Research On Molecular Property Prediction And Optimization Methods Through Graph Attention Adversarial Networks

This paper introduces the first and critical step in modern drug discovery-the computational discovery of ideal lead compounds,which often includes hit compound screening,molecular property prediction,and molecular optimization.Compared to traditional methods,the deep learning model-based lead compound discovery can simulate the interaction between candidate compounds and targets through hit compound screening and property prediction,and improve the properties of candidate drugs through molecular optimization,thereby improving drug discovery efficiency.Firstly,a molecular property prediction method based on graph attention adversarial network is proposed.The method transforms the linear input of molecules into a molecular structure graph and uses a graph attention network to obtain neighbor node information and learn the high-dimensional representation of molecules.Then,the extracted features are projected into two subspaces through bidirectional adversarial network with variable perturbation gradients,which dynamically generates a large number of feature representations in one space.The direction of feature perturbation that affects the activity value the most is found,the maximum difference in biological activity value is minimized,and the predicted molecular property is finally outputted.The experiments were conducted on 33 GPCRs datasets,and the results were compared with the graph attention network algorithm.The experimental results show that our method has improved to varying degrees in r~2,RMSE,Tau,EF10,EF20,and EF30 indicators.In addition,the experiments on the ADMET datasets showed that our method had improvement in AUC,ACC,MCC,specificity,and average sensitivity decreased by 2.35%.Overall,our method performed significantly better than other graph neural network-based methods.Furthermore,a molecular optimization method based on graph attention adversarial generative network is proposed.The method first generates new features for atoms and bonds through the graph reconstruction network module,and then finds the key atoms whose positions can be replaced with alternative elements through the graph node classification module.Finally,the validity optimization module is used to ensure the effectiveness of the molecular optimization process.Experiments were conducted on multiple datasets such as GPCRs,ADMET,and COVID-19.The experimental results show that our proposed molecular optimization method can generate novel and high-activity MMP-Cliffs molecules,demonstrating the effectiveness and superiority of our method.Overall,this paper emphasizes the importance of lead compound discovery based on deep learning models in modern drug discovery,and proposes a molecular property prediction method and molecular optimization method based on bidirectional adversarial graph attention network.The experimental results show that both methods can effectively improve the efficiency and quality of drug discovery.In summary,these methods provide a new approach and tool for drug research and development.