Predicting microbial community compositions using compressive sensing

A primary goal of microbiome engineering is to use synthetic and natural microbial communities for improving outcomes in human health, agriculture, and climate. For these methods to succeed, it is crucial to reliably predict the final community composition from the initial composition, in a well-defined environment. However, our ability to learn community compositions is hindered by the vast number of experiments required to assemble all communities possible from a given pool of species. Thus far, predictive methods have focussed on fitting parametric, mechanistic models, like the generalized Lotka-Volterra model, to limited empirical data.

In this work, we present a new model-agnostic approach to predicting community compositions using compressive sensing. Working on a community-structure landscape, we discovered a sparse representation of the community compositions. We then leveraged this sparsity to predict community compositions from limited data by applying techniques from compressive sensing. By sampling just ~1% of all possible communities, we recovered community compositions with high accuracy on synthetic datasets. We demonstrate the usefulness of this method in diverse real-world settings by applying it to four experimental datasets, with two datasets associated with soil microbiome, one relevant to the fruit fly gut microbiome, and one consisting of species relevant to the human gut.