Bayesian mixture analysis for metagenomic community profiling.

Sofia Morfopoulou; Vincent Plagnol

To limit the complexity of the fit, we used the two step procedure described in the Methods and fully implemented in metaMix. We first fitted the mixture model with the complete set of 2,500 species and a limited run length of 500 iterations. Based on this analysis, we identified 1,312 species supported by at least one read and explored this state space. To limit the computational time, we also considered a stronger approximation, including only the 374 potential species supported by at least 10 sequencing reads. Both approaches generated similar results, albeit the more complex one with 1,312 potential species required the quadruple of the computation time (12h instead of 3h). metaMix identified 116 species, detecting successfully all the members of the metagenomic community (Supplementary Table S1 ). These were detected on the strain level except in four instances where a different strain of the same species, or different species within the same genus was detected. Four species were identified and not in the simulated dataset, hence can be considered as false positives (Supplementary Table S1 ). enforces the parsimonious results (any thetaPrior greater than 10), thereby removing the unique read penalty, Pathoscope behaves as a standard mixture model and identifies 165 species. With these settings, it identifies all but one members of the community. The organisms are identified at the strain level, except in three instances where it identified different species within the same genus. The major interpretation issue is the presence of a long tail of species (54 species) that are actually not present in the mixture (Supplementary Table S1 ). Pathoscope produced the results in one minute.