Coordination-informed machine learning enables virtual screening of phenanthroline ligands by predicting Am/Eu binding preferences.

A coordination-informed machine-learning workflow reconstructs metal-ligand graphs from design-stage ligands for log β1 prediction. Privileged 3D/quantum-chemical pretraining enables transfer to f-element-rich data and virtual screening of 300 000 phenanthroline-derived ligands for predicted Am(III)-over-Eu(III) binding preference, with Δlog β1 used as a thermodynamic ranking metric rather than a complete extraction-selectivity descriptor.