MeCoBi: A fast and accurate transition metal complex builder that generates 3D structures at scale and enables structure-based machine learning

The three-dimensional arrangement of ligands around a metal centre determines its electronic structure, reactivity, and biological activity, yet generating accurate three-dimensional (3D) structures from a ligand list remains difficult. Existing tools tend to be slow, narrow in coverage, or require manual specification. We present MeCoBi (Metal Complex Builder), an automated pipeline that builds and optimises 3D structures of mononuclear metal complexes from ligand SMILES across 36 d-block and main-group metals and coordination numbers 2–9. On 30,830 experimental structures from the Cambridge Structural Database, MeCoBi generated a structure for 99.6% of complexes with a median error in metal–ligand bond lengths of 0.06 Å. We then used the generated structures as inputs for machine learning. For the cytotoxicity of ruthenium complexes from MetalCytoToxDB, a model trained only on the structures came within 0.02 in ROC-AUC of the best model based on two-dimensional descriptors; for the antibacterial activity of a library of 1,800 iridium complexes prepared in our laboratory, it ranked the new complexes better than the 2D descriptor models (ROC-AUC 0.88 against 0.84). For the dipole moments and HOMO–LUMO gaps of 10,175 complexes from the tmQM dataset, adding the results of a GFN2-xTB calculation on the MeCoBi structures to a two-dimensional model lowered its mean absolute error from 1.68 to 1.24 D and 0.367 to 0.311 eV, and the improvement remained for structures built from the SMILES alone. MeCoBi is open source, with a web interface at mecobi.york.ac.uk.