Improving DDI Prediction Performance of Node2vec Embedding Using Graph Neural Network Based Models

A drug-drug interaction (DDI) is a reaction between two or more drugs that can reduce or increase the reaction of a medicine synergistically or cause adverse side effects. DDI detection, therefore, is an important objective in patient safety and pharmaceutical industry. Many researchers try to predict the DDI of unknown drugs by training the known DDI data in-silico approaches. In-silico approaches can be categorized into three groups: knowledge-based, similarity-based, and graph-based. Among them, graph-based approaches are known to have achieved great performance by casting DDI prediction as a link prediction problem on DDI graphs. In this paper, we explore how we can improve DDI prediction performance of the embedding learning method node2vec[6] using representation learning algorithms of graph neural networks (GNNs). We first created and trained node2vec model to obtain initial drug features; then we used three GNN based models to improve the learned node2vec drug embedding; finally, we used four different classifiers to implement link prediction, which is DDI prediction. Our experimental results showed that all four classifiers performance were improved using GNN learned embedding.