Data-Driven Prediction of Enantioselectivity for the Sharpless Asymmetric Dihydroxylation: Model Development and Experimental Validation

The Sharpless asymmetric dihydroxylation remains a key transformation in chemical synthesis, yet its success hides unexpected cases of lower selectivity. A chemoinformatic workflow was developed to allow data-driven analysis of the reaction. A database of 1007 reactions employing AD-mix α and β was curated from the literature, and an alignment-dependent, fragment-based featurization of alkenes was implemented for modeling. This platform converged on machine learning models capable of predicting the magnitude of enantioselectivity for multiple alkene classes, achieving Q2F3 values ≥ 0.8, test r2 values ≥ 0.7 and mean absolute errors (MAE) ≤ 0.3 kcal/mol. The features of alkenes contributing to model performance were assessed with SHapley Additive exPlanations (SHAP) analysis to gather insight into factors underlying predictions. Experimental validation demonstrated that the models could achieve meaningful predictions on numerous out-of-sample alkenes.