Modelling Microbiome Association with Host Phenotypes Using a Bayesian Dirichlet Process Model

Dysbiosis in the human gut microbiome has been shown to be intimately involved in the pathogenesis of a wide range of communicable and non-communicable diseases. As microbiome wide association study becomes the workhorse for identifying association between microbial taxa and human diseases/traits, proper modelling of microbial taxa abundances is critical. In particular, statistical frameworks need to effectively model correlation among microbial taxa as well as latent heterogeneity across samples. Here, a Bayesian method using the Dirichlet process random effects model is devised for microbiome association study. The proposed method uses a weighted combination of phylogenetic and radial basis function kernels to model taxa effects, and a non-parametrically modelled latent variable to model latent heterogeneity among samples. Using simulated and real microbiome datasets, it is shown that the method has high statistical power for association inference. SoftwareThe R codes to implement the method has been incorporated into a script phy-loDPM.R, and is available online at https://github.com/AwanyDenis/phyloDPM.