ccsb-scripps/AutoDock-GPU

What it solves

AutoDock-GPU accelerates the process of molecular docking—predicting how a small molecule (ligand) binds to a receptor protein. It solves the computational bottleneck of the original serial AutoDock 4.2 by leveraging GPUs and other accelerators to process multiple ligand-receptor poses in parallel, significantly reducing the time required for virtual screenings.

How it works

The software is an OpenCL and CUDA accelerated version of AutoDock 4.2.6. It uses a Lamarckian Genetic Algorithm (LGA) that is "embarrassingly parallelizable," allowing it to distribute the processing of poses across multiple compute units. It supports various target platforms including GPUs (Nvidia, AMD, Intel), multicore CPUs, and FPGAs. To improve search efficiency, it implements gradient-based local search methods like ADADELTA and an improved Solis-Wets algorithm.

Who it’s for

It is designed for computational chemists, biologists, and drug discovery researchers who need to perform high-throughput virtual screenings of large ligand libraries against protein receptors.

Highlights

• Significant Speedups: Achieves up to 56x speedup on GPUs and 4x on quad-core CPUs compared to the original serial AutoDock 4.2.
• Broad Hardware Support: Compatible with CUDA and OpenCL, enabling it to run on a wide range of GPU and CPU architectures.
• Batched Pipeline: Includes a dedicated pipeline for running virtual screenings on the same receptor across multiple ligands.
• Advanced Search: Features gradient-based local search and an automatic stopping criterion (AutoStop) to prevent wasting computation once convergence is reached.