k-fold-structure/repsp-benchmark

RepSP Benchmark

Everything needed to run the folding / co-folding / probing benchmarks except per-encoder features (extract those from k-fold-structure/repsp-triprorep-tokens with your encoder, see the GitHub release README).

The benchmark goes apo → holo: predict properties of the bound homodimer ("holo") from the unbound monomer ("apo"). The two chains are identical (homodimer), so each residue position carries a single label.

Paper: Atom-level Protein Representation Learning Improves Protein Structure Prediction (arXiv:2605.22133). GitHub release: hsjang0/TriProRep.

Contents

splits/
    folding/{train,valid,test}.txt        # 390,627 / 400 / 1,000 homodimer AFids
    probing/{train,valid,test}.txt        # 39,100 / 400 / 1,000 (10% of folding train)
probing/
    labels.pkl                            # per-residue labels: binding_site / delta_sasa_mean / levy_tier / bond_type_plip
REPSP_PDB/
    monomer/{valid,test}.tar.gz           # <AF-id>_monomer.pdb, chain A (apo)
    monomer/train_NNN.tar.gz              # sharded (~40 GB each)
    homodimer/{valid,test}.tar.gz         # <AF-id>.pdb, chain A + B (holo)
    homodimer/train_NNN.tar.gz            # sharded (~40 GB each)
boltz_apo_tokens/   shard*.tar            # Boltz-tokenized apo monomers (folding input)
boltz_apo_targets/  shard*.tar            # apo structure targets
boltz_holo_tokens/  shard*.tar            # Boltz-tokenized holo dimers (co-folding input)
boltz_holo_targets/ shard*.tar            # holo structure targets

The AFid in every filename is the same homodimer identifier taken from splits/{folding,probing}/{train,valid,test}.txt. A monomer PDB is resolved as REPSP_PDB/monomer/<AF-id>_monomer.pdb, the homodimer as REPSP_PDB/homodimer/<AF-id>.pdb.

Download

# Small assets (splits, labels, Boltz tokens)
hf download k-fold-structure/repsp-benchmark --repo-type dataset --local-dir ./benchmark \
    --exclude "REPSP_PDB/*"
cd ./benchmark/boltz_holo_tokens && for t in shard*.tar; do tar xf "$t"; done && rm shard*.tar

# Monomer PDBs, test split only (about 45 MB compressed)
hf download k-fold-structure/repsp-benchmark --repo-type dataset --local-dir ./benchmark \
    --include "REPSP_PDB/monomer/test.tar.gz"
mkdir -p ./REPSP_PDB/monomer
tar xzf ./benchmark/REPSP_PDB/monomer/test.tar.gz -C ./REPSP_PDB/monomer/

# For folding / co-folding training, grab train shards + the homodimer side.

Notes

• Splits are LMDB-cleaned. A small number of AFids that fail Boltz tokenization are already dropped, so downstream stages won't hit KeyErrors.
• The boltz_* dirs use the tokenized layout the folding trunk (SimpleFold) reads: manifest.json + records/ + tokens/ (or structures/ for targets). Point the folding datamodule's tokenized_dir / target_dir at these folders after extracting.

License and attribution

• Code, splits, Boltz tokens, and probing labels: MIT.
• Structures under REPSP_PDB/: the homodimer PDBs are AFDB-Multimer predictions provided by NVIDIA to the AlphaFold Protein Structure Database; the apo monomers are AlphaFold-2 single-chain predictions we generated. Both are redistributed here under CC BY 4.0 with attribution to DeepMind and EMBL-EBI, per the AFDB terms of use. Cite Jumper et al., 2021 and Varadi et al., 2022, 2024 alongside the paper below.

Citation

@misc{triprorep,
  title  = {Atom-level Protein Representation Learning Improves Protein Structure Prediction},
  author = {Kim, Taewon and Jang, Hyosoon and Seo, Hyunjin and Seo, Seonghwan and Kim, Hyeongwoo and Zhung, Wonho and Shin, Mingyeong and Kim, Wooyoun and Ahn, Sungsoo},
  year   = {2026},
  eprint = {2605.22133},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url    = {https://arxiv.org/abs/2605.22133}
}