Enerzyme: A Framework for Efficient Training of Reactive Neural Network Potentials for Enzyme Catalysis with Application to Methyltransferases

Abstract: Quantum mechanical (QM) cluster models provide an effective framework for mechanistic studies of enzymatic reactions but remain computationally demanding. Neural network potentials (NNPs) offer a promising route to reduce this cost, but enzymes present challenges beyond small molecules, including large system sizes, implicit-solvent environments, substantial polarization, and charge transfer. Here, we present an integrated software framework for efficient NNP training for mechanistic studies of enzymes, demonstrated on QM cluster models of S-adenosyl-L-methionine-dependent methyltransferases (MTases). Our Enerzyme code introduces modular electrostatics-aware NNP architectures and combines automated QM-cluster construction with reactive dataset generation. The Enerzymette subpackage automates reaction pathway exploration at both NNP and DFT levels. We show that iterative flexible scans and nudged elastic band calculations impose stricter requirements on NNPs than conventional dataset metrics. Nevertheless, NNPs trained on fewer than 1,000 system-specific datapoints reproduce reaction energetics and transition-state structures for MTase clusters containing up to 545 atoms with near-chemical accuracy. Direct supervision of atomic charges and consistent dielectric screening substantially improve simulation stability and accuracy, while multitask-learned atomic charges capture charge transfer and polarization trends and provide chemically meaningful descriptors of reactivity. Finally, transferability across chemically diverse catechol O-methyltransferase substrates indicates that NNPs learn generalizable reactivity patterns as training data expand across multiple enzymes. Together, these results establish a foundation for accelerating enzyme mechanistic studies and guide future NNP development for biomolecular reactivity.

• The paper demonstrates that incorporating electrostatics directly into NNP architectures enhances the accuracy of enzyme catalysis simulations.
• It presents an automated QM cluster data generation pipeline and a modular workflow that reduces computation time by 2-3 orders of magnitude compared to traditional DFT methods.
• The framework achieves sub-2.5 kcal/mol error in transition state barriers and shows promising transferability across diverse methyltransferase systems.

Enerzyme: A Modular, Electrostatics-Aware Framework for NNPs in Enzyme-Catalyzed Reaction Modeling

Motivation and Challenges in Enzyme NNP Development

The accurate and efficient modeling of enzymatic catalysis at quantum mechanical (QM) accuracy is a computationally intensive task due to the large system size, complex electrostatics, and prevalent charge transfer phenomena in proteins. While QM cluster models have driven progress by enabling mechanistic insights, their computational cost restricts systematic explorations across large-scale enzyme and substrate spaces. Recent progress in neural network potentials (NNPs) suggests a potential for significant acceleration, but existing NNPs built for small molecules struggle with enzyme-scale systems where environment-induced polarization, implicit solvation, and long-range electrostatics critically govern reactivity.

Enerzyme directly addresses these obstacles by introducing a modular, workflow-integrated software platform tailored for efficient training and deployment of reactive NNPs for enzymatic QM clusters. The study specifically demonstrates the framework on S-adenosyl-L-methionine-dependent methyltransferases (MTases), a mechanistically and functionally representative enzyme superfamily.

Automated Data Generation and the Enerzyme/Enerzymette Framework

An essential innovation is the integrated, reproducible pipeline for automatic QM cluster construction and chemically diverse dataset generation. Starting from PDB structures, the QuantumPDB package generates QM clusters centered on catalytic sites with explicit atomistic context, including ligands, cofactors, and coordinating ions. Reactive configurational sampling is accomplished via high-temperature steered MD combined with efficient quantum or universal NNP-level interatomic potentials (e.g., GFN1-xTB, UMA). The approach robustly samples rare transition region geometries and diverse protein-ligand conformers, providing less than 1,000 high-quality DFT-labeled points per system as sufficient for NNP training.

The Enerzyme platform modularizes message-passing GNN-based NNP architectures—supporting PhysNet, SpookyNet, and MACE—and introduces bespoke pre- and post-core layers for embedding system-level features (e.g., total charge) and explicit electrostatic computations. Enerzymette, the workflow manager, provides fully automated and standardized reaction-path and NEB exploration, harmonizing optimizer and workflow logic between NNPs and DFT, thus allowing direct fidelity benchmarking.

Electrostatics-Aware NNP Architecture: Physical Priors and Multitask Supervision

A central advance is in treating electrostatics as a first-class component of the NNP. Enerzyme generalizes explicit charge prediction (using the 1.2xCM5 charge scheme) and ensures global charge-conservation through linear correction. Beyond direct charge supervision, the architecture incorporates a long-range electrostatic energy module consistent with the implicit-solvent dielectric used in reference QM calculations. Further, the dipole moment is regressed jointly with energy, force, and charge supervisions, promoting multi-modal learning and rigorous physical prior incorporation. This is essential for stability and transferability, particularly for systems with extended charged moieties and mixed explicit/implicit solvation.

Numerically, direct charge supervision and inclusion of a physically consistent dielectric are shown to be vital for accurate energetics and recovery of chemically correct charge-transfer phenomena throughout methyl transfer pathways in MTases. Weaknesses or inconsistencies in these design choices degrade both structural accuracy and energetic reliability, sometimes leading to failure in converging NEB simulations or yielding unphysical TS geometries.

Rigorous Evaluation: Energetic, Structural, and Electronic Descriptors

Enerzyme's NNPs, trained system-specifically on less than 1,000 data points, achieve sub-2.5 kcal/mol errors in transition state barriers and reaction energies relative to DFT in large enzyme clusters (up to 545 atoms; e.g., HcgC), and bond/angle deviations in the TS region usually under 0.05 Å/2°. Importantly, simulation stability and correct mechanistic predictions under iterative NEB exploration are obtained, even where NNPs with minimal capacity or lacking explicit charge supervision fail. These workflows holistically benchmark NNP reliability for tasks that transcend mere single-point fitting.

Multitask-learned atomic charges do not simply reproduce the trained scheme, but emerge as robust descriptors that faithfully capture reaction-centric charge redistribution, including polarization of both the substrate and proximal residues. Intriguingly, the learned NNP charges show stronger correlation with reaction barrier trends across substrate sets than conventional DFT charge schemes, despite the latter serving as the training target. This can be attributed to the joint physics- and reactivity-driven supervisory signal.

Generalization and Data Efficiency: Transferability Across Enzyme/Substrate Space

Transfer learning and data efficiency are rigorously probed on the catechol O-methyltransferase (COMT) family with five diverse substrate clusters. While NNPs system-specifically trained on one substrate show limited zero-shot transfer to others with divergent chemotypes or reactivity, training on a multi-system combined dataset significantly boosts in-distribution and out-of-distribution accuracy, yielding mean absolute barrier errors near 2 kcal/mol across nine systems. Notably, the MACE-based NNPs can sustain this accuracy with lower data requirements and better out-of-distribution stability than SpookyNet variants.

A cost analysis reveals that, including both dataset labeling and model training, Enerzyme's workflow reduces wall-clock time by 2-3 orders of magnitude compared to direct DFT-based exploration when amortized across an enzyme family. However, for large-capacity NNPs, model training can approach or even exceed DFT compute cost, indicating future optimization targets related to training stability, active learning, and hyperparameter tuning.

Practical and Theoretical Implications

Enerzyme delivers a practical paradigm for rapid, physically reliable mechanistic simulation across enzyme families, enabling high-throughput explorations of reaction pathways, substrate screening, and potentially mutant or drug design that require QM-level fidelity. Its architectural and evaluation innovations—modular electrostatics-aware layers, consistent dielectric integration, and workflow-driven benchmarking—directly inform the design and evaluation of next-generation biomolecular NNPs, as well as ML/MM embedding approaches.

On the fundamental side, the empirical observation that physically constrained, multi-supervised NNP charges outperform reference DFT-derived charges as reactivity descriptors suggests the emergence of novel, possibly more chemically relevant representations within multitask learning. The scaling of transferability with chemically diverse training composition points to the validity of data-centric strategies for universal reactive NNPs for biomolecules, provided that proper workflow evaluation (e.g., NEB convergence) is incorporated.

Conclusion

Enerzyme sets a new standard for NNP-driven enzyme catalysis modeling, combining architectural advances with workflow automation and rigorous task-specific benchmarking (2607.01362). It achieves close to chemical accuracy in transition state energetics and mechanistic detail with minimal data and compute for individual systems, and demonstrates promising progress toward generalizable, physically consistent, and data-efficient simulation protocols for whole enzyme families. The established approaches, particularly regarding electrostatics treatment and physically motivated multitask supervision, will be instrumental as the field progresses toward large-scale, interpretable, and ML-empowered mechanistic studies in biocatalysis and beyond.