A New Paradigm for Applying Deep Learning to Protein-Ligand Interaction Prediction

Abstract

Protein-ligand interaction prediction poses a significant challenge in the field of drug design. Numerous machine learning and deep learning models have been developed to identify the most accurate docking poses of ligands and active compounds against specific targets. However, the current models often suffer from inadequate accuracy and lack practical physical significance in their scoring systems. In this research paper, we introduce IGModel, a novel approach that leverages the geometric information of protein-ligand complexes as input for predicting the root mean square deviation (RMSD) of docking poses and the binding strength (the negative value of the logrithm of binding affinity, pKd) with the same prediction framework. By incorporating the geometric information, IGModel ensures that its scores carry intuitive meaning. The performance of IGModel has been extensively evaluated on various docking power test sets, including the CASF-2016 benchmark, PDBbind-CrossDocked-Core, and DISCO set, consistently achieving state-of-theart accuracies. Furthermore, we assess IGModels generalization ability and robustness by evaluating it on unbiased test sets and sets containing target structures generated by AlphaFold2. The exceptional performance of IGModel on these sets demonstrates its efficacy. Additionally, we visualize the latent space of protein-ligand interactions encoded by IGModel and conduct interpretability analysis, providing valuable insights. This study presents a novel framework for deep learning-based prediction of protein-ligand interactions, contributing to the advancement of this field. O_TEXTBOXKey MessagesO_LIWe introduce the first framework for simultaneously predicting the RMSD of the ligand docking pose and its binding strength to the target. C_LIO_LIIGModel can effectively improve the accuracy of identifying the near-native binding poses of the ligands, and can still outperform most baseline models in scoring power, ranking power and screening power tasks. C_LIO_LIIGModel is still ahead of other state-of-the-art models in the unbiased data set and the target structure predicted by AlphaFold2, proving its excellent generalization ability. C_LIO_LILatent space provided by IGModel learns the physical interactions, thus indicating the robustness of the model. C_LI C_TEXTBOX