PILOT: Equivariant diffusion for pocket conditioned de novo ligand generation with multi-objective guidance via importance sampling

This is the official repository for PILOT - a model for guided structure-based drug discovery via equivariant (continuous and discrete) denoising diffusion. If you have any questions, feel free to reach out to us: [email protected], [email protected].

Installation

Install the main environment via mamba

mamba env create -f environment.yml

For preparing pdbqt files, install a new environment

conda create -n mgltools -c bioconda mgltools

We also recommend installing a separate environment for running the docking

mamba env create -f environment_vina.yml

Data

Activate the main environment

conda activate e3mol

CrossDocked

Download the CrossDocked data as described in https://github.com/pengxingang/Pocket2Mol/tree/main/data

Create the CrossDocked data

python experiments/data/ligand/process_crossdocked.py --basedir /path/to/crossdocked_pocket10-folder --outdir /your/data/folder --no-H --dist-cutoff 7 

Kinodata-3D

Download the Kinodata-3D dataset here (only kinodata_docked_with_rmsd.sdf.gz needed) https://zenodo.org/records/10410259

Create Kinodata-3D dataset

python experiments/data/ligand/process_kinodata.py --basedir /path/to/kinodata_folder --outdir /your/data/folder --no-H --dist-cutoff 5 

PDBQT files for docking

Create the pdbqt files for the test complexes Activate the mgltools environment

conda activate mgltools

python experiments/docking_mgl.py path/to/test_dir /where/to/store/pdbqt_files dataset

(replace dataset with "crossdocked" or "kinodata")

Training

Activate the main environment

conda activate e3mol

Pocket-conditioned diffusion training

Train PILOT from scratch on CrossDocked

python experiments/run_train.py --conf configs/diffusion_crossdocked.yaml --save-dir /your/save/dir

Train PILOT from scratch on Kinodata-3D

python experiments/run_train.py --conf configs/diffusion_kinodata.yaml --save-dir /your/save/dir

Sampling

Model checkpoints

Currently, we provide the model weights upon request. Please contact us via email.

Test set (on multiple nodes using SLURM's job array)

Sample de novo ligands given the CrossDocked (Kinodata-3D) test set, the sampling can be started on multiple GPU nodes:

Modify scripts/generate_ligands_multi.sl (scripts/generate_ligands_multi_kinodata.sl):

- num-gpus: Number of GPU nodes you want to use (number of test files divided by num-gpus)
- model-path: Set the path to the trained model (normally save_dir/best_valid.ckpt)
- save-dir: Where the sampled molecules as SDF files shall be saved
- test-dir: Path to test directory containing .pdb, .sdf and .txt files
- pdb-dir: Path to the pre-processed pdb files (see above: experiments/data/ligand/fetch_pdb_files.py)
- dataset-root: Main path to the dataset
- batch-size: Batch size (40-50 on a V100 GPU)
- n-nodes-bias: The ligand sizes are sampled from the ligand size distribution extracted from the training data. With n-nodes-bias an additional number of atoms is added (for crossdocked: 10)
- num-ligands-per-pocket-to-sample: 100 [default on CrossDocked 100]
- num-ligands-per-pocket-to-save: 100 [default on CrossDocked 100]
- max-sample-iter: 50 [max. number of iterations to fulfill num-ligands-per-pocket-to-sample]
- batch-size: 40 
- n-nodes-bias: 0 [increase sampled/fixed ligand size by the number provided]
- vary-n-nodes: [0, n-nodes-bias] is added randomly (uniform)
- fix-n-nodes [whether or not to use the ground truth ligand size for number of atoms (hence no sampling of ligand sizes)]
- prior-n-atoms: targetdiff [conditional or targetdiff - sample ligand size from pocket conditional ligand size distribution]
- property-importance-sampling [whether or not to use property importance sampling]
- property-importance-sampling-start: 200 [when on the diffusion trajectory to start importance sampling]
- property-importance-sampling-end: 300 [when on the diffusion trajectory to end importance sampling]
- property-every-importance-t: 5 [every n-th step perform importance sampling]
- property-tau 0.1 [temperature for importance sampling]
- sa-importance-sampling [whether or not to use SA importance sampling]
- sa-importance-sampling-start: 0 [when on the diffusion trajectory to start importance sampling]
- sa-importance-sampling-end: 300 [when on the diffusion trajectory to start importance sampling]
- sa-every-importance-t: 5 [every n-th step perform importance sampling]
- sa-tau: 0.1 [temperature for importance sampling]

sbatch scripts/generate_ligands_multi.sl

After sampling is finished, aggregate the results from all jobs to print the full evaluation

python experiments/aggregate_results.py --files-dir /your/sampling/save_dir

All ligands per target are saved in sdf files. The molecules in the sdf files contain all properties as well.

Docking of generated ligands (on multiple nodes using SLURM's job array)

As soon ligands are generated for the respective pockets, we can start docking.

Modify scripts/docking_multi.sl (scripts/docking_multi_kinodata.sl):

- num-cpus: Number of CPU nodes you want to use (number of generated sdf files divided by num-cpus; see IMPORTANT note below)
 -sdf-dir: Path to the generated ligands 
 -save-dir Path where all evaluations are saved at
 -pdbqt-dir Path where all pdbqt files are stored (see above: experiments/docking_mgl.py)
 -pdb-dir: Path to the pre-processed pdb files (see above: experiments/data/ligand/fetch_pdb_files.py)
 -dataset: Which dataset, e.g., crossdocked
 -docking-mode: vina_dock or qvina2 (default)

sbatch scripts/docking_multi.sl

After docking is finished, aggregate the results from all jobs to print the full evaluation

python experiments/aggregate_results.py --files-dir /your/docking/save_dir --docked --docking-mode qvina2

Single PDB file

Activate the main environment

conda activate e3mol

In general, we assume a ground truth ligand docked to a protein, from which the binding site can be extracted. Otherwise the binding site must be found first.

To get all necessary files for sampling, run

python experiments/data/ligand/process_pdb.py --main-path /path/to/main_folder --pdb-id PDB_ID --ligand-id LIGAND_ID --no-H --dist-cutoff 7

Activate the mgltools environment and create the pdbqt file

conda activate mgltools

python experiments/docking_mgl.py path/to/pdb_dir /where/to/store/pdbqt_file dataset

(replace dataset with "pdb_file")

Then for sampling run

sbatch scripts/generate_ligands_multi_pdb_file.sl

(specify all arguments to your needs as before)

Afterwards, for docking run

sbatch scripts/docking_multi_pdb_file.sl

(again, specify the arguments; the docking mode can be set to "vina_dock", "vina_score", "qvina2". Default is "qvina2".)

Acknowledgement

This study was partially funded by the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie Actions grant agreement “Advanced machine learning for Innovative Drug Discovery (AIDD)” No. 956832.