DSDPFlex: An Improved Flexible-Receptor Docking Method with GPU Acceleration

Virtual screening constitutes an indispensable phase in the pharmaceutical industry. Among the different methods, the most popular docking strategy, semi-flexible docking (rigid-receptor and flexible-ligand model), ignores the conformational changes of proteins. When the binding interaction between ligands and proteins adheres to an "induced-fit" paradigm, fully flexible methods such as molecular dynamics simulation are needed, but they can be computationally demanding. In an effort to strike a balance between computational speed and precision, flexible docking techniques, as exemplified by AutoDock Vina and AutoDockFR, treat selected protein side chains as flexible parts. However, these methods are still not fast enough in high throughput virtual screening tasks, and some of the algorithms are not specifically optimized for flexible-receptor sampling. In this work, we propose DSDPFlex, an improved receptor-flexible docking method. We accelerated the flexible docking process using GPU parallelization. We also implemented optimizations with respect to sampling, scoring, and search space to prevent redundant search and false-positive solutions, to further improve the efficiency and accuracy. To estimate the performance of DSDPFlex, a cross-docking dataset was constructed. Upon cross-docking tests, DSDPFlex obtained a 19.1% top-1 success rate, which is higher than AutoDock Vina, and the average runtime is ~1 second per task, representing a ~250-fold acceleration with respect to Vina. The generalizability of the method is validated on other commonly used test sets; the adaptability to different systems is also discussed. With improvement in both speed and accuracy, DSDPFlex is expected to show potential in docking-based studies.