Figure 3

IRON normalization. Scatterplots of log₁₀ non- background-subtracted probe intensities are used to demonstrate the IRON normalization algorithm. Points are colored by density (red: high, blue: low) in A, D, E, F. Initial points (A), are filtered in (B) to remove extreme low- and high- intensity points (grey). Iterative rank-order pruning (B) further removes outlier points at each iteration (red: high %Δ rank, blue: low %Δ rank), leaving the final training set (magenta) differing by ≤ 1 %Δ rank. Sparsely sampled regions (red) within the training set are up-weighted in (C), prior to fitting a smoothed piece-wise linear curve (green) in (D). Non-linear intensity-dependent scaling is applied to the sample (GSM467826) using the fit curve, so as to shift the fit curve onto the X,Y diagonal (E). The non-linear scaling resulting from quantile normalization (F), is unable to cope with the asymmetry between the samples, effectively fitting a line (diagonal in green) between the highest density distribution and the lesser density subpopulation, resulting in greater non-linear distortion than originally present in the unprocessed data.