Table 2 Comparisons of average accuracies (%) against state-of-the-art models for retrosynthesis on the USPTO-50k dataset
From: CNN-based two-branch multi-scale feature extraction network for retrosynthesis prediction
Model | Top-k accuracy % | |||
|---|---|---|---|---|
1 | 3 | 5 | 10 | |
Template-free | ||||
Seq2Seq | 37.4 | 52.4 | 57.0 | 61.7 |
Transformer | 42.7 | 63.9 | 69.8 | \(\setminus\) |
Hasic’s | 47.2 | 55.7 | 61.5 | 65.1 |
GTA | 47.3 | 67.8 | 73.8 | 80.1 |
G2Gs | 48.9 | 67.6 | 72.5 | 75.5 |
Tetko’s | 53.5 | 69.4 | 81.0 | 85.7 |
Template-based | ||||
RetroSim | 37.3 | 54.7 | 63.3 | 74.1 |
NeuralSym | 38.5 | 55.7 | 61.3 | 66.6 |
GLN | 52.5 | 69.0 | 75.6 | 83.7 |
GraphRetro | 53.7 | 68.3 | 72.2 | 75.5 |
EBMs | 55.2 | 74.6 | 80.5 | 86.9 |
CNN-TMN (Plain) | 49.1 | 64.4 | 67.6 | 72.6 |
CNN-TMN (Aug) | 61.1 | 79.1 | 83.9 | 87.7 |