Skip to main content

Table 2 Number of steps taken before the mean exceeds certain internal and external thresholds (earliest sample exceeding threshold is shown in brackets)

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

 

Threshold

Number of steps required for optimization beyond prior at a given threshold

Number of steps required for optimization beyond external thresholds

120%

140%

160%

180%

200%

Inactive mean

Active mean

80% precision threshold

DRD2

REINVENT

> 500

(15)

> 500

(685)

> 500

(22,292)

> 500

(> 32,000)

> 500

(> 32,000)

1

(1)

163

(15)

 > 500

(15)

Augmented Hill-Climb + DF2

19

(2)

6

(49)

105

(1248)

> 500

(3009)

> 500

(23,150)

2

(2)

19

(2)

48

(2)

OPRM1

REINVENT

133

(7)

> 500

(868)

> 500

(7663)

> 500

(> 32,000)

> 500

(> 32,000)

4

(2)

80

(4)

> 500

(7)

Augmented Hill-Climb + DF2

3

(16)

17

(22)

45

(29)

150

(34)

> 500

(2759)

6

(16)

17

(22)

33

(28)

AGTR1

REINVENT

> 500

(25)

> 500

(510)

> 500

(5,596)

> 500

(> 32,000)

> 500

(> 32,000)

1

(2)

> 500

(8)

419

(6)

Augmented Hill-Climb + DF2

62

(27)

318

(869)

396

(3,404)

> 500

(5,207)

> 500

(27,979)

2

(1)

62

(27)

46

(2)

OX1R

REINVENT

5

(1)

52

(1)

> 500

(7)

> 500

(142)

> 500

(490)

1

(2)

9

(1)

> 500

(490)

Augmented Hill-Climb + DF2

9

(1)

15

(2)

31

(2)

87

(31)

382

(557)

2

(1)

14

(2)

494

(557)

Average fold improvement

19.8

(2.5)

11.2

(38.7)

8.3

(71.8)

2.8

(240.6)

1.1

(3.8)

0.5

(1.0)

5.5

(2.1)

9.7

(3.2)

  1. The final row lists the Augmented Hill-Climb in combination with DF2 fold improvement over REINVENT. Where a threshold was not reached within the maximum number of training steps (or samples) it has been annotated as being greater than 500 (or 32,000)