Learning from Deep Representations of Multiple Networks for Predicting Drug–Target Interactions

Abstract

Many computational approaches have been developed to predict drug-target interactions (DTIs) based on the use of different similarity networks that connect drugs and targets. However, such approaches do not fully exploit all the information available in all similarity networks which can be considered as multiple domain representations of DTIs. As more comprehensive understanding of the latent knowledge underlying the DTI networks requires combining insights obtained from multiple, diverse networks, there is a need for a computational approach to be developed to learn hidden patterns from multiple DTI networks simultaneously for more complete understanding of DTIs. In this paper, we propose such an approached based on a deep multiple DTI network fusion algorithm, called DDTF. to take into consideration all relevant DTI networks. With this DDTF, the identification of DTIs can be made more effective. The DDTF performs its tasks in several steps. Given a set of complex heterogeneous networks, DDTF first uses the network completion algorithm to reconstruct the data representation information to obtain the best network description. To do so, a matrix factorization technique is first used. Based on this approach, the networks obtained from multiple domains are first represented by several similarity matrices and the feature vectors of each pair of drug and protein of the DTI networks are obtained. With these features and representations, we introduce here a novel approach based on non-negative matrix factorization to rescale similarity networks to ensure that the data are reliable. DDTF algorithm constructs a new network to represent the similarity between two vertices. The new similarity network is calculated from the heterogeneous information embedded by a new fusion algorithm. As a final step, DDTF finds a deep representation of each drug or protein in the fused network and use such information for the inference of DTIs. Given the fused deep representations, DDTF can discover optimal projection from a drug network onto a target network. The DDTF algorithm has been tested with real data and experimental results show that DDTF outperforms sophisticated network integration approaches and others significantly. Based on the experiments, it is discovered that the network representation inferred by DDTF has a higher correlation than those yielded by previous work. Moreover, it is noted that completing similarity network based on known networks is a promising direction for drug-target predictions.