Feature combination networks for the interpretation of statistical machine learning models: application to Ames mutagenicity

Table 5 PubChem descriptor model performance for split chemical space validation sets

Algorithm	Measure	Aliphatic halogen	Aromatic nitro	Aziridine	Bay region PAH	Carboxylic acid	Epoxide	Aromatic amine (primary)	Aromatic amine (secondary)	Aromatic amine (tertiary)
SVM	AUC	0.83	0.78	---	0.73	0.96	0.85	0.82	0.91	0.77
	BAC	0.77	0.64	NaN	0.58	0.91	0.77	0.75	0.82	0.73
	SEN	0.83	0.96	1.00	0.97	0.88	0.87	0.86	0.82	0.83
	SPEC	0.71	0.32	NaN	0.20	0.94	0.67	0.64	0.82	0.64
RF	AUC	0.8	0.8	---	0.72	0.94	0.87	0.84	0.95	0.82
	BAC	0.75	0.64	NaN	0.70	0.9	0.73	0.74	0.85	0.68
	SEN	0.87	0.96	1.00	1.00	0.88	0.92	0.9	0.88	0.78
	SPEC	0.63	0.32	NaN	0.40	0.92	0.54	0.58	0.82	0.57
DT	AUC	0.66	0.58	---	0.65	0.87	0.56	0.71	0.82	0.57
	BAC	0.66	0.55	NaN	0.50	0.82	0.54	0.63	0.82	0.70
	SEN	0.7	0.96	0.92	1.0	0.74	1.00	0.76	0.65	0.83
	SPEC	0.63	0.13	NaN	0.00	0.9	0.08	0.50	1.00	0.57
kNN	AUC	0.77	0.79	---	0.74	0.91	0.70	0.78	0.88	0.78
	BAC	0.73	0.64	NaN	0.5	0.82	0.65	0.69	0.8	0.70
	SEN	0.69	0.96	1.00	1.00	0.79	0.71	0.85	0.88	0.83
	SPEC	0.77	0.32	NaN	0.00	0.85	0.58	0.53	0.73	0.57

Where NaN = not a number result as all predictions were true positive, AUC = area under curve, BAC = balanced accuracy, SEN = sensitivity, SPEC = specificity.

ISSN: 1758-2946