README.md

---
license: mit
library_name: deepbioisostere
tags:
  - chemistry
  - drug-discovery
  - molecular-optimization
  - bioisostere
  - graph-neural-network
---

# DeepBioisostere

Model weights and fragment library for **DeepBioisostere**, a deep generative
model that performs end-to-end bioisosteric replacement: it selects which
fragment of a molecule to replace, what to replace it with, and how to attach
it, so as to satisfy several property targets at once.

- Paper: [Nature Communications (2026)](https://doi.org/10.1038/s41467-026-75512-9)
- Code: https://github.com/Hwoo-Kim/DeepBioisostere
- Project page: https://hwoo-kim.github.io/DeepBioisostere/

## Usage

These files are fetched automatically; you do not normally download them by hand.

```bash
pip install deepbioisostere
deepbioisostere generate -s "Cc1ccc2cnc(N(C)CCc3ccccn3)nc2c1" -t mw=0 -t logp=-1
```

```python
from deepbioisostere import Conditioner, DeepBioisostere, Generator

properties = ["logp", "mw"]
model = DeepBioisostere.from_pretrained(properties=properties)
generator = Generator(
    model=model,
    conditioner=Conditioner(phase="generation", properties=properties),
    properties=properties,
)
df = generator.generate([("Cc1ccc2cnc(N(C)CCc3ccccn3)nc2c1", {"mw": 0, "logp": -1})])
```

## Contents

### Checkpoints

Each checkpoint is conditioned on a specific set of properties. The filename
encodes the property set, sorted alphabetically.

| File | Properties | Variant |
|---|---|---|
| `DeepBioisostere_logp.pt` | logP | standard |
| `DeepBioisostere_mw.pt` | MW | standard |
| `DeepBioisostere_qed.pt` | QED | standard |
| `DeepBioisostere_sa.pt` | SA | standard |
| `DeepBioisostere_logp_mw.pt` | logP + MW | standard |
| `DeepBioisostere_mw_qed.pt` | MW + QED | standard |
| `DeepBioisostere_qed_sa.pt` | QED + SA | standard |
| `DeepBioisostere_logp_mw_ablation.pt` | logP + MW | ablation |
| `DeepBioisostere_mw_qed_ablation.pt` | MW + QED | ablation |
| `DeepBioisostere_qed_sa_ablation.pt` | QED + SA | ablation |

The remaining pairs (logP+QED, logP+SA, MW+SA) were not trained. Ablation
variants set `use_subgraph_AMPN=False`, which changes the architecture, and are
selected with `from_pretrained(..., ablation=True)`.

### Fragment library

| File | Description |
|---|---|
| `fragment_library.csv` | 145,854 insertion fragments with frequency, BRICS type and train/val/test split |
| `frag_features.pkl` | Pre-parsed graph features, keyed by the csv SMILES |
| `frag_brics_maskings.pkl` | BRICS-type compatibility masks (training only) |

The two `.pkl` files are derived from the csv and are regenerated automatically
if absent, which takes roughly an hour of CPU. They are published so that a
first run is a download instead.

> **Note on fragment identity.** The features are keyed on the exact SMILES
> string in `fragment_library.csv`, not on a re-canonicalised form. RDKit's
> canonical SMILES writer changed after the library was built under 2022.03:
> under 2026.03, 1,712 of the fragments are written differently (stereocentre
> descriptors flip). The underlying molecules and their parsed features are
> unchanged — this is a serialisation difference only — but the csv string is
> the authoritative key, since training data references fragments by row index.

## Citation

```bibtex
@article{kim2026deepbioisostere,
  title   = {Autonomous bioisosteric replacement for multi-property optimization in drug design},
  author  = {Kim, Hyeongwoo and Moon, Seokhyun and Zhung, Wonho and
             Kim, Shinwoo and Lim, Jaechang and Kim, Woo Youn},
  journal = {Nature Communications},
  year    = {2026},
  doi     = {10.1038/s41467-026-75512-9},
}
```

License: MIT.