Figure 5. Performance on networks known to be enriched with true interactions a priori and depleted of true interactions.

The precision-recall curve (A), receiver operating characteristic curve (B), the significance level of filtered networks (C), the degree of IPA enrichment in the top ranked set of genes (D), number of edges inferred (E), and the connectivity distributions at varying confidence score thresholds are shown for two different types of filters for the IPA enriched and IPA depleted networks. The fidelity score cutoff was normalized to facilitate comparison between the two filters. The black line represents the results obtained when filtering the networks with only the confidence metric for the IPA enriched networks. The blue line shows the results obtained when the networks are filtered with both the confidence metric and the semantic similarity metric using the optimal filter settings for the IPA enriched networks. The red line, when present, signifies the results obtained when using the networks filtered with the optimal filter settings. All genes in the IPA depleted network had semantic similarity metrics of 0, so this metric cannot be used to filter the network. For (C), the missing segments after the vertical dotted line correspond to p-values = 0. For the (A) and (B), the results obtained for the IPA depleted networks are not shown due to the sparse IPA validated interactions. The dotted lines represent random prediction for (A) and (B) and the 0.05 significance level for (C).