DiffSBDD: Structure-based Drug Design with Equivariant Diffusion Models

Official implementation of DiffSBDD, an equivariant model for structure-based drug design, by Arne Schneuing*, Yuanqi Du*, Charles Harris, Arian Jamasb, Ilia Igashov, Weitao Du, Tom Blundell, Pietro Lió, Carla Gomes, Max Welling, Michael Bronstein & Bruno Correia.

1. Dependencies
   1. Conda environment
   2. QuickVina 2
   3. Pre-trained models
2. Benchmarks
   1. CrossDocked Benchmark
   2. Binding MOAD
   3. Sampled molecules
3. Training
4. Inference
   1. Sample molecules for a given pocket
   2. Test set sampling
   3. Fix substructures
   4. Metrics
   5. QuickVina2
5. Citation

Dependencies

Conda environment

conda create -n sbdd-env
conda activate sbdd-env
conda install pytorch cudatoolkit=10.2 -c pytorch
conda install -c conda-forge pytorch-lightning
conda install -c conda-forge wandb
conda install -c conda-forge rdkit
conda install -c conda-forge biopython
conda install -c conda-forge imageio
conda install -c anaconda scipy
conda install -c pyg pytorch-scatter
conda install -c conda-forge openbabel

The code was tested with the following versions

QuickVina 2

For docking, install QuickVina 2:

wget https://github.com/QVina/qvina/raw/master/bin/qvina2.1
chmod +x qvina2.1 

We need MGLTools for preparing the receptor for docking (pdb -> pdbqt) but it can mess up your conda environment, so I recommend to make a new one:

conda create -n mgltools -c bioconda mgltools

Pre-trained models

Pre-trained models can be downloaded from Zenodo.

• CrossDocked, conditional $C_\alpha$ model
• CrossDocked, joint $C_\alpha$ model
• CrossDocked, conditional full-atom model
• CrossDocked, joint full-atom model
• Binding MOAD, conditional $C_\alpha$ model
• Binding MOAD, joint $C_\alpha$ model
• Binding MOAD, conditional full-atom model
• Binding MOAD, joint full-atom model

Benchmarks

CrossDocked

Data preparation

Download and extract the dataset as described by the authors of Pocket2Mol: https://github.com/pengxingang/Pocket2Mol/tree/main/data

Process the raw data using

python process_crossdock.py <crossdocked_dir> --no_H

Binding MOAD

Data preparation

Download the dataset

wget http://www.bindingmoad.org/files/biou/every_part_a.zip
wget http://www.bindingmoad.org/files/biou/every_part_b.zip
wget http://www.bindingmoad.org/files/csv/every.csv

unzip every_part_a.zip
unzip every_part_b.zip

Process the raw data using

python -W ignore process_bindingmoad.py <bindingmoad_dir>

Add the --ca_only flag to create a dataset with $C_\alpha$ pocket representation.

Sampled molecules

Sampled molecules can be found on Zenodo.

Training

Starting a new training run:

python -u train.py --config <config>.yml

Resuming a previous run:

python -u train.py --config <config>.yml --resume <checkpoint>.ckpt

Inference

Sample molecules for a given pocket

To sample small molecules for a given pocket with a trained model use the following command:

python generate_ligands.py <checkpoint>.ckpt --pdbfile <pdb_file>.pdb --outdir <output_dir> --resi_list <list_of_pocket_residue_ids>

For example:

python generate_ligands.py last.ckpt --pdbfile 1abc.pdb --outdir results/ --resi_list A:1 A:2 A:3 A:4 A:5 A:6 A:7 

Alternatively, the binding pocket can also be specified based on a reference ligand in the same PDB file:

python generate_ligands.py <checkpoint>.ckpt --pdbfile <pdb_file>.pdb --outdir <output_dir> --ref_ligand <chain>:<resi>

Optional flags:

Sample molecules for all pockets in the test set

test.py can be used to sample molecules for the entire testing set:

python test.py <checkpoint>.ckpt --test_dir <bindingmoad_dir>/processed_noH/test/ --outdir <output_dir> --sanitize

There are different ways to determine the size of sampled molecules.

• --fix_n_nodes: generates ligands with the same number of nodes as the reference molecule
• --n_nodes_bias <int>: samples the number of nodes randomly and adds this bias
• --n_nodes_min <int>: samples the number of nodes randomly but clamps it at this value

Other optional flags are equivalent to generate_ligands.py.

Fix substructures

inpaint.py can be used for partial ligand redesign with the conditionally trained model, e.g.:

python inpaint.py <checkpoint>.ckpt --pdbfile <pdb_file>.pdb --outdir <output_dir> --ref_ligand <chain>:<resi> --fix_atoms C1 N6 C5 C12

--add_n_nodes controls the number of newly generated nodes

Metrics

For assessing basic molecular properties create an instance of the MoleculeProperties class and run its evaluate method:

from analysis.metrics import MoleculeProperties
mol_metrics = MoleculeProperties()
all_qed, all_sa, all_logp, all_lipinski, per_pocket_diversity = \
    mol_metrics.evaluate(pocket_mols)

evaluate() expects a list of lists where the inner list contains all RDKit molecules generated for one pocket.

For computing docking scores, run QuickVina as described below.

QuickVina2

First, convert all protein PDB files to PDBQT files using MGLTools

conda activate mgltools
cd analysis
python docking_py27.py <bindingmoad_dir>/processed_noH/test/ <output_dir> bindingmoad
cd ..
conda deactivate

Then, compute QuickVina scores:

conda activate sbdd-env
python analysis/docking.py --pdbqt_dir <docking_py27_outdir> --sdf_dir <test_outdir> --out_dir <qvina_outdir> --write_csv --write_dict

Citation

@article{schneuing2022structure,
  title={Structure-based Drug Design with Equivariant Diffusion Models},
  author={Schneuing, Arne and Du, Yuanqi and Harris, Charles and Jamasb, Arian and Igashov, Ilia and Du, Weitao and Blundell, Tom and Li{\'o}, Pietro and Gomes, Carla and Welling, Max and Bronstein, Michael and Correia, Bruno},
  journal={arXiv preprint arXiv:2210.13695},
  year={2022}
}