README.md

DrugRepurposing

This repo is for the paper "Eravacycline, an Antibacterial Drug, Repurposed for Pancreatic Cancer Therapy: Insights from a Molecular-Based Deep Learning Model" which contains the code to train a drug reporpusing model for cancer treatments

Data

The data is located under the data folder:

1. labels_training_set_w_drugbank_id.csv- contains the data collated manually of drugs that are classified as anti-cancer and those that have indication to not have anti-cancer activity
2. all_data_infer_labels_preds_w_drug_bank_info_no_features.csv- all drugs in DrugBank excluding those that were used to train the model

Code

The anti-cancer model was train using the chemprop library: https://github.com/chemprop/chemprop/

To reproduce the anti-cancer without addtional features run the following command from the command-line:

chemprop_train --data_path ./data/labels_training_set_w_drugbank_id.csv \
--dataset_type classification \
--number_of_molecules 1 --smiles_columns Smiles \
--metric auc --extra_metrics prc-auc accuracy mcc \
--save_preds --save_smiles_splits \
--config_path ./data/full_data_hyperparams_w_rkdit.json \
--loss_function binary_cross_entropy \
--split_sizes 0.7 0.1 0.2 --target_columns cancer \
--split_type scaffold_balanced --save_dir ./model \
--num_folds 3 --ensemble_size 2 \

or use code/multimodal_learning/target_main.py to train via python code

To reproduce the anti-cancer with additional features run the multimodal learning code using the code/multimodal_learning/interactions_main.py to train a DDI model then train the chemprop model using code/multimodal_learning/target_main.py with the task cancer`` and uncomment lines 111 and 112 to add the addtional features

or from the commandline

chemprop_train --data_path ./data/labels_training_set_w_drugbank_id.csv \
--dataset_type classification \
--number_of_molecules 1 --smiles_columns Smiles \
--metric auc --extra_metrics prc-auc accuracy mcc \
--save_preds --save_smiles_splits \
--config_path ./data/full_data_hyperparams_w_rkdit.json \
--loss_function binary_cross_entropy \
--split_sizes 0.7 0.1 0.2 --target_columns cancer \
--split_type scaffold_balanced --save_dir ./model_DDI_DTI \
--num_folds 3 --ensemble_size 2 \
--features_path ./data/labels_training_set_w_drugbank_id_DDIFeature_256.csv ./data/labels_training_set_w_drugbank_id_TargetPCAFeature_64.csv

Predict with chemprop model

to predict with the chemprop model, please refer to https://github.com/chemprop/chemprop/#predicting on how to use saved trained models