Deep Generative Model for the Dual-Objective Inverse Design of Metal Complexes

Deep generative models yielding transition metal complexes (TMCs) remain scarce despite the key role of these compounds in industrial catalytic processes, anticancer therapies, and energy transformations. Compared to drug discovery within the organic molecular space, TMCs pose further challenges including the encoding of chemical bonds of higher complexity and the optimization of multiple properties, in a context in which synthesizability is affected by additional, complex factors. In this work, we developed a junction tree variational autoencoder (JT-VAE) model for the generation of metal ligands. After implementing a SMILES-based encoding of the metal–ligand bonds, the model was trained with the tmQMg-L ligand library, allowing for the random generation of thousands of monodentate and bidentate ligands with full validity and high novelty. The generated ligands were labeled with two target properties of the associated [IrL4]+ and [IrL2]+ homoleptic TMCs; namely the HOMO-LUMO gap (ϵ) and the metal charge (qIr), both computed at a DFT level. This data was used to implement a conditional JT-VAE model generating ligands from a prompt, with the single or dual objective of optimizing either one or both properties in Y = (ϵ, qIr). Conditional ligand generation was able to navigate both central and extreme regions of this bidimensional property space, allowing for chemical interpretation based on the step-wise analysis of the decoded optimization trajectories.