Bayesian mixture analysis for metagenomic community profiling.

Sofia Morfopoulou; Vincent Plagnol

The main challenge behind such a formulation is computational. Even with a relatively small number of species to consider, the number of subsets of this space that could explain the mixture grows exponentially. Efficient computational strategies are required to make this problem tractable. Here we show that this inference can be achieved for modern scale metagenomics datasets. Our strategy is based on parallel tempering, a Monte Carlo Markov Chain technique, using parallel computing to speed up the inference. We implemented these ideas in a user friendly R package called metaMix. metaMix produces posterior probabilities for various models as well as the relative abundances under each model. We demonstrate its potential using datasets from clinical samples as well as benchmark metagenomic datasets.