Table 2.
Performance of our RetroExplainer and the state-of-the-art methods on USPTO-FULL and USPTO-MIT benchmarks
| Model/Dataset | Top-k accuracy (%) | |||
|---|---|---|---|---|
| Reaction class unknown | ||||
| k = 1 | 3 | 5 | 10 | |
| USPTO-FULL | ||||
| FRetroSim41 | 32.8 | - | - | 56.1 |
| FNeuralSym8 | 35.8 | - | - | 60.8 |
| GGLN36 | 39.3 | - | - | 63.7 |
| SRetroPrime64 | 44.1 | - | - | 68.5 |
| GRetroXpert32 | 49.4 | 63.6 | 67.6 | 71.6 |
| SR-SMILES65 | 48.9 | 66.6 | 72.0 | 76.4 |
| GRetroExplainer (Ours) | 51.4 | 70.7 | 74.7 | 79.2 |
| USPTO-MIT | ||||
| FRetroSim41 | 47.8 | 67.6 | 74.1 | 80.2 |
| SAutoSynRoute60 | 54.1 | 71.8 | 76.9 | 81.8 |
| SRetroTRAE28 | 58.3 | - | - | - |
| SR-SMILES65 | 60.3 | 78.2 | 83.2 | 87.3 |
| GLocalRetro57 | 54.1 | 73.7 | 79.4 | 84.4 |
| GRetroExplainer (Ours) | 60.3 | 81.6 | 86.4 | 90.5 |
S: Denotes sequence-based models.
G: Denotes graph-based models.
F: Denotes finger-prints-based models. The best-performing results are marked in bold.