README.md

metadata

pretty_name: Protenix Data
size_categories:
  - 100K<n<1M
task_categories:
  - other
tags:
  - biology
  - protein
  - rna
  - structure-prediction
  - tar
  - datasets
configs:
  - config_name: files
    default: true
    data_files:
      - split: train
        path: metadata.csv
  - config_name: shards
    data_files:
      - split: train
        path: shards.csv

Protenix Data

Protenix is ByteDance's open-source PyTorch reproduction of AlphaFold3, a biomolecular structure predictor that handles proteins, DNA, RNA, ligands, ions, and modifications under a unified all-atom diffusion model. Alongside the model code and weights, the team released the full preprocessed training dataset used to train Protenix and its successors, making it one of the largest publicly available AF3-style training corpora. The released data is built from the wwPDB and is intended to be drop-in usable for AF3-style training. The original v0.5.0 release uses a PDB cutoff of 2021-09-30 (matching AF3), and Protenix-v1-20250630 extends the cutoff to 2025-06-30 on a larger curated set. MSAs are generated with the ColabFold pipeline against UniRef30 and the ColabFold environmental database, with UniRef sequences retained for pairing via taxonomy IDs and the remainder folded into the unpaired MSA features. Protenix-v1 also adds RNA MSAs and HMMER-derived structural templates aligned with the AF3 template strategy.

Summary

Original files	931,270
Original payload	1.05 TiB (1,150,960,949,915 bytes)
Tar shards	52
Archive size	1.05 TiB (1,152,633,989,120 bytes)
Max shard payload	20.00 GiB
Metadata generated	2026-05-24T16:01:55Z

Source Layout

Directory	Files	Size
common	9	1.18 GiB
indices	4	28.58 MiB
mmcif	220,113	282.13 GiB
mmcif_bioassembly	219,912	34.89 GiB
mmcif_msa_template	475,724	624.02 GiB
posebusters_bioassembly	308	41.52 MiB
posebusters_mmcif	429	360.26 MiB
recentPDB_bioassembly	8,808	1.44 GiB
rna_msa	5,959	39.07 GiB
search_database	4	88.75 GiB

Common File Types

Extension	Files	Size
.a3m	479,867	658.57 GiB
.cif	220,543	283.34 GiB
.fasta	4	88.75 GiB
.pkl.gz	229,028	36.37 GiB
.csv	1,816	4.53 GiB
.pkl	2	255.97 MiB
.json	4	58.64 MiB
.csv.gz	1	25.70 MiB
.txt	2	20.70 MiB
.npz	2	435.23 KiB
	1	1.50 KiB

Shards

Shard	Files	Payload	Archive
shards/shard-00000.tar	3,743	19.99 GiB	20.00 GiB
shards/shard-00001.tar	2,219	19.69 GiB	19.69 GiB
shards/shard-00002.tar	26,895	20.00 GiB	20.04 GiB
shards/shard-00003.tar	17,553	20.00 GiB	20.03 GiB
shards/shard-00004.tar	23,622	19.91 GiB	19.95 GiB
shards/shard-00005.tar	21,546	20.00 GiB	20.03 GiB
shards/shard-00006.tar	18,365	20.00 GiB	20.03 GiB
shards/shard-00007.tar	12,514	20.00 GiB	20.02 GiB
shards/shard-00008.tar	16,486	20.00 GiB	20.03 GiB
shards/shard-00009.tar	14,994	20.00 GiB	20.02 GiB
shards/shard-00010.tar	12,684	19.99 GiB	20.01 GiB
shards/shard-00011.tar	10,438	19.97 GiB	19.99 GiB
shards/shard-00012.tar	13,431	20.00 GiB	20.02 GiB
shards/shard-00013.tar	9,903	20.00 GiB	20.01 GiB
shards/shard-00014.tar	10,788	20.00 GiB	20.02 GiB
shards/shard-00015.tar	9,355	20.00 GiB	20.02 GiB
shards/shard-00016.tar	142,265	20.00 GiB	20.24 GiB
shards/shard-00017.tar	81,080	20.00 GiB	20.13 GiB
shards/shard-00018.tar	15,468	20.00 GiB	20.02 GiB
shards/shard-00019.tar	14,558	20.00 GiB	20.02 GiB
shards/shard-00020.tar	14,807	20.00 GiB	20.02 GiB
shards/shard-00021.tar	13,074	20.00 GiB	20.02 GiB
shards/shard-00022.tar	12,558	19.99 GiB	20.01 GiB
shards/shard-00023.tar	15,179	19.99 GiB	20.02 GiB
shards/shard-00024.tar	13,741	20.00 GiB	20.02 GiB

Only the first 25 shards are shown here. See shards.csv for all 52 shards.

Metadata

metadata.csv columns:

Column	Description
path	Original relative path in the source folder.
shard_path	Tar shard containing the file.
member_path	Path of the file inside the tar shard.
top_level	First directory under the source folder.
directory	Parent directory of the file.
filename	File basename.
extension	File extension, preserving compound suffixes such as .csv.gz.
size_bytes	Original file size in bytes.
size_human	Human-readable original file size.
modified_utc	Local file modification timestamp captured during packaging.

shards.csv lists one row per tar shard. _MANIFEST.json contains the aggregate build summary.

Download Everything

pip install -U huggingface_hub
hf download LiteFold/protenix-data --repo-type dataset --local-dir ./protenix-data

Extract all shards:

mkdir -p ./data
for shard in ./protenix-data/shards/*.tar; do
  tar -xf "$shard" -C ./data
done

Use With datasets

Use the datasets API to query file metadata, then use huggingface_hub to download the shard that contains the file.

from datasets import load_dataset
from huggingface_hub import hf_hub_download
import tarfile

files = load_dataset("LiteFold/protenix-data", "files", split="train")
row = files[0]

shard = hf_hub_download(
    repo_id="LiteFold/protenix-data",
    repo_type="dataset",
    filename=row["shard_path"],
)

with tarfile.open(shard) as archive:
    archive.extract(row["member_path"], path="./data")

For streaming metadata:

from datasets import load_dataset

files = load_dataset("LiteFold/protenix-data", "files", split="train", streaming=True)
for row in files:
    print(row["path"], row["shard_path"])
    break

Notes

The tar shards are uncompressed by design. This keeps packaging and random extraction simple and avoids spending CPU compressing data that is often already compressed.

Citation

@article{bytedance2025protenix,
  title     = {Protenix - Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction},
  author    = {ByteDance AML AI4Science Team and Chen, Xinshi and Zhang, Yuxuan and Lu, Chan and Ma, Wenzhi and Guan, Jiaqi and Gong, Chengyue and Yang, Jincai and Zhang, Hanyu and Zhang, Ke and Wu, Shenghao and Zhou, Kuangqi and Yang, Yanping and Liu, Zhenyu and Wang, Lan and Shi, Bo and Shi, Shaochen and Xiao, Wenzhi},
  journal   = {bioRxiv},
  year      = {2025},
  publisher = {Cold Spring Harbor Laboratory},
  doi       = {10.1101/2025.01.08.631967},
  url       = {https://www.biorxiv.org/content/10.1101/2025.01.08.631967v1}
}