Unlike genomic sequence data, transcriptomic data points are highly contextual and require a lot of detailed spatiotemporal annotation to realize their full value. Unfortunately such annotation is inconsistent at best and this example serves to illustrate that. The brain is the most complex organ in the body, with expression levels of factors varying widely across different regions and cell types, and even on a circadian level. Without knowing exactly what region of the brain was being assayed, it's difficult to compare two expression levels in two different databases. There's a wider issue here concerning the need for greater granularity of data annotation in transcriptomic datasets but until such times as standards are properly observed, the safest approach is to use these resources as a guide while bearing the above caveats in mind.

I have not seen this, but I would imagine it has to do with different probe sets? If I saw that sort of result, I would likely attempt to confirm with a literature search as well. Are you looking strictly in human or in other species as well?

While not knowing the details of the data sources you cite, it seems a reasonable expectation that such data sources will reflect the origin of the data sourced. Different researchers reporting different results should not be expected to provide correspondences between their disparate data.

Unlike genomic sequence data, transcriptomic data points are highly contextual and require a lot of detailed spatiotemporal annotation to realize their full value. Unfortunately such annotation is inconsistent at best and this example serves to illustrate that. The brain is the most complex organ in the body, with expression levels of factors varying widely across different regions and cell types, and even on a circadian level. Without knowing exactly what region of the brain was being assayed, it's difficult to compare two expression levels in two different databases. There's a wider issue here concerning the need for greater granularity of data annotation in transcriptomic datasets but until such times as standards are properly observed, the safest approach is to use these resources as a guide while bearing the above caveats in mind.

I agree with what was written above. Also, one has to consider if in the original experiment the tissue of interest was included - such as brain in the SAGE analysis. In general, I would think that RNAseq data are the most reliable.

Well, you can have a lot of different reasons. For example, Illumina Body Map is only done on human tissues and I guess not at every developmental stage, so some developmentally regulated genes could be under-estimated in some tissues. CGAP SAGE experiments is also only in mouse and human, so you can miss some data. Moreover SAGE technique relies on 3'end sequencing of polyA mRNA which have a specific consensus cleaving site for a restriction enzyme, so you miss 1) mRNA with no polyA, 2) some which don't have this restriction site, 3) shorter isoforms.

Then you have the differences between NGS and micro-array data and the different sets of probes used for each dataset. If one tissue has been mainly investigated with an incomplete gene probe set, some genes expressed in this tissue will be under-estimated. I did not look much more into BioGPS functioning and I don't know how they translate a bunch of datasets into a single expression level, but some cut-off or threshold effects of these databases might also be at work.

I would therefore try a Pubmed search with your gene(s) and your tissue(s) of interest to look for any previous gene expression data publication. Or maybe look in details the datasets selected by bioGPS.

Tissues contain many different cell-types. I think cell-type specific expression is more informative. RNAseq and real time PCR give more dymatic ranges thus more reliable.

Many thanks for all your helpful comments! I also checked the details about these datasets:

BioGPS - Measurements were obtained for 76 normal human tissues and compartments hybridized against HG-U133A (Affymetrix).

Illumina body map - RNA obtained from 16 normal human tissues was sequenced and mapped to genes via their transcripts.

SAGE - 19 normal human tissues. The expression level of a particular gene in a particular tissue was calculated as the number of appearances of the corresponding tag divided by the total number of tags in libraries derived from that tissue.

So, as I understand it: microarray data (BioGPS) should theoretically more coincide with RNASeq as with SAGE, since SAGE reflects rather a "portion" of gene of interest within particular tissue, right? and, as correctly mentioned by Klaus, 76 tissues were analyzed in BioGPS versus 16 tissues in RNASeq, one should definitely check whether the tissue was generally included in the analysis.

Another thing - protein expression pattern. That would give us probably more conclusive results to compare it with mRNA expression. And that what I found actually very interesting: liver-specific transcription factors, for example, are present also in the other tissues on the mRNA level, but not as protein. What´s the purpose of mRNA synthesis without translation?

Maria- to speak to your last point, a lot of tissue-specific expression is a matter of degree (liver-specific factors are abundant in liver, but may be expressed at very low levels in other tissues), and a lot of control is exerted at the translational level (detection of a particular mRNA does not mean that it will result in detectable levels of protein expression). The sensitivity of newer methods to measuring even low level of RNA transcripts can lead to the presence of signal that may not be biologically relevant in some cases.