A deeper understanding of the relationship between the antimicrobial resistance (AMR) gene carriage and phenotype is necessary to develop effective response strategies against this global burden. AMR phenotype is often a result of multi-gene interactions; therefore, we need approaches that go beyond current simple AMR gene identification tools. Machine-learning (ML) methods may meet this challenge and allow the development of rapid computational approaches for AMR phenotype classification. To examine this, we applied multiple ML techniques to 16,950 bacterial genomes across 28 genera, with corresponding MICs for 23 antibiotics with the aim of training models to accurately determine the AMR phenotype from sequenced genomes. This resulted in a >1.5-fold increase in AMR phenotype prediction accuracy over AMR gene identification alone. Furthermore, we revealed 528 unique (often species-specific) genomic routes to antibiotic resistance, including genes not previously linked to the AMR phenotype. Our study demonstrates the utility of ML in predicting AMR phenotypes across diverse clinically relevant organisms and antibiotics. This research proposes a rapid computational method to support laboratory-based identification of the AMR phenotype in pathogens.

Abstract Background Antimicrobial resistance (AMR) genes are found to be ubiquitous within the microbiome, even when antimicrobial usage is absent. To identify the AMR phenotype, the most common method is to use a laboratory-based assay. Yet, when dealing with samples from the microbiome, many species are difficult to culture within the laboratory. The vast quantity of strains would be time-consuming to culture. To avoid this, a computational approach may be a more favourable choice. AMR gene finder tools are efficient at determining the AMR genotype. Despite this, how a genotype relates to the AMR phenotype is still an open question. Methods To evaluate the relationship between the AMR phenotype and the AMR genotype, 16 950 genomes from BV-BRC which had corresponding MIC values were analysed. Using Weka’s J48 decision tree model, the relationship between the AMR phenotype and the AMR genotype was analysed. The role of accessory genes in relation to the AMR phenotype was analysed in the same way. Results The J48 models could predict the AMR phenotype accurately using AMR genes and accessory genes; the average accuracy was 91.7% and 92.2%, respectively. The results found that gene co-occurrence, presence and absence of genes are key factors to analyse when identifying the AMR phenotype from genomic data. These factors are not evaluated by commonly used AMR gene finder tools, which could miss vital information to determine the correct phenotype. Conclusions Our results highlight why we should continue to research the relationship between the AMR phenotype and genomic data.