Machine learning-based approach for identification of new resistance associated mutations from whole genome sequences of Mycobacterium tuberculosis.

Abstract

Motivation: Currently available methods for the prediction of genotypic drug resistance in Mycobacterium tuberculosis utilize information on known markers of drug resistance. Hence, machine learning approaches are needed that can discover new resistance markers.

Results: Whole genome sequences with known phenotypic drug resistance profiles have been utilized to train XGBoost and ANN classifiers for 5 first-line and 8 second-line tuberculosis drugs. Benchmarking on a completely independent dataset from CRyPTIC database revealed that our method has high sensitivity (90%-95%) and specificity (94%-99%) for five first-line drugs and robust performance for six second-line drugs with a sensitivity of 77%-89% at over 95% specificity. An explainable AI method, SHapley Additive exPlanations, has successfully identified resistance mutations for each drug in a completely automated way. This approach could not only identify known resistance associated mutations in agreement with the WHO mutation catalogue, but also predicted >100 other potential resistance associated mutations for 13 antibiotics in new genes outside the known resistance loci. Identification of new resistance markers opens up the opportunity for the discovery of novel mechanisms of drug resistance.

Availability and implementation: Our prediction method has been implemented as TB-AMRpred webserver and command line tool, available freely at http://www.nii.ac.in/TB-AMRpred.html and https://github.com/Ankitapal1995/TB-AMRprd.