Redesign to Mechanism: Interpretable AI Reveals Determinants of Protein Hydrate Binding

Natural antifreeze proteins (AFPs) demonstrate exquisite structure-function relationships, while in silico designs often struggle to achieve a balance among structural stability, expression efficiency, and functional activity. This study establishes a computational framework integrating deep learning, molecular dynamics simulations, and interpretable machine learning for the rational design of the AFPs. The Chill+ algorithm showed that 83.4% of the designed peptides exhibited superior hydrate inhibition activity relative to wild-type Tenebrio molitor AFP (TmAFP). XGBoost and SHAP analyses revealed that spatially defined structural features more accurately predicted inhibitory activity than statistical sequence features. Asn₂₉ stabilizes the hydrate lattice via bifunctional hydrogen bonding and hydrophobic guest mimicry, while Ser₄-Ile₁₇ mutations optimize the hydrophilic water ordering and hydrophobic anchoring. Val mutations further revealed two functional regions within the hydrate-binding site: an ordered anchoring region and a dynamic perturbation region. This work establishes an interpretable and generalizable framework for engineering high-performance AFPs.