Performance evaluation of different statistical models using simulated data. (a) The observed false positive rate at the significance level of 0.05 for the linear model (lm), generalized linear model (glm), and GLiMMPS. Data were simulated with different sequencing depth with mean total junction reads ranging from 5 to 80, as described in Supplementary Methods in Additional file 1. (b) Receiver operating characteristic (ROC) curve analysis demonstrates that GLiMMPS outperforms the lm and glm models. The ROC curve plots the fraction of true positives called correctly and the fraction of false positives called incorrectly using P-values from each model. The zoomed-in figure shows the part of the ROC curve where the false positive rate is in the range of (0, 0.2).
Zhao et al. Genome Biology 2013 14:R74 doi:10.1186/gb-2013-14-7-r74