Rapid antibiotic resistance predictions from genome sequence data for S. aureus and M. tuberculosis

Rapid and accurate detection of antibiotic resistance in pathogens is an urgent need, affecting both patient care and population-scale control. Microbial genome sequencing promises much, but many barriers exist to its routine deployment. Here, we address these challenges, using a de Bruijn graph comparison of clinical isolate and curated knowledge-base to identify species and predict resistance profile, including minor populations. This is implemented in a package, Mykrobe predictor, for S. aureus and M. tuberculosis, running in under three minutes on a laptop from raw data. For S. aureus, we train and validate in 495/471 samples respectively, finding error rates comparable to gold-standard phenotypic methods, with sensitivity/specificity of 99.3%/99.5% across 12 drugs. For M. tuberculosis, we identify species and predict resistance with specificity of 98.5% (training/validating on 1920/1609 samples). Sensitivity of 82.6% is limited by current understanding of genetic mechanisms. We also show that analysis of minor populations increases power to detect phenotypic resistance in second-line drugs without appreciable loss of specificity. Finally, we demonstrate feasibility of an emerging single-molecule sequencing technique.