Modeling time-series data from microbial communities

Abstract

As sequencing technologies have advanced, the amount of information regarding the composition of bacterial communities from various environments (e.g. skin, soil) has grown exponentially. To date, most work has focused on cataloging taxa present in samples and determining whether the distribution of taxa shifts with exogenous covariates. However, important questions regarding how taxa interact with each other and their environment remain open, thus preventing in-depth ecological understanding of microbiomes. Time-series data from 16S rDNA amplicon sequencing are becoming more common within microbial ecology, but given the big data nature of these studies, there are currently no methods capable of utilizing the breadth of the data to infer ecological interactions from these longitudinal data. We address this gap by presenting a method of analysis using Poisson regression fit with an elastic-net penalty that 1) takes advantage of the fact that the data are time series; 2) constrains estimates to allow for the possibility of many more interactions than data; and 3) is scalable enough to handle data consisting of thousands of taxa. We test the method on gut microbiome data from white-throated woodrats (Neotoma albigula) that were fed varying amounts of the plant secondary compound oxalate over a period of 22 days to estimate interactions between OTUs and their environment.