Font Size: a A A

Research On Drug Virtual Screening Based On Bi-LSTM And PSO Improved Algorithms

Posted on:2024-06-13Degree:MasterType:Thesis
Country:ChinaCandidate:W H ZhaoFull Text:PDF
GTID:2544307058981859Subject:Engineering
Abstract/Summary:PDF Full Text Request
Drug screening and discovery is an indispensable step in drug development,aiming to find compounds that can bind to specific targets and have activity.In order to overcome the disadvantages of traditional drug screening and development methods,such as high cost,long time and relatively low success rate,virtual screening technology has emerged.With the rapid development of deep learning technology,virtual screening technology has gradually become popular and has become an important part of modern drug discovery,which can be used in various stages of drug discovery,including drug molecule design,molecular docking,and efficacy assessment.At the same time,virtual screening technology is very important for accelerating drug discovery process,reducing R&D cost and improving R&D success rate.In this thesis,we propose a new virtual screening method and neural network based on Bidirectional Long Short-Term Memory Network(Bi-LSTM),Transformer module,feature fusion,Particle Swarm Optimization(PSO)improved convolutional Neural network with the above mentioned algorithms and mechanisms into the deep learning neural network,a new virtual screening method and neural network is proposed.The main research contents of this thesis are:(1)In this thesis,a Molecular Directed Information Feature Extraction Network(MDIFEN)based on Bi-LSTM module is proposed for feature extraction and classification of molecular directed message.Compared with traditional molecular description methods(e.g.,fingerprints,descriptors,etc.),the molecular directed information generated by Directed Message Passing Neural Networks(D-MPNN)is more flexible and accurate,and performs well in molecular description and property prediction.This network feeds the output sequence of the convolutional neural network into the Bi-LSTM module,which can improve the performance of the model by solving the problem of long-range dependence of sequence information such as molecular directed message passing information while maintaining the extracted local feature information.(2)In this thesis,we further propose a Molecular Information Fusion Neural Network(MIFNN)based on feature fusion,Transformer module and PSO algorithm in order to improve the screening effect.MDIFEN extracts the features of molecular targeted information transfer,and on the other hand,the molecular Morgan fingerprints are extracted by a two-dimensional convolutional neural network,and the obtained dual stream features are cascaded and fused,and sent to the Transformer module for weight redistribution.Finally,it is sent to the Support Vector Machine(SVM)based on the improved PSO algorithm for classification,and the PSO algorithm is used to optimize the parameters such as weights or bias values in the SVM,which improves the training speed and accuracy.MIFNN improves the prediction ability,stability and generalization ability of the model through feature fusion on the one hand;on the other hand,the Transformer module with the help of PSO algorithm can effectively alleviate the overfitting phenomenon that often occurs in the virtual screening of drugs.(3)This thesis is based on publicly available datasets for virtual screening,such as HIV,BACE,Tox21,Tox Cast,etc.These datasets contain rich compound information,have high quality labeled data,diverse data types and are easily accessible,which have high practical value.In this thesis,we designed complete control experiments to validate the two virtual screening neural networks proposed above.Firstly,we set up a baseline model comparison experiment to compare the newer and more effective models at the current stage;in addition,we also design different self-control experiments to verify the accuracy and effectiveness of the virtual screening by adding modules and algorithms to the original networks.
Keywords/Search Tags:Bi-LSTM, Feature fusion, Transformer, Particle Swarm Optimize
PDF Full Text Request
Related items