Figure 2. Statistical validations and comparisons to single-species expression networks. A. Meta-gene permutations. Shown is the number of meta-gene interactions (y-axis) exceeding a P-value cutoff (x-axis) in networks constructed from: real meta-genes (blue line), a random distribution (red line), and randomly permuted meta-genes (green line). P-values are shown in log-scale. Red arrow denotes P < .05 used in the gene co-expression network. B. Random Halves. We randomly divided the databases of each species into two equally sized sets, and then generated new networks derived from each half of the data for a series of P-values. Shown is the percent of meta-gene pairs with P<p in the first half that have P<0.05 in the second half, for each P-value p. P-values are shown in log-scale. Three additional randomizations gave identical results. Red arrow denotes P < .05 used in the gene co-expression network. C-F. Comparison of multiple species to single-species expression networks. We constructed a co-expression network from each species by selecting a Pearson correlation cutoff of k and linked every pair of genes with a correlation of k or higher. We re-iterated this procedure at various settings of k to generate expression networks with differing degrees of coverage and predictive power. We also constructed co-expression networks from multiple species as described above, using not only the cutoff of P < .05 used in the network discussed above but for varying P-values. Shown is a comparison of all methods in terms of their ability to predict functional categories from KEGG. For each functional category, we combined the neighbors in the network of all genes from the category, and plotted the percent of genes from the category that were included (x-axis; coverage) versus the percent of interactions that were between two genes in that category (y-axis; accuracy). We varied the Pearson threshold for constructing the network in each case to obtain different networks that result in different coverage and accuracy.