bisectgroup/Leak-CURBER-enzyme-classification

Leak-CURBER enzyme classification P-seq

Summary

Leak-CURBER is a benchmark for enzymatic reaction tasks (NeurIPS 2026). This repository holds the leaderboard split enzyme classification P-seq at threshold 0.385. It contains the paper's train and validation rows, merged and shuffled with seed 42. The test set is held out for the leaderboard.

Rows

281,529 rows in one train split.

Split

The P-seq split at threshold 0.385: test proteins have a maximum sequence identity to the training proteins below 0.385 (MMseqs2).

Columns

Column	Type	Meaning
uniprot_id	string	UniProt accession.
sequence	string	Amino acid sequence (protein input).
ec_number	string	Label: one full four-level EC number. A protein with several ECs has several rows.
uniprot_date	timestamp[ns]	UniProt entry creation date (helper column).
pdbs	list<element: string>	List of structure ids. Predicted structures hold the UniProt accession. Experimental ones hold PDB:accession.

Target

ec_number, a full four-level EC number (multi-label: one row per protein-EC pair).

Sources and license

This dataset is released for academic, non-commercial use under CC BY-NC-SA 4.0. The terms of each source database also apply. Users must follow them.

• BRENDA, CC BY 4.0 (https://www.brenda-enzymes.org/license.php).
• Rhea, CC BY 4.0 (https://www.rhea-db.org/help/license-disclaimer).
• SABIO-RK, HITS terms for internal non-commercial research and academic use (https://sabiork.h-its.org/layouts/content/termscondition.gsp). This source restricts use to academic, non-commercial research.
• KEGG, Kanehisa Laboratories terms (https://www.kegg.jp/kegg/legal.html). This source restricts use to academic, non-commercial research.
• UniProt, CC BY 4.0 (https://www.uniprot.org/help/license): protein sequences and annotation dates.
• RCSB PDB, CC0 1.0 (https://www.rcsb.org/pages/usage-policy), and AlphaFold DB, CC BY 4.0 (https://alphafold.ebi.ac.uk/faq): structure ids in the pdbs column refer to these databases, where present.

Reaction records come from these four databases. Proteins and their ECs are linked through these records. Rows by source: rhea 255,320; kegg 70,773; sabio 46,173; brenda 11,618 (joined on protein and EC; a row can touch several sources, so the counts do not add up to the row total).

Citation

@inproceedings{ahmed2026leakcurber,
  title     = {Leak-CURBER: A Leakage-Controlled Multimodal Evaluation Benchmark for Enzymatic Reaction Tasks},
  author    = {Ahmed, Adhil and Balaji, Roshan and Bhatt, Nirav Pravinbhai},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

Code: https://github.com/bisect-group/Leak-CURBER

Please also cite the sources this dataset uses:

@article{changBRENDAELIXIRCore2021,
    title = {{BRENDA}, the {ELIXIR} core data resource in 2021: new developments and updates},
    volume = {49},
    issn = {0305-1048},
    url = {https://doi.org/10.1093/nar/gkaa1025},
    doi = {10.1093/nar/gkaa1025},
    number = {D1},
    journal = {Nucleic Acids Research},
    author = {Chang, Antje and Jeske, Lisa and Ulbrich, Sandra and Hofmann, Julia and Koblitz, Julia and Schomburg, Ida and Neumann-Schaal, Meina and Jahn, Dieter and Schomburg, Dietmar},
    month = jan,
    year = {2021},
    pages = {D498--D508},
}

@article{bansalRheaReactionKnowledgebase2022,
    title = {Rhea, the reaction knowledgebase in 2022},
    volume = {50},
    issn = {0305-1048},
    url = {https://doi.org/10.1093/nar/gkab1016},
    doi = {10.1093/nar/gkab1016},
    number = {D1},
    journal = {Nucleic Acids Research},
    author = {Bansal, Parit and Morgat, Anne and Axelsen, Kristian B and Muthukrishnan, Venkatesh and Coudert, Elisabeth and Aimo, Lucila and Hyka-Nouspikel, Nevila and Gasteiger, Elisabeth and Kerhornou, Arnaud and Neto, Teresa Batista and Pozzato, Monica and Blatter, Marie-Claude and Ignatchenko, Alex and Redaschi, Nicole and Bridge, Alan},
    month = jan,
    year = {2022},
    pages = {D693--D700},
}

@article{wittigSABIORKUpdatedResource2018,
    title = {{SABIO}-{RK}: an updated resource for manually curated biochemical reaction kinetics},
    volume = {46},
    issn = {0305-1048},
    url = {https://doi.org/10.1093/nar/gkx1065},
    doi = {10.1093/nar/gkx1065},
    number = {D1},
    journal = {Nucleic Acids Research},
    author = {Wittig, Ulrike and Rey, Maja and Weidemann, Andreas and Kania, Renate and Müller, Wolfgang},
    month = jan,
    year = {2018},
    pages = {D656--D660},
}

@article{kanehisaKEGGBiologicalSystems2025,
    title = {{KEGG}: biological systems database as a model of the real world},
    volume = {53},
    issn = {1362-4962},
    url = {https://doi.org/10.1093/nar/gkae909},
    doi = {10.1093/nar/gkae909},
    number = {D1},
    journal = {Nucleic Acids Research},
    author = {Kanehisa, Minoru and Furumichi, Miho and Sato, Yoko and Matsuura, Yuriko and Ishiguro-Watanabe, Mari},
    month = jan,
    year = {2025},
    pages = {D672--D677},
}

@article{theuniprotconsortiumUniProtUniversalProtein2025,
    title = {{UniProt}: the {Universal} {Protein} {Knowledgebase} in 2025},
    volume = {53},
    issn = {1362-4962},
    url = {https://doi.org/10.1093/nar/gkae1010},
    doi = {10.1093/nar/gkae1010},
    number = {D1},
    journal = {Nucleic Acids Research},
    author = {{The UniProt Consortium}},
    month = jan,
    year = {2025},
    pages = {D609--D617},
}