RetroRanker: leveraging reaction changes to improve retrosynthesis prediction through re-ranking

Table 1 Single-step retrosynthesis models and datasets

Model	Dataset	Pairs	Model type
Model	Dataset	Pairs	Template-based	Template-free	Graph	Sequence
RetroXpert [40]	USPTO-50k*	1.9 million		\(\checkmark\)	\(\checkmark\)	\(\checkmark\)
GLN [26]	USPTO-50k*	1.8 million	\(\checkmark\)		\(\checkmark\)
R-SMILES [15]	USPTO-50k	1.2 million		\(\checkmark\)		\(\checkmark\)
AT [14]\(^\dag\)	USPTO-full	12.3 million		\(\checkmark\)		\(\checkmark\)
R-SMILES [15]	USPTO-full	16.3 million		\(\checkmark\)		\(\checkmark\)

\(^{*}\) Duplicated reactions are further removed, the sizes of train/valid/test are slightly changed to 39713/4989/5005 [23]
\(^{\dag }\) AT is for Augmented Transformer

ISSN: 1758-2946