Parametric Treatment of cDNA Microarray Data

As each set of microarray data is affected by variations in experimental conditions, appropriate normalization processes are required. Various approaches towards such normalization have been proposed and generally involve adjustments to every pair of the data sets, often between the simultaneously hybridizing R and G probes. Chen et al. [1] introduced a model in which each the signal ratio between the probes is normally distributed; the same concept is basically used in the process of the Stanford Microarray Database [6]. Recently, Yang et al. [4] extended this method by stabilizing the average of the signal ratio, which could be biased based on signal intensity. Such non-parametric methods have been further improved by introduction of a parameter that also maintains the variance in the signal ratios [2, 3]. An alternative and simple parametric method that assumes lognormal distribution of data has also been widely employed. Besides its simple ease in performing the calculations, it is also capable of determining the data z-scores, a possible common unit for data comparisons. However, microarray data often have a skewed distribution, and this low fidelity to the distribution model severely limits the accuracy of the data. In the parametric process, the estimation of background, which is defined as the constant part of additive noise [1], can be a major source of normalization inaccuracies. In most cases, the background is estimated from the area outside the DNA spot of the image data; based on the assumption that the background on a tip is uniform. However, as DNA spots and also the intact surface of the tip can bind free dyes at different densities, such estimations are clearly prone to errors. To check this possibility, an alternative estimation method that is based on signal intensity of control DNA is tested against the conventional image-based method. The distributions of both sets of processed data are investigated using probability plots.