Machine learning method for the prediction of Bedaquiline-resistant Mycobacterium tuberculosis

Area of Science:

• Genomics
• Machine Learning
• Drug Resistance

Background:

• Increasing Bedaquiline (BDQ) resistance in Mycobacterium tuberculosis (MTB) poses a significant challenge.
• The molecular basis of BDQ resistance is not fully understood due to undefined resistance loci and target variations.

Purpose of the Study:

• To predict BDQ resistance in MTB using whole-genome sequencing data and machine learning.
• To identify novel drug targets and genetic features associated with BDQ resistance.

Main Methods:

• Applied machine learning models (Multilayer Perceptron, Random Forest) to predict BDQ resistance from MTB whole-genome sequencing data.
• Utilized variant calling format data, including adapter trimming and alignment to the H37Rv reference genome.
• Employed explainable AI (XAI) methods like SHapley Additive exPlanations to interpret model predictions.

Main Results:

• Machine learning models achieved high prediction accuracies (83.60% for MLP, 79.64% for RF).
• Identified 15 new antibiotic-resistant genes by mapping top features to the H37Rv genome, including non-coding regions.
• XAI facilitated the pinpointing of specific mutations linked to BDQ resistance.

Conclusions:

• Machine learning models effectively predict BDQ resistance in MTB.
• Explainable AI enhances understanding of the key genetic features driving BDQ resistance.
• The study reveals potential new targets for combating drug-resistant tuberculosis.