Assigning the Stereochemistry of Natural Products by Machine Learning

Nature has settled for L-chirality for proteinogenic amino acids and D-chirality for the carbohydrate backbone of nucleotides. Here we asked the question whether stereochemical patterns might also exist among natural products (NPs) such that their stereochemistry could be assigned automatically. Indeed, we report that a machine learning model can be trained to assign the stereochemistry of NPs using the open access NP database COCONUT. In detail, our transformer model, called NPstereo, translates an NP structure written as absolute SMILES into the corresponding isomeric SMILES notation containing stereochemical information with >85 % overall accuracy and >95 % accuracy per stereocenter across various NP classes including secondary metabolites such as alkaloids, polyketides, lipids and terpenes. NPstereo might be useful to assign or correct the stereochemistry of newly discovered NPs.