Research On The Prediction Method Of Drug-protein Interaction Based On Heterogeneous Network

Identifying the proteins that interact with drugs plays an important role in the initial period of developing drugs.At the same time,it also plays an important role in drug repositioning.Traditionally,drug-protein interactions are often identified and verified through biological experiments,which require high time and economic costs.Wet lab experiments can be performed with candidate proteins predicted by computational methods,which can not only reduce the cost of drug development,but also reduce the time required and accelerate the process of drug development.At present,most methods for predicting the interactions between drugs and proteins focus on exploiting various data about drugs and proteins.However,these methods failed to completely learn and integrate the attribute information,the topology information of a pair of drug and protein nodes and their attribute distribution.This will affect the performance of the model.Three new methods,CVDTI,GVDTI and MGVDTI,are proposed to predict the interactions between drugs and proteins.CVDTI is a model based on convolutional variance autoencoder and multi-layer convolution neural network.It integrates the similarities,interactions and correlations between drugs,proteins and diseases,deeply integrates the relevant information of drug and protein nodes,and learns the attribute distribution and attribute representation of a pair of drug and protein nodes.The initial attribute information of drug and protein nodes is sparse and has a high dimension.In this paper,the attribute information related to drugs and proteins is projected into the low dimensional attribute space through CVDTI,and it is subject to Gaussian distribution to improve the accuracy of the model.The experimental results show that CVDTI is superior to several latest drug-protein interaction prediction methods.In recent years,more and more studies have proved the importance of the topology information and attention mechanism.Therefore,this paper proposes a drug-protein interaction prediction method based on graph convolutional autoencoder with attributelevel attention,GVDTI.It extracts and embeds the hidden topological structure from drug similarity,drug-disease and drug-protein sub-networks.Since individual attribute in a node’s attribute vector have different contributions to topological embedding,we propose the new attention mechanism at node attribute level to adaptively learn and reflects the discriminative contributions of each sub-network’s node attribute.The comparison results show that the prediction performance of GVDTI is better.In order to comprehensively consider the topology of drug and protein related networks.This paper also proposes method,MGVDTI,which is based on multi-channel graph convolutional autoencoder.A new way is added to the graph convolution module of GVDTI to extract the topological representation of heterogeneous networks.It not only constructs a more comprehensive attribute network,but also obtains the topological representation of each node from the drug similar network,protein similar network,drugprotein heterogeneous network and their corresponding attribute network.Then,pairwise topological representation,attribute representation and attribute distribution are fused by convolution to predict drug-protein interaction scores.The results show that compared with other methods,the recall rate of MGVDTI has been greatly improved,the prediction results are more accurate,and more interactions can be retrieved.In addition,case studies of five drugs,quetiapine,verapamil,amitriptyline,clozapine,and ziprasidone,were conducted to further demonstrate the ability of MGVDTI to discover potential candidate proteins.