Skip to main content
Fig. 2 | Genome Biology

Fig. 2

From: Do count-based differential expression methods perform poorly when genes are expressed in only one condition?

Fig. 2

The effect of scale that signal-to-noise is calculated on. ac Mean–variance relationships for different scales of the original all-zero-in-one-condition data. df Corresponding ROC curves for the ENCODE dataset (GM12892 cells to H1-hESC), using S/N to set the true labels. Here, the signal-to-noise (S/N) is calculated from (trimmed mean of M-values-normalized) counts-per-million and used for all methods. Linear is equivalent to Rapaport’s method, where S/N is calculated on the counts-per-millions. Log represents S/N calculated on log-transformed counts-per-million. vsn represents S/N calculated on variance-stabilized data [14]. ROC curves employ the same labels across all methods: the top 40 % of S/N are used as true DE genes whereas the lowest 20 % are false. Each method’s P value is used for ranking the genes. FPR false positive rate, TPR true positive rate

Back to article page