Personalized Peptide–Disease Recommendation Using Knowledge Graphs and Graph Neural Networks

Abstract

Personalized peptide–disease recommendation represents a complex socio-technical challenge that sits at the intersection of biomedical knowledge representation, machine learning, clinical decision support, and regulatory governance. This paper examines how knowledge graphs and graph neural networks can be integrated into recommendation infrastructures that connect peptide candidates, disease mechanisms, patient phenotypes, and therapeutic evidence. Rather than treating recommendation as a purely predictive task, the analysis frames it as a systems problem involving heterogeneous data ingestion, ontology alignment, multi-modal patient representation, uncertainty management, fairness, auditability, deployment, and long-term sustainability. Knowledge graphs provide a structured substrate for integrating molecular, phenotypic, clinical, and literature-derived evidence, while graph neural networks offer inductive and relational learning capabilities suited to sparse and evolving biomedical networks. The paper discusses architectural trade-offs among centralized, federated, and hybrid deployment models, with attention to privacy, institutional trust, and resource asymmetry. It also analyzes the risks of bias amplification, evidence provenance loss, and overfitting to well-studied diseases or populations. The argument is that personalized peptide–disease recommendation should be governed as a learning health system, where model updates, clinical validation, pharmacovigilance, and stakeholder oversight are coupled. The conclusion emphasizes that technical performance alone is insufficient; robust recommendation requires transparent evidence pathways, continuous evaluation, fairness-aware design, and policy mechanisms that align algorithmic outputs with clinical accountability and equitable access.