Microbiome Enterotype Detection via a Latent Variable Allocation Model

Human gut microbiome studies are of critical importance to evaluate the health of an individual. Understanding the interactions within microbial communities has a crucial role in comprehending human biological systems. Topic modelling techniques are usually employed to analyze textual documents and exploit a latent variable framework to find hidden topics in the data. Quite interestingly, a clear semantic correspondence between text and microbiome analyses allows applying these techniques to detect enterotypes. One of the most commonly used models in the field is the Latent Dirichlet Allocation which, however, suffers from some limitations due to the stiffness of its standard topic prior distribution, the Dirichlet. This study proposes a flexible Dirichlet distribution as topic prior distribution in the context of microbial systems. The goal is to detect the important taxa characterizing different enterotypes.