Table 5.
Comparison of retrosynthesis recently published methods for retrosynthesis prediction on USPTO-50 k.
| Model | Top-1 | Top-2 | Top-5 | Top-10 | Ref. # | Comments |
|---|---|---|---|---|---|---|
| Seq2Seq | 37.4 | 57.0 | 61.7 | 12 | 40/5/5 split; splitting any reactions with multiple products into multiple single product and removal of trivial products | |
| Transformer (3*6) | 42.7 | 52.5 | 69.8 | – | 13 | 45/5 split: no validation set was used |
| Transformer (6*8), (self corrected) | 43.7 | 65.2 | 68.7 | 19 | 40/5/5 split, reagents from reactants are removed | |
| Transformer, augmentation | 44.8 | 57.1 | 57.7 | 79.4 | 32 | Same as in ref. 12. |
| Similarity-based | 37.3 | 63.3 | 74.1 | 20 | Same as in ref. 12. | |
| Graph Logic Network | 52.5 | 75.6 | 83.7 | 24 | Same as in refs. 12,19. | |
| G2Gs | 48.9 | 72.5 | 75.5 | 25 | Same as in ref. 12. | |
| ATa | 53.5 | 69.4 | 81 | 85.7 | Same as in ref. 13. | |
| AT | 53.2 | 68.1 | 80.5 | 85.2 | Only 40 k samples were used as training set to match the other results | |
| AT MaxFragb | 58.5 | 73 | 85.4 | 90 | Same as in ref. 13. | |
| AT MaxFrag | 58 | 73.4 | 84.8 | 89.1 | Only 40 k samples were used as training set to match the other results |
aThe results of the reference model applied to x100 augmented dataset using beam size = 10.
bThe classical retro-synthesis accuracy was estimated as accuracy for prediction of the largest fragment (MaxFrag).