ad 1)

could you specify what exactly is meant by "absolute" (mathematically)? A very direct interpretation is to take the values as positive, independent of their sign (i.e. the "absolute values"). If logFC is a vector of log fold-changes, then abs(logFC) will return the vector of absolute log fold-changes.

ad 2)

I suggest to evaluate gene expressions only *withi* a platform, even also only *within* an experiment. Otherwise there will be a lot of not-well understood systematic differences between data from different experiments and platforms that can easily and severely confound the conclusions.

Normalization is generally a critical and difficult task. Within an experiment and for "biologically comparable samples" quantile normalization seems to work fast and reliable, leading to sensible results. However, I am not aware of a good way to normalize data from different kinds of samples or plattforms in a sensible way so that they can be analyzed together.

A batch correction may be appropriate, the package svg provides the function ComBat. But I am not sure if and how this is applicable in your situation.

I assume that an absolute fold change is the one that uses ratios, rather than negative numbers. So an absolute fold change of 0.5 corresponds to a (conventional) fold change of -2. You take the negative reciprocal to convert from one to the other. However limma works with log 2 values which are negative when less than one. This is the usual way that limma does things. I'm afraid that you will have to try and figure things out from all this. But I'm am suspicious of the claim that they used limma for absolute fold changes.

My advice for anyone trying to trying to normalize data from different platforms is: DON'T! Although you can remove a lot of batch effects with ComBat, you are swamping the system with huge amount of technical variation: different samples, done in different places, done on different days, done on different technologies which is the worst type of technical variation of them all. The purpose of normalization is to remove technical variation and leave only biological variation, but you are throwing every conceivable piece of technical variation together. I personally would have no confidence in any result you derive from this and I suspect your combined gene list will be a piece of junk representing lists of genes could well correspond to the use of different technologies or different days of the week etc.

If you want to do comparisons between different experiments, analyse each of them individually and compare the resultant gene lists that you derive. What you lose in n-numbers you will more than compensate for with reduced technical variation. That's my advice anyway.

I absolutely agree with Timothy!

Just one note:

fold-changes are neccesarily positive. There are no negative fold-changes. There are only negative log fold-changes. Some people seem to transform negative log fold-changes (LFC) to "FC" = -2^(-LFC) = -1/(2^LFC) to get a "negative fold-change value", what is really really confusing.

If one insists on using this kind of transformation (so that for inteance +3 indicates a three-fold induction and -3 indicates a repression to 1/3 [note: this is not a three-fold repression!!]) it should not be names "fold-change". Use another name, so that these values will not be confused with fold-changes.

I'd add there is no such thing as "absolute" fold change. Fold change by definition is a relative measure of change, relative to some baseline be it a wild type, vehicle control, or whatever. And any time you change that reference, you change the estimate of differential expression.

Differential gene expression is a method of relative determination of changes in expression of one population, relative to another. LIMMA and eBayes determine relative differential expression. You can talk about absolute measures of expression in a sample or in a population, but the minute you shift to discussing differential gene expression, you must include the context of the relative reference of that determination.

I also agree that you should analyze different platforms independently and then compare results.  The different technologies simply do not lend themselves to combined meta-analysis easily (normalization being just one issue).  You can always compare gene lists from different studies or platforms non-parametrically, or by correlation of response (concordance or discordance of differential expression), or by pathway enrichment overlap or several other approaches.

Actually the problem is that I cannot define the "absolute logFC" because in the paper they did not specify and give any reference either. But, as I have understood, their "absolute logFC" is just a limma calculated logFC. 

Thank you for answers, it helps me a lot!  

Why don't you just ask the authors directly?

I did, but I did not get any answer from them.

Oh, sorry. I did not expect that authors would not respond.

So it would be helpful if you'd give a reference to the paper you are talking about. So we could try to understand the author's intentions and the context in what they used the "absolute logFC".

Most likely it was for the simple purpose of selecting candidates that are differentially expressed by a certain minimal factor. For instance, logFC (treated vs. control) of higher than +1 indicates that the expression in the treated samples is more than 2 times higher than in the controls, whereas a logFC (treated vs. control) of below -1 indicates that the expressin in the control samples is more than 2 times higher than in the treated samples. If one wants to select all genes with this "factor of 2 difference" to either side one would have to select all genes with logFC+1. Alternatively one can simply select all genes with |logFC|>1, wher |x| means "the absolute of x" = always the positive value, independent of the sign of x).

In Genes & Cancer 4(11-12) 486–500, 2013 (see link) you can find such an example:

"Final sets were derived by further requiring that the absolute log2FC DGE be greater than 1 for at least one of the conditions in the comparison." (bold face by me)

http://gan.sagepub.com/content/early/2013/10/31/1947601913506115.full

Once, I met with "absolute log2FC" where to the values was added some constant value to make all values positive. So, as can be seen there is a lot of different explanations for the term "absolute", despite it should not.

Anyway, I am so grateful that you are trying to help me understand what the authors ment. Here is the reference to the paper 

http://www.biomedcentral.com/1471-2164/15/S1/S13

Ok, in this publication it is just what I assumed earlier. The authors write in Methods / Differentially expressed genes:

"Differentially expressed genes (DEG) [...] were detected [...] with absolute log2-fold-change > 2 [...]."

This simply means that the considered genes as differentially regulated only when the signal ratio between the compared groups was larger than 4 (2(+2)) or smaller than 0.25 (2(-2)). This means: "a more than 4-fold difference in either direction".

Got it?

Yes. Probably it was really obvious, but I did not notice it. I appreciate your help and finally understood.

Glad Jochen got that straightened out.  I'm not a mathematician, but it seems to be the authors could have saved a lot of confusion by simply saying

"Differentially expressed genes (DEG) [...] were detected [...] with absolute value of log2-fold-change > 2 [...]." (my edit in bold).

The term "absolute value" is a formal mathematical concept and avoids any confusion of meaning, especially when throwing around the word "absolute" in a study of  "relative" differential expression.

As I said before, it seems that the term "absolute" has different meaning related to the context, despite it should not. It is really confusing when trying to guess what the authors ment, especially when the term is more mathematical than biological.

Anyway, they could, but they did not.

Thanks for involvement!