Figure 6

Measures of performance. (a) Plot of mean log2 fold changes for FC = 2, empty and FC = 1 probesets for all 300 methods to generate probeset expression values. The mean was generated from the probesets with the lowest Cyber-T P values, the lowest 90% for TPs (125 out of 139 for FC = 2) and the lowest 1% for FPs (101 out of 10,104 for empty; 25 out of 2,495 for FC = 1). (b) Plot of ratio of mean fold change of TPs (FC = 2) divided by mean fold change of FPs (empty or FC = 1). (c) Plot of AUC scores for all probesets and for probesets that can be aligned to present transcripts (FC = 1, FC > 1 and mixed probesets). TPs were FC > 1 probesets and mixed probesets that could be aligned to spiked-in transcripts. All other probesets are true negatives. The plot also includes AUC scores using the FP rate of empty probesets to show which methods work best to reduce FDRs associated with present or absent transcripts. (d) Plot of observed FDR (true q-value) based on the calculated q-values below 0.10 when considering only probesets with present transcripts. To show the contribution of probesets with absent transcripts to FDRs, the plot also includes the observed FDR when all probesets are used.