A database of reaction properties containing the atom-mapped reaction SMILES and the target properties. This repository belongs to the manuscript Machine learning of reaction properties via learned representations of the condensed graph of reaction, a detailed description of the datasets and their respective preprocessing is available in the article.
All reactions can be found in the data folder as CSV files.
Data extracted from DOI:10.1088/2632-2153/aba822, CCSD level of theory. Atom-mapped manually.
Data extracted from DOI:10.1038/s41467-020-19267-x. Unbalanced reactions were discarded. Atom-mapped using Grzybowski's algorithm. Only forward reactions are given in the CSV file (the manuscript also includes reverse reactions).
Data extracted from DOI:10.1021/acs.jpca.7b07361. Atom-mapped using Grzybowski's algorithm.
Data extracted from DOI:10.1073/pnas.1423570112, atom-mapped manually. Only substrates with a single phosphate group were kept. One-hot encoding of the enzymes given in the same order as the Reaction SMILES.
To reproduce the exact results in the manuscript, the folder data_splits holds the exact data splits into training, validation and test set. Per database, five folds of splits are reported. The folder models reports the CGR GCNN (default hyperparameters) models for each dataset, one for each data split. Unpack the directories you plan to use. Note that for USPTO-1K-TPL, the individual folds are archieved instead of the full folder to limit the filesize. The data_splits and models folders also hold splits for three databases not reported in the data folder, namely:
Data extracted from DOI:10.1038/s41597-020-0460-4, at the ωB97X-D3/def2-TZVP level of theory (and B97-D3/def2-mSVP for pretraining).
Data extracted from DOI:10.1039/D0SC04896H.
Data extracted from DOI:10.1038/s42256-020-00284-w.
To predict with one of the models, for example to utilize the first model for the first data split of the lograte database, install Chemprop and run:
chemprop_predict --test_path data_splits/lograte/fold_0/aam_test.csv --checkpoint_path models/lograte/fold_0/model.pt --preds_path preds_lograte_0.csv
to save the predictions to the file preds_lograte_0.csv (or choose any other file name). To use all five models for an ensemble prediction on your own set of reactions (provided as atom-mapped SMILES string, refer to the manuscript or the files in data_splits whether to include explicit hydrogens), use:
chemprop_predict --test_path <path to your csv> --checkpoint_dir models/lograte/ --preds_path <path to save preds>
The models snar and phosphatase furthermore require the input of molecular features, refer to the manuscript for an in-detail explanation. To predict with the first model on the first data split of the snar database, run:
chemprop_predict --test_path data_splits/snar/fold_0/aam_test.csv --features_path data_splits/snar/fold_0/features_test.csv --checkpoint_path models/snar/fold_0/model.pt --preds_path preds_snar_0.csv
Copyright (c) 2021, Esther Heid. For individual licenses, please refer to the respective manuscripts.