Relational completion based non-negative matrix factorization for predicting metabolite-disease associations

Abstract Metabolite, also known as intermediate metabolite, refers to substances produced or consumed in the metabolic processes. There are growing evidences that metabolites play an important role in the study of diseases. Due to the traditional experiments, it is time-consuming and luxurious to find the associations between metabolite and disease, we proposed a computational method, called RCNMF, to predict metabolite-disease associations. Firstly, we calculate the disease semantic similarity and the molecular fingerprint similarity of metabolite. The molecular fingerprint similarity of metabolite makes full use of the molecular structure internal information of metabolites. Then, we modify the original metabolite-disease associations matrix to replace some values of 0 with numbers between 0 and 1. Finally, we use the non-negative matrix factorization algorithm to predict potential metabolite-disease associations. We adopt the cross-validation mechanism to verify the performance of our proposed method. The AUC values of based the Leave-one-out cross validation measurement and the Five-fold cross validation measurement reach 0.9566 and 0.9430, respectively. What is more, case studies of common diseases also illustrate the effectiveness of our method. Thus, the superior experimental results show that our method can effectively predict the potential disease-metabolites associations.