Improvement of deep neural network-based retrosynthetic prediction through label augmentation

Abstract

Designing the reactants that can produce the desired material is an important task in chemical and material science. Since the trial and error approach based on the intuition and experience of experimental experts requires a lot of time and money, many studies to apply deep learning technology for efficient retrosynthesis planning have been reported. To plan an optimal synthesis pathway by searching the large chemical space using the retrosynthesis prediction model, the ability to predict various reactants that yield the desired target product is required. In this work, we propose a training strategy to augment the reaction labels by adding the various chemical reactions to a limited dataset through virtual reaction sampling. Using a new training dataset constructed through this label augmentation, we improved the performance of predicting more diverse reactants that can produce the target molecule. We applied the label augmentation strategy to the previously reported graph and template-based retrosynthesis prediction model, and the retrained model showed an improved top-10 average round-trip accuracy by 8.2%, although the top-10 exact match accuracy was decreased by 4.7%. In addition, we found that the correlation between the prediction score and accuracy was improved.