Machine learning models for whole genome based prediction of drug resistance in Mycobacterium tuberculosis: a systematic review

Abstract

Tuberculosis remains a leading infectious disease globally, with drug-resistant strains posing significant clinical challenges. While whole-genome sequencing enables comprehensive resistance profiling, traditional rule-based interpretation systems struggle with complex genotype-phenotype relationships, rare variants, and epistatic interactions. Machine learning approaches offer potential to leverage high-dimensional genomic data for improved resistance prediction beyond current catalogue-based methods. This systematic review evaluated machine learning approaches for predicting M. tuberculosis drug resistance using whole-genome sequencing data. We analyzed 15 studies encompassing 20 distinct models, examining algorithmic approaches, feature engineering strategies, bioinformatics pipelines, and validation methodologies. Performance was assessed using sensitivity, specificity, accuracy, and area under the ROC curve for first-line anti-tuberculosis drugs. Gradient boosting methods and attention-based neural networks demonstrated strong predictive performance. Among models reporting sensitivity values for isoniazid, 13 of 18 achieved values of > = 90%; 16 of 17 models reporting rifampicin sensitivity achieved > = 90%. The Hierarchical Attention Neural Network with Task Transfer (HANN-TT) model achieved high discriminatory ability (AUC: 97.9% for isoniazid, 99.1% for rifampicin). Pyrazinamide prediction proved most challenging, with sensitivity varying widely across models (56–98%). The whole-genome XGBoost (WG-XGB) approach demonstrated substantially improved pyrazinamide prediction on the BV-BRC dataset (95% sensitivity, 99% specificity). Machine learning models represent a promising and increasingly robust approach for Whole genome sequences (WGS) based tuberculosis resistance prediction, with gradient boosting and deep learning architectures demonstrating high diagnostic performance for well-characterized drugs. Success depends critically on comprehensive feature engineering, standardized bioinformatics pipelines, and rigorous external validation, including comparative benchmarking against existing clinical tools. Future progress requires formal risk-of-bias assessment, pre-registered protocols, and prospective clinical validation before deployment in diagnostic settings.