Figure 2

Dependency of precision and recall on score threshold and level of conservation. a) Results from eight ncRNA search tools run on our test set of pairwise UCSC alignments of ncRNAs compiled from human, mouse, worm, fly, and yeast ncRNA databases (n = 3046). We calculated recall and precision for 100 output score thresholds distributed linearly over the range of output scores, and for five ranges of sequence identity (50–100%). (Only 5.5% of the alignments fall below 50% sequence identity.) Precision and recall were calculated using a shuffled test set with of 100 shuffles for each real ncRNA alignment. The middle column depicts the product of precision and recall (*intensity scaled such that white = 0.3). The two columns on the right show the actual distribution of scores attained for real alignments (third from right) and shuffled alignments (second from right) normalized to the count of the most recurrent score. The far right column shows the shuffled score distribution normalized to the real score counts (100× brighter) to reveal presence of high-scoring shuffles (i.e. false positives that limit precision). b) Distribution of the sequence identity of the pairwise alignments. (Similar figures for ncRNAs in eight different classes are shown in Supplementary Figures 1-8 [Additional file 1]).