Protein-ligand Binding Affinity Prediction Based On Graph Neural Networks

In the process of drug development,studying the affinity between small molecule drugs and proteins is a critical step that helps researchers evaluate the drug’s potency,toxicity,and dosage,and quickly screen potential drug candidates.Traditional wet lab methods for evaluating protein-ligand affinity are not only time-consuming and labor-intensive but also involve significant experimental costs.The development of modern computer technology and bioinformatics has provided powerful tools for predicting protein-ligand affinity,greatly accelerating research progress in this field.Common methods for predicting protein-ligand affinity include molecular docking,molecular dynamics simulation,machine learning,and deep learning.Molecular docking is one of the commonly used prediction methods,which generates a large number of protein-ligand conformations through computer simulation.Then,stable and reliable conformations are selected for docking experiments,and the interaction forces and affinity of the resulting complex are calculated.Molecular dynamics simulation is a physics-based prediction method for protein-ligand affinity,which first simulates the molecular movements of small molecules and pr oteins during the binding process.By analyzing their structural changes and dynamic behavior,the interaction forces and affinity between them are calculated in both spatial and temporal domains to obtain the final protein-ligand binding affinity.These two computer-based methods have high requirements for computing resources and time.In recent years,machine learning and deep learning have been widely used in the field of protein-ligand affinity prediction.These methods use known protein-ligand binding data to train complex neural networks to learn their features and binding patterns.Machine learningbased methods mainly include random forests,gradient boosting trees,and others,while deep learning-based methods mainly include convolutional neural networks(CNNs),attention models(Transformers),graph neural networks(GNNs),and others.These methods have decent prediction accuracy and speed,but there is still room for improvement and enhancement,such as some models based on sequence information that do not utilize structural information,and some methods that require pre-obtaining protein-ligand complexes and cannot process monomers.This article proposes training graph neural networks to address the problems in this field of research.By integrating molecular chemical bonds and amino acid distances,the model has stronger learning and generalization capabilities.In addition,the overall architecture of the model is more lightweight than other methods,requiring low computational resources and faster prediction speed.On two recognized public test sets in this field,the proposed method in this article outperformed existing methods in terms of prediction accuracy.Relevant ablation experiments were also conducted to further demonstrate the rationality and feasibility of the method’s design.In summary,the method proposed in this article can serve as a novel tool for predicting the affinity between proteins and small molecules,thereby assisting in drug development.