Figure 11

Performance measures for low, medium and high intensity probesets. For each intensity subset of probesets (low, medium and high), 301 methods (Figure 5) were used to generate expression values. All probes that could not be mapped to the subset of interest were masked, and the remaining probesets were normalized with the FC = 1 probesets within each subset. AUC performance scores for (a) low, (c) medium and (e) high intensity probesets were generated for each method using FC > 1 probesets and mixed probesets that can be aligned to spiked-in transcripts as TPs. True negatives were FC = 1 probesets, empty probesets and mixed probesets that can be aligned only to FC = 1 probesets. (b,d,f) Plots of observed FDR (true q-value) for low (b), medium (d) and high (f) intensity probesets based on the calculated q-values below 0.10. AUC and q-values were generated for each subset of probesets using a FC = 1 normalization using all probesets with present transcripts (Figure 6). These values show that some methods (for example, GC-NSB with vsn probeset-level normalization) perform very differently depending on the subset of probesets used for normalization.