INGOT-DR: an interpretable classifier for predicting drug resistance in M. tuberculosis

Area of Science:

• Microbiology
• Computational Biology
• Genetics

Background:

• Predicting drug resistance in Mycobacterium tuberculosis is crucial for effective tuberculosis treatment.
• Current methods for predicting drug resistance lack transparency, accuracy, or flexibility.
• Existing machine learning models often lack interpretability, while interpretable methods may have lower accuracy.

Purpose of the Study:

• To develop a novel, transparent, accurate, and flexible predictive model for bacterial drug resistance.
• To address the limitations of existing methods in predicting drug resistance and identifying its mechanisms.
• To provide a tool that can be optimized for various evaluation metrics simultaneously.

Main Methods:

• A novel technique inspired by group testing and Boolean compressed sensing was developed.
• The approach integrates high predictive accuracy with intrinsic interpretability.
• The method is flexible and can be customized for different evaluation metrics.

Main Results:

• The proposed method demonstrated high or comparable accuracy to common machine learning models for predicting resistance to first- and second-line tuberculosis antibiotics.
• The technique successfully identified gene variants associated with drug resistance.
• The method is intrinsically interpretable and adaptable to various evaluation metrics.

Conclusions:

• The novel technique offers a significant advancement in predicting drug resistance in Mycobacterium tuberculosis.
• The approach provides a transparent, accurate, and flexible alternative to existing methods.
• The developed implementation is publicly available and compatible with Scikit-learn tools.