ChemBFN: Bayesian Flow Network for Chemistry

This is the repository of the PyTorch implementation of ChemBFN model.

Build State

Features

ChemBFN provides the state-of-the-art functionalities of

• SMILES or SELFIES-based de novo molecule generation
• Protein sequence de novo generation
• Template optimisation (mol2mol)
• Classifier-free guidance conditional generation (single or multi-objective optimisation)
• Context-guided conditional generation (inpaint)
• Outstanding out-of-distribution chemical space sampling
• Fast sampling via ODE solver
• Molecular property and activity prediction finetuning
• Reaction yield prediction finetuning

in an all-in-one-model style. A built-in MLFF can further convert the generated SMILES strings to the corresponding 3D conformations.

News

• [26/12/2025] We were invited to submit a short report about ChemBFN for CICSJ Bulletin.
• [09/10/2025] A web app chembfn_webui for hosting ChemBFN models is available on PyPI.
• [30/01/2025] The package bayesianflow_for_chem is available on PyPI.
• [21/01/2025] Our first paper has been accepted by JCIM.
• [17/12/2024] The second paper of out-of-distribution generation is available on arxiv.org.
• [31/07/2024] Paper is available on arxiv.org.
• [21/07/2024] Paper was submitted to arXiv.

Install

$ pip install -U bayesianflow_for_chem

Usage

You can find example scripts in 📁example folder.

Pre-trained Model

You can find pretrained models (linked to pretraining datasets) on our 🤗Hugging Face model page.

Dataset Handling

We provide a Python class CSVData to handle data stored in CSV or similar format containing headers to identify the entities. The following is a quickstart.

1. Download your dataset file (e.g., ESOL from MoleculeNet) and split the file:

>>> from bayesianflow_for_chem.tool import split_data

>>> split_data("delaney-processed.csv", method="scaffold")

2. Load the split data:

>>> from bayesianflow_for_chem.data import smiles2token, collate, CSVData

>>> dataset = CSVData("delaney-processed_train.csv")
>>> dataset[0]
{'Compound ID': ['Thiophene'], 
'ESOL predicted log solubility in mols per litre': ['-2.2319999999999998'], 
'Minimum Degree': ['2'], 
'Molecular Weight': ['84.14299999999999'], 
'Number of H-Bond Donors': ['0'], 
'Number of Rings': ['1'], 
'Number of Rotatable Bonds': ['0'], 
'Polar Surface Area': ['0.0'], 
'measured log solubility in mols per litre': ['-1.33'], 
'smiles': ['c1ccsc1']}

3. Create a mapping function to tokenise the dataset and select values:

>>> import torch

>>> def encode(x):
...   smiles = x["smiles"][0]
...   value = [float(i) for i in x["measured log solubility in mols per litre"]]
...   return {"token": smiles2token(smiles), "value": torch.tensor(value)}

>>> dataset.map(encode)
>>> dataset[0]
{'token': tensor([  1, 151,  23, 151, 151, 154, 151,  23,   2]), 
'value': tensor([-1.3300])}

4. Wrap the dataset in torch.utils.data.DataLoader:

>>> dataloader = torch.utils.data.DataLoader(dataset, 32, collate_fn=collate)

Cite This Work

@article{2025chembfn,
    title={Bayesian Flow Network Framework for Chemistry Tasks},
    author={Tao, Nianze and Abe, Minori},
    journal={Journal of Chemical Information and Modeling},
    volume={65},
    number={3},
    pages={1178-1187},
    year={2025},
    doi={10.1021/acs.jcim.4c01792},
}

@article{2025chembfn_report,
    title={Molecular Structure Design via Bayesian Flow Network},
    author={Tao, Nianze and Nagai, Touma and Abe, Minori},
    journal={CICSJ Bulletin},
    volume={43},
    number={1},
    pages={10-14},
    year={2025},
    doi={10.11546/cicsj.43.10},
}

Out-of-distribution generation and fast sampling:

@misc{2024chembfn_ood,
    title={Bayesian Flow Is All You Need to Sample Out-of-Distribution Chemical Spaces}, 
    author={Tao, Nianze and Abe, Minori},
    year={2024},
    eprint={2412.11439},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    url={https://arxiv.org/abs/2412.11439}, 
}