Figure 2.

Although the augmented dataset contains roughly 7 times more molecules than the original dataset, the histograms show that the augmentation strategy largely preserves the distribution of multiple molecule metrics. For Synthesizability, generated molecules for data augmentation were required to have a score above 0.3, resulting in the observed sharp decline in the histogram. For drug-likeness, no constraints were placed on the augmented data, which resulted in a decrease in typical scores relative to Enamine.