Prediction Of Phage-Host Interactions Based On Deep Learning

Bacteria have developed drug resistance during the extensive use of antibiotics,which has become a critical problem affecting human health worldwide.With the large increase in drugresistant bacteria,the therapeutic effect of antibiotics is not obvious,and people need to take some breakthrough approaches to solve this problem.As one of the most common and diverse organisms in the biosphere,the phage can not only destroy specific bacteria host but also multiply it.These characteristics make phage therapy one of the most promising methods to replace antibiotics.The key to phage therapy is to correctly match the target bacteria host with the corresponding therapeutic phage.Experimental methods to verify the correlation between the phage and the target bacteria host are time-consuming,labor-intensive,and expensive.Therefore,it is necessary to develop computational methods that predict the interactions between phage and target bacteria host.At present,similarity-based computational methods have low prediction accuracy,and machine learning-based prediction methods have poor stability,due to the random selection of negative samples.According to the current research status of prediction of phage-host interaction methods,this thesis develops two phage-host interaction prediction methods based on deep learning algorithms,which use the data from PhagesDB and GenBank databases.The detailed work of this thesis is as follows:(1)This thesis constructs a phage-host interaction prediction method based on randomly selected negative samples and deep learning.Because there are many unknown phage-host interactions(negative samples)data,this thesis randomly selects a part of negative samples from the negative sample set and constructs a data set with positive samples(known phagehost interactions).At the same time,this thesis quantifies multiple types of features(amino acid composition,chemical element abundance,and molecular weight),then this thesis uses the deep convolutional neural network to construct a prediction model PredPHI(Predicting phage host interactions).The experimental results show that PredPHI has more advantages than previous methods in predicting phage-host interaction.(2)This thesis constructs a phage-host interaction prediction method based on highquality negative samples and deep learning.To avoid the problem of poor model stability due to the random selection of negative samples,this thesis first designs three negative sample selection methods to construct three different training sets: random selection,K-Means-based clustering,and phage similarity method.Based on the data set and features of the previous work,this thesis uses more stringent conditions for data screening(removing phage-and host-encoded proteins that are hypothetical proteins).At the same time,this thesis compares the performance of different classifiers on the training set,and the results show that the deep learning model works best on all three training sets.Then this thesis compares the performance of the models built on the three training sets on the independent test set.The final experimental results show that the model constructed by selecting negative samples based on K-Means clustering method(PredPHI-V2)can get better performance,and the stability of the model is better than the model constructed by the method of randomly selecting negative samples.Finally,this thesis constructs a high-performance phage-host interaction prediction method that can provide phages with interactions for infected bacteria,provide potential phages for their personalized treatments,and facilitate further experiments by researchers.