[BioC] Limma

[Kevin R. Coombes]

On 4/6/2011 11:34 AM, Steve Lianoglou wrote:
> Howdy,
>
> Just some thoughts:
>
> On Wed, Apr 6, 2011 at 11:59 AM, James W. MacDonald
> <jmacdon at med.umich.edu>  wrote:
>> Hi Seraya,
>>
>> On 4/6/2011 11:16 AM, Seraya Maouche wrote:
> [snip]
>> I think the reviewer's point is that you are computing a fold-change based
>> on something that is primarily due to signal divided by something that is
>> primarily due to noise. In other words, the numerator is signal, and the
>> denominator is noise.
>>
>> In addition, the value due primarily to noise is very close to zero, so you
>> end up with a huge fold change that can vary widely, depending on the
>> variability of your noise signal. So the actual value of the ratio is
>> probably not that meaningful, but the fact that it is reliably large
>> probably means that the gene is truly differentially expressed. You just
>> cannot say reliably by how much.
> I don't know ... It's not entirely clear to me what is causing the
> reviewer to be so upset about this.
>
> Seraya, is there some assertion you are making regarding the actual
> *value* of the fold change in these cases (where the denominator is
> "absent"), or are you just saying that these (this) gene is
> differentially express in cond_a vs. cond_b?
>
> Anyway, in order to sidestep this "technicality", maybe one could put
> the fold change into an "appropriate" scale after-the-fact?
>
> For instance, in these situations where the gene is absent in one
> expt, and not in the other, maybe one could set the expression of the
> "absent" probes/genes in a given condition to be the minimum value of
> a detected/expressed gene in that condition. Then calculate the fold
> change post-facto? Where by post-facto I mean to just let limma do
> "it's thing" without any data manipulation. Then, when Seraya is
> listing some fold change value (in a table in the paper(?)), replace
> the fold change of XX/0 (= Inf) with XX/(min gene expression in expt
> the gene is absent in)?
One can plausibly argue that using this minimum is equivalent to using 
the detection limit of the array, and thus the resulting fold-change 
estimate is a lower bound in the case of "absent" genes.  One might even 
be able to make a similar argument using the "noise" denominator that 
the reviewer objected to.

> I mean, the whole thing is a bit strange to me, unless you are indeed
> asserting something based on the value of the fold change -- in which
> case, the value itself in these situations I guess doesn't make much
> sense at all and I reckon I can see the point the reviewer is
> nitpicking over (?).
>
> Just a thought,
> -steve
>