Bayesian optimisation for additive screening and yield improvements in chemical reactions – beyond one-hot encodings

Reaction additives play a significant role in controlling the reactivity and outcome of chemical reactions. For example, a recent high-throughput additive screening identified a phthalimide ligand additive Ni-catalysed photoredox decarboxylative arylations. This discovery enabled a 4-fold yield improvement by stabilising oxidative addition complexes and breaking up deactivated catalyst aggregates. However, such large-scale screenings are currently inaccessible to most research groups. This work demonstrates how these discoveries can be made under much lower experimental budgets using Bayesian optimisation. We consider a unique reaction screening setting with 720 additives which forces us to go beyond simple one-hot encoding of the reaction components. We investigate a range of molecular representations and demonstrate convincing improvements over baselines. Our approach is not limited to Ni-catalysed reactions but can be generally applied to, for example, achieving yield improvements in diverse cross-coupling reactions or unlocking access to new chemical spaces of interest to the chemical and pharmaceutical industries. Presented at ELLIS Workshop on Machine Learning for Molecules (ML4Molecules 2022).