SemVac: A Semantic Vaccinology Paradigm Powered by LLMs for Antigen Discovery

Abstract

Reverse vaccinology identifies vaccine antigens from pathogen genomes, yet existing methods rely mainly on sequence and structure and overlook the functional and immunological knowledge recorded in the published literature. We introduce semantic vaccinology, a paradigm in which large language models (LLMs) reason over literature-derived protein descriptions to predict protective antigens. Implemented as SemVac, the workflow retrieves publications linked to each protein through PaperBLAST, condenses the evidence into a structured semantic profile, and prompts an LLM to return an antigenicity probability. Benchmarked against a curated 246-protein bacterial benchmark and the specialized protein-language and geometric-deep-learning predictor PLGDL, the best of 14 general-purpose LLMs matched or exceeded the precision of PLGDL; the open-weight Kimi K2 0905 offered the strongest performance-cost balance. Predictions were robust to masking of vaccine keywords, reproducible across repeated inference, and generalized to a 1,200-protein cross-pathogen dataset. Surprisingly, explicit chain-of-thought reasoning increased recall but lowered precision in every model tested, revealing over-reasoning in biological scoring. Applied to the mpox virus proteome, SemVac recovered the established mpox antigen repertoire and prioritized uncharacterized candidates. Two of these, A30L and C19L, received independent experimental support in recent orthopoxvirus vaccine development, providing external validation. For one candidate, B20R, the model generated a coherent but false TNF-decoy narrative unsupported by curated annotations, demonstrating that LLM confabulation can be detected when reasoning traces are cross-checked against curated resources. Semantic vaccinology therefore establishes the literature as an explicit, auditable third modality alongside sequence and structure, while making its failure modes transparent and correctable.