Table 5.

Classification Results for SVM Models with 18 Additional Variable Combinations

Noa.	Variable Set	Data Setb	Sensitivity (%)	Specificity (%)	Accuracy (%)
7	h-CLAT + Toolbox + 6 properties	Training	97	97	97
		Test	95	100	96
8	KeratinoSens + Toolbox + Avg.Lys.Cys + 6 properties	Training	99	100	99
		Test	84	100	89
9	KeratinoSens + h-CLAT + Avg.Lys.Cys + 6 properties	Training	97	92	96
		Test	90	86	89
10	h-CLAT + Toolbox + Avg.Lys.Cys + 6 properties	Training	96	96	96
		Test	84	100	89
11	KeratinoSens + h-CLAT + Toolbox + 6 properties	Training	96	96	96
		Test	90	86	89
12	h-CLAT + KeratinoSens + 6 properties	Training	94	89	93
		Test	90	86	89
13	h-CLAT + Avg.Lys.Cys + KeratinoSens + Toolbox + Log P	Training	91	96	93
		Test	90	86	89
14	h-CLAT + Avg.Lys.Cys + 6 properties	Training	96	92	95
		Test	84	86	85
15	Avg.Lys.Cys + Toolbox + 6 properties	Training	91	100	94
		Test	79	100	85
16	h-CLAT + 6 properties	Training	87	89	87
		Test	90	86	89
17	h-CLAT + Toolbox + Log P	Training	81	92	84
		Test	84	100	89
18	Avg.Lys.Cys + KeratinoSens + 6 properties	Training	93	96	94
		Test	74	86	77
19	Avg.Lys.Cys + KeratinoSens + Toolbox + Log P	Training	88	92	89
		Test	79	86	81
20	Avg.Lys.Cys + 6 properties	Training	85	100	89
		Test	74	100	81
21	Toolbox + 6 properties	Training	90	81	87
		Test	84	71	81
22	KeratinoSens + Toolbox + 6 properties	Training	91	85	89
		Test	74	86	77
23	KeratinoSens + 6 properties	Training	79	89	82
		Test	74	86	77
24	6 properties only	Training	68	89	73
		Test	73	71	73

Abbreviations: 6 properties = molecular weight, log octanol:water partition coefficient, log water solubility, log vapor pressure, melting point, and boiling point; Avg.Lys.Cys = average depletion for lysine and cysteine peptides from the direct peptide reactivity assay; h-CLAT = human cell line activation test; log P = log octanol:water partition coefficient; No. = number; SVM = support variable machine; Toolbox = read-across using QSAR Toolbox.

^aModels are listed in descending order of the average accuracy of the test and training sets.

^bThe training set of 94 substances contains 68 LLNA sensitizers and 26 LLNA nonsensitizers. The test set of 26 substances contains 19 LLNA sensitizers and 7 LLNA nonsensitizers.