Fig. 3From: RetroRanker: leveraging reaction changes to improve retrosynthesis prediction through re-rankingOverview of the reaction encoder. The encoder takes the predicted reactants and the given product as input. The features of each atom are composed of its molecule-level features (e.g., atom degree, bond order) and reaction change features (e.g., number of reacted atoms in the neighborhood). After featurization, the reactant molecular graph and the product molecular graph are updated by GNN backbones, respectively. The reaction representation is obtained by concatenating the GNN output of the reactant and productBack to article page