Skip to main content
. 2020 Jan 28;10:1631. doi: 10.3389/fphar.2019.01631

Table 6.

Prediction accuracies of hERG blockade classification models developed by different methods with the same Doddareddy's training set.

Model SE SP MCC Q (%) AUC
Doddareddy's test set (255/P:108, N:147)
Conv-CapsNet 94.4% 89.8% 0.835 91.8% 0.940
RBM-CapsNet 91.7% 92.5% 0.840 92.2% 0.944
CNN 87.0% 85.0% 0.715 85.9% 0.933
MLP 82.4% 86.4% 0.687 84.7% 0.920
DBN 72.2% 80.8% 0.533 80.8% 0.903
SVM 90.7% 84.4% 0.743 87.1% 0.933
kNN 69.4% 96.6% 0.703 85.1% 0.928
Logistic regression 88.8% 83.7% 0.710 85.5% 0.858
LightGBM 79.6% 82.3% 0.617 81.2% 0.810
Doddareddy's external validation (60/P:18, N:42)
Conv-CapsNet 88.9% 71.4% 0.554 76.7% 0.806
RBM-CapsNet 94.4% 71.4% 0.604 78.7% 0.811
CNN 94.4% 52.4% 0.441 65.0% 0.725
MLP 88.9% 57.1% 0.426 66.7% 0.707
DBN 88.9% 52.4% 0.386 63.3% 0.683
SVM 88.9% 52.4% 0.386 63.3% 0.660
kNN 77.8% 52.4% 0.279 60.0% 0.624
Logistic regression 83.3% 52.4% 0.332 61.7% 0.623
LightGBM 61.1% 59.5% 0.190 60.0% 0.609

(TP, true positive; TN, true negative; FP, false positive; FN, false negative; SE (%), sensitivity, SE = TP/(TP + FN); SP (%), specificity, SP = TN/(TN + FP); Q (%), overall accuracy, Q = [TP + TN)/(TP + TN + FP + FN)].