MolDiff: Addressing the Atom-Bond Inconsistency Problem in 3D Molecule Diffusion Generation (ICML 2023)
This is A Tailored Diffusion Framework for Generating 3D Drug-Like Molecules, with sampling success rate of >99%, almost threefold increase compared to the previous diffusion model.
This can serve as a better backbone for other applications of 3D molecule diffusion models such as pocket-based generation and linker generation.
More information can be found in our paper.
Update Jul 23, 2024
Add the trained MolDiff checkpoint on the QM9 dataset and the corresponding sampling configuration file. See theckpt
directory.
The codes have been tested in the following environment:
Package | Version |
---|---|
Python | 3.8.13 |
PyTorch | 1.10.1 |
CUDA | 11.3.1 |
PyTorch Geometric | 2.0.4 |
RDKit | 2022.03.2 |
conda env create -f env.yaml
conda activate MolDiff
conda create -n MolDiff python=3.8 # optinal, create a new environment
conda activate MolDiff
# Install PyTorch (for cuda 11.3)
conda install pytorch==1.10.1 cudatoolkit=11.3 -c pytorch -c conda-forge
conda install pyg -c pyg
# Install other tools
conda install -c conda-forge rdkit
conda install pyyaml easydict python-lmdb -c conda-forge
# Install tensorboard only for training
conda install tensorboard -c conda-forge
You can download the processed data geom_drug.tar.gz
(~2GB) from here and unzip them (tar -zxvf
)
in the data
folder as:
data
├── geom_drug
│ ├── mol_summary.csv
│ ├── split_by_molid.pt
│ ├── processed.lmdb
│ └── processed_molid2idx.pt
If you want to process the data from the sdf files, you have to further download the sdf.tar.gz
(~4GB) from here. Unzip it (~28GB after unzipping) in the data/geom_drug
folder and remove the processed.lmdb
and processed_molid2idx.pt
:
data
├── geom_drug
│ ├── mol_summary.csv
│ ├── split_by_molid.pt
│ └── sdf
│ ├── 0.sdf
│ ├── 1.sdf
│ ├── ...
Then by running any sampling, training or evaluation script, the data will be processed automatically.
We provide the sampling config file sample_MolDiff.yml
in configs/sample
folder. We also provide a simplified version of MolDiff sample_MolDiff_simple.yml
that does not use the bond guidance for sampling and uses the model trained without the new bond noise schedule proposed in the paper.
To sample molecules using pretrained models, please first download the pretrained model weights from here and put them in the ./ckpt
folder. There are three model weight files:
MolDiff.pt
: the pretrained complete MolDiff model.MolDiff_simple.pt
: the pretrained simplified MolDiff model that was trained without using the new bond noise schedule.bond_predictor.pt
: the pretrained bond predictor that is used for bond guidance during sampling.
After preparing the pretrained weights (either the downloaded files or trained by yourself) and setting the correct model weight paths in the config file, you can run the following command to sample molecules:
python scripts/sample_drug3d.py --outdir <output_directory> --config <path_to_config_file> --device <device_id> --batch_size <batch_size>
The parameters are:
outdir
: the root directory to save the sampled molecules.config
: the path to the config file.device
: the device to run the sampling.batch_size
: the batch size for sampling. If set to 0 (default), it will use the batch size specified in the config file.
An example command is:
python scripts/sample_drug3d.py --outdir ./outputs --config ./configs/sample/sample_MolDiff.yml
After sampling, there will be two directories in the outdir
folder that contains the meta data and the sdf files of the sampling, respectively.
To evaluate the generated molecules, run the following command:
python scripts/evaluate_all.py --result_root <result_root> --exp_name <exp_name> --from_where generated
The parameters are:
result_root
: the parent directory of the directory of the sampled molecules (i.e, the same as theoutdir
parameter when runningsample_drug3d.py
).exp_name
: the name (or prefix) of the directory of the molecules (excluding the suffix_SDF
).from_where
: be one ofgenerated
ofdataset
.
An example command to calculate metrics for the sampled molecules is:
python scripts/evaluate_all.py --result_root ./outputs --exp_name sample_MolDiff_20230101_000000 --from_where generated
You also need to calculate some metrics for the test dataset to calculate the Jensen-Shannon divergence (JSD) between the generated molecules and the test dataset. To do so, run the following command:
python scripts/evaluate_all.py --exp_name test --from_where dataset
Then you can use the interactive notebook script/analyze_generated.ipynb
to analyze all the metrics defined in the paper. But make sure to set the directory of the sampled molecules and the test dataset correctly in the notebook.
We also provide two versions of training config files: the complete MolDiff train_MolDiff.yml
and the simplified one train_MolDiff_simple.yml
(not use the new bond noise schedule). To train the model from scratch, run the following command:
python scripts/train_drug3d.py --config <path_to_config_file> --device <device_id> --logdir <log_directory>
For example, to train the complete MolDiff model, run:
python scripts/train_drug3d.py --config ./configs/train/train_MolDiff.yml --device cuda:0 --logdir ./logs
To use the bond predictor for guidance during sampling, you also need to train a bond predictor:
python scripts/train_bond.py --config ./configs/train/train_bondpred.yml --device cuda:1 --logdir ./logs
@InProceedings{pmlr-v202-peng23b,
title = {{M}ol{D}iff: Addressing the Atom-Bond Inconsistency Problem in 3{D} Molecule Diffusion Generation},
author = {Peng, Xingang and Guan, Jiaqi and Liu, Qiang and Ma, Jianzhu},
booktitle = {Proceedings of the 40th International Conference on Machine Learning},
pages = {27611--27629},
year = {2023},
editor = {Krause, Andreas and Brunskill, Emma and Cho, Kyunghyun and Engelhardt, Barbara and Sabato, Sivan and Scarlett, Jonathan},
volume = {202},
series = {Proceedings of Machine Learning Research},
month = {23--29 Jul},
publisher = {PMLR},
pdf = {https://proceedings.mlr.press/v202/peng23b/peng23b.pdf},
url = {https://proceedings.mlr.press/v202/peng23b.html},
}
If you have any question, feel free to contact me at [email protected]