Carbohydrate Transformer: Predicting Regio- and Stereoselective Reactions Using Transfer Learning

Organic chemistry is central to society because it enables the synthesis of complex molecules and materials used in all fields of science and technology. The synthetic methods represent a vast body of accumulated knowledge optimally suited for deep learning. Indeed, most organic reactions involve distinct functional groups and can readily be learned by deep learning models and chemists alike. The task is, however, much more challenging for regio- and stereoselective transformations because their outcome also depends on functional group surroundings in subtle ways. Here, we challenge the Molecular Transformer model to predict reactions in carbohydrate chemistry, a field of central importance in the life sciences and for vaccine development and where regio- and stereoselectivity are notoriously difficult to predict even for experienced chemists. We show that transfer learning of the general USPTO model with a small set of carbohydrate reactions produces a specialized Carbohydrate Transformer model, returning predictions for carbohydrate reactions with remarkable accuracy. We validate these predictions experimentally with a previously unpublished synthesis of a lipid-linked oligosaccharide, involving regioselective protecting group operations and stereoselective glycosylations that are typical for complex carbohydrate synthesis. The chemical reaction transfer learning methods presented in this work are generally applicable to any reaction class of interest.