bisectgroup/Leak-CURBER-embeddings

Leak-CURBER embeddings

This repository holds precomputed embeddings for all Leak-CURBER entities: proteins, molecules and reactions. Leak-CURBER is a NeurIPS 2026 benchmark for enzymatic reaction tasks.

The keys cover all splits of all tasks. The repository has inputs only. It has no targets and no split labels.

The viewer shows the key index of each store (key_hash, shard_id, row_idx). The vectors are in .npy shards.

Stores

Each store is a folder with meta.json, index.parquet, shard_*.npy and, for some stores, failures.parquet. Sizes are in GB (10^9 bytes). The repository has 47.6 GB in total.

Path	Modality	Model	Key (type)	Vector	Keys	Shards	Size
esmc/ESMC-6B/hf_v1	protein	ESM C 6B (biohub/ESMC-6B)	sequence (raw sequence)	float16 x 2560	283,552	3	1.46
esm3/open_small_structure/v1	protein structure	ESM3 open small (structure input)	acc_id (accession and structure source)	float16 x 1536	425,815	5	1.32
unimol_sdf/unimolv2/v1	molecule (3D)	Uni-Mol2 1.1B	SDF conformer stem (one row per conformer)	float16 x 1536	12,635,437	127	39.12
molr/2d/v1	molecule (2D)	MolR	smiles (canonical SMILES)	float16 x 1024	1,178,136	5	2.43
smited/materials_smi_ted_fork/v1	molecule (SMILES)	SMI-TED, an open-source family of large encoder-decoder foundation models for chemistry (bisectgroup/materials-smi-ted-fork)	smiles (canonical SMILES)	float16 x 768	1,179,350	4	1.83
ecfp/radius2_nbits1024/v1	molecule	ECFP (RDKit Morgan)	smiles (canonical SMILES)	uint8 x 1024	1,179,350	3	1.23
drfp/radius3_min0_len2048_rings0_h0/v1	reaction	DRFP	rxn_key	uint8 x 2048	86,075	2	0.18
rxnfp/default/v1	reaction	rxnfp (default BERT)	rxn_smiles	float16 x 256	88,521	1	0.05

How each vector is computed:

• ESM C: mean of the last hidden state over the residue positions (the start and end tokens are not used). The model runs in bfloat16. The vector is stored as float16. The key is the raw amino acid sequence.
• ESM3: mean of the ESM3 open small structure encoder output over the residue positions. The input is the structure of the protein, not only its sequence. The encoder runs in bfloat16. The key is <accession>|<source>, where the source is esm, alphafold, experimental or esmfold2. An experimental chain has the key <PDB id>:<UniProt accession>|experimental.
• Uni-Mol2: the molecule-level representation of Uni-Mol2 1.1B for one 3D conformer (no atom-level output). The key is the name of the SDF conformer file. It has the form pubchem_<sha256(smiles_key)> for the PubChem conformer or datamol_<sha256(smiles_key)>_conf0 to _conf9 for up to ten RDKit conformers. Each conformer has its own row, so one molecule has up to 11 rows. The store is indexed by conformer, not by SMILES. The paper mean-pools the rows that share one SMILES key into one vector per molecule. Do this pooling yourself. smiles_key is the canonical SMILES in the Leak-CURBER tables, and its SHA-256 is the smiles_hash column.
• MolR: output of the MolR graph encoder on the 2D molecular graph, with the fixed scale factor 1536.0. The key is the canonical SMILES.
• SMI-TED: output of the autoencoder encoder of SMI-TED on the token embeddings of the SMILES. The key is the canonical SMILES. The vector is computed with bfloat16 autocast and stored as float16.
• ECFP: Morgan fingerprint, radius 2, 1024 bits, chirality off. Values are 0 or 1. The key is the canonical SMILES.
• DRFP: differential reaction fingerprint, radius 3, minimum radius 0, 2048 bits, no rings, no hydrogens. Values are 0 or 1. The key is the reaction key used for grouping in the Leak-CURBER tables.
• rxnfp: the [CLS] vector of the default rxnfp BERT model. The key is the reaction SMILES.

Format and loading

meta.json describes the store: embedder, model name, version, embedding_dim, storage_dtype, key_field, key_type and max_shard_bytes (512 MiB).

index.parquet has three columns. key_hash is the first 24 hex characters of the SHA-256 of the UTF-8 key. shard_id and row_idx give the shard file and the row of the vector. The index does not store the raw keys, so you must hash your key.

Shards are shard_<shard_id:06d>.npy. Each shard is a 2D array that you can load with numpy memory mapping.

import hashlib

import numpy as np
import pandas as pd
from huggingface_hub import snapshot_download

store = "esmc/ESMC-6B/hf_v1"
root = snapshot_download(
    "bisectgroup/Leak-CURBER-embeddings", repo_type="dataset", allow_patterns=f"{store}/*"
)
index = pd.read_parquet(f"{root}/{store}/index.parquet").set_index("key_hash")

key = "ADDRNPLEECFRETDYEEFLEIAKNGLSTT"  # a protein sequence in the index
key_hash = hashlib.sha256(key.encode("utf-8")).hexdigest()[:24]
shard_id, row = index.loc[key_hash, ["shard_id", "row_idx"]]
shard = np.load(f"{root}/{store}/shard_{shard_id:06d}.npy", mmap_mode="r")
vector = np.asarray(shard[row])  # float16, shape (2560,)

For the ESM3 store, hash "<accession>|<source>". For the Uni-Mol2 store, hash the SDF file stem of each conformer, then mean-pool the conformer vectors of one SMILES key.

Notes

• Empty fingerprints give all-zero rows: DRFP has 1,641 such rows. ECFP has none.
• failures.parquet lists the keys that could not be embedded, with raw_key and error. ESM C has 2 failures (out of memory). ESM3 has 10,500: 10,255 without a structure file and 245 where the ESMFold2 sequence differs from the dataset sequence (these 245 keys are also in the index). MolR has 1,214 SMILES that could not be parsed. The other stores have no failures file.
• The Uni-Mol2 shards hold 31,535 rows that have no index entry. Ignore them. The index also has conformer keys of molecules that are not in the current Leak-CURBER tables.
• ESM3 keys are structure accessions with a source suffix, not sequences.
• rxnfp vectors are stored as float16. Values differ slightly with batch composition (max abs about 2e-3).

License

This dataset is released for academic, non-commercial use under CC BY-NC-SA 4.0.

The keys are sequences, SMILES and reactions from the source databases (BRENDA, SABIO-RK, BindingDB including ChEMBL-derived records, Rhea, KEGG, MetaCyc and UniProt). The terms of these databases also apply.

Model licenses: ESM C and ESM3 are MIT, and the Biohub Acceptable Use Policy also applies. Uni-Mol2, MolR, rxnfp and DRFP are MIT. SMI-TED is Apache-2.0. ECFP uses RDKit (BSD-3).

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

The keys come from the Leak-CURBER sources. Cite the sources through the dataset repositories listed below.

Please also cite the models this dataset uses:

@misc{EvolutionaryScaleESM,
    title = {Evolutionary {Scale} · {ESM} {Cambrian}: {Revealing} the mysteries of proteins with unsupervised learning},
    url = {https://www.evolutionaryscale.ai/blog/esm-cambrian},
    language = {en},
}

@misc{candidoLanguageModelingMaterializes2026,
    title = {Language {Modeling} {Materializes} a {World} {Model} of {Protein} {Biology}},
    copyright = {© 2026, Posted by openRxiv. This pre-print is available under a Creative Commons License (Attribution 4.0 International), CC BY 4.0, as described at http://creativecommons.org/licenses/by/4.0/},
    url = {https://www.biorxiv.org/content/10.64898/2026.06.03.729735v1},
    doi = {10.64898/2026.06.03.729735},
    language = {en},
    publisher = {bioRxiv},
    author = {Candido, Salvatore and Hayes, Thomas and Derry, Alexander and Rao, Roshan and Lin, Zeming and Verkuil, Robert and Wu, Bryan Z. and Lee, Jin Sub and Bruguera, Elise S. and Keval, Jehan A. and Kopylov, Mykhailo and Pak, John E. and Wu, Wesley and Thomas, Neil and Mataraso, Samson and Hsu, Alvin and Trotman-Grant, Ashton C. and Fatras, Kilian and Costa, Allan dos Santos and Badkundri, Rohil and Akın, Halil and Oktay, Deniz and Deaton, Jonathan and Montabana, Elizabeth and Sitwala, Hrishita and Yu, Yue and Wiggert, Marius and Carlin, Dylan Alexander and Goering, Anthony W. and Blazejewski, Tomasz and Sandora, McCullen and Hla, Michael and Jia, Tina Z. and Kloker, Leon H. and Sofroniew, Nicholas J. and Uehara, Masatoshi and Pannu, Jassi and Bachas, Sharrol and Liu, Daniel S. and Sercu, Tom and Rives, Alexander},
    month = jun,
    year = {2026},
}

@article{hayesSimulating500Million2025,
    title = {Simulating 500 million years of evolution with a language model},
    volume = {387},
    url = {https://www.science.org/doi/10.1126/science.ads0018},
    doi = {10.1126/science.ads0018},
    number = {6736},
    journal = {Science},
    publisher = {American Association for the Advancement of Science},
    author = {Hayes, Thomas and Rao, Roshan and Akin, Halil and Sofroniew, Nicholas J. and Oktay, Deniz and Lin, Zeming and Verkuil, Robert and Tran, Vincent Q. and Deaton, Jonathan and Wiggert, Marius and Badkundri, Rohil and Shafkat, Irhum and Gong, Jun and Derry, Alexander and Molina, Raul S. and Thomas, Neil and Khan, Yousuf A. and Mishra, Chetan and Kim, Carolyn and Bartie, Liam J. and Nemeth, Matthew and Hsu, Patrick D. and Sercu, Tom and Candido, Salvatore and Rives, Alexander},
    month = feb,
    year = {2025},
    pages = {850--858},
}

@inproceedings{jiExploringMolecularPretraining2024,
    title = {Exploring {Molecular} {Pretraining} {Model} at {Scale}},
    volume = {37},
    url = {https://proceedings.neurips.cc/paper_files/paper/2024/hash/53923bb44655a7defb31c7744c01b62b-Abstract-Conference.html},
    doi = {10.52202/079017-1489},
    booktitle = {Advances in {Neural} {Information} {Processing} {Systems}},
    publisher = {Curran Associates, Inc.},
    author = {Ji, Xiaohong and Wang, Zhen and Gao, Zhifeng and Zheng, Hang and Zhang, Linfeng and Ke, Guolin and E, Weinan},
    year = {2024},
    pages = {46956--46978},
}

@misc{wangChemicalReactionAwareMoleculeRepresentation2021,
    title = {Chemical-{Reaction}-{Aware} {Molecule} {Representation} {Learning}},
    url = {http://arxiv.org/abs/2109.09888},
    doi = {10.48550/arXiv.2109.09888},
    publisher = {arXiv},
    author = {Wang, Hongwei and Li, Weijiang and Jin, Xiaomeng and Cho, Kyunghyun and Ji, Heng and Han, Jiawei and Burke, Martin D.},
    month = sep,
    year = {2021},
}

@article{soaresOpensourceFamilyLarge2025,
    title = {An open-source family of large encoder-decoder foundation models for chemistry},
    volume = {8},
    copyright = {2025 The Author(s)},
    issn = {2399-3669},
    url = {https://www.nature.com/articles/s42004-025-01585-0},
    doi = {10.1038/s42004-025-01585-0},
    language = {en},
    number = {1},
    journal = {Communications Chemistry},
    publisher = {Nature Publishing Group},
    author = {Soares, Eduardo and Vital Brazil, Emilio and Shirasuna, Victor and Zubarev, Dmitry and Cerqueira, Renato and Schmidt, Kristin},
    month = jul,
    year = {2025},
    pages = {193},
}

@article{schwallerMappingSpaceChemical2021,
    title = {Mapping the space of chemical reactions using attention-based neural networks},
    volume = {3},
    copyright = {2021 The Author(s), under exclusive licence to Springer Nature Limited},
    issn = {2522-5839},
    url = {https://www.nature.com/articles/s42256-020-00284-w},
    doi = {10.1038/s42256-020-00284-w},
    language = {en},
    number = {2},
    journal = {Nature Machine Intelligence},
    publisher = {Nature Publishing Group},
    author = {Schwaller, Philippe and Probst, Daniel and Vaucher, Alain C. and Nair, Vishnu H. and Kreutter, David and Laino, Teodoro and Reymond, Jean-Louis},
    month = feb,
    year = {2021},
    pages = {144--152},
}

@article{probstReactionClassificationYield2022,
    title = {Reaction classification and yield prediction using the differential reaction fingerprint {DRFP}},
    volume = {1},
    issn = {2635-098X},
    url = {https://pubs.rsc.org/en/content/articlelanding/2022/dd/d1dd00006c},
    doi = {10.1039/D1DD00006C},
    language = {en},
    number = {2},
    journal = {Digital Discovery},
    publisher = {RSC},
    author = {Probst, Daniel and Schwaller, Philippe and Reymond, Jean-Louis},
    month = apr,
    year = {2022},
    pages = {91--97},
}

@article{rogersExtendedConnectivityFingerprints2010,
  title = {Extended-Connectivity Fingerprints},
  author = {Rogers, David and Hahn, Mathew},
  journal = {Journal of Chemical Information and Modeling},
  volume = {50},
  number = {5},
  pages = {742--754},
  year = {2010},
  doi = {10.1021/ci100050t}
}

Dataset repositories

• https://huggingface.co/datasets/bisectgroup/Leak-CURBER-kcat
• https://huggingface.co/datasets/bisectgroup/Leak-CURBER-Km
• https://huggingface.co/datasets/bisectgroup/Leak-CURBER-Ki
• https://huggingface.co/datasets/bisectgroup/Leak-CURBER-Kd
• https://huggingface.co/datasets/bisectgroup/Leak-CURBER-EC50
• https://huggingface.co/datasets/bisectgroup/Leak-CURBER-IC50
• https://huggingface.co/datasets/bisectgroup/Leak-CURBER-enzyme-classification
• https://huggingface.co/datasets/bisectgroup/Leak-CURBER-enzyme-retrieval
• https://huggingface.co/datasets/bisectgroup/Leak-CURBER-reaction-outcome