Learning microbial interaction networks from metagenomic count data

TL;DR

This paper introduces a hierarchical Poisson-multivariate normal model to accurately infer direct microbial interactions from metagenomic count data, outperforming existing methods in synthetic and real plant-associated microbiome experiments.

Contribution

The paper presents a novel hierarchical model that controls for confounders and captures direct microbial interactions using an $\\ell_1$ penalized precision matrix, improving inference accuracy.

Findings

Outperforms SparCC and glasso in synthetic tests

Correctly identifies direct interactions in plant microbiome data

Provides a structured approach for modeling count-based microbial data

Abstract

Many microbes associate with higher eukaryotes and impact their vitality. In order to engineer microbiomes for host benefit, we must understand the rules of community assembly and maintenence, which in large part, demands an understanding of the direct interactions between community members. Toward this end, we've developed a Poisson-multivariate normal hierarchical model to learn direct interactions from the count-based output of standard metagenomics sequencing experiments. Our model controls for confounding predictors at the Poisson layer, and captures direct taxon-taxon interactions at the multivariate normal layer using an $\ell_1$ penalized precision matrix. We show in a synthetic experiment that our method handily outperforms state-of-the-art methods such as SparCC and the graphical lasso (glasso). In a real, in planta perturbation experiment of a nine member bacterial community, we show our model, but not SparCC or glasso, correctly resolves a direct interaction structure among three community members that associate with Arabidopsis thaliana roots. We conclude that our method provides a structured, accurate, and distributionally reasonable way of modeling correlated count based random variables and capturing direct interactions among them.