. 2020 Nov 4;11:5575. doi: 10.1038/s41467-020-19266-y

Table 6.

Top-k accuracy for retrosynthesis prediction on USPTO-full dataset.

	Retrosim²⁰	Neuralsym³	GLN²⁴	AT^a
Top-1	32.8	35.8	39.3	46.2 (44.4)^b
Top-2				57.2 (54.9)
Top-10	56.1	60.8	63.7	73.3 (70.4)
MaxFrag top-1				54
MaxFrag top-2				66.3
MaxFrag top-5				77.3
MaxFrag top-10				80.6

^aModel was trained using a filtered training set of 769 k from ref. ²⁰.

^bAccuracies in parentheses correspond to those recalculated for the test set by assuming that AT failed for all 4% of excluded reactions. Results for retrosim and neuralsym approaches as reported by Dai et al.²⁴.