Integrating Machine Learning-Based Pose Sampling with Established Scoring Functions for Virtual Screening.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in cheminformatics

Background:

• Physics-based docking is a traditional method for structure-based virtual screening (VS).
• Machine learning (ML) approaches are emerging as powerful tools to enhance VS technologies.

Purpose of the Study:

• To integrate DiffDock-L, an ML-based pose sampling method, into VS workflows.
• To evaluate the performance of DiffDock-L when combined with Vina, Gnina, and RTMScore scoring functions.
• To compare the integrated approach with traditional physics-based docking methods like AutoDock Vina.

Main Methods:

• Utilized DiffDock-L for pose sampling in virtual screening.
• Integrated DiffDock-L with Vina, Gnina, and RTMScore scoring functions.
• Assessed performance on the DUDE-Z benchmark dataset using cross-docking scenarios.

Main Results:

• DiffDock-L demonstrated competitive virtual screening performance and pose sampling quality.
• The ML-based method generated physically plausible and biologically relevant poses.
• The choice of scoring function was found to significantly impact virtual screening success.

Conclusions:

• DiffDock-L presents a viable alternative to traditional physics-based docking algorithms for VS.
• The integration of ML methods holds significant potential for advancing virtual screening technologies.
• Optimizing scoring function selection is crucial for maximizing the effectiveness of ML-based VS approaches.