Table 1 Performance summary of various Transformer-based models trained with different fragmentation schemes in unimolecular test set and a comparison with the BiLSTM-based models.
From: Retrosynthetic reaction pathway prediction through neural machine translation of atomic environments
BiLSTM-based | Transformer-based | |||||||
|---|---|---|---|---|---|---|---|---|
MACCS | ECFP2 | ECFP4 | MACCS | ECFP2 | ECFP4 | AE2 | AE0 ∪ AE2 | |
Tc = 1.0 | 29.9 | 35.6 | 9.1 | 30.1 | 54.9 | 26.0 | 50.9 | 55.4 |
Tc ≥ 0.85 | 57.7 | 50.7 | 28.4 | 57.5 | 67.6 | 50.1 | 59.9 | 68.1 |
\(\overline{{T}_{c}}\) | 0.84 | 0.80 | 0.66 | 0.85 | 0.88 | 0.73 | 0.84 | 0.88 |
