DNA Microarrays are widely used to monitor gene expression differences. The data resulting from a single array experiment reflect gene expression levels of thousands of genes. Some of the observed data points are however not the reflection of true expression levels, but are the consequence of experimental biasses. The most well know is the color-bias, which is the consequence of the molecular affinity of the Cy3 and/or Cy5 fluorochromes to certain DNA sequences resulting in aberrantly measured ratios. Other biasses are related to the use of enzymes and can be the result of the choice of the type of array or the labeling method. In a set of well controlled experiments we have studied the nature of these experimental biasses and we present ways to either eliminate or circumvent the problems they induce. We prepared a large batches of total RNA and amplified antisense RNA of human cell lines (Jurkat cells and U2OS cells) and mouse tissues (spleen, heart, embryo and kidney) and used either enzymatic or chemical methods to label this RNA with Cy-dyes. Our data analysis software (http://dexter.nki.nl ; NKI Normalization Tools, u/p: KWF) employes a Rosetta-like error model to assign p-values to spots and can select ratios that are statistically significant outliers. If we combine Dye-Swap or Reverse Fluor experiments we can specifically eliminate dye-bias effects and leave genuine (gene-expression based) ratio differences intact. In combined experiments, we have measured significant (p<0.01) differences in gene expression ratios less than 1.4 fold if the genes expression levels are sufficiently high enough. The number of features or sequences on a microarray that show dye-bias is not small as is still generally believed. The figure below (panel A) shows the expected multitude of genes that were selected as genuine significant outliers in a mouse spleen-versus-heart microarray experiment (combination of three dye-swap pairs). Strikingly, from the same data we selected those features that showed significant dye-bias (panel B). None of the highlighted features in both panels overlap. This clearly indicates the need for dye-bias removal. We will discuss the methods we use for this and the relevance of other types of experimental biasses during the meeting. The microarrays are prepared by the Central Microarray Facility of the Netherlands Cancer Institute. Mouse cDNA arrays contain 22K NIA clones, Human cDNA arrays contain 18K Invitrogen clones. Website: http://microarrays.nki.nl

[Proc Amer Assoc Cancer Res, Volume 45, 2004]