| The determination of lead compounds plays an important role in the development of new drugs.Based on computer virtual screening of lead compounds,it is helpful to use the advantages of computer high-speed and automation to improve the efficiency of new drug research and development and reduce costs.Molecular docking is an important technology of virtual screening,but there are many technical problems in the process of molecular docking,such as many parameters,complex steps,and difficult to identify active sites.In recent years,the rapid development of deep learning model provides a new method for virtual screening.Compared with molecular docking technology,the deep learning model has the advantages of relatively simple steps,high degree of automation and low requirements for background knowledge of new drug research and development in virtual screening of lead compounds.This article encodes the data(Encoding Atoms/Bonds/Sequence,E-ABS),we construct E-ABS-LSTM-MHA,which combines the Long Short-Term Memory(LSTM)with the Multi-Head Attention(MHA)mechanism and Fully Connected Neural Network(FCNN),and train the binding data of Binding DB in the order of millions to realize the extraction of binding characteristics from compounds/ligands to human targets proteins/receptors.On this basis,the binding data of Binding DB in the order of millions are classified according to the structural similarity of compounds/ligands.According to the descending order of the volume size of the bound data categories after classification,the first 21 groups of bound data categories are modeled respectively: EABS-LSTM is composed of LSTM and fully connected neural network(FCNN);EABS-CNN is composed of CNN and FCNN.Train the above three models to extract binding features from compounds/ligands to human targets proteins/receptors.The optimal candidate model is determined according to the threshold that the Precision is greater than 94%.Through a large number of experiments,this paper realizes the construction,training,prediction of three deep learning models and the determination of the best candidate model.Among them,the Accuracy of E-ABS-LSTM-MHA on the whole sample is 99.68%,the false positive rate is 0.29%,and the Precision rate is only 1.87%;E-ABS-CNN has the highest prediction Accuracy and Precision in the first,second and third categories of bound data,and the lowest false positive rate.The prediction Accuracy in the first category of bound data is 99.92%,the false positive rate is 0.13%,and the Precision rate is 99.99%;In the second category,the Accuracy rate of binding data prediction is 99.86%,the false positive rate is 0.07%,and the Precision rate is99.80%;In the third category,the Accuracy rate of binding data prediction is 99.99%,the false positive rate is 0.12%,and the Precision rate is 99.68%.It meets the requirement that the Precision rate is greater than 94%,and is selected as the final model.For the prediction results of indazole series of compounds,this paper uses molecular docking technology to verify the results.The results show that the binding energy of indazole compounds predicted by E-ABS-CNN with protein Cytochrome P450 2D6,d UTP pyrophosphatase,Sphingosine Kinase1 and Tyrosine-protein kinase Fes/Fps is less than ‐7.00 kcal/mol,which is the recognized threshold,indicating that E-ABSCNN has prediction potential.In this study,the work of computer virtual screening of lead compounds in the process of new drug research and development has been explored based on structure classification,and feature extraction has been carried out by using deep learning related technology.Integrate the best results of all experiments,The Precision rate has been improved to 99.99%,the false positive rate has been reduced to 0.07%,and the molecular docking verification has been preliminarily passed. |
