Link prediction using deep learning approach for type 2 diabetes drug repurposing

Abstract

There is still no effective treatment for type 2 diabetes, which has been on the rise for years. By repositioning current medications for new indications, drug repurposing can aid in the discovery of novel medications. Deep learning has recently been applied to this problem via link prediction utilizing a graph representation that learns from either the structure of a graph or the semantic meaning of entity text. However, because they used a single representation as the basis for their work without making any model improvements, earlier attempts still had restricted performance. In this study, we suggest a new deep-learning approach for the drug repurposing of entities associated with type 2 diabetes. Transformer, a current deep learning network, serves as the foundation of our model's architecture. Regarding our link prediction in the graph, each entity is embedded utilizing both (1) structural information embedded from the node and its neighbor nodes and (2) semantic information retrieved from its name and descriptions. The experiment was conducted using type 2 diabetes data gathered from PubMed and UMLS Metathesaurus. The findings demonstrated that our combined model can outperform other models that only contain a single module, i.e. StAR and HittER, by exhibiting an increase of 77.17% on the mean reciprocal rank score for the drug discovery task. Finally, using the model for drug repurposing, we can identify several medications that may be employed to treat type 2 diabetes.