Bayesian mixture analysis for metagenomic community profiling.

Sofia Morfopoulou; Vincent Plagnol

Methods focused on the statistical inference of the set of present species as well as the estimation of their relative proportions, incorporate knowledge from all reads to assign each individual read to a species. From a statistical standpoint, this identification and quantification question can be thought of as an application of mixture models. These ideas have been applied in the metagenomics context in frequentist (GRAMMy (Xia et al., 2011) ) and Bayesian (Pathoscope (Francis et al., 2013) ) settings. GRAMMy formulates the problem as a finite mixture model, using the Expectation-Maximization (EM) algorithm to estimate the relative genome abundances. Pathoscope refines this process by penalizing reads with ambiguous matches in the presence of reads with unique matches and enforcing parsimony within a Bayesian context. Both methods work with unassembled sequence data and they are not currently setup to incorporate an initial short read assembly step, which could be achieved by assigning a higher weight to contigs formed by multiple reads. Fitting a mixture model is useful for the species relative abundance estimation, as well as the read to species assignment. A related but distinct question concerns the set of species which should be included in the mixture model. This question is closely related to the biological question of asking what species are present in the mixture. Including all species flagged as potential matches by the read classification can introduce a large number of species, often in the low thousands. Mixture models will, in this situation, identify a large number of species at low levels. This interpretation is appropriate in some applications. In many other cases, the expectation is that the underlying species set should be parsimonious and that some divergence with database species or sequencing errors can explain a large fraction of the non matching reads.