A generalisable approach to drug susceptibility prediction for M. tuberculosis using machine learning and whole-genome sequencing

Abstract

AbstractRapid and up-to-date drug susceptibility testing is urgently needed to address the threat of multidrug resistant tuberculosis. We developed a composite machine learning system to predict susceptibility from whole-genome sequences for 13 anti-tuberculosis drugs. We trained, validated and externally tested the system, and assessed its performance against a previously validated mutation catalogue, existing molecular assays, and World Health Organization Target Product Profiles. 174,492 phenotypes and 26,328 isolates from 34 countries were studied. The sensitivity of the model was greater than 90% for all drugs except ethionamide, clofazimine and linezolid. Specificity was greater than 95% for all drugs except ethambutol, ethionamide, bedaquiline, delamanid and clofazimine. The machine learning system was more sensitive than the mutation catalogue and molecular assays. For rifampicin-resistant samples, it correctly predicted a pan-susceptible second-line regimen with 98% accuracy. The proposed system can help guide therapy and be updated automatically as new resistance determinants emerge.