Improved accuracy of antibiotic resistance determination from Enterococcus faecium whole-genome sequences: a population-based genotypic and phenotypic analysis

Background: DNA sequencing could become an alternative to in vitro antibiotic susceptibility testing (AST) methods for determining antibiotic resistance by detecting genetic determinants associated with decreased antibiotic susceptibility. This is exemplified by the success in predicting antibiotic resistance (ABR) from the genomes of bacteria such as Mycobacterium tuberculosis, Staphylococcus aureus or Streptococcus pneumoniae. Here, we aimed to assess and improve the accuracy of ABR determination from Enterococcus faecium genomes for diagnosis and surveillance purposes. Methods: We conducted a literature search to compile a catalogue of genes and mutations predictive of antibiotic resistance in E. faecium. We evaluated the diagnostic accuracy of this database to determine susceptibility to 12 different antibiotics using a large and diverse population of 4,382 E. faecium isolates with available whole-genome sequences and in vitro culture-based AST phenotypes. We kept isolates tested with broth microdilution, Vitek2 and disk diffusion; and antibiotics with at least 50 susceptible and 50 resistant isolates. Phenotypic resistance was derived from raw minimum inhibitory concentrations (MICs) and measured inhibition diameters; and harmonised using the breakpoints set by the European Committee on Antimicrobial Susceptibility Testing (EUCAST). A bioinformatics pipeline was developed to process raw sequencing reads, identify ABR genetic determinants and report genotypic resistance. We used our curated database, as well as ResFinder, AMRFinderPlus and LRE-Finder, to assess the accuracy of genotypic predictions against phenotypic resistance. Results: We curated a catalogue of 228 genetic markers involved in resistance to 12 different antibiotics in E. faecium. Very accurate genotypic predictions could be obtained for ampicillin, ciprofloxacin, vancomycin and linezolid resistance, with sensitivity and specificity well above 95%. High sensitivity (above 90%) was obtained for tetracyclines, teicoplanin, and high-level aminoglycoside resistance, although at lower specificity (60 to 90%). Sensitivity was expectedly lower for daptomycin (73.6%) and tigecycline (38.3%). Compared to available ABR databases and bioinformatic tools, our curated database produced comparable accuracy in the detection of ciprofloxacin (98.0 and 98.8% sensitivity and specificity, respectively), linezolid (100 and 98.3%), and high-level streptomycin (97.7 and 88.6%) and gentamicin (96.8 and 82.2%) resistance; improved sensitivity for ampicillin (99.7%), tigecycline (38.3%), daptomycin (73.6%) and quinupristin-dalfopristin resistance (89.2%); and improved specificity for ampicillin (97.9%), vancomycin (98.8%), teicoplanin (93.3%) and tetracycline (71.0%) resistance. In a validation dataset of 382 isolates, comparable or improved diagnostic accuracies were also achieved. Interpretation: This work represents the largest published effort to date at evaluating the accuracy of antibiotic susceptibility predictions from E. faecium genomes. The results and resources made available here will facilitate the adoption of whole-genome sequencing as a tool for the diagnosis and surveillance of AMR in E. faecium. A complete characterization of the genetic basis of resistance to last-line antibiotics, and the mechanisms mediating antibiotic resistance silencing, are needed to close the remaining sensitivity and specificity gaps in genotypic predictions.Funding: Wellcome Trust, UK Department of Health, British Society for Antimicrobial Chemotherapy, Academy of Medical Sciences and the Health Foundation, Medical Research Council Newton Fund, Vietnamese Ministry of Science and Technology, and European Society of Clinical Microbiology and Infectious Disease