Research On Prediction Of Protein–protein Interaction Based On Deep Learning

Protein-protein interaction(PPI)is the physical contact between two or more protein molecules.A thorough understanding of PPIs is important for annotating protein function,establishing protein-protein interaction network,studying the molecular mechanism of disease,and developing new disease treatment methods.With the rapid development of high-throughput sequencing technology and the great success of the Human Genome Project,the rapid growth of protein and related data has laid a data foundation for the identification of protein-protein interactions using computational methods such as deep learning.Due to the difficulty of obtaining protein structure data,the known methods based on deep learning mainly focus on protein sequence data.However,the features extracted only from one-dimensional sequences are very limited,and the function of proteins is determined by their structure.Whether or not proteins can interact with each other is closely related to the structure of the protein.Therefore,how to obtain spatial features from the structure to help improve the ability to identify protein interactions is a problem to be considered.Another current dilemma is that with the increased complexity of the disease,the difficulty and cost of drug discovery are increasing rapidly.Drug discovery is technically difficult,requires a lot of investment and a long development cycle,and the problem of drug resistance is a huge potential hazard.Drugs are able to exert their effects,which is essentially a kind of interaction.In order to solve this problem,artificial intelligence to assist the development of antimicrobial peptide drugs is a new attempt.To this end,this paper has carried out the following work in response to the above two problems:(1)Aiming at the problem that the known methods for predicting protein-protein interactions can’t effectively utilize structural information,this paper proposes a multi-dimensional feature fusion deep learning model,TAGPPI,which fuses the one-dimensional sequence features and three-dimensional structural features of proteins,applying for the downstream protein-protein interaction prediction task.The TAGPPI model consists of three modules,namely a sequence feature extraction module,a structural feature learning module and a classifier module.The sequence feature extraction module applies a one-dimensional convolutional neural network and pooling operations to capture the information in the one-dimensional space of proteins.The structural feature learning module extracts the spatial structure information of proteins through a graph neural network.The role of the classifier is to fuse the features learned in the two dimensions and output the classification result.Although the number of known protein structures is limited at present,high-precision protein structures have been obtained by applying the latest research on protein structure prediction methods.We ought to reasonably apply this research result in various scientific research directions related to proteins.A series of experimental results have shown that the performance of the TAGPPI model is better than those methods that only use sequence information,and has achieved high accuracy in binary and multi-classification tasks.(2)In view of the current difficulties in the drug discovery and the great harm of drug resistance,we propose a method based on deep learning to assist in the design of antimicrobial peptide drugs.The method proposed in this paper can generate antibacterial peptides and predict whether antibacterial peptide drugs can act on bacteria when the target of drug is unknown.After pre-training in large-scale protein sequence data set,the model performs transfer learning on antimicrobial peptide data set,in order to allow the model to learn the semantic features required for excellent antimicrobial peptides.At the same time,an SVM(Support Vector Machine)discriminator is trained to identify whether the antibacterial peptides generated by the model have antibacterial activity,that is,whether the peptide drugs can interact with bacteria.In addition,through the design of iterative optimization process and the screening of SVM discriminator,the ability of the model to generate highly active antimicrobial peptides is further improved.The new antibacterial peptides generated and selected under this process have been verified by biological experiments to have excellent antibacterial ability,which also proves the effectiveness of the method proposed in this paper.