Reaxys data was downloaded as an RDFile, which was filtered to only contain formal borylation reaction using borylation_cleanup.py.
Where the script failed due to presence of Markush structures, those were removed using markush_remover.py.
SMILES strings can be processed in the same manner using functions from borylation.py.
To convert RDFiles into CSV files rdf_to_csv.py was used.
Automated resolution of reagents, catalysts, and solvents was performed using csv_and_discard.py. Any entry with at least one unresolved chemical entity was discarded.
As an example of SMILES data preparation for T5Chem model, see roche_preprocessing.ipynb.
For the anatomy of the dataset, see data/roche
Setup, training, and prediction instructions are provided at https://yzhang.hpc.nyu.edu/T5Chem/tutorial.html or called using t5chem {train, predict} -h upon successful installation.