PharDiff: Pharmacophore-aware Diffusion Model for Pocket Specific 3D Molecular Generation

Diffusion process of proposed method

Note: For a better understanding of the process, protein pockets have been excluded in the video below.

Forward process	Reverse process

Requirements

We include key dependencies below. Our detailed environmental setup is available in [environment.yml] The code has been tested in the following environment:

Package	Version
Python	3.8
PyTorch	1.13.1
CUDA	11.6
PyTorch Geometric	2.2.0
RDKit	2022.03.2

Install via Conda

conda create -n PharDiff python=3.8
conda activate PharDiff
conda install pytorch pytorch-cuda=11.6 -c pytorch -c nvidia
conda install pyg -c pyg
conda install rdkit openbabel tensorboard pyyaml easydict python-lmdb -c conda-forge

Data

The data used for training/evaluation would have been provided through the submission site in a folder named Data or Google Drive folder.

Data
|__Training Data  
|   |  # Raw complex structures of protein-ligand available from the CrossDocked2020 dataset. Proteins are specified in .pdb format, and Ligands in .sdf format.
|   |__crossdocked_v1.1_rmsd1.0.tar.gz 
|   |
|   |  # Processed data that can be used for model training, obtainable through the execution of the ./Anonymous/datasets/pl_pair_dataset.py file
|   |__crossdocked_v1.1_rmsd1.0_pocket10_processed_final.lmdb 
|   |
|   |  # Index storage files for each sample, used for splitting the train set and test set, or for other preprocessing purposes.
|   |__index.pkl
|   
|__Vocab
|   |   # The set of Fragments constituting the ligands of CrossDocked2020 samples, extracted through the 3D Fragmentizing Algorithm.
|   |___vocab.txt
|   |
|   |   # The set of pharmacophores playing a crucial role in protein-ligand interactions, extracted through the 3D Pharmacophore Decomposition proposed in this paper.
|   |___vocab_super.pickle
|   
|__Split
|   |   # Names and index numbers of samples used directly for training and validation.
|   |___crossdocked_pocket10_pose_split.pt
|   |
|   |   # Raw file for creating the crossdocked_pocket10_pose_split.pt file. It is split through pdb id.
|   |___split_by_name.pt
|
|__Test Data
|   |...
|

To train the model from scratch, you need the preprocessed lmdb file and split file:

• crossdocked_v1.1_rmsd1.0_pocket10_processed_final.lmdb
• crossdocked_pocket10_pose_split.pt

To evaluate the model on the test set, you need to unzip the test_set.zip in Data folder. It includes the original PDB files that will be used in Vina Docking.

If you want to process the dataset from scratch, you need to download CrossDocked2020 v1.1 from here, save it into data/CrossDocked2020, and run the scripts in scripts/data_preparation:

• clean_crossdocked.py will filter the original dataset and keep the ones with RMSD < 1A. It will generate a index.pkl file and create a new directory containing the original filtered data (corresponds to crossdocked_v1.1_rmsd1.0.tar.gz in the drive). You don't need these files if you have downloaded .lmdb file.

• extract_pockets.py will clip the original protein file to a 10A region around the binding molecule. E.g.

• split_pl_dataset.py will split the training and test set. We use the same split split_by_name.pt as AR and Pocket2Mol, which can also be downloaded in the Google Drive - data folder.

• get_pharmacophore.ipynb will identify pharmacophores in complex data comprising a protein's pocket and the ligand. Through this process, it is possible to extract information about the indices and position of atoms that make up the pharmacophore, as well as details about the intermolecular interactions they undergo.

  python scripts/data_preparation/clean_crossdocked.py --source data/CrossDocked2020 --dest data/crossdocked_v1.1_rmsd1.0 --rmsd_thr 1.0
  
  python scripts/data_preparation/extract_pockets.py --source data/crossdocked_v1.1_rmsd1.0 --dest data/crossdocked_v1.1_rmsd1.0_pocket10
  
  python scripts/data_preparation/split_pl_dataset.py --path data/crossdocked_v1.1_rmsd1.0_pocket10 --dest data/crossdocked_pocket10_pose_split.pt --fixed_split data/split_by_name.pt

Note

If you want to directly understand the preprocessing process of the dataset for training a diffusion model that considers pharmacophores, please refer to ./Anonymous/datasets/pl_pair_dataset.py.

If the .lmdb file does not exist, this code will automatically execute, but it will not run if it already exists.

If you have already moved all the datasets provided through the submission site to the directory, this code will not execute.

Training

Training from scratch

python scripts/train_diffusion.py configs/training.yml

Sampling

Sampling for pockets in the testset

python scripts/sample_diffusion.py configs/sampling.yml --data_id {i}

Evaluation

Evaluation from sampling results

python scripts/evaluate_diffusion.py {OUTPUT_DIR} --docking_mode vina_score --protein_root data/test_set

The docking mode can be chosen from {qvina, vina_score, vina_dock, none}

Note: It will take some time to prepare pqdqt and pqr files when you run the evaluation code with vina_score/vina_dock docking mode for the first time.

Real-world Validation

If you want to generate molecules for a new protein not in the test set, you should run ./Anonymous/scripts/real_world_inference.ipynb. Remember that you need to prepare the ligand's .sdf file for creating the protein pocket and the .pdb file containing the structural information of the protein.

Typically, the above process is also necessary for performing MD simulation.

Result

The main results for the proposed model are presented in the table below. For a more comprehensive overview of the results obtained with our model, please refer to the Report.

Evaluation of Generated Molecule

Model	VinaScore	VinaMin	VinaDock	HighAiffinity	VinaScoreSA	SR
AR	-5.75	-6.18	-6.75	0.379	-5.59	74.7%
LiGAN	-	-	-6.33	0.21	-	-68.4%
GraphBP	-	-	-4.80	0.14	-	57.1%
Pocket2Mol	-5.15	-6.42	-7.15	0.48	-5.12	88.7%
DiffSBDD	52.78	16.45	-6.65	0.452	-51.53	83.0%
DrugGPS	28.18	6.33	-3.74	0.12	-27.32	48.1%
TargetDiff	-5.47	-6.64	-7.80	0.57	-5.31	91.9%
ResGen	13.79	-1.53	-4.90	0.23	-13.73	40.7%
PharDiff	-6.64	-7.20	-8.16	0.62	-5.93	98.1%
Testset	-6.36	-6.71	-7.45	-	-6.28	-

Distribution of RMSD before and after Docking	Demo video of Molecular Dynamics about Generated Molecule