Background: Tuberculosis disease, caused by Mycobacterium tuberculosis, is a major public health problem. The emergence of M. tuberculosis strains resistant to existing treatments threatens to derail control efforts. Resistance is mainly conferred by mutations in genes coding for drug targets or converting enzymes, but our knowledge of these mutations is incomplete. Whole genome sequencing (WGS) is an increasingly common approach to rapidly characterize isolates and identify mutations predicting antimicrobial resistance and thereby providing a diagnostic tool to assist clinical decision making.

Methods: We applied machine learning approaches to 16,688 M. tuberculosis isolates that have undergone WGS and laboratory drug-susceptibility testing (DST) across 14 antituberculosis drugs, with 22.5% of samples being multidrug resistant and 2.1% being extensively drug resistant. We used non-parametric classification-tree and gradient-boosted-tree models to predict drug resistance and uncover any associated novel putative mutations. We fitted separate models for each drug, with and without “co-occurrent resistance” markers known to be causing resistance to drugs other than the one of interest. Predictive performance was measured using sensitivity, specificity, and the area under the receiver operating characteristic curve, assuming DST results as the gold standard.

Results: The predictive performance was highest for resistance to first-line drugs, amikacin, kanamycin, ciprofloxacin, moxifloxacin, and multidrug-resistant tuberculosis (area under the receiver operating characteristic curve above 96%), and lowest for third-line drugs such as D-cycloserine and Para-aminosalisylic acid (area under the curve below 85%). The inclusion of co-occurrent resistance markers led to improved performance for some drugs and superior results when compared to similar models in other large-scale studies, which had smaller sample sizes. Overall, the gradient-boosted-tree models performed better than the classification-tree models. The mutation-rank analysis detected no new single nucleotide polymorphisms linked to drug resistance. Discordance between DST and genotypically inferred resistance may be explained by DST errors, novel rare mutations, hetero-resistance, and nongenomic drivers such as efflux-pump upregulation.

Conclusion: Our work demonstrates the utility of machine learning as a flexible approach to drug resistance prediction that is able to accommodate a much larger number of predictors and to summarize their predictive ability, thus assisting clinical decision making and single nucleotide polymorphism detection in an era of increasing WGS data generation.