Drug Repurposing for Type 2 Diabetes Using Combined Textual and Structural Graph Representation Based on Transformer
Sothornin Mam, Duangdao Wichadakul, Peerapon Vateekul · IEEE Access · 2023
Type 2 diabetes has been predominant for years. There is still no cure to help patients completely recover from the sickness. Drug repurposing can help to discover a new drug by repositioning existing drugs for new medications. Prior attempts, however, to resolve this issue are found to be limited. Performance relied on a single representation without any improvement being achieved. To surmount this task, a deep learning approach is applied. Link prediction using graph neural networks is used. Our model’s architecture is based on a "Transformer": a recent and popular deep learning network. For our link prediction in the graph, each entity is embedded by using both: (1) structural information sourced from a node and its neighbor’s node, and (2) semantic information extracted from its name and descriptions. The experiment is conducted via type 2 diabetes data that we extracted and prepared from PubMed and UMLS Metathesaurus. Results demonstrate that our combined model, by showing an increase of 77.17% on the mean reciprocal rank (MRR) score for the drug discovery task, can outperform other models that only have a single module i.e. StAR and HittER. Finally, after deploying the model for drug repurposing, we can discover some drugs that have the potential to be used for type 2 diabetes treatment.