DRAGON: Drug Repurposing via Graph Neural Networks with Drug and Protein Embeddings as Features

Since the pandemic, drug repurposing has become a helpful technique for associating treatments to new or already known diseases. Drug repurposing finds new uses for existing drugs, leading to a more affordable solution than de novo drug development. The reduction in time and costs that drug repurposing provides makes it an effective technique to accelerate the process of discovering new treatments. In the present study, we apply a graph deep learning approach to a heterogeneous biomedical graph, aiming to predict a specific link type that connects diseases with drugs, in order to put forward drug repurposing opportunities. In particular, we generate a new model called DRAGON, which builds upon a two-layered Graph Neural Network pipeline. In the encoder stage, drug and protein nodes are initialized with embeddings representing molecular structure and amino acid sequence information. We compare the proposed model to previous baselines that studied the disease-drug prediction approach but did not consider the initialization with embeddings of these two node types. DRAGON reports an improvement of 0.02 in the area under the precision-recall curve when compared to these baselines. We hypothesize that the repurposing model may benefit from the inclusion of multimodal information from different sources.