Prediction of Novel Antimicrobial Resistance Genes (ARGs) Using Machine Learning Methods on Metagenomic Data

Nowadays, most countries are completely dependent on antimicrobial drugs for medicine and food security, thus causing a decline in antibiotics’ effectiveness by over-consumption. The increase in pathogen resistance to antimicrobials poses a threat of fatal infections and limits the potential health benefits of medical treatments. Therefore, it is important to devise a method for the early detection of antimicrobial resistance genes (ARGs). However, most methods for novel ARGs discovery are based on sequence similarity to a predefined database, and therefore limit the gene repertoire that can be discovered. Here, we developed machine learning (ML) methods to predict new ARGs that share no similarity to any previously identified resistance gene. Our approach is based on various features, such as protein physicochemical properties, genomic context, and phylogenetic patterns. We predicted 157 ARG candidates with no functional annotation and no close or distant characterized homologs. This work demonstrates the potential of ML methods on biological signatures for the discovery of novel ARG without relying on sequence similarity to known genes. Our approach has the potential to improve our understanding of the resistance gene repertoire and contribute to better surveillance of AMR genes across the globe.