Abstract
A catalyst selection method for the optimization of an asymmetric, vinylogous Mukaiyama aldol reaction is described. A large library of commercially available and synthetically accessible copper-bis(oxazoline) catalysts was constructed in silico. Conformer-dependent, grid-based descriptors were calculated for each catalyst, defining a chemical feature space suitable for machine learning. Selection of a diverse subset of catalyst space produced an initial training set of 26 novel bis(oxazoline) ligands which were synthesized and tested for stereoselectivity in the copper-catalyzed, vinylogous Mukaiyama aldol reaction for five substrate combinations. One ligand in the training set provided 88% average enantiomeric excess, exceeding the performance of catalysts identified through an initial optimization campaign. Supervised and semi-supervised catalyst selection methods, including quantitative structure-selectivity relationship modelling, nearest neighbors analysis, and a focused analogue clustering strategy, were employed to identify an additional 12 novel bis(oxazoline) ligands. The selected ligands outperformed the initial training set hit in four out of five product classes, and in some cases demonstrated excellent enantiocontrol exceeding 95% ee. The effectiveness of the unsupervised training set selection process is discussed, and the expediency of the nearest neighbor and focused analogue approaches are contrasted with the supervised quantitative structure-selectivity relationship modelling approach.