caijihuize/sp-3di-dataset

Swiss-Prot AA–3Di Dataset v3

This release pairs amino-acid sequences with Foldseek 3Di sequences derived from AlphaFold Database Swiss-Prot PDB v6 models. It provides sequence-cluster-aware train, validation, and test splits, split ID lists, and predicted PDB structures for the validation and test examples.

Load the sequence dataset

from datasets import load_dataset

base = "hf://datasets/caijihuize/sp-3di-dataset/data/"
dataset = load_dataset("parquet", data_files={
    "train": base + "train-00000-of-00001.parquet",
    "validation": base + "validation-00000-of-00001.parquet",
    "test": base + "test-00000-of-00001.parquet",
})
print(dataset["train"].column_names)
# ['sequence_3di', 'sequence_aa', 'id']

sequence_3di is the Foldseek 3Di sequence; sequence_aa is the amino-acid sequence. Each id is the stable AlphaFold model identifier. The corresponding ids/{train,valid,test}_ids.txt files list IDs in the same order as their split rows. The validation split is named validation in the Parquet dataset and valid in the ID and provenance filenames.

Contents

• data/*.parquet: one Parquet shard for each split.
• ids/: row-aligned split ID lists.
• structures/validation/ and structures/test/: 1,000 AlphaFold PDB files per split; filenames match the corresponding id. Training PDB files are not included in this release.
• provenance/: source archive checksum, software/build details, split counts, cluster membership, exclusions, and audit/validation summaries.
• checksums/SHA256SUMS: SHA-256 for every other file in this release.

Dataset construction

The source archive contained 550,122 paired AA/3Di/pLDDT records. Filtering retained 503,440 candidates and excluded 46,682. Final splits contain 364,752 training, 1,000 validation, and 1,000 test records. The evaluation structures were checked against their AA sequences; all 2,000 matched.

MMseqs2 clustering used 30% minimum sequence identity and 80% coverage of both sequences. Final cross-split sequence audits used the same thresholds and reported no alignments meeting them. Foldseek structural neighbors were recorded as an audit and were not used to remove examples. Therefore, the split supports sequence-level isolation under the stated search protocol; it does not claim fold novelty. See provenance/build_manifest.json and the processing repository for full details.

Source and attribution

Structures and source sequences were obtained from the AlphaFold Database Swiss-Prot PDB v6 archive. The source archive SHA-256 is recorded in provenance/source_manifest.json. AlphaFold Database data are made available under CC BY 4.0; please retain attribution to the AlphaFold Database and cite the AlphaFold methods paper.

Processing code and scripts: caijihuize/sp-3di-dataset on GitHub.

Limitations

The structures are AlphaFold predictions, not experimentally determined structures. pLDDT is a local confidence estimate and does not establish biological correctness. Sequence isolation does not guarantee structural-fold novelty. Pretraining or fine-tuning exposure of a model checkpoint is outside this dataset's audit.

Citations

• Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). https://doi.org/10.1038/s41586-021-03819-2
• van Kempen, M. et al. Fast and accurate protein structure search with Foldseek. Nature Biotechnology 42, 243–246 (2024). https://doi.org/10.1038/s41587-023-01773-0