Research On Drug Target Feature Prediction Based On Multi-network Fusion

Drug target interaction prediction aims to explain the current mechanism of action of drugs and the possible unknown target activity,which is an important part of the development of pharmacology,and is also the focus of current research.Traditional biological experiments can effectively detect drug-target interactions,but they are time-consuming and costly.With the increasing and updating of data of various drugs,targets and interactions,computational methods to predict possible drug target interactions have gradually become a research hotspot as an economical and effective alternative means,providing high reliability for further experiments.Existing related research is mainly based on the calculation method of a variety of drug targets by fusion associated network to predict drug targets,but most of the method is easy to overlook contained information between different network topology,it is difficult to predict accurately extract the characteristics of the different network information,and does not consider drug target correlation matrix sparse sex influence on the calculation results,It is difficult to meet the increasing demand of drug target research.Therefore,on the basis of previous work,this paper will study from the following two aspects:(1)Aiming at the problems of feature fusion and dimension reduction existing in multi-network fusion,this paper proposes a drug target prediction algorithm MDADTI based on multimodal denoising autoencoder.Based on the topology similarity matrix of drug and target calculated by random walk with restart and positive point mutual information,the algorithm adaptively generates compressed low-dimensional vectors from high-dimensional vertex vectors by applying multimodal denoising autoencoder according to different input matrices.Finally,the drug target matrix is reconstructed to achieve the prediction effect.In order to verify the effectiveness of the method,we compare the algorithm with the traditional machine learning algorithm,and the model effect is far better than other methods.Compared with other advanced algorithms,the accuracy rate is far higher than other methods.(2)Aiming at the problems of feature update and matrix sparse in multinetwork fusion,this paper proposes a drug target prediction algorithm NAGCN based on attention diagram convolution.Based on the initial drug target features obtained by diffusion component analysis,the attention map convolution was used to update the drug target features and reconstruct the drug target prediction model by NAGCN.The 10-fold cross validation method is used to compare NAGCN with other existing prediction algorithms based on multi-network fusion feature updating,and the results show that our method is superior.In addition,we further verify the validity of NAGCN through robustness test and actual case analysis.