Fig. 4
From: Human-in-the-loop active learning for goal-oriented molecule generation
EPIG-guided active learning improves the generation of molecules with desired penalized LogP values according to the oracle compared to the “No Feedback” baseline. This improvement is visible in the concentration of points within the green rectangular target area and the increasing linear correlation between oracle and predicted LogP values. The plot shows the predicted LogP values better aligning with the oracle values over the course of the generation runs. We report these values for all generated molecules with a transformed predicted LogP score above 0.5. The generation runs with AL updates using EPIG as the acquisition criterion and noisy simulated experts are shown in red, while the “No Feedback” baseline is shown in gray