---
license: cc-by-4.0
task_categories:
- fill-mask
language:
- en
tags:
- chemistry
- molecules
- selfies
- chembl
- pre-training
---

# ChEMBL 36 SELFIES

Pre-training dataset for [ModernMolBERT](https://github.com/HauserGroup/ModernMolBERT). Contains ~2.4M drug-like small molecules from ChEMBL 36 represented as [SELFIES](https://github.com/aspuru-guzik-group/selfies) strings.

## Dataset details

| field | value |
|-------|-------|
| source | [lukaskim/ChEMBL-36](https://huggingface.co/datasets/lukaskim/ChEMBL-36) |
| representation | SELFIES |
| train rows | 2,390,314 |
| validation rows | 24,228 |
| total rows | 2,414,542 |
| min heavy atoms | 3 |
| max heavy atoms | 100 |
| max MW | 1000.0 |
| deduplicated by | InChIKey |
| split method | deterministic hash on InChIKey |
| valid fraction | 0.01 |

## Preparation stats

- Input rows: 2,854,815
- After deduplication: 2,854,815
- Valid SELFIES conversions: 2,854,762
- After physicochemical filters: 2,414,542
- Dropped (invalid or filtered): 440,273

## Columns

| column | description |
|--------|-------------|
| `selfies` | SELFIES string (primary pre-training input) |
| `canonical_smiles` | original ChEMBL SMILES |
| `smiles_canonical_clean` | RDKit-canonicalized SMILES |
| `standard_inchi_key` | InChIKey used for deduplication and splitting |
| `chembl_id` | ChEMBL compound identifier |
| `qed_weighted` | QED drug-likeness score |
| `heavy_atoms`, `mw_freebase`, `alogp`, ... | Physicochemical descriptors from ChEMBL |

## Usage

```python
from datasets import load_dataset

ds = load_dataset('HauserGroup/ChEMBL36-SELFIES')
print(ds['train'][0]['selfies'])
```

## Versions

- datasets: 4.8.5
- rdkit: 2026.03.1
- selfies: 2.1.1
- pandas: 3.0.3
- pyarrow: 24.0.0

## License

ChEMBL data is released under [CC BY-SA 3.0](https://creativecommons.org/licenses/by-sa/3.0/). The prepared dataset files in this repository are released under CC BY 4.0.