ArtiDock: fast and accurate machine learning approach to protein-ligand docking based on multimodal data augmentation

Abstract

We present ArtiDock - the deep learning technique for predicting ligand poses in the protein binding pockets (aka “AI docking”), which is based on augmenting inherently limited training data with algorithmically generated artificial binding pockets and the ensembles of representative conformations of the ligand-protein complexes obtained from MD simulations. Performance of ArtiDock is compared systematically with other AI docking techniques and conventional docking programs on the PoseBusters dataset, which is dedicated for benchmarking the AI pose prediction algorithms. ArtiDock outperforms the best AI docking techniques and the major conventional docking programs, being at least an order of magnitude faster while providing superior accuracy in terms of RMSD and additional ligand pose correctness metrics. The influence of data augmentation on the model performance is evaluated and the perspectives of further development are discussed.