Code for "Retrosynthetic reaction prediction using neural sequence-to-sequence models" paper

http://pubs.acs.org/doi/full/10.1021/acscentsci.7b00303

Requirements

All requirements for https://github.com/google/seq2seq

• python 3.5
• tensorflow 1.0.1

Additionally:

• numpy
• networkx
• rdkit

Model training

The data has already been preprocessed for training the seq2seq model

Model training can be started by running the google_seq2seq_train_jul_5_2017_1.sh script

Also, a pretrained model checkpoint can be found at: http://deepchem.io.s3-website-us-west-1.amazonaws.com/trained_models/jul_5_2017_1.tar.gz. Copy the folder into model_working_directory. Then from within the train_options.json file, modify the vocab_target and vocab_source flag to the actual vocab path in the processed data folder

Model inference

Model inference can be performed by running the google_seq2seq_infer_jul_5_2017_1_x.sh scripts.

The google_seq2seq_infer_jul_5_2017_1_nobeam.sh file will perform inference with no beam search. The output will be a prediction text file

The google_seq2seq_infer_jul_5_2017_1_x.sh will perform inference with beam search. The outputs will be a prediction text file, as well as a .npz that contains the raw beams which will be processed during the model evaluation

The ---checkpoint_path flag within the script should be changed to the relevant checkpoint file that you want to perform inference on.

Model evaluation

The prediction can be evaluated using reaction_evaluation.ipynb in the evaluation folder. For the no beam search predictions, need to change pred_path to the relevant prediction file. For the beam search predictions, need to change beam_width and beam_path to the relevant beam widths used for each particular .npz beam prediction file.