Bayesian Modeling on Microbiome Data Analysis: Application to Subgingival Microbiome Study.

The study of microbiome data has been widely used to investigate associations between the abundance of microbial taxa and human diseases. Identifying and understanding these relationships precisely gives the microbiome a key role in human health, disease status, and the development of new diagnostics and targeted therapeutics. Due to its unique features such as compositional data, excessive zero counts, overdispersion, and complexed structure between taxa, undertaking effective microbiome data analytics presents numerous obstacles. To quantify covariate-taxa effects on the subgingival microbiome study, we proposed a refined Bayesian zero-inflated negative binomial (ZINB) regression model with random subject effects. This proposed approach not only accommodates inflated zero counts and overdispersion similar to the existing ZINB model developed by Jiang et al. (Biostatistics 22(3):522-540, 2021), but also accounts for subject-level heterogeneity through the inclusion of random subject effects. In addition, an efficient Markov chain Monte Carlo (MCMC) sampling algorithm was developed for Bayesian computation. Overall effects of pre-selected group variables on predicted taxa abundance were estimated and tested under the proposed model. We conduct simulation studies and demonstrate that the proposed model outperforms the competing models in achieving a better power with controlling the type I error. The usefulness of the proposed model is applied to a real subgingival microbiome study.