Retrosynthesis, the task of synthesizing a complex molecule using simple precursors, is one of the fundamental problems in organic chemistry. Recently, there has been a significant development in applying machine learning to retrosynthesis planning, particularly in predicting reaction substrates for a given target molecule. These methods can be divided into template-based and template-free approaches. The former, although often fast and robust, can require a large number of templates to cover the diverse space of possible reactions. The latter, while achieving state-of-the-art accuracy, can be computationally expensive and provide less interpretability. We present METRO (Masked Edit-based Templates for RetrOsynthesis) - a machine learning model for retrosynthesis prediction that represents reactions as minimal templates which encode only the changes during a reaction. This definition allows for a smaller number of templates required to cover the reaction space and increases the accuracy in comparison to the previous models. Additionally, during both training and inference, we mask out templates that are highly improbable, which further increases the model performance. Our method achieves state-of-the-art accuracy on standard benchmarks, including large-scale datasets, while being significantly faster in execution than the previous state-of-the-art approaches.

METRO predicts reaction center together with edit-based template to generate reaction substrates for a target molecule.

Top-K single-step retrosynthesis accuracy on USPTO-50K dataset

Piotr Helm

Appearances