Latent Tree based Inference of Ecological Network using the Poisson Log-Normal Model
Abstract
In the past decade, ecological networks have become a key tool to describe interactions between species and better understand the dynamics of a whole ecosystem or anticipate its response to a given change. Such interaction networks can be inferred based on the observation of the respective abundance of each species. Metagenomics relies on Next-Generation Sequencing (NGS) technologies to evaluate the (relative) abundance of microbial species in a given medium under varying experimental conditions or across replicates. A typical metabarcoding experiment results in a vector of read counts associated with each species under study. From a statistical perspective, network inference is usually considered in the framework of probabilistic graphical models. A huge statistical literature exists about this problem in the Gaussian case, that is when the data consists in continuous observations. These methods need to be adapted to count data. In this work, we propose a comprehensive statistical framework for the inference of ecological networks based on metagenomic counts. To this aim, we use the Poisson log-normal (PLN) model which provides a generic description for multivariate count data. The PLN model ac- counts for the specificities of metagenomic data such as over-dispersion or sequencing depth heterogeneity. More importantly, the PLN model allows to correct for the effect of covariates, which is critical to avoid the detection of spurious edges in the graph.