A Hybrid Docking and Machine Learning Approach to Enhance the Performance of Virtual Screening Carried out on...

Area of Science:

• Computational chemistry and drug discovery.
• Structural biology and molecular modeling.

Background:

• Protein-protein interactions (PPIs) are crucial in cellular functions and disease.
• Targeting PPIs with small molecules is a significant challenge in drug development.
• Previous work identified DLIGAND2 as a useful scoring tool for virtual screening of PPIs.

Purpose of the Study:

• To evaluate novel virtual screening strategies for protein-protein interaction (PPI) targets.
• To assess the utility of solvent accessible surface area (SASA) descriptors derived from docking poses.
• To investigate the performance of machine learning (ML) models in conjunction with SASA descriptors for virtual screening.

Main Methods:

• Rescoring of Surflex docking poses using GOLD scoring functions and consensus scoring.
• Derivation of protein and ligand SASA descriptors in bound and unbound states.
• Training and validation of eight ML models (including random forest and neural networks) using SASA descriptors.

Main Results:

• SASA descriptors outperformed standard scoring functions (Surflex and GOLD) in virtual screening performance and early recovery.
• Machine learning models significantly improved enrichment factors, achieving up to a seven-fold increase.
• Neural network and random forest-based ML models demonstrated superior performance for PPI datasets.

Conclusions:

• Docking-pose derived SASA descriptors are valuable for structure-based virtual screening.
• Machine learning models incorporating SASA descriptors offer robust and attractive tools for hit-finding in PPI inhibitor design.
• This approach enhances the rational design of small-molecule PPI inhibitors.