Enerzyme: The AI That Can Watch Enzymes Work

Inside every living cell, enzymes are running the chemistry of life, snipping, splicing, and assembling molecules with a precision that industrial chemists can only dream of. A single enzyme can catalyze a million reactions per second. The problem? Understanding how they do it has been painfully slow. Even with GPU-accelerated quantum chemistry, simulating a single step in an enzymatic reaction on a cluster of 300 atoms can take hours. Mapping a full reaction pathway takes days.

Now researchers at MIT have built something that changes the math entirely.

Heather Kulik’s group released Enerzyme, an open-source framework that trains neural networks to simulate enzyme catalysis with near-quantum-chemical accuracy, but at a fraction of the cost. The key breakthrough isn’t a better algorithm. It’s that Enerzyme handles everything that makes enzymes uniquely hard to model: their size, their solvent environment, and the complex charge transfers that drive their chemistry.

Background & Context

Quantum mechanical (QM) cluster models have been the gold standard for studying enzyme mechanisms. Researchers carve out the active site of an enzyme (typically 200 to 600 atoms) and run density functional theory (DFT) calculations to map out reaction pathways. These calculations are accurate but expensive. A single geometry optimization on a 300-atom cluster can take hours on a GPU cluster. A full reaction pathway exploration, with flexible scans and nudged elastic band calculations, can take days.

Neural network potentials (NNPs) emerged a decade ago as a faster alternative. Train a neural network on DFT data, and it learns to predict energies and forces without re-running the quantum chemistry. The catch: NNPs worked well on small molecules in a vacuum, but enzymes are a different beast. They’re large, surrounded by implicit solvent that mimics the protein environment, and their reactions involve dramatic charge redistribution that simple NNPs can’t capture.

“Enzymes introduce new challenges not faced by NNPs for small molecules, including large system sizes, implicit-solvent environments, substantial polarization, and charge transfer,” the authors write. Existing NNP foundation models trained on millions of small-molecule configurations fail when dropped into an enzyme active site because the physics is fundamentally different.

What the Researchers Did

Kulik’s team built Enerzyme as a modular framework with two main components. The first, simply called Enerzyme, handles NNP architecture, training, and inference. It wraps three popular graph neural network backbones (PhysNet, SpookyNet, and MACE) with customized layers that add electrostatic awareness. The network is trained to predict not just energies and forces, but also atomic partial charges. This multitask learning forces the network to understand charge transfer and polarization, which are central to how enzymes catalyze reactions.

The second component, Enerzymette, automates the tedious workflow of reaction pathway exploration. It builds QM clusters directly from Protein Data Bank (PDB) structures using a companion package called QuantumPDB, generates reactive training data through high-temperature steered molecular dynamics, and then runs flexible scans and nudged elastic band (NEB) calculations to map out reaction pathways, all without human intervention.

They tested the framework on SAM-dependent methyltransferases (MTases), a family of enzymes that transfer methyl groups and play roles in gene regulation, drug metabolism, and natural product biosynthesis. The MTase family is a particularly demanding test case because electrostatic interactions and charge transfer dominate their catalytic mechanism.

Enzyme	System Size	Training Data	Energy MAE	Force MAE
PfPMT (malaria)	284 atoms	956 configs	0.48 kcal/mol	0.94 kcal/mol/\u00c5
HcgC (hydrogenase)	307 atoms	988 configs	0.42 kcal/mol	0.89 kcal/mol/\u00c5
COMT (mammalian)	~400 atoms	920 configs	0.51 kcal/mol	0.97 kcal/mol/\u00c5

What They Found

The results surprised even the researchers. With fewer than 1,000 training configurations per system (a tiny dataset by NNP standards), the trained models reproduced DFT-level reaction energetics to within 0.5 kcal/mol. That’s approaching chemical accuracy, the oft-cited 1 kcal/mol threshold for meaningful predictions. On the largest test system, a COMT cluster with 545 atoms, the NNP still held within 1 kcal/mol.

But accuracy on standard metrics wasn’t the real test. The researchers subjected their models to the same workflows that DFT calculations run: flexible scans along the reaction coordinate and climbing-image NEB calculations to find transition states. These are more demanding than simple energy prediction. An NNP that scores well on random test configurations can fail catastrophically when asked to navigate a reaction path.

Here’s where the electrostatic supervision paid off. Models trained without explicit charge prediction produced erratic potential energy surfaces during scans. Models with charge supervision (where the network learned to output atomic charges alongside energies) produced smooth, physically meaningful reaction profiles that matched DFT within 0.7 kcal/mol across the entire path.

The charge predictions themselves proved chemically interpretable. The multitask-learned charges correctly identified which atoms donate and accept charge during the reaction, matching the charge-transfer patterns the group had previously mapped out in expensive QM/MM studies. In essence, the neural network learned not just the energetics, but the underlying physics of how electrons move during catalysis.

Why It Matters

This is a genuine shift in how enzyme mechanics can be studied. Right now, a thorough computational study of a single enzyme’s mechanism is a weeks-long project requiring significant computational resources. With Enerzyme, a researcher could train an accurate surrogate model in a day (one afternoon of calculation on a single GPU) and then run reaction pathway explorations in minutes.

The implications for drug discovery are direct and large. Many drugs work by inhibiting enzymes. COMT inhibitors, for instance, are used to treat Parkinson’s disease. Understanding exactly how an inhibitor binds and how it might be improved is currently a slow, iterative process of simulation and crystal structure analysis. A fast, accurate surrogate model could let medicinal chemists screen modifications in hours instead of weeks.

How It Could Change Our Lives

The most immediate impact will likely be in understanding drug metabolism. Methyltransferases process about 10% of all clinically used drugs. When your body metabolizes a medication, these enzymes are often the ones doing the work. A computational model that accurately predicts how different enzymes handle different drug molecules could help predict drug interactions, side effects, and individual variability in drug response, before a single clinical trial begins.

Beyond pharmacology, the same framework could accelerate enzyme engineering for industrial chemistry. Companies are already using engineered enzymes to produce everything from biofuels to biodegradable plastics. Each new enzyme design currently requires extensive experimental characterization. A reliable computational model could cut that cycle from months to days.

The Bigger Picture

Enerzyme is part of a broader movement in computational chemistry. Universal NNP foundation models trained on millions of configurations across thousands of molecules have become remarkably good at general chemistry. But they hit a wall on large, charged, solvated systems like enzyme active sites. The MIT team’s approach of system-specific fine-tuning with physics-aware training objectives suggests a practical path forward: start from a foundation model, then specialize with targeted data and electrostatic supervision.

Limitations & What’s Next

The current work focused on a single enzyme family (methyltransferases). The framework is designed to be general, but it hasn’t been tested on other enzyme classes with different chemistries (proteases, kinases, or oxidoreductases). The training data generation still requires DFT calculations, which means it’s not free; the advantage is in amortizing that cost across many subsequent calculations. The authors also note that some NNP architectures (particularly PhysNet) showed failure modes in certain systems, suggesting that architecture choice matters and the best setup may be system-dependent.

The Enerzyme code is open-source, available on GitHub. If it generalizes to other enzyme families (and there’s reason to think it will), it could become a standard tool in computational enzymology, turning what was a weeks-long computation into an afternoon’s work.

---

\ud83d\udcc4 Source: Luo, W. & Kulik, H. J. “Enerzyme: A Framework for Efficient Training of Reactive Neural Network Potentials for Enzyme Catalysis with Application to Methyltransferases.” arXiv:2607.01362 (2026). https://arxiv.org/abs/2607.01362