A Two-Way Bayesian Mixture Model for Clustering in Metagenomics

Abstract

Abstract. We present a new and efficient Bayesian mixture model based on Poisson and Multinomial distributions for clustering metagenomic reads by their species of origin. We use the relative abundance of different words along a genome to distinguish reads from different species. The distribution of word counts within a genome is accurately represented by a Poisson distribution. The Multinomial mixture model is derived as a standardized Poisson mixture model. The Bayesian network efficiently encodes the conditional dependencies between word counts in a DNA due to overlaps and hence is most consistent with the data. We present a two-way mixture model that captures the high dimensionality and sparsity associated with the data. Our method can cluster reads as short as 50 bps with accuracy over 80%. The Bayesian mixture models clearly outperform their Naive Bayes counterparts on datasets of varying abundances, divergences and read lengths. Our method attains comparable accuracy to that of state-of-art Scimm and converges at least 5 times faster than Scimm for all the cases tested. The reduced time taken, by our method, to obtain accurate results is highly significant and justifies the use of our proposed method to evaluate large metagenome datasets.