Background: Extracting biological knowledge from large amounts of gene expression information deposited in public databases is a major challenge of the postgenomic era. Additional insights may be derived by data integration and cross-platform comparisons of expression profiles. However, database meta-analysis is complicated by differences in experimental technologies, data post-processing, database formats, and inconsistent gene and sample annotation.
Results: We have analysed expression profiles from three public databases: Gene Expression Atlas, SAGEmap and TissueInfo. These are repositories of oligonucleotide microarray, Serial Analysis of Gene Expression and Expressed Sequence Tag human gene expression data respectively. We devised a method, Preferential Expression Measure, to identify genes that are significantly over- or under-expressed in any given tissue. We examined intra- and inter-database consistency of Preferential Expression Measures. There was good correlation between replicate experiments of oligonucleotide microarray data, but there was less coherence in expression profiles as measured by Serial Analysis of Gene Expression and Expressed Sequence Tag counts. We investigated inter-database correlations for six tissue categories, for which data were present in the three databases. Significant positive correlations were found for brain, prostate and vascular endothelium but not for ovary, kidney, and pancreas.
Conclusion: We show that data from Gene Expression Atlas, SAGEmap and TissueInfo can be integrated using the UniGene gene index, and that expression profiles correlate relatively well when large numbers of tags are available or when tissue cellular composition is simple. Finally, in the case of brain, we demonstrate that when PEM values show good correlation, predictions of tissue-specific expression based on integrated data are very accurate.