ProtQueSt: Query-Conditioned Retrieval-Augmented Generation for Protein Function Annotation

Protein function prediction from sequence is inherently query-dependent: the same protein may be characterized by its catalytic activity, domain architecture, or cellular localization depending on the biological question. Prior work has shown that large language models consistently underperform retrieval-based methods on this task, yet simple retrieval transfers annotations from embedding-space neighbors without adapting to the query. We introduce ProtQueSt, a retrieval-augmented framework that pairs a structure-aware retriever with a query-conditioned contrastive retriever that aligns protein and annotation representations via Feature-wise Linear Modulation (FiLM). FiLM conditioning and query-pooled negative sampling prove jointly essential, as neither alone improves over the structural baseline. ProtQueSt achieves the highest Entity-BLEU (48.79, +37% over RAPM) and LLM-as-a-judge score reported on Prot-Inst-OOD. This result supports reframing text-based protein understanding as a query-conditioned retrieval problem rather than a single fixed sequence-to-text mapping.