A Bayesian Multi-Task Approach for Detecting Global Microbiome Associations

MotivationThe human gut microbiome has been shown to be associated with a variety of human diseases, including cancer, metabolic conditions and inflammatory bowel disease. Current statistical techniques for microbiome association studies are limited by relying on measures of ecological distance, or only allowing for the detection of associations with individual bacterial species, rather than the whole microbiome. ResultsIn this work, we develop a novel Bayesian multi-task approach for detecting global microbiome associations. Our method is not dependent on a choice of distance measure, and is able to incorporate phylogenetic information about microbial species. We apply our method to simulated data and show that it allows for consistent estimation of global microbiome effects. Additionally, we investigate the performance of the model on two real-world microbiome studies: a study of microbiome-metabolome associations in inflammatory bowel disease (Beamish, 2017), and a study of associations between diet and the gut microbiome in mice (Turnbaugh et al., 2009). We show that we can use the method to reliably detect associations in real-world datasets with varying numbers of samples and covariates. AvailabilityOur method is implemented using the R interface to the Stan Hamiltonian Monte Carlo sampler. Software for running our methods is available at https://github.com/FrankD/MicrobiomeGlobalAssociations. Contactf.dondelinger@lancaster.ac.uk