This paper presents a new methodology. It describes procedures for normalizing Affymetrix probe-level microarray data and compares their effects on gene expression estimates.

When a microarray experiment is carried out there are two major sources of variation: biological and non-biological. Only the first is of interest, with ideally the non-biological variability being small. Typically, sources of technical variability—for example the settings on the scanner, the amount of cRNA hybridized to the array and the order in which the arrays are hybridized—contribute to this unwanted variation. A normalization procedure attempts to reduce this unwanted variability without also removing the biological differences we are seeking to identify.

This paper presents a number of normalization algorithms and compares their performance using microarray data. After normalization you'd like the variability of non-differential genes to be reduced without also losing the differential genes. We found that methods which used the complete dataset to establish the normalization were superior to those which normalized to a single baseline array. The quantile normalization method was the fastest of the complete data methods and proved to be effective in the bias and variance comparisons.

I became involved with this research as part of my doctoral studies. It is one component of our work on the RMA expression measure for Affymetrix oligonucleotide data.