Ensemble learning from ensemble docking: revisiting the optimum ensemble size problem

Area of Science:

• Computational chemistry and structural biology
• Machine learning applications in drug discovery

Background:

• Protein flexibility is a key challenge in accurate protein-ligand binding affinity prediction.
• Existing ensemble docking methods face limitations in selecting optimal receptor conformations, leading to computational costs and false positives.

Purpose of the Study:

• To develop a novel strategy combining ensemble learning and ensemble docking to address receptor flexibility in affinity prediction.
• To rank protein conformations based on their importance for model accuracy and improve the efficiency of drug screening.

Main Methods:

• Utilized available X-ray structures of cyclin-dependent kinase 2 (CDK2) as an initial receptor ensemble.
• Employed graph-based redundancy removal for efficient and objective selection of representative receptor conformations.
• Applied random forest ensemble learning to energetic features from docked poses to predict binding affinity.

Main Results:

• A few key receptor conformations were identified as sufficient for achieving high accuracy (1 kcal/mol) in affinity prediction.
• The proposed method demonstrated significant improvements in early enrichment power compared to standard ensemble docking.
• Machine learning effectively selected important experimental conformers, enhancing prediction accuracy while managing computational resources.

Conclusions:

• The developed strategy offers a clear approach for machine learning to select crucial receptor conformations for improved protein-ligand affinity prediction.
• This method enhances the accuracy and efficiency of virtual screening, providing valuable insights for receptor-specific docking-rescoring strategies.