AI-Driven Hotspot Prediction and Rational Substitution Engineering for Enhanced Enantioselectivity in R/S Transaminases

Pyridoxal 5′-phosphate (PLP)-dependent amine transaminases (ATAs) are key biocatalysts for the asymmetric synthesis of chiral amines, with broad applications in pharmaceutical manufacturing. While (S)-selective ATAs are well characterized, (R)-selective counterparts—critical for industrial processes such as sitagliptin synthesis—remain comparatively underexplored. Here, we present a comparative structural and computational analysis of fold type I and fold type IV ATAs to elucidate the determinants of enantioselectivity. Structural superposition and AlphaFold2-assisted modeling reveal a mirrored “handshake-like” inversion of substrate-binding pockets that underpins enantiocomplementarity between R- and S-selective enzymes. To complement structural insights, we explored a structure-informed machine learning framework incorporating residue-level interaction descriptors and physicochemical mutation features. Using an enzyme-level GroupKFold cross-validation strategy, the XGBoost classifier achieved a mean ROC–AUC of 0.72, demonstrating moderate predictive performance for identifying stereocontrolling hotspot residues from integrated structural interaction descriptors and physicochemical mutation features. Feature attribution suggests that stereocontrolling hotspots are primarily located within flexible loop regions and pocket-defining residues rather than conserved catalytic motifs. Collectively, this study presents an integrated comparative framework that combines structural superposition, molecular docking, molecular dynamics simulations, curated engineering datasets, and structure-informed machine learning to systematically compare fold I and fold IV transaminases. By integrating newly generated computational analyses with curated literature evidence, this work identifies transferable structural determinants and engineering hotspots that can guide future transaminase engineering.