Predictive modeling of microbial data with interaction effects

Microbial interactions are of fundamental importance for the functioning and the maintenance of microbial communities. Deciphering these interactions from observational data or controlled lab experiments remains a formidable challenge due to their context-dependent nature, i.e., their dependence on (a)biotic factors, host characteristics, and overall community composition. Here, we present a statistical regression framework for microbial data that allows the inclusion and parsimonious estimation of species interaction effects for an outcome of interest. We adapt the penalized quadratic interaction model to accommodate common microbial data types as predictors, including microbial presence-absence data, relative (or compositional) abundance data from microbiome surveys, and quantitative (absolute abundance) microbiome data. We study the effect of including hierarchical interaction constraints and stability-based model selection on model performance and propose novel interaction model formulations for compositional data. To illustrate our framework’s versatility, we consider prediction tasks across a wide range of microbial datasets and ecosystems, including metabolite production in model communities in designed experiments and environmental covariate prediction from marine microbiome data. While we generally observe superior predictive performance of our interaction models, we also assess limits of these models in presence of extreme data sparsity and with respect to data type. On a large-scale gut microbiome cohort data, we identify sparse family-level interaction models that accurately predict the abundance of antimicrobial resistance genes, enabling the formulation of novel biological hypotheses about microbial community interactions and antimicrobial resistance.