Machine Learning Predicts Accurately Mycobacterium tuberculosis Drug Resistance From Whole Genome Sequencing Data

Area of Science:

• Genomics
• Machine Learning
• Antimicrobial Resistance

Background:

• Tuberculosis (TB) remains a global health challenge, exacerbated by drug-resistant strains.
• Mutations in Mycobacterium tuberculosis genes confer resistance, but knowledge of these mutations is incomplete.
• Whole genome sequencing (WGS) offers a rapid method for characterizing isolates and predicting antimicrobial resistance.

Purpose of the Study:

• To apply machine learning (ML) to WGS data for predicting drug resistance in Mycobacterium tuberculosis.
• To identify novel mutations associated with drug resistance.
• To evaluate the utility of ML in assisting clinical decision-making for TB treatment.

Main Methods:

• Utilized ML models (classification trees, gradient-boosted trees) on 16,688 M. tuberculosis isolates with WGS and drug susceptibility testing (DST) data.
• Modeled resistance to 14 antituberculosis drugs, incorporating "co-occurrent resistance" markers.
• Assessed predictive performance using sensitivity, specificity, and area under the receiver operating characteristic curve (AUC).

Main Results:

• ML models achieved high predictive performance (AUC > 96%) for first-line drugs and multidrug-resistant TB.
• Performance was lower for some third-line drugs (AUC < 85%).
• Gradient-boosted-tree models outperformed classification-tree models; inclusion of co-occurrent markers improved predictions for some drugs.

Conclusions:

• Machine learning is a valuable tool for predicting drug resistance in M. tuberculosis using WGS data.
• This approach can accommodate numerous predictors, aiding in SNP detection and clinical decision-making.
• Discordance between genotypic and phenotypic resistance may stem from DST errors, rare mutations, or non-genomic factors.