Figure 2
Relative performance of the network and the classifier methods by AUC. Each filled circle represents a specific Gene Ontology (GO) annotation term in the biological process (BP), cellular component (CC), or molecular function (MF) class, plotting AUC for the slim network on the x-axis and for the classifier on the y-axis. For less frequent categories ('3 to 10', '11 to 30'), AUC tends to be higher in the networkslim than in the classifier, with most points falling under the diagonal. For more frequent terms ('31 to 100', '101 to 300'), most points are concentrated near the diagonal, suggesting that the two approaches perform similarly. This network-bias for less frequent terms was observed across all GO classes of BP, MF and CC. It is notable that for the most specific GO terms ('3 to 10'), many GO terms were predicted effectively by the network but not by the classifier with AUCnetwork >> 0.5 and AUCclassifier, although the two approaches used the same data set for training and testing. AUC, area under the receiver operating characteristic.