Deep-Learning-Driven Discovery and Engineering of Imine Reductase for the Efficient Synthesis of Chiral Opine Derivatives

Abstract The narrow amine donor scope of imine reductases (IREDs) limits their application in the synthesis of opine compounds. Herein, we report a machine learning (ML)-guided protein engineering strategy to address this limitation. An IRED from an unclassified Ensifer, IRED5, was identified through deep-learning-based screening for its strict enantioselectivity toward methyl glycinate. Subsequent engineering using a few-shot learning for protein fitness prediction approach yielded the M178T/A208L/P241V/A36E/S67Q/Q187N (M6) variant with six mutations across the active site and protein surface. Kinetic analysis showed 156-fold toward methyl pyruvate and 72-fold improvements toward methyl glycinate in catalytic efficiency, respectively, over the wild-type (WT) enzyme. M6 accepted diverse keto esters and amino acid esters, enabling stereoselective formation of opine derivatives bearing two stereocenters. Structural and computational analyses revealed the molecular basis for activity enhancement. In preparative synthesis, M6 achieved complete conversion at a 3 M substrate loading, corresponding to a space time yield of 366 g·L–1·d–1. At a 4 M substrate loading, conversion remained high at 95.4% with a space time yield of 525 g·L–1·d–1, affording 10.5 g of enantiopure methyl N-(2-methoxy-2-oxoethyl)-d-alaninate (1a) in 75.0% isolated yield with > 99.9% ee. This work establishes an efficient biocatalytic route to opine compounds and demonstrates the utility of ML-guided enzyme engineering in expanding IRED synthetic capability.