Directed Generation And Application Of Antimicrobial Peptides Based On Generative Adversarial Networks

In recent years,the issue of antibiotic resistance has increasingly threatened human health.Antimicrobial peptides,as a type of biological material,are excellent new antibiotic alternatives due to their unique non-specific membrane rupture mechanism.However,natural antimicrobial peptides have many problems such as high toxicity and difficult identification.Therefore,artificial design of antimicrobial peptides has become a hot research topic.However,the traditional methods for designing antimicrobial peptides are difficult,and existing methods mostly target a single property of the peptides,making it difficult to generate peptides with multiple properties simultaneously.Efficient identification of antimicrobial peptides to save time and manpower is also necessary.However,there are several challenging issues in the classification of antimicrobial peptides,such as poor classifier generalization ability,small data volume,and difficulty in evaluating classification performance,which have become limiting factors in antimicrobial peptide research.To address these issues,this article mainly performs the following tasks:Firstly,regarding the difficulty of traditional methods for designing antimicrobial peptides and the inability to generate peptides with multiple properties simultaneously,we propose a generative model Multi-CGAN based on Generative Adversarial Networks.The Multi-CGAN model can generate peptides with multiple specific properties simultaneously by training on peptide data with multiple single attribute labels,thus improving the efficiency of targeted design of antimicrobial peptides.Through experiments,we demonstrate that Multi-CGAN has good generative performance and generates peptides with good diversity and low homogeneity to training data.Furthermore,Multi-CGAN can improve the performance of certain downstream tasks,and we further explore the generative ability of the model and the interpretable analysis of antimicrobial peptide generation.Multi-CGAN can capture information from the sequence and physicochemical property level,and establish a relationship from noise distribution to antimicrobial peptide generation distribution.Secondly,to improve the performance of antimicrobial peptide classification tasks,we propose Perturbation GAN,a model that borrows ideas from transfer learning and adversarial learning,to augment the antimicrobial peptide dataset to address the issue of insufficient data volume.The Perturbation GAN model can utilize perturbations of negative class samples that are similar to positive class properties to enhance model generalization ability.The generated positive samples can not only learn the rough distribution of the positive training data,but also learn the information around the positive training data by adjusting the loss weight of negative samples.Perturbation GAN model demonstrates that finding appropriate loss weight of negative samples and adding a suitable number of generated peptide sequences to the original dataset can effectively improve the performance of the original antimicrobial peptide binary classification problem.Finally,we developed a multi-peptide sequence generation platform that can adjust the model parameters to obtain peptide sequence data that meets different requirements,greatly facilitating the use of related practitioners.Through this research,we provide new ideas and methods for the design and classification of antimicrobial peptides,and provide useful references and guidance for research and applications in related fields.