Scalable Profiling and Visualization for Characterizing Microbiomes

Abstract

Metagenomics is the study of the combined genetic material found in microbiome samples, and it serves as an instrument for studying microbial communities, their biodiversities, and the relationships to their host environments. Creating, interpreting, and understanding microbial community profiles produced from microbiome samples is a challenging task as it requires large computational resources along with innovative techniques to process and analyze datasets that can contain terabytes of information. The community profiles are critical because they provide information about what microorganisms are present in the sample, and in what proportions. This is particularly important as many human diseases and environmental disasters are linked to changes in microbiome compositions. In this work we propose novel approaches for the creation and interpretation of microbial community profiles. This includes: (a) a cloud-based, distributed computational system that generates detailed community profiles by processing large DNA sequencing datasets against large reference genome collections, (b) the creation of Microbiome Maps: interpretable, high-resolution visualizations of community profiles, and (c) a machine learning framework for characterizing microbiomes from the Microbiome Maps that delivers deep insights into microbial communities. The proposed approaches have been implemented in three software solutions: Flint, a large scale profiling framework for commercial cloud systems that can process millions of DNA sequencing fragments and produces microbial community profiles at a very low cost; Jasper, a novel method for creating Microbiome Maps, which visualizes the abundance profiles based on the Hilbert curve; and Amber, a machine learning framework for characterizing microbiomes using the Microbiome Maps generated by Jasper with high accuracy. Results show that Flint scales well for reference genome collections that are an order of magnitude larger than those used by competing tools, while using less than a minute to profile a million reads on the cloud with 65 commodity processors. Microbiome maps produced by Jasper are compact, scalable representations of extremely complex microbial community profiles with numerous demonstrable advantages, including the ability to display latent relationships that are hard to elicit. Finally, experiments show that by using images as input instead of unstructured tabular input, the carefully engineered software, Amber, can outperform other sophisticated machine learning tools available for classification of microbiomes.