Awesome

Energy-Inspired Molecular Conformation Optimization

This repository is the official implementation of Energy-Inspired Molecular Conformation Optimization (ICLR 2022). [PDF]

Installation

Dependency

The code has been tested in the following environment:

Install via Conda and Pip

conda create -n confopt python=3.7
conda install rdkit -c rdkit
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.9.0+cu111.html
pip install dgl-cu111 tensorboard easydict ase
pip install lie_learn # for SE(3)-Transformer baseline

You may need to install pytorch, pytorch-geometric and dgl by yourself to make it compatible with your CUDA version.

Data and Pre-trained Models

The processed data and the pre-trained models can be found in this folder. You can also download them via the following command:

# for dataset (10.79 GB)
wget -O data.zip 'https://www.dropbox.com/sh/bcidyj2mbgy5dp2/AAB_lXSjadWI1wUk6WZgLEBGa?dl=1' 
unzip data.zip -d data

# for pre-trained models (1.49 GB)
wget -O pretrained_models.zip 'https://www.dropbox.com/sh/cq6ho0imyynkfpg/AACq0GW_auRdLXAIQicnG56wa?dl=1' 
unzip pretrained_models.zip -d pretrained_models

If you meet difficulties in downloading the data / pretrained models, please try this Google Drive link.

Training

Conformation Optimization

To train a conformer optimization model:

python train_conf.py --config configs/{qm9, drug}_default.yml --model_type {equi_se3trans, egnn, ours_o2, ours_o3}

Conformation Generation

To train a conformer generation model:

• with random initial conformers:

python train_sampling.py --config configs/{qm9, drug}_sampling.yml --model_type ours_o2 --propose_net_type random --noise_std 0.028 --eval_propose_net_type random --eval_noise 0.028

• with RDKit initial conformers:

python train_sampling.py --config configs/{qm9, drug}_sampling.yml --model_type ours_o2 --propose_net_type gt --noise_std {0.5, 1.0} --eval_propose_net_type online_rdkit --eval_noise 0.

Property Prediction

To train a property prediction model:

python train_prop.py --config configs/qm9_prop_default.yml --model_type ours_o2 --pos_type {gt, rdkit, ours} --target_name homo

Evaluation

Conformation Optimization Evaluation

To evaluate the conformer optimization model:

python eval_conf.py --ckpt_path <model_path> --test_dataset <dataset_path>

For example, with data and pretrained model prepared (see above), you can run the following command to evaluate our two-atom model on the QM9 dataset:

python eval_conf.py --ckpt_path pretrained_models/conf_opt/qm9_our_o2 --test_dataset data/qm9/qm9_test.pkl

One can also dump optimized conformers and reproduce the results with error bars reported in the paper with the following command. Dumped conformers can also be used in downstream tasks like molecular property prediction.

python dump_confs.py \
  --test_dataset data/qm9/qm9_test.pkl \
  --ckpt_path_list pretrained_models/conf_opt/qm9_our_o2 \
  --dump_dir dump_confopt_results \
  --filter_pos False --rdkit_pos_mode all

Conformation Generation Evaluation

To evaluate the conformer generation model:

python eval_sampling.py --ckpt_path <model_path> --eval_propose_net_type random --eval_noise 0.028

or:

python eval_sampling.py --ckpt_path <model_path> --eval_propose_net_type online_rdkit --eval_noise 0.

Property Prediction Evaluation

To evaluate the property prediction model:

python eval_prop.py --ckpt_path pretrained_models/prop_pred_with_gt/qm9_homo

Results

Conformation Optimization Results

The conformation optimization performance of baseline models and our models on the QM9 and GEOM-Drugs datasets:

Conformation Generation Results

The conformation generation performance of our models on the GEOM-QM9 and GEOM-Drugs datasets:

Citation

@inproceedings{guan2022energy,
  title={Energy-Inspired Molecular Conformation Optimization},
  author={Guan, Jiaqi and Qian, Wesley Wei and Liu, Qiang and Ma, Wei-Ying and Ma, Jianzhu and Peng, Jian},
  booktitle={International Conference on Learning Representations},
  year={2022}
}