Dynamic conformation-aware protein language modeling enables structure-guided engineering of Candida antarctica lipase B

Abstract

This study presents a dynamic conformation-aware protein language modeling (DC-PLM) strategy for engineering Candida antarctica lipase B (CALB), a lipase whose catalytic efficiency is constrained by lid-dependent structural transitions. By integrating dual 3Di structural alphabet sequences derived from both open- and closed-lid conformations, together with a differential feature derived from dual-state SaProt embeddings and modulated via a Feature-wise Linear Modulation (FiLM) mechanism, the model effectively learns the coupled relationship between the sequence, structure, conformation and function. Iterative PLM-experiment feedback across five rounds led to the identification of high-performing mutants, culminating in a quintuple variant (M5) exhibiting an 8.6-fold increase in V_max and a 16.5-fold enhancement in k_cat. Molecular dynamics simulations revealed that beneficial mutations facilitate lid opening, stabilize the open conformation, and promote catalytically competent substrate orientations. When applied to industrial waste oil, engineered mutants significantly accelerated long-chain fatty acid hydrolysis, with M4 and M5 achieving reaction equilibrium within 1 hour compared to 5 hours for the wild type. These results demonstrate that incorporating dynamic structural information into PLMs enables mechanism-guided optimization of enzymes and provides a new framework for protein engineering.