For a given bacterial strain, antimicrobial resistance (AMR) genes are those that have been identified to confer antibiotic resistance. Usually, the bacterial whole genome sequence or AMR genes are used to predict antibiotic resistance. Interestingly, it has recently been shown that machine learning methods can predict antibiotic resistance also using sets of conserved genes chosen at random.

This study investigates a possible reason for this phenomenon. We hypothesized that conserved genes closely related to AMR genes in terms of physical interactions could contain information about antibiotic resistance. To test our hypothesis, we quantified the relation between conserved genes and AMR genes in terms of distances on the interactome that were measured using kernel methods on protein-protein interaction networks. This allowed us to select sets of conserved genes that were evaluated in terms of their ability to predict antibiotic resistance using a machine learning model based on decision trees.

Our experiments could not conclusively establish the existence of a relation between the distance of conserved genes from AMR genes and antibiotic resistance. While our results can possibly be justified by the poor quality of the currently available data, our work also suggests the existence of compensatory alterations in bacterial genomes that could be used in the future for antibiotic resistance prediction.