Skip to main content

Table 3 CPU hours required for RL strategies to optimize the DRD2 docking score benchmark task to different thresholds

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

 

CPU hours required for optimization beyond prior at a given threshold

CPU hours required for optimization beyond external thresholds

Threshold

120%

140%

160%

180%

200%

Inactive mean

Active mean

80% precision

REINFORCE

74 (0)

173 (0)

– (20)

– (34)

– (96)

2 (0)

103 (0)

177 (0)

REINFORCE + KL regularization

183 (0)

– (0)

– (33)

– (74)

– (216)

22 (0)

204 (0)

– (0)

REINVENT

79 (0)

– (0)

– (8)

– (164)

– (–)

4 (0)

93 (0)

– (0)

REINVENT 2.0

38 (0)

202 (0)

– (16)

– (53)

– (92)

12 (0)

51 (0)

198 (0)

BAR

– (0)

– (0)

– (32)

– (32)

– (–)

4 (0)

0 (0)

– (0)

Hill-Climb

44 (0)

114 (0)

177 (0)

218 (24)

– (85)

16 (0)

57 (0)

99 (0)

Hill-Climb + KL regularization

45 (0)

106 (0)

157 (0)

– (45)

– (45)

8 (0)

58 (0)

99 (0)

Hill-Climb*

11 (0)

31 (1)

52 (6)

– (15)

– (31)

2 (0)

11 (0)

24 (0)

Hill-Climb* + KL regularization

14 (0)

28 (0)

74 (1)

– (17)

– (17)

6 (0)

17 (0)

31 (0)

Augmented Hill-Climb

9 (0)

16 (0)

72 (0)

151 (14)

216 (15)

2 (0)

13 (0)

27 (0)

  1. Time is representative of when the batch mean exceeds the respective internal / external threshold (time of the earliest sample exceeding threshold is shown in brackets). Run using an AMD Threadripper 1920 × CPU and Nvidia GeForce RTX 2060 super GPU. Failing to reach a threshold is marked by a “–”