Bayesian mixture analysis for metagenomic community profiling.

Sofia Morfopoulou; Vincent Plagnol

A practical prior for the mixing parameters w is the Dirichlet distribution owing to its conjugate status to the multinomial distribution. Despite the use of conjugate priors, the probabilistic assignment of reads to species involves the expansion of the likelihood into K n terms which is computationally infeasible through direct computation. An efficient estimation can be performed by the introduction of unobserved latent variables that code for the read assignments. In this framework, either the Gibbs sampler (Marin et al., 2005) , a Monte Carlo Markov Chain technique, or the Expectation-Maximization (EM) (Dempster and . CC-BY-NC-ND 4.0 International license is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/007476 doi: bioRxiv preprint Laird, 1977) algorithm can be used to estimate the mixture weights w. EM returns a point estimate for w while the Gibbs sampler the distribution of w (Supplementary Text 1 for details of implementation). Both methods were implemented and provided comparable results.