Table 1. Template Top-k Accuracy (%) of Different Method Variants on USPTO-sm and USPTO-lg^*.

^*“Model” indicates how the templates were ranked. “Filter” specifies if and how templates were excluded from the ranking via FPF or an applicability check (App). Pre-train indicates whether a model was pre-trained on the applicability task. Error bars represent confidence intervals on binomial proportions. The gray rows indicate methods specifically proposed here or in prior work.

^aWidth of 95% confidence interval <1.3%.

^bWidth of 95% confidence interval <0.4%.

^cNote that the applicability filter violates modeling constraints from the section entitled “Single-Step Retrosynthesis”.