Integrating functional and taxonomic data types for microbiome data analysis

Abstract

Communities of microbes in natural environments, referred to as microbiomes, are commonly profiled with DNA sequencing approaches. Sequencing results are typically partitioned so that the relative abundances of microbes (taxonomic data) and genes (functional data) are analyzed separately. It is challenging to biologically interpret these data, partially due to the lack of computational frameworks for joint analysis of taxonomic and functional data. Herein, I addressed this issue from three perspectives. First, I did so in the context of an investigation into the microbiome of pediatric Crohn's disease patients. Our main goal with this work was to compare the performance of microbiome data types for classifying samples in both independent and combined models. We found that genera identified through marker-gene sequencing performed best in these models, but that in combined models functions performed best for classifying treatment response. Although these and other insights were valuable, it became clear that improved methods for generating and analyzing taxa-function links were needed. One method for generating these links is through metagenome prediction methods. Although these approaches are widely used, they suffer from several major caveats and have been inconsistently validated. Accordingly, I developed a new bioinformatic method, PICRUSt2, for generating predicted taxa-function links based on several hypothesized improvements. Although I confirmed that this new approach performed moderately better than alternative methods, I also identified issues with analyzing metagenome predictions in general. My final project focused on partially addressing these and related problems in functional data analysis. I did this by developing a novel method to better integrate taxonomic and functional data types to identify functional biomarkers. This tool, POMS, accurately identified genes under selection in simulated data and performed well when applied to actual case-control metagenomics datasets. Taken together, this thesis represents several valuable developments in joint taxa-function analysis that enabled improved interpretation of microbiome data. In several instances, particularly with the application POMS, this joint analysis approach yielded novel insights that would be overlooked by analyzing each data type individually.