles (GSEP) and (2) pathway-based integrative Bayesian Analysis of Genomics data (piBAG). GSEPs integrate methylation and expression data to determine which genes are strongly regulated by methylation and identify which CpG sites are driving the association. Our approach involves penalized regression that induces sparsity and penalizes CpG based on likelihood of importance based on characteristics learned from genome-level analyses. We apply to CRC data. piBAG is a multi-level Bayesian hierarchical model with a mechanistic part decomposing mRNA expression into components explained by upstream platforms, a clinical part regressing clinical outcome on platform-specific gene components, and a novel hierarchical shrinkage prior that induces sparsity, borrows strength across genes in a common pathway, and produces pathway scores measuring prognostic importance of that pathway. We demonstrate by simulation gains engendered by the model's integrative and pathway components, and apply to GBM data.
|
