High-throughput physics-based enzyme engineering 1. The study presents a scalable, physics-based enzyme-engineering workflow that reaches quantum-mechanical accuracy for enzyme electrostatics while staying fast enough for high-throughput, atomistic free-energy calculations of multistep catalytic mechanisms. 2. A central contribution is MACEPOL-EF, an electric field-aware machine-learning interatomic potential (MLIP) designed for the unusually strong and heterogeneous electric fields found in protein active sites (roughly 0.49–1.60 V Å−1, consistent with Stark-effect measurements). 3. To train for that regime, the authors build BioPol-EF: a DFT-labeled dataset (~280k entries) where external electrostatic fields are systematically reconstructed around molecules to match protein-like, element-dependent field statistics, addressing a known failure mode of field-agnostic ML potentials in ML/MM. 4. Architecturally, MACEPOL-EF extends MACEPOLAR-1 with (i) per-atom charge–potential (qV) coupling so atoms can sense anisotropic external potentials/fields, and (ii) an adaptive polarization head that captures linear polarization plus higher-order response, while retaining physically meaningful atomic charges. 5. The paper implements a bidirectional electrostatic-embedding ML/MM scheme in AMBER: the ML region responds to MM-derived potentials/fields, while the MM region responds to on-the-fly ML electrostatics via analytic force propagation, yielding stable energy conservation and a physically consistent hybrid potential. 6. The practical target is an engineered oxidative amidase (OxiAm) derived from aldehyde dehydrogenase (ALDH) that forms amides from aldehydes + amines, competing against native hydrolysis (acid formation). Using enhanced sampling (OPES-MetaD), they compute full reaction free-energy landscapes and extract minimum-energy pathways and transition-state ensembles. 7. Mechanistically, they resolve a multistep pathway: Cys293 thiolate attack → tetrahedral intermediate → hydride transfer to NAD(P)+ → metastable thioester → branch point where either water (hydrolysis) or amine (aminolysis) attacks, with proton transfer mediated by a preorganized water-assisted H-bond network. Ensemble analysis shows early steps are rigid/preorganized, while hydrolysis/aminolysis have broad, heterogeneous TS ensembles. 8. A key modeling advance is microkinetic integration of multiple barriers (thioester formation, aminolysis, hydrolysis) rather than assuming a single rate-determining step. This improves kcat agreement with experiment (|Δlog| ~0.67 vs ~2.28 when approximating with aminolysis alone) and enables quantitative prediction of selectivity (amide/acid), including a saturation-mutagenesis panel at position 259. 9. The framework generalizes across substrate scope and homologous scaffolds: calculated log10(kcat) correlates with experimental yields across 25 aldehyde–amine combinations (r = 0.48), and the same four mutations (L163G, R166M, T234H, E259L) shift selectivity toward amidation across multiple ALDHs via scaffold-dependent atomistic mechanisms (sometimes improving selectivity at the cost of turnover, sometimes improving both). 10. In a pharma-oriented case study (ABBV-318 route), the authors use transition-state interaction energy decomposition to nominate redesign positions (including a non-obvious site F454), run virtual saturation mutagenesis with ML/MM free-energy barriers, and experimentally validate improved variants (e.g., G163C ~80% yield at pH 11; F454H ~60% yield at pH 7), illustrating how TS-ensemble physics can guide mutations beyond previously explored sites. 💻Code: github.com/ClickFF/MLMM4AMBE… ; github.com/Hsuchein/AutoNACC ; github.com/ClickFF/MACEPOL-E… 📜Paper: biorxiv.org/content/10.64898… #ComputationalBiology #EnzymeEngineering #Biocatalysis #MolecularSimulation #FreeEnergy #MLMM #QMMM #AMBER #MachineLearning #Microkinetics