That is because most scientists will choose a few genes to analyse in detail. They cant analyse all of them you know. And the actual fold change is not very accurate or reproducible using other techniques. 

Thanks Abel. So, that is one answer:

To select a minimum relevant set of genes for a more detailed analysis.

How large could be this minimal set? Approximately.

depends on how much resources lab has, usually its not more than 10.

Thanks, Jochen,

You are talking about what we can understand about diseases' mechanisms by analyzing transcriptomic data, and in what conditions. However, in most diseases, multiple genes are altered (e.g., mutated or up/down regulated). I am working mainly in cancer. So, do you believe that we can focus on just on one gene, to understand or control the disease by treatment?

 Abel,

Please tell me, what you mean by analyzing in detail a couple of genes? It seems that you are talking about some costly investigations. This justifies selecting a small number of genes. Probably, the main questions remain: which small subset? I will come soon with some suggestions. However, I hope that other investigators will tell us their opinion.

Discriminative problems, like cancer vs. normal, or progressive vs. non-progressive disease, are encountered in many fields. Usually, these are treated as classification problems, and machine learning methods are used. I did not see t-test (or similar tests) based approaches. Obviously, they know these techniques but do not consider them the most adequate for a discriminative (classification) problem. So, while a list of differentially expressed genes (DEG) and a selected subset (using an arbitrary cutoff ) could be of interested, is not the most informative one. Also, it has a low translational impact. In my opinion, by far more informative, are predictive genes (PEG). Our bioinformatics analysis pipeline includes in-house machine learning techniques to identify ALL RELEVANT GENES for a given biomedical problem. This is a ranked list of genes (usually less than 100). In most studies, there is a much smaller subset of genes with high relative importance, followed by a long tail of genes with small and decreasing importance (probably indicating redundancy, but this is another story). So, one can select the subset of genes with relatively high importance for further study. I think that this list (there are other lists too, and I will probably talk about them later) is much more informative than the DEG list. Why? It is based on a powerful classification algorithm, and its accuracy gives the credibility of the list. I will reject a list of genes if the accuracy is 65-70%, but I will trust one if the accuracy is 95-99%.