MA plots (average intensity of the signal vs log-ratio) of (a) ChIP-chip and (b) its corresponding mock control data without normalization. Both show significant dye-bias that cannot be ignored. A direct subtraction of mock control from experiment corrects the systematic errors to some extent, but it also introduces additional noise. This can be seen by how the probes marked in green in (a) are spread out in (b). Conventional within-array normalization using lowess fitting does not work well due to the presence of a large cluster of probes that show binding.