Research On LncRNA And Disease Association Prediction Method Based On Multi-layer Heterograph

Human diseases are often closely associated with aberrant expression of long noncoding RNAs,and identification of disease-associated lncRNAs is essential to help elucidate complex pathogenesis.However,designing experiments to identify diseaseassociated IncRNAs is usually resource-intensive.In this paper,we propose three approaches to predict lncRNA-disease correlations at different levels of LncRNAs and diseases.The experimental results show that the prediction results of all the models proposed in this paper are better than the previous ones.The main work accomplished is described as follows.(1)The first approach is a predictive model based on a fully connected autoencoder and convolutional neural network for extracting,encoding and adaptively integrating multiple representations.First,we constructed a multilayer heterogeneous graph using different types of nodes and edges of lncRNAs,diseases and miRNAs to reflect the similarities,interactions and associations between lncRNA,disease and miRNA nodes.Then we defined two representations,which are low-dimensional feature attribute representation and local topology representation.The low-dimensional feature attribute representation is obtained by extracting the low-dimensional features of lncRNAs,diseases and miRNAs nodes in the heterogeneous graph by a fully-connected autoencoder based on body-in.We propose attention mechanisms at the node feature level and meta-path level to learn more informative features.Convolutional neural networkbased modules are built to further learn local topological representations of lncRNAs and diseases at the meta-path level.Tested on public datasets,the models proposed in this dissertation are found to outperform lncRNA-disease prediction models by multiple evaluation criteria and case study analysis.(2)The second approach is a prediction model based on generative adversarial networks and convolutional neural networks,used to integrate the properties of lncRNAs.We propose a model based on generative adversarial network,i.e.,a generator based on a fully connected neural network autoencoder to encode the low-dimensional topological representation of lncRNA-disease pairs,and a discriminator based on a convolutional neural network to discriminate between true and false neighboring topological embeddings.Comparison results with six lncRNA-disease association methods show that our model and technical contributions have advantages in terms of AUC and AUPR.Case studies of three cancers further demonstrate the ability to detect potential candidate lncRNAs.(3)The third approach is a model based on graph convolutional networks and convolutional neural networks.The global topological representation of the heterogeneous graph is extracted by using a graph convolutional neural network,and the local topological representation of lncRNA and disease node pairs is obtained by a convolutional neural network.The outstanding achievement of GCLDA in predicting potential lncRNA candidates associated with diseases is demonstrated by comparison with six recent prediction models and case studies of three cancers.