Research On The Prediction Algorithm Of Metabolite-disease Association Based On Similarity Network

Metabolism is a general term for a series of orderly chemical reactions and plays a vital role in maintaining human life,such as the growth and reproduction of organisms and the response to the external environment of the body.A large number of studies and experiments have shown that the concentration of some metabolites varies greatly compared with healthy people.Therefore,the associations between metabolites and disease are important judgments of doctor's diagnosis and treatment.With the improvement of high-throughput metabolomics technology,researchers can obtain information about the associations between metabolites and diseases.At the same time,the establishment of metabolomics database such as HMDB promotes the development of metabolomics.However,because of the diversity of metabolites and diseases,thousands of associations still need to be detected,and the known data are still only the tip of the iceberg.Traditional biological experiments can test and verify some assumptions,but it usually takes a long time to get results.If the deviation between the results and assumptions is too large or the results are not significant,the experimenter may have to bear the corresponding economic losses.Therefore,the development of computational methods which save experimental time,money and provide available prediction results can mine metabolites strongly associated with disease to provide reference for biological researchers in the field of reducing the scope of the experiment and reducing the waste of costs.In this article,we use different types of computing methods to mine disease-related metabolites based on the network topology and biological information fusion of different networks,the specific work is as follows:Firstly,a prediction algorithm based on KATZ model is proposed to predict disease associated metabolites.First,biological information(disease semantic information)is fused with topological information.Subsequently,the known disease-metabolite association,disease similarity network disease,metabolite similarity network are input into the KATZ model.The KATZ model calculates the scores for two nodes according to the number of paths between each node and the length of each path.According to the experimental analysis,the algorithm can be used as a tool to mine disease related metabolites.Secondly,an improved bipartite network projection algorithm based on linear neighbor similarity is proposed to predict disease-related metabolites.The initial similarities are obtained by fusion of multi-data biological information(disease and gene information,disease semantic information,etc.)and topological information.After constructing a new feature network,we calculated linear neighbor similarity as the final similarity network.Finally,the improved bipartite network projection algorithm is used to obtain the final prediction relationship.The experimental results show that this method can identify disease-related metabolites well.Thirdly,a LightGBM(Light Gradient Boosting Machine)model is proposed to predict metabolite disease association.Firstly,the metabolite functional similarity is extracted from the biological information of the association between metabolites and pathways,and the disease functional similarity is obtained according to associations of diseases and symptoms.Secondly,the final features are extracted by statistical method,graph theory method,matrix decomposition method and PCA dimensionality reduction method.Finally,the features and labels are input into the LightGBM classifier to find the potential associations between disease and metabolites.The results from the corresponding performance evaluation method and the analysis of the experimental results show that the prediction algorithm has better performance.Three prediction algorithms are designed to mine the association relationship between potential metabolites and disease based on the known metabolite-disease association,metabolites(disease)topology information and their corresponding multiple biological information in this article.These algorithms lay the foundation for researchers in related fields to verify disease-related metabolites.