Affinity Engineering of a Hyaluronic Acid-Specific Carbohydrate-Binding Module Using a Rule-Constrained Large Language Model Agent

Abstract Hyaluronic acid (HA) is widely used in food and nutritional supplements, yet reliable quantification in complex matrices remains limited by the lack of cost-effective, high-affinity recognition elements. Carbohydrate-binding modules (CBMs) are stable, recombinantly accessible, and intrinsically selective for polysaccharides. However, their native affinity is often insufficient for sensitive analysis. Here, we developed a rule-constrained large language model (LLM)-based agent for affinity optimization of the HA-binding CBM SrCBM70. The agent converts structural, energetic, and evolutionary information into rule constraints that compress the mutation search space, then refines design through iterative feedback from experimental affinity data. After two rounds of iterative optimization, the double mutant E29W/T112S reached a Kd of 154 ± 7 nM, a 20.9-fold improvement over WT SrCBM70, and retained HA specificity. Mechanistic analysis suggested that E29W enhanced CH−π interactions and T112S expanded the protein–HA contact interface. This work provides a nanomolar HA-recognition element for food analysis and establishes an agentic protein engineering strategy with potential extension to CBM affinity engineering.