Prediction Of LncRNA-disease Association Based On Graph Neural Network

Long non coding RNA(lncRNA)is defined as non coding-RNA with length more than 200 nucleotides.More and more studies have shown that lncRNA plays an important role in many life processes.Exploring the relationship between lncRNA and diseases can not only understand the pathogenesis from the molecular level,but also contribute to the diagnosis and treatment of diseases.Traditional experimental methods can accurately identify the relationship between lncRNA and disease,but they are time-consuming and laborious.With the establishment of a large number of lncRNA and disease-related databases,and in order to overcome the shortage of traditional experimental methods,researchers use computational methods to predict the associations between lncRNA and disease.However,most of the existing methods only consider the shallow feature information of lncRNA-disease,and ignore the deep topological information and node feature information.In order to solve the above problem,based on graph neural network,this article integrates the biological features of lncRNAs and diseases,and extracts the topological information and node feature information of lncRNA-disease.After that,it obtains the low dimensional representation of nodes,and then predicts the associations of lncRNA-disease according to the node representations.It includes the following two aspects.(1)We proposed a graph attention network based framework(GANLDA)for predicting the associations between lncRNA and disease.In this framework,the known lncRNA-disease associations is represented as a bipartite graph.We first integrate the lncRNA-miRNA,lncRNA-GO and lncRNA-Gene data to form the biological features of lncRNA nodes,and integrate the disease-miRNA and disease-Gene data to form the biological features of disease nodes.Then we use principal component analysis(PCA)to denoise the original features of lncRNA and disease,and employ the multi head graph attention network to learn the latent feature representations of lncRNA and disease.Finally,the embedding vectors of lncRNA and disease are concatenated,and the score of the relationship between each lncRNA and disease is obtained by using the multi-layer perception(MLP).The ten fold cross validation experiment and De novo test results show that this method is superior to the comparison method.At the same time,the case study also verifies the effectiveness of the method.(2)We proposed a lncRNA-disease associations prediction framework(GAMCLDA)based on graph autoencoder matrix completion.At first,the bipartite graph and node features are encoded by graph convolution network to obtain the embedding vector of nodes.For a new disease without any known association,the top k similar diseases are selected to pre-fill in the association matrix by calculating the Gaussian interaction profile kernel similarity between diseases.Then,we decode the embedding vector to obtain new lncRNA-disease associations by inner product.Finally,the weight of positive samples is improved by cost sensitive learning.The ten fold cross validation experiment and De novo test results show that our method is superior to the existing method.In addition,the case study further prove the ability of this method to predict unknown lncRNA disease associations.