Prediction of antimicrobial susceptibility of pneumococci based on whole-genome sequencing data: a direct comparison of two genomic tools to conventional antimicrobial susceptibility testing

Determination of antimicrobial resistance (AMR) in pneumococcal isolates is important for surveillance purposes and in a clinical context. Antimicrobial susceptibility testing (AST) of pneumococci is complicated by the need for exact minimal inhibitory concentrations (MICs) of beta-lactam antibiotics. Two next-generation sequencing (NGS) analysis tools have implemented the prediction of AMR in their analysis workflow, including the prediction of MICs: Pathogenwatch (https://pathogen.watch/) and AREScloud (OpGen). The performance of these tools in comparison to phenotypic AST following EUCAST guidelines is unknown. A total of 538 isolates were used to compare both tools with phenotypic AST for penicillin, amoxicillin, cefotaxime/ceftriaxone, erythromycin, trimethoprim-sulfamethoxazole, and tetracycline. Disk diffusion was performed for all isolates, and broth microdilution was performed for isolates with reduced beta-lactam susceptibility. Demultiplexed FASTQ files from Illumina sequencing, covering the whole genome of pneumococci, were used as input for the NGS tools. Categorical agreement (CA), major error (ME), and very major error (VME) rates were calculated. For beta-lactam antibiotics, CA was high (>94%) associated with none or only one ME and VME (93% for predictions by AREScloud, while for Pathogenwatch, this ranged around 88%. For trimethoprim-sulfamethoxazole, CA was for both tools