The official implementation of Learning Gradient Fields for Molecular Conformation Generation (ICML 2021 Long talk)
# Clone the environment
conda env create -f env.yml
# Activate the environment
conda activate confgf
# Install Library
git clone https://github.com/DeepGraphLearning/ConfGF.git
cd ConfGF
python setup.py install
# Create conda environment
conda create -n confgf python=3.7
# Activate the environment
conda activate confgf
# Install packages
conda install -y -c pytorch pytorch=1.7.0 torchvision torchaudio cudatoolkit=10.2
conda install -y -c rdkit rdkit==2020.03.2.0
conda install -y scikit-learn pandas decorator ipython networkx tqdm matplotlib
conda install -y -c conda-forge easydict
pip install pyyaml
# Install PyTorch Geometric
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cu102.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-1.7.0+cu102.html
pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-1.7.0+cu102.html
pip install torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.7.0+cu102.html
pip install torch-geometric==1.6.3
# Install Library
git clone https://github.com/DeepGraphLearning/ConfGF.git
cd ConfGF
python setup.py install
The offical raw GEOM dataset is avaiable [here].
We provide the preprocessed datasets (GEOM, ISO17) in a [google drive folder]. For ISO17 dataset, we use the default split of [GraphDG].
Download the raw GEOM dataset and unpack it.
tar xvf ~/rdkit_folder.tar.gz -C ~/GEOM
Preprocess the raw GEOM dataset.
python script/process_GEOM_dataset.py --base_path GEOM --dataset_name qm9 --confmin 50 --confmax 500
python script/process_GEOM_dataset.py --base_path GEOM --dataset_name drugs --confmin 50 --confmax 100
The final folder structure will look like this:
GEOM
|___rdkit_folder # raw dataset
| |___qm9 # raw qm9 dataset
| |___drugs # raw drugs dataset
| |___summary_drugs.json
| |___summary_qm9.json
|
|___qm9_processed
| |___train_data_40k.pkl
| |___val_data_5k.pkl
| |___test_data_200.pkl
|
|___drugs_processed
| |___train_data_39k.pkl
| |___val_data_5k.pkl
| |___test_data_200.pkl
|
iso17_processed
|___iso17_split-0_train_processed.pkl
|___iso17_split-0_test_processed.pkl
|
...
All hyper-parameters and training details are provided in config files (./config/*.yml
), and free feel to tune these parameters.
You can train the model with the following commands:
python -u script/train.py --config_path ./config/qm9_default.yml
python -u script/train.py --config_path ./config/drugs_default.yml
python -u script/train.py --config_path ./config/iso17_default.yml
The checkpoint of the models will be saved into a directory specified in config files.
We provide the checkpoints of three trained models, i.e., qm9_default
, drugs_default
and iso17_default
in a [google drive folder].
You can generate conformations of a molecule by feeding its SMILES into the model:
python -u script/gen.py --config_path ./config/qm9_default.yml --generator ConfGF --smiles c1ccccc1
python -u script/gen.py --config_path ./config/qm9_default.yml --generator ConfGFDist --smiles c1ccccc1
Here we use the models trained on GEOM-QM9
to generate conformations for the benzene. The argument --generator
indicates the type of the generator, i.e., ConfGF
vs. ConfGFDist
. See the ablation study (Table 5) in the original paper for more details.
You can also generate conformations for an entire test set.
python -u script/gen.py --config_path ./config/qm9_default.yml --generator ConfGF \
--start 0 --end 200 \
python -u script/gen.py --config_path ./config/qm9_default.yml --generator ConfGFDist \
--start 0 --end 200 \
python -u script/gen.py --config_path ./config/drugs_default.yml --generator ConfGF \
--start 0 --end 200 \
python -u script/gen.py --config_path ./config/drugs_default.yml --generator ConfGFDist \
--start 0 --end 200 \
Here start
and end
indicate the range of the test set that we want to use. All hyper-parameters related to generation can be set in config files.
Conformations of some drug-like molecules generated by ConfGF are provided below.
The results of all benchmark tasks can be calculated based on generated conformations.
We report the results of each task in the following tables. Results of ConfGF
and ConfGFDist
are re-evaluated based on the current code base, which successfully reproduce the results reported in the original paper. Results of other models are taken directly from the original paper.
The COV and MAT scores on the GEOM datasets can be calculated using the following commands:
python -u script/get_task1_results.py --input dir_of_QM9_samples --core 10 --threshold 0.5
python -u script/get_task1_results.py --input dir_of_Drugs_samples --core 10 --threshold 1.25
Table: COV and MAT scores on GEOM-QM9
QM9 | COV-Mean (%) | COV-Median (%) | MAT-Mean (\AA) | MAT-Median (\AA) |
---|---|---|---|---|
ConfGF | 91.06 | 95.76 | 0.2649 | 0.2668 |
ConfGFDist | 85.37 | 88.59 | 0.3435 | 0.3548 |
CGCF | 78.05 | 82.48 | 0.4219 | 0.3900 |
GraphDG | 73.33 | 84.21 | 0.4245 | 0.3973 |
CVGAE | 0.09 | 0.00 | 1.6713 | 1.6088 |
RDKit | 83.26 | 90.78 | 0.3447 | 0.2935 |
Table: COV and MAT scores on GEOM-Drugs
Drugs | COV-Mean (%) | COV-Median (%) | MAT-Mean (\AA) | MAT-Median (\AA) |
---|---|---|---|---|
ConfGF | 62.54 | 71.32 | 1.1637 | 1.1617 |
ConfGFDist | 49.96 | 48.12 | 1.2845 | 1.2827 |
CGCF | 53.96 | 57.06 | 1.2487 | 1.2247 |
GraphDG | 8.27 | 0.00 | 1.9722 | 1.9845 |
CVGAE | 0.00 | 0.00 | 3.0702 | 2.9937 |
RDKit | 60.91 | 65.70 | 1.2026 | 1.1252 |
The MMD metrics on the ISO17 dataset can be calculated using the following commands:
python -u script/get_task2_results.py --input dir_of_ISO17_samples
Table: Distributions over distances
Method | Single-Mean | Single-Median | Pair-Mean | Pair-Median | All-Mean | All-Median |
---|---|---|---|---|---|---|
ConfGF | 0.3430 | 0.2473 | 0.4195 | 0.3081 | 0.5432 | 0.3868 |
ConfGFDist | 0.3348 | 0.2011 | 0.4080 | 0.2658 | 0.5821 | 0.3974 |
CGCF | 0.4490 | 0.1786 | 0.5509 | 0.2734 | 0.8703 | 0.4447 |
GraphDG | 0.7645 | 0.2346 | 0.8920 | 0.3287 | 1.1949 | 0.5485 |
CVGAE | 4.1789 | 4.1762 | 4.9184 | 5.1856 | 5.9747 | 5.9928 |
RDKit | 3.4513 | 3.1602 | 3.8452 | 3.6287 | 4.0866 | 3.7519 |
pymol -R
Display - Background - White
Display - Color Space - CMYK
Display - Quality - Maximal Quality
Display Grid
- by object: use
set grid_slot, int, mol_name
to put the molecule into the corresponding slot - by state: align all conformations in a single slot
- by object-state: align all conformations and put them in separate slots. (
grid_slot
dont work!)
- by object: use
Setting - Line and Sticks - Ball and Stick on - Ball and Stick ratio: 1.5
Setting - Line and Sticks - Stick radius: 0.2 - Stick Hydrogen Scale: 1.0
-
To show molecules
hide everything
show sticks
-
To align molecules:
align name1, name2
-
Convert RDKit mol to Pymol
from rdkit.Chem import PyMol v= PyMol.MolViewer() rdmol = Chem.MolFromSmiles('C') v.ShowMol(rdmol, name='mol') v.SaveFile('mol.pkl')
- load a sequence of pymol objects named
traj*.pkl
into the PyMol, wheretraji.pkl
is thei-th
conformation in the trajectory. - Join states:
join_states mol, traj*, 0
- Delete useless object:
delete traj*
Movie - Program - State Loop - Full Speed
- Export the movie to a sequence of PNG files:
File - Export Movie As - PNG Images
- Use photoshop to convert the PNG sequence to a GIF with the transparent background.
Please consider citing the following paper if you find our codes helpful. Thank you!
@inproceedings{shi*2021confgf,
title={Learning Gradient Fields for Molecular Conformation Generation},
author={Shi, Chence and Luo, Shitong and Xu, Minkai and Tang, Jian},
booktitle={International Conference on Machine Learning},
year={2021}
}
Chence Shi ([email protected])