PU-GRAIL: residue-level graph learning for identifying protective bacterial antigens under positive-unlabeled supervision.

motivationThe identification of protective antigens is fundamentally constrained by sparse annotations and pervasive label uncertainty in reverse vaccinology. Protective antigen-antibody interactions are mediated by a limited subset of surface-accessible residues that form spatially coherent epitopes. This motivates modeling antigenicity at the residue level, where 3D structure provides critical context for functional immune recognition. Moreover, antigen datasets are inherently positive unlabeled: unannotated proteins may contain hidden positives, making reliable negatives difficult to obtain.

resultsWe present PU-GRAIL, a graph neural network framework that integrates protein language model embeddings with predicted 3D structures under positive-unlabeled learning. Trained and evaluated on three benchmark datasets, PU-GRAIL achieves competitive performance compared with existing methods. Importantly, the model's attention mechanism enables residue-level interpretation, identifying putative epitope regions that correspond to experimentally validated antibody-binding sites. Beyond standard benchmarks, we demonstrate practical utility through (i) severity-associated antigenicity patterns in SARS-CoV-2 patient cohorts, and (ii) proteome-wide vaccine candidate prioritization across 11 bacterial species. AVAILABILITY AND IMPLEMENTATION: The PU-GRAIL software is available at https://github.com/jaeminjj/PU-GRAIL.