Home | Business News | Browse by Publication | J | Journal of the American Statistical Association

Semilinear high-dimensional model for normalization of microarray data: a theoretical analysis and partial consistency; Comment.

Article Excerpt
The normalization of cDNA arrays is an issue of high practical relevance, being an initial necessary step in the analysis of microarray data, which are nowdays used extensively in genetics and biological research. The impact of statistics on this area can be described as both considerably high and disappointingly low. Statisticians (e.g., Tseng, Oh, Rohlin, Liao, and Wong 2001; Yang et al. 2002) first pointed out that the "global normalization" technique proposed by the developers of array technology overlooked a series of measurable effects and in turn suggested more flexible and sensitive strategies whose value has been recognized by practitioners. Indeed, even commercial software currently incorporates methodology inspired by such contributions as those by Tseng et al. (2001) and Yang et al. (2002). Fan, Peng, and Huang further enlarge the bag of statistical tools available for normalization. This represents a clear success. Nonetheless, disappointing that evidence from statistics has not motivated--so far--the development of more reliable technology or better understanding of the nature of these biases. For example, the results from different normalization strategies, including the one described in the present article, clearly indicate a strong intensity effect (m) that is highly variable from slide to slide. Although statistical correction is possible, there is very little understanding of the biochemical process that produces such biases. I wish that this statistically detected distortion was taken more seriously and that more energy was invested in the understanding of its basis and development of better technology.

The normalization strategy described in this article exploits the presence on one array of multiple spots measuring the expression value of the same gene. This design allows the authors to do without the assumption of zero average expression change across all (or a crudely identified large subset of) genes with similar intensity. The analysis of the experiments for which this assumption is more problematic will benefit most from the suggested methodology. One such case is represented by custom arrays, where the spotted genes are preselected according to the "suspicion" that they may be affected by the treatment under investigation. For example, consider arrays that are printed only with genes believed to be expressed in a specific tissue ("brain array") and used to study the effects of biological treatments known to impact...

View this article FREE - Now for a Limited Time, try Goliath Business News
Free for 3 Days!