Large Image

Abstract

Recent advances in AI-driven protein design, including ProteinMPNN for inverse folding, have created opportunities for enzyme engineering. ProteinMPNN can produce artificial enzymes with high thermal stability and soluble expression, but they often reduce activity. To address how such activity can be enhanced, we performed stepwise redesign of NAD⁺-dependent L-threonine 3-dehydrogenase (TDH) using ProteinMPNN combined with structure- and sequence-based redesign. Using Cupriavidus necator TDH (CnTDH) as the starting point, we generated MPNNTDH by ProteinMPNN through setting frozen residues corresponding to 33% of highly conserved residues among CnTDH homologs. MPNNTDH showed a T_m over 20 °C higher than CnTDH, but its k_cat and k_cat/K_m were ~5,000-fold and ~100,000-fold lower, respectively. Introducing evolutionarily adaptive mutations and correcting sequence-context discrepancies improved activity, with the best variant showing >50-fold and >1,000-fold increases in k_cat and k_cat/K_m, respectively, compared with those of MPNNTDH. However, the k_cat and k_cat/K_m values of the best mutants obtained through redesign of MPNNTDH remain two orders of magnitude lower than those of CnTDH. These results indicate that introducing evolutionarily adaptive mutations into ProteinMPNN-designed inactive variants can improve activity, but is insufficient to restore activity to native levels.