# AF2_UniRef50: SaProt Structure-Aware Pre-training Set

## Facts

- Provider: Westlake University REPL lab (Su, Han, Zhou, Shan, Zhou and Yuan; SaProt)
- License: CC-BY-4.0 (https://alphafold.ebi.ac.uk/download)
- Size: 8 files
- Format: CSV (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/westlake-repl/AF2_UniRef50
- Access: public, readable without an account
- Category: Structural Biology
- Source: https://openreview.net/forum?id=6MRm3G4NiU
- Page: https://scigantic.com/archive/dd60cee2-dc22-59ac-92b6-b4767e923164
- Keywords: Structural Biology, protein structure, protein language model, CSV, Hugging Face

## About

Pre-training dataset for SaProt (Protein Language Modeling with Structure-aware Vocabulary, ICLR 2024) from the Westlake University REPL lab, linked from the SaProt repository as the dataset used to pre-train the model. Proteins are UniRef50 representatives with AlphaFold2 predicted structures from the AlphaFold Protein Structure Database, each encoded as a structure-aware sequence that interleaves the amino acid letter with the Foldseek 3Di structural letter for every residue; per the SaProt documentation the character # marks residues in low-confidence regions (pLDDT below 70). Measured 2026-11-14: train.csv (21.9 GB) and valid.csv (11 MB) with columns uniprot_id and seq, plus the same training data as an LMDB database (train/data.mdb, 36.2 GB); repository storage about 62 GB. Open the CSVs with pandas, pyarrow or datasets; open the LMDB with the lmdb package as shown on the dataset card, where each record is JSON. Uses include pre-training or fine-tuning protein language models on combined sequence and structure tokens, Foldseek 3Di modelling, and transfer to mutation effect, stability and function prediction. The structures derive from AlphaFold DB, which EMBL-EBI and Google DeepMind distribute under CC BY 4.0; cite SaProt and AlphaFold DB.

## Load the data

```python
import pandas as pd
df = pd.read_csv("hf://datasets/westlake-repl/AF2_UniRef50/train.csv")
df.head()
```

## Data structure

File types in sampled files: CSV

Example file: `train.csv`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/dd60cee2-dc22-59ac-92b6-b4767e923164
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/dd60cee2-dc22-59ac-92b6-b4767e923164
- Over MCP: `get_schema_card` with id `dd60cee2-dc22-59ac-92b6-b4767e923164` at https://api.scigantic.com/mcp

## Related topics

- [CSV datasets](https://scigantic.com/topics/csv)
- [Structural Biology datasets](https://scigantic.com/topics/structural-biology)
- [Protein structure datasets](https://scigantic.com/topics/protein-structure)
- [Protein language model datasets](https://scigantic.com/topics/protein-language-model)
