Inverse ligand design: a generative data-driven model for optimizing vanadyl-based epoxidation catalysts

We present a machine learning (ML) model for the inverse, de novo generative design of vanadyl-based catalyst ligands for the epoxidation of small alkenes and alcohols. Leveraging molecular descriptors calculated using the RDKit library, our model achieves high performance in validity (64.7 %), uniqueness (89.6 %), and RDKit similarity (91.8 %) after training on a curated dataset of six million structures. Focusing on the modular nature of vanadyl catalyst scaffolds – VOSO 4 , VO(O i Pr) 3 and VO(acac) 2 – the model generates feasible ligands optimized for catalytic performance. The VOSO 4 ligands were consistent with high-yield reactions, while VO(O i Pr) 3 and VO(acac) 2 revealed greater structure variability. Unlike conventional generative approaches, the ligand inverse design also aims to co-design the reaction system, including substrate SMILES and reaction conditions. The model framework was investigated using multiple generative model architectures, resulting in the deep-learning transformer as the most powerful, and uncovering clustering patterns in electronic and structural descriptors related to yield predictions. High synthetic accessibility scores supported the feasibility of the generated ligands. Although the experimental dataset lacks modularity and has few negative data, the model compensates this by using structured descriptor encoding and compatibility scoring. This generative ML framework shows a scalable approach to catalyst ligand design, paving the way for efficient data optimization in synthetic chemistry.