Machine Learning-Guided Scope Selection to Balance Performance and Substrate Similarity

Abstract

Abstract The determination of a reaction substrate scope enables downstream users to assess whether a specific transformation is suitable for an intended application. The information content of this scope, characterized by its performance and structural diversity, is crucial to ensuring the quality of such decisions. Despite their importance, scopes have historically been evaluated subjectively, leading to well-documented selection biases that compromise the scope quality. Herein, we report a scope score to analyze scopes quantitatively within a specific chemical search space, enabling the benchmarking of common selection methods. Additionally, we developed a broadly applicable and user-friendly machine learning algorithm, ScopeBO, for selecting scopes that effectively balance substrate performance and diversity. The scope score is used to optimize the hyperparameters of ScopeBO and validate its performance across several reaction data sets. Using statistical tests, we found that ScopeBO outperforms all other selection methods considered. Consequently, ScopeBO provides an objective and standardized approach to scope selection that maximizes the information content of the evaluated substrates, addressing a long-standing challenge in reaction development.