A0398
Title: A general testing method for inference of microbial networks with compositional data
Authors: Yijuan Hu - Emory University (United States) [presenting]
Abstract: Inference of microbial networks reveals inter-dependencies or interactions among microbial taxa within communities. The compositional, sparse, high-dimensional, highly overdispersed, and sometimes clustered sequencing data pose significant challenges to this task. There is a lack of testing methods that control the false discovery rate (FDR) and thus calibrate the discoveries. A novel testing method called TestNet has been introduced. It is based on the empirical covariance and distance covariance of the centred-log-ratio data for capturing linear and nonlinear dependencies, respectively. A permutation procedure is developed for generating null replicates that account for the compositional effects and the extensive zero counts in microbiome data, assuming sparse dependencies in a microbial community. The permutation procedure readily allows schemes that preserve clustering structures in the samples, e.g., longitudinal samples. Therefore, the method applies to general scenarios involving the inference of microbial networks. The extensive simulation studies indicate that TestNet controls the FDR well while achieving high efficiency in a wide range of scenarios; the results from existing methods are not calibrated for any error rate.
