Retrosynthetic Planning with Generative Models

Recently, retrosynthetic planning with machine learning and deep learning has been studied actively. One of the powerful approach was the retrosynthesis based on the molecular similarity proposed by Coley et al. In this approach, similarities between the target product and products in the reaction database were calculated to find similar products. Next, candidate reactions were generated by modifying reactions of the similar targets. This method by Coley et al. is more accurate than other methods. However, its search space is limited because it is based on the matching with the existing reactions.
In this presentation, we propose a method with generative models, such as GAN(generative adversarial network) or VAE( variational autoencoder ) in order to expand the search space. The idea of the proposed method is to learn a generative model with GAN or VAE, generate reactants with the generative model, and then a reaction with reaction prediction. We got new reactants with the proposed method. We applied our proposed method to US patent dataset. As a result, reactants which did not exist in the reaction database were produced. Besides, the accuracy of retrosynthesis was higher than the previous methods.