Table 1.

Results of the benchmark dataset: comparative analysis

Method classifier	Our approach (SATuRNo)	PC-based method			Partial PC-based method			Matisse method
		IB1	C4.5	NB	IB1	C4.5	NB	IB1	C4.5	NB
Number of genes	34	29			79			20
Representative Acc.	90.9%	87.87%	78.78%	87.87%	93.93%	63.63%	84.84%	90.9%	87.87%	87.87%
Weighted Avg. TPR	0.90	0.87	0.78	0.87	0.93	0.63	0.84	0.90	0.87	0.87
Weighted Avg. FPR	0.08	0.16	0.24	0.13	0.06	0.39	0.15	0.14	0.13	0.13
Specificity	0.84	0.76	0.69	0.84	0.92	0.53	0.84	0.76	0.84	0.84
Sensitivity	0.95	0.95	0.85	0.90	0.95	0.70	0.85	1	0.90	0.90
AUC	0.89	0.86	0.70	0.91	0.937	0.64	0.92	0.88	0.94	0.93

We compared our approach to other published techniques on the basis of two tasks: network inference and classification using the inferred networks. For network inference, we applied a PC-based method (?), Partial PC-based method (De la Fuente et al., 2004) and the Matisse tool (Ulitsky and Shamir, 2007). After building class-specific networks, the genes involved in these networks were used as inputs to several classifiers: nearest neighbors (IB1), decision trees (C4.5 algorithm) and Naive Bayes classifiers. Several measures as representative accuracy, TPR, FPR, specificity, sensitivity and AUC values are shown. The representative accuracy is the proportion of correctly classified patients. The TPR and FPR are the weighted average true and positive rate. The specificity is the proportion of control patients, which were recognized as control category. The sensitivity is the proportion of disease patients, which were recognized as disease category. Finally, the AUC values represent the area under the receiver operating characteristic curve. Avg., average.