Reactive machine-learning force field running DFT-accurate molecular dynamics in which bonds break and form, from proton hopping to enzyme catalysis.
No providers recorded yet. Browse all providers
A classical force field treats a covalent bond as a spring that cannot snap. That is fine for folding and binding, but it rules out the events that define enzyme chemistry: a serine attacking an ester, a proton hopping along a chain of waters, a histidine picking up and dropping a hydrogen as the local pH shifts. Quantum mechanics captures those events directly, at a cost that confines it to a few hundred atoms and picoseconds. QuantaMind, from MoleculeMind, is a machine-learning force field (MLFF) built to close that gap: it predicts density functional theory (DFT)-level energies, forces and dipoles fast enough to run nanosecond molecular dynamics on solvated proteins, with bond breaking and formation emerging from the dynamics rather than being scripted.
The key design choice is the training data. Most biomolecular MLFFs learn from equilibrium or near-equilibrium conformations sampled by classical MD, so they have never seen a stretched bond or a transition state and fail once a reaction starts. QuantaMind is trained on transition-state-centered configurations drawn from quantum chemical studies of three enzyme reactions, alongside protein-fragment data. Because those sources were computed with different DFT functionals and basis sets, the model takes the level of theory as an explicit input, which lets it pool heterogeneous QM data without the labels contradicting one another.
The work was first posted as a bioRxiv preprint in September 2025 and published in Science Advances in September 2026 by Song Xia, Deqiang Zhang, Xu Shang and Jinbo Xu. A companion preprint applies the same model to the full PETase catalytic cycle.
QuantaMind builds on the SpookyNet architecture. An embedding module encodes atomic numbers, total charge, spin multiplicity and the DFT method; three interaction layers each pair a message-passing local sublayer, using Bernstein-polynomial radial bases and spherical harmonics, with an attention-based non-local sublayer for long-range correlation. The readout predicts per-atom energies and partial charges; forces are the gradient of total energy, and dipoles follow from the charges. Training minimizes RMSE on energy, force and dipole with AMSGrad and stochastic weight averaging, oversampling the small datasets so they are not swamped by the large one.
The base model is trained on the GEMS bottom-up and crambin top-down protein-fragment datasets plus three transition-state sets (5,286 configurations) from phenylalanine ammonia-lyase, D-psicose 3-epimerase and epoxide hydrolase reactions. The preprint additionally describes a 2,476-structure phosphorus dataset. For the constant-pH work, a tuned variant adds 1,780 histidine, 3,698 sodium-ion and 1,833 phosphate-buffer fragments. Retrained on the SO3LR data alone, QuantaMind performs comparably to SO3LR. Inference costs 4.2 × 10⁻⁶ s per atom per step on an A100 80GB GPU, and 1-ns NVE runs on water boxes of 648 and 9,999 atoms show total-energy drift below 10⁻⁴ kcal/mol per atom per picosecond.
The model is aimed at reaction-level questions in biomolecular simulation. In water boxes of up to 24,001 atoms it gives an extrapolated proton diffusion coefficient of 0.87 ± 0.10 Ų/ps against experimental values of 0.94–0.96, and an estimated pH of 6.34 for pure water at 373 K against 6.13 measured. A 6-ns simulation of an ~18,000-atom system, steered by bias potentials on the reacting atoms, recapitulates the PETase cycle for MHET hydrolysis, and a 20-ns run of hen egg-white lysozyme in phosphate buffer yields a His15 pKa of about 5.81, close to the NMR value of 5.5. The preprint also uses it to refine predicted protein structures and protein-ligand pockets.
QuantaMind sits among general-purpose biomolecular potentials such as AQuaRef and MACE-POLAR-1, with its emphasis on reactivity rather than equilibrium structure. The authors note its training domain is still largely protein-derived, so small molecules with unfamiliar chemistry need the iterative refinement loop, and the histidine titration relied on a tuned variant. The model and MD code are released under Apache 2.0, without training scripts or data; the trained weights are distributed only on request to academic researchers under non-commercial terms, and commercial use requires a separate license from MoleculeMind.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.