Forward Reaction Prediction as Reverse Verification: A Novel Approach to Retrosynthesis

27 October 2020, Version 1
This content is a preprint and has not undergone peer review at the time of posting.


Pupils' intuitive knowledge can lead them to verify multiplication by means of division. Based on this analogy, this study introduces the basic reverse verification concept to verify retrosynthesis through forward reaction. In this work, we present a "combined" model approach for retrosynthetic reaction prediction, where the first model is applied to retrosynthesis, and the second model, which is a "verified" model, is applied to the forward reaction prediction to verify the top-n reactants predicted by the retrosynthetic model. Using a "combined" model borrowed from human language translation, sequence-to-sequence (seq2seq) + transformer models, we improve the top-1 accuracy of retrosynthetic prediction by 4.3% (37.4% vs 41.7%). The application of the similarity + seq2seq models increases the top-1 accuracy by 4.6% (52.9% vs 57.5%). In this way, we can not only improve the accuracy but also automate the evaluation of the synthetic route.


reaction prediction
reverse verification
deep learning
organic synthesis
organic chemistry


Comments are not moderated before they are posted, but they can be removed by the site moderators if they are found to be in contravention of our Commenting Policy [opens in a new tab] - please read this policy before you post. Comments should be used for scholarly discussion of the content in question. You can find more information about how to use the commenting feature here [opens in a new tab] .
This site is protected by reCAPTCHA and the Google Privacy Policy [opens in a new tab] and Terms of Service [opens in a new tab] apply.