Predictive model for the Aromatic ThianThrenation reaction for C-H functionalization (PATTCH)

PATTCH is a hybrid SQM/ML workflow for the prediction of the reaction feasibility and site-selectivity of the direct thianthrenation reaction of aromatic C-H bonds.
For further information, please refer to our preprint or publication.

L. M. Sigmund, T. Seifert, R. Halder, G. Bergonzini, M. J. Johansson, P.-O. Norrby, K. Jorner, and M. Kabeshov, Predicting Reaction Feasibility and Selectivity of Aromatic C─H Thianthrenation with a QM–ML Hybrid Approach, Angew. Chem. Int. Ed. 2025, e202510533, DOI: doi.org/10.1002/anie.202510533.

Installation

1) Installation of PATTCH

Clone the repository.

git clone https://github.com/MolecularAI/thianthrenation_prediction.git

Create a conda environment with the provided pattch_env.yaml file and install the pattch package into the environment.

conda env create -f pattch_env.yml -n pattch_env
conda activate pattch_env
pip install .

After the installation, change line 34 in handler.py of the qmdesc dependency to make it compatible with Pytorch 2.
From

        state = torch.load(stream, lambda storage, loc: storage)

to

        state = torch.load(stream, lambda storage, loc: storage, weights_only=False)

2) Installation of Multiwfn

The PATTCH workflow requires the noGUI version of Multiwfn as a mandatory dependency. Go through the following steps to install the program (if not already installed). Additional information on the installation can be found in section 2.1 of the Multiwfn manual.
a) Dowload the Multiwfn_3.8_dev_bin_Linux_noGUI.zip file from this webpage.
b) Unzip the folder and place it in an appropriate directory which can be accessed by PATTCH.
c) Ensure that the downloaded Multiwfn_noGUI file has executable permission. If not, this can be added by changing the current working directory to where Multiwfn was saved followed by executing the following command.

chmod +x Multiwfn_noGUI

d) Add the following lines to your .bashrc file. Replace <path to parent directory> with the correct path. This will make Multiwfn available to the command line, regardless of the current working directory.

export Multiwfnpath=<path to parent directory>/Multiwfn_3.8_dev_bin_Linux_noGUI
export PATH=$PATH:<path to parent directory>/Multiwfn_3.8_dev_bin_Linux_noGUI

e) Open a new terminal and run the program (for testing without any input file) by executing the following command.

Multiwfn_noGUI

This should result in output printed to the terminal containing the following text.

Multiwfn -- A Multifunctional Wavefunction Analyzer
Version 3.8(dev), release date: 2023-Oct-8
Developer: Tian Lu (Beijing Kein Research Center for Natural Sciences)
Below paper ***MUST BE CITED*** if Multiwfn is utilized in your work:
         Tian Lu, Feiwu Chen, J. Comput. Chem., 33, 580-592 (2012)

Usage

Set environment variables before running the model by executing the following commands in the shell.

export OMP_STACKSIZE=3G
export OMP_NUM_THREADS=1
export OMP_MAX_ACTIVE_LEVELS=1
export MKL_NUM_THREADS=1

After activating the environment, the pattch package can be used as follows by simply specifying the input *.csv file (-i). The input file must at least contain two columns labeled substrate_name (identification string for each compound) and substrate_smiles (SMILES string of each compound). For testing, you can run the fullowing commands.

cd examples/sample_run
python -m pattch -i sample_input.csv

The exact run type of the PATTCH workflow can be configured through the -i, -p, -a and -x flags. Run the following command to see their functionality.

python -m pattch --help

The prediction results can be found in the prediction_results folder. Output files generated along the prediction pipeline are saved to the output_files folder.