[BioC] stability of gene lists with gcrma and limma upon changing the number of samples in the normalization

[Sean Davis]

On Wed, Apr 2, 2008 at 6:51 PM, Richard Friedman
<friedman at cancercenter.columbia.edu> wrote:
> Dear Bioconductor users.
>
>         I am using gcrma and limma on a dataset of 3 conditions with 2
>  replicates
>  each.
>
>  1. First I normalized 2 conditons together (4 arrays) and got 312
>  probesets with B>0
>  using limma comparisons. I used no filtering.
>
>  2. Then I normalized all 3 conditions together (6 arrays) and
>  compared the
>  same 2 conditions that I did before. I got 301 probesets with B>0. I
>  used no filtering.
>
>  3. Comparing the probesets I got before I got in runs 1 and 2 I got
>  91 in common.
>
>  Does this sound right?
>  Shouldn't the list be more stable than that?

Hi, Richard.  I don't think anyone can comment fully on this without
access to the data.  However, it is not unusual for gene lists to
change significantly when you add 1/3 more data, particularly when you
have a small number of samples to begin with.  Both gcrma and limma
will be affected by the additional data, even though you are
apparently making the same comparison in both cases 1 and 2.

Hope that helps.

Sean