Skip to main content

Table 4 Compounds generated using reinforcement learning

From: Memory-assisted reinforcement learning for diverse molecular de novo design

Target

Memory type

Generated active compounds

Unique BM scaffolds

Unique carbon skeletons

HTR1A

No memory

9323

7312

5446

Compound similarity

16,779

13,304

9887

IdenticalBMScaffold

17,390

13,863

9941

IdenticalCarbonSkeleton

17,597

15,531

12,408

ScaffoldSimilarity

17,383

15,296

12,082

DRD2

No memory

5143

2635

1949

Compound similarity

21,486

17,844

12,749

IdenticalBMScaffold

22,312

14,850

8220

IdenticalCarbonSkeleton

22,115

19,096

12,562

ScaffoldSimilarity

22,784

20,712

16,434

  1. The generative models were directed towards predicting active compounds using RL for 300 iterations. During each iteration, a model generated 100 compounds resulting in a total of 30.000 compounds. Only compounds with a prediction score of at least 0.7 were considered active