Hi Hristina

the first thing to do is to compare both designs in order to check if there are enough similarities for further analysis. The second will be to compare the QC of both studies, preventing the comparison of carrots to tomatoes  ;)

after that, I would take some housekeeping genes to adjust both data sets...

fred

Normalising, or attempting to normalise 2 independent experiments as one is fraught with issues, even if the same chip platform is being used for both experiments. It's hard enough making sure any within experiment batch effects are compensated for, let's alone trying to do this for 2 separate experiments.

You should not compare these two datasets directly in a straightforward “lets just normalise it together” method.

There are methods for cross-study normalisation, lots of them - whether you're combining gene expression measures across independent studies. [http://www.ncbi.nlm.nih.gov/pubmed/15231529] [http://www.ncbi.nlm.nih.gov/pubmed/15774008] or combining other measures such as rank-ordering (as in RankProd as discussed) or p-values [http://www.ncbi.nlm.nih.gov/pubmed/12154050] Then there are the Bayesian approaches [http://www.ncbi.nlm.nih.gov/pubmed/16677390]

Every method has caveats, issues and the more trivial the solution the more caveats. There is a BioConductor package called MADAM [http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2848045/] which implements some meta-analysis methods including the RankProducts approaches. There are also cross-study approaches implemented in the web-based analysis tool ArrayMining.net [http://www.arraymining.net/] including empirical Bayes approaches (as in ComBat) [http://jlab.byu.edu/ComBat/Abstract.html], median rank score normalisation, normalised discretization, quantile discretization.

It is better to analyze these two datasets separately. If you want to normalize expression, then you will normalize it using Z-score scaling.

Pooling samples profiled with HTA 2.0 and U133, even if both Affymetrix, is 100% prone to errors. This happens for several reasons.

By design, they probe a different number of genes (47,000 transcripts vs >245.000 transcripts), with a very different number of probes that, in the latter chip type, are distributed along all the exons of a gene and across exon-exon junctions). I guess that even the sensitivity of probes is different between the two.

Thus, an attempt might be that of using any of the tools suggested above for batch effect removal. But. by my experience, even matching transcripts of the two microarray sets, normalizing the former and applying the obtained normalization distribution to the latter, a significant unavoidable batch effect remain. You should have planned this beforehand, and profile a couple of very same samples with both technologies and then using these two profiles to normalize the first and second sets of arrays.

In the fortunate case of having cases and controls in both sets, you might trivially make cases vs controls for the two sets, separately, and working directly with fold-changes. Relative measures should be more reliable in this case.

https://tools.thermofisher.com/content/sfs/brochures/hta_array_2_0_datasheet.pdf