Selectivity profiling of BCRP versus P-gp inhibition: from automated collection of polypharmacology data to multi-label learning

Table 2 10-fold cross-validation results of selected classifiers built on the 4-class classification dataset

Algorithm	Descriptor set	Accuracy (%)	Kappa	AUC
MultilayerPerceptron, training time = 600	MOE2D	55.3	0.39	0.81
MultilayerPerceptron, training time = 600	MACCS	59.0	0.44	0.79
K nearest neighbours, k = 5	MOE2D	60.2	0.46	0.80
K nearest neighbours, k = 5	MACCS	57.1	0.41	0.78
K nearest neighbours, k = 5	ECFP	54.7	0.40	0.80
RotationForest, iterations = 50	MOE2D	60.9	0.47	0.85
RotationForest, iterations = 50	MACCS	62.7	0.50	0.84
SVM with polynomial kernel	MOE2D	57.1	0.42	0.78
SVM with polynomial kernel	MACCS	62.7	0.49	0.80
SVM with polynomial kernel	ECFP	70.2	0.60	0.84
RandomForest, 200 trees	MOE2D	60.9	0.47	0.85
RandomForest, 200 trees	MACCS	60.9	0.47	0.86
RandomForest, 200 trees	ECFP	67.1	0.55	0.88

When no results are shown for the descriptor set ECFP, it is because the computational time and/or memory needed were too large. In italic letters, the model that gives the best cross-validation results

ISSN: 1758-2946