Title: Subject-specific Bayesian hierarchical model for microbiome data analysis
Authors: Matteo Pedone - University of Florence (Italy) [presenting]
Francesco Stingo - University of Florence (Italy)
Abstract: Recent biomedical evidence suggests that the knowledge of microbiome composition and its function has a huge potential as a diagnostic tool. Motivated by the availability of microbiome abundance counts collected from different sources, clinical factors and diet-related covariates, the purpose is to explore associations between the microbial composition and the diet. Within the Dirichlet-multinomial regression framework, we propose a Bayesian hierarchical model that accounts for the complex structure of the interactions between diet and clinical factors. This leads to a high-dimensional framework, where sparsity is strongly induced via suitable priors and a thresholding mechanism. The model incorporates subject-specific regressions defined by coefficients that can vary flexibly with the covariates; the model effectively allows the effects of the covariates on the microbiome to be heterogeneous even when the sample size is small. An analysis of the microbiome abundance from patients affected by colorectal adenocarcinoma illustrates how the proposed approach can be used to determine the heterogeneous effects of diet and clinical factors on the microbiome.