Table 3.
Evaluation of single-step retrosynthetic models on different train-test splits of USPTO-50k dataset
| Data split | Model | Top-k accuracy (%) | |||
|---|---|---|---|---|---|
| k = 1 | 3 | 5 | 10 | ||
| Original random split | MEGAN | 48.1 | 70.7 | 78.4 | 86.1 |
| GraphRetro | 53.7 | 68.3 | 72.2 | 75.5 | |
| Graph2Edits (D-MPNN) | 55.1 | 77.3 | 83.4 | 89.4 | |
| Tanimoto similarity <0.6 | MEGANa | 47.0 | 69.2 | 76.2 | 83.6 |
| GraphRetroa | 49.1 | 63.2 | 66.9 | 69.1 | |
| Graph2Edits (D-MPNN) | 52.0 | 75.6 | 83.2 | 89.4 | |
| Tanimoto similarity <0.4 | MEGANa | 45.4 | 68.4 | 76.9 | 84.6 |
| GraphRetroa | 44.2 | 56.2 | 58.7 | 59.6 | |
| Graph2Edits (D-MPNN) | 47.5 | 71.7 | 80.1 | 88.0 | |
Graph2Edits (D-MPNN) uses the directed message passing neural network (D-MPNN) as graph encoder.
aDenotes that the result is implemented by the open-source code with well-tuned hyperparameters.