Predicting Phage-Host Association By Heterogeneous Graph Attention Neural Network

Bacteria are common microbes in human bodies and they are closely related to our health.But the bacteria in the human body have developed drug resistance during the abuse of antibiotics,and we are helpless about the emergence of super bacteria.Bacteriophage is a group of viruses that attacks bacteria,each of which may affect one type or a few types of bacteria.They can multiply themselves in the host.The pheromones secreted by them have the advantages of safety,strong bactericidal effect and difficulty in producing drug resistance,etc.The bactericidal efficiency is hundreds or even tens of thousands of times that of conventional antibiotics at present.Phage therapy has brought dawn to the medical community,but finding effective phages in traditional biological experiments is like finding a needle in a haystack.Bio informatics predict potential phage-host associations based on similarity computation.Matrix factorization,traditional machine learning and neural network algorithms have been widely used in biological association prediction.In this paper,we predict the potential phage-host interaction by heterogeneous graph attention neural network algorithm as well as the quorum sensing similarity of bacteria,respectively.First,we predict phage-host associations based on the heterogeneous graph attention neural network algorithm(HGAT-PH).We used the known phage and host associations to calculate the Gaussian Interaction Profile(GIP)kernel similarity to construct a phage similarity network and a host similarity network,respectively.Then,we combine the two similarity networks with the known phage-host relationship pair to form a phage-host heterogeneous graph.After that,we input the heterogeneous graph into graph attention neural network algorithm to obtain the embedding of the phage and the host,and then the embedding is reconverted into the form of adjacency matrix according to tensor decomposition.In the form of adjacency matrix,the value of elements in the matrix indicates the potential association probability between the phage and the host.In this paper,we used three phage and bacterial genome datasets which are extracted from NCBI and phagesDB.Finally,the performance of HGAT-PH is verified better than the other three methods under five-fold cross-validation.The experiment shows that the proposed method in this paper has some improvement compared with the comparison method.Second,we construct a priori information based on the similarity of bacterial quorum sensing as the prediction of phage-host association.The generation of biofilms on the surface of bacterial populations is responsible for the development of drug resistance in bacteria,making it difficult for antibiotics to enter.Since the formation of biofilm is closely related to quorum sensing in bacteria,we constructed the quorum sensing(QS)similarity based on the data of the quorum sensing database.Then the existing phage oligonucleotide frequency(ONF)similarity was also added as a prior information to reconstruct the phage similarity network and the host similarity network.Finally,we put the reconstructed phage-host heterogeneous network to the graph attention neural network to predict potential interactions.The final experiments show that the performance of the prediction model that integrates QS similarity is improved to a certain extent compared with the previous one,and the HGAT-PHqs algorithm achieves the best results.In summary,we propose two effective phage-host interaction prediction methods in this paper,which have achieved the best prediction results with existing models,and provide new methods and tools for predicting potential phage-host associations.