Research On The Prediction Method Of MicroRNA And Disease Association Based On Two-layer Network

MicroRNAs(miRNAs)are a class of endogenous non-coding RNAs with a length of approximately 22-24 nucleotides.Accumulating evidence has demonstrated a close relationship between miRNAs and the occurrence and development of various diseases.Therefore,identifying candidate disease-related miRNA may contribute to exploring the pathogenesis of a disease.Reliable disease-related candidate miRNAs can be obtained by biological experiments;however,these methods are costly and time-consuming.In recent years,an increasing number of researchers have proposed computerized prediction methods to screen potential miRNA-disease associations.However,there are many types of connected edges in the miRNA-disease bilayer network,and these connected edges have nonlinear complex relationships.The aforementioned methods are all shallow prediction models which cannot discover deep relations between miRNA and disease nodes.The attributes of miRNA nodes including their family and cluster belonging information,however,have not been deeply integrated.In this paper,we firstly construct a bi-layer heterogeneous network of miRNA and disease nodes,and it contains multiple types of connections among these nodes which reflect neighbour topology of miRNAdisease pairs,and the attributes of miRNA nodes,especially miRNA-related families and clusters.We propose three miRNA-disease association prediction methods based deep learning.Comparison results with previous miRNA-disease association methods demonstrated the superior performance of our model.The unique contributions of our model are as follows:Research on neighbor topology in miRNA-disease bilayer network.This work proposes a prediction model,GMDA,to extract and integrate multiple representations of miRNA and disease nodes.Firstly,we construct a bilayer heterogeneous network with node attributes to facilitate the learning of the neighbor topology representations of miRNA-disease nodes.The network consists of multiple types of connections to embed the similarity and association between miRNAs and diseases,miRNAs-related family and cluster attributes.We also design an embedding mechasim to extract the pairwise neighbor topology from the network.we exploit the idea of generative and adversarial to learn enhanced representations of a pair of miRNA and disease nodes.The generator consists of an automatic encoder and decoder to generate false neighbor topology feature embedding of the node pair.The encoder based on a multi-layer convolution neural networks encoded a neighbor topology representation of the node pair.This was followed by reconstruction of the neighbor topology embedding of node pair based on a multi-layer transposed convolution decoder.The discriminator is based on a multi-layer convolutional neural network to discriminate the false neighbor topology embedding and the original true topology embedding generated by the generator.This discriminant strategy benefits the generator to generate neighbor topology embedding as close to the true topology embedding as possible and obtain the final neighbor topology representation of miRNA-disease node pairs.Test on the public dataset,through a number of evaluation criteria and case study analysis,it is found that GMDA is better than the five state-of-the-art miRNA-disease prediction models.Research on network representation learning in miRNA-disease bilayer network.Our main goal is to obtain the network embedding representation of miRNA-disease node pairs.Therefore,a miRNA-disease association prediction model,NEMDA,based on network representation learning is proposed.First of all,using the same strategy as GMDA,a bilayer network with node attributes is constructed by using miRNA similarity,miRNA node attributes,disease similarity and the relationship between them,which is used to embed the neighbor topology information and attribute distribution information of miRNA and disease nodes.The neighbor topology information of the node pair is captured by the multi-layer convolutional autoencoder,and the low-dimensional vector representation of the neighbor topology is obtained.The attribute distribution of miRNA and disease nodes is learned by multi-layer convolutional variational autoencoder,and the output of the coding part is represented as a low-dimensional vector of node attribute distribution.Then,through the attention mechanism module at the presentation level,the two representations and the corresponding miRNA family and cluster information are adaptively fused.Finally,a pair of miRNA and disease association scores were obtained through fully connected layer.The experimental results show that NEMDA has better performance than other miRNA-disease association prediction models.Research on topology and attribute information in miRNA-disease bilayer network.This work proposed a miRNA-disease association prediction model,DGMDA,based on graph convolutional networks(GCN)and generative adversarial network(GAN).First of all,a bilayer heterogeneous network is constructed to integrate miRNA and disease nodes,as well as various connections between these nodes,including miRNA similarity,disease similarity,miRNA and disease association.Then,a pairwise topology representation module based on GCN is introduced to learn the topology information of a pair of miRNA and diseases in the miRNA-disease bilayer network.After that,the node attribute representation is generated by learning the original attribute information of the node pairs based GAN module.Finally,a score evaluation module with a representation level attention mechanism is used to distinguish the contribution between the topological representation and attribute representation of node pairs and output a miRNA-disease pairwise association prediction scores.