RotNet: A Rotationally Invariant Graph Neural Network for Quantum Mechanical Calculations

Please cite

@article{tu2023rotnet,
  title={RotNet: A Rotationally Invariant Graph Neural Network for Quantum Mechanical Calculations},
  author={Tu, Hongwei and Han, Yanqiang and Wang, Zhilong and Chen, An and Tao, Kehao and Ye, Simin and Wang, Shiwei and Wei, Zhiyun and Li, Jinjin},
  journal={Small Methods},
  pages={2300534},
  year={2023},
  publisher={Wiley Online Library}
}

Updates

More recent updates are available from GitHub.

• [Sep 15, 2023] We revised RotNet and added training tricks, which eliminate the instability issue reported in the paper.

Datasets

The two datasets used in the experiments can be found at:

• Protein fragments
  • Google Drive
  • 百度网盘（提取码: jymi）
• Electronic materials
  • Google Drive
  • 百度网盘（提取码：y21t）

Please put the downloaded dataset(s) in ./data/raw.

File Hierarchy

• ./data: Contains datasets used by RotNet.
  • ./data/raw: (Needs to be downloaded beforehand. See above.) Contains our two datasets.
  • ./data/atom_init_embedding.json: Initial embeddings for atoms.
• ./exp: (Will be created automatically on first run.) Contains training logs and saved model weights.
• ./dataset.py: Dataset class definition.
• ./evaluate.py: Evaluates a model. See ./eval.sh for usage.
• ./loss_weights.py: Weights for loss terms.
• ./models.py: Model definition.
• ./train.py: Trains RotNet. See ./run.sh for usage.
• ./transforms.py: Implements the rotationally invariant transformation (RIT).
• ./utils.py: Helper functions or classes.
• ./eval.sh: Usage of ./evaluate.sh.
• ./run.sh: Usage of ./train.py.
• ./requirements.txt: Package specification for environment setup.

Usage

1. Clone this repository.

2. Set up an environment as per ./requirements.txt.

3. Download our datasets and put them in ./data/raw. You may use custom datasets that have been formatted like ours.

4. Process raw data. To process protein fragments, for example, run

   python dataset.py --dataset_name protein_fragments --cutoff 4

   If you want to overwrite previously processed datasets, run

   python dataset.py --dataset_name protein_fragments --cutoff 5 --remake

5. (Optional) Modify weights for loss terms in ./loss_weights.py. You will need to modify them if another dataset is in use.

6. Run ./run.sh to train RotNet. You may modify arguments passed to ./train.py.

7. Run ./eval.sh to evaluate models. Arguments passed to ./evaluate.py must be consistent with those passed to ./train.py.

License

Distributed under the Apache 2.0 License. See LICENSE for more information.

Acknowledgements

This work was supported by the National Key R&D Program of China [No. 2021YFC2100100]; the National Natural Science Foundation of China [No. 21901157] and the Shanghai Science and Technology Project [No. 21JC1403400].