Figure 6

Cross validation results for function prediction methods in the All network. a) Illustration of Methods 1–7 on example cluster, b-d) Boxplots of the percentage of correct function assignment, over 100 random partitions into training (90%) and test (10%) sets, using the following prediction methods: 1- SAMPLEUNIF (sampled uniformly at random), 2- SAMPLEGLOBAL (sampled based on the global distribution), 3- MAJORITYNEIGH (the majority assignment to immediate neighbors), 4- MAJORITYCLUST (the majority assignment within the cluster), 5- SAMPLECONNECT (sampled from the distribution of functions for a given connectivity), Method 6- SAMPLENEIGH (sampled from preference bias for non-like pairs by determining the majority function of immediate neighbors, selecting proteins in the cluster with that function and sampling from the distribution of functions assigned to their neighbors), and 7- NEIGHCONNECT (sampled from the combined distribution of SAMPLENEIGH and SAMPLECONNECT). Methods 8–10 apply Methods 5–7 to proteins incorrectly predicted by Method 4-MAJORITYCLUST while Methods 11–15 apply Methods 3–7 to hub proteins. a) KEGG b) COG c) TIGR.