transferable-samplers/many-peptides-md

1---2license: mit3 4---5 6> [!IMPORTANT]7> **Critical Update**8> 9> The original 8AA TICA models within `subsampled_trajectories/*/8AA/*.npz` employed a CA-only atom selection. **These models are not valid for comparison to results in our paper.**10> 11> **Updated files (uploaded 15/12/2025)** now contain corrected models. If you previously downloaded this dataset, please re-download to ensure accurate results.12>13> Note: Codebase references to `tica_features_ca` must now be replaced with `tica_features`. **This was resolved in our codebase by [PR #26](https://github.com/transferable-samplers/transferable-samplers/pull/26).**14>15> Note: Unguarded `snapshot_download` calls will automatically redownload the relevant files when it detects a change in the repo.16> 17> We sincerely apologize for any inconvenience this may have caused.18 19# ManyPeptidesMD Dataset20 21Welcome to the **ManyPeptidesMD** dataset!22 23This dataset was generated as part of the work **Amortized Sampling with Transferable Normalizing Flows** [*NeurIPS 2025*](https://neurips.cc/virtual/2025/loc/san-diego/poster/118702).24 25**ManyPeptidesMD** is a collection of molecular dynamics (MD) trajectories for 21,700 randomly sampled peptide sequences, generated to support research in molecular simulation, machine learning for molecular dynamics, and unnormalized density sampling. 26 27The length distribution of the dataset is as follows:28| Sequence length | 2    | 3     | 4     | 5     | 6     | 7     | 8      |29|-----------------|------|-------|-------|-------|-------|-------|--------|30| Training        | 200  | 1,000 | 1,500 | 2,000 | 3,000 | 4,000 | 10,000 |31| Validation      | 10   | ---   | 10    | ---   | ---   | ---   | 10     |32| Testing      | 30   | ---   | 30    | ---   | ---   | ---   | 30     |33 34---35 36## Usage37 38The easiest way to use this dataset is to work from the accompanying [codebase](https://github.com/transferable-samplers/transferable-samplers).39 40In this codebase the training webdataset is streamed and cached, and the evaluation data is downloaded automatically.41 42## Data Organization43 44### Full Trajectories45 46* **Location:** `trajectories/`47* **Sampling rate:**48 49  * Training: Positions and velocities saved every **1 ps**50  * Validation & Testing: Positions and velocities saved every **10 ps**51* **Trajectory length:**52 53  * Training: **200 ns**54  * Validation & Testing: **5 μs**55 56### PDB Files57 58* **Location:** `pdb_tarfiles/`59  * **Note:** Due to Hugging Face repo limits these are provided as `.tar` files for each subset.60 61### Additional Formats62 63For ease-of-use we additionally provide:64 651. **Webdataset**66 67   * **Path:** `webdatasets/single_frames/`68   * **Format:** Each `.tar` contains 4 randomly selected position frames per sequence, from a 10ps/frame subsample of the full training trajectories. The samples are preshuffled within each `.tar`. The sample filenames are formatted as `{SEQUENCE}_{TIME}.bin` where `{TIME}` is the time **in picoseconds** from the original trajectory.69   * **Coming soon:** Webdataset for pairs of samples (e.g a 10ps interval) and chunks of trajectory sample.70 712. **Subsampled Validation & Test Sets**72 73   * **Path:** `trajectories_subsampled/`74   * **Description:** 500 ps downsample of validation/testing trajectories, giving sets of size 10,000 (as used in the paper’s evaluation).75   * **Additional info:** Includes **TICA projection data** computed from full 10 ps interval trajectories for metric calculation.76 77---78 79## Simulation Details80 81All simulations were performed using **OpenMM** with the following configuration:82 83```python84forcefield = openmm.app.ForceField("amber14-all.xml", "implicit/obc1.xml")85nonbondedMethod = openmm.app.CutoffNonPeriodic86nonbondedCutoff = 2.0 * openmm.unit.nanometer87temperature = 310  # Kelvin88 89# Initialize forcefield system90system = forcefield.createSystem(91    self.pdb_dict[sequence].topology,92    nonbondedMethod=nonbondedMethod,93    nonbondedCutoff=nonbondedCutoff,94    constraints=None,95)96 97# Initialize integrator98integrator = openmm.LangevinMiddleIntegrator(99    temperature * openmm.unit.kelvin,100    0.3 / openmm.unit.picosecond,101    1.0 * openmm.unit.femtosecond,102)103```104 105---106 107## Citation108 109If you use this dataset please cite our work [*Amortized Sampling with Transferable Normalizing Flows*](https://arxiv.org/abs/2508.18175).110 111 112## Acknowledgement113 114We greatly thank Hugging Face for hosting this large dataset!115