Deconvoluting the Effects of Substituents on Reaction Barriers through Machine Learning: The Case of Brønsted Acid-Mediated Nazarov Cyclizations

The barrier heights of Nazarov cyclizations are influenced by substituents in ways that can be challenging to predict. We explore the ability of machine learning (ML) models to predict the barriers by learning the activating and deactivating contributions of different substituents on the divinyl ketone. Random forest and graph neural network models are shown to achieve good predictive accuracies, with mean absolute errors of <i>ca</i>. 2 kcal/mol against density functional theory-calculated barriers, after training on a compact data set of 272 reactions using simple and interpretable features that do not require performing quantum mechanical calculations. Substituents' electronic and steric feature contributions to the barriers are quantified through feature importance analysis using the SHAP model interpretation framework. To achieve a low barrier for cyclization, the most impactful substrate design strategy entails placing substituents α to the carbonyl group to sterically control the accessible reactant conformations and electronically stabilize the forming oxyallyl cation. The analysis also explains low or high barriers for several substrates that do not conform to empirical rules. This work illustrates the ability of ML models to capture complicated synergistic and antagonistic effects on chemical reactivity, and demonstrates how ML model feature importance analysis complements insights available from density functional theory (DFT).