agronahm@mcmaster.ca

1. Power Analysis for Differential Abundance Microbiome Studies

Abstract: Differential abundance studies play a crucial role in identifying genes that exhibit differential expression between control and treatment groups. The identification of such genes holds great potential for clinical applications and aiding in treatment decisions. However, it is imperative to ensure the reliability of the obtained results. This necessitates the use of power analysis to determine the reliability of significance results obtained in a differential abundance studies. We propose a novel framework for simulating microbiome count data based on actual data and demonstrate that the average power reported in the literature for differential abundance studies might be overestimated. Our framework allows for the estimation of required sample sizes to achieve desired power levels and calculation of the expected number of significant Operational Taxonomic Units (OTUs) in a given study design. By providing a framework and simulation-based evidence, we contribute to advancing the field’s understanding of power analysis and its implications for the accurate interpretation of results. 

2. Longitudinal Analysis of Microbiome Data including Correlation Structure

Human microbiome is dynamic in nature. Understanding the dynamics in longitudinal microbiome datacan help explain the mechanisms that underpin human health and disease. However, analysing longitudinal microbiome data, whether 16S rRNA or metagenome shotgun, is challenging. Along with the difficulties presented by the characteristics of microbiome data such as sparsity, overdispersion, and compositionality, repeated measure introduces correlation between observations from the same individual at different time points. The majority of methods used in the literature to analyse longitudinal microbiome data only model one taxon at a single time point and do not handle the correlation structure between individual taxa across time. In this work, a multivariate Negative Binomial Generalized Mixed Model (MVNBMM) is proposed for analyzing longitudinal microbiome data. A special feature of MVNBMM is that is can handle the correlation structure of individual taxa over time. Modelling the correlation between repeated measures in individual samples is however, very challenging because the number of parameter to be estimated for a negative binomial model is computationally expensive and almost impossible for microbiome dataset which is typically of high dimensionality. In this project, a reduced rank method, which reduces the number of parameter estimates is used for modelling the correlation structure of taxa.