ArtiDock: Accurate Machine Learning Approach to Protein-Ligand Docking Optimized for High-Throughput Virtual...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Classical protein-ligand docking methods have plateaued in accuracy and performance.
• Machine learning (ML) docking shows promise but faces challenges in accuracy, benchmarking, and pose validity.
• High-throughput virtual screening (HTVS) requires efficient and accurate docking tools.

Purpose of the Study:

• Introduce ArtiDock, an ML-based docking technique optimized for HTVS.
• Develop a robust benchmark for evaluating docking performance in realistic scenarios.
• Compare ArtiDock's accuracy and speed against established classical and AI-based docking methods.

Main Methods:

• Developed ArtiDock, an ML-based protein-ligand docking algorithm.
• Created the PLINDER dataset for pocket-specific rigid docking benchmark.
• Evaluated ArtiDock using PLINDER and PoseX benchmarks against AutoDock, Vina, Glide, and AI docking methods.

Main Results:

• ArtiDock demonstrated 29-38% higher accuracy than classical docking methods (AutoDock, Vina, Glide).
• Achieved competitive accuracy with higher throughput compared to AI docking and cofolding methods on PoseX.
• ArtiDock showed superior performance in challenging cases, including unbound proteins and sites with ions/water.

Conclusions:

• ArtiDock offers a significant advancement in ML-based docking for drug discovery.
• The method provides a favorable accuracy-to-speed trade-off for HTVS applications.
• ArtiDock is a recommended tool for HTVS, particularly for complex docking scenarios.