This is the official code repository for the paper Drug Discovery with Dynamic Goal-aware Fragments (ICML 2024).
In our paper, we introduce:

• Fragment-wise Graph Information Bottleneck (FGIB), a goal-aware fragment extration method using the graph information bottleneck (GIB) theory to construct a fragment vocabulary for target chemical properties.

  

• Goal-aware fragment Extraction, Assembly, and Modification (GEAM), a generative framework consists of FGIB, soft-actor critic (SAC), and genetic algorithm (GA), each responsible for fragment extraction, fragment assembly, and fragment modification.

  

Run the following commands to install the dependencies:

conda create -n geam python==3.7
conda activate geam
conda install cudatoolkit=10.2
pip install torch==1.8.1+cu102 -f https://download.pytorch.org/whl/torch_stable.html
pip install torch-scatter==2.0.8 torch-sparse==0.6.12 torch-cluster==1.5.9 -f https://data.pyg.org/whl/torch-1.8.1+cu102.html
pip install torch-geometric==2.0.4 rdkit requests urllib3==1.26.6 more_itertools gym==0.18
conda install -c conda-forge openbabel
conda install -c dglteam dgl-cuda10.2==0.8

Run the following command to preprocess the ZINC250k dataset:

python utils_fgib/data.py

We provide the pretrained FGIB for target proteins parp1, fa7, 5ht1b, braf, and jak2, respectively (parp1.pt, fa7.pt, 5ht1b.pt, braf.pt, and jak2.pt, respectively), in the folder ckpt.

To train your own FGIB, run the following command:

python train_fgib.py -g ${gpu_id} -t ${target}
# e.g., python train_fgib.py -g 0 -t parp1

We provide the initial fragment vocabularies for target proteins parp1, fa7, 5ht1b, braf, and jak2, respectively (parp1.txt, fa7.txt, 5ht1b.txt, braf.txt, and jak2.txt, respectively), in the folder data.

To construct the initial fragment vocabulary from scratch, run the following command:

python get_frags.py -g ${gpu_id} -t ${target} -m ${gib_path} -v ${vocab_path}
# e.g., python get_frags.py -g 0 -t parp1 -m ckpt/parp1.pt -v data/parp1.txt

To generate molecules with GEAM, run the following command:

python run.py -g ${gpu_id} -t ${target} -m ${gib_path} -v ${vocab_path}
# e.g., python run.py -g 0 -t parp1 -m ckpt/parp1.pt -v data/parp1.txt

To evaluate the generated molecules, run the following command:

python eval.py ${file_name} -t ${target}
# e.g., python eval.py results/file_name.csv -t parp1

If you find this repository and our paper useful, we kindly request to cite our work.

@article{lee2024GEAM,
  author    = {Seul Lee and Seanie Lee and Kenji Kawaguchi and Sung Ju Hwang},
  title     = {Drug Discovery with Dynamic Goal-aware Fragments},
  journal   = {Proceedings of the 41th International Conference on Machine Learning},
  year      = {2024}
}