Practical Machine Learning-Assisted Design Protocol for Protein Engineering: Transaminase Engineering for the Conversion of Bulky Substrates

Protein engineering is essential for improving the catalytic performance of enzymes for applications in biocatalysis, in which machine learning provides an emerging approach for variant design. Transaminases are powerful biocatalysts for the stereoselective synthesis of chiral amines but one major challenge is their limited substrate scope. We present a general and practical variant design protocol for protein engineering to combine the advantages of three strategies, including directed evolution, rational design, and machine learning, and demonstrate the application of the protocol in the protein engineering of transaminases with higher activity toward bulky substrates. A high-quality data set was obtained by rational design of selected key positions, which was then applied to create a machine learning model for transaminase activity. This model was applied for the data-assisted design of optimized variants, which showed improved activity (up to 3-fold over parent) for three bulky substrates, maintaining enantioselectivity of the starting enzyme scaffold as well as improving the enantiomeric excess (up to >99%ee).