Multi‐network logistic matrix factorization for metabolite–disease interaction prediction

Abstract

Edited by Qinghua CuiIdentifying disease-related metabolites is of great significance for the diagnosis, prevention, and treatment of disease. In this study, we propose a novel computational model of multiple-network logistic matrix factorization (MN-LMF) for predicting metabolite-disease interactions, which is especially relevant for new diseases and new metabolites. First, MN-LMF builds disease (or metabolite) similarity network by integrating heterogeneous omics data. Second, it combines these similarities with known metabolite-disease interaction networks, using modified logistic matrix factorization to predict potential metabolite-disease interactions. Experimental results show that MN-LMF accurately predicts metabolite-disease interactions, and outperforms other state-of-the-art methods. Moreover, case studies also demonstrated the effectiveness of the model to infer unknown metabolite-disease interactions for novel diseases without any known associations.

Metabolite-disease interaction prediction based on logistic matrix factorization and local neighborhood constraints

“…To improve the performance of the logical matrix factorization algorithm, researchers ( 12 , 17 ) introduced the local neighbor constraint. They sorted the nodes by their similarities to find neighbor nodes, but they ignored the diffusion and propagation of label information carried by neighbor nodes, which limited the performance enhancement.…”

“…However, in the training process, the latent vectors of some unobserved metabolites and diseases are obtained based on negative samples, which may not be accurate enough. Ma et al ( 12 ) presented an effective solution. Let represent the set of metabolites interacting with any disease, and let represent the set of K nearest neighbors of metabolites in .…”

“…Recently, some researchers have used machine learning methods to predict the interactions between metabolites and diseases ( 1 , 2 , 9 – 12 ). The majority of these methods work as follows: First, a metabolite-related heterogeneous network is built by integrating multi-omics information; second, the candidate metabolites are scored via a random walk-based method ( 4 , 9 , 13 ); finally, the ranking of disease-related metabolites is obtained according to the score.…”

“…Several studies have verified that compared with the global similarity network, the local structure information (neighbors) of nodes can significantly improve the algorithm’s performance. Ma et al ( 12 ) adopted the nearest neighbor regularization to eliminate the noise information in the metabolite-disease interaction network, and obtained good prediction results, which proved the effectiveness of the local structure information in the prediction of metabolite-disease interaction. Zhou et al ( 14 ) achieved the accurate classification of unlabeled nodes by introducing local neighbor information.…”

Data-Driven Imputation of Miscibility of Aqueous Solutions via Graph-Regularized Logistic Matrix Factorization

“…The basic idea behind the graph regularization is to incorporate prior knowledge into LMF learningthat similar categories of solutes tend to exhibit similar miscibility behaviorby promoting vector representations of solutions of similar solutes to be near each other in the latent space. (GR-MF has been applied to single-cell RNA-seq clustering and predicting drug–drug, drug–target, and metabolite–disease interactions ,− and side effects of drugs and (2) the rough grouping of the compounds into the categories of polymer, protein, surfactant, or salt, we show that GR-LMF learns latent representations of the solutions that give ATPS predictions on missing entries outperforming (i) ordinary LMF and (ii) a standard supervised machine learning approach using random forest classifiers taking as input physicochemical features of the compounds in the solutions.…”