[BioC] Limma: All probes come out as significant

[Sean Davis]

On Thursday 12 April 2007 09:09, Daniel Brewer wrote:
> Sean Davis wrote:
> > On Thursday 12 April 2007 06:24, Daniel Brewer wrote:
> >> Hello,
> >>
> >> I have a curious problem involving Limma.  I have an ExpressionSet
> >> object (called Seminoma) that contains the results of 18 samples (12
> >> tumours and 6 normals).  The only strange thing I have done is to join
> >> Affymetrix U133A and B results (renaming the probes so that there is no
> >> overlap).
> >> ...
> >> As you can see, all the probes appear to be significantly differentially
> >> expressed.  I am sure this should not be the case, especially after
> >> examining a number of different probes.  For example, on probe "117_at"
> >> if I run a t-test() it produces a p-value of 0.1034 (no adjustment)
> >> whereas limma suggests it is 8.4e-07 (or 2.45e-06).
> >>
> >> My only thought is that something must be happening in the eBayes step.
> >>  Can anyone help me on what I might be doing wrong?
> >
> > Your design vector will test that the probes are significantly different
> > from 0.  You probably want to include an intercept term or (I find it
> > more natural for two groups) define the two groups explicitly and then
> > use a contrast matrix to get the difference between the two.  In either
> > case, your design matrix needs at least two columns.
> >
> > Sean
>
> Thanks for that.  You are completely correct and it seems I have made a
> fundamental error with limma when working with 1-channel arrays (I
> learnt it using 2-channel).  So was  my design matrix basically
> detecting probes that are different from zero in only the tumour
> samples? 

Yes, I believe so.  

Sean