Microarrays are powerful and cost-effective tools for large-scale analysis of gene expression. While the utility of this technology has been established [1, 2], analytical methods are evolving and a matter of contention. Key among the more controversial aspects is the treatment of data from weak spots, which significantly influences outcomes. For example, ratio analysis is commonly employed to determine expression differences between two samples. However any procedure that uses raw intensities to infer relative expression is limited due to the fact that accuracy is signal level dependent, with variation increasing dramatically for low intensity signals [1, 3]. Several methods have been developed to diminish the influence of additive noise. One solution is to ignore any genes whose transcripts are present at a low total abundance, to exclude weak spots – arbitrarily (in Kooperberg etal., [3] an intensity cutoff was used such that the relative error in ratios was less than 25%) or with some statistical procedures [4, 5]. Other methods proposed for discriminating expressed genes from those not expressed, such as the method of Greller and Tobin [6], are suitable only for bimodal distributions in which the distribution of intensities for these two subsets are non-overlapping, unlike many empirical data sets. However even procedures for flagging and exclusion of weak spots based on solid statistical background [4] remains problematic as these methods discard potentially valuable data. This issue is compounded by the fact that in biological systems several key regulators may be expressed at low levels presumably so that modulation of these regulators can be tightly controlled [7].

The two main sources of heterogeneity in gene expression variations are indicated in Rocke and Durbin [7] as, the "additive component", prominent at low expression levels, and the "multiplicative component", prominent at high expression levels. The intensity measurement *y*
_{
i, j
}for gene *i* ∈ I = {*i*
_{1}, ..., *i*
_{
n
}} in sample *j* ∈ J = {*j*
_{1}, ..., *j*
_{
m
}} is modeled by the equation: *y*
_{
i, j
}= α_{
i, j
}+ μ_{
i, j
}× *e*
^{η} + ε_{ι, j
}, where α – is the normal background (and independent of expression level), μ – the expression level in arbitrary units, ε – is first within spot error term (additive), and η – is the second error term, which represents the proportional error (multiplicative) [8, 9]. Gene expression data obtained with standard procedure of the local background subtraction will include noisy spots – spots at which expression level is ignorably low and whose intensity ε_{ι, j
}presents additive noise.

We have previously demonstrated the presence of normally distributed noise spots in radioactive labeled Clontech macroarrays and proposed an iterative algorithm for obtaining the parameters of this distribution [2, 10]. Herein we have extended the utility of this approach by demonstrating the noncorrelative nature of these spots in both internal and external comparisons. We also present new algorithm modifications for locating the additive noise in gene expression histograms and for estimation of its distribution parameters. Quantization of additive noise variation can therefore be used as a statistically robust criterion to identify measurable but low-level gene expression. It becomes possible to select even genes that are stably expressed at the additive noise level that can be discriminated from additive noise due to their stability.