DyScore

DyScore is the implementation for the paper DyScore: A Boosting Scoring Method with Dynamic Properties for Identifying True Binders and Non-binders in Structure-based Drug Discovery.

How does it work?

DyScore consists of several steps:

• Molecular Docking: [Optional] Dock a ligand to a target protein to identify the ligand binding conformation.
• Data Processing: Generate the static and dynamic features for the protein-ligand complex.
• Model Prediction: Predict the likelihood of whether a given compound is a true binder.

Running DyScore

1. Install Docker and correctly configure it.

*Note: The installation of docker requires root privileges

2. Clone this repository and navigate to the repository folder with cd dyscore

git clone https://github.com/YanjunLi-CS/dyscore.git
cd dyscore

3. Download the model parameters to user specific <DOWNLOAD_DIR> path.

./download_weight.sh <DOWNLOAD_DIR>

The download_weight script will download the model parameter files for DyScore and DyScore-MF. Once the script has finished, you should have the following directory structure:

<DOWNLOAD_DIR>/                            # Total: ~ 18 GB
    woMF/                                  # ~ 12.1 GB
        # 3 files
    wMF/                                   # ~ 6.2 GB
        # 3 files

4. Install the docker/run_docker.py dependencies in a virtualenv and activate the virtualenv:

conda env create --name dyscore -f docker/dyscore.yml
conda activate dyscore

5. Build the Docker image:

docker build -f docker/Dockerfile -t dyscore ./

6. Run docker/run_docker.py to predict whether the given ligands can be bound to the target protein. The receipt file with PDB format is required to pass to the argument --rec_file. Multiple ligand files with different file names can be stored in the same directory, and the directory is required to pass to the argument --lig_dir. The prediction results will be outputted to the user-specific directory with the argument --out_dir.

It is essential to specify the correct ligand binding site. Here, we provide three different methods to find or assign binding site. The first one is to use a reference ligand, where users need to pass the reference ligand file to the argument --ref_file. For example,

python docker/run_docker.py \ 
    --rec_file ./example/3p0g_protein.pdb \
    --lig_dir ./example/ligand \
    --ref_file ./example/3p0g_ligand.mol2 \
    --out_dir ./example_output \
    --weight_dir <DOWNLOAD_DIR>

• The reference ligand should be extracted from the X-ray complex or docked complex, and please make sure the reference ligand is correctly bound to the binding site of the target protein.

The second one is to pass the 3D coordinates (i.e., min_x,min_y,min_z,max_x,max_y,max_z) of the box for ligand binding site to the argument --box. For example,

python docker/run_docker.py \ 
    --rec_file ./example/3p0g_protein.pdb \
    --lig_dir ./example/ligand \
    --box 51.00,3.00,-1.00,78.00,30.00,26.00 \
    --out_dir ./example_output \
    --weight_dir <DOWNLOAD_DIR>

The third one is to automatically detect the most likely ligand binding site using the Cavity program. But the detecting process may take a long time. The binding site definition is critical for any analysis for protein-ligand binding, so we DO NOT recommend the user to use the automatic detection method, which is possible to miss the correct binding site and mess up all the following processes. For example,

python docker/run_docker.py \ 
    --rec_file ./example/3p0g_protein.pdb \
    --lig_dir ./example/ligand \
    --detect \
    --out_dir ./example_output \
    --weight_dir <DOWNLOAD_DIR>

*Note that the three methods are mutually exclusive to each other. If users use any combination of them, they will receive the argument error.

Although DyScore is designed for post-analysis for results from virtual screening, we also provide the molecular docking option and by default, it is set as True. In this case, DyScore will automatically perform 3D conformation generation, protonation refinement, and molecular docking for input molecules. If the input molecules are already docked to the protein binding site by the user, just pass the argument --no_dock to disable the automatic docking. This will save a lot of time. For example,

python docker/run_docker.py \ 
    --rec_file ./example/3p0g_protein.pdb \
    --lig_dir ./example/ligand_docked \
    --ref_file ./example/3p0g_ligand.mol2 \
    --out_dir ./example_output \
    --no_dock \
    --weight_dir <DOWNLOAD_DIR>

Running DyScore-MF

We also provide the fingerprint-based version of DyScore, named DyScore-MF. The DyScore-MF could be activated by passing the argument --mf. For example,

python docker/run_docker.py \ 
    --rec_file ./example/3p0g_protein.pdb \
    --lig_dir ./example/ligand \
    --ref_file ./example/3p0g_ligand.mol2 \
    --out_dir ./example_output \
    --mf \
    --weight_dir <DOWNLOAD_DIR>

DyScore Results

The outputs will be saved in the directory provided via the --out_dir argument of run_docker.py

• example_output/input.csv (generated input file for DyScore)
• example_output/predicted.csv (Output file predicted by DyScore)
• example_output/input_mf.csv (generated input file for DyScore-MF)
• example_output/predicted_mf.csv (Output file predicted by DyScore-MF)

Larger prediction value in predicted.csv or predicted_mf.csv indicate higher possibility to be real-binder, and vice versa

Example output of example_output/predicted.csv

sampleID,prediction
example2,0.13840055465698242
example3,0.015093784779310226
example1,0.004906294867396355

So compound example2 would be the best and compound example1 would be the worst.

Detailed defination of all items in the CSV could be found in supplementary/experimental_data/README.md file.

Note: Due to the randomness in 3D conformation generation, molecular docking, and structure optimization, the prediction value for each compounds could change to a certain extent in each repeat runs. However, the statsitic of enrich factor would be stable in repeat runs.

Supplementary

• experiment_data

Input CSV file generated from docked structure, as well as the output data from DyScore and DyScore-MF

• train_example

The full flowchart of DyScore training is demonstrated in model_train.py

• predict_example

Example for using DyScore/DyScore-MF model with provided CSV file

Citation

If you use DyScore, please consider citing:

@article{li2022dyscore,
  title={DyScore: A boosting scoring method with dynamic properties for identifying true binders and nonbinders in structure-based drug discovery},
  author={Li, Yanjun and Zhou, Daohong and Zheng, Guangrong and Li, Xiaolin and Wu, Dapeng and Yuan, Yaxia},
  journal={Journal of Chemical Information and Modeling},
  volume={62},
  number={22},
  pages={5550--5567},
  year={2022},
  publisher={ACS Publications}
}