Development of machine learning models to identify potentially active compounds against tuberculosis.

Area of Science:

• Computational chemistry and bioinformatics
• Drug discovery and development
• Machine learning in medicinal chemistry

Background:

• Tuberculosis (TB) remains a global health crisis, exacerbated by multi-drug-resistant (MDR) and extensively drug-resistant (XDR) strains.
• The urgent need for novel anti-TB drug candidates necessitates efficient lead identification strategies.
• Machine learning offers powerful tools for accelerating the drug discovery pipeline.

Purpose of the Study:

• To explore the application of machine learning algorithms for identifying potential anti-TB compounds.
• To develop and evaluate classification models for predicting compound activity against TB targets.
• To interpret the contribution of molecular descriptors to bioactivity using explainable AI methods.

Main Methods:

• A dataset of 23,791 molecules from the ChEMBL database was utilized.
• 103 classification models were built using six molecular representations (RDKitDes, MACCSFP, MorganFP, PairsFP, PubChemFP, RDKitFP).
• Seven machine learning algorithms (RF, XGBoost, DT, KNN, GNB, LR, ANN) were employed and evaluated using tenfold cross-validation and AUC metrics, with SHAP for interpretability.

Main Results:

• Multiple machine learning models demonstrated predictive capability for identifying anti-TB compounds.
• The study established a framework for rapid lead identification using diverse molecular representations and algorithms.
• SHAP analysis provided insights into key molecular features driving compound activity against TB targets.

Conclusions:

• Machine learning approaches are effective in accelerating the identification of novel drug candidates for tuberculosis.
• The developed models and methodologies can significantly aid in combating drug-resistant TB.
• Integrating molecular representations, ML algorithms, and explainable AI enhances the efficiency of drug discovery.