Genomic prediction of tuberculosis drug-resistance: benchmarking existing databases and prediction algorithms

BACKGROUND: It is possible to predict whether a tuberculosis (TB) patient will fail to respond to specific antibiotics by sequencing the genome of the infecting Mycobacterium tuberculosis (Mtb) and observing whether the pathogen carries specific mutations at drug-resistance sites. This advancement has led to the collation of TB databases such as PATRIC and ReSeqTB that possess both whole genome sequences and drug resistance phenotypes of infecting Mtb isolates. Bioinformatics tools have also been developed to predict drug resistance from whole genome sequencing (WGS) data. Here, we evaluate the performance of four popular tools (TBProfiler, MyKrobe, KvarQ, PhyResSE) with 6746 isolates compiled from publicly available databases, and subsequently identify highly probable phenotyping errors in the databases by genetically predicting the drug phenotypes using all four software. RESULTS: Our results show that these bioinformatics tools generally perform well in predicting the resistance status for two key first-line agents (isoniazid, rifampicin), but the accuracy is lower for second-line injectables and fluoroquinolones. The error rates in the databases are also non-trivial, reaching as high as 31.1% for prothionamide, and that phenotypes from ReSeqTB are more susceptible to errors. CONCLUSIONS: The good performance of the automated software for drug resistance prediction from TB WGS data shown in this study further substantiates the usefulness and promise of utilising genetic data to accurately profile TB drug resistance, thereby reducing misdiagnoses arising from error-prone culture-based drug susceptibility testing. ELECTRONIC SUPPLEMENTARY MATERIAL: The online version of this article (10.1186/s12859-019-2658-z) contains supplementary material, which is available to authorized users.