Machine learning-based prediction of antimicrobial resistance and identification of AMR-related SNPs in Mycobacterium tuberculosis.

BACKGROUND

Mycobacterium tuberculosis (MTB) is a human-specific pathogen that primarily infects humans, causing tuberculosis (TB). Antimicrobial resistance (AMR) in MTB presents a formidable challenge to global health. The employment of machine learning on whole-genome sequencing data (WGS) presents significant potential for uncovering the genomic mechanisms underlying drug resistance in MTB.

METHODS

We used 18 binary matrices, each consisting of genotypes and antimicrobial susceptibility testing phenotypes from a specific MTB-antimicrobial dataset. By constructing training and test datasets on all SNPs, intersected SNPs, and randomly generated SNPs, we developed a Machine learning (ML) framework using twelve different algorithms. Then, we compared the performances of the various ML models and used the SHapley Additive exPlanations (SHAP) framework to decipher why and how decisions are made within the optimal algorithm. Lastly, we applied the models to predict the resistance phenotype to rifampicin (RIF) and isoniazid (INH) in the additional independent MTB isolate datasets from India and Israel.

RESULTS

In our study, the Gradient Boosting Classifier (GBC) model was the best in terms of correctly identified percentages (97.28%, 96.06%, 94.19%, and 92.81% for the four first-line drugs, RIF, INH, pyrazinamide, and ethambutol respectively). By estimating the contributions of AMR-related SNPs by SHAP values, we found that position 761,155 (rpoB_p.Ser450), 2,155,168 (katG_p.Ser315) rank top in RIF and INH, their higher values (1 for alternative allele) tend to predict the resistance trait for these two drugs. In addition, the best model GBC generalizes well in predicting the resistance phenotypes for RIF and INH in the external independent MTB isolate datasets from India and Israel.

CONCLUSIONS

This study integrates ML methods into antimicrobial resistance research, develops a framework for predicting resistance phenotypes, and explores AMR-related SNPs in MTB. Quantifying the important SNPs' contribution to model decisions makes the ML algorithmic process more transparent, interpretable enabling and enables clinical practice.