k = 5 cross validation results from SN2-TS. Cross validation results for TSNet and TSNet-shared upon the SN2-TS dataset, pre-trained and trained from scratch. Loss values are coordinate angstroms (Å), while distance error values are Euclidean distance Å. Standard deviation values are in parentheses. Overall best distance error values are achieved by TSNet trained from scratch, while the best loss values are achieved by TSNet when pre-trained with the midpoint prediction task from the QM9. TSNet-distance which trains using distance error performs the worst.
| Model | Train | Validation | ||
|---|---|---|---|---|
| Mean loss | Mean distance error | Mean loss | Mean distance error | |
| TSNet | 0.1455 (0.01) | 0.02672 (0.01) | 0.4576 (0.03) | 0.1831 (0.03) |
| Pre-trained TSNet (midpoint) | 0.06956 (0.004) | 0.07868 (0.006) | 0.3631 (0.04) | 0.2095 (0.03) |
| TSNet-shared | 0.2977 (0.03) | 0.07301 (0.02) | 0.6598 (0.1) | 0.1950 (0.04) |
| Pre-trained TSNet-shared (midpoint) | 0.07941 (0.007) | 0.04978 (0.01) | 0.4181 (0.05) | 0.1925 (0.06) |
| Pre-trained TSNet (energy) | 0.1327 (0.03) | 0.04387 (0.01) | 0.5124 (0.07) | 0.2838 (0.1) |
| TSNet-distance | — | 0.6198 (0.2) | — | 1.828 (0.3) |