Table 2.
Prediction accuracy of the reference model for different subsets of the test set of USPTO-50k using a beam search of size 10.
| Top-1 | Top-5 | Top-10 | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Test set augmentation | All | Stereo (20%) | No stereo (80%) | All | Stereo (20%) | No stereo (80%) | All | Stereo (20%) | No stereo (80%) |
| x1 | 48.3 | 44.7 | 49.2 | 73.4 | 67.3 | 74.9 | 77.4 | 71 | 79 |
| x20 | 53.4 | 47.3 | 55 | 80 | 73.3 | 81.9 | 84.2 | 79.2 | 85.4 |
| x100 | 53.5* | 47.1 | 55.1 | 81* | 74.6 | 82.6 | 85.7* | 81.2 | 86.8 |
| MaxFrag,a x1 | 53.5 | 48.7 | 54.7 | 79.2 | 72.7 | 80.9 | 81.6 | 75.1 | 83.3 |
| MaxFrag, x20 | 58.5 | 52 | 60.1 | 84.7 | 79 | 86.1 | 88.6 | 83.6 | 89.8 |
| MaxFrag, x100 | 58.5* | 51.2 | 60.3 | 85.4* | 79.4 | 86.9 | 90* | 85.1 | 91.2 |
aThe classical retro-synthesis accuracy was estimated as the percentage of correctly predicted largest fragments, i.e., “maximum fragment” (MaxFrag) accuracy.
* Indicate the best results.