Abstract
The selection of an effective adjuvant is a critical bottleneck in vaccine development, particularly for emerging diseases where experimental data is sparse. While computational methods exist for antigen discovery, the task of directly ranking adjuvants for a given disease context remains largely unaddressed. In this work, we frame disease-adjuvant selection as a top-k recommendation task on a heterogeneous knowledge graph grounded in biomedical ontologies. We propose VaxjoGNN, a graph neural network that learns predictive representations of diseases and adjuvants by integrating curated facts, mechanistic pathways, and textual evidence. A key innovation is our listwise training objective, which combines ApproxNDCG and ListNet surrogate losses to directly optimize for ranking performance with improved stability. On a public benchmark, our model achieves a Normalized Discounted Cumulative Gain (NDCG@10) of 0.59 on seen diseases and demonstrates strong generalization to novel diseases, achieving a 5.4-fold improvement over random baselines (NDCG@10=0.27). Our work provides a novel, ontology-anchored, AI-driven framework for prioritizing vaccine adjuvants, with the potential to accelerate vaccine research and development.