jamesgleave / Deep-Docking-NonAutomated

Deep-Docking-NonAutomated

Deep docking (DD) is a deep learning-based tool developed to accelerate docking-based virtual screening. Using a docking program of choice, one can screen extensive chemical libraries like ZINC15 (containing > 1.3 billion molecules) 50 times faster than typical docking. For further details into the processes behind DD, please refer to our paper (https://doi.org/10.1021/acscentsci.0c00229). This repository provides all the scripts required to run the DD process, except ligand preparation and molecular docking (which can be done with many licensed and/or free programs).

This version of DD is non-automated and thus requires users to use their own docking programs and ligand preparation tools.

If you use DD, please cite:

Gentile F, Agrawal V, Hsing M, Ton A-T, Ban F, Norinder U, et al. Deep Docking: A Deep Learning Platform for Augmentation of Structure Based Drug Discovery. ACS Cent Sci 2020:acscentsci.0c00229.

Requirements

• rdkit
• tensorflow >= 1.14.0 (1.15 GPU version recommended. If you are using cuda11, please use nvidia-tensorflow)
• pandas
• numpy
• keras
• matplotlib
• scikit-learn
• Ligand preparation tool
• Docking program

Help

For help with the options for a specific script, type

python script.py -h

Before Starting

You need to fill in the activation_script.sh file in phase_2-3 so that it can automatically activate your conda environment when training.

Preparing a database for Deep Docking

Databases for DD should be in SMILE format. For each compound of the database, DD requires the Morgan fingerprints of radius 2 and size 1024 bits, represented as the list of the indexes of bits that are set to 1.

First, it is recommended to split the SMILES into a number of evenly populated files to facilitate other steps such as random sampling and inference, and place these files into a new folder. This reorganization can be achieved for example with the split command in bash, and the resulting files should have .txt extensions. For example, consider a smiles.smi file with a billion compounds, to obtain 1000 evenly split .txt files of 1 million lines each we run:

split -d -l 1000000 smiles.smi smile_all_ --additional-suffix=.txt

Ideally the number of split files should be equal to the number of CPUs used for random sampling (phase 1, see below), but always larger than the number of GPUs used for inference (phase 3, see below).

Morgan fingerprints can be then generated in the correct DD format using the Morgan_fing.py script (located in the utilities folder):

python Morgan_fing.py -sfp path_smile_folder -fp path_to_morgan_folder -fn name_morgan_folder -tp num_cpus

which will create all the fingerprints and place them in path_to_morgan_folder/name_morgan_folder.

Phase 1. Random sampling of molecules

In phase 1 molecules are randomly sampled from the database to build/augment a training set. During the first iteration, this phase also samples molecules for the validation and test sets.

Runing phase 1

To run phase 1, run the following sequence of scripts to randomly sample the database, and to extract Morgan fingerprints and SMILES of the sampled molecules:

python molecular_file_count_updated.py -pt project_name -it current_iteration -cdd left_mol_directory -t_pos num_cpus -t_samp molecules_to_dock
python sampling.py -pt project_name -fp path_to_project_without_name -it current_iteration -dd left_mol_directory -t_pos total_processors -tr_sz train_size -vl_sz val_size
python sanity_check.py -pt project_name -fp path_to_project_without_name -it current_iteration
python Extracting_morgan.py -pt project_name -fp path_to_project_without_name -it current_iteration -md morgan_directory -t_pos total_processors
python Extracting_smiles.py -pt project_name -fp path_to_project_without_name -it current_iteration -smd smile_directory -t_pos num_cpus

• molecular_file_count_updated.py determines the number of molecules to be sampled from each file of the database, according to the desired number of molecules to sample. The sample sizes (per million) are stored in Mol_ct_file_updated.csv file created in the left_mol_directory directory.

• sampling.py randomly samples the desired number of molecules for the training, validation and testing sets (again note that only during the first iteration do we generate the validation and testing sets).

• sanity_check.py removes overlaps between sampled sets.

• Extracting_morgan.py and Extracting_smiles.py extract morgan fingerprints and SMILES for the compounds that have been randomly sampled, and organize them in morgan and smiles folders inside the directory of the current iteration.

IMPORTANT: For molecular_file_count_updated.py AND sampling.py the option left_mol_directory is the directory where molecules are sampled; for iteration 1, left_mol_directory is the directory storing the Morgan fingerprints of the database; BUT for subsequent iterations this must be the path to morgan_1024_predictions folder of the previous iteration.

For example, in iteration 2:

python molecular_file_count_updated.py -pt project_name -it current_iteration -cdd /path_to_project/project_name/iteration_1/morgan_1024_predictions -t_pos num_cpus -t_samp molecules_to_dock
python sampling.py -pt project_name -fp path_to_project_without_name -it current_iteration -dd /path_to_project/project_name/iteration_1/morgan_1024_predictions -t_pos total_processors -tr_sz train_size -vl_sz val_size

This will ensure that sampling is done progressively on better scoring subsets of the database over the course of DD.

After phase 1. Docking

After phase 1 is completed, molecules grouped in the smiles folder need to be prepared and docked to the target. Use your favourite workflow for this step. It is important that docking are stored as SDF files in a docked folder in the current iteration directory, keeping the same name convention of the files in the smile folder (names can be slightly changed but the name of the set (eg validation, testing, training) should always be present in the name of the respective SDF file).

Phase 2. Neural network training

In phase 2, deep learning models are trained on the docking scores from the previous phase.

Runing phase 2

Again we just need to run the following in succession:

python Extract_labels.py -if True/False -n_it current_iteration -protein project_name -file_path path_to_project_without_name -t_pos num_cpus -score score_keyword
python simple_job_models.py -n_it current_iteration -mdd morgan_directory -time 00-04:00 -file_path project_path -nhp num_hyperparameters -titr total_iterations -n_mol num_molecules --percent_first_mols percent_first_molecules -ct recall_value --percent_last_mols percent_last_mols

• Extract_labels.py extracts docking scores and organizes them to be used for model training. It should generate three comma-spaced files, training_labels.txt, validation_labels.txt and testing_labels.txt inside the current iteration folder.

• simple_job_models.py creates bash scripts to run model training using the progressive_docking.py script. These scripts are generated inside the simple_job folder in the current iteration. Note that if -ct is not specified, the recall value will be set to 0.9.

The bash scripts generated by simple_job_models.py should be then run on GPU nodes to train DD models. The resulting models will be stored in the all_models folder in the current iteration.

Phase 3. Selection of best model and prediction of the entire database

In phase 3 the models from phase 2 are evalauted and the best performing one is chosen for predicting scores of all the molecules in the database. This step will create a morgan_1024_predictions subfolder which will contain all the molecules that are predicted as virtual hits in the current iteration.

Run phase 3

To run phase 3,

python -u hyperparameter_result_evaluation.py -n_it current_iteration --data_path project_path -mdd morgan_directory -n_mol num_molecules
python simple_job_predictions.py -protein project_name -file_path path_to_project_without_name -n_it current_iteration -mdd morgan_directory

• hyperparameter_result_evaluation.py evaluates the models generated in phase 2 and select the best (most precise) one.

• simple_job_predictions.py creates bash scripts to run the predictions over the full database using the Prediction_morgan_1024.py script. These scripts will be stored in the simple_job_predictions folder of the current iteration.

The generated bash scripts can be run on GPU nodes to predict virtual hits from the full database. Predicted compounds will be stored in morgan_1024_predictions folder of the current iteration.

After Deep Docking. The final phase

After the last iteration of DD is complete, SMILES of all or a ranked subset of the predicted virtual hits can be obtained for the final docking. Ranking is based on the probabilities of being virtual hits. Use the following script (availabe in final_phase).

python final_extraction.py -smile_dir path_to_smile_dir -prediction_dir path_to_predictions_last_iter -processors n_cpus -mols_to_dock num_molecules_to_dock

Executing this script will return the list of SMILES of all the predicted virtual hits of the last iteration or the top num_molecules_to_dock molecules ranked by their probabilities, whichever is smaller. Probabilities will also be returned in a separated file.