Flexible Clustering Models for Next-Generation Metagenomic Sequence Samples

• Activity Number: 646
• Type: Topic Contributed
• Date/Time: Thursday, August 4, 2011 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Karin S. Dorman*+ and Ranjan Maitra
• Companies: Iowa State University and Iowa State University

Abstract

Next generation metagenomic sequence samples consist of many short sequences sampled from multiple organisms. It is a classic clustering or classification problem, where the goal is to assign each sequence to one of a finite number of organisms. Clustering is relevant when the sample may include unexpected or unknown species. Classification assumes all possible organisms are known. Typically, sequence data is modeled, explicitly or effectively, as Markovian. As data become abundant, the models have grown in complexity. One such flexible and popular model is the interpolated Markov model (IMM), borrowed from the field of speech recognition. We formulate IMMs as a mixture over the complexity (generally order) of the Markovian dependence structure. Using mixtures over dependence structure avoids the heuristics of IMM paremeter estimation, and naturally allows model selection using any standard criterion or formal model comparison using bootstrap. Metagenomic samples are then modeled as two-dimensional mixtures over context length and species. Implementation for large samples and complex models is challenging, but results for simulation and small datasets are promising.