Figure 3.

Top-1 and round-trip accuracy for the reactant prediction task, using the USPTO-MIT testing dataset, for different levels of data augmentation. Top. Top-1 accuracy. “Strict” requires an exact match between the model prediction and the target. “Lenient” requires that at least one molecule predicted by the model matches a target molecule. Bottom. Round-trip accuracy. The diagram shows how round-trip accuracy was computed. When reagents were part of the datasets, the true reagents were added to the predicted reactants before being sent to the product prediction model. P—true product, [R_c]—true reactant(s), [R_g]—true reagent(s), P̂—predicted product, and —predicted reactant(s).