Yukk1Zz/if3-crossdocked2020

CrossDocked2020 — IF3 preprocessed

Preprocessed CrossDocked2020 dataset for training the IF3 EGNN-CFM pocket-conditioned 3D molecular generator. Data is derived from the Pocket2Mol / 3D-SBDD / DiffSBDD / TargetDiff community release.

Contents

File	Size	Purpose
train.lmdb/	2.7 GB	84,393 pocket–ligand pairs for training
val.lmdb/	297 MB	9,362 pairs for validation (10% of train pool)
split_by_name.pt	15 MB	Original train/test split (by pocket family)
index_train.csv	12 MB	90,000 train pairs before RDKit filtering
index_val.csv	1.4 MB	10,000 val pairs
index_test.csv	16 KB	100 held-out test pairs (pocket2mol/targetdiff standard)
crossdocked_pocket10.tar.gz	1.6 GB	Raw archive (pocket PDBs within 10 Å + ligand SDFs)

LMDB schema

Each entry in train.lmdb / val.lmdb is a pickled dict:

{
    "lig_pos":       np.ndarray [N_lig, 3]   float32,  # ligand atom coords (Å)
    "lig_atom_type": np.ndarray [N_lig]      int64,    # ATOM_VOCAB index 0..10 (C,N,O,S,F,Cl,Br,P,I,B,Se)
    "pock_pos":      np.ndarray [N_pock, 3]  float32,  # pocket atom coords (Å)
    "pock_feat":     np.ndarray [N_pock, 44] float32,  # element one-hot (20) + residue one-hot (20) + is_backbone (1) + secondary structure (3)
    "name":          str,                              # "<pocket_stem>_<ligand_stem>"
}

Plus a metadata key b"__meta__" with {"n_samples", "pocket_radius", "max_lig_atoms"}.

Filtering applied during preprocessing

• rmsd < 1.0 (docked-pose quality gate, from index.pkl inside the raw tarball)
• 3 ≤ N_heavy_atoms ≤ 50
• Non-empty pocket within 10 Å of ligand centroid

Final yield: 84,393 train + 9,362 val (from 90,000 / 10,000 candidates).

Usage

from huggingface_hub import snapshot_download
local = snapshot_download(repo_id="Yukk1Zz/if3-crossdocked2020", repo_type="dataset")

from if3.data.crossdocked import CrossDockedDataset
from if3.data.collate import MolGraphCollator
from torch.utils.data import DataLoader

ds = CrossDockedDataset(f"{local}/train.lmdb")
loader = DataLoader(ds, batch_size=32, shuffle=True, collate_fn=MolGraphCollator())
batch = next(iter(loader))  # MolGraphBatch

Reproducibility

Regenerate from scratch using the IF3 repo:

git clone https://github.com/NakashimaYuki/if3
cd if3
python scripts/download_crossdocked.py --step split       # needs raw tarball
python scripts/download_crossdocked.py --step preprocess  # writes train/val LMDB

Citation

Please cite the original CrossDocked2020 work and the downstream papers that popularized this pre-processing (Luo et al. 3D Generative Model for SBDD, NeurIPS 2021; Peng et al. Pocket2Mol, ICML 2022; Guan et al. TargetDiff, ICLR 2023; Schneuing et al. DiffSBDD, 2023).

License

Inherited from CrossDocked2020 (CC BY 4.0).