The official implementation of the paper REBIND (ICLR 2025), By Taewon Kim, Hyunjin Seo, Sungsoo Ahn, and Eunho Yang (*: equal contribution)
We present a novel force-based edge-rewiring framework for ground-state molecular conformation prediction, mitigating the limitation of prior works which have inadequately addressed non-bonded interactions
- python == 3.10.16
- pytorch == 2.1.0+cu121
- torch-geometric == 2.3.0
- rdkit == 2024.3.3
For the rest of the dependencies, please refer to the requirements.txt.
The original official datasets can be found at Molecule3D and Qm9.
For your conveniecne, we've added a run.py file. The details for each steps are specified below.
bash experiments/conformer_prediction/rebind.sh [device number]python -m evaluate \
--data_dir datasets/ \
--dataset QM9 \
--mode random \
--split test \
--log_file logs/[log file name] \
--REBIND_checkpoint [trained ckpt] \
--device cuda:{args.device} \
--removeHsWe provide model weights for QM9: REBIND_QM9
Please cite our paper if you find our work useful for your research:
@article{kim2024rebind,
title={REBIND: Enhancing ground-state molecular conformation via force-based graph rewiring},
author={Kim, Taewon and Seo, Hyunjin and Ahn, Sungsoo and Yang, Eunho},
journal={arXiv preprint arXiv:2410.14696},
year={2024}
}Our codebase is based upon the works of GTMGC. We thank the authors for releasing their code.