HeteroRepur: Efficient Modeling of Heterogeneous Disease Graph for Depression Drug Repurposing using Heterogeneous Graph Neural Networks

Abstract

ObjectiveMajor Depressive Disorder (MDD) is one of the most prevalent psychiatric disorders, yet existing treatments often result in suboptimal responses and encounter high recurrence rate. Repositioning existing medications, which often significantly reduces the required time and budget, has been demonstrated to be a promising strategy for drug discovery of multiple central nervous system (CNS) disorders. However, an efficient modeling strategy of heterogeneous semantic information among biological networks has yet to be developed for drug repurposing. Material and MethodsIn this study, we proposed HeteroRepur, by integrating heterogeneous graph neural networks to learn interaction features of Depression Drug Repurposing Graph (DDRG) and then classify drugs. ResultsThe DDRG embeddings learned by HeteroRepur accurately capture graph information, including 89 types of interaction or associations among drugs, genes, pathways, and GOs, as well as disease genetic network of MDD. In drug classification task, HeteroRepur outperforms various baseline models, including existing graph neural network approaches that do not consider heterogeneous network information. Finally, HeteroRepur predicts new candidates for treating MDD, including CNS drugs that have the potential to be extended to MDD, and dietary supplements that might benefit depression patients by elevating essential nutrient levels. ConclusionHeteroRepur outperforms traditional graph-based approaches, and can be used to rank promising drugs among complex disease networks with heterogeneous topology.