An AI trained on breaking bonds simulated a 17,000-atom plastic-eating enzyme's full reaction

The hard part of simulating chemistry was never the compute. It was teaching a model what a breaking bond actually looks like. A Science Advances paper this week from Molecule Mind (分子之心) reports a reactive machine-learning force field, a model that learns the forces between atoms well enough to drive a molecular-dynamics simulation, trained on transition-state-centered reaction data and conditioned on the underlying quantum-mechanical method. Used on a 17,792-atom system containing the PETase enzyme, its substrate, and roughly 5,000 waters, the model drove a 6-nanosecond reactive simulation through the full catalytic cycle and matched quantum-mechanical reference calculations at better than 0.99 agreement.

PETase is the enzyme some bacteria use to break down PET plastic. Simulating its catalytic cycle means watching proton transfers, bond breaking, and bond formation propagate over time, the kind of dynamics that classical force fields handle poorly and that earlier machine-learning potentials either avoided or could not stabilize at enzyme scale.

What makes the new model different is the training set. Most machine-learning force fields learn from snapshots of stable molecules, then get asked to predict what happens when those molecules react. Molecule Mind's team, led by founder Xu Jinbo, a RaptorX-Contact pioneer from the AlphaFold era, built their training corpus around non-equilibrium conformations and transition states: 5,286 reactive events centered on the moment a bond breaks or forms, per the company's announcement on PRNewswire. The model was also conditioned on the specific density-functional method and basis set used to generate the reference data, so it learns the local chemistry rather than memorizing one quantum-chemistry code's output.

That second ingredient matters because the bottleneck in reactive molecular dynamics has not been raw GPU throughput. It has been the model's stability over long simulations: a force field that gets the equilibrium geometry right can still blow up the moment a proton transfers, because the chemistry of the transition state is a different problem from the chemistry of the resting state. By training on transition states directly and conditioning on the underlying quantum method, the lab argues it has broken what it calls the "impossible triangle" of accuracy, scale, and simulation length that has kept reactive dynamics out of the everyday drug-discovery workflow.

The paper-validated result is the 17,792-atom PETase simulation. The 100,000-atom, single-GPU, roughly 0.25-second-per-step claim is a separate company disclosure, reported in the same PRNewswire release and discussed in earlier bioRxiv preprints of the methodology. The distinction matters: a peer-reviewed paper supports the enzyme-scale result; the larger scale and the single-GPU accessibility story are company benchmarks until independent groups reproduce them.

The arc is the point. Xu's career began with RaptorX-Contact, the 2016 contact-prediction work that helped make AlphaFold possible, and Molecule Mind was founded in 2022 to push AI from static structure prediction into dynamic chemistry. The company reported a cumulative Series A above USD 100 million in June 2026, per the qbitai write-up, and launched MoleculeOS, a platform layer for AI molecular R&D, in July 2026. Each step widens the kind of chemistry a machine-learning model can credibly simulate, first fold a protein, then design a binder, now watch it catalyze a reaction. The trajectory points at drug discovery, enzyme engineering, and eventually the in-silico reaction screening that has historically required a supercomputer.

What to watch next: independent benchmarks of the reactive force field against established reactive MLFFs such as MACE, Allegro, and NewtonNet, and whether the 100,000-atom, single-GPU throughput holds up outside the company's own runs. The paper is the more durable claim. The accessibility story is the more interesting one if it survives contact with a third-party lab.