Assisting In-Silico Drug Discovery Through Protein-Ligand Binding Affinity Prediction By Convolutional Neural Networks

Context In this paper, we present a new feature-generating method based on distance-weighted atomic contact to predict the binding affinity between ligands and proteins in structure-based drug design, which can distinguish between weak and strong interactions. We discuss the significance and challenges of predicting binding affinity and how recent advances in hardware and deep learning algorithms have led to the surge of applying these techniques in drug design. Methods: We utilize the PDBbind 2016 dataset for training our model. Our approach involves using a convolutional neural network architecture to predict the binding affinity between ligands and proteins. We investigate the impact of choosing the architecture on the performance of the model. Our best model, the Sequential Model, produces Pearson's correlation of 0.79 on the independent core set. Our results suggest that a vanilla and shallow convolutional network has more acceptable performance than a more complicated architecture specifically for this problem.