I'm a bit confused as to how exactly the median polish in the RMA works in Robin and I'm hoping you could clarify it for me. I understand that a matrix is compiled with the row and column effect subtracted until the matrix is zero and the row and column effects are then added. I'm uncertain though as to what is in the column and what is in the row? Is a matrix compiled for each probe set for a single gene across the replicated microarrays to give a single value for a probe or is the matrix compiled for all the probe sets for one microarray? Any help will be greatly appreciated.

Hi Martin,

we recently published a paper discussing the median polish procedure used in RMA and proposed a
modification named tRMA. Please have a look at the paper (http://www.biomedcentral.com/1471-2105/11/553)
for a detailed discussion.

best regards,
Marc

Thank you Marc. That helped a lot. So it carries out the median polish not just across the replicates for a particular treatment, but across all the microarrays regardless of treatment? Surely a lot of biological information is lost in that case?

Best regards,
Martin

Dear Martin

as posted below (don't know why it only showed up now) (t)RMA indeed uses all microarrays.
The rest is done in differential expression analysis.
One could surely also think of whether to fit models incorporating sample groups (this is likely similar to "ANOVA normalization" for 2 color arrays). But (t)RMA coupled to limma will in any case give you very good results.
So the upshot is, if there is any statistically recoverable biology (since you would have to introduce some statistical models, taking this pseudo-technical replicates into account) lost, it usually doesn't hurt.

If you wanted to you could try leaving things unsummarized and then treat the probes as technical replicates in limma. (This is not well thought through though). But my hunch is that you will likely rather have too many probe-sets declared changed then.

Cheers,
Björn

Hi Martin,

certainly, the method affects probe sets with a weaker signal or inconsistent probe sets that match more than
one target transcript more strongly than probe sets that display a strong, consistent signal. Hence, on the lower
end of the intensities measured, some biological information might be lost (but this information is less reliable
anyway because it is closer to the noise).

bests,
Marc

Hi Marc,
What I don't understand in the pdf you sent me (thank you for sending me that by the way) is that the column effects are reduced to zero and the grand effect is added to this. But since the column's are samples it seems reducing all the columns to zero is removing the variation between the samples and making them the same.

Thanks,
Martin

To be a bit more verbose: median polish is part of the RMA (and tRMA) procedure to summarize probes into a single probe-set value. This is the "summarization step". In this case for each Probe set (consisting of often 11 probes) and across ALL arrays the median polish procedure is performed. You might wonder why all arrays are used, but bear in mind that this is normalization where you don't need to specify groups just yet.

As RMA shows some (for us) unwanted behavior for small, odd sample numbers, we introduced tRMA (Giorgi et al). Now RMA and tRMA just differ by where the sweep goes through first. Probably you want to have a look at this figure to get a grip on the median polish procedures http://www.springerlink.com/content/b583112874337q52/MediaObjects/1471-2105-11-553-S8.PDF.


Best Wishes,
Björn