The covalent bond type (e.g. single, double, aromatic, delocalized, etc) is a fundamental chemical concept whose evaluation is necessary in a vast array of common applications, such as calculating formal charges or constructing arrow-pushing diagrams to describe chemical reactions. In the cheminformatics and machine learning spaces, they are commonly used as descriptors to encode molecular structure information. Although determining bond types in organic molecules is trivial, their evaluation in metal-containing periodic structures, given only the atomic coordinates, can be challenging. For this, the tools in the CSD Python API are widely used. However, this library is proprietary, requiring paid licensing to be used, and is known to produce poor and inconsistent results for organometallic compounds and metal-organic frameworks (MOFs).

We are developing an edge-centric message-passing graph neural network called GraNMa (GRAph Neural network MAchine learned) Bond Typer to compute the covalent bond types in molecular and periodic structures containing metal-ligand bonds, such as organometallic compounds and MOFs, using only the 3D atomic coordinates. Current efforts are directed towards refining the model and improving bond type assignment in cases where the CSD Python API is known to be incorrect, such as metal-coordinating heterocycles. Ultimately, this metal-aware machine learning approach will allow GraNMa Bond Typer to be an open-source replacement for the CSD Python API bond typer.