Evaluating Predictive Accuracy in Asymmetric Catalysis: A Machine Learning Perspective on Local Reaction Space

Machine learning (ML) models are increasingly being employed in asymmetric catalysis to predict reaction outcomes and optimize enantioselective processes. Despite the trend of expanding data set sizes to improve model performance, asymmetric catalysis presents unique challenges, including the difficulty of acquiring high-quality experimental data and the often-limited availability of structurally diverse examples. Consequently, rational data set design requires the practitioner to choose whether to collect data that maximizes diversity in the training set or data that maximizes representation around a target prediction. A key challenge in these studies is understanding the role of local reaction space─specifically, how much predictive accuracy is driven by nearest neighbors (structurally and electronically similar data points) and the next-nearest neighbors? This study investigates the predictive power of ML models trained with varying levels of local representation in the reaction space. We provide a framework, a radius-based random forest (RaRF) algorithm, to systematically probe the effects of including diverse reactions dissimilar to a target prediction. We show that when the training set is representative of the target reaction, the gains from increasing data set diversity are modest─typically less than 0.1 kcal/mol in predictive error─and increasing to only 0.5 kcal/mol for extrapolative tests, highlighting the need for targeted data set design. Furthermore, these findings hold even for complex architectures and features. Finally, we demonstrate that a targeted, neighborhood-oriented strategy greatly accelerates the identification of predictive models compared to diversity-driven approaches.