Machine learning-based analysis of drug resistance mutations in Mycobacterium tuberculosis

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Not reported in abstract — specific accuracy, AUC-ROC, or F1-score values not provided

• Primary Outcome: Random Forest classifier demonstrated the highest predictive performance among eight supervised ML algorithms for classifying drug-resistance type (Rifampicin Resistance, Isoniazid Resistance, MDR-TB, Pre-XDR-TB) from M. tuberculosis mutation profiles in 3,065 cases

• Nnt Or Sensitivity: Sensitivity and specificity not reported in abstract. Ten novel candidate resistance mutations identified: rpoB-I480T, rpoC-G332R, rpoC-L527V, gyrA-D94V, KatG-G99E, KatG-A106V, KatG-W191R, KatG-W328C, KatG-T380I, KatG-M420T — none currently listed in WHO TB mutation catalogues

• Confidence Interval: Not reported

Clinical Application

The model relies on genomic sequencing data from M. tuberculosis isolates, which is increasingly available in high-income settings and reference laboratories but remains limited in many high-burden, resource-constrained settings. Implementation would require integration with existing TB sequencing pipelines and clinical decision support infrastructure. The ten novel mutations identified require experimental validation before any clinical use. In Australia, TB management is guided by the Communicable Diseases Network Australia (CDNA) guidelines and RACGP recommendations. The National Mycobacterium Reference Laboratory network performs phenotypic drug-susceptibility testing and whole-genome sequencing for complex cases. The TGA has not approved any ML-based TB resistance prediction tool. PBS-listed first-line TB therapy (Rifampicin, Isoniazid, Pyrazinamide, Ethambutol) and second-line agents for MDR-TB are available through state health departments. This computational study does not directly change current Australian practice but may inform future genomic surveillance tools. The predominantly overseas-born Australian TB case profile means that strain lineage diversity (particularly Lineages 2 and 4) is relevant, and external validation in Australian isolates would be required before any local implementation. Patients with confirmed or suspected drug-resistant M. tuberculosis infection undergoing whole-genome sequencing or targeted mutation profiling, particularly in settings with high MDR-TB or Pre-XDR-TB burden

Abstract

Tuberculosis is a deadly airborne disease caused by Mycobacterium tuberculosis. Drug-resistant tuberculosis presents significant challenges for treatment and control of the disease. Resistant strains of Mycobacterium tuberculosis arise from specific mutations in the bacterium. Identification and characterization of resistance-associated mutations are crucial for effective treatment strategies because the first- and second-line drugs for the disease target distinct genes in the bacterium and serve different purposes. Our study developed a machine learning prediction model to analyze mutations across multiple drug-resistance types. The proposed framework predicts drug-resistance mutations across four drug-resistance types, including Rifampicin Resistance, Isoniazid Resistance, Multidrug Resistance, and Pre-extensively Drug-Resistant tuberculosis. The NIAID-NIH TB portal is a publicly available dataset of tuberculosis patients, including drug-resistance information. Our study analyzed 3,065 cases of drug-resistant TB. Eight supervised ML algorithms were implemented for the study. A Random Forest classifier with 10-fold cross-validation shows higher predictive performance than the other seven algorithms considered for further analysis. Significant drug resistance mutations were identified using SHapley Additive exPlanations feature importance. The World Health Organisation mutation catalogues, considered the gold standard for drug-resistant mutations, were used to evaluate prediction results. Mutations not reported in the WHO catalogues were identified during the post-prediction comparative analysis stage, as they may represent potential resistance-conferring markers warranting further investigation, including structural and functional validation or experimental validation. The mutations are rpoB-I480T, rpoC-G332R, L527V, gyrA-D94V, KatG-G99E, A106V, W191R, W328C, T380I, and M420T. The study further checks the stability and pathogenicity of the mutations using computational tools, including I-Mutant 2.0 and PredictSNP. The findings added more clarity and further evidence for the significance of the mutation, based on its contribution to drug resistance.

References

1. 1.Thankamani, A., Biji, C. L., & Doss, C. G. P. (2026). Machine learning-based analysis of drug resistance mutations in Mycobacterium tuberculosis. PLOS ONE. https://doi.org/10.1371/journal.pone.0352863