Protein Language Model‐Guided Engineering of a 2,3‐Butanediol Dehydrogenase for the Enantioselective Synthesis of Cyclic <i>α</i> ‐Hydroxy Ketones

(2R,3R)-butanediol dehydrogenases (BDHs) are promising catalysts for the production of α-hydroxy ketones, which are highly valuable compounds in the synthesis of fine chemicals and pharmaceuticals. However, (2R,3R)-BDHs display limited stereoselectivity, thus restricting wider applications. In this study, we engineered a (2R,3R)-BDH from Bacillus subtilis (BsBDH) to enhance and invert its stereoselectivity toward 1,2-cyclohexanediol (1,2-CHD) for the production of chiral 2-hydroxycyclohexanone. The hot spots 115, 118, 293 of BsBDH were initially identified using the protein language model ESM-1v. Subsequently, to obtain a stable scaffold to engineer stereoselectivity, we devised a strategy of position analysis and source search, achieving a true-positive rate of 88.2% in designing thermostable single variants. Furthermore, iterative saturation mutagenesis was applied to the hot spots of the thermostable variant 6M2, and obtained a trans-CHD preference variant LTF (ee > 99%) and a cis-CHD preference variant 10M (ee > 99%). Several high-activity variants were also obtained, including 6M2/F115C/L118F and 6M2/F115L/L118M, which demonstrated the activity improvements toward 25 substrates, with the highest enhancement reaching 5183.1-fold. Additionally, molecular dynamics (MD) simulations and the incorporation of non-canonical amino acids (ncAAs) were utilized to elucidate the mechanisms underlying the variants. The engineered BsBDH variants exhibit promising potential for the biocatalytic production of α-hydroxyketones.