Explainable link prediction (drug – disease interaction) on Biomedical Knowledge Graphs with the use of Graph Neural Networks

Abstract

This thesis addresses the challenge of drug repurposing by developing a pipeline for explainable link prediction on biomedical knowledge graphs. The study utilizes Graph Neural Networks to tackle the issue of graph incompleteness and predict novel drug-disease interactions from the iASiS Open Data Graph. A Graph Convolutional Network is trained to identify potential therapeutic links. Then, the GNNExplainer method is employed to provide interpretable explanations for each prediction by highlighting the most influential subgraph structures and node features. The model achieved strong predictive performance, identifying 44 potential novel drug-disease interactions, with the generated explanations allowing for qualitative assessment and enhancing confidence in the results.