JAMUN

Molecular dynamics has to resolve bond vibrations, so its integration timestep sits at a couple of femtoseconds — while the changes that open a cryptic pocket or release a bound loop unfold over microseconds to milliseconds. That gap is an artifact of high-frequency motions that barely affect the ensemble, not something fundamental. JAMUN attacks the timestep rather than the force field: it adds a single fixed amount of Gaussian noise (σ = 0.4 Å) to every atomic coordinate, runs Langevin dynamics in that smoothed space where barriers are blurred and much larger steps are stable, then uses a learned denoiser to "jump" any noisy configuration back to a physically valid all-atom structure.

That noise–walk–jump construction is walk-jump sampling, the framework Prescient Design applied to antibody sequence generation in Walk-Jump Sampling. JAMUN carries it into continuous 3D space and models every atom explicitly, with the walk and jump steps decoupled so that one Langevin trajectory yields many independent structures. Training the denoiser on MD data at a single noise level, rather than over a schedule, keeps the physical priors of the underlying simulations — which is what lets one checkpoint sample peptides that were never in its training set, including peptides longer than any it saw.

JAMUN was developed by Prescient Design, the machine learning group inside Genentech, with a first author at MIT; it appeared on arXiv in October 2024 and was accepted at NeurIPS 2025. It sits alongside MDGen, BioEmu-1 and AlphaFlow in the move from single-structure prediction toward generated ensembles, and is scoped to peptides of two to six residues, where converged reference MD can be run for comparison.

#Key Features

• Sampling in a smoothed space: A one-time noising step flattens the landscape enough that Langevin dynamics decorrelates one to two orders of magnitude faster than MD.
• All-atom and SE(3)-equivariant: The denoiser is equivariant to rotations and translations but not reflections, so it distinguishes chirality natively and needs no post-hoc chirality fixer.
• Transfers to unseen and longer peptides: A model trained only on tetrapeptides recovers most metastable states and their relative populations for pentapeptides it was never shown.
• Handles macrocycles: Because it models atoms directly rather than backbone torsions, the same architecture applies to cyclic peptides, reproducing basins that MD samples on the CREMP macrocycle sets.
• A Boltzmann emulator, not a Boltzmann generator: Samples are not reweighted and carry no convergence guarantee; importance reweighting is left to future work.

#Technical Details

The denoiser is a geometric graph neural network in the NequIP family, built on e3nn, with roughly 10.5 million parameters. Edges connect atoms within 10 Å in the noisy coordinates; features combine bondedness, Bessel-embedded interatomic distances, atomic number, PDB atom name and residue code. The residue index is deliberately excluded, which the authors found necessary for generalizing to longer chains. Training ran for three days on two 40 GB NVIDIA A100 GPUs, with identical hyperparameters and the same σ = 0.4 Å throughout; separate checkpoints exist per dataset only because the force fields, solvent models and temperatures differ. Training sets span Timewarp 2AA and 4AA, MDGen 4AA-Explicit, the authors' own capped 2AA and uncapped 5AA OpenMM simulations, and CREMP macrocycles.

On a single A100, JAMUN draws a sample in 2 ms for dipeptides and 8 ms for pentapeptides — 100,000 conformations in 3 to 12.5 minutes against 40 minutes to 258 hours for the corresponding reference MD. Markov state model metastable-state probabilities correlate with reference MD at R² = 0.873 and 0.881 on the two Timewarp sets, and 94.7% of samples pass PoseBusters checks. On unseen pentapeptides the Jensen–Shannon distance over all torsions is 0.197, against 0.411 for Boltz-1 and 0.420 for BioEmu.

#Applications

The immediate use is cheap conformational ensembles for short peptides: screening for cryptic or transient pockets, characterizing a peptide binder's flexibility, and generating starting structures for downstream physics-based work. The macrocycle results extend this to a therapeutic modality where classical force fields are unreliable and quantum-mechanical conformer searches are expensive. The model takes an initial PDB structure and returns a trajectory, so it drops into existing MD analysis pipelines.

#Impact

JAMUN's contribution is less a leaderboard position than a demonstration that keeping the physics in the representation buys generalization. Approaches that tokenize backbone torsions or fix a pair representation inherit a fixed system size; modelling all atoms in a mildly noised Cartesian space keeps the sampler system-agnostic, and so extends to longer peptides and macrocycles without retraining. The limits are stated candidly: the demonstrations cover two- to six-residue peptides, the model does not transfer to chains as long as ten residues because its training peptides lack secondary structure, and it occasionally hallucinates a basin. Scaling to proteins would need far more MD data generation, the open problem the authors identify.