Microarray Normalization Revisited for Reproducible Breast Cancer Biomarkers

Figure 10

Effect of bias correction via mismatch probes: MAS5 versus PLIER versus RMA, applied to single studies. (a) Distribution profiles of log₂-normalized expression values of all 54675 probe sets (normal distribution kernel method). 80%-quantiles, q_0.80, are shown as solid circles. (b) Histogram of correlation distances of log₂-normalized expression values between pairs of pipelines. e.g., for the comparison MAS5Single vs. RmaRSingle (red histogram, labelled comparison “B” in Figure 1 and Table 3): All samples are normalized by MAS5RSingle and also by RmaRSingle, and then examined in pairs: Within each pair, the correlation distance is computed from all 54675 probe sets (see text and caption of Table 3). The frequency distribution of distances is shown as a histogram. These distributions are quantitatively characterized by mean and quantiles (median, q_0.15, q_0.85) in Table 3. For transparency, we show the median value also aside the histogram. Panel (c): For all 54675 probe sets of each sample, the 80% quantile (q_0.80) is computed. This is done for each of the 3753 samples and plotted over sample-ID (-axis). Sample-ID runs consecutively over all 38 studies. The overall q_0.80 (over all probe-sets and samples) is marked as full circle in the expression profile of each pipeline in panel (a).

(a)

(b)

(c)