Bayesian Zero-Constrained Regression Modeling with Structured Prior for Microbiome Feature Selection

• Activity Number: 423 - Bayesian Microbiomics
• Type: Invited
• Date/Time: Thursday, August 12, 2021 : 4:00 PM to 5:50 PM
• Sponsor: International Society for Bayesian Analysis (ISBA)
• Author(s): Liangliang Zhang* and Yushu Shi and Kim-Anh Do and Robert Jenq and Christine B. Peterson
• Companies: The University of Texas MD Anderson Cancer Center and The University of Missouri Columbia and The University of Texas MD Anderson Cancer Center and The University of Texas MD Anderson Cancer Center and The University of Texas MD Anderson Cancer Center
• Keywords: Bayesian regression; feature selection; microbiome; compositional data; Ising prior; complex structure

Abstract

With the enormous advances in microbiome studies, researchers have suggested that the human microbiota is becoming a crucial role in understanding health and diseases. However, the applicability of standard variable selection methods is complicated by the main challenges of microbiome data. First, the OTU data are compositional, as the abundances within each sample sum up to a constant. I proposed a novel generalized transformation and zero-constrained prior to handle the compositionality. This novel prior transform the parameters and consistently shrink their summation to zero. Second, different OTUs can sometimes have a similar impact on the outcome, as they share phylogenetic similarities. I utilized the Ising prior to encourage the joint selection of OTUs that are similar to each other. In addition, as binary outcomes and survival data are very common in medical research, we extend the Bayesian linear regression framework to Bayesian logistic regression and survival models. We demonstrate the better performance of our method compared with existing variable selection method in both simulation and real data application.