/DTI_PDBbind

Primary LanguagePython

Code for the paper: "PIGNET: A physics-informed deep learning model toward generalized drug-target interaction predictions" by Seokhyun Moon, Wonho Zhung, Soojung Yang, Jaechang Lim, Woo Youn Kim

Table of Contents

Install Dependencies

PIGNet needs conda environment. After installing conda,

you can manually install the packages for our code. The package list is shown as follows

  • rdkit
  • pytorch
  • numpy
  • biopython
  • ase
  • anaconda scikit-learn
  • scipy
  • smina

or you can just execute the followings in the command line.

./dependencies
conda activate pignet

Running

Data Preprocessing

To prepare our dataset for train and test, change directory to data, and follow the instructions.

Train

Important: Default values of the rest of the arguments are set as the best parameters, which generated our results in the paper. Also, set interaction_net argument as True. It will affect the model enormously.

PIGNet uses four dataset: original(data_dir), docking(data_dir2), random_screening(data_dir3), and cross_screening(data_dir4), and each dataset has three arguments for train.

  • data_dir: Directory consists of the preprocessed pickle data.
  • key_dir: Directory consists of the keys for pickle data.
  • filename: A path of the text file which consists of complex key and its binding affinity.

During the training, the results will be written at following arguments:

  • train_result_filename
  • test_result_filename
  • train_result_docking_filename
  • test_result_docking_filename
  • train_result_screening_filename
  • test_result_screening_filename

To train with the uncertainty, add the following arguments:

  • dropout_rate: Set this argument as 0.2 unlike the train without uncertainty which dropout_rate is 0.1.
  • with_uncertainty: Flag of using uncertainty during train.
  • mc_dropout: Should set as True to use Monte-Carlo dropout.

We also realized 3D CNN model of KDEEP. To train 3D CNN model, you can use following arguments with the default values:

  • potential: Select a model to train. Default value is "harmonic".
  • grid_rotatin: Whether rotate the grid or not during the train.
  • lattice_dim: Size of the 3D lattice.
  • scaling: Interval of the lattice points.

Model benchmark with CASF2016 benchmark and csar

Important: To benchmark the model with CASF2016 benchmark, you should prepare CASF-2016, csar1 and csar2 data in data directory first.

Inside the benchmarks directory, execute ../test.py with the corresponding test datasets. We did several benchmark study with CASF-2016 and csar datasets. To test the model with specific dataset, give three arguments for test dataset.

  • data_dir: Directory consists of the preprocessed pickle data.
  • key_dir: Directory consists of the keys for pickle data.
  • filename: A path of the text file which consists of complex key and its binding affinity.

The test result will be written at following arguments:

  • test_result_filename

To test the model with uncertainty, add the following arguments:

  • with_uncertainty: Flag of model trained with or without uncertainty.
  • n_mc_sampling: Number of the Monte-Carlo sampling.

Important: For different test dataset, executing ../test.py code will generate several different result files. Followings are the ways to get the score for each benchmark. Execute the commands at the benchmarks directory. To get the R score, you should forward the std output to output_file for each benchmark.

  • Csar1

You can just execute csar1_test.sh at the benchmarks/csar1 directory. To get the R value from the forwarded csar1_output_file, just execute following:

grep 'R:' {csar1_output_file}*
  • Csar2

You can just execute csar2_test.sh at the benchmarks/csar2 directory. To get the R value from the forwarded csar2_output_file, just execute following:

grep 'R:' {csar2_output_file}*

Important: To test our save file(save/save_1000.pt), you should set {epoch} as 1000.

  • Scoring power

You can just execute scoring_test.sh at the benchmarks/scoring directory. Then, execute the following command, where scoring_result_file is argument value of test_result_filename of test.py.

python ../../casf2016_benchmark/scoring_power.py {scoring_result_file} {epoch}
  • Ranking power

Ranking power uses scoring_result_file. Just execute the following command, where scoring_result_file is argument value of test_result_filename of test.py.

python ../../casf2016_benchmark/ranking_power.py {scoring_result_file} {epoch}
  • docking power

You can just execute docking_test.sh at the benchmarks/docking directory. Then, execute the following command, where docking_result_file is argument value of test_result_filename of test.py.

python ../../casf2016_benchmark/docking_power.py {docking_result_file} {epoch}
  • screening power

To compute the screening power, you should iterate keys from 0 to 99 as shown in benchmarks/screening/screening_test.sh. Then, execute the following commands inside benchmarks/screening directory.

cat result_* > total_result.txt
python ../../casf2016_benchmark/screening_power.py total_result.txt {epoch}

Command Examples

train.sh in this directory is the code that we used for training the model.

Train

python -u train.py \
--save_dir=save \
--tensorboard_dir=run \
--train_result_filename=result_train.txt \
--test_result_filename=result_test.txt \
--train_result_docking_filename=result_train_docking.txt \
--test_result_docking_filename=result_test_docking.txt \
--train_result_screening_filename=result_train_screening.txt \
--test_result_screening_filename=result_test_screening.txt \
--data_dir={original data dir} \
--filename={original data file path} \
--key_dir={original data key dir} \
--data_dir2={docking data dir} \
--filename2={docking data file path} \
--key_dir2={docking data key dir} \
--data_dir3={random_screening data dir} \
--filename3={random_screening data file path} \
--key_dir3={random_screening data key dir} \
--data_dir4={cross_screening data dir} \
--filename4={cross_screening data file path} \
--key_dir4={cross_screening data key dir} \
--potential='harmonic' \
--interaction_net \
> output 2> /dev/null

Train With Uncertainty (Just add to train command)

--dropout_rate=0.2
--train_with_uncertainty
--mc_dropout=True

Test

test.sh in benchmarks directory is the basic code that we used for training the model.

We use casf2016_scoring, casf2016_ranking, casf2016_docking, casf2016_screening, csar1, csar2 to benchmark the model. Use different {benchmark name}, {benchmark data dir}, {benchmark data file path}, and {benchmark data key dir} for each benchmark, in the following code example.

Also, for the casf2016_docking and casf2016_screening, we recommend to use ngpu=1 option.

python -u ../test.py \
--batch_size=64 \
--num_workers=0 \
--restart_file=../save/save_1000.pt \
--n_gnn=3 \
--dim_gnn=128 \
--test_result_filename=result_{benchmark name}_harmonic_1000 \
--ngpu=0 \
--interaction_net \
--potential="harmonic" \
--data_dir={benchmark data dir} \
--filename={benchmark data file path} \
--key_dir={benchmark data key dir} \
> test_{benchmark name}_harmonic_1000

Citing this work

@article{moon2020pignet,
    title={PIGNet: A physics-informed deep learning model toward generalized drug-target interaction predictions},
    author={Moon, Seokhyun and Zhung, Wonho and Yang, Soojung and Lim, Jaechang and Kim, Woo Youn},
    journal={arXiv preprint arXiv:2008.12249},
    year={2020}
}