DrugANNs

Global AI Challenge solution Overall pipeline:

• data preprocessing (removed unneeded parts of molecules)
• generated Morgan, MACCS and Estate fingerprints
• applied MolCLR graph neurla network
• applied RandomForest to the features described before
• the models' results were merged and averaged
• the results from the previous point were also passed to the Lipinski rule checker

Repository structure

• notebooks - contains all the notebooks which were used during the analysis
• data - folder with all the data we used
• MolCLR - directory with MolCLR model
• YouGraphRF - directory with random forest model

Model running

1. Run data_preprocessing.ipynb to make canonical SMILES
2. Run ogb-rdk-transform.ipynb to get preprocessed dataset
3. Go to YouGraphRF and run python random_forest.py --smiles_file ... --smiles_test_file ...
4. Take predictions from rf_preds/rf_final_pred.npy
5. Go to MolCLR
6. Place preprocessed molecules data to data/covid/COVID.csv and data/covid/COVID-test.csv for train and test subsets correspondingly.
7. Run python finetune_contrast.py
8. Finally, run predict-molclr.ipynb. You need to change model path with your checkpoint. Or you can find checkpoint used for submission in finetune folder
9. The final predictions should be passed to lipinski_rule_application.ipynb

Requirements

You can find the requirements in requirements.txt file