A new pairwise kernel for biological network inference with support vector machines

BMC Bioinformatics

Table 1 Performance on reconstruction of the yeast metabolic networks.

	MLPK		TPPK		MLPK + TPPK		Direct
Data	Accuracy	AUC	Accuracy	AUC	Accuracy	AUC	AUC
Expression	77.9 ± 1.2	84.8 ± 1.2	77.4 ± 0.9	84.1 ± 0.4	78.2 ± 0.9	84.9 ± 1.3	51.9 ± 1.6
Localization	63.8 ± 2.2	67.5 ± 3.0	62.4 ± 1.0	65.6 ± 0.8	64.4 ± 0.9	66.3 ± 1.0	55.1 ± 1.4
Phylogenetic profile	79.5 ± 0.9	84.3 ± 0.9	77.7 ± 1.6	83.6 ± 1.7	80.7 ± 0.8	85.4 ± 1.1	60.7 ± 1.4
Yeast two-hybrid	75.9 ± 1.2	82.5 ± 1.4	59.4 ± 1.0	65.4 ± 1.7	76.7 ± 0.8	83.0 ± 0.4	51.6 ± 1.4
Sum	83.9 ± 0.7	91.6 ± 0.5	84.0 ± 0.7	91.2 ± 0.4	83.9 ± 0.9	91.5 ± 0.6	60.6 ± 1.3
Pairwise sum	81.4 ± 0.5	89.0 ± 0.4	80.7 ± 1.1	88.6 ± 0.6	81.6 ± 0.7	89.2 ± 0.8	-

The table lists, for each type of data, the accuracy and area under the ROC curve obtained by each pairwise kernel. Values in the tables are means and standard errors in a 3 × 5 cv experiment. TPPK is the tensor product pairwise kernel, and MLPK is the metric learning pairwise kernel. The column MLPK + TTPK shows the results when an SVM is trained with the sum of the MLPK and TPPK pairwise kernels. The row Sum shows the results when the kernel between the genes is the sum of the expression, localization, phylogenetic profile and yeast two-hybrid kernels. The line Pairwise sum shows the results obtained with the SVM when the pairwise kernel used is the sum of pairwise kernels derived from the expression, localization, phylogenetic profile and yeast two-hybrid kernels, respectively. The Direct column shows the result of the direct method, where gene pairs are ranked according to their distance as defined by each kernel to predict edges.

ISSN: 1471-2105