Fig. 3
From: Explaining and avoiding failure modes in goal-directed generation of small molecules
\(S_{opt}\) and \(S_{dc}\) in the EGFR, JAK2 and DRD2 Topliss-augmented datasets. From top to bottom: hexbin plots (log scale) of data control as a function of optimization score; Mean Average Difference between \(S_{dc}\) and \(S_{opt}\) as a function of \(S_{opt}\) (at absciss x, the MAD plotted is the MAD for molecules with optimization scores higher than x); distribution of \(S_{dc}\) (95 CI) as a function of optimization score. For the second and third row, the lines and boxplots stop at absciss \(x_{max}\) for which there is no more samples with optimization scores higher than \(x_{max}\)