Research On The Prediction Of Potential Association Between Microorganisms And Diseases Based On Heterogeneous Networks And Deep Learnin

Experiments show that there are a large number of microorganisms in human body,which has a great impact on human health.It plays an important role in the formation and development of various complex diseases.Therefore,mastering the potential relationship between microorganisms and diseases will help us understand the pathogenesis of diseases and have great significance for disease prevention,diagnosis and treatment.However,designing traditional biological experiments to verify disease-related microorganisms is usually expensive and time-consuming.Therefore,developing efficient computational methods to predict the relationship between microorganisms and diseases and providing biologists with reliable disease-related candidate microorganisms can effectively reduce biological costs and improve the efficiency of experiments.How to accurately predict the relationship between microorganisms and diseases and shorten the training time of the model is a problem worth studying.In this paper,three prediction methods of potential association between microorganisms and diseases are proposed.These three methods have different performance in prediction accuracy and calculation speed:(1)Bidirectional Heterogeneous Selection Algorithm Predict Association Between Microorganisms And Diseases.This method has great advantages in prediction accuracy,and it is the highest prediction accuracy of the three methods in this paper.First,we build a bidirectional heterogeneous microbial disease network based on the similarity between microorganisms and diseases.Different similarities can more comprehensively reflect the relationship between microorganisms and diseases from different perspectives.Secondly,based on the enhanced two-way random walk,we learn the neighbor topology information of microorganisms and disease nodes in the two-way heterogeneous network from the directions of microorganism-disease and disease-microorganism.Multifeature fusion of microorganism and disease nodes is helpful for the final prediction of the relationship between microorganism and disease.In heterogeneous networks,if all neighbor information of each node is aggregated,the information of different types of nodes will also be aggregated,which will lead to information redundancy.Therefore,we propose a selection model based on graph convolution to selectively aggregate the neighbor information of disease(microorganism)nodes.Finally,compared with the most advanced models,ablation experiments and case studies based on LOOCV and five-fold cross-validation proved the predictive performance of our model.(2)Prediction Method of Potential Association Between Microorganism and Disease Based On Multiple Matrix Decomposition.This method has great advantages in computing speed.It is the method with the best calculation speed in this paper.First,we not only deeply integrate the disease features based on the multiple non-negative matrix decomposition method to obtain low-dimensional feature vectors,but also integrate the diversity of disease features.We assume that the possibility of the potential association between microorganisms and diseases is greatly related to the distance between them.The interaction similarity of Gaussian kernel spectrum of disease,the semantic similarity of disease,the symptom similarity of disease,and the functional similarity of disease reflect the characteristics of disease from different angles,and the retention of multiple disease features can fully integrate the information of different disease levels from different levels.Finally,by comparing with the most advanced model,the case study based on LOOCV and five-fold cross-validation proves the prediction performance and calculation speed of our model.(3)Asymmetric Random Walk and Convolution Neural Network For Microbial Disease Potential Prediction.This method has good performance in prediction accuracy and calculation speed.We build a heterogeneous network of microorganisms and diseases by integrating multiple connections related to microorganisms and diseases,as well as their similarity,interaction and original feature matrix,and then use asymmetric random walk and convolution neural network model to predict the potential relationship between microorganisms and diseases.ARWCNP not only uses a variety of similarities between microorganisms and diseases,but also obtains topology information of heterogeneous networks.The topological feature vectors of microorganism and disease nodes are obtained by asymmetric random walk algorithm,and then the convolution neural network is used to further integrate the deeper feature information.Finally,compared with the most advanced models,ablation experiments and case studies based on LOOCV and fivefold cross-validation proved the excellent prediction performance and good computing speed of our model.