Predicting MmpR-based bedaquiline resistance using sequence- and structure-based features

Abstract Accurate molecular detection of resistance to bedaquiline, a core drug for the treatment of drug resistant tuberculosis, remains challenging. In this study, we investigated whether three sequence- and nine structure-based features describing the impact of Rv0678 variants on the MmpR transcriptional repressor could predict the bedaquiline phenotype of isolates containing Rv0678 variants. Paired genotypic and phenotypic data was used to train a binary random forest classifier. The mean value of the individual features was similar for resistant and susceptible variants (p≥0.05). Leave one out cross validation showed predictability of bedaquiline resistance from Rv0678 variants when using a binary classifier with a combination of sequence and structural features (ROC AUC = 0.766; f1 score = 0.746), but the performance was too low for clinically use. Evolutionary conservation of the affected residue was the most important individual feature (mean decrease in impurity = 0.226 ± 0.003) to discriminate bedaquiline resistance from susceptibility. Prediction of bedaquiline resistance was only possible when restricting the data to missense variants, as the selected features could not be applied to Rv0678 insertions and deletions. Additionally, prediction of bedaquiline resistance was only possible when restricting the data to isolates for which the phenotype was determined using Mycobacterial Growth Indicator Tubes, suggesting possible misclassification of the bedaquiline phenotype by other methods. The results of this study suggest that structural features describing Rv0678 missense variants could be used to predict bedaquiline resistance, albeit not yet at the performance level required for clinical practice. Author summary Guidelines by the World Health Organization indicate that antibiotic resistant tuberculosis should be treated with bedaquiline, a newly discovered antibiotic. However, resistance to bedaquiline has already emerged and is spreading rapidly. More than 500 different variants have been discovered in the bedaquiline resistance genes in Mycobacterium tuberculosis. It is not known which DNA variants specifically drive bedaquiline resistance due to insufficiency of data. Here, we used a machine learning technique to predict whether a variant in the Rv0678 gene confers bedaquiline resistance. To this end, for each variant we defined features that describe the impact of the variant on the MmpR (encoded by the Rv0678 gene) protein structure. These features were then fed into a classification algorithm to predict bedaquiline resistance. We successfully developed a bedaquiline resistance prediction model, although performance was limited. We also found that accurate predictions were only possible when restricting the data to samples that were tested on the MGIT platform, a World Health Organization endorsed method to test for bedaquiline resistance. Our study shows that predicting bedaquiline resistance from protein structural features is feasible. Furthermore, our study provides new insights into phenotypic testing platforms to assess bedaquiline resistance.