Indeed, we used machine learning to predict reactions. DFT and molecular dynamics are much more computationally expensive (magnitude of ~10^4s) whereas here you can predict a reaction in less than a sec.
Obviously we are not the only one working on this topic, but our approach is pretty flexible and general. Compared to other approaches, we handle stereochemistry (graph methods are struggling with that) and we don't require data with atom mapping for training.
Obviously we are not the only one working on this topic, but our approach is pretty flexible and general. Compared to other approaches, we handle stereochemistry (graph methods are struggling with that) and we don't require data with atom mapping for training.
The research behind the approach has been published in a peer reviewed journal if you want more details: https://pubs.rsc.org/en/Content/ArticleLanding/2018/SC/C8SC0...