Perception of Chemical Bonds via Machine Learning
Preprints are manuscripts made publicly available before they have been submitted for formal peer review and publication. They might contain new research findings or data. Preprints can be a draft or final version of an author's research but must not have been accepted for publication at the time of submission.
An approach based on machine-learning is presented that is able to identify chemical bond types such as single, double, triple and aromatic bonds based on spatial atomic coordinates only, as provided for example from quantum chemical calculations or from crystallographic data.
The basic idea behind this work is to exploit the various chemical knowledge already assembled in molecular data files in form of connection tables and bond blocks. Rules for novel chemistry or particular functional groups can be learned automatically by training on structure data (.sd/.sdf) files with the respective bond information. Provided that the underlying database is sufficiently large and diverse, the approach is able to identify chemical bond orders in molecules with an accuracy comparable to classical bond-perception tools using hard-coded rules and cheminformatic algorithms. The workflow is implemented in Python using the open source packages RDKit, scikit-learn and pandas (https://github.com/CHLoschen/mamba).