Figure 2

fRMA can mimic RMA. Distribution of difference in expression estimates for 22283 probesets across 200 breast tumor arrays (GSE11121) between 3 different fRMA implementations and RMA. The 3 fRMA implementations are as follows: (1) the default fRMA implementation (dotted line), (2) fRMA trained on a balanced random sample from the arrays being analyzed (dashed line), and (3) fRMA trained on a balanced random sample from the arrays being analyzed and using the same reference distribution for quantile normalization as RMA.