Fig. 8

Per-molecule optimization of different RL strategies against different objective tasks of varying difficulty: (a) number of heavy atoms, (b) Similarity to Risperidone (DRD2 inverse agonist), (c) predicted probability of DRD2 activity, (d) Glide-SP docking score against DRD2, (e) predicted probability of dual activity against DRD2 and (f) predicted probability of selective activity towards DRD2 over DRD3. Standard deviation can be seen in Additional file 1: Figure S8. In all cases, except the number of heavy atoms, AHC outperforms all other RL strategies with respect to objective optimization while maintaining validity and uniqueness. Only valid molecules are plotted, therefore gaps seen with HC* denote regions where no valid molecules were generated