An alternative to the Dirichlet-multinomial regression model for microbiome data analysis

The prevalence of human microbiome data in biomedical research has increased due to the association observed between microbiome composition and several diseases. The Dirichlet-multinomial distribution is frequently used to analyze this type of data, but it often fails to adequately model real microbiome datasets due to the restrictive parameterization imposed on its covariance matrix. This work proposes a novel distribution to be considered in microbiome data analysis, which can be used to define an alternative regression model for multivariate count data (e.g., microbiome data). The new distribution is a structured finite mixture model with Dirichlet-multinomial components. We show how this mixture model can enhance microbiome data analysis by clustering patients into “enterotypes”, a classification based on the taxa composition of gut microbiota. Finally, we consider an application based on a real gut microbiome dataset.