Prediction of Reaction Orthogonality using Machine Learning

We present a statistical learning model relying on a small dataset to predict the selectivity of a two state system toward the same substrate, specifically of redox-switchable metal complexes in the ring opening polymerization of e-caprolactone or trimethylene carbonate. We mapped the descriptor space of several switchable metal complexes and surveyed a set of supervised machine learning algorithms using different train/test validation methods on a limited dataset based on experimental studies of ca. 10 metal complexes. Linear discriminant analysis showed an accuracy of >80% and a F1 score of 0.86 on a test mixture of experimental and predicted molecules, and successfully predicted the reactivity of three new metal complexes. The established method will be used to guide future studies in recommending promising new metal complexes for related substrates, reducing the need for blind synthetic trial and error efforts.