Molecule Edit Graph Attention Network: Modeling Chemical Reactions as Sequences of Graph Edits

The central challenge in automated synthesis planning is to be able generate and predict outcomes of a diverse set chemical reactions. In particular, many cases, the most likely pathway cannot applied due additional constraints, which requires proposing alternative With this mind, we present Molecule Edit Graph Attention Network (MEGAN), an end-to-end encoder–decoder neural model. MEGAN inspired by models that express reaction as sequence graph edits, akin arrow pushing formalism. We extend model retrosynthesis prediction (predicting substrates given product reaction) scale it up large data sets. argue representing edits enables efficiently explore space plausible reactions, maintaining flexibility modeling fashion achieving state-of-the-art accuracy standard benchmarks. Code trained are made available online at https://github.com/molecule-one/megan.