niklastr/biased_peptides

Biased Peptides

Prefix-sampled training WebDatasets derived from ManyPeptidesMD. Each configuration keeps the ManyPeptidesMD train split and the single-frame WebDataset layout, but draws unique frames only from the first 1/d of every 200,000-frame trajectory. Use these sets to measure how incomplete equilibration changes model training.

Configurations

Path	Prefix	Unique frames / molecule	Shards	Members / shard
prefix_1_0/	1	20,000	5,000	86,800
prefix_0_5/	1/2	20,000	5,000	86,800
prefix_0_25/	1/4	20,000	5,000	86,800
prefix_0_125/	1/8	20,000	5,000	86,800
prefix_0_0625/	1/16	10,000	5,000	43,400

prefix_0_0625 cannot fill 20,000 unique frames. A 1/16 prefix of 200,000 frames has 12,500 frames, and 10,000 is the largest count that divides 5,000 shards.

Layout

Each configuration contains 0000.tar through 4999.tar. Member names are {sequence}_{frame_id:06d}.bin. The .bin payload is little-endian float32 coordinates with shape (N, 3) in nanometres. This matches ManyPeptidesMD single_frames.

The train split has 21,700 molecules. Frames are globally shuffled inside each shard. Sequence identity, atom order, and the Amber-ff14SB / OBC1 implicit-water setting follow ManyPeptidesMD.

Use the many_peptides adapter with source_format: webdataset_tar.

Sampling

For prefix 1/d, eligible frames are indices 0 .. floor(200000/d)-1. The exporter draws unique indices from that prefix with seed 0, packs them evenly across 5,000 shards, and shuffles each shard. This is a time prefix, not a random draw from the full trajectory.

Source

• Trajectories: ManyPeptidesMD train NPZ files, 200,000 frames per molecule.
• Reference WebDataset: ManyPeptidesMD single_frames (5,000 shards, 4 frames per molecule per shard).

Please cite ManyPeptidesMD when you use these sequences or coordinates:

@misc{tan2025amortized,
      title={Amortized Sampling with Transferable Normalizing Flows},
      author={Charlie B. Tan and Majdi Hassan and Leon Klein and Saifuddin Syed and Dominique Beaini and Michael M. Bronstein and Alexander Tong and Kirill Neklyudov},
      year={2025},
      eprint={2508.18175},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2508.18175},
}

@misc{rindtorff_2026_biased_peptides,
      title={Biased Peptides},
      author={Rindtorff, Niklas},
      year={2026},
      publisher={Hugging Face},
      howpublished={\url{https://huggingface.co/datasets/niklastr/biased_peptides}},
}

This derived dataset is released under CC BY 4.0. ManyPeptidesMD is distributed under the terms stated in its dataset repository.