Guan, J.; Shang, J.; Peng, C.; Sun, Y.

Viruses play indispensable roles in ecosystems and human health, yet deciphering their molecular functions remains challenging. Many viral protein annotations are incomplete or poorly characterized. Existing tools typically predict functional categories without linking to verifiable evidence, hindering the credibility of functional interpretation. Here, we present VirProtRAG, a viral protein function annotation framework that integrates information retrieval with evidence-grounded knowledge generation. It introduces three task-adapted components: a hybrid retrieval module combining keyword-based and semantic dense retrieval to maximize literature coverage, synonym-expanded and rank-aware retrieval with reciprocal rank fusion for improved search effectiveness, and literature quality and evidence-oriented re-ranking to enhance reliability and interpretability. Results show that hybrid retrieval strategy performed best, with quality and evidence features further enhancing re-ranking. Compared with direct LLM prompting without retrieved literature, it consistently improves generation performance, underscoring the critical role of external knowledge. Finally, we built a searchable database comprising all 17,484 reviewed Swiss-Prot viral proteins, supporting both sequence- and text-based queries. VirProtRAG introduced 32.53% non-overlapping function annotations beyond existing expert curation, and independently supported 56.34% of sequence-inferred function points with retrieved literature. Case studies further demonstrate its capability to augment and refine the characterization of previously unannotated or poorly understood viral proteins.