2026-09-15
MoleculeMind’s QuantaMind ML force field runs reactive MD at DFT accuracy: a 17,792-atom PETase cycle in ~20 h/ns on one A100, proton D=0.87 Ų/ps vs ~0.95 experiment.
Reactive molecular dynamics sits between two bad options. Classical force fields scale, but bonds are hardcoded, so proton hops and bond making/breaking are off the table. Quantum forces are accurate and unaffordable beyond tiny regions and short times. Machine learning force fields (MLFFs) fit a potential energy surface and try to keep both. Many are trained near equilibrium, or on small-molecule reactions with a handful of heavy atoms. Accuracy falls off in enzyme pockets, solvent, and transition-state geometries.
MoleculeMind’s Science Advances paper introduces QuantaMind, a SpookyNet-style MLFF with a categorical embedding for the DFT recipe (functional plus basis). It is trained on transition-state-centered, nonequilibrium structures so bond events can appear during MD, without carving out a QM region for QM/MM.
Inputs are atomic numbers, coordinates, total charge, spin multiplicity, and a discrete DFT-method index. Atom embeddings pass through three local/nonlocal interaction layers. Per-atom energies are read out and summed with nuclear repulsion, long-range electrostatics, and D4 dispersion. Forces are analytic derivatives. Method embeddings exist because TS labels are mostly B3LYP/def2-SVP while the main corpus is PBE0/def2-TZVPP+MBD; the network is told which surface a point came from.
Training mixes GEMS protein fragments, Crambin, and three enzyme TS sets (TS1/TS2/TS3, 5,286 frames). Extra fragments cover histidine, Na+, phosphate buffer, and PETase. Full boxes are too large for DFT, so a 5 Å sphere around a point of interest is cut, hydrogens are added, and fragments stay under 300 atoms. Trajectories can be resampled, labeled, and used to fine-tune, an active-learning loop. The loss is RMSE on energy, force, and dipole. Training uses PyTorch Lightning; MD runs in SchNetPack with the whole integrator on GPU.
On one NVIDIA A100 80GB, inference is 4.2×10⁻⁶ s/atom/step. For the 18,000-atom PETase box that is about 20 hours per nanosecond. The largest water box they report is 24,001 atoms.
Proton diffusion: six boxes from 649 to 24,001 atoms, three 1 ns runs each, extrapolated to infinite size at 0.87±0.10 Ų/ps against experimental 0.94 and 0.96. Huo et al.’s earlier MLFF reported 0.375±0.242. Water autoionization at 373 K in a 9,999-atom box over 3 ns gives pH 6.34±0.06 (experiment 6.13±0.03). Acid-base neutralization over 100×50 ps runs yields 4.9×10¹¹ to 7.1×10¹¹ M⁻¹ s⁻¹ (experiment 1.4×10¹¹; same order, high).
The enzyme demo is a full catalytic cycle: an LCC PETase variant (257 residues), one MHET, 4,898 waters, 17,792 atoms. Every atom, solvent included, is on the same MLFF. A 6 ns trajectory covers acylation, deacylation, and triad restoration. Umbrella sampling along chosen atom pairs, then WHAM: acylation 11.6±0.1 kcal/mol, deacylation nucleophilic attack 10.5, ester cleavage 5.0. The text marks acylation nucleophilic attack as rate-limiting at 14.5±0.1 kcal/mol. Experimental LCC barriers are 14.5 to 16.5 kcal/mol on pNP and 18.7 on BHET; the simulated substrate is MHET, so the numbers are not the same assay. 342 frames drawn from the trajectory match independent PBE0 single points with Pearson r > 0.99 on force components.
Hen egg-white lysozyme in phosphate buffer for 20 ns gives His15 pKa 5.81 versus NMR 5.5. NVE energy drift over 1 ns is about 7×10⁻⁵ kcal/mol/atom/ps.
For enzyme mechanisms, constant-pH work, and reaction paths, this replaces a QM/MM patchwork with one reactive MLFF at tens of thousands of atoms and tens of nanoseconds. There is no QM region to choose and no need to freeze protonation states. At 20 hours per nanosecond on one A100, denser umbrella windows become realistic, sampling that DFT-in-the-loop cannot buy. Matched against SO3LR on the same data, accuracy is similar; the claim is stable reactive MD, not a leaderboard win. Code and trajectories are on GitHub and Zenodo.
This is an incremental, usable step: the field can now run reactive, solvated, nanosecond enzyme cycles on a single GPU. It is not yet a general simulator for arbitrary chemistry.
The authors are blunt. Chemical coverage is protein-heavy; unseen chemistries degrade and need per-system iteration. The TS sets were not recomputed at PBE0; method embeddings only paper over that. Bootstrap error bars on free energies and diffusion are likely too tight. His15 protonation flips cluster between 5 and 10 ns, so the pKa is unevenly sampled. The PETase cycle is not unbiased dynamics: each step uses a stiff bias (k = 10.0 eV/Ų) on a one-dimensional coordinate, and the authors note that acylation and deacylation may be concerted, which wants a 2D surface. The neutralization rate is biased by sharing one initial geometry across 100 runs. The largest system in the paper is 24,001 atoms; 100,000-atom figures circulating outside the paper are not in this text. Three related patents are pending from a company lab.