SculptDrug is a novel spatial condition-aware generative model designed for structure-based drug design (SBDD). It leverages a Bayesian Flow Network (BFN) to integrate hierarchical structural information and spatial protein constraints, enabling the generation of high-affinity, geometry-consistent ligands.
The implementation is under active reorganization.
We plan to release the full codebase shortly to support reproducibility and future research. For evaluation, we provide generated results in the meantime.
git clone https://github.com/decisionintelligence/SculptDrug.git
cd SculptDrugThe following are the key dependencies required to run SculptDrug.
| Library | Version |
|---|---|
| Python | 3.9.17 |
| PyTorch | 2.1.0+cu118 |
| PyTorch Geometric | 2.3.1 |
| RDKit | 2023.9.5 |
| PyTorch Lightning | 2.3.0 |
Download the processed CrossDocked dataset and split files from Google Drive, originally provided by TargetDiff.
Place the following files under the data/ directory:
crossdocked_v1.1_rmsd1.0_pocket10_processed_final.lmdbcrossdocked_pocket10_pose_split.pt
To extract protein surface accessibility points:
conda create -n pymesh python=3.9.10
conda activate pymesh
conda install -c conda-forge/label/cf202003 pymesh2
# Run surface extraction
python ./profiles_extract/surface_extract_fast.py /path/to/crossdocked_v1.1_rmsd1.0_pocket10The fully preprocessed dataset will be released after the camera-ready version of the paper. For now, we provide a test subset for reproducibility.
To train SculptDrug from scratch:
python ./train_bfn.py --config_file configs/default.yaml --exp_name train_model --revision 0 --devices 0To generate ligands using a trained model checkpoint:
python train_bfn.py --test_only --num_samples 100 --sample_steps 100 --ckpt_path ./checkpoints/checkpoint.ckptOutput will be saved to:
./outputs/generated.pt
Visualize the generated ligands:
To evaluate the generated ligands, run:
bash script/test_baseline.sh --generated_file outputs/generated.pt --exhaustiveness 16The evaluation log will be saved to:
outputs/log.txt
This project builds upon ideas and code from:
We thank the original authors for their excellent contributions.
The code and datasets are released under MIT, CC BY-NC-SA 4.0, and GPL-3.0 licenses where applicable.