Fig. 5

The CS-driven stacked ensemble framework classification performance. The stacking framework collects uncorrelated predictions of base classifiers, strengthening diverse predictions and reduce overfitting in the final predicted model. The results of internal testing and independent validation of the prognostic model were assessed by area under the curve- receiver operating characteristics (AUC-ROC). Herein, the AUC-ROC plots illustrate the augmented classification performance achieved by stacking framework instead of implementing a specific classification algorithm. For internal evaluation, the designated super-learner (DNNs) has obtained 98.8% AUC-ROC (a) while base-learners RF and XGB achieved 88.6% (b) and 79.6% (c) AUC-ROC, respectively. The trained and tested prognostic model administered to identify hits from a small independent validation dataset has achieved a remarkable 83.90% AUC-ROC for the stacked framework (d). In contrast, base-learners, RF, and XGB obtained 81.80% (e) and 80.82% (f) AUC-ROC. The benchmark performance AUC-ROC plots for the CS-driven stacked ensemble obtained 90.2% (g) and base-learners RF and XGB obtained 82.2% (h) and 81.3% (i). Our results from implementing different machine and deep learning algorithms suggested that if any of the algorithm cannot handle input data well, the super-learner could handle the classification and data tasks. From the independent dataset, CS module of A-HIOT identified 35 hit molecules those demand further optimization as per receptor structure and will be considered as input for PS module of A-HIOT