Validating and benchmarking deep learning model for protein-ligand binding affinity prediction

Abstract

Drug discovery is a complex, time-consuming, and challenging process, but the introduction of computational methods, especially machine learning, and deep learning models, has accelerated the entire process. OnionNet is a deep Convolutional Neural Network model that showed promising results in predicting the protein-ligand binding affinities based on the structural features of protein-ligand complexes. As the performance of any machine learning/deep learning model depends on the quality of the datasets used, we need to validate and benchmark the model using various datasets to ensure its accuracy and reliability. The OnionNet model is trained using the PDBbind v2016 dataset. The prediction power of the model is evaluated with the PDBbind v2016 core set, the BAPPL dataset, and the BindingDB dataset. The performance of OnionNet is impressive on the PDBbindv2016 dataset and BAPPL dataset, which have high-quality experimentally determined 3D protein-ligand structures. The performance of the model is below average on the BindingDB dataset with 3D structures of ligands generated by the program Vconf. Still, it is better when compared to the performance of AutoDock and AutoDock Vina. The performance of the OnionNet, AutoDock, and AutoDock Vina in finding the best ligand binding pose is also evaluated, and AutoDock Vina outperformed the others.