Identifying opportunities for late-stage C-H alkylation with in silico reaction screening and high-throughput experimentation

This repository contains a reference implementation to preprocess the data train and apply the graph machine learning models for in silico reaction screening as introduced by David F. Nippa, Kenneth Atz, Alex T. Müller, Clemens Isert, Jens Wolfard, Martin Binder, Oliver Scheidegger, David B. Konrad, Uwe Grether, Rainer E. Martin & Gisbert Schneider, Journal, X, XX (2023) (see preprint).

Minisci chemistry:

Graph machine learning:

1. Environment

Create and activate the minisci environment. If you don't use a GPU, leave out the line - cudatoolkit=11.7 in envs/environment.yml.

cd envs/
conda env create -f environment.yml
conda activate minisci
poetry install

Add the "minisci path" as PYTHONPATH to your ~/.bashrc file.

export PYTHONPATH="${PYTHONPATH}:<YOUR_PATH>/minisci/"

Source your ~/.bashrc.

source `~/.bashrc`
conda activate minisci

Test your installation by running test_pyg.py.

python test_pyg.py 
>>> torch_geometric.__version__: 2.3.0
>>> torch.__version__: 1.13.1
>>> rdkit.__version__: 2022.09.5

2. Data

The data/ directory contains 7 files:

1. 20230830_minisci_pub_all_data.tsv containing all 1148 reactions published within the scope of this study,
2. 20230830_minisci_pub_decoy.tsv containing all decoy reactions,
3. 20230830_minisci_pub_hte.tsv containing the HTE data
4. 20230830_minisci_pub_hte_new.tsv containing the HTE reaction data of the validated literature reactions,
5. 20230830_minisci_pub_literature.tsv containing the literature data,
6. template_acids.tsv containing the SMILES-strings of all carboxylic acids, and
7. template_substrates.tsv cointaining the SMILES-strings of all substrates.

cd minisci/
ll data/
20230830_minisci_pub_all_data.tsv
20230830_minisci_pub_decoy.tsv
20230830_minisci_pub_hte.tsv
20230830_minisci_pub_hte_new.tsv
20230830_minisci_pub_literature.tsv
template_acids.tsv
template_substrates.tsv

3. Data preprocessing

preprocess.py loads the reaction data from data/20230830_minisci_pub_all_data.tsv, loops over the substrates, carboxylic acids and reaction conditions and calculates molecular graphs and different 3D conformations.

python preprocess.py 

The generated files for substrates, carboxylic acids and reaction conditions are stored in data/.

ll data/
lsf_rxn_carbacids_20230830_minisci_pub_all_data.h5
lsf_rxn_cond_dicts_20230830_minisci_pub_all_data.pt
lsf_rxn_conditions_20230830_minisci_pub_all_data.h5
lsf_rxn_substrate_20230830_minisci_pub_all_data.h5
rxn_smi_dict_20230830_minisci_pub_all_data.pt

3. Model training and validation

The hyperparameters are stored in config files in config/. Model training using a 75 to 25 % dataset split can be started via train.py. config_700 runs the reaction yield prediction models and config_800 the binary reaction outcome models. -cv defined the cross-validation (1-3) and -testset the test set (1-4).

cd gml/
python train.py -config 700 -mode a -cv 1 -testset 1

The training script generates two directories where the models (models/) and results (results/) are stored.

After training all models using 3 cross-validations and all 4 test sets, model performance can be analyzed via confusion_binary.py (confusion matrix binary prediction), confusion_yield.py (confusion matrix reaction yield prediction), loss_curve.py (loss curves during training), and get_outliers.py (printing out the failed ML predictions with MAE > 75%). The plots are subsequently saved in plots/.

ll analysis/
confusion_binary.py
confusion_yield.py
get_outliers.py
loss_curve.py

4. Model application and in silico reaction screening

SMILES lists for substrates and carboxylic acids can be stored in 2 .tsv files as the two examples shown in data/.

ll data/
template_acids.tsv
template_substrates.tsv

Once the two template files are defined and one model is trained, the production.py file can be run using the following parameters: -config: config file for the model; -model_postfix: model postfix, i.e., in this case, cross-validation number and test set number from training; -acids: acid template file; and -substrates: substrate template file.

cd gml/
python production.py -config 700 -model_postfix 1_2 -acids template_acids -substrates template_substrates

The prediction will be stored in prediction/ generating 5 files, i.e. the scored molecules for substrates (rxn_substrate_scoring_700.xlsx, and rxn_substrates_scoring_700.csv), the scored molecules for the carboxylic acids (rxn_acids_scoring_700.csv and rxn_acids_scoring_700.xlsx), and the predictions for the individual reactions (rxn_pred_700.csv). In rxn_pred_700.csv, every combination of substrate + carboxylic acid and their prediction is stored.

ll predictions/
rxn_acids_scoring_700.csv
rxn_acids_scoring_700.xlsx
rxn_pred_700.csv
rxn_substrate_scoring_700.xlsx
rxn_substrates_scoring_700.csv

5. License

The software was developed at ETH Zurich and is licensed by the AGPL-3.0 license, i.e. discribed in LICENSE.

6. Citation

@article{nippa_atz2023graph,
  title={Graph transformer neural network for chemical reactivity prediction},
  author={Nippa, David F and Atz, Kenneth and M{\"u}ller, Alex T and Wolfard, Jens and Isert, Clemens and Binder, Martin and Scheidegger, Oliver and Konrad, David B and Grether, Uwe and Martin, Rainer E and Schneider, Gisbert},
  year={2023},
  journal={ChemRxiv preprint 10.26434/chemrxiv-2023-8hdmv},
}
To Do: To be updated with DOI and journal.