---
license: other
pretty_name: LIT-PCBA Docked Poses (gnina virtual screening)
tags:
  - chemistry
  - molecular-docking
  - virtual-screening
  - drug-discovery
  - gnina
  - lit-pcba
configs:
  - config_name: default
    data_files:
      - split: train
        path: litpcba_gnina_scores.parquet
---

# LIT-PCBA Docked Poses (gnina virtual screening)

The docked poses from `lit-pcba_vs.tgz`, produced by the Koes lab at the
University of Pittsburgh for *Virtual Screening with Gnina 1.0*.

The original is one large archive that must be read start to finish. Here it
is split by target into `lit-pcba-vs/`, one archive per LIT-PCBA target, so
you can download only the targets you need:

```bash
hf download Kingldore/lit-pcba-vs lit-pcba-vs/FEN1.tar.gz --repo-type dataset
tar -xzf lit-pcba-vs/FEN1.tar.gz
```

15 targets, 833 GB in total. The contents are unchanged from the original;
only the packaging differs.

## Cross-docking

Unlike DUD-E, LIT-PCBA targets have several crystal structures (1 to 15 per
target), and every screening compound was docked into *every* structure for
its target, not just one. So a compound has one score per receptor structure
it was docked into, not one score total — the row key below is
`(target, receptor_pdb, compound_id)`.

Each archive also contains a small number of redock files, where a
structure's own co-crystallized ligand is re-docked into every structure for
that target. These are a pose-recovery check, not screening compounds — they
carry no active/decoy label and are left out of the score table below.

## Score table

`litpcba_gnina_scores.parquet` summarises every docked pose into one row per
`(target, receptor_pdb, compound_id)` — 16,672,528 rows, 74,950 actives and
16,597,578 inactives.

| column | meaning |
|---|---|
| `protein_id` | LIT-PCBA target name |
| `receptor_pdb` | archive path to the receptor structure this row was docked against |
| `compound_id` | PubChem CID |
| `label` | 1 = active, 0 = inactive |
| `n_poses` | poses the statistics were computed over |
| `gnina_best_*` | value at the pose with the highest CNNscore |
| `gnina_mean_*`, `gnina_std_*` | mean and population standard deviation over all poses |
| `docked_pose_sdf_path` | archive path to the SDF containing this compound's poses against this receptor |

The three score families are `minimized_affinity`, `cnn_score`, and
`cnn_affinity`. `minimized_affinity` is the empirical docking score, from the
column gnina's summary files label `Vina`.

All three `gnina_best_*` values come from the **same** pose — the one with
the highest CNNscore — rather than each column being maximised independently.

`gnina_std_*` is null for compounds with a single pose. One row (IDH1) has a
null `gnina_best_cnn_score`/`gnina_best_cnn_affinity`: gnina's own scoring
output was missing those two values for that pose.

**Original download:** <https://bits.csb.pitt.edu/files/gninavs/lit-pcba_vs.tgz>

## Citation

Please cite both the method and the underlying benchmark.

**gnina virtual screening** — the source of these docked poses:

```bibtex
@article{sunseri2021virtual,
  title   = {Virtual Screening with Gnina 1.0},
  author  = {Sunseri, Jocelyn and Koes, David Ryan},
  journal = {Molecules},
  volume  = {26},
  number  = {23},
  pages   = {7369},
  year    = {2021},
  doi     = {10.3390/molecules26237369}
}
```

**LIT-PCBA** — the benchmark set that was docked:

```bibtex
@article{tran2020lit,
  title   = {LIT-PCBA: An Unbiased Data Set for Machine Learning and Virtual
             Screening},
  author  = {Tran-Nguyen, Viet-Khoa and Jacquemard, C{\'e}line and Rognan, Didier},
  journal = {Journal of Chemical Information and Modeling},
  volume  = {60},
  number  = {9},
  pages   = {4263--4273},
  year    = {2020},
  doi     = {10.1021/acs.jcim.0c00155}
}
```

## Links

- Paper: <https://doi.org/10.3390/molecules26237369>
- LIT-PCBA paper: <https://doi.org/10.1021/acs.jcim.0c00155>
- gnina software: <https://github.com/gnina/gnina>
- LIT-PCBA: <https://drugdesign.unistra.fr/LIT-PCBA/>
- Koes lab file index: <https://bits.csb.pitt.edu/files/>

## License

This is a mirror hosted for convenience; the data belongs to its original
authors and no additional licence is granted here.