CoBdock-2: enhancing blind docking performance through hybrid feature selection combining ensemble and multimodel...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in bioinformatics

Background:

• Accurate identification of orthosteric binding sites and prediction of small molecule affinities are crucial for virtual screening.
• Traditional blind docking methods face challenges due to large search spaces, while cavity detection-guided docking relies heavily on the quality of cavity detection tools.

Purpose of the Study:

• To develop an improved machine learning-based blind docking method, CoBDock-2, that enhances binding site and ligand pose prediction accuracy.
• To identify key molecular characteristics of orthosteric binding sites through advanced ensemble feature selection.

Main Methods:

• CoBDock-2 extracts 1D numerical representations from protein, ligand, and interaction structural features.
• It employs advanced ensemble feature selection techniques, evaluating 21 methods across 9,598 features.
• The method integrates molecular docking and cavity detection results for enhanced prediction.

Main Results:

• CoBDock-2 achieved 77% binding site identification accuracy and 55% ligand pose prediction accuracy (RMSD ≤ 2 Å).
• It demonstrated a 19% reduction in mean distance to ground truth ligands and an 18.5% decrease in mean pose RMSD.
• The Weighted Hybrid Feature Selection variant further increased binding site accuracy to 79.8%, with significant statistical improvements (p < 0.05).

Conclusions:

• CoBDock-2 significantly improves binding site and pose prediction accuracy compared to previous methods.
• The enhanced reliability and generalizability of CoBDock-2 are highlighted by reduced prediction variability.
• CoBDock-2 shows promise as a robust tool for virtual screening and drug discovery, positioning it against modern deep learning strategies.