DSDP: A Blind Docking Strategy Accelerated by GPUs.

Area of Science:

• Computational chemistry
• Drug discovery and development
• Bioinformatics

Background:

• Virtual screening, particularly molecular docking, is crucial for identifying drug candidates.
• Traditional docking methods are time-consuming and struggle with blind docking accuracy.
• Machine learning docking methods offer speed but often lack sufficient accuracy.

Purpose of the Study:

• To develop a novel method, Deep Site and Docking Pose (DSDP), that integrates traditional and machine learning techniques.
• To enhance the performance of blind docking in drug discovery.
• To improve the accuracy and efficiency of identifying ligand binding sites and poses.

Main Methods:

• DSDP predicts protein binding sites, providing accurate search shapes and initial ligand positions.
• It utilizes a scoring function and a modified AutoDock Vina search strategy.
• GPU acceleration is employed to speed up conformational sampling.

Main Results:

• DSDP achieved a 29.8% top-1 success rate in blind docking on a challenging dataset (RMSD < 2 Å) in 1.2 seconds.
• On the DUD-E dataset, DSDP showed a 57.2% top-1 success rate with 0.8 seconds per system.
• Performance on the PDBBind dataset yielded a 41.8% top-1 success rate in 1.0 second per system.

Conclusions:

• DSDP significantly outperforms state-of-the-art methods in blind docking accuracy and speed.
• The method effectively predicts binding sites and optimizes ligand pose sampling.
• DSDP represents a substantial advancement for virtual screening in drug discovery.