minimol-onnx

MiniMol molecular fingerprints, frozen into a portable ONNX graph. The goal is a straightforward install and a single call to get fingerprints, without the dependency conflicts of the original package.

Given the same featurized input, the fingerprints match the original MiniMol to within 6.4e-6 max absolute difference across 300 diverse molecules (1–190 atoms).

Install

pip install minimol-onnx

The 32 MB ONNX graph ships inside the wheel — no download step, no cache directory, works offline.

Use

from minimol_onnx import MinimolONNX

model = MinimolONNX()
fingerprints = model(["c1ccc(-c2cccnc2)cc1", "CCO"])   # (2, 512) float32

model accepts a single SMILES string or any iterable of them, and returns an (n, 512) float32 array. Batching is a performance knob only — results are identical either way.

MinimolONNX(
    model_path=...,             # override the bundled graph
    featurization_kwargs=...,   # dict, or path to JSON
    batch_size=100,
    providers=["CPUExecutionProvider"],   # any onnxruntime provider
)

For GPU inference, pip install minimol-onnx[gpu] and pass providers=["CUDAExecutionProvider"].

Dependencies

onnxruntime numpy scipy pandas networkx loguru rdkit datamol torch torch-geometric, verified against torch 2.13, numpy 2.4, scipy 1.17, pandas 3.0, onnxruntime 1.28. torch and torch-geometric are pulled in by the featurizer, not by inference. Torch runs entirely in onnxruntime but featurization still builds a PyG Data object, so the install is ~2 GB.

Development

The ONNX export pipeline lives in export/ and ships in the sdist. It needs an environment with minimol 1.3.5 and graphium 2.4.7 installed; see export/README.md.

git clone https://github.com/a-pouplin/minimol-onnx
cd minimol-onnx
pip install -e ".[dev]"
pytest

License and attribution

Apache-2.0 AND MIT. This package combines two upstream sources:

• The vendored featurizer in minimol_onnx/minimol_feat/ derives from graphium 2.4.7 (Apache-2.0, Valence Labs / Recursion Pharmaceuticals / Graphcore), modified to drop the compiled extensions.
• The weights in minimol_v1.onnx are exported from minimol 1.3.5 (MIT, Graphcore Ltd.) and are numerically unchanged.

See NOTICE for the full breakdown. If you use this, cite the original work:

@article{klaser2024minimol,
  title   = {{MiniMol}: A Parameter-Efficient Foundation Model for Molecular Learning},
  author  = {Kl{\"a}ser, Kerstin and Banaszewski, B{\l}a{\.z}ej and Maddrell-Mander, Samuel and
             McLean, Callum and M{\"u}ller, Luis and Parviz, Ali and Huang, Shenyang and
             Fitzgibbon, Andrew},
  journal = {arXiv preprint arXiv:2404.14986},
  year    = {2024},
  url     = {https://arxiv.org/abs/2404.14986}
}