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

Introduction: Predicting the binding affinity of ligands and proteins is a vital yet difficult part of structure-based drug design. Recent progress in hardware, particularly GPUs, and the development of efficient deep learning algorithms have significantly increased the application of these technologies to address challenges in drug design. In this study, we introduce a new feature-generation method based on distance-weighted atomic contact, which effectively differentiates between weak and strong interactions. We also examine how the choice of convolutional neural network (CNN) architecture impacts this problem.Materials and Methods: We used the PDBbind ۲۰۱۶ dataset to train our models. Our approach involved testing different CNN architectures, focusing on a simple, shallow sequential model. The feature-generation method was created to capture key interaction patterns between ligands and proteins. We validated the model's performance using the independent core set of CASF-۲۰۱۶.Results: Our best model, the Sequential Model, achieved a Pearson's correlation coefficient of ۰.۷۹ on the CASF-۲۰۱۶ core set. These results show that a simple, shallow convolutional network paired with a basic feature-generation method can outperform more complex models in this specific case.Conclusions: This study demonstrates that a vanilla CNN architecture and a simple feature-generation technique can effectively predict ligand-protein binding affinity. The findings indicate that simpler models can deliver highly acceptable performance in structure-based drug design. The Python code needed to reproduce this research is available at https://github.com/miladrayka/convolutional_neural_networks.

کلیدواژه ها:

Deep Learning ، Molecular docking ، PDBbind ، Scoring Function