Figure 1
Overall performance of the various algorithms' capacity to predict mouse gene GO annotation. The performance of each general strategy ('networkfull', network-based prediction including expression data; 'networkslim', network-based prediction excluding expression data; and 'classifier', naïve Bayes classifiers) as well as several methods of combining the networkfull and classifier scores ('mean', arithmetic mean of network and classifier scores; 'min', minimum of their scores; and 'max', maximum of their scores) is plotted as (a) the mean AUC and (b) the average APR across all Gene Ontology (GO) annotations averaged across ten-fold predictions in the indicated hierarchies (BP, biological process; CC, cellular component; MF, molecular function) and annotation specificities (terms annotating 3 to 10, 11 to 30, 31 to 100, or 101 to 300 genes). The network approach clearly outperforms the classification approach on the infrequent annotations ('3 to 10' and '11 to 30'), while the two methods perform nearly equivalently on the frequent annotations ('31 to 100' and '101 to 300'). The mean and max combinations generally perform slightly better than either of their constituents (networkfulland classifier). The full network shows a significant advantage over the slim network for CC terms and, to a lesser degree, for BP and MF terms. AUC, area under the receiver operating characteristic.