Table 2.
Parameters of each model tuned through GridSearchCV (tenfold cross validation).
| Model | Best parameter |
|---|---|
| RFC* [CTL]** (All***) | (criterion = 'entropy', max_features = None, n_estimators = 250) |
| Logistic* [CTL] (SHAP 0.01***) | (class_weight: None, penalty: 'l2', solver: 'liblinear') |
| Logistic [CTL] (SHAP 0.02***) | (class_weight: None, penalty: 'l2', solver: 'liblinear') |
| RFR* [Dys]** (All) | (max_depth = 25, n_estimators = 700) |
| Linear* [Dys] (SHAP 0.01) | (default) |
| Linear [Dys] (SHAP 0.02) | (default) |
| RFR [Exc] (All) | (max_depth = 20, n_estimators = 300) |
| Linear [Exc] (SHAP 0.01) | (default) |
| Linear [Exc] (SHAP 0.02) | (default) |
| Stepwise model [CTL → Dys] (all) |
RFC [CTL]: (criterion = 'entropy', max_features = None, n_estimators = 250) RFR [Dys]: (max_depth = 25, n_estimators = 700) |
| Stepwise model [CTL → Exc] (all) |
RFC [CTL]: (criterion = 'entropy', max_features = None, n_estimators = 250) RFR [Exc]: (max_depth = 20, n_estimators = 300) |
| Stepwise model [CTL → Dys] (SHAP 0.01) |
Logistic [CTL]: (class_weight: None, penalty: 'l2', solver: 'liblinear') Linear [Dys]: (default) |
| Stepwise model [CTL → Exc] (SHAP 0.01) |
Logistic [CTL]: (class_weight: None, penalty: 'l2', solver: 'liblinear') Linear [Exc]: (default) |
| Stepwise model [CTL → Dys] (SHAP 0.02) |
Logistic [CTL]: (class_weight: None, penalty: 'l2', solver: 'liblinear') Linear [Dys]: (default) |
| Stepwise model [CTL → Exc] (SHAP 0.02) |
Logistic [CTL]: (class_weight: None, penalty: 'l2', solver: 'liblinear') Linear [Exc]: (default) |
*RFC: Random forest classifier, RFR: Random forest regression model, Logistic: Logistic regression classifier, Linear: Linear regression model.
**[CTL]: model for predicting CTL level, [Dys]: model for predicting dysfunction score, [Exc]: model for predicting exclusion score, [CTL → Dys]: model for predicting the ICB response using [Dys] model when [CTL] model predicts the CTL level as high, [CTL → Exc]: model for predicting the ICB response using [Exc] model when [CTL] model predicts the CTL level as low.
***All: All miRNA features, SHAP 0.01: miRNAs with feature importance of mean (|Shapley values|) > 0.01, SHAP 0.02: miRNAs with feature importance of mean (|Shapley values|) > 0.02.