Physics-Inspired Accuracy Estimator for Model-Docked Ligand Complexes

Abstract

Model docking, which refers to ligand docking into the protein model structures, is becoming a promising avenue in drug discovery with advances in artificial intelligence (AI)-based protein structure prediction. However, a significant challenge remains; even when sampling was successful in model docking, typical docking score functions failed to identify correct solutions for two-thirds of them. This discrepancy between scoring and sampling majorly arises because these scoring functions poorly tolerate minor structural inaccuracies. In this work, we propose a deep neural network named DENOISer to address the scoring challenge in model-docking scenarios. In the network, ligand poses are ranked by the consensus score of two independent subnetworks: the Native-likeness prediction and the Binding energy prediction networks. Both networks incorporate physical knowledge as an inductive bias in order to enhance pose discrimination power while ensuring tolerance to small interfacial structural noises. Combined with Rosetta GALigandDock sampling, DENOISer outperformed existing docking tools on the PoseBusters model-docking benchmark set, as well as on a broad cross-docking benchmark set. Further analyses reveal that the physics-based components and the consensus ranking approach are the two most crucial factors contributing to its ranking success. We expect that DENOISer may assist future drug discovery endeavors by providing more accurate structural models for protein-ligand complexes.