Retro-BLEU

Retro-BLEU is designed to assess the quality of generated retrosynthesis routes. You can evaluate retrosynthesis routes either by utilizing our pre-extracted positive bigrams from the USPTO dataset or you may also extract your own bigrams

Requirement

Submodules

(These submodules are prepared for extracting bigrams from USPTO dataset. For leveraging Retro-BLEU for evaluation, you can simply skip to the next section.)

• aizynthfinder (v3.7.0)
• PaRoutes (benchmark dataset V2)
• aizynthtrain(optional if you want to extract negative bigrams)

Please use git clone --recursive https://github.com/catalystforyou/Retro-BLEU.git to correctly load the submodules.

Enviornments

conda create -n Retro-BLEU python=3.8
conda activate Retro-BLEU
pip install reaction-utils
conda install rdkit -c conda-forge
pip install rxnmapper

Evaluating Retrosynthesis Routes

Please ensure that the routes are saved in SMILES format within a JSON file. You can use test_routes/routes_retrograph.json as a reference. To test your routes, execute the calc_retrobleu.py script and provide the path to the JSON file containing your saved routes.

Extracting Bigrams

Step 1: Preparaing the patent routes

conda activate paroutes-env (see submodule PaRoutes)
mkdir data
cp PaRoutes/data/download_data.py data/
cd data
python download_data.py
cp uspto_template_library.csv ../PaRoutes/data
cd ../PaRoutes/setup
python extract_uspto_data.py --template_library ../data/uspto_template_library.csv --all_reactions ../data/selected_reactions_all.csv
python extract_routes.py --max-workers 24
python analyze_routes.py
cd ../../preprocess
python convert_route.py

Step 2: Extracting reaction templates for patent routes

cd preprocess
conda activate Retro-BLEU
python extract_templates.py

(if you want to examine template on other radii for Retro*-190, run the commands below)

sh extract_training_template.sh
python merge_train_templates.py

Step 3: Performing single-step predictions

cd preprocess
conda activate aizynth-dev (see submodule aizynthfinder)
python single_step.py

Step 4: Generating retrosynthesis routes for set-n1 & n5 via aizynthfinder

You may use the script retrosynthesis.py in preprocess/regen_routes/ and you can modify the config file and targets to predict. (There is a generated version under ../routes/{n5routes, n1routes}/{mcts_v2, retrostar_v2} and you don't actually need to regenerate the routes.)

Step 5: Evaluating bigram overlap

cd preprocess
python extract_first_routes.py

Then you may use the evaluate_algorithms.ipynb to examine different retrosynthesis searching algorithms across PaRoutes & Retro*-190.

Step 5: Selecting the patent-extracted route from set-n1 & set-n5

python generate_vocab_bigrams.py
python calc_acc.py