Table 2. Comparison of classification performance cross datasets.

Training set	Test set
GSE8511	Benign–PCA			PCA–Mets
	GSE3325	GSE32269	GSE35988	GSE3325	GSE32269	GSE35988
DRW-GM	0.9899 ± 0.0294	0.8522 ± 0.1990	0.9836 ± 0.0216	0.9958 ± 0.0170	0.9011 ± 0.0724	0.9994 ± 0.0070
DRW-GM-NMa	0.9866 ± 0.0363	0.8413 ± 0.2080	0.9817 ± 0.0271	0.9947 ± 0.0187	0.8995 ± 0.0786	0.9991 ± 0.0073
DRW	0.9780 ± 0.0404	0.8254 ± 0.2113	0.9827 ± 0.0196	0.9851 ± 0.0419	0.8918 ± 0.0750	0.9989 ± 0.0081
PAC	0.9634 ± 0.0643	0.8139 ± 0.2093	0.9675 ± 0.0405	0.9482 ± 0.1035	0.6645 ± 0.0836	0.9911 ± 0.0300
Mean	0.9450 ± 0.0733	0.6663 ± 0.2530	0.9351 ± 0.0446	0.9105 ± 0.0298	0.5172 ± 0.0788	0.9713 ± 0.0274
Median	0.9402 ± 0.0838	0.5315 ± 0.2121	0.8995 ± 0.0793	0.9036 ± 0.0671	0.7295 ± 0.0689	0.9694 ± 0.0512
Genes	0.9014 ± 0.1231	0.9023 ± 0.1321	0.9682 ± 0.0186	0.8609 ± 0.0703	0.7492 ± 0.0835	0.8654 ± 0.0744

Shown are the average AUC and the standard deviation. The classification evaluation are performed according to cross-dataset experiments. The training set is GSE8511. Three independent test sets are GSE3325, GSE32269 and GSE35988. The classifications between Benign and PCA samples, and between PCA and Mets samples are carried out respectively. The AUC shown in bold is the best AUC for the corresponding paired training-test dataset.

^aDRW-GM-NM: The classification method that uses gene–metabolite graph, but not incorporates differential metabolites for topological importance evaluation.