---
license: cc0-1.0
task_categories:
  - other
tags:
  - drug-discovery
  - molecular-generation
  - protein-ligand
  - structural-biology
size_categories:
  - 1M<n<10M
---

# CrossDocked2020

Pre-processed CrossDocked2020 dataset containing raw receptor PDB and ligand
SDF.gz files, organized for efficient loading.

## Dataset Summary

- **Unique pairs**: 2,533,288
- **Unique receptor PDB files**: 24,533
- **Source types**: cdonly, it0, it2, it2_redocked, it2_10p20n, mod_it2,
  it2_completeset, it2_redocked_completeset, mod_it2_completeset
- **Fold splits**: 3 folds (0, 1, 2) per source type category
  (completeset categories only populate fold0)


## Pose Counts

Each ligand `.sdf.gz` contains **multiple docked poses** concatenated as SDF
records. The 5 filename patterns correspond to different gnina iterations
(`tt_min` is the energy-minimized crystal pose; the others are docked poses
from successive scoring iterations). Pose totals across the 2,533,288
uploaded files:

| Filename pattern | Files | Poses | Mean poses/file |
|---|---:|---:|---:|
| `tt_docked` | 594,185 | 11,394,561 | 19.18 |
| `it1_tt_docked` | 585,876 | 11,235,334 | 19.18 |
| `it2_tt_docked` | 415,654 | 8,050,604 | 19.37 |
| `it1_it2_tt_docked` | 339,656 | 6,672,859 | 19.65 |
| `tt_min` | 597,917 | 597,917 | 1.00 |
| **Total** | **2,533,288** | **37,951,275** | **14.98** |

So the dataset holds **37,951,275 poses** distributed across
2,533,288 `.sdf.gz` files (≈ 38.0 M poses).

## Repository Structure

```
receptors/          Unique receptor PDB files in tar.gz archives
ligands/            Ligand SDF.gz files in tar shards (WebDataset-compatible)
manifest.parquet    Pair index with metadata and fold split info
```

## Ligand Tar Shard Format

Each shard is a tar file containing pairs of files per sample:
- `{pair_idx:07d}.sdf.gz` — original ligand SDF.gz (all conformers)
- `{pair_idx:07d}.json` — metadata (receptor_path, complex_dir, source_type)

## Manifest Schema

| Column | Type | Description |
|--------|------|-------------|
| pair_idx | uint32 | Global unique pair ID |
| complex_dir | string | Complex directory name |
| receptor_pdb | string | Receptor PDB filename |
| ligand_sdf_gz | string | Ligand SDF.gz filename |
| source_type | string | One of 9 categories (see Source types above) |
| shard_idx | uint16 | Ligand shard number |
| label | int8 | Types file label (0/1) |
| score1 | float32 | Types file score 1 |
| score2 | float32 | Types file score 2 |
| {cat}_fold{n} | string | "train" / "test" per category and fold |

## Original Source

Paul G. Francoeur, Tomohide Masuda, Jocelyn Sunseri, Andrew Jia,
Richard B. Iovanisci, Ian Snyder, David R. Koes.
*J. Chem. Inf. Model.* 2020, 60(9), p.4200-4215.