Research On CircRNA-Disease And Drug-Disease Association Prediction Methods Using Graph Neural Networks

The birth of molecular biology and the completion of the Human Genome Project have propelled the development of life sciences and led to the explosive growth of biological data.To reveal the biological information and biological natural laws contained in the massive data,a comprehensive discipline,bioinformatics,has been formed by integrating biology with computer science,mathematics and physics.Biological entity association prediction is an effective way to explore the internal connections and roles of biological systems based on bioinformatics,aiming to study the interactions between biological entities such as diseases,RNAs,drugs and proteins.Disease research is the focus of human medicine research.Predicting the association between diseases and other biological entities can contribute to finding the pathogenesis of diseases,exploring the specific expression of RNAs in diseases,and providing scientific basis for the identification of therapeutic targets and biomarkers.Since biological experiment-based association prediction methods are time-consuming,expensive and cannot be performed on a large scale,traditional machine learning and deep learning-based association prediction methods are widely applied.To explore the molecular mechanism of disease pathogenesis and discover new applications of marketed drugs,this thesis conducts research on circRNA-disease association prediction and drag-disease association prediction based on graph neural networks.The details are as follows:Firstly,a circRNA-disease association prediction algorithm with sparse interaction information based on a multi-source homogeneous graph structure is proposed.To obtain the initial features containing multi-source information,the circRNA and disease homogeneous graph structures are constructed by circRNA Gaussian interaction profile kernel similarity,disease Gaussian interaction profile kernel similarity,and disease semantic similarity.The variational graph auto-encoder suitable for handling sparse data is used to extract the deep features of circRNAs and diseases,respectively.CircRNA and disease features are weighted fusion,and then fed into the random forest to perform the prediction task.Experimental results show that the variational graph auto-encoder can effectively extract the deep features of circRNAs and diseases,and the random forest has a strong classification ability in the circRNA-disease association prediction.The proposed model achieves better circRNA-disease association prediction performance compared with other existing methods,and the case study can further prove the practical value of the model.Then,a circRNA-disease association prediction model based on a heterogeneous graph embedding model is developed to solve the problem of missing heterogeneous information in the existing methods.Given that circRNAs can act as miRNA "sponges" to affect diseases,miRNA nodes are introduced to construct a circRNA-miRNA-disease heterogeneous graph structure containing heterogeneous association information.To obtain the heterogeneous semantic information of the graph structure,four meta-paths are designed,and the meta-path-based random walk is applied to sample in the graph structure.The path embedding model based on Skip-gram and random negative sampling is constructed to convert the walk sampling information into the initial feature vectors of circRNAs and diseases.To extract the deep circRNA-disease features and perform prediction tasks,the CosMulformer model based on linear self-attention mechanism and Hadamard product is developed.Experimental results indicate that the introduction of miRNAs can effectively complement heterogeneous information,and the CosMulformer model can effectively obtain circRNA-disease deep interaction features.Compared with other existing methods,the proposed method has better performance in circRNA-disease association prediction,and the case study can clearly show the practical application value of the model.Finally,a drug-disease association prediction algorithm based on graph convolutional neural network and graph attention network is proposed for the problem that the same drug has different mechanisms of action for different diseases.To obtain the initial drug-disease information,a drug-disease heterogeneous graph structure is constructed based on the known associations.Based on the heterogeneous graph structure,subgraphs of each drug-disease association pair are extracted separately to distinguish different background information of different diseases for the same drug.The GnnAp model is constructed by combining the graph convolution layer with the graph attention layer to learn drug-disease interaction features and predict the association between drugs and diseases.Experimental results demonstrate that adding subgraph extraction can effectively improve the prediction accuracy of the model,and the graph representation learning module can fully extract the deep drug-disease features.Compared with other algorithms,the proposed model can obtain better prediction results of drug-disease association.