Benchmarking Artificial Intelligence Algorithms for Drug Repurposing on Open Targets Platform

<p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" class="first" dir="auto" id="d6297330e179">Computational methods have become essential for prioritizing drug repurposing candidates by integrating diverse biological and chemical data sources, enabling efficient and systematic identification of potential therapeutic applications. However, the lack of standardized benchmarks hinders the reliable evaluation of Deep Learning algorithms for drug repurposing. In this study, we designed and evaluated a benchmark framework for drug-disease association prediction using an integrative and publicly available platform, the Open Targets dataset.First, we developed an Open Targets Knowledge Graph (KG) based on entities like drug, disease, gene, and their annotations like disease therapeutic area, drug type, and gene pathway. Since the dataset is updated multiple times per year, it offers a natural separation between versions, enabling retrospective validation over time. We ensured models were tested on real-world scenarios where predictions were validated against newly available information. Additionally, we refined the disease and drug dataset by incorporating key ontological relationships to prevent data leakage.We applied various graph neural networks (GNNs), including Graph Convolutional Networks (GCN), GraphSAGE, and a Transformer-based GNN model. Our results demonstrated that the models successfully predicted novel drug-disease associations. The models showed strong and consistent performance in drug-disease prediction, achieving AUC values around 0.85 in retrospective validation. Our model also identified potentially novel drug-disease associations yet to be validated, providing valuable insights for drug repurposing.This retrospective benchmark ensures reproducibility, enables fair model comparison, and serves as a foundation for developing more interpretable and clinically relevant AI-driven approaches. Future research will explore integrating multi-omics data to improve effectiveness for specific rare diseases.