| Traditional drug discovery techniques have made significant contributions to the treatment of human diseases,but such research processes are time-consuming,expensive and laborious.In view of this,it is very necessary to improve the efficiency of drug development.Drug repositioning technology is also proposed under this background,and has gained more and more attention.It is a new drug development model to find new indications for drugs already on the market.In recent years,since the application of highthroughput biotechnology has enabled the rapid accumulation of biological data,at the same time due to advance in artificial intelligence methods and computing science and technology,drug repositioning technology based on computational methods has an opportunity for development.Compared with other drug development strategies,it realizes the new use of old drugs by jointly analyzing biomedical data from different sources to find the association between existing drugs and new indications,thus reducing the time and cost of drug development.In this paper,two neural recommendation algorithm models based on graph neural network are proposed to predict potential Drug-Disease Association(DDA).Compared with the latest evaluation models,our models achieves excellent performance.The main work of this paper is as follows:1)A new end-to-end model CFDDA is proposed,which adopts the graph based neural collaborative filtering algorithm to perform entity encoding for DDA prediction.Drug/disease characteristics at different levels were modeled by neural collaborative filtering,graph convolutional network and weighted random walk,and DDA classification features were obtained after feature fusion.Finally,potential DDA are determined by prediction function.The results of 10-fold cross-validation on open data sets show that CFDDA has obvious advantages over other recent methods.In particular,the performance of AUPR evaluation index on the two data sets reaches 0.539 and 0.5917,respectively,which is much higher than other baseline models.2)A new end-to-end model BiGNN is proposed.This method uses bilevel graph neural network to predict DDA.The first layer uses the graph convolutional network method to obtain the features of different perspectives based on different data resources.The second layer uses the attention mechanism to integrate the features of different perspectives to obtain the classification features of DDA.Finally,potential DDA are predicted by classification function.Experimental results using 10-fold cross-validation on public data sets show that both AUROC values and AUPR values of BiGNN exceed the latest baseline models.To sum up,two methods of calculating drug repositioning based on graph neural network and integrating heterogeneous data resources were proposed for DDA prediction.The test results on the public data set exceeded the most recent baseline model for the same period.This indicates that advanced computational methods and multi-modal data fusion can effectively discover the relationship between potential biomedical entities and provide new ideas for drug repositioning research. |
