[BioC] Limma: All probes come out as significant

Sean Davis sdavis2 at mail.nih.gov
Thu Apr 12 13:05:16 CEST 2007


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).
>
> > design
>
>  [1] 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0
>
> I run the following to get the adjusted p-values for differential
> expression.
>
> > fit <- lmFit(Seminoma,design)
> > fit <- eBayes(fit)
> > tempo <- topTable(fit,n=nrow(exprs(Seminoma)),adjust="BH")
> > tail(tempo)
>
>              ID    logFC  AveExpr        t     P.Value   adj.P.Val        
> B 36857 236888_at 4.050545 5.896910 2.848388 0.007679017 0.007679872
> -2.489166 39997 240028_at 3.590545 5.153103 2.795185 0.008761283
> 0.008762063 -2.602027 5638  206112_at 4.702957 7.202596 2.749623
> 0.009800000 0.009800654 -2.697717 37129 237160_at 3.356252 4.824457
> 2.721845 0.010488579 0.010489046 -2.755610 9742  210262_at 4.657111
> 7.232671 2.690246 0.011326586 0.011326838 -2.821047 23669 223673_at
> 3.837049 5.837806 2.648655 0.012524609 0.012524609 -2.906478
>
> 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



More information about the Bioconductor mailing list