AI and the World's Hard Problems · Life sciences · arXiv · 2026-10-23
QProtFuse: Query-Guided Layer Fusion for Retrieval-Augmented Protein Function Prediction
LLMs have limited protein understanding, and existing protein retrieval-augmented methods largely expand external retrieval signals while overlooking knowledge inside a protein language model. The authors propose QProtFuse, which uses queries to adaptively fuse features from multiple layers and jointly learns protein and annotation retrieval. On Prot-Inst-OOD it achieves the best retrieval scores, and it yields the highest LLM-as-a-judge scores for downstream function descriptions with Qwen3-8B and DeepSeek-V4-Pro.