Generative AI models reverse X-to-y predictions into y-to-X inverse design, where X is a chemical system and y is a target property. In the deep approach to generative AI, models are based on neural networks. We hereby developed a JT-VAE autoencoder method that encodes molecules as graphs and junction trees, addressing the challenge of supporting the complex bonding patterns between transition metals and organic molecules acting as ligands. The resulting model enabled inverse design tasks in which the generation of novel metal complexes and ligands was conditioned by one or two target properties. Further, navigation in latent space ( that is, molecule optimization within the compressed representation learned by the JT-VAE ) was projected and interpreted in chemical space.
This work was initiated by Magnus Strandgaard, who visited the Hylleraas Centre in a 3-month PhD exchange in 2023, coming from the group of Prof. Jan H. Jensen in the University of Copenhagen, Denmark.
Abstract
Deep generative models yielding transition metal complexes (TMCs) remain scarce despite the key role of these compounds in industrial catalytic processes, anticancer therapies, and the energy transition. Compared to drug discovery within the chemical space of organic molecules, TMCs pose further challenges, including the encoding of chemical bonds of higher complexity and the need to optimize multiple properties. In this work, we developed a generative model for the inverse design of transition metal ligands and complexes, based on the junction tree variational autoencoder (JT-VAE). After implementing a SMILES-based encoding of the metal–ligand bonds, the model was trained with the tmQMg-L ligand library, allowing for the generation of thousands of novel, highly diverse monodentate (κ1) and bidentate (κ2) ligands, including imines, phosphines, and carbenes. Further, the generated ligands were labeled with two target properties reflecting the stability and electron density of the associated homoleptic iridium TMCs: the HOMO–LUMO gap (ϵ) and the charge of the metal center (qIr). This data was used to implement a conditional model that generated ligands from a prompt, with the single- or dual-objective of optimizing either or both the ϵ and qIr properties and allowing for chemical interpretation based on the optimization trajectories. The optimizations also had an impact on other chemical properties, including ligand dissociation energies and oxidative addition barriers. A similar model was implemented to condition ligand generation by solubility and steric bulk.
Reference
Strandgaard, M.; Linjordet, T.; Kneiding, H.; Burnage, A. L.; Nova, A.; Jensen, J. H.; Balcells, D. A Deep Generative Model for the Inverse Design of Transition Metal Ligands and Complexes JACS Au 2025, 5, DOI: 10.1021/jacsau.5c00242
First published online: 23 April 2025.