Study On Disease-related LncRNA Association Prediction Based On Multi-layer Heterogeneous Networ

Identifying disease-related lnc RNAs can help to deeply understand the pathogenesis of complex diseases,discover markers of complex diseases,study new drugs targeting lnc RNAs,as well as promote disease prevention,diagnosis,and treatment,and improve human health.However,traditional biological experimental methods have many disadvantages in predicting candidate lnc RNAs for diseases,such as long study periods and complicated experimental procedures.Using computational methods to predict lnc RNA-disease associations can not only reduce the time and cost of experiments but also provide biologists with highly accurate lnc RNAs.Therefore,it is a very interesting topic to predict the disease-associated lnc RNAs accurately.In this paper,based on the heterogeneous network structure constructed from lnc RNA and disease-related multi-sourced data,three computational models are proposed to predict disease-related candidate lnc RNAs.All models incorporate deep learning methods,and all of our models achieved optimal prediction performance compared to advanced methods for lnc RNA-disease association prediction.The main work accomplished is as follows,A lnc RNA-disease association prediction method,CASCLDA,based on the convolutional autoencoder with skip connections.First,for the similarity,association and interaction relationships between lnc RNA and disease multi-sourced data,we constructed the lnc RNA-disease-mi RNA three-layer heterogeneous network.Then,we constructed a matrix representation of the multi-source data,and by splicing multiple attributes of lnc RNA and disease,we generated a node-pair attribute embedding matrix.We propose a node pair attribute encoding module to learn the attribute representation of node pairs,integrating the detailed features and representative features of node pairs.Finally,the attribute representation of node pairs is used for association prediction.The results of comprehensive experiments demonstrate that CASCLDA outperforms several state-of-the-art lnc RNA-disease association prediction models.A lnc RNA-disease association prediction model,AMPLDA,based on the semantic meta-path and graph convolutional autoencoder.The multiple connections in the heterogeneous network contain rich semantic information.However,existing methods do not take the important information into consideration,so we propose AMPLDA to encode and integrate the semantic information of multiple meta-paths,the global topology of heterogeneous network.First,we construct a three-layer heterogeneous network to integrate similarities,associations,and interactions among multiple-sourced data.Multiple meta-paths associated with lnc RNAs and diseases are constructed to represent various semantics.Then,we build an embedding strategy to represent meta-path-specific semantics.For each meta-path and heterogeneous network,we build graph convolutional autoencoders for them separately to learn the local topological representation and the global topological representation of each lnc RNA and disease node.Subsequently,we designed a meta-path-level attention mechanism to learn the importance of each topological representation.Finally,the convolutional neural network is used to fuse the topological representations of the lnc RNA and disease nodes,and the topological representation is used for the final association prediction.The results under several evaluation metrics indicate that AMPLPA has superior predictive performance than other methods.The recall rate at different k also demonstrates that AMPLDA can retrieve more true lnc RNA-disease associations.A lnc RNA-disease association prediction model,NSLDA,based on the heterogeneous network structure learning.Based on the original heterogeneous network,we propose a new model NSLDA,which aims to learn the new heterogeneous network and provide more accurate network data for the subsequent graph convolutional neural network.First,we construct a three-layer heterogeneous network.For the multiple connection relations contained in the heterogeneous network,we generate the relational networks.In each relational network,feature similarity network and feature topology network are constructed based on the similarity between node features and the relationship between features and topologies.Then,we establish a channel-level attention mechanism to give different weights to feature similar network and topological networks to obtain feature network.In addition,we build multiple meta-paths to represent the multi-hop interaction information in heterogeneous network.Similarly,in each relational network,we generate the semantic network based on each meta-path and fuse multiple semantic networks to get the semantic network of relation through the channel layer attention mechanism.Finally,the original relational network,the feature network and the semantic network are integrated to get the new relational network.Then,we feed the new heterogeneous network into the graph convolutional neural network for the final association prediction task.Ablation experiments demonstrate the effectiveness of network structure learning for each part.The average AUC and AUPR demonstrate that NSLDA consistently outperforms the other six comparison methods.