Official implementation of DiffLinker, an Equivariant 3D-conditional Diffusion Model for Molecular Linker Design by Ilia Igashov, Hannes Stärk, Clément Vignac, Arne Schneuing, Victor Garcia Satorras, Pascal Frossard, Max Welling, Michael Bronstein and Bruno Correia.
Given a set of disconnected fragments in 3D, DiffLinker places missing atoms in between and designs a molecule incorporating all the initial fragments. Our method can link an arbitrary number of fragments, requires no information on the attachment atoms and linker size, and can be conditioned on the protein pockets.
Animations
The code was tested in the following environment:
| Software | Version |
|---|---|
| Python | 3.10.5 |
| CUDA | 10.2.89 |
| PyTorch | 1.11.0 |
| PyTorch Lightning | 1.6.3 |
| OpenBabel | 3.0.0 |
You can create a new conda environment using provided environment.yaml file:
conda env create -f environment.ymlor manually creating the base environment:
conda create -c conda-forge -n difflinker rdkitand installing all the necessary packages:
biopython
imageio
networkx
pytorch
pytorch-lightning
scipy
scikit-learn
tqdm
wandbActivate the environment:
conda activate difflinkerNormally, the whole installation process takes 5-10 min.
Please find the models here or use direct download links:
- [ZINC] DiffLinker
- [ZINC] DiffLinker (given anchors)
- [ZINC] Size GNN
- [GEOM] DiffLinker
- [GEOM] DiffLinker (given anchors)
- [GEOM] Size GNN
- [Pockets] DiffLinker (full pocket)
- [Pockets] DiffLinker (full pocket, given anchors)
- [Pockets] DiffLinker (backbone atoms, given anchors)
- [Pockets] DiffLinker (unconditioned, given anchors)
First, download necessary models and create directories (we recommend to use GEOM models as they are the most generic):
mkdir -p models
wget https://zenodo.org/record/7121300/files/geom_difflinker.ckpt?download=1 -O models/geom_difflinker.ckpt
wget https://zenodo.org/record/7121300/files/geom_size_gnn.ckpt?download=1 -O models/geom_size_gnn.ckptGenerate linkers for your own fragments:
python -W ignore generate.py --fragments <YOUR_PATH> --model models/geom_difflinker.ckpt --linker_size models/geom_size_gnn.ckptIf you have the full target protein and want the pocket to be computed automatically based on the input fragments:
mkdir -p models
wget https://zenodo.org/records/10988017/files/pockets_difflinker_full_no_anchors_fc_pdb_excluded.ckpt?download=1 -O models/pockets_difflinker_full.ckpt
python -W ignore generate_with_protein.py --fragments <FRAGMENTS_PATH> --protein <PROTEIN_PATH> --model models/pockets_difflinker_full.ckpt --linker_size <DESIRED_LINKER_SIZE> --anchors <COMMA_SEPARATED_ANCHOR_INDICES> If you want to use the file with pocket you computed yourself:
mkdir -p models
wget https://zenodo.org/records/10988017/files/pockets_difflinker_full_no_anchors_fc_pdb_excluded.ckpt?download=1 -O models/pockets_difflinker_full.ckpt
python -W ignore generate_with_pocket.py --fragments <FRAGMENTS_PATH> --pocket <POCKET_PATH> --model models/pockets_difflinker_full.ckpt --linker_size <DESIRED_LINKER_SIZE> --anchors <COMMA_SEPARATED_ANCHOR_INDICES> mkdir -p models
wget https://zenodo.org/record/7121300/files/pockets_difflinker_backbone.ckpt?download=1 -O models/pockets_difflinker_backbone.ckpt
python -W ignore generate_with_pocket.py --fragments <FRAGMENTS_PATH> --pocket <POCKET_PATH> --backbone_atoms_only --model models/pockets_difflinker_backbone.ckpt --linker_size <DESIRED_LINKER_SIZE> --anchors <COMMA_SEPARATED_ANCHOR_INDICES>- Fragment file should be passed in one of the following formats:
.sdf,.pdb,.mol,.mol2 - Protein should be passed in
.pdbformat - Currently pocket-conditioned generation does not support prediction and sampling of the linker size (will be added later)
- To obtain correct anchor indices for your fragments, you can open the file in PyMOL and click
Label -> atom identifiers -> ID. You can select anchor atoms and pass the corresponding IDs to the generation script - For more options check help:
python generate.py --helporpython generate_with_pocket.py --help
First, download datasets:
mkdir -p datasets
wget https://zenodo.org/record/7121271/files/zinc_final_train.pt?download=1 -O datasets/zinc_final_train.pt
wget https://zenodo.org/record/7121271/files/zinc_final_val.pt?download=1 -O datasets/zinc_final_val.ptNext, create necessary directories:
mkdir -p models
mkdir -p logsRun trainig:
python -W ignore train_difflinker.py --config configs/zinc_difflinker.ymlIn this example, we will consider the training and testing process on the ZINC dataset.
All the instructions about downloading or creating datasets from scratch can be found in data directory.
python -W ignore train_size_gnn.py \
--experiment zinc_size_gnn \
--data datasets \
--train_data_prefix zinc_final_val \
--val_data_prefix zinc_final_val \
--hidden_nf 256 \
--n_layers 5 \
--batch_size 256 \
--normalization batch_norm \
--lr 1e-3 \
--task classification \
--loss_weights \
--device gpu \
--checkpoints models \
--logs logsThere are the distributions of numbers of atoms in linkers used for training linker size prediction GNNs:
First, download test dataset:
mkdir -p datasets
wget https://zenodo.org/record/7121271/files/zinc_final_test.pt?download=1 -O datasets/zinc_final_test.ptDownload the necessary models:
mkdir -p models
wget https://zenodo.org/record/7121300/files/zinc_difflinker.ckpt?download=1 -O models/zinc_difflinker.ckpt
wget https://zenodo.org/record/7121300/files/zinc_size_gnn.ckpt?download=1 -O models/zinc_size_gnn.ckptNext, create necessary directories:
mkdir -p samples
mkdir -p trajectoriesIf you want to sample 250 linkers for each input set of fragments, run the following:
python -W ignore sample.py \
--checkpoint models/zinc_difflinker.ckpt \
--linker_size_model models/zinc_size_gnn.ckpt \
--samples samples \
--data datasets \
--prefix zinc_final_test \
--n_samples 2 \
--device cuda:0You will be able to see .xyz files of the generated molecules in the directory ./samples.
If you want to sample linkers and save trajectories, run the following:
python -W ignore sample_trajectories.py \
--checkpoint models/zinc_difflinker.ckpt \
--chains trajectories \
--data datasets \
--prefix zinc_final_test \
--keep_frames 10 \
--device cuda:0You will be able to see trajectories as .xyz, .png and .gif files in the directory ./trajectories.
First, you need to download ground-truth SMILES and SDF files of molecules, fragments and linkers from the relevant test sets (recomputed with OpenBabel) + SMILES of the training linkers. Check this resource for finding the right ones. Here, we will download files for ZINC:
mkdir -p datasets
wget https://zenodo.org/record/7121448/files/zinc_final_test_smiles.smi?download=1 -O datasets/zinc_final_test_smiles.smi
wget https://zenodo.org/record/7121448/files/zinc_final_test_molecules.sdf?download=1 -O datasets/zinc_final_test_molecules.sdf
wget https://zenodo.org/record/7121448/files/zinc_final_train_linkers.smi?download=1 -O datasets/zinc_final_train_linkers.smi Next, you need to run OpenBabel to reformat the data:
mkdir -p formatted
python -W ignore reformat_data_obabel.py \
--samples samples \
--dataset zinc_final_test \
--true_smiles_path datasets/zinc_final_test_smiles.smi \
--checkpoint zinc_difflinker \
--formatted formatted \
--linker_size_model_name zinc_size_gnnThen you can run evaluation scripts:
python -W ignore compute_metrics.py \
ZINC \
formatted/zinc_difflinker/sampled_size/zinc_size_gnn/zinc_final_test.smi \
datasets/zinc_final_train_linkers.smi \
5 1 None \
resources/wehi_pains.csv \
diffusionAll the metrics will be saved in the directory ./formatted.
Igashov, I., Stärk, H., Vignac, C. et al. Equivariant 3D-conditional diffusion model for molecular linker design. Nat Mach Intell (2024). https://doi.org/10.1038/s42256-024-00815-9
@article{igashov2024equivariant,
title={Equivariant 3D-conditional diffusion model for molecular linker design},
author={Igashov, Ilia and St{\"a}rk, Hannes and Vignac, Cl{\'e}ment and Schneuing, Arne and Satorras, Victor Garcia and Frossard, Pascal and Welling, Max and Bronstein, Michael and Correia, Bruno},
journal={Nature Machine Intelligence},
pages={1--11},
year={2024},
publisher={Nature Publishing Group UK London}
}
If you have any questions, please contact me at [email protected]
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent