Molecular docking is a type of computer simulation aimed at finding the binding mode between a small molecule or ligand and a larger receptor molecule. The success of docking simulation depends on the efficiency of the search algorithm that explores the space of possible conformations and the scoring function that evaluates their binding potential within the binding site. Due to the ability to learn complex dependencies from data, we decided to integrate a deep-learning model into the existing docking protocol of the CmDock tool. We chose the AQDnet model because it achieved excellent results on the CASF-2016 benchmark and used it as a scoring function during docking. We tried to integrate it in three different ways, with the model being successful only in rescoring docked poses. Docking using the AQDnet model achieved a slightly better average RMSD on 90 complexes from the DUD-E dataset compared to the CmDock scoring function. However, the CmDock scoring function was still more successful when docking some of the complexes. We also accelerated the algorithm for feature generation required by the AQDnet model for scoring by about 20 times.
|
