Graph Neural Networks Reveal Candidate Protein Biomarkers Underlying Abdominal Aortic Aneurysm Biology

Abdominal aortic aneurysm (AAA) remains difficult to detect early due to its silent progression and lack of reliable molecular biomarkers. Computational prioritization of disease-associated proteins from protein-protein interaction (PPI) networks offers a scalable alternative to traditional experimental discovery, yet no prior study has applied graph neural network (GNN) architectures to AAA-specific biomarker prediction. We address this gap by developing a positive-unlabeled (PU) learning framework combining message-passing GNNs with PU-specific loss functions to rank candidate AAA-associated proteins across a proteome-wide PPI network, requiring only literature-curated positive labels without plasma samples or experimental validation data. Using 4 PU models on a PPI network of 8,300 proteins and 29,744 interactions, we identified 182 candidate proteins, of which 19 were prioritized through network topology, functional enrichment, and disease relevance. These proteins, including integrins, extracellular matrix regulators, and inflammatory mediators, map to core vascular processes implicated in AAA. Gene Ontology and Kyoto Encyclopedia of Genes and Genomes enrichment supported their mechanistic roles, and transcriptomic and microRNA evidence corroborated several predictions. Notably, COL6A3 has been independently identified as the strongest causal protein signal for AAA in a proteome-wide Mendelian randomization study, providing external validation of our framework. We additionally report 96 novel candidates not retained after enrichment filtering, including 17 predicted independently by all 4 models, representing potential new AAA biology for experimental follow-up. All candidates require prospective experimental validation before clinical application. This work provides a systems-level computational framework for mechanistically grounded AAA biomarker discovery and highlights high-priority candidates for experimental and translational follow-up.