To install the environment with plenty of uneccessary packages (the output of conda env export --no-build --name retrogfn), run:
conda env create -f environment_full.yaml
To install the minial set of packages extracted with pipreqs, run:
conda create -n python=3.11.4 retrogfn pip
conda activate retrogfn
pip install -r requirements.txt
Checkpoints can be downloaded from here and should be placed in the checkpoints directory.
See notebooks/example.ipynb.
To prepare the training datasets, run the following notebooks under notebooks/created_dataset:
create_positive.ipynb. It removes the atom mapping from the raw USPTO dataset. We call this dataset "positive".extract_forward_templates.ipynb. It extracts the forward templates from the USPTO dataset.create_negative_forward.ipynb. It creates the negative reactions by applying the forward templates to reactants from the positive dataset.create_negative_shuffle.ipynb. It creates the negative reactions by shuffling the reactants from the positive dataset. A product from a positive dataset is assigned with a reactants coming from a similar (in terms of Tanimoto distance) reaction.merge_files.ipynb. It merges the positive and negative datasets into a single one.