Identification of active molecules against Mycobacterium tuberculosis through machine learning

Area of Science:

• Computational chemistry and bioinformatics
• Drug discovery and development
• Machine learning applications in medicine

Background:

• Tuberculosis (TB) remains a global health threat, with drug-resistant strains posing a significant challenge.
• The urgent need for novel anti-TB drug candidates necessitates advanced computational approaches.
• Existing treatments are challenged by extensively drug-resistant tuberculosis (XDR-TB).

Purpose of the Study:

• To develop and evaluate machine learning models for predicting Mycobacterium tuberculosis (Mtb) inhibitors.
• To identify the most effective machine learning algorithms and consensus strategies for Mtb inhibitor prediction.
• To create a publicly accessible webserver for identifying potential TB drug candidates.

Main Methods:

• Four machine learning algorithms (SVM, RF, XGBoost, DNN) were trained using diverse molecular representations.
• Consensus strategies, including stacking, were employed to integrate predictions from multiple models.
• Shapley additive explanations (SHAP) were used to interpret model predictions and identify key molecular descriptors.

Main Results:

• The XGBoost model demonstrated superior individual prediction performance.
• A consensus model stacking RF, XGBoost, and DNN predictions achieved high accuracy (AUC 0.842 training, 0.942 external test).
• The developed ChemTB webserver provides a free tool for Mtb inhibitor detection.

Conclusions:

• Machine learning, particularly consensus modeling, is effective for predicting Mtb inhibitors.
• The ChemTB webserver offers a valuable resource for accelerating TB drug discovery.
• Interpretable AI methods enhance understanding of structure-activity relationships for anti-TB compounds.