With SMI2GCS you can generate atomic descriptors from SMILES. The atomic descriptors are based on convolutions of CM5 atomic charges computed using semiempirical tight binding (GFN1-xTB).

More information about the method is available in the RegioML paper. Including: 1, 2, 3, 4, and 5.

We recommend using anaconda to install the Python 3 environment:

conda env create -f environment.yml && conda activate smi2gcs

Then download the binaries of xtb version 6.4.0:

mkdir dep; cd dep; wget https://github.com/grimme-lab/xtb/releases/download/v6.4.0/xtb-210201.tar.xz; tar -xvf ./xtb-210201.tar.xz; cd ..

Sort each shell according to a modified version of the Cahn-Ingold-Prelog (CIP) priority rules and the CM5 charges if CIP is unambiguous:

1. Sort according to atomic number in descending order.
2. If (1) is not unique, for each atom with the same priority (A*):
   1. Go to bound and yet not included atoms and sum up atomic numbers. Set the priority of A* according to the sum of the atomic numbers.
   2. If (2i) did not give an unambiguous result expand the shell of each atom A* by one bond.
   3. Repeat (2ii) until a unique order is found.
3. If no unique order is found in (2) and all bound atoms are included, then sort atoms according to the CM5 charges in descending order.

@article{Ree2022,
  title = {RegioML: predicting the regioselectivity of electrophilic aromatic substitution reactions using machine learning},
  volume = {1},
  ISSN = {2635-098X},
  url = {http://dx.doi.org/10.1039/D1DD00032B},
  DOI = {10.1039/d1dd00032b},
  number = {2},
  journal = {Digital Discovery},
  publisher = {Royal Society of Chemistry (RSC)},
  author = {Nicolai Ree and Andreas H. G\"{o}ller and Jan H. Jensen},
  year = {2022},
  pages = {108–114}
}