On the relevance of technical variation due to building pools in microarray experiments

Pooled samples are frequently used in experiments measuring gene expression. In this method, RNA from different individuals sharing the same experimental conditions and explanatory variables is blended and their concentrations are jointly measured. As a matter of principle, individuals are represented in equal shares in each pool. However, some degree of disproportionality may arise from the limits of technical precision. As a consequence a special kind of technical error occurs, which can be modelled by a respective variance component. Previously published theory - allowing for variable pool sizes - has been applied to four microarray gene expression data sets from different species in order to assess the practical relevance of this type of technical error in terms of significance and size of this variance component.

The number of transcripts with a significant variance component due to imperfect blending was found to be 4329 (23 %) in mouse data and 7093 (49 %) in honey bees, but only 6 in rats and none whatsoever in human data. These results correspond to a false discovery rate of 5 % in each data set. The number of transcripts found to be differentially expressed between treatments was always higher when the blending error variance was neglected. Simulations clearly indicated overly-optimistic (anti-conservative) test results in terms of false discovery rates whenever this source of variability was not represented in the model.

Imperfect equality of shares when blending RNA from different individuals into joint pools of variable size is a source of technical variation with relevance for experimental design, practice at the laboratory bench and data analysis. Its potentially adverse effects, incorrect identification of differentially expressed transcripts and overly-optimistic significance tests, can be fully avoided, however, by the sound application of recently established theory and models for data analysis.