COVID-19 Drug Design using Generative RNN-LSTM

COVID-19 is an infectious disease caused by a newly discovered strain of coronavirus (SARS-CoV-2), a type of virus known to cause respiratory infections in humans. This new strain was unknown before December 2019, when an outbreak of a pneumonia of unidentified cause emerged in Wuhan, China.

Basic Local Alignment Search Tool (BLAST) results show close homology to the bat Coronavirus. A crystal structure of the main protease of the virus was obtained by Liu et al., found at https://www.rcsb.org/structure/6LU7

Since the outbreak, researchers have been collaborating and working closely to stop the spread of the disease and to propose possible treatment plans. New advances in machine intelligence have introduced algorithms that can learn important patterns from vast amounts of data, approaching expert-level of ability in some tasks. This means that anyone with these models can contribute to the global research effort.

This project uses many ideas and implementations developed by others, and bring them together towards a common task. My main reference was Topazape's repo which implements the paper Generative Recurrent Networks for De Novo Drug Design.

The aim of this project is to find drug candidates (ligand) with a high binding affinity with the COVID-19 main protease using deep learning.

1. Outline of the problem and introduction

2. Dataset preparation

3. Train LSTM-based RNN model

4. Generate SMILES strings

5. Use transfer learning to fine-tune model, generating molecules that are structurally similar to potential protease inhibitors of COVID-19

6. Use PyRx to get binding scores of molecules with SARS-CoV-2 main protease

7. Report highest scoring candidates

Youtube video for this project

https://www.youtube.com/watch?v=EPkQrNHKIX8&t=3s

Requirements

This model is built using Python 3.7, and utilizes the following packages;

• numpy
• pandas
• tensorflow
• tqdm
• Bunch
• matplotlib
• RDKit
• scikit-learn

Dataset Preparation

Datasets from two sources: i) Moses data set and ii) ChEMBL data set were combined. Together these two data sets represent about 2.5 million smiles.

Preprocess dataset to remove duplicates, salts, stereochemical information, nucleic acids and long peptides.

In terminal, cd to the file and run python cleanup_smiles.py datasets/all_smiles.txt datasets/all_smiles_clean.txt

After cleaning the smiles using the cleanup_smiles.py script and only retaining smiles between 34 to 128 characters in length, './datasets/all_smiles_clean.txt' contains the final list of 180793 smiles on which the model was trained.

Potential COVID-19 protease inhibitors were included for model fine-tuning using transfer learning

According to this paper - Binding site analysis of potential protease inhibitors of COVID-19 using AutoDock

SMILES obtained from PubChem

Protease Inhibitor	SMILES
Remdesivir	CCC(CC)COC(=O)C(C)NP(=O)(OCC1C(C(C(O1)(C#N)C2=CC=C3N2N=CN=C3N)O)O)OC4=CC=CC=C4
Nelfinavir	CC1=C(C=CC=C1O)C(=O)NC(CSC2=CC=CC=C2)C(CN3CC4CCCCC4CC3C(=O)NC(C)(C)C)O
Lopinavir	CC1=C(C(=CC=C1)C)OCC(=O)NC(CC2=CC=CC=C2)C(CC(CC3=CC=CC=C3)NC(=O)C(C(C)C)N4CCCNC4=O)O
Ritonavir	CC(C)C1=NC(=CS1)CN(C)C(=O)NC(C(C)C)C(=O)NC(CC2=CC=CC=C2)CC(C(CC3=CC=CC=C3)NC(=O)OCC4=CN=CS4)O
Darunavir	CC(C)CN(CC(C(CC1=CC=CC=C1)NC(=O)OC2COC3C2CCO3)O)S(=O)(=O)C4=CC=C(C=C4)N
Atazanavir	CC(C)(C)C(C(=O)NC(CC1=CC=CC=C1)C(CN(CC2=CC=C(C=C2)C3=CC=CC=N3)NC(=O)C(C(C)(C)C)NC(=O)OC)O)NC(=O)OC

These protease inhibitors SMILES were added into datasets/protease_inhibitors_for_fine-tune.txt

Train LSTM-based RNN model to generate SMILES

Configuration

See config.json in base_experiment.

parameters	meaning
exp_name	experiment name (default: LSTM_Chem)
data_filename	filepath for training the model (SMILES file with newline as delimiter)
data_length	number of SMILES for training. If you set 0, all the data is used (default: 0)
units	size of hidden state vector of two LSTM layers (default: 256, see the paper)
num_epochs	number of epochs (42)
optimizer	optimizer (default: adam)
seed	random seed (default: 71)
batch_size	batch size (default: 512)
validation_split	split ratio for validation (default: 0.10)
varbose_training	verbosity mode (default: True)
checkpoint_monitor	quantity to monitor (default: val_loss)
checkpoint_mode	one of {auto, min, max} (default: min)
checkpoint_save_best_only	the latest best model according to the quantity monitored will not be overwritten (default: False)
checkpoint_save_weights_only	If True, then only the model's weights will be saved (default: True)
checkpoint_verbose	verbosity mode while ModelCheckpoint (default: 1)
tensorboard_write_graph	whether to visualize the graph in TensorBoard (defalut: True)
sampling_temp	sampling temperature (default: 0.75, see the paper)
smiles_max_length	maximum size of generated SMILES (symbol) length (default: 128)
finetune_epochs	epochs for fine-tuning (default: 12, see the paper)
finetune_batch_size	batch size of finetune (default: 1)
finetune_filename	filepath for fine-tune the model (SMILES file with newline as delimiter)

Docking procedure with PyRx:

Download here: https://pyrx.sourceforge.io

PyRX ligand docking tutorial https://www.youtube.com/watch?v=2t12UlI6vuw

1. Open the structure of the protein and ligand complex (.cif crystallographic information file) https://www.rcsb.org/3d-view/6LU7
2. Select the ligand chain, delete the ligand, and save the file as a .pdb
3. Process generated SMILES and save it as .sdf file
4. Follow the video tutorial to get binding scores and save it as a csv file

Final Results

Ligand	Binding Affinity (kcal/mol)
Lopinavir	-6.9
Generated_3	-6.8
Generated_5	-6.3
Darunavir	-6.1
Generated_1	-6.0
Generated_6	-5.7
Generated_2	-5.5
Nelfinavir	-5.5
Atazanavir	-5.4
Remdesivir	-5.3
Ritonavir	-5.3
Generated_4	-5.2

COVID-19 Main Protease (6LU7) in complex with Generated_3 Molecule

References

Generative Recurrent Network for De Novo Drug Design https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5836943/ https://github.com/topazape/LSTM_Chem

Binding site analysis of potential protease inhibitors of COVID-19 using AutoDock https://link.springer.com/article/10.1007/s13337-020-00585-z

PubChem data related to COVID-19 https://pubchemdocs.ncbi.nlm.nih.gov/covid-19

Refer here for COVID-19 drugs in clinical trials https://pubchem.ncbi.nlm.nih.gov/#tab=compound&query=covid-19%20clinicaltrials

Crystal structure of COVID-19 main protease https://www.rcsb.org/structure/6LU7

RDKit https://www.rdkit.org/docs/GettingStartedInPython.html

https://github.com/topazape/LSTM_Chem

https://github.com/forkwell-io/fch-drug-discovery

https://github.com/mattroconnor/deep_learning_coronavirus_cure

https://github.com/tmacdou4/2019-nCov

https://dash-gallery.plotly.host/dash-drug-discovery/