Research On The Prediction Method Of MiRNA-disease Potential Correlation Based On Generative Adversarial Networ

miRNA is a kind of small molecule RNA,which can regulate gene expression and affect physiological function of cells by targeting and blocking mRNA expression.It plays an important role in tumor,cardiovascular disease,nervous system disease,diabetes,liver disease,etc.miRNA can also affect the body’s sugar metabolism by regulating the expression of insulin receptors and signal transduction pathway,and also affect the body’s resistance to pathogens by regulating the immune response pathway.Identifying the mirNA-disease pair relationship helps to discover biomarkers and disease mechanisms.The construction of computational models to identify potential correlation has become a research focus,and the use of neural networks to mine deep correlation information between mirnas and diseases has achieved remarkable results.Integrating multi-source data to extract effective features is a challenging task.Traditional MLP multi-layer perceptron structures or deep neural network models have been faced with bottlenecks in extracting features of mirNa-disease deep association information.In order to solve this problem,this paper aims to construct the mirNa-disease association characteristic matrix,and proposes to generate antagonistic network to learn the potential association relationship of mirNa-disease heterogeneous characteristic network,so as to obtain high-dimensional feature vectors,and realize the truth positive rate of generated data approaching the real data.In this paper,each miRNA and disease node were constructed to realize the correlation characteristics between nodes.Three methods for predicting mirNa-disease potential correlation based on generative adversarial network were proposed.Experimental results showed that the proposed model achieved good results in predicting mirNa-disease potential correlation,which was better than previous methods.The main contributions of this paper are as follows:(1)The first method is a prediction model based on bidirectional adversarial generation network(BGANMDA for short).First,based on the known mirNA-disease association and the comprehensive similarity of miRNA(disease),the miRNA similarity network,disease similarity network and Gaussian interaction spectrum kernel similarity were constructed.Next,a similar feature network with a complete potential mirNadisease relationship was obtained.Bidirectional adversarial generation network(BiGAN)was constructed to predict the mirNA-disease association.BiGAN consists of an encoder,a generator and a discriminator.The BGANMDA encoder maps the feature vector x to the potential representation z,and the BGANMDA generator captures the features previously in space to generate new mirNa-disease associations.On the basis of generating adversarial network(GAN),BGANMDA input data of mirNa-disease as real samples,and realize data compression through the full connection layer so that the discriminator can better capture high-dimensional features.The results were finally tested in a public data set,and good results were obtained in a variety of evaluation indicators compared with the latest 10 mirNA-disease prediction models.(2)The second method is based on the Self-Attention generative adversarial network model(SAGANMDA),which introduces the self-attention mechanism into the generative adversarial network and uses the self-attention mechanism to calculate the long correlation of the global feature learning features in the mirNa-disease similar feature network.The feature weighting of global feature matrix position is realized.First,using the same data preprocessing method as BGANMDA,miRNA functional similarity,miRNA sequence similarity and disease semantic similarity were used to construct mirNa-mirNA correlation matrix and disease-disease correlation matrix,and thus mirNadisease correlation characteristic matrix was obtained.The self-attention mechanism is introduced into GAN network,and the self-attention mechanism is introduced into the generator to realize the learning and training of the global characteristics of mirNa-disease,so that the network can pay attention to the global information of mirNA-disease.Finally,it is verified on the public test set,and better results are obtained compared with the previous prediction models.(3)The third method is mirNa-disease prediction model(MMAGMDA)based on multi-source attribute fusion.We designed a multi-source attribute learning framework to predict mirNa-disease associations using graph machine neural networks(GCN)and selfattention generation adversarial networks.Specifically,GCN focuses on local mirNadisease association features to capture local association matrix topology information and learn compression dimension features.Self-attention generation adversarial networks focus on global features and capture global mirNA-disease data features.Finally,the final verified scores of the two models were weighted by scoring strategies to calculate the mirNa-disease interaction scores.Evaluation of MMAGMDA under different experimental conditions showed that the proposed method achieved better results than the latest 10 mirNA-disease models.