Table 3 The performance of RetroRollout* compared to other approaches on the natural products dataset

From: A virtual platform for automated hybrid organic-enzymatic synthesis planning

Method name

Multi step method

Single step model

Single step type

Success rate (%)

Avg. solutiond

ChemEnzyRetroPlanner

RetroRollout*

Pistachio ringbreaker

O + TB

5.2

0.1

BKMS metabolic

E + TB

58.7

5.2

Reaxys+BKMS metabolic

O + E + TB

64.4

6.0

Notea

O + E + TF

98.4

9.8

USPTO-all-remappedb

O + TB

37.2

3.4

Reaxys

O + TB

45.4

4.2

Reaxys + Reaxys biocatalysis

O + E + TB

53.0

4.8

Retro*-GraphFP

Retro*

USPTO-all-remappedb

O + TB

29.1

1.7

Retro*-Original29

Retro*

Notec

O + TB

28.8

1.6

RetroPathRL40

MCTS

–

E + TB

59.8

6.1

RetroBioCat27

MCTS

–

E + TB

33.2

9.4

RXN4Chem59

Hyper-graph exploration

–

O + TF

42.7

0.8

BioNavi-NP39

Retro*

–

O + E + TF

89.4

2.9

ASKCOS41

MCTS

–

O + TB

36.4

8.7

BioNavi28

Retro*

–

O + E + TF

97.0

9.5

  1. The single-step retrosynthesis prediction models are labeled based on the name of their training dataset. O Organic single-step model, E Enzymatic single-step model, TB Template-based single-step model, TF Template-free single-step model. Bold indicates the best.
  2. aUsing the model weights trained in the BioNavi-NP work.
  3. bUsing the GraphFP model trained on USPTO-all-remapped.
  4. cUsing the single-step retrosynthesis prediction model from the original Retro* paper29.
  5. dThis metric represents the average number of successful synthetic routes proposed by the method for a given target molecule.
Back to article page