Generality-driven optimization of enantio- and regioselective catalysis by high-throughput experimentation and machine learning

The longstanding quest for achieving substrate generality stems from the unpredictability of single-model optimization. Leveraging the potential of high-throughput experimentation (HTE), we present a practical multi-substrate screening strategy that enables the general asymmetric mono-reduction of unsymmetrical 1,2-dicarbonyl compounds. This study combines quantitative 1H NMR with simultaneous chiral analysis facilitated by 19F NMR for pooled samples without any purification steps. Remarkably, this approach accelerated the workflow by eight times compared to the traditional method, enabling robust screening of 31 chiral oxazaborolidinium ion (COBI) variants across eight 1,2-dicarbonyl compounds. Extensive substrate screening successfully addressed even the most challenging substrates with subtle steric differentiation, such as Et/Me. Moreover, the HTE data were utilized in a machine learning (ML) model with descriptors based on the CGR (Condensed Graphs of Reaction), which identified catalysts with enhanced reactivity for previously unexplored substrates, independent of quantum chemical calculations. In this approach, the ARMS (Automated Reaction Mapping for various Substituents) system was introduced to streamline SMILES (Simplified Molecular Input Line Entry System) preprocessing for multi-substrate reactions, improving data efficiency and enabling CGR generation. The resulting chiral α-silyloxy ketones, obtained in excellent yields (up to >99%) with excellent selectivities (up to >99% ee, up to >20:1 r.r.), were then easily transformed to form various natural products and pharmaceuticals, such as (S)-bupropion—originally sold as a racemic mixture but now available in its chiral form.