Bayesian mixed effects models for zero-inflated compositions in microbiome data analysis

摘要

Detecting associations between microbial composition and samplecharacteristics is one of the most important tasks in microbiome studies. Mostof the existing methods apply univariate models to single microbial speciesseparately, with adjustments for multiple hypothesis testing. We propose aBayesian nonparametric analysis for a generalized mixed effects linear modeltailored to this application. The marginal prior on each microbial compositionis a Dirichlet Processes, and dependence across compositions is induced througha linear combination of individual covariates, such as disease biomarkers orthe subject's age, and latent factors. The latent factors capture residualvariability and their dimensionality is learned from the data in a fullyBayesian procedure. We propose an efficient algorithm to sample from theposterior and visualizations of model parameters which reveal associationsbetween covariates and microbial composition. The proposed model is validatedin simulation studies and then applied to analyze a microbiome dataset forinfants with Type I diabetes.