A Study On Phage-host Association Prediction Method Based On Graph Embedding Representation Learning

The ability of phages to lyse bacteria and inactivate them has led to studies suggesting that phage therapy may be a new alternative to antibiotic therapy.Investigating phage-host interactions could provide strong support for phage therapy,but currently phage-host association information is insufficient,and finding phages of specific bacteria requires expensive and time-consuming biological experiments.In contrast,computational-based association prediction methods can provide a reference direction for biological experiments and shorten the experimental cycle.The phage-host association database is still incomplete and biological data are missing,resulting in few studies on phage-host association prediction at present.Therefore,how to effectively predict phage-host association becomes an important challenge for current phage therapy.In this thesis,we use multi-source data to construct phage-host heterogeneous networks and learn to mine potential associations in phage-host association networks based on graph-embedded representations,and the main contributions are as follows.First,this thesis proposes a phage-host association prediction algorithm based on graph attention mechanism embedded representation learning(GERMAN-PHI).Firstly,the multi-source similarity networks of phages and hosts are constructed based on the diversity of biometric information of phages and hosts,respectively.Secondly the phagehost association network connects the two similarity networks to form the phage-host heterogeneous network.Then considering the different importance of different neighbor nodes to the central node in the heterogeneous graph,the talking-head strategy is introduced to improve the multi-head attention mechanism to encode the node representation.Finally,the phage-host association matrix is complemented by the neural induction matrix.The experimental results and visual analysis show that GERMAN-PHI can effectively mine the potential phage-host association.Second,this thesis proposes a phage-host association prediction algorithm based on feature aggregation strategy(biGERMAN-PHI).Considering that the microbial feature representations of multiple sources contain complex topological information of phage-host heterogeneous networks,a bi-directional aggregation strategy is proposed to fuse the node representations in the multi-source feature space,which in turn encourages the information transfer of nodes in complex neural networks,thus realizing the task of phage-host association prediction in sparse non-connected graphs.The experimental results show that compared with other five prediction algorithms,bi GERMAN-PHI can tap the potential phage-host correlation more accurately.The effectiveness of biGERMAN-PHI algorithm in non-connected graphs is further verified by example analysis.