Table 2. Top-k Accuracy for Retrosynthesis Prediction on the USPTO-Full Data Seta.
| model |
Top-k accuracy (%) |
||||
|---|---|---|---|---|---|
| model type | methods | 1 | 3 | 5 | 10 |
| template-based | RetroSim4 | 32.8 | - | - | 56.1 |
| NeuralSym5 | 35.8 | - | - | 60.8 | |
| GLN7 | 39.3 | - | - | 63.7 | |
| semi-template-based | RetroPrime12 | 44.1 | 59.1 | 62.8 | 68.5 |
| template-free | MEGAN26 | 33.6 | - | - | 63.9 |
| aug. transformer* | 44.4 | - | - | 70.4 | |
| NAG2G (ours) | 47.7 | 62.0 | 66.6 | 71.0 | |
| aug. transformer*◦20 | 46.2 | - | - | 73.3 | |
| G2GT*◦25 | 49.3 | - | 68.9 | 72.7 | |
| NAG2G (ours)◦ | 49.7 | 64.6 | 69.3 | 74.0 | |
a
Models denoted by an asterisk (*) used supplementary data sets for training or incorporated techniques to improve accuracy during inference. For models denoted by a circle (◦), the invalid reactions are excluded from the test set, following the setting of the augmented transformer.20 To align our methods with the previous baselines, we adopted the approach from the augmented transformer,20 assuming that the methods failed on the removed test data, as evidenced by the results of our methods without a circle (◦).