Prediction Of Adverse Drug Reactions Based On Multi-kernel Graph Convolutional Networks

With the continuous progress in medicine,pharmacology and pharmaceuticals,drugs can be used to treat more and more diseases and achieve better results.However,the process of drug use may have adverse effects that can cause certain health risks to patients,so it is necessary to find out whether the human body will have adverse effects after taking a certain drug.Although a large number of clinical experiments are conducted in the process of drug development,traditional clinical experiments on drugs are expensive and less efficient.With the development of information technology,a large amount of clinical data has been aggregated,and researchers have started to use computers to predict the potential association between drugs and adverse reactions,so as to guide drug development and clinical drug use.In this paper,we propose a multikernel graph convolutional networks based adverse drug reaction prediction method to improve the shortcomings of current prediction methods and achieve better prediction results.Firstly,based on the fact that similar drugs have similar adverse reactions,the similarity matrix of drugs and adverse reactions needs to be calculated.In order to cover more comprehensive drug information,the chemical structure of the drug and the known association information of the drug and the adverse reactions need to be extracted;the known association information of the selected drug and the adverse reactions are extracted for the adverse reaction information.After calculation by four similarity measures,multiple kernel synthesis will be performed using maximum cosine.Secondly,there are more than one type of nodes and edges on the topological graph between drugs and adverse reactions,and the construction of the heterogeneity matrix will be performed based on the similarity matrix of drugs and adverse reactions and the known drug-adverse reaction adjacency matrix obtained above.Finally,the heterogeneity matrix is input into the graph convolutional networks model,and a graph embedding is extracted from each layer of the graph convolutional networks,which in turn extracts a drug kernel and an adverse reaction kernel,and the drug kernel and adverse reaction kernel of each layer are synthesized with average weighting and input into the Dual Laplacian regularized least squares model for prediction.In this paper,the multiple kernel learning method used to calculate the similarity matrix of drugs and the similarity matrix of adverse reactions has better prediction effect than single kernel.The proposed multi-kernel graph convolutional networks based adverse drug reaction prediction method also showed AUC values of 0.9767,0.9763,and 0.9746 for Pauwels,Liu,and Mizutani datasets,respectively,and AUPR values of 0.8348,0.8401,and 0.8457,respectively.learning methods and existing literature methods using the same dataset,the proposed method has a more significant improvement in both AUC and AUPR.Also,when compared with the traditional graph convolutional networks prediction method,the prediction effect is also better than that of the graph convolutional networks,achieving a more comprehensive acquisition of adverse drug reaction information.