niklastr/vacuum_peptides

Vacuum Peptide Ensembles

Molecular-dynamics trajectories derived from ManyPeptidesMD. This repository provides vacuum ensembles for a small subset of peptides, initialized from ManyPeptidesMD conformers.

Overview

Path	Model	Use
trajectories/v1/	Amber14, peptide only	Gas-phase sampling
webdatasets/v1/300K/	Streamable samples from the 300 K state	Model training
webdatasets/v1/300-500K/	Streamable samples from all REMD states	Temperature-conditioned training

Files follow:

trajectories/v1/<split>/<length>/<sequence>/state_YY_<temperature>K.npz

Each NPZ contains positions, smiles, id, temperature, potential energy, and metadata. positions has shape (frames, atoms, 4) with rows [atomic_number, x_nm, y_nm, z_nm].

Files aggregate independent REMD bursts initialized from up to 100 source conformers. Saved frames within each burst are 10 ps apart. Please note that the data were simulated using REMD, which breaks temporal ordering.

Simulation

Starting structures came from the corresponding ManyPeptidesMD trajectories. Up to 100 conformers per sequence were selected by uniform stride from the end of each source trajectory. Peptide sequences, atom order, and train/validation/test assignments follow ManyPeptidesMD.

Replica Exchange

Parameter	Value
Replicas	8, geometrically spaced from 300 to 500 K
Temperatures	300.00, 322.71, 347.14, 373.42, 401.69, 432.10, 464.81, 500.00 K
Integrator	Langevin, 2 fs, 0.3 ps^-1 collision rate
Warm-up / production	0.5 ns / 100 ns per replica
Save interval	5,000 steps = 10 ps
Exchange interval	5,000 steps = 10 ps

Outputs were demultiplexed by thermodynamic state. Each state file therefore represents a fixed temperature, not one replica's path through the ladder. Select state_00_300p00K.npz for 300 K data.

Energy Function

• Peptide: Amber14.
• Vacuum: NoCutoff; H-bond constraints.

Citation

Please cite this dataset and ManyPeptidesMD when using its peptide sequences or starting conformers:

@misc{rindtorff_2026_vacuum_peptide_trajectories,
  author       = {Rindtorff, Niklas},
  title        = {Vacuum Peptide Ensembles},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/niklastr/vacuum_peptides}},
  note         = {Dataset in progress; accessed 2026-07-23}
}

Source dataset: ManyPeptidesMD.

License

This derived dataset is released under CC BY 4.0. ManyPeptidesMD is distributed under the terms stated in its dataset repository. See the source repository for authoritative terms and provenance caveats.