Kilig33/DuoPose-database

DuoPose Processed Database

This repository contains the processed target-level database used by the DuoPose project and its evaluated model variants, including GeoPose and the DuoPose family. The archive is reconstructed from the DecoyDB-derived structural source. It is not a copy of the complete upstream DecoyDB repository. The processing pipeline used the project code to convert receptor, native-ligand, and decoy PDBQT structures into one PyTorch/PyTorch Geometric object per target.

Repository files

processed_full.tar
processed_full.tar.sha256
README.md

• processed_full.tar: archive containing the processed target-level files.
• processed_full.tar.sha256: SHA-256 checksum for the archive.
• README.md: dataset description, reconstruction details, download and verification instructions.

The archive contains one top-level directory, processed_full/, and exactly 61,085 target-level .pt files.

Dataset statistics

Item	Value
Complete raw target directories identified	61,104
Successfully processed targets	61,085
Raw targets excluded from formal statistics	19
Excluded fraction	0.031095%
Processed coverage of complete raw targets	99.968905%
Extracted processed tensor bytes	5,675,197,832
Train targets	48,868
Validation targets	6,108
Test targets	6,109
Total fixed-split targets	61,085

The 19 excluded raw targets did not produce valid processed objects. Because the excluded fraction is very small, no imputation was performed. Only successfully reconstructed targets were included in the formal processed dataset and the fixed train/validation/test split lists.

The split lists are maintained in the companion DuoPose source-code repository. The data archive itself does not create separate train/, val/, or test/ directories; all target files are stored together under processed_full/.

Upstream source and reconstruction

The original structural source is DecoyDB:

• DecoyDB GitHub repository
• DecoyDB Hugging Face dataset card

This project reconstructed a target-level pose-ranking database rather than redistributing the complete upstream repository. The main processing stages were:

1. Scan the DecoyDB-derived structural archive and group files by target.
2. Parse receptor, native-ligand, and decoy coordinates from PDBQT files.
3. Compute the geometric center of the native ligand.
4. Extract receptor atoms within an 8 Å pocket radius.
5. Construct 32-dimensional protein atom features.
6. Parse decoy MODEL blocks and retain decoys with atom counts compatible with the native ligand.
7. Recompute decoy RMSD values using Kabsch alignment.
8. Save one target-level .pt dictionary for each successfully processed target.
9. Re-read native ligands to add 32-dimensional ligand atom features and ligand element labels.

The corresponding source-code entry points are:

scripts/process_full_dataset.py
scripts/augment_ligand_features.py
dataset.py::PocketSubgraphExtractor
dataset.py::DTIDataset
scripts/verify_processed_dataset.py

The upstream auxiliary CSV RMSD column is not used as the label source for the processed .pt objects. RMSD values are recomputed from the coordinates and are used for decoy selection, curriculum/ranking weights, and reporting; RMSD is not provided directly as a model node feature.

Processed file structure

Each file is a target-level dictionary serialized with PyTorch. The expected schema is:

<target_id>.pt

protein_graph: torch_geometric.data.Data
  x: [N_protein, 32]       # protein atom features
  pos: [N_protein, 3]      # coordinates in Å

true_ligand_coords: [N_ligand, 3]       # native pose coordinates in Å
ligand_atom_features: [N_ligand, 32]    # ligand atom features
ligand_elements: list[str]

decoys: list[dict]
  coords: [N_ligand, 3]    # decoy coordinates in Å
  rmsd: float              # Kabsch-aligned RMSD

target_id: str

The cached target files contain the protein graph node features and positions, but the final model radius graph is constructed dynamically during model execution. The model pipeline uses a dynamic graph cutoff of approximately 12 Å.

Task semantics

Each target contains:

• one receptor pocket;
• one native ligand pose, treated as the positive pose;
• multiple decoy ligand poses, treated as negative candidates;
• recomputed RMSD values for pose quality and ranking-related procedures.

The dataset is used for native-versus-decoy pose ranking rather than as a pre-expanded binary-classification table.

Download

The processed DuoPose dataset can be downloaded from Hugging Face with:

hf download Kilig33/DuoPose-database processed_full.tar \
  --repo-type dataset \
  --local-dir .

Download the archive checksum with:

hf download Kilig33/DuoPose-database processed_full.tar.sha256 \
  --repo-type dataset \
  --local-dir .

The repository ID is Kilig33/DuoPose-database.

Archive integrity verification

The archive checksum currently recorded for processed_full.tar is:

70d69cb4b8aa74653d0863d43c16ae73389ebee71b7145760e00b17d2b656387  processed_full.tar

Windows PowerShell

$expected = ((Get-Content .\processed_full.tar.sha256).Trim() -split '\s+')[0].ToLower()
$actual = (Get-FileHash -Algorithm SHA256 .\processed_full.tar).Hash.ToLower()

if ($expected -ne $actual) {
    throw "SHA-256 verification failed"
}
Write-Output "SHA-256 verification passed"

Linux or macOS

sha256sum -c processed_full.tar.sha256

The expected result is:

processed_full.tar: OK

Extraction

Windows, Linux, or macOS

tar -xf processed_full.tar

After extraction, the expected directory is:

processed_full/
|-- <target_id>.pt
|-- <target_id>.pt
`-- ...

The archive contains exactly 61,085 .pt files under processed_full/.

Expected project location

After extraction, place or link the directory into the companion source-code repository as:

DuoPose/data/processed_full/

For example:

DuoPose/
|-- data/
|   `-- processed_full/
|       |-- <target_id>.pt
|       `-- ...
|-- dataset.py
|-- scripts/
`-- ...

The code repository provides the authoritative train/validation/test lists and the per-file SHA-256 manifest. After restoring the archive, the processed data can be checked with:

python scripts/verify_processed_dataset.py \
  --processed-dir data/processed_full \
  --manifest checksums/processed_full_files.sha256 \
  --split-dir data

Data integrity audit

The processed directory used to create this archive was audited on August 6, 2026. The audit found:

• 61,085 processed .pt files;
• 0 zero-byte files;
• 0 missing files;
• 0 extra files;
• 0 SHA-256 mismatches against the per-file release manifest;
• disjoint train, validation, and test splits;
• complete split coverage of all 61,085 processed targets;
• no ZIP-container integrity errors in the serialized target files;
• all required target-level key markers present.

The per-file manifest and detailed audit record are maintained in the companion source-code release. The archive-level checksum in this repository must be regenerated whenever processed_full.tar is replaced.

Licensing and citation

The DuoPose processed dataset is released under the Apache License 2.0.

This dataset is derived from DecoyDB. Users of the dataset should cite both the associated DuoPose paper and the original DecoyDB work, and should preserve applicable upstream attribution and license notices.

The processed archive is a project-derived reconstruction and should not be presented as the complete upstream DecoyDB dataset.