Network medicine provides a powerful framework for understanding complex diseases by integrating diverse biomedical entities into a unified graph structure. This systems-level perspective enables predictive modeling of disease mechanisms and facilitates the identification of therapeutic interventions. Recent advances in deep learning combined with biomedical knowledge graphs (KGs) have shown promise in various predictive tasks for understanding disease mechanisms and drug discovery. However, most existing models remain task-specific, limiting their ability to generalize across diverse tasks within a unified framework. Here, we introduce NetMedGPT, a transformer-based foundation model trained on a large-scale biomedical KG using masked token prediction. By learning contextualized representations of biomedical nodes, NetMedGPT enables unified, zero-shot inference across different drug discovery tasks. Specifically, in five tasks, including predicting the association of drugs with indications, targets, adverse drug reactions, contraindications, and off-label uses, NetMedGPT consistently outperforms all specialized baselines, achieving AUPRC gains of between 2% and 35%. When applied on ClinicalTrials.gov data, it preferentially predicts drug–disease pairs with real-world therapeutic relevance. Notably, NetMedGPT's generative capacity enables it to construct plausible and druggable disease mechanisms, offering interpretable biological insights via an interactive web tool (https://prototypes.cosy.bio/netmedgpt/). These findings highlight NetMedGPT as a network medicine foundation model that unifies diverse biomedical signals, supports scalable hypothesis generation, and holds strong potential to accelerate drug repurposing, even in data-scarce or rare disease contexts.
Jan Baumbach (Mon,) studied this question.