MoleculeMind pushes QuantaMind to 100,000-atom reaction simulations on a single GPU

MoleculeMind, a Chinese AI protein-design company, says its reactive machine-learning force field QuantaMind has been scaled to systems of 100,000 atoms while running on a single GPU. According to a report by QbitAI, the latest extended tests simulate hundreds of nanoseconds of reaction in such systems at close to quantum-chemistry accuracy, at roughly 0.25 seconds per step, with no need for an expensive supercomputing cluster.

The underlying work is published in Science Advances. In that paper, QuantaMind ran a continuous 6-nanosecond simulation of a complete PET hydrolase (PETase) system containing 17,792 atoms, including the enzyme, its substrate and nearly 5,000 water molecules, with no preset reaction path: proton transfer, bond breaking and bond formation, and the full enzymatic catalytic cycle all emerged step by step inside the simulation.

Molecular simulation has long been trapped in an impossible triangle of accuracy, time and scale. High-precision quantum-mechanical methods resolve every detail of a reaction but can only handle a few hundred to a few thousand atoms over extremely short moments, while classical simulations reach tens of thousands to millions of atoms but assume fixed bonds and are blind to processes such as enzyme catalysis and acid-base chemistry. Earlier AI force fields were mostly trained on molecules in their stable states, so they had never seen the instant a reaction happens.

QuantaMind's answer is to let the model watch those instants at scale. The paper trained it on 5,286 key configurations of chemical reactions, integrated computed data from different sources and levels of accuracy with quality-based weighting, and then looped between running simulations and performing fresh high-precision calculations on configurations the model was unsure about. The report says total energy drift stays minimal over long runs, allowing stable simulations lasting hundreds of nanoseconds.

Across three very different stress tests, proton hopping in water, spontaneous water autoionization and near-instant acid-base neutralization, results matched experiments closely even though the model was never specifically trained on them. A further test moved into a protein environment: the model placed a protein in a buffer close to real experimental conditions for 20 nanoseconds and let key sites accept or release protons on their own, estimating a critical point of 5.81 against a nuclear magnetic resonance experimental value of 5.5.

The team also randomly sampled configurations from each stage and re-checked them with quantum-mechanical calculations, finding agreement above 0.99, meaning the AI's judgments were nearly identical to quantum mechanics. The paper's corresponding author, Xu Jinbo, is widely described as a founder of AI protein folding: his 2016 RaptorX-Contact was the first effective AI protein structure prediction algorithm. The company he founded, MoleculeMind, was established in 2022, completed a Series A round in June 2026 for more than 100 million dollars in cumulative funding, and opened its AI-native bioeconomy operating system MoleculeOS in July.

Capabilities beyond the paper are already showing up in real projects. In a project to design a pH-sensitive antibody with an extended half-life, a candidate antibody designed with QuantaMind's help achieved a dissociation rate at pH 6.0 that was 62 times that at pH 7.4, a result that previously tended to require multiple rounds of wet-lab screening. In enzyme engineering, MoleculeMind has also used the model to clarify reaction mechanisms of complex catalytic enzymes and to pinpoint key mutation sites, narrowing the scope of wet-lab work.

For industry, the significance is that verification of chemical reactions can move from the laboratory into the computer: for the first time there is a sign that accuracy, time and scale can be pushed outward together, which could lower the trial-and-error cost of long-cycle work such as enzyme engineering and antibody design. MoleculeMind says QuantaMind is expanding all three boundaries at once, from tens of thousands to hundreds of thousands of atoms, from tens of nanoseconds to hundreds of nanoseconds of simulation, and with computation speed doubling again. Whether those capabilities can be reproduced stably in more real systems is the key thing to watch.