---
license: other
license_name: bigbind
tags:
- molecular-docking
- gnina
- drug-discovery
- structure-based-virtual-screening
pretty_name: BigBind GNINA-Docked Poses
size_categories:
- 1M<n<10M
configs:
- config_name: default
  data_files:
  - split: train
    path: bigbind_gnina_scores.parquet
---

# BigBind GNINA-Docked Poses

GNINA redocking of every protein-ligand pair in BigBind (train/val/test, including SNA
decoy variants). Each ligand is docked into its BigBind pocket and rescored with
GNINA's CNN scoring function.

## Files

| File | Contents |
|---|---|
| `results.tar.gz` | All docked poses, one folder per protein target: `results/{protein_id}/{ligand}.sdf` |
| `bigbind-docked-vs/{protein_id}.tar.gz` | Same poses, re-split per target so a single protein can be fetched without downloading the whole archive |
| `bigbind_gnina_scores.parquet` | One row per ligand, with pose-derived score columns (below) |

## Scores schema

| Column | Description |
|---|---|
| `protein_id` | BigBind pocket identifier |
| `ligand_smiles` | Ligand SMILES |
| `label` | 1 = active, 0 = inactive/decoy (from BigBind) |
| `n_poses` | Number of poses GNINA returned |
| `gnina_best_*` | Score at the pose GNINA's CNN rates most likely correct (`minimized_affinity`, `cnn_score`, `cnn_affinity`, `cnn_vs`, `cnn_affinity_var`) |
| `gnina_mean_*`, `gnina_std_*` | Mean / population std of each score across all poses |
| `gnina_docked_status` | Whether GNINA returned any pose for this ligand |
| `docked_pose_sdf_path` | Pose path relative to `results/`, e.g. `1433S_HUMAN_1_233_0/mol_4699.sdf` |
| `compressed_docked_pose_sdf_path` | Where to find that same pose inside `bigbind-docked-vs/`, e.g. `bigbind-docked-vs/1433S_HUMAN_1_233_0.tar.gz/results/1433S_HUMAN_1_233_0/mol_4699.sdf` |

Column layout follows the [Kingldore/dude-vs](https://huggingface.co/datasets/Kingldore/dude-vs) schema.

## Docking parameters

Docked with GNINA v1.3.2 (`master:f23dd2b`), run with `--cnn_scoring rescore --exhaustiveness 8 --num_modes 9 --seed 42`.

## License

Derived from BigBind; usage is subject to BigBind's original license terms.

## Citation

Brocidiacono, M., Francoeur, P., Aggarwal, R., Popov, K. I., & Koes, D. R. (2024).
BigBind: Learning from Nonstructural Data for Structure-Based Virtual Screening.
*Journal of Chemical Information and Modeling*. https://doi.org/10.1021/acs.jcim.3c01211