DWPREPHI: a novel deep learning-based computational model to predict phage-host interaction via complex multi-dimensional biological information.

Many studies project that, due to antibiotic misuse, phage therapy has been considered as one of the most promising alternatives for the treatment of human diseases infected by antibiotic-resistant bacteria. The identification of phage-host interactions (PHI) helps to explore the mechanisms by which bacteria respond to phages and provides new insights into effective therapeutic approaches. Computational models for predicting PHI are not only time/cost saving, but also more efficient and economical than traditional wet experiments. In this work, we proposed a deep learning based computational model named DWPREPHI to predict PHI through the combining DNA and protein sequence information. More specially, DWPREPHI first extracted information about the node properties of the interaction network by a natural language processing algorithm that initialized the node representations of the phage and the target bacterial host. The graph embedding algorithm, Deepwalk, was then used to extract link behavior information from the PHI network, and finally a deep neural network was applied to accurately detect interactions between phages and their bacterial hosts. On the drug-resistant bacteria dataset ESKAPE, DWPREPHI achieved a prediction accuracy of 92.25% and an AUC value of 0.9674 under the 5-fold cross-validation method, which was significantly better than other methods. In addition, three case studies were conducted for E. coli, Pseudomonas aeruginosa, and Salmonella enterica to further demonstrate the utility of the proposed model. Among the top 10 phages associated with these hosts, 7, 8, and 8 have been reported. These excellent experimental results suggested that the DWPREPHI model could provide reasonable candidates for sensitive bacteria for biological experiments in phage therapy. [ABSTRACT FROM AUTHOR]