Prediction Of Drug-protein Interaction Based On Multi-source Data Fusio

Drugs exert their effect in treating diseases by interacting with corresponding proteins,so the discovery of drug-protein interactions(DTI)is crucial to the development of drugs.Accurate identification of potential drug-protein interactions reduces the cost and time required for drug development in biological experiments.The computational prediction approach of DTI demonstrates that screening of reliable candidate proteins can help biologists in experimental validation.Traditional methods usually integrate multi-source data of drugs and proteins to predict DTI.The multiple connections between two drugs(proteins)reflect their internal relationship from different perspectives.However,most of previous methods failed to deeply integrate these connections.In addition,multiple drug-protein heterogeneous networks can be constructed based on multiple kinds of connections.The diverse topological structures of these networks are still not exploited completely.This paper proposes the following three innovative methods for multi-source information,topological information,attribute information and semantic information of drugs and proteins.(1)For the attribute information of drugs and proteins,the paper proposes TDCNDTI based on deep learning to predict drug-protein interactions.The paper proposes a new embedding strategy to form 3D attribute embeddings of drug and protein node pairs that cover drug and protein related similarities and interactions.The three-dimensional attribute embedding is used as the input of the three-dimensional convolutional neural network,and the deep features embedded in width,length and height are extracted by two layers of convolution-pooling operations.Finally,the features are fed as input to the fully connected layer to obtain the interaction score between a drug-protein pair.This paper validates the ability of the method in predicting drug-protein interactions by case studies and comparative experiments.(2)The paper also proposes a new method named GADTI to predict the interaction tendency between drugs and proteins.First,four drug-protein heterogeneous networks are established.Then,the neighbor node sequence of a drug or protein node is formed by random walk to reflect the node’s hidden neighbor topology.Second,a framework based on graph neural network is designed to learn the topological features of each node’s neighbors in different types.The paper also builds a network-level attention mechanism to enhance the context dependence among multiple neighbor topologies of drug-protein node pairs.Compared with several other advanced drug-protein interaction prediction methods,it is shown that GADTI achieves better prediction performance.Evaluation of the top-ranked recall and case studies of five drugs demonstrate the predictive ability of GADTI.(3)Based on the constructed drug-protein heterogeneous network,a new method named MGDTI is proposed in this paper.We can obtain multiple meta-paths of drug and protein nodes by the heterogeneous network.Different meta-paths have different semantic structures and contain rich semantic information.Therefore,the paper proposes a method to calculate the similarity between nodes of the same type based on multiple semantic structures.And in order to enhance the original node features,the paper uses the graph convolutional neural network to integrate the data rich in semantic information and the original node features to obtain the final node representation.In the experiments,case studies and comparisons with other state-of-the-art methods can demonstrate the feasibility of the method.