Protenix Data: AlphaFold3-Style Biomolecular Structure Training Corpus

Protenix Data is ByteDance's open-source, preprocessed training dataset for its Protenix model, a PyTorch reproduction of AlphaFold3. The dataset is described as one of the largest publicly available AF3-style training corpora, built from the wwPDB. LiteFold released the dataset on Hugging Face, with a last recorded update on 2026-05-27.

Use Cases

• Training a unified all-atom diffusion model based on the described dataset scope covering proteins, DNA, RNA, ligands, ions, and modifications.
• Benchmarking structure prediction models against a large, publicly available AF3-style training corpus.
• Reproducing or extending the Protenix model's training pipeline using the released preprocessed data.

Strengths

• Described as one of the largest publicly available AF3-style training corpora.
• Covers multiple biomolecular types including proteins, DNA, RNA, ligands, ions, and modifications.
• Full preprocessed training dataset released alongside model code and weights.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license are unknown, which may limit suitability assessment.
• Freshness should be verified as the last update timestamp is from the future (2026-05-27).

Provenance

Source

ByteDance's Protenix project, released by LiteFold.

Collection Method

Built from the wwPDB and preprocessed for training.

Freshness

Last updated 2026-05-27 11:43:56; freshness should be verified.