Predicting Reactivity and Enantioselectivity from Small Experimental Datasets: A Case Study in Asymmetric Magnesium Catalysis

Abstract Predicting reactivity and enantioselectivity from limited experimental data remains a central challenge in asymmetric catalysis, where reaction development is often guided by small, internally generated scope datasets rather than large screening campaigns. Here, we show that such datasets can be converted into practical decision-support tools when model development is aligned with the intrinsic limitations of low-data chemistry. Using magnesium-catalyzed chalcone epoxidation and thia-Michael addition as experimentally grounded case studies, we developed a machine-learning workflow that combines reactivity classification, where appropriate, with enantioselectivity regression trained on internally consistent laboratory data. Comparative model evaluation reveals that reactivity classification can remain robust even in highly homogeneous high-enantiomeric excess (ee) datasets, whereas reliable enantioselectivity regression requires broader selectivity distributions and careful assessment of the applicability domain. Prospective experimental validation on previously untested substrates confirms that the models capture useful structure–reactivity and structure–selectivity trends rather than merely fitting retrospective data. These results establish a realistic framework for applying machine learning to asymmetric reaction-scope development under small-data conditions, providing practical guidance for substrate prioritization and AI-assisted catalysis.