Table 3.
Accuracy of the desired physical and chemical properties of molecules generated by DRAGONFLY
| DRAGONFLY method | MW | Rot. B. | HBA | HBD | PSA | LogP |
|---|---|---|---|---|---|---|
| Unit | g mol−1 | # | # | # | Å2 | – |
| MAD | 75.52 | 2.81 | 0.981 | 1.69 | 27.08 | 1.25 |
| Ligand-SMILES | ||||||
| MAE | 7.7 ( ± 0.2) | 0.29 ( ± 0.01) | 0.23 ( ± 0.01) | 0.027 ( ± 0.005) | 4.4 ( ± 0.2) | 0.252 ( ± 0.004) |
| MAD / MAE | 9.8 ( ± 0.2) | 9.9 ( ± 0.48) | 4.3 ( ± 0.21) | 63 ( ± 11) | 6.1 ( ± 0.2) | 4.94 ( ± 0.08) |
| Ligand-SELFIES | ||||||
| MAE | 8.0 ( ± 0.2) | 0.88 ( ± 0.040) | 0.40 ( ± 0.014) | 0.230 ( ± 0.007) | 6.8 ( ± 0.21) | 0.380 ( ± 0.006) |
| MAD / MAE | 9.4 ( ± 0.2) | 3.2 ( ± 0.14) | 2.5 ( ± 0.09) | 7.3 ( ± 0.3) | 4.0 ( ± 0.1) | 3.27 ( ± 0.05) |
| Structure-SMILES | ||||||
| MAE | 12.1 ( ± 0.5) | 0.42 ( ± 0.02) | 0.28 ( ± 0.02) | 0.046 ( ± 0.007) | 4.6 ( ± 0.1) | 0.315 ( ± 0.008) |
| MAD / MAE | 6.2 ( ± 0.3) | 6.7 ( ± 0.3) | 3.5 ( ± 0.2) | 37 ( ± 6) | 5.9 ( ± 0.2) | 4.0 ( ± 0.1) |
| Structure-SELFIES | ||||||
| MAE | 15 ( ± 0.4) | 1.12 ( ± 0.04) | 0.50 ( ± 0.03) | 0.27 ( ± 0.02) | 7.4 ( ± 0.3) | 0.426 ( ± 0.008) |
| MAD / MAE | 5.03 ( ± 0.1) | 2.5 ( ± 0.09) | 2.0 ( ± 0.1) | 6.3 ( ± 0.4) | 3.6 ( ± 0.1) | 2.92 ( ± 0.05) |
Bold indicates if the SELFIES- or the SMILES-based models achieved a higher value for the investigated property. Abbreviations: MAD Mean absolute deviation, MAE Mean absolute error, MW Molecular weight, Rot. B. Number of rotatable bonds, HBA Hydrogen bond acceptors, HBD Hydrogen bond donors, PSA Polar surface area. The numbers are presented as the mean and standard deviation, with a sample size of N = 3, i.e., 3 DRAGONFLY runs, each sampling 2000 SMILES-strings. MAD / MAE yields a number that indicates by which factor a model is better than the MAD.