Combinatorial Protein Language Model-Guided Engineering of TEV Protease for Enhanced Stability and Production

Abstract

Industrial enzyme engineering focuses on improvement of enzyme production yield, stability, catalytic activity, and substrate specificity, but often suffers from low efficiency with time-consuming and labor-intensive design and screening processes of massive libraries. Recent advances in AI and machine learning created protein language models trained by numerous datasets and shed new lights to speed up the enzyme engineering processes with high accuracy structural prediction. Here, we developed a highly efficient enzyme engineering strategy combining three protein language models (xTrimoMPNN-Thermo, ESM-IF, and MPNNsol) and use it to generate TEV protease variants with improved expression, stability, and function. The results indicated that a small number of TEV protease designs (<50 designs) were sufficient to develop variants with desired properties, demonstrating its high efficiency. Our strategy could be broadly applied to accelerate designing and engineering various industrial enzymes.