Top-K accuracy of forward synthesis on the USPTO-MIT dataset. “Separated” and “mixed” denote whether reagents are separated from reactants or not.
| USPTO-MIT top-K accuracy (%) | ||||||
|---|---|---|---|---|---|---|
| Setting | Model | K = 1 | 2 | 5 | 10 | 20 |
| Separated | MT22,28 | 90.5 | 93.7 | 95.3 | 96.0 | 96.5 |
| MEGAN22 | 89.3 | 92.7 | 95.6 | 96.7 | 97.5 | |
| AT17 | 91.9 | 95.4 | 97.0 | — | — | |
| Chemformer38 | 92.8 | — | 94.9 | 95.0 | — | |
| Ours | 92.3 | 95.8 | 97.5 | 98.0 | 98.6 | |
| Mixed | MT22,28 | 88.7 | 92.1 | 94.2 | 94.9 | 95.4 |
| MEGAN22 | 86.3 | 90.3 | 94.0 | 95.4 | 96.6 | |
| AT17 | 90.4 | 94.6 | 96.5 | — | — | |
| Chemformer38 | 91.3 | — | 93.7 | 94.0 | — | |
| Ours | 91.0 | 95.0 | 96.8 | 97.0 | 97.3 | |