A Modelling Framework for Embedding-based Predictions for Compound-Viral Protein Activity

The guidelines to setup the environment and all the packages required to run our framework is available in installation_guide.md file

Here we provide the scripts, data, model and results for predicting compound-viral activity scores using our pipeline.

Details about how to obtain the training and test set for the predictive models is provided in the data folder.

We consider the SARS-COV-2 virus as a use-case and provide ranked list of compounds based on the 3 main proteases:

a) 3CL-pro
b) PL-pro
c) Spike protein

An installation guide ("installation_guide.txt") is included in this repository which should complete all system requirements.

This package contains eight individual machine learning models. The four traditional machine learning models are as follows:

a) GLM - scripts/supervised_glm_on_ls_protein_compound.py
b) Random Forests - scripts/supervised_rf_on_ls_protein_compound.py
c) SVM - scripts/supervised_svm_on_ls_protein_compound.py
d) XGBoost - scripts/supervised_xgb_on_ls_protein_compound.py

How to run:

• cd scripts

• python supervised_<method>_on_ls_protein_compound.py Train_Compound_Viral_interactions_for_Supervised_Learning_with_<compound_features>_LS.csv <type>_Compound_Viral_interactions_for_Supervised_Learning_with_<compound_features>_LS.csv <method>_<compound_features>_Compound_LS_Protein_supervised_<type>_predictions.csv

Here <method> can be either glm, rf, svm, xgb, <compound_features> can be either MFP or LS and <type> can either Test or sars_cov_2.

The files for training and testing (Test or sars_cov_2) are produced in data folder by following the instructions in the README available in the data folder.

Outputs:

a) GLM - results/GLM_MFP_Compound_LS_Protein_supervised_Test_predictions.csv, results/GLM_LS_Compound_LS_Protein_supervised_Test_predictions.csv, results/GLM_supervised_sars_cov2_predictions.csv

b) RF - results/RF_MFP_Compound_LS_Protein_supervised_Test_predictions.csv, results/RF_LS_Compound_LS_Protein_supervised_Test_predictions.csv and results/RF_supervised_sars_cov2_predictions.csv

c) SVM - results/SVM_MFP_Compound_LS_Protein_supervised_Test_predictions.csv, results/SVM_LS_Compound_LS_Protein_supervised_test_predictions.csv and results/SVM_supervised_sars_cov2_predictions.csv
d) XGB - results/XGB_MFP_Compound_LS_Protein_supervised_Test_predictions.csv, results/XGB_LS_Compound_LS_Protein_supervised_Test_predictions.csv and results/XGB_supervised_sars_cov2_predictions.csv

The four end-to-end deep learning models:

a) CNN - scripts/torchtext_cnn_supervised_learning.py
b) LSTM - scripts/torchtext_lstm_supervised_learning.py
c) CNN-LSTM - scripts/torchtext_cnn_lstm_supervised_learning.py
d) GAT-CNN - scripts/torchtext_gat_cnn_supervised_learning.py

Runs on test mode:

1. data/Test_Compound_Viral_interactions_for_Supervised_Learning.csv
2. data/sars_cov_2_Compound_Viral_interactions_for_Supervised_Learning.csv

How to run:

• cd scripts

• python torchtext_<method>_supervised_learning.py Train_Compound_Viral_interactions_for_Supervised_Learning.csv <type>_Compound_Viral_interactions_for_Supervised_Learning.csv <method>_supervised_<type>_predictions.csv

Here <method> can be either cnn, lstm, cnn_lstm, gat_cnn and <type> can be either Test or sars_cov_2.

Ouput Files with location:

a) CNN - results/cnn_supervised_Test_predictions.csv and results/cnn_supervised_sars_cov_2_predictions.csv
b) LSTM - results/lstm_supervised_Test_predictions.csv and results/lstm_supervised_sars_cov_2_predictions.csv
c) CNN-LSTM - results/cnn_lstm_supervised_Test_predictions.csv and results/cnn_lstm_supervised_sars_cov_2_predictions.csv d) GAT-CNN - results/gat_cnn_supervised_Test_predictions.csv and results/gat_cnn_supervised_sars_cov_2_predictions.csv

To compare the performance of the methods on the test set:
a) Install R, Librarires- ggplot2, ggthemes
b) Run R make_error_correlation_plots.R

To get a ranked list of compounds for SARS-COV-2 viral proteins:
a) Run sars_cov_2_postprocessing.py

Outputs:
a) 3CL-Pro - results/3CL_Pro_Top_Ranked_Compounds.csv
b) PL-Pro - results/PL_Pro_Top_Ranked_Compounds.csv
c) Spike - results/Spike_Pro_Top_Ranked_Compounds.csv