ERRalpha-Predictor

A Framework of Ensemble Models for Predicting ERRα Binders, Antagonists, and Agonists Using Artificial Intelligence. You can use it simply at https://err1p-web.streamlit.app

Overview:

• datasets/ contains the three dataset files ("xxx_train" used for 8:1:1 data splitting, "xxx_external" used for external validation);
• ensemble_models/ take the external validation set of three datasets as examples to use ERRα-Predictor;
• graph_models/ contains the final selected GNN models for the three datasets;
• ml_final_models/ contains the final selected ML models for three datasets;
• mmpa_and_representative_substructure/ contains reference repository code
• streamlit/ contains local web deployment of ERRα-Predictor;
• supplementary_information/ contains supplementary materials;
• ml_best_hyperparameters.pkl optimal hyperparameters for ML models;
• best_hyperparameters.json optimal hyperparameters for GNN models;
• GNNExplainer_viz.ipynb visualization of GNN models

Conda environments:

python = 3.7
deepchem = 2.7.1
imbalanced-learn = 0.7.0
lightgbm = 4.5.0
numpy = 1.21.6
optuna = 4.0.0
pytorch = 1.13.1
pytorch-cuda = 11.7
rdkit = 2023.3.2
scikit-learn = 1.0.2
torch-geometric = 2.3.1
xgboost = 1.6.2
mol2vec = 0.2.2
openpyxl = 3.1.3
streamlit = 1.23.1
streamlit-ketcher = 0.0.1

Experiments:

ML models

You need to download the mol2vec pkl file from here first, then for the three datasets, run the machine learning codes molecular_representation.py and model_ml_construct.py, get representation and model training results respectively.

GNN models

Given the json file of the best hyperparameters, you only need to run model_graph_merge_hyper.py; if you want to try this hyperparameter exploration process, run graph_hyper_tune.py.

Ensemble models

You can look at the examples in ensemble_models：for binders, antagonists, and agonists, ERRα-Predictor was used to predict them respectively.

Usage:

If you want to use ERRα-Predictor, after configuring the required Python environment and downloading the mol2vec pkl file, it is recommended that you enter the streamlit for local web deployment.

Use the command:

cd streamlit
streamlit run app.py

The web page incorporates functionalities for data retrieval and prediction. The former involves querying the collected dataset of ERRα ligands, while the latter utilizes ERRα-Predictor models to perform predictions.
Users can choose to submit either a single SMILES string or multiple SMILES strings for querying or prediction, and the results can be downloaded.