Fig. 3

Generative model performance during optimization for the Glide-Agent (green) and the SVM-Agent (red), calculated every 100 steps. Mean optimization of scores—docking score and predicted probability of activity—are shown in (a) and (b) respectively, as well as the 95% confidence interval. Additional metrics shown are (c) validity, (d) uniqueness, (e) novelty, (f) internal diversity, (g) scaffold diversity, (h) sphere exclusion diversity, (i) Fréchet ChemNet Distance, (j) single nearest neighbour similarity and (k) fragment similarity. As the most important observation, the SVM-Agent reaches very high scores much more quickly, which comes at the cost of a significant reduction in uniqueness and diversity, when compared to the Glide-Agent. For definitions and detailed discussion see main text