This repository is the official implementation of Reprogramming Pretrained Target-Specific Diffusion Models for Dual-Target Drug Design (NeurIPS 2024). [PDF]
Our codebase is built upon TargetDiff.
Please refer to the official repo of TargetDiff for details about installation.
Differently, the following dependencies are required:
python -m pip install numpy==1.23.1
python -m pip install meeko==0.3.0
The curated dataset is provided at https://huggingface.co/datasets/zhouxiangxin/DualDiff.
Download data.tar.gz, uncompress it, and put it under the folder of the project.
The data is organized as follows:
data/processed/dock/index_dict.pkl
data/processed/drug_synergy/synergy_idx_list.pkl
...
data/processed/dock/ligand_protein_dataset_v2/CC12CCC3C(CCC4=CC(=O)C=CC43C)C1CCC(=O)O2/
data/processed/dock/ligand_protein_dataset_v2/CC12CCC3C(CCC4=CC(=O)C=CC43C)C1CCC(=O)O2/P11511/
data/processed/dock/ligand_protein_dataset_v2/CC12CCC3C(CCC4=CC(=O)C=CC43C)C1CCC(=O)O2/P11511/ligand.sdf
data/processed/dock/ligand_protein_dataset_v2/CC12CCC3C(CCC4=CC(=O)C=CC43C)C1CCC(=O)O2/P11511/pocket_10A.pdb
data/processed/dock/ligand_protein_dataset_v2/CC12CCC3C(CCC4=CC(=O)C=CC43C)C1CCC(=O)O2/P11511/protein.pdb
data/processed/dock/ligand_protein_dataset_v2/CC12CCC3C(CCC4=CC(=O)C=CC43C)C1CCC(=O)O2/P11511/protein_clean.pdb
...
data/processed/dock/ligand_protein_dataset_v2/c1cnc2cc3c(cc2n1)C1CNCC3C1/
data/processed/dock/ligand_protein_dataset_v2/c1cnc2cc3c(cc2n1)C1CNCC3C1/P32297/
data/processed/dock/ligand_protein_dataset_v2/c1cnc2cc3c(cc2n1)C1CNCC3C1/P32297/ligand.sdf
data/processed/dock/ligand_protein_dataset_v2/c1cnc2cc3c(cc2n1)C1CNCC3C1/P32297/pocket_10A.pdb
data/processed/dock/ligand_protein_dataset_v2/c1cnc2cc3c(cc2n1)C1CNCC3C1/P32297/protein.pdb
data/processed/dock/ligand_protein_dataset_v2/c1cnc2cc3c(cc2n1)C1CNCC3C1/P32297/protein_clean.pdb
...
where index_dict.pkl is the mapping file from indices (from 0 to 437 inclusively) to smiles of the reference ligands, and synergy_idx_list.pkl is a list where each element is a tuple of two indices, indicating that the two targets are synergistic and there are poteinal dual-target drugs for these two targets.
CC12CCC3C(CCC4=CC(=O)C=CC43C)C1CCC(=O)O2 is the smiles of a reference ligand and P11511 is the UniProt ID of the corresponding target.
*/ligand.sdf is the sdf file of the reference ligand. pocket_10A is the pocket. (The residues within a 10 Å range of the reference molecule are retained.)
protein.pdb and protein_clean.pdb are complete proteins.
Please note that the preprocessed data is for research purposes only.
DualDiff and CompDiff are sampling algorithms. The base model is targetdiff. Please follow the instructions of official repo of TargetDiff to download the pretrained model.
The download link is also provided as follows:
https://drive.google.com/drive/folders/1-ftaIrTXjWFhw3-0Twkrs5m0yX6CNarz?usp=share_link
For DualDiff and CompDiff, the pockets of the dual targets need to be aligned. The alignment is based on generated ligands which can be selected by different criteria.
Please see dual_scripts/align/align_protein_ligand_score.py for example.
As an example, a pair of aligned pockets based on Max Vina Dock is provided at outputs/composition_score/8/104.
The following visualization shows the pocket, the reference lgiand, and the generated ligand of target A (left) (idx=8) and target B (middle) (idx=104) and the aligned pockets (right).
 |
 |
 |
|---|
Examples of sampling algorithms are provided as follows:
To sample for a single target, please refer to dual_scripts/example/sample_targetdiff.sh.
To sample for dual targets by DualDiff, please refer to dual_scripts/example/sample_targetdiff.sh.
To sample for dual targets by CompDiff, please refer to dual_scripts/example/sample_compdiff.sh.
The evaluation is the same as TargetDiff. We also provided examples at dual_scripts/evaluate.
@inproceedings{
zhou2024reprogramming,
title={Reprogramming Pretrained Target-Specific Diffusion Models for Dual-Target Drug Design},
author={Xiangxin Zhou and Jiaqi Guan and Yijia Zhang and Xingang Peng and Liang Wang and Jianzhu Ma},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
year={2024},
}