Bayesian mixture analysis for metagenomic community profiling.

Sofia Morfopoulou; Vincent Plagnol

We approximate this penalty factor based on a user-defined parameter r that represents the species read support required by the user to believe in the presence of this species. We compute the logarithmic penalty value as the log-likelihood difference between two models: one where all N reads belong to the "unknown" category and one where r reads have a perfect match to some unspecified species and the remaining N − r reads belong to the "unknown" category. In the nucleotide similarity situation, the p ij probabilities for the r reads originating from this unspecified species are approximated by 1/(median genome length in the reference database). This parameter essentially reflects how many reads are required to provide credible support that a species is present in the mixture and acts as a probabilistic threshold as opposed to a deterministic one applied on a ranked list.