Transition metal complexes (TMCs) are widely used as homogeneous catalysts and functional materials, with applications spanning sustainability, energy conversion, and biological systems. The combinatorially large space of possible TMCs makes exhaustive exploration intractable, motivating high-throughput virtual screening (HTVS) campaigns driven by electronic structure calculations and machine learning predictions. Such approaches are highly sensitive to the initial structure and connectivity of the provided TMC, with metal–ligand coordination being particularly difficult to determine a priori. Researchers thus primarily rely on subsets of complexes with experimentally-resolved connectivity or on error-prone heuristics. To explore beyond such limited regions of chemical space, we present pydentate, a framework for the accelerated discovery of TMCs by integrating graph neural networks, comprehensive handling of metal–ligand bonding, and electronic structure calculations. Our graph neural network models accurately predict metal–ligand coordination from only SMILES string inputs, while analysis of the learned representations reveals chemically meaningful trends in coordination. We extend our approach to handle hemilabile ligands with variable coordination modes across chemical environments. While these ligands with dynamic coordination are out of distribution from the coordination model training data, we demonstrate that fine-tuning our coordination models to identify hemilabile ligands results in predictions with high overall and balanced accuracy. Synergistic use of our models for predicting ligand coordination and hemilability enables identification of experimentally observed coordination, even for challenging systems with multiple chemically viable binding modes. Algorithms we developed to map the predicted connectivity graphs to denticity and hapticity are integrated with the high-throughput screening software molSimplify for the rapid generation of de novo, physically realistic TMCs and subsequent screening through electronic structure calculations. We anticipate our workflow will significantly accelerate computationally driven discovery of functional TMCs.
