Application of machine learning techniques to tuberculosis drug resistance analysis

Area of Science:

• * Computational biology and bioinformatics
• * Infectious disease diagnostics
• * Antimicrobial resistance surveillance

Background:

• * Timely identification of Mycobacterium tuberculosis (MTB) drug resistance is crucial for reducing mortality and preventing further antibiotic resistance.
• * Machine learning (ML) has shown promise in predicting MTB resistance and identifying resistance markers.
• * Previous ML applications lacked validation on large, diverse, multi-center cohorts.

Purpose of the Study:

• * To validate and compare various machine learning classifiers and dimension reduction techniques for MTB drug resistance prediction.
• * To assess the performance of ML models on a large, multi-center global cohort of MTB isolates.
• * To identify potential novel resistance and susceptible markers through mutation ranking.

Main Methods:

• * Development and comparison of several machine learning classifiers and linear dimension reduction techniques.
• * Analysis of a cohort comprising 13,402 MTB isolates from 16 countries across 6 continents.
• * Testing of 11 different anti-tuberculosis drugs for resistance prediction.

Main Results:

• * Machine learning classifiers significantly improved the area under the curve (AUC) for predicting resistance to pyrazinamide, ciprofloxacin, and ofloxacin compared to conventional methods.
• * Logistic regression and gradient tree boosting demonstrated superior performance among the tested classifiers.
• * Combining ML classifiers with sparse principal component analysis or non-negative matrix factorization further enhanced predictive performance (F1-score and AUC) for several drugs.

Conclusions:

• * Machine learning provides a robust and accurate approach for predicting drug resistance in large, diverse tuberculosis datasets.
• * The validated ML models offer improved diagnostic capabilities compared to conventional methods.
• * Mutation ranking analysis suggests the potential for discovering new genetic markers associated with drug resistance or susceptibility.