Prediction of pyrazinamide resistance inMycobacterium tuberculosisusing structure-based machine learning approaches

Abstract

Pyrazinamide is one of four first-line antibiotics currently used to treat tuberculosis and has been included in newer treatment regimens undergoing clinical trials due to its unique sterilizing effects and synergy with newer drugs. However, phenotypic antibiotic susceptibility testing for pyrazinamide is problematic. Resistance to pyrazinamide is primarily driven by genetic variation in pncA, which encodes PncA, an enzyme that converts pyrazinamide into its active form. We curated a derivation dataset of 291 non-redundant, missense amino acid mutations in PncA with associated high-confidence phenotypes from studies of clinical isolates and in vitro/in vivo screening studies and then trained machine learning models to predict pyrazinamide resistance based on sequence-and structure-based features of each missense mutation. The clinical relevance of the models was tested by predicting the binary resistance phenotype of 2,292 clinical isolates harboring missense mutations in PncA to pyrazinamide. The probabilities of resistance predicted by the model were also compared with in vitro pyrazinamide minimum inhibitory concentrations of 27 isolates to determine whether the machine learning model could predict the degree of resistance. Finally, we predicted the effect on pyrazinamide resistance of the remaining 814 possible missense mutations caused by single nucleotide polymorphisms in PncA that have not yet been observed in public databases. Overall, this work offers an approach to improve the sensitivity and specificity of pyrazinamide resistance prediction in genetics-based clinical microbiology workflows for tuberculosis, highlights novel mutations for future biochemical investigation, and is a proof of concept for using this approach in other drugs such as bedaquiline.

Deciphering Bedaquiline and Clofazimine Resistance in Tuberculosis: An Evolutionary Medicine Approach

“…Another method for prospectively identifying mutations that confer resistance is structure-based modeling, which has recently been combined with machine learning methods to design algorithms for predicting resistance to rifampicin, isoniazid, and pyrazinamide [50][51][52][53][54] . These approaches rely on the fact that resistance-conferring mutations are often located in drug binding pockets or active sites and have distinct biophysical consequences as compared to their susceptible counterparts 55 .…”

How To Optimally Combine Genotypic and Phenotypic Drug Susceptibility Testing Methods for Pyrazinamide

“…We set out to explore the limitations of our original approach in light of the most important studies in this area. In particular, we aimed to increase the limited sensitivity by including types of data that were beyond the scope of the original review (e.g., results from engineered strains, quantitative pDST results, and interpretative approaches based on alternative statistical methods) and six expert rules (see Supplementary methods in the supplemental material) ( 3 , 8 – 14 ).…”

How quickly can we predict trimethoprim resistance using alchemical free energy methods?

“…Although we have focused primarily on calculating the effect of the mutation on the binding free energy of the antibiotic, there are a range of other methods that could be brought to bear. Machine learning methods, using genetic, structural and/or chemical features, are likely to be sufficiently accurate and also fast [11,25,26]. Such methods could be used to screen out mutations that have no effect on an antibiotic, leaving only the marginal cases for computationally more intensive approaches such as proposed here.…”