TABLE 2.
Performance metrics of various machine learning models on the stress set when correlating the PANTHER Scores with known experimental ΔG values
| Model | Stress set | |||
|---|---|---|---|---|
| r | MAE | rs | ρ-value | |
| Random Forest Regression | 0.64 | 1.63 | 0.64 | 6.02 × 10−14 |
| GBoosting Regression | 0.59 | 1.97 | 0.56 | 1.34 × 10−10 |
| Neural Network | 0.53 | 2.01 | 0.59 | 7.03 × 10−12 |
| Stacked Ensemble | 0.59 | 1.68 | 0.59 | 3.02 × 10−12 |
| XGBoosting Regression | 0.59 | 1.67 | 0.59 | 8.54 × 10−12 |
| Linear Regression | 0.44 | 3.58 | 0.57 | 3.02 × 10−12 |
| PredPRBA | 0.18 | 2.16 | 0.19 | 0.05 |
| PRA-Pred | 0.18 | 2.10 | 0.25 | 9.19 × 10−3 |
-
Metrics include Pearson correlation coefficient (r), mean absolute error (MAE), Spearman's rank correlation coefficient (rs), and ρ-value. Random Forest Regression demonstrates the highest predictive accuracy on both data sets. The table also includes the performance achieved by the other two tested methods (i.e., PredPRBA and PRA-Pred), which will be discussed in the subsection “Comparison of PANTHER Score with existing functional software.”
