Accelerating AutoDock4 with GPUs and Gradient-Based Local Search.

Area of Science:

• Computational chemistry
• Molecular modeling
• Drug discovery

Background:

• AutoDock4 is a widely used molecular docking program.
• Its physics-inspired scoring function is valuable for drug discovery.
• AutoDock4's long execution times limit its use in large-scale screenings.

Purpose of the Study:

• To accelerate AutoDock4's execution time for large-scale virtual screening.
• To improve the search efficiency of the docking process.

Main Methods:

• Developed AutoDock-GPU, an OpenCL implementation of AutoDock4 leveraging GPU hardware.
• Introduced ADADELTA, a gradient-based local search method.
• Implemented an improved Solis-Wets random optimizer.

Main Results:

• AutoDock-GPU reduces docking runtime by up to 350-fold compared to single-threaded AutoDock4.
• New local search algorithms significantly decrease scoring function calls.
• Achieved substantial improvements in docking throughput and search efficiency.

Conclusions:

• AutoDock-GPU significantly enhances molecular docking performance.
• The optimized algorithms facilitate large-scale virtual screening using the AutoDock4 scoring function.
• This work accelerates drug discovery pipelines.