A deep-learning model that predicts product yields of organic reactions

What is it about?

The study of chemical transformations is an important part of chemistry. Chemists spend a great deal of time testing the possibility of different reactions and assessing how successful they are at forming a product. Owing to the vast number of chemical combinations and factors that can affect the formation of a product, this task is simply not possible by experiments alone. Instead, machine learning and artificial intelligence are used to predict the most optimal reactions. This paper presents a reaction yield prediction model for reactions involving organic compounds that predicts how much of a reactant will turn into a product. The model, based on a transformer model, can take inputs of chemical reactions represented by character sequences or strings (SMILES) and can predict product yields for all types of organic reactions.

Featured Image

Photo by Markus Spiske on Unsplash

Photo by Markus Spiske on Unsplash

Why is it important?

Most machine learning models predict the chance of chemical reactions taking place. These models do not indicate a successful reaction as they leave out the yield of the product. In this paper, the researchers have developed a model that predicts the product yields for reactions written in the SMILES format. SMILES is a widely accepted notation for entering chemical structures into a computer. The model can predict the product yield with less training data. When trained on highly consistent data obtained from high-throughput experiments, the model performed better than data-driven models that used different chemical descriptors. KEY TAKEAWAY: Organic reactions are complex. Thus, it can take a considerable amount of time and effort to find reactions that yield high amounts of a product. The model developed in this study is one of the few that predicts the yield of chemical reactions. It can be used as a screening tool to find optimal organic reactions.