SNF-CVAE: Computational method to predict drug-disease interactions using similarity network fusion and collective variational autoencoder
MetadataShow full item record
CitationJarada, T. N., Rokne, J. G. ve Alhajj, R. (2021). SNF-CVAE: Computational method to predict drug-disease interactions using similarity network fusion and collective variational autoencoder. Knowledge-Based Systems, 212. https://dx.doi.org/10.1016/j.knosys.2020.106585
Drug repositioning is an emerging approach to identify novel therapeutic potentials for approved drugs and discover therapies for previously untreatable diseases. Drug repositioning has also attracted considerable attention in the pharmaceutical industry due to its time and cost efficiency in the drug development process compared to the traditional de novo drug discovery process. Recent advances in genomics, the tremendous growth of large-scale publicly available data, the availability of high-performance computing capabilities, along with the rise of machine learning, have further motivated the development of computational drug repositioning approaches. Investigating the relationship between different biomedical entities (e.g., drugs, diseases, genes) is one vital part of most recent studies in the drug repositioning field. Drug-disease interaction (R-DI) prediction is another main issue in drug repositioning research. Combining these relationships and interactions when introducing computational methods to identify novel drug-disease interactions with high accuracy is very challenging. In this study, we propose a robust approach, SNF-CVAE, for predicting novel drug-disease interactions using drug-related similarity information and known drug-disease interactions. SNF-CVAE integrates similarity measures, similarity selection, similarity network fusion (SNF), and collective variational autoencoder (CVAE) to conduct a non-linear analysis and improve the drug-disease interaction prediction accuracy. We evaluated the robustness of SNF-CVAE using different information models, drug similarity calculation measures, and drug similarity information. Moreover, we compared SNF-CVAE performance with four state-of-the-art machine learning models. SNF-CVAE achieved outstanding performance in stratified 5-fold cross-validation (Prec = 0.902, Rec = 0.883, F1 = 0.893, AUC-ROC = 0.958, and AUC-PR=0.970). Furthermore, we showed the efficiency of SNF-CVAE in predicting novel drug-disease interactions by validating the top-ranked interactions against pharmaceutical indications and clinical trial studies, which resulted in substantial pieces of evidence for almost all of RDIs predicted by our proposed model. To further demonstrate the reliability and robustness of SNF-CVAE, we conducted two case studies on the top predicted drug candidates for potentially treating Alzheimer's disease and Juvenile rheumatoid arthritis, which were successfully validated against clinical trials and published studies. In conclusion, we strongly believe that computational drug repurposing research could significantly benefit from integrating similarity measures and deep learning models to predict novel drug-disease interactions in heterogeneous networks.