Table 2 Comparison of performance among different models for the test set.

From: Deep learning driven biosynthetic pathways navigation for natural products with BioNavi-NP

Methods

Success rate

Hit rate of building blocks

Hit rate of pathways

Longest length

Avg. solutiona

Time (h)b

BioNavi-NP (MCTS)

34.8%

16.3%

1.9%

3

1.0

92

RetroPathRL

52.7%

4.8%

3.8%

3

2.8

2

BioNavi-NP

90.2%

56.0%

24.7%

6

4.9

18

RetroPathRL_UDB

10.8%

5.1%

4.1%

3

2.8

3

BioNavi-NP_UDB

74.7%

72.8%

26.1%

6

4.9

28

  1. UDB user-defined building blocks.
  2. aDenotes the average number of pathways found, only the top-1 result is supported by the MCTS algorithm, while for RetroPathRL, it outputs all pathways it can find. The output option for Retro* is set as top-5 (default is top-10).
  3. bIt is an about 4-times computational time for outputting top-10 in comparison to top-5, that is, the time consuming of BioNavi-NP (if only requesting the top-3) is comparable to RetroPathRL (the average number of pathways returned by RetroPathRL is close to 3).
Back to article page