Recognition And Prediction Of Antimicrobial Peptides Based On Deep Learning

With the proliferation of antibiotics,many bacteria have become resistant to antibiotics,threatening people’s health.Antibacterial peptide is an alkaline polypeptide substance with antibacterial activity.It has the advantages of high antibacterial activity and broad antibacterial spectrum.It is a good substitute for antibiotics with promising applications.At present,many researchers have used machine learning algorithms to predict and identify antimicrobial peptides,and achieved excellent results,but there is still some room for improvement.In this thesis,deep network models are used for antimicrobial peptide prediction and identification,and an antimicrobial peptide prediction platform is constructed to provide antimicrobial peptide prediction services,assist researchers in antimicrobial peptide prediction and identification,and improve the efficiency and identification of antimicrobial peptide prediction and identification.First,an end-to-end network including an embedding layer,a convolutional layer,a maximum pooling layer,a bidirectional LSTM layer,and a fully connected layer was constructed,and four different antimicrobial peptide data sets were acquired,and on these four data sets train end-to-end networks separately and compare with other antimicrobial peptide prediction models.The effect of this model is better than BiLSTM,iAMP-2L and MAMP-Pred,and the prediction accuracy is 0.67%,3.93%and 3.17%higher respectively.In order to analyze the contribution of each structure of the model to the model,the embedding layer,convolution layer and bidirectional LSTM of the model were removed and retrained and tested.The results show that when the convolution layer and bidirectional LSTM layer are removed,the prediction accuracy of the model Reduced by 3.06%and 0.73%respectively,the corresponding area under the ROC curve decreased by 0.0084 and 0.0115,respectively,while removing the embedding layer,the model prediction accuracy and the corresponding area under the ROC curve decreased by 0.62%and 0.0001,which indicates that for prediction For antimicrobial peptides,it is important to use convolutional layers and bidirectional LSTM to obtain the local and global information of the sequence.Secondly,a BERT model is pre-trained on a total of 556,603 protein sequences obtained from UniProt and then is fine-tuned on four different antimicrobial peptide data sets.In comparison with AMPScan,Bi-LSTM,iAMP-2L and MAMP-Pred,the prediction accuracy is higher than 0.93%,0.36%,4.21%and 1.51%respectively.In addition,all antimicrobial peptide data sets are synthesized and cross-validated by 50%.The verification results show that the specificity,sensitivity and accuracy of the model are all higher than 85%.The above experiments show that the antimicrobial peptide prediction model based on BERT is feasible.Therefore,on the integrated antimicrobial peptide data set,an antimicrobial peptide prediction model was finetuned to provide antimicrobial peptide prediction services.Finally,through demand analysis and research,an antimicrobial peptide prediction service platform was designed and implemented based on the Django framework.The model used by the platform was a BERT-based antimicrobial peptide prediction model that was fine-tuned on the comprehensive data set.There are still some deficiencies in the work of this thesis,so in the future,we will rebuild the data set,build a multi-classification model,use the BERT model on other protein tasks,and further optimize the antimicrobial peptide prediction platform.