Prediction of Drug-Disease Associations and Their Effects by Signed Network-Based Nonnegative Matrix Factorization

Predicting drug-disease associations using computational methods benefits drug repositioning. Drug-disease associations are events that drugs exert effects on diseases, there are different effects about drug-disease associations. For example, drug-disease associations are annotated as therapeutic or marker/mechanism (non-therapeutic) in Comparative Toxicogenomics database (CTD). However, existing association prediction methods ignore effects that drugs exert on diseases. In this paper, we propose a signed network-based nonnegative matrix factorization method (SNNMF) to predict drug-disease associations and their effects. First, drug-disease associations are represented as a signed bipartite network with two types of links for therapeutic effects and non-therapeutic effects. After decomposing the network into two subnetworks, SNNMF aims to approximate the association matrix of each subnetwork by two nonnegative matrices, which are low-dimensional latent representations for drugs and diseases respectively, and diseases in two subnetworks share the same latent representations. In the computational experiments, SNNMF performs well in predicting effects of drug-disease associations. Moreover, SNNMF accurately predicts drug-disease associations and outperforms existing association prediction methods. Case studies show that SNNMF helps to find out novel drug-disease associations that are not included in CTD, and simultaneously predicts their therapeutic effects.