Comparison of structure- and ligand-based scoring functions for deep generative models: a GPCR case study

Table 1 Probability of recovering known DRD2 active and inactive molecules

Origin of dataset	Probability of generating active molecule (× 10⁻⁶)		Probability of generating inactive molecule (× 10⁻⁶)		Active bias (fold change from Prior)
Active DRD2 chemical space relative to training data	In	All	In	All	In	All
Prior	10 (30)	10 (30)	5055 (604)	5957 (495)	0.002 (1)	0.002 (1)
Glide-Agent	11 (32)	63 (84)	422 (125)	917 (175)	0.025 (12.5)	0.069 (40.6)
SVM-Agent	34 (72)	79 (72)	256 (124)	486 (168)	0.130 (64.9)	0.163 (95.7)

The reported probability values are the mean (and standard deviation) across ten samples of 10,000 de novo molecules drawn from the model. The Glide- and SVM-Agent have a similar probability of recovering known active molecules, therefore the SVM-Agent bias towards generating active molecules over inactivate molecules is mostly driven by the lower probability of generating inactive molecules

ISSN: 1758-2946