Skip to main content

Advertisement

Table 2 Comparison of AD methods applied to the test set of CAESAR BCF model

From: Defining a novel k-nearest neighbours approach to assess the applicability domain of a QSAR model for reliable predictions

Approach IN AD Q 2 OUTSIDE AD
All samples inside (no AD approach) 95 0.797 None
Proposed approach (Euclidean dist., k = 15) 91 0.803 33 61 82 83
Bounding box 95 0.797 None
PCA bounding box 93 0.804 33 40
Convex hull 73 0.789 3 7 9 13 18 33 34 36 37 38 39 40 41 43 51 56 61 72 79 91 92 94
Euclidean dist (95 percentile) 88 0.802 3 33 36 37 40 42 61
Mahalanobis dist (95 percentile) 89 0.791 18 43 54 61 83 91
Classical kNN (Euclidean dist., k = 5) 87 0.797 3 33 34 40 61 82 83 94
Fixed Gaussian kernel 85 0.794 3 24 33 34 40 61 82 83 91 94
Optimized Gaussian kernel 66 0.831 3 912 22 24 33 34 38 40 45 47 51 53 54 56 61 68 69 75 76 80 82 83 87 89 91 93 94 95
Variable Gaussian kernel (k = 15) 81 0.790 3 24 33 34 40 43 61 80 82 83 89 91 94 95
Adaptive Gaussian kernel 88 0.801 3 33 43 61 82 83 91
Fixed Epanechnikov kernel 87 0.799 3 33 40 43 61 83 91 94
Nearest neighbour density estimator (k = 15) 91 0.806 3 33 61 91