xwang112358/BioBlobs-data

BioBlobs data

The ProteinShake and VenusX data behind Unsupervised discovery of Functional Substructures (Xin Wang, Kaiwen Shi and Carlos Oliver, NeurIPS 2026), in the layout the BioBlobs code reads. Paper · Code · Model

It holds the protein structures, labels and train, validation and test splits. For VenusX it also holds the paper's ESM2-T30 and SaProt-650M residue embeddings. Build the ProteinShake embeddings with the code, as described below.

Contents

Folder	Task	Structures	Source
ec_proteinshake/	EC-L3, enzyme class	15,603	ProteinShake, RCSB PDB
pfam_proteinshake/	Pfam, protein family	31,109	ProteinShake, RCSB PDB
go_molecular_function_proteinshake/	GO-MF, molecular function	32,633	ProteinShake, RCSB PDB
scop_fam_proteinshake/	SCOP-FAM, SCOP family	10,066	ProteinShake, RCSB PDB
scop_sf_proteinshake/	SCOP-SF, SCOP superfamily	10,066	ProteinShake, RCSB PDB
venusx_site_fragment/{Act,BindI,Evo,Motif}/MF50/	VenusX functional sites	1,743 / 2,038 / 12,776 / 2,368	VenusX, AlphaFold DB

Each ProteinShake task has a sequence and a structure split, which assign clusters of similar sequences or structures (0.7 similarity) to disjoint splits. VenusX uses its Mixed-Family MF-50 split. SCOP-FAM and SCOP-SF share the same structures, which are included in both folders.

The VenusX embeddings are the ones behind the paper's VenusX results. They cover 1,732 / 2,028 / 12,723 / 2,358 proteins of Act / BindI / Evo / Motif. The other 84 proteins have AlphaFold2 structures with residues past the end of their VenusX site annotation, and the paper's embedding build skipped them. The code leaves them out of every split, which gives the paper's test sets of 186 / 203 / 1,291 / 249 samples. The embeddings take 33.0 GB: 11.0 GB for ESM2-T30 and 22.0 GB for SaProt-650M, stored as float16.

<task>_proteinshake/
  pdb.tar.gz                          pdb/<pdb_id>.pdb
  labels.csv, token_maps.json         EC-L3
  targets.jsonl, token_map.json       Pfam, GO-MF, SCOP-FAM, SCOP-SF
  sequence/{train,val,test}_split.csv
  structure/{train,val,test}_split.csv
venusx_site_fragment/<target>/MF50/
  pdb.tar.gz                          pdb/<uniprot_id>.pdb
  samples.csv                         one row per protein and InterPro site
  {train,val,test}_split.csv
  token_maps.json, metadata.json
  esm2_static_cache_mmap/             ESM2-T30 embeddings: embeddings.bin, meta.pt
  saprot_static_cache_mmap/           SaProt-650M embeddings: embeddings.bin, meta.pt

embeddings.bin stacks the per-residue embeddings of all proteins. meta.pt holds the protein names and row offsets and loads with torch.load(..., weights_only=True).

The download is 38.2 GB, of which 33.0 GB are the VenusX embeddings. The structures unpack to 23.6 GB. SHA256SUMS lists a checksum for every file (sha256sum -c SHA256SUMS).

Usage

Download the repository as the code's data/ folder and unpack the structures:

hf download xwang112358/BioBlobs-data --repo-type dataset --local-dir data
# To skip the 33 GB of VenusX embeddings, add --exclude "*_static_cache_mmap/*"
for f in data/*_proteinshake/pdb.tar.gz data/venusx_site_fragment/*/MF50/pdb.tar.gz; do
  tar -xzf "$f" -C "$(dirname "$f")"
done

Then follow docs/REPRODUCE.md in the code repository. The data-preparation scripts find the unpacked files, skip the download and parse the structures. For ProteinShake, then build the ESM2-T30 or SaProt-650M residue embeddings with preprocessing/prepare_plm_cache.py. SaProt also needs Foldseek, which reads the PDB files. VenusX training reads the downloaded embeddings, so skip this step for VenusX.

Licenses

• ProteinShake files (*_proteinshake/): CC-BY-4.0, as released by ProteinShake. ProteinShake builds these tasks from RCSB PDB structures and their annotations. Its documentation lists the source of each task.
• VenusX files (venusx_site_fragment/): Apache-2.0, as released by VenusX (AI4Protein/VenusX_Res_*_MF50 and AI4Protein/VenusX_*_AlphaFold2_PDB). The site annotations come from InterPro, and the structures are AlphaFold DB models (CC-BY-4.0).
• VenusX embeddings (venusx_site_fragment/*/MF50/*_static_cache_mmap/): Apache-2.0, like the VenusX files they are computed from. They are outputs of ESM2-T30 and SaProt-650M, both released under the MIT License.

Citation

If you use this data, please cite BioBlobs and the original datasets.

@inproceedings{wang2026bioblobs,
  title     = {Unsupervised discovery of Functional Substructures},
  author    = {Wang, Xin and Shi, Kaiwen and Oliver, Carlos},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

Also cite ProteinShake (Kucera et al., NeurIPS 2023 Datasets and Benchmarks), VenusX (Tan et al., 2025), InterPro and AlphaFold DB. If you use the VenusX embeddings, also cite ESM2 (Lin et al., Science 2023) and SaProt (Su et al., ICLR 2024).