Poisson-Markov Mixture Model and Parallel Algorithm for Binning Massive and Heterogenous DNA Sequencing Reads

A major computational challenge in analyzing metagenomics sequencing reads is to identify unknown sources of massive and heterogeneous short DNA reads. A promising approach is to efficiently and sufficiently extract and exploit sequence features, i.e., k-mers, to bin the reads according to their sources. Shorter k-mers may capture base composition information while longer k-mers may represent reads abundance information. We present a novel Poisson-Markov mixture Model (PMM) to systematically integrate the information in both long and short k-mers and develop a parallel algorithm for improving both reads binning performance and running time. We compare the performance and running time of our PMM approach with selected competing approaches using simulated data sets, and we also demonstrate the utility of our PMM approach using a time course metagenomics data set. The probabilistic modeling framework is sufficiently flexible and general to solve a wide range of supervised and unsupervised learning problems in metagenomics.