B1182
Title: Penalized graphical models for cross-sectional and longitudinal microbiome data
Authors: Zachary Kurtz - Lodo Therapeutics (United States)
Christian L. Mueller - Simons Foundation (United States) [presenting]
Abstract: Learning statistical association networks from microbial targeted amplicon sequencing (TAS) data holds the promise to unravel the organizational structure of microbes in their natural habitats, ranging from human gut to large-scale marine ecosystems. TAS count data are compositional (or relative abundance) data due to experimental limitations, thus requiring dedicated statistical methods for network inference. For cross-sectional microbiome data, a popular statistical framework is based on log-ratio data transformations from the field of compositional data analysis, followed by shrinkage-based graphical model inference. We extend this framework to longitudinal and time series data using a latent variable approach. For short longitudinal data sets we model the autocorrelations in the samples via structured low-rank components, followed by shrinkage-based graphical model inference. For sufficiently long time series, we first employ a Bayesian non-parametric approach that takes into account the compositionality of the time series, and then employ shrinkage-based latent graphical models on the residuals. We illustrate the efficacy of these approaches on several longitudinal and time series data sets from the human gut microbiome.
