SE-OnionNet: A Convolution Neural Network for Protein–Ligand Binding Affinity Prediction

Abstract

Deep learning methods, which can predict the binding affinity of a drug-target protein interaction, reduce the time and cost of drug discovery. In this study, we propose a novel deep convolutional neural network called SE-OnionNet, with two squeeze-and-excitation (SE) modules, to computationally predict the binding affinity of a protein-ligand complex. The OnionNet is used to extract a feature map from the three-dimensional structure of a protein-drug molecular complex. The SE module is added to the second and third convolutional layers to improve the non-linear expression of the network to improve model performance. Three different optimizers, stochastic gradient descent (SGD), Adam, and Adagrad, were also used to improve the performance of the model. A majority of protein-molecule complexes were used for training, and the comparative assessment of scoring functions (CASF-2016) was used as the benchmark. Experimental results show that our model performs better than OnionNet, Pafnucy, and AutoDock Vina. Finally, we chose the macrophage migration inhibitor factor (PDB ID: 6cbg) to test the stability and robustness of the model. We found that the prediction results were not affected by the docking position, and thus, our model is of acceptable robustness.