Fig. 3

Comparison between REINVENT and Augmented Hill-Climb learning strategies to optimize DRD2 docking scores at varying levels of σ. (a) Augmented Hill-Climb is more efficient at optimizing docking score at all levels of σ but (b) undergoes increased mode collapse via a drop in uniqueness. (c) Docking score optimization can be stabilized and (d) mode collapse rescued by applying a diversity filter. (e–g) Augmented Hill-Climb in combination with DF1 is more sensitive to changes in σ, this affects the extent to which de novo molecules occupy property space which is not present in the prior training set (grey shaded area) i.e., extrapolation