[BioC] Background correction -- was: Illumina data problem

[Wolfgang Huber]

Dear Allen,

are you saying that your estimated fold changes are compressed towards 1
because you did not do background correction?
There is a reason why background correction is done: to increase
sensitivity of the analysis. Simply to omit background correction
because you don't like negative net intensities misses the point, and it
causes exactly what you describe, namely loss of sensitivity.

There are two approaches to this:
(i) use a biased background correction method that does not produce
negative values, then you can log-transform
(ii) use an unbiased background correction method that can produce
negative values, then do a glog-transform
The net effect of these approaches is often quite similar in practice.

I would like to refer to the documentation of, or to the authors of, the
beadarray and lumi packages regarding the best set of functions to
perform these (e.g. "vst" is along the lines of (ii)).

Best wishes
 Wolfgang

PS Allen, Please do consider using more specific (=useful) subject lines.

------------------------------------------------------------------
Wolfgang Huber  EBI/EMBL  Cambridge UK  http://www.ebi.ac.uk/huber



affy snp wrote:
> Hi Sean,
> 
> Thanks a lot for pointing out this.
> 
> We knocked down a gene of our interest so I would at least expect
> that gene would be shown of down-regulated expression in the test
> sample, that was what I mean "it should not be the case which no
> genes were significantly up- or down- regulated".
> 
> We only have one normal and one test sample in our experiment
> so I just tried testing the fold change for test samples vs. normal
> sample at the beginning to get a flavor of how many genes would
> be affected by the knock-down experiment at the expression level.
> Of this, the median fold change of expression level of around 46713
> probes is around 1.
> 
> That was done for the data which was not processed as the microarray
> facility did previously. What they did before is to do the background
> subtraction etc and returned us the data with some negative values.
> For that set of data, after normalization, they were some genes
> which were significantly expressed and fit our experimental design.
> 
> So I am wondering, based on the raw data, did I do sth wrong so that
> it resulted in the conclusion which I got?
> 
> Best,
>     Allen
> 
> On Jan 2, 2008 1:10 PM, Sean Davis <sdavis2 at mail.nih.gov> wrote:
> 
>>
>> On Jan 2, 2008 12:31 PM, affy snp <affysnp at gmail.com> wrote:
>>
>>> Hello, dear list:
>>>
>>> I recently had questions about Illumina data two months ago because of
>>> a large portion of negative values found. Suggested by some expertises
>>> here, I requested our microarray facility to re-process the data by not
>>> doing
>>> any global or background subtraction etc. This way, in deed, there is no
>>> negative values existing. But very surprisingly, by applying two
>>> packages,
>>> Lumi and Beadarray respectively, after normalization, the fold change
>>> between the test and normal samples is around 1. Apparently, it should
>>> not be this case. Can anybody here offer some possible reasons for this?
>>>
>> Hi, Allen.
>>
>> Do you mean that there are no differentially expressed probes?  What test
>> did you do (limma, t-test, etc.)?  And what information suggests that "it
>> should not be the case"?
>>
>> If there are not differentially expressed probes, there could be many
>> reasons, both biological and technical for this being the case.  You will
>> probably need to dig into the data to check the quality closely.  Also, make
>> sure that sample labels were not switched, etc.  In short, when you get a
>> result that you don't expect, you will need to do some troubleshooting and
>> if you come up empty (no technical explanation), then you (unfortunately)
>> need to explain to your lab collaborators (or yourself) that the experiment
>> did not show differences.
>>
>> Hope that helps,
>> Sean
>>
>>