Large scale statistical analysis of GEO datasets

The problem addressed here is that of simultaneous treatment of several gene expression datasets, possibly collected under different experimental conditions and/or platforms. Using robust statistics, a large scale statistical analysis has been conducted over 20 datasets downloaded from the Gene Expression Omnibus repository. The differences between datasets are compared to the variability inside a given dataset. Evidence that meaningful biological information can be extracted by merging different sources is provided. Background Many genomewide expression datasets have been published during the past ten years. Repositories, such as the Gene Expression Omnibus (GEO) database [1], have made available an impressive wealth of data. Using them as a whole, instead of restricting statistical studies to one particular dataset, is tantalizing. Two recently published R/Bioconductor packages [2, 3] provide various tools for merging datasets coming from different studies. However, a serious doubt has been cast by Haibe-Kains et al. [4], after comparing two large scale pharmacogenomic studies: whereas both studies had a good overall correlation, important discordances could be observed. Thus, the following crucial question remains to be answered: is it statistically legitimate to merge datasets coming from different studies? An attempt at answering this question is reported here. Merging different datasets, requires prior checking that the information they contain is compatible, and hence that detected differences between gene expressions under different conditions are not artifacts, due to experimental or data processing methods. An obvious obstacle to simultaneous treatment is that expression data collected under different experimental conditions and/or platforms usually have incompatible