[BioC] limma: get all sig genes from multiple contrasts

[Ivan Baxter]

Thanks Jim- that will help  I think I am still going to get useful 
information out of the clustering. I have 7 different combinations of 
treatments, so the number of different patterns is going to be quite 
large. Clustering seems to be  a good way to get a feel for which 
patterns are there and  of interest. For a package like goCluster, 
wouldn't I want to reduce the number of genes which are in the set that 
is analyzed (say from 22k to ~1k)? While there might be genes that fit a 
certain pattern that isn't significantly different in any of my 
contrasts, it seems likely that it would be in the minority. After I 
have identified interesting patterns from the significantly changed 
genes, I could then go back in and see if other genes match that pattern?

thanks

Ivan


James W. MacDonald wrote:
> Hi Ivan,
>
> Ivan Baxter wrote:
>> I realize this may be a silly question, but I have gone through all 
>> the case studies in the limma users manual and I can't seem to find 
>> the answer to this anywhere.   I have a 3x3 factorial experiment and 
>> I followed the case studies to make my linear model and designate 13 
>> contrasts of interest.  My question is: is there a simple way to get 
>> all the genes that show significant differences in any one of my 
>> contrasts? (for use in clustering, for ex.)
>>
>
> You can do this easily enough without needing any added functionality 
> in limma.
>
> results <- decideTests(fit2)
> index <- apply(results, 1, any)
> sigchange <- eset[index,]
>
> As an aside, unless you are planning to cluster your data to show 
> patterns you have extracted using the different contrasts, this is 
> probably not what you want to do. You are extracting only those genes 
> that fulfill a certain set of criteria, so any resulting clustering 
> solution will by definition show a pattern that reflects that.
>
> As an example, if you do a t-test comparing two sample types and then 
> cluster the significant genes, you will get a heatmap showing that the 
> two samples are quite different from each other, with very little 
> variation within each sample type (which is what the t-test is testing 
> for).
>
> HTH,
>
> Jim
>
>
>