Prediction of antibacterial compounds by machine learning approaches

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Machine learning

Background:

• Quantitative structure-activity relationship (QSAR) and machine learning (ML) methods show promise for predicting antibacterial compound activity.
• Further investigation into diverse ML methods, robust feature selection, and rigorous model evaluation is needed.

Purpose of the Study:

• To evaluate three ML methods (SVC, k-NN, C4.5 decision tree) for predicting antibacterial activity.
• To identify optimal molecular descriptors using feature selection.
• To rigorously assess model performance via cross-validation and independent testing.

Main Methods:

• Trained and tested Support Vector Classification (SVC), k-nearest neighbor (k-NN), and C4.5 decision tree models.
• Utilized a feature selection method to identify representative molecular descriptors from a large pool.
• Evaluated model performance using 5-fold cross-validation and an independent test set.

Main Results:

• SVC achieved high prediction accuracies: 96.66% for antibacterial and 99.50% for non-antibacterial compounds during cross-validation.
• Independent testing showed SVC accuracies of 98.15% (antibacterial) and 98.02% (non-antibacterial).
• SVC outperformed k-NN and C4.5 decision tree models and showed slight improvements over existing ML/QSAR models.

Conclusions:

• Machine learning, particularly SVC, is a valuable tool for predicting antibacterial activity.
• The developed SVC model demonstrates potential for accelerating the discovery of novel antibacterial agents.
• Rigorous evaluation and feature selection enhance the reliability of ML-based drug discovery models.