Family-Specialized Transformer for L-cystathionine gamma-lyase Engineering and Its Structural Interpretation

The diversity of protein structures and reaction mechanisms complicates general-purpose artificial intelligence models for enzyme engineering, motivating family-specialized models. In this study, we developed EnzFormer, a specialized artificial intelligence pipeline for engineering Staphylococcus aureus L-cystathionine gamma-lyase (SaMccB). To overcome the scarcity of experimental labels, we used GPT-4o to generate putative activity labels for cystathionine gamma-lyase homologs, leveraging species-level ecological and evolutionary metadata as a proxy for functional selection. Using these labels, we trained a Transformer classifier on embeddings from the ESM Cambrian protein language model. From an exhaustive single-mutant library, in silico prioritization nominated 4 variants for testing and identified SaMccB V129G with a ~2-fold increase in catalytic turnover relative to the wild type. Val129 is distal to the active site; crystallographic and biochemical analyses suggest that V129G weakens local packing, thereby increasing the conformational flexibility of the active site loop, consistent with faster conformational steps in the catalytic cycle. Together, these results suggest that combining large language model-derived evolutionary priors with a family-specialized predictive model can identify distal mutations that modulate enzyme dynamics.