Table 6 The top-N accuracy of our model and the proportion of invalid SMILES in the N-th prediction with different decoding strategies in product prediction on USPTO MIT (A) and in single-step retrosynthesis on USPTO 50K (B). The difference in accuracy between standard and speculative beam search is negligible
(A) Product prediction | Top-1, % | Top-3, % | Top-5, % | |
|---|---|---|---|---|
Accuracy | Beam search | 88.425 | 93.690 | 94.733 |
SBS | 88.425 | 93.690 | 94.720 | |
Pred. 1, % | Pred. 3, % | Pred. 5, % | ||
Invalid SMILES | Beam search | 0.232 | 8.270 | 13.182 |
SBS | 0.232 | 8.275 | 13.285 |
(B) Single-step retrosynthesis | Top-1, % | Top-5, % | Top-10, % | |
|---|---|---|---|---|
Accuracy | Beam search | 52.077 | 82.069 | 88.918 |
SBS | 52.077 | 82.069 | 89.038 | |
SBS (smart drafts) | 52.077 | 82.069 | 88.978 | |
Pred. 1, % | Pred. 5, % | Pred. 10, % | ||
Invalid SMILES | Beam search | 0.799 | 3.534 | 8.107 |
SBS | 0.799 | 3.534 | 8.007 | |
SBS (smart drafts) | 0.799 | 3.534 | 8.187 |