Expressive Analysis of Gut Microbiota in Pre- and Post- Solid Organ Transplantation Using Bayesian Topic Models

There is a growing evidence that variation in gut microbial communities has important associations with overall host health, and that the diversity and the richness of such communities is helpful in distinguishing patients at high risk of life-threatening post-transplantation conditions. The aim of our paper is to provide an expressive and highly interpretable characterization of microbiome alterations, with the goal of achieving more effective transplantations characterized by a rejection rate as low as possible, and to avoid more severe complications by treating patients at risk in a timely and effective way. For this purpose, we propose using topic models to identify those bacterial species that have the most important weight under the two different experimental conditions (healthy and transplanted patients, or patients whose fecal microbiota has been sampled both in pre- and post-transplantation phases). Topic models are Bayesian statistical models that are not affected by data scarcity, because conclusions we can draw borrow strength across sparse gut microbiome samples. By exploiting this property, we show that topic models are expressive methods for dimensionality reduction which can help analyze variation and diversity in gut microbial communities. With topic models the analysis can be carried out at a level close to natural language, as the output can be easily interpreted by clinicians, since most abundant species are automatically selected and the microbial dynamics can be tracked and followed over time.