Statistical Methods for Association Analysis of Microbiome Data

Abstract

The human microbiome is an integral component of the human body. High-throughput sequencing techniques have provided detailed information on abundance and phylogeny of individual taxa in the human microbiome. A variety of association studies based on microbiome data has emerged in recent years, revealing important relationships among microbial features as well as between the microbiome and host health. Challenges specific to microbiome data, such as high-dimensionality and sparsity, call for novel statistical approaches. Meanwhile, common practical needs in association analyses, such as covariate adjustment and analysis of clustered data, can be extended to microbiome data. Here we present four projects on novel statistical methods for association analyses of microbiome data. In Project 1, we propose a powerful kernel-based approach for microbiome genome-wide association studies (GWASs), where we evaluate the covariate-adjusted association between groups of genetic variants at the gene level and the overall microbiome composition at the community level. In Project 2, we develop a kernel-based multivariate independence test for clustered data and apply the test to evaluate the association between the overall microbiome composition and a multivariate trait based on longitudinal data. In Project 3, we propose a multivariate approach to construct microbial association networks, where we develop a conditional independence test to assess the pairwise association between multivariate microbial features, such as bacterial genera composed of multiple species. In Project 4, we propose a novel approach for one-sample Mendelian randomization with a microbial exposure, which allows us to evaluate the causal effect of individual microbial taxa on a continuous health outcome with an improved power.