Machine learning-driven lipase engineering for highly enantioselective synthesis of chiral florfenicol intermediate via enzymatic resolution-chemical racemization coupling system

(2 S ,3 R )- p -methylsulfonylphenyl serine ethyl ester [(2 S ,3 R )- p -MPSE], a key chiral building block of the aryl β-hydroxy-α-amino ester family, serves as one of the critical intermediates for the synthesis of the veterinary antibiotic florfenicol. Kinetic resolution of rac - threo - p -MPSE by lipases represents a highly promising route. However, low catalytic efficiency of natural enzymes restricted its application. In this study, a lipase BSL from Bacillus subtilis was mined and rational design was performed to obtain a series of mutants. A robust Partial Least Squares (PLS) regression model, integrated with multi-dimensional encoding from the Amino Acid Index Database (AAindex) and Fast Fourier Transform (FFT)-based data processing technology, was constructed and trained. Using this machine learning (ML)-driven enzyme engineering approach, a mutant E115F/L67D/I421Q/P317I was successfully created, showing a 15.4-fold increase in catalytic efficiency ( k cat / K m ). Under the load of 1 M high substrate concentration, this mutant could achieve complete resolution of rac - threo - p -MPSE within 4 h, yielding optically pure (2 S ,3 R )- p -MPSE ( e.e. s  > 99%) with a yield of 97.4% (based on the theoretical amount of single enantiomer). Additionally, the inactive isomer (2 R ,3 S )- p -methylsulfonylphenyl serine [(2 R ,3 S )- p -MPS] was recycled via chemical racemization based on a reversible retro-aldol cleavage mechanism and re-esterification, which facilitated the recycling of the substrate. This study not only breaks through the bottleneck of low catalytic efficiency of natural lipases through a machine learning-mediated precise enzyme engineering strategy, but also establishes a high-benefit chiral preparation process for (2 S ,3 R )- p -MPSE via an enzymatic resolution-chemical racemization coupling system.