Skip to main content
Fig. 4 | Journal of Cheminformatics

Fig. 4

From: Augmented Hill-Climb increases reinforcement learning efficiency for language-based de novo molecule generation

Fig. 4

Improved learning efficiency of Augmented Hill-Climb against four targets: (a) DRD2, (b) OPRM1, (c) AGTR1 and (d) OX1R. (top left panel) Distribution of known active and inactive molecule docking scores. (top right panel) Optimization of de novo molecule docking score via reinforcement learning. (bottom right panel) The top 500 REINVENT generated scaffolds with the corresponding time of generation by REINVENT or by Augmented Hill-Climb (in combination with DF2) if co-generated. Blue lines represent scaffolds generated by REINVENT first and green lines generated by Augmented Hill-Climb (in combination with DF2) first. Scaffolds with a difference in generation time of < 100 RL updates are more transparent. Augmented Hill-Climb in combination with DF2 shows improved learning efficiency compared to REINVENT and optimizes past a docking score threshold corresponding to a retrospective classification precision of 80% (black dashed line) in all cases

Back to article page