chem-dl-rxn

This repository containing [1] training deep learning models [2] a flask inference server.

We use the similar methods as these publications:

• Retrosynthesis
• Reaction Prediction

1. Training deep learning models

we use pytorch and fairseq for the transformer model.

$ cd training_forward_retro

Then, please follow steps in:

1. training_forward_retro/INSTALL_TRAINING.md for installation
2. training_forward_retro/README.md for training reaction prediction
3. training_forward_retro/README-RETRO.md for training retrosynthesis

2. Run a flask inference server

Please follow steps in:

1. INSTALL_SERVER.md for installation

3. Some detail

Dataset of reaction prediction

• USPTO_STEREO dataset, mixed starting materials and reactants SMILES as the model input; products SMILES as the model output.
• Train/Valid/Test = 902K / 50K / 50K

Dataset of retrosynthesis

• USPTO_STEREO dataset, products SMILES as the model input; starting materials SMILES as the model output.
• Train/Valid/Test = 902K / 50K / 50K

Accuracy of reaction prediction = 72.2%

Accuracy of retrosynthesis = 41.2%