Dr. Michael Wu - Statistics for Keeping your Microbiome Analyses out of the Toilet

Abstract: Microbiome profiling studies of hundreds to thousands of individuals are being conducted within existing epidemiologic cohorts. Analysis of data from these studies offers comprehensive identification of bacterial taxa related to a plethora of health outcomes. However, key characteristics of these studies (e.g. large sample size) also induce serious statistical challenges, particularly in combination with the difficulties inherent to microbiome data (e.g. high-dimensionality, sparsity, compositionality). Some challenges include accommodating batch effects and robustly identifying taxa related to outcomes. To address these problems, we propose novel batch correction and individual-taxon differential abundance testing frameworks. Our work is based on using two-part zero-inflated quantile regression which makes minimal distributional assumptions while accommodating the zero-inflated nature of the data. We illustrate our work through simulations and application to data from a number of large-scale microbiome studies including the CARDIA cohort and HIV Reanalysis Consortium