marcossantanaioc / De_novo_design_SARSCOV2

Notebooks and files from the paper De novo Design and Bioactivity Prediction of SARS-CoV-2 Main Protease Inhibitors using Recurrent Neural Network-based Transfer Learning

Supplementary information

Notebooks and files from the paper "De novo Design and Bioactivity Prediction of SARS-CoV-2 Main Protease Inhibitors using Recurrent Neural Network-based Transfer Learning"

data/

Generative models data

• AID_1706_misc_combined_processed.zip: data used to train the generative model and classifier

• ChEMBL_v1.zip: ChEMBL data used to train the general chemical model

Classifier data

• external_validation_mpro_xchem.csv: external validation set of fragments screened against SARS-CoV-2 Mpro
• random_splits.tar.gz: Random splits used to train the classifier
• scaffold_splits.tar.gz: Scaffold-based split used to train the classifier

notebooks/

ULMFit notebook

• LSTM_generative.ipynb: notebook used for ULMFit.

models/

General chemical model

• general_model_weights.pth: Pretrained weights of the general chemical model.
• vocab.pkl: Chemical vocabulary of the general chemical model. This vocabulary is necessary to fine-tune the model.

Fine-tuned model

• finetuned_encoder.pth: Encoder (all layers except the last one) of the fine-tuned chemical model
• finetuned_model.pkl: Fine-tuned chemical model
• classifier.pkl: Final fine-tuned classifier

Chemprop models

• SARS.zip: Chemprop model as available at the original website
• SARS_balanced.zip: Balanced Chemprop available at the original website

results/

Generative models data

• crossvalidation_results.csv: Cross-validation results for validation, test and external sets using the ULMFit method.
• generated_70k_90.csv: Generated molecules using the fine-tuned chemical model (N = 70000).
• top20_druglike.smi: Top-20 drug-like generated molecules.