Inferring strain-level mutational drivers of phage-bacteria interaction phenotypes

Abstract

AbstractThe enormous diversity of bacteriophages and their bacterial hosts presents a significant challenge to predict which phages infect a focal set of bacteria. Infection is largely determined by complementary -and largely uncharacterized-genetics of adsorption, injection, and cell take-over. Here we present a machine learning (ML) approach to predict phage-bacteria interactions trained on genome sequences of and phenotypic interactions amongst 51Escherichia colistrains and 45 phage λ strains that coevolved in laboratory conditions for 37 days. Leveraging multiple inference strategies and withouta prioriknowledge of driver mutations, this framework predicts both who infects whom and the quantitative levels of infections across a suite of 2,295 potential interactions. The most effective ML approach inferred interaction phenotypes from independent contributions from phage and bacteria mutations, predicting phage host range with 86% mean classification accuracy while reducing the relative error in the estimated strength of the infection phenotype by 40%. Further, transparent feature selection in the predictive model revealed 18 of 176 phage λ and 6 of 18E. colimutations that have a significant influence on the outcome of phage-bacteria interactions, corroborating sites previously known to affect phage λ infections, as well as identifying mutations in genes of unknown function not previously shown to influence bacterial resistance. While the genetic variation studied was limited to a focal, coevolved phage-bacteria system, the method’s success at recapitulating strain-level infection outcomes provides a path forward towards developing strategies for inferring interactions in non-model systems, including those of therapeutic significance.