Fig. 2From: RetroRanker: leveraging reaction changes to improve retrosynthesis prediction through re-rankingOverview of re-ranking with RetroRanker. Each prediction from the single-step retrosynthesis model is paired with the product as input to RetroRanker (the potential reaction changes are highlighted in dashed boxes). RetroRanker consists of a reaction encoder and a scorer module. The reaction encoder takes the molecular features and the potential reaction changes as input, which are further updated with GNN backbones. The re-ranking score is calculated based on the high-dimensional representation from GNNBack to article page