COMPLEMENTATION OF GENOTYPE-TO-PHENOTYPE ANTIMICROBIAL RESISTANCE PREDICTION OF ENTEROCOCCI WITH EXPERIMENTAL APPROACHES

Abstract

Antimicrobial resistance (AMR) is a global-health crisis that currently causes more than 1 million deaths annually. Whole-genome sequencing (WGS) data can provide insights into AMR transmission and mechanisms to guide surveillance and prevention strategies. Machine learning (ML) has shown great promise in predicting resistance phenotypes from WGS data, but predictions must be rigorously examined and experimentally corroborated prior to clinical implementation. We used ML and feature selection to examine the key drivers of prediction accuracy in 647 genomes of Enterococcus faecium and Enterococcus faecalis for vancomycin, doxycycline, and erythromycin resistance. Known resistance genes were predictive of vancomycin and doxycycline resistance, but not erythromycin. Mobile genetic elements (MGEs) alone also yielded highly accurate predictions, highlighting their association with key resistance genes. We experimentally tested this linkage by quantifying rates of plasmid transfer among Enterococcus species. The vancomycin resistance gene-carrying MGEs existed as a transposable unit alone or nested within a plasmid. Conjugation was successful with all of the tested intra- and interspecies hosts. Doxycycline and erythromycin resistance genes were always transferred via a plasmid, with limited instances of interspecies conjugation. Transcriptome profiling of enterococci exposed to vancomycin or doxycycline revealed upregulation of key predictive drivers, including genes and MGEs, associated with resistance. Doxycycline demonstrated the highest number of differentially regulated genes encoding proteins involved in general stress response and metabolic adaptation, which the ML prediction models did not capture. In the case of erythromycin, the key resistance gene ermB was not differentially expressed, consistent with the established role of mutations in regulatory regions leading to its constitutive expression. Our study found that vancomycin and doxycycline resistance can be accurately predicted using AMR genes and MGEs, with MGEs posing a particular risk due to their conjugative capacity across different Enterococcus isolates. Accurate prediction of erythromycin and other resistance phenotypes may require additional features based on allelic variation and regulatory-sequence mutations that are not represented in the pan-genome profiling. Complementing ML prediction models with experimental results on the dissemination and expression of resistance determinants helps identify the most consistent and relevant features for model development, bringing ML technology closer to real-world implementation.