Machine Learning Meta-analysis of Large Metagenomic Datasets: Tools and Biological Insights

Abstract

Shotgun metagenomic analysis of the human associated microbiome provides a rich set of microbial features for prediction and biomarker discovery in the context of human diseases and health conditions. However, the use of such high-resolution microbial features presents new challenges, and validated computational tools for learning tasks are lacking. Moreover, classification rules have scarcely been validated in independent studies, posing questions about the generality and generalization of disease-predictive models across cohorts. In this paper, we comprehensively assess approaches to metagenomics-based prediction tasks and for quantitative assessment of the strength of potential microbiome-phenotype associations. We develop a computational framework for prediction tasks using quantitative microbiome profiles, including species-level relative abundances and presence of strain-specific markers. A comprehensive meta-analysis, with particular emphasis on generalization across cohorts, was performed in a collection of 2424 publicly available metagenomic samples from eight large-scale studies. Cross-validation revealed good disease-prediction capabilities, which were in general improved by feature selection and use of strain-specific markers instead of species-level taxonomic abundance. In cross-study analysis, models transferred between studies were in some cases less accurate than models tested by within-study cross-validation. Interestingly, the addition of healthy (control) samples from other studies to training sets improved disease prediction capabilities. Some microbial species (most notably Streptococcus anginosus) seem to characterize general dysbiotic states of the microbiome rather than connections with a specific disease. Our results in modelling features of the “healthy” microbiome can be considered a first step toward defining general microbial dysbiosis. The software framework, microbiome profiles, and metadata for thousands of samples are publicly available at http://segatalab.cibio.unitn.it/tools/metaml. The human microbiome–the entire set of microbial organisms associated with the human host–interacts closely with host immune and metabolic functions and is crucial for human health. Significant advances in the characterization of the microbiome associated with healthy and diseased individuals have been obtained through next-generation DNA sequencing technologies, which permit accurate estimation of microbial communities directly from uncultured human-associated samples (e.g., stool). In particular, shotgun metagenomics provide data at unprecedented species- and strain- levels of resolution. Several large-scale metagenomic disease-associated datasets are also becoming available, and disease-predictive models built on metagenomic signatures have been proposed. However, the generalization of resulting prediction models on different cohorts and diseases has not been validated. In this paper, we comprehensively assess approaches to metagenomics-based prediction tasks and for quantitative assessment of microbiome-phenotype associations. We consider 2424 samples from eight studies and six different diseases to assess the independent prediction accuracy of models built on shotgun metagenomic data and to compare strategies for practical use of the microbiome as a prediction tool.