Table 1.
The best model for each electrostatic feature on a subset with 1000 data points
| Electrostatic feature | Sensitivity | Specificity | Precision | F1 |
|---|---|---|---|---|
| ESP_T (Feature 14) | 0.06 | 0.99 | 0.63 | 0.10 |
| AVE_ESP (Feature 17) | 0.10 | 0.98 | 0.63 | 0.26 |
| AVE_ESP1 (Feature 18) | 0.10 | 0.98 | 0.53 | 0.17 |
| RANK_AVEESP1 (Feature 20) | 0.16 | 0.96 | 0.44 | 0.23 |
| ESP_T_0.1 (Feature 21) | 0.08 | 0.98 | 0.54 | 0.14 |
| AVE_ESP_0.1 (Feature 24) | 0.13 | 0.98 | 0.55 | 0.21 |
| AVE_ESP1_0.1 (Feature 25) | 0.09 | 0.99 | 0.59 | 0.16 |
| RANK_AVEESP1_0.1 (Feature 27) | 0.18 | 0.97 | 0.53 | 0.27 |
| ESP_T_0.3 (Feature 28) | 0.25 | 0.87 | 0.31 | 0.28 |
| AVE_ESP_0.3 (Feature 31) | 0.25 | 0.88 | 0.31 | 0.28 |
| AVE_ESP1_0.3 (Feature 32) | 0.30 | 0.88 | 0.35 | 0.32 |
| RANK_AVEESP1_0.3 (Feature 34) | 0.22 | 0.93 | 0.42 | 0.29 |
| ESP_T_0.5 (Feature 35) | 0.33 | 0.88 | 0.38 | 0.35 |
| AVE_ESP_0.5 (Feature 38) | 0.23 | 0.97 | 0.60 | 0.33 |
| AVE_ESP1_0.5 (Feature 39) | 0.31 | 0.94 | 0.54 | 0.40 |
| RANK_AVEESP1_0.5 (Feature 41) | 0.26 | 0.94 | 0.47 | 0.33 |
The number of the feature, as listed in Supplementary Table S3, is given in parentheses. Models were trained on the individual features related to electrostatics, in search of individual features with high predictive value. Predictive value in this case was measured using the F1 Score, which favored models having high Specificity but low to moderate Sensitivity. The best results were obtained for features calculated using the shell between the surfaces offset 0.5 Å from the van der Waals and solvent accessible surfaces. Based on this observation, we later checked whether use of more distant surfaces improved our final model, but this was not the case. The predictive value of the models trained on these individual features has F1 Scores that are fairly low, but in combination with other features, we will derive significantly better models.