Prediction Of Adverse Drug-Drug Interaction Based On Graph Neural Network

Adverse Drug-Drug Interactions(DDIs)refers to when two or more drugs are applied simultaneously or within a certain period of time,with the participation of body factors,the drugs occur due to the interaction between each other Changes in pharmacokinetics or(and)pharmacodynamics.Adverse drug interactions are a very important risk factor in drug therapy,which may cause serious side effects to patients and even death.Although a large number of in vivo and in vitro experiments will be used to screen out some combinations of adverse drug interactions during the drug development stage,this process requires huge costs and requires a certain test cycle.However,after the drug is put on the market,due to the complexity of human vital signs,there may still be a large number of new side effects that threaten human health and life.With the rapid development and continuous improvement of medical informatization and the abundant medical data resources,people began to use data analysis and machine learning to predict adverse interactions between drugs in advance during drug design and treatment planning.Although many computational research methods have been proposed,there are still some problems that we need to pay attention to and solve.The limitations of these methods mainly include(i)the inability to better cope with the highly unbalanced data of DDI(ii)mostly two-stage rather than end-to-end models(?)the inability to capture higher-order similarities between drugs(iv)The specific DDI type cannot be predicted.In order to solve the above problems,the article adopts a graph(network)representation structure.The problem of predicting adverse drug interactions can also be modeled naturally using graph data structures.Nodes represent drugs and other drug-related molecules,and edges represent adverse interactions between the two drugs or connections between drugs and other molecules.The graph is a powerful and complex tool that can mine more hidden information for our prediction task by collecting more neighbor information and even learning the information in the entire graph.The main work of the article includes:1.Proposed a model of adverse drug interaction prediction based on graph neural network autoencoder“GNN Autoencoder".In order to train the network end-to-end,we use the self-encoder model.The encoder part uses the algorithm of the graph neural network to nonlinearly learn the topology information and feature information of each node neighbor node on the graph to obtain the hidden layer of the node.Vector representation,and then use the decoder to reconstruct the graph structure,the new edges appearing in the graph may be the predicted new adverse drug interactions2.Proposed a model for predicting adverse drug interactions based on multi-relational link prediction”R-VGAE Autoencoder" is proposed.In order to make the predicted results really affect the patients' combination medication,we need to predict the specific types of adverse effects,so we have considered the specific types of edges in the model settings,and each type represents a different type of side effects.At the same time,we added protein information related to drugs to the graph to construct a heterogeneous graph,and then used the multi-relational prediction method based on the heterogeneous graph to realize the prediction of specific drug interaction types.Finally,we perform method validation and experiments on DrugBank,Twosides and other data sets.The accuracy of prediction is greatly improved compared with previous methods,and the specific DDI type can also be predicted.In addition,we investigated the predicted false positive results,and some cases of adverse interactions have appeared and discussed in some drug-related papers,indicating that this research has indeed greatly helped the discovery of new DDIs.