Using GNN property predictors as molecule generators (DIDgen)

This is the repository for the paper: Using GNN property predictors as molecule generators.

You can use DIDgen (Direct Inverse Design generator) to generate diverse molecules with a specific property by inverting a GNN that predicts that property.

Install

pip install git+https://github.com/ftherrien/inv-design

Usage

If you have trouble using or installing DIDgen please create an issue or ask a question. I will be happy to help!

As a command-line interface (cli)

didgenerate [-h] [-n N] [-c CONFIG] [-o OUTDIR]

The results are organized in OUTDIR as such

OUTDIR
├── drawings
│   └── generated_mol_0.png     # An image(s) of the generated graph(s)
├── final_performance_data.pkl
├── final_performance.png
├── initial_mol.png
├── model_weights.pth
├── property_value_list.txt     # A list of smiles strings and corresponding predicted property
├── qm9/
└── xyzs/
    ├── generated_mol_0.pickle  # A RDKit mol object
    └── generated_mol_0.xyz     # A molecular conformer with 3D positions

As a Python API

from didgen import generate

out = generate(number_of_samples, outdir, config_dict)

This creates the same output directory as the cli. out is a list of python dictionaries containing the generated graphs, their corresponding smiles and the predicted property.

Parameters

You can find a list of parameters and their description in the documentation.

Generate molecules online using Colab

Train a GNN to predict the energy gap on a subset of QM9 and generate a molecule with an energy gap of 4.1 eV

Citation

@Article{Therrien2025,
author={Therrien, F{\'e}lix and Sargent, Edward H. and Voznyy, Oleksandr},
title={Using GNN property predictors as molecule generators},
journal={Nature Communications},
year={2025},
month={May},
day={08},
volume={16},
number={1},
pages={4301},
issn={2041-1723},
doi={10.1038/s41467-025-59439-1},
url={https://doi.org/10.1038/s41467-025-59439-1}
}