[NEWS] Please check our latest work on molecular conformation generation, which has been accepted in ICML'2021 (Long Talk): Learning Gradient Fields for Molecular Conformation Generation. [Code]

[OpenReview] [arXiv] [Code]

This is the official code repository of our ICLR paper "Learning Neural Generative Dynamics for Molecular Conformation Generation" (2021).

Step 1: Create a conda environment named CGCF from env.yml :

conda env create --file env.yml

Step 2: Install PyTorch Geometric :

conda activate CGCF
./install_torch_geometric.sh

# Create conda environment
conda create --name CGCF python=3.7

# Activate the environment
conda activate CGCF

# Install packages
conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch
conda install rdkit==2020.03.3 -c rdkit
conda install tqdm networkx scipy scikit-learn h5py tensorboard -c conda-forge
pip install torchdiffeq==0.0.1

# Install PyTorch Geometric
pip install --no-index torch-scatter -f https://pytorch-geometric.com/whl/torch-1.6.0+cu101.html
pip install --no-index torch-sparse -f https://pytorch-geometric.com/whl/torch-1.6.0+cu101.html
pip install --no-index torch-cluster -f https://pytorch-geometric.com/whl/torch-1.6.0+cu101.html
pip install --no-index torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.6.0+cu101.html
pip install torch-geometric

The official datasets are available here.

The dataset file is a pickled Python list consisting of rdkit.Chem.rdchem.Mol objects. Each conformation is stored individually as a Mol object. For example, if a dataset contains 3 molecules, where the first molecule has 4 conformations, the second one and the third one have 5 and 6 conformations respectively, then the pickled Python list will contain 4+5+6 Mol objects in total.

The output format is identical to the input format.

Example: generating 50 conformations for each molecule in the QM9 test-split with the pre-trained model.

python generate.py --ckpt ./pretrained/ckpt_qm9.pt --dataset ./data/qm9/test.pkl --num_samples 50 --out ./generated.pkl

More generation options can be found in generate.py.

Example: training a model for QM9 molecules.

python train.py --train_dataset ./data/qm9/train.pkl --val_dataset ./data/qm9/val.pkl

More training options can be found in train.py.

Please consider citing our work if you find it helpful.

@inproceedings{
  xu2021learning,
  title={Learning Neural Generative Dynamics for Molecular Conformation Generation},
  author={Minkai Xu* and Shitong Luo* and Yoshua Bengio and Jian Peng and Jian Tang},
  booktitle={International Conference on Learning Representations},
  year={2021},
  url={https://openreview.net/forum?id=pAbm1qfheGk}
}

If you have any question, please contact me at minkai.xu@umontreal.ca or xuminkai@mila.quebec.

• Feb 4, 2021. Code is coming soon.
• Feb 20, 2021. Code is released.