Automated prediction of organic reaction products based on machine learning

Abstract

Chemical reaction product prediction is a fundamental exercise for chemists. Various product prediction models have been developed, but because most models are focused on main product prediction, the prediction of toxic byproducts was difficult. Therefore, we developed a binary classification model for toxic byproducts at first. The model performed better than the previous model in predicting the production of specific toxic substances. In addition, as a second study, we expand the reaction prediction range of the Reaction Mechanism Generator. The new model can predict reactions that are not concluded in chemical reaction templates, and reactions containing Si can be additionally predicted to C, H, O, and N. Our models cannot completely replace past models. However, by using both models together, we can predict reaction products more accurately and reliably.