UniGEN-DDI: Computing Drug-drug Interactions Using a Unified Graph Embedding Network

Abstract

Finding drug-drug interaction is crucial for patient safety and treatment efficacy. Two drugs may show a synergistic effect but may sometimes cause a severe health issue, including lethality. Wet lab studies are often performed to understand such interactions but are limited by cost and time. However, the biochemical data generated can be explored to compute unknown interactions. Here, we developed a computational model named UniGEN-DDI (Unified Graph Embedding Network for Drug-Drug Interaction) for the estimation of interactions between drugs. It is a simple unified network model containing the biochemical information of drug association developed using the compiled data from DrugBank 5.1.0. The feature learning of the drugs was carried out using a combination of GraphSAGE and Node2Vec algorithms, which were found efficient in extracting diverse features. The simple architecture of our model led to a significant reduction in computational time compared to the baselines, while maintaining a high prediction accuracy. The model performed well on the data, which was equally distributed between interacting and non-interacting drugs. As a more challenging evaluation, we performed non-overlapping splitting of the data based on the drug action on different parts of the body, and our model performed well in both interaction estimation and time efficiency. Our model successfully identified 15 previously unknown drug interactions in DrugBank 5.1.0 in the top 20 estimates (75% correct prediction), which were confirmed from experimental studies in updated DrugBank 6.0. In the top 50 estimates, 68% of the unknown drug interactions were correctly identified that showcased a strong performance of our model.