Structure Prediction · LiteFold
protenix-data
Preprocessed biomolecular dataset for training AlphaFold3-style all-atom structure prediction models.
What protenix-data is
Protenix Data is a multimodal biomolecular training corpus produced by ByteDance's AML AI4Science Team. It provides the preprocessed structural and sequence data used to train Protenix, an open-source PyTorch reproduction of AlphaFold3 designed to predict 3D structures across proteins, DNA, RNA, ligands, ions, and chemical modifications.
Training all-atom biomolecular structure predictors requires vast collections of assemblies, paired multiple sequence alignments, and structural templates that are computationally intensive to assemble from scratch. Derived from the Worldwide Protein Data Bank (wwPDB), this dataset offers a standardised, drop-in training corpus for researchers developing and evaluating biomolecular deep learning architectures.
How it was made
The dataset is curated from wwPDB macromolecular crystallographic information files (mmCIF) and biological assemblies. The initial v0.5.0 release implements a structural cutoff date of 2021-09-30 to match AlphaFold3, whilst the Protenix-v1-20250630 release extends the curation cutoff to 2025-06-30. Multiple sequence alignments (MSAs) are generated via the ColabFold pipeline using UniRef30 and the ColabFold environmental database. UniRef sequences are paired across chains using taxonomy identifiers, with remaining sequences assigned to unpaired MSA features. Protenix-v1 additionally incorporates RNA MSAs and HMMER-derived structural templates.
The corpus contains 931,270 files totalling 1.05 TiB, distributed across 52 uncompressed tar shards with maximum payloads of 20 GiB each. Uncompressed packaging facilitates rapid random-access extraction of individual files without decompressing full archives. Files and their internal archive locations are catalogued in companion metadata files, including metadata.csv and shards.csv.
Key facts
- Total file count
- 931,270
- Dataset size
- 1.05 TiB (1,150,960,949,915 bytes)
- Archive format
- 52 uncompressed tar shards
- Supported modalities
- Proteins, DNA, RNA, ligands, ions, and chemical modifications
- Primary source
- Worldwide Protein Data Bank (wwPDB)
- PDB cutoff dates
- 2021-09-30 (v0.5.0) and 2025-06-30 (Protenix-v1)
- Common file formats
- .a3m, .cif, .pkl.gz, .fasta, .csv
What you can use it for
- AlphaFold3-style model training
- Train all-atom biomolecular diffusion models on paired protein, nucleic acid, ion, and ligand assemblies using preprocessed features.
- Biomolecular MSA and template retrieval
- Retrieve precomputed ColabFold multiple sequence alignments and HMMER structural templates across hundreds of thousands of PDB assemblies without running search pipelines locally.
- Complex prediction benchmarking
- Evaluate macromolecular modelling and small-molecule pose prediction performance using curated subsets such as PoseBusters and recent PDB biological assemblies.
Explore the data
No licence declared
The authors have not published a licence, which means no permission to redistribute the data has been granted to anyone. The rows are not served here; read them on the original dataset page.How the columns are distributed
Over 931,270 rows of files/train
directory
string_text- min
- 5
- median
- 23
- mean
- 17.94
- max
- 25
extension
string_label- .a3m52%
- .pkl.gz25%
- .cif24%
0.00% empty
filename
string_text- min
- 4
- median
- 11
- mean
- 11.41
- max
- 78
member_path
string_text- min
- 14
- median
- 36
- mean
- 30.35
- max
- 101
modified_utc
datetimepath
string_text- min
- 14
- median
- 36
- mean
- 30.35
- max
- 101
shard_path
string_label- shards/shard-00016.tar15%
- shards/shard-00017.tar9%
- shards/shard-00002.tar3%
size_bytes
int- min
- 0
- median
- 370k
- mean
- 1.2M
- max
- 81.0B
size_human
string_text- min
- 3
- median
- 9
- mean
- 9.09
- max
- 11
top_level
string_label- mmcif_msa_template51%
- mmcif24%
- mmcif_bioassembly24%
Strengths and limits
Good at
- Provides an all-atom training corpus covering proteins, nucleic acids, small-molecule ligands, ions, and chemical modifications.
- Includes precalculated ColabFold MSAs with taxonomy-based chain pairing alongside HMMER structural templates.
- Organised into uncompressed tar shards with file-level metadata indices to allow selective downloading and extraction of specific records.
Watch out for
- Requires substantial local storage, with the complete set of uncompressed shards occupying over 1.05 TiB.
- Restricted to structures deposited prior to the specified cutoff dates of September 2021 for v0.5.0 and June 2025 for Protenix-v1.
Questions
Can I look inside protenix-data without downloading it?
Not here: it declares no licence, so no permission to redistribute its rows has been granted. Its description and statistics are still shown.
What licence is protenix-data under?
Its authors have not declared a licence, which means no permission to reuse or redistribute it has been granted.
What biomolecules are included in the dataset?
The dataset contains coordinates, alignments, and templates for proteins, DNA, RNA, small-molecule ligands, ions, and modified residues curated from wwPDB entries.
How are the sequence alignments constructed?
MSAs are generated using the ColabFold pipeline against UniRef30 and the ColabFold environmental database. UniRef hits are paired across chains using taxonomy identifiers, while non-paired sequences are retained as unpaired MSA features.
What is the difference between release v0.5.0 and Protenix-v1?
Release v0.5.0 mirrors AlphaFold3 by applying a PDB deposition cutoff of 2021-09-30, whereas Protenix-v1 expands the curated set to a cutoff of 2025-06-30 and adds dedicated RNA MSAs and HMMER-derived templates.
Why are the tar shards left uncompressed?
Shards are intentionally uncompressed to avoid processing overhead on files that are already compressed, such as pickled archives, and to enable straightforward random extraction of individual files.
How to cite it
ByteDance AML AI4Science Team, Chen, Xinshi, Zhang, Yuxuan, Lu, Chan, Ma, Wenzhi, Guan, Jiaqi, Gong, Chengyue, Yang, Jincai, Zhang, Hanyu, Zhang, Ke, Wu, Shenghao, Zhou, Kuangqi, Yang, Yanping, Liu, Zhenyu, Wang, Lan, Shi, Bo, Shi, Shaochen, Xiao, Wenzhi. (2025). Protenix - Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction. bioRxiv. doi: 10.1101/2025.01.08.631967
Written from protenix-data's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.