MolFLAE is a codebase for the paper "Manipulating 3D Molecules in a Fixed-Dimensional E(3)-Equivariant Latent Space" (NeurIPS 2025).
This release provides code and scripts for latent-space experiments under Latent_Experiments/.
Training code and support for GEOM-Drugs and QM9 datasets will be released in a follow-up update.
- Latent-space experiments:
Latent_Experiments/latent_experiment - Model implementation:
Latent_Experiments/model - Instructions for running latent-space experiments:
Latent_Experiments/README.md - Environment spec:
environment.yaml
Create a conda environment (recommended):
conda env create -f environment.yaml
conda activate MolFLAENote: RDKit and OpenBabel are used in the repo and are included in the environment file; ensure they are available. Schrodinger utilities (used for some molecule repairs) are not included and are optional.
For questions, please open an issue on this repository or contact the authors listed in the NeurIPS paper.
If you use this work, please cite:
@inproceedings{chen2025manipulating,
title={Manipulating 3D Molecules in a Fixed-Dimensional E(3)-Equivariant Latent Space},
author={Zitao Chen and Yinjun Jia and Zitong Tian and Wei-Ying Ma and Yanyan Lan},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=6xL4MRFeJo}
}