Experimental Approaches to the Corey-Bakshi-Shibata and Dakin-West Reactions as Model Systems for Machine Learning Applications in the Small-Data Regime

Predicting enantioselectivity remains a fundamental challenge in asymmetric organocatalysis. While Machine Learning (ML) offers data-driven pathways for catalyst optimization, its application is hindered by data scarcity and publication bias. This work investigates the integration of experimental data into predictive ML models within the small-data regime, using the Corey-Bakshi-Shibata (CBS) reduction and the asymmetric variant of the Dakin-West reaction as model reactions. The primary objective was the development of robust predictive models from limited, unbiased datasets, and to delineate the operational limits of current molecular representations as well as ML algorithms. Initial investigations of the CBS reduction of ketones yielded a graph-based ML model that predicts stereochemical outcomes and identifies promising catalyst precursors. However, the synthesis and investigation of additional precursors, enabled by a newly developed synthetic route, revealed deviations from previous predictions. This demonstrates that the reliability of small-data ML models depends on high-quality datasets and robust analytics to avoid the compensation for empirical artifacts. Further investigations of CBS catalysts into ketimine substrates revealed that an unselective background reaction consistently outpaces the catalyzed pathway. This inherent lack of stereocontrol renders ketimines unsuitable for modeling, making the suppression of the background pathway an essential prerequisite for future work. For the asymmetric Dakin-West reaction, the research strategy prioritized empirical data generation over premature modeling. Through systematic parameter optimization, a comprehensive data set was established, which maps complex structure-activity relationships and provides a solid, high-quality foundation for future ML applications. Ultimately, this work demonstrates that advancing ML-powered catalyst design within the small-data regime depends on the synergy of high-precision experimental data and rigorous analytical validation. By focusing on data quality in addition to increasing data quantity, the operational boundaries of predictive modeling can be pushed to address complex synthetic challenges.