Skip to main content

Table 3 Comparison of the performance of the different methods in the target-specific case

From: DrugEx v2: de novo design of drug molecules by Pareto-based multi-objective reinforcement learning in polypharmacology

Rewarding scheme

Dataset

Validity

Desirability

Uniqueness

Diversity

Purine ring

Furan ring

Benzene ring

 

LIGAND

100.00%

14.63%

100.00%

0.67

28.27%

50.61%

71.84%

PF

DrugEx v1

98.07%

48.42%

87.32%

0.73

29.65%

61.61%

70.99%

DrugEx v2

99.53%

89.49%

90.55%

0.73

23.73%

56.23%

67.40%

ORGANIC

98.29%

86.98%

80.30%

0.64

10.60%

89.27%

65.28%

REINVENT

99.59%

70.66%

99.33%

0.79

3.85%

33.82%

92.53%

WS

DrugEx v1

97.61%

44.96%

95.89%

0.68

78.92%

80.21%

68.02%

DrugEx v2

99.62%

97.86%

90.54%

0.31

19.58%

98.56%

51.87%

ORGANIC

98.97%

88.14%

84.13%

0.49

9.68%%

96.66%

71.48%

REINVENT

99.55%

81.27%

98.87%

0.34

25.13%

97.52%

74.61%

  1. Shown are validity, desirability, uniqueness, and substructure distributions of SMILES generated by four different methods in the target-specific case with PF and WS rewarding schemes. For the validity, desirability and uniqueness, the highest values are bold, while for the distribution of substructures, the bold data are labeled as the most closed to the values in the LIGAND set