Filming Chemical Reactions: An Enzyme Catalytic Cycle on a Single NVIDIA A100

Forum topic · QianXun · 2026-09-30

Summary

QuantaMind MD, a machine-learning force field from Molecule Design (Shanghai), simulates reactive molecular dynamics at near quantum-mechanical accuracy. Published in Science Advances (DOI: 10.1126/sciadv.aeg3595), the paper reports that a single simulation step for a 100,000-atom system—which would theoretically take DFT methods roughly 13 million days—runs in 0.25 seconds on one NVIDIA A100 80GB GPU. Key designs include transition-state-centered training data (5,286 enzyme catalytic transition-state configurations), DFT-method-class embedding, and active learning loops. Validation covers proton diffusion in water, spontaneous water autoionization pH, lysozyme His15 pKa, and a full PETase catalytic cycle (14.5±0.1 kcal/mol barrier in a 17,792-atom system over 6 ns), with 342 randomly sampled configurations independently verified against DFT forces at Pearson correlation above 0.99. Company-reported post-review results include pH-sensitive antibodies with 62-fold faster dissociation at pH 6.0, though these are not peer-reviewed.

From 35,000 Years to a Quarter Second

To build intuition: traditional quantum chemistry (DFT) would theoretically need about 13 million days to simulate a single step of a 100,000-atom system—over 35,000 years, longer than the entire history of human civilization since wheat domestication. A paper published September 11 in Science Advances reports an alternative figure: 0.25 seconds for the same system and step, on a single NVIDIA A100 80GB.

The paper, *Bridging the accuracy-speed divide in reactive molecular dynamics with QuantaMind MD* (DOI: 10.1126/sciadv.aeg3595), is by Deqian Zhang, Song Xia, Xu Shang, and Jinbo Xu from Molecule Design in Shanghai. Dividing the two numbers yields roughly a 4.5-trillion-fold speedup—the figure cited in media coverage, and the arithmetic checks out.

> Background: QuantaMind is a machine-learning force field (MLFF): a neural network that learns interactions between atoms. Molecular dynamics simulation is conceptually simple—at each step, compute the force on every atom and advance positions via Newton's equations, repeating millions of times to produce a trajectory. The difficulty is entirely in "where the forces come from": quantum-mechanical forces are accurate but require an expensive electronic-structure calculation every step; classical force fields are cheap but hard-code bond connectivity, so they fail when bonds break.

The Impossible Triangle and Three Designs

Molecular simulation has long faced a trade-off triangle: accuracy (quantum precision), speed (engineering usability), and scale (biological systems). Existing approaches pick two:

DFT and quantum chemistry: accuracy ceiling, but cost explodes with system size.
Classical force fields: cheap per step, can run tens of thousands of atoms for microseconds, but preset bond connectivity fails for chemical reactions.
QM/MM partitioning: quantum mechanics for the reaction center, classical for the environment—workable but boundary treatment is error-prone.
Existing MLFFs: unstable over long simulations and undertrained on reactive regions.

QuantaMind's three key designs target these weaknesses:

Reactivity-centered data: instead of learning only from equilibrium conformations, the model incorporates non-equilibrium configurations and transition-state-centered reaction data (5,286 enzyme catalytic transition states), covering states before, during, and after the reaction.
DFT method-class embedding: different reaction types use different-accuracy DFT teachers for energies and forces.
Stability benchmark: in 1-nanosecond NVE simulations of pure water, energy drift stays at the 10⁻⁵ kcal/mol/atom/ps level.

Four Validation Cases, from Water to Enzymes

| Target | System & Duration | QuantaMind Result | Experimental Reference | |---|---|---|---| | Proton diffusion coefficient | 24,001-atom water | 0.87±0.10 Ų/ps | 0.94 / 0.96 | | Water autoionization pH (373K) | 9,999-atom water, 3 ns | 6.34±0.06 | 6.13±0.03 | | Lysozyme His15 pKa | 20 ns simulation | ~5.81 | 5.5 (NMR) | | PETase full catalytic cycle | 17,792 atoms, 6 ns | Barrier 14.5±0.1 kcal/mol | Comparable computational reference |

Key points:

Proton diffusion: the paper claims ~87% lower deviation versus comparable studies.
Water autoionization is a rare-event observation: in a 3 ns trajectory the model sampled spontaneous water splitting and derived pH 6.34 versus the experimental 6.13.
PETase is the centerpiece: this enzyme degrades PET plastic, and the simulation completed the full acylation, deacylation, and catalytic-site recovery cycle in a 17,792-atom system, with a rate-limiting barrier of 14.5±0.1 kcal/mol.
Independent verification: 342 configurations sampled from the trajectory were re-checked with standalone DFT single-point calculations; Pearson correlation coefficients for atomic force components all exceeded 0.99.

Paper Numbers vs. Company Numbers

These two sets of figures should be kept separate:

| Item | In paper (peer-reviewed) | Post-submission extension (company) | |---|---|---| | Largest system | 24,001 atoms (water) | ~100,000 atoms | | Longest stable simulation | 20 ns (lysozyme) | Hundreds of nanoseconds | | Inference speed | 4.2×10⁻⁶ s/atom/step | 2.1×10⁻⁶ s/atom/step | | 100k-atom single step | — | 0.25 s |

The ~18,000-atom PETase system needs about 20 hours per nanosecond of simulation (paper data); the 6 ns cycle corresponds to roughly 5 days on a single card—feasible for a lab.
The "100,000 atoms, hundreds of ns" and "0.25 s per step" figures come from post-submission tests and are not peer-reviewed.
The 4.5-trillion-fold speedup arithmetic is valid (13 million days ≈ 1.12×10¹² seconds, divided by 0.25 s), but its premise—that DFT would need 13 million days for one step—is a theoretical estimate that varies with hardware and implementation.

From Paper to Antibodies

On the industrial side, Molecule Design reports two application tracks:

Enzyme engineering and biomanufacturing: using mechanistic energy barriers to screen mutants computationally.
Drug and biomolecule R&D: in a pH-sensitive antibody design project, QuantaMind-assisted candidates showed a dissociation rate at pH 6.0 that was 62× the rate at pH 7.4 (company-reported, not in the paper).

The broader product matrix includes the protein generation model NewOrigin, structure prediction model MMfold, de novo design platform MMdesign, and QuantaMind, planned for integration into the AI-native bio-economy operating system MoleculeOS—closing the loop of generation, design, prediction, and mechanistic understanding. Founder Jinbo Xu previously built RaptorX-Contact for protein structure prediction.

Why This Team, and What Remains

AI for Science's past three years have focused on predicting outcomes: AlphaFold predicts structures, diffusion models generate molecules. Their shared weakness is the inability to explain "why"—exactly what drug review demands in the form of mechanistic evidence. QuantaMind pushes simulation toward "seeing the process," adding a layer of verifiable mechanism between prediction and experiment.

The method's moat lies in data organization rather than network architecture: transition-state-centered datasets, DFT class embedding, and active learning loops—all demanding deep chemical understanding rather than parameter scaling.

Open questions: PETase is a single-enzyme, single-substrate validation; generalization to other enzymes awaits follow-up work; hundred-nanosecond timescales remain insufficient for slow processes like some allosteric events; and the 62-fold antibody claim needs wet-lab validation.

One accelerator card under $30,000, 0.25 seconds for a step that once took 35,000 years. The economics of molecular simulation changed its denominator that day.

---

Sources and caveats

Paper: *Bridging the accuracy-speed divide in reactive molecular dynamics with QuantaMind MD*, Science Advances, DOI 10.1126/sciadv.aeg3595, published 2026-09-11 (PubMed 42726868)
Company disclosures and Jinbo Xu interview coverage cross-checked via NetEase Tech, Guangming Net, China.com.cn, and AI123 (2026-09-14 to 09-30)
The four validation figures come from the paper; 100k-atom/hundreds-of-ns results and the 62-fold antibody figure are company extension tests, not peer-reviewed
"4.5 trillion-fold" is a media figure; its arithmetic has been recomputed and its premise noted; the ~20 hours/ns PETase cost is paper data for an A100 80GB