retrosynthesis_readme.md

Retrosynthesis

These scripts train a regressor from a molecule to the latent space. We can then use this network and the decoder of the MoleculeChef to try to predict what reactants were used to create these molecules. Even if we do not get the correct reactants we hope that the resulting product has similar properties/structure.

Steps to run.

1. Train the regressor by running: regress_product_to_latents.py
2. Run perform_retrosynthesis.py to run the test set through this regressor and reassemble reactant bags.
3. Run the MolecularTransformer https://github.com/pschwllr/MolecularTransformer to predict the products associated with these reactant bags. You can run the Transformer with the following command (in the transformer repo):

   python translate.py -model <saved_transformer_model_path> \
                       -src <tokenized_reactants_file_path> \
                       -output <output_file_name> \
                       -batch_size 300 -replace_unk -max_length 400 -fast -gpu 1 -n_best 5
   

4. Run create_retrosynthesis_plots.py to create the plots.