[BioC] nested design in limma?

[Jenny Drnevich]

Hi Gordon,

Thanks for your response. I started checking the correlations at each 
level: spot correlation is 0.81 and dye-swap pairs is weaker, -0.20, but 
perhaps not so weak as to be ignorable. The big problem occurred when 
trying to estimate correlations within dams as a block effect, because the 
arrays are direct comparisons, and of the three offspring from dam C1, one 
is compared to an offspring from dam T1, one to an offspring from dam T2 
and one to an offspring from dam T3 - so there are no good blocking groups! 
Going to a separate channel analysis requires yet another level of 
correlation - intra-spot, so that's probably not an option either.

>Just as an aside, I am continually amazed at how common technical 
>dye-swaps are. As far as I can see, they just complicate the analysis to 
>no advantage, yet they have captured the imagination of many biologists. 
>My guess is that this an attempt to balance the dyes, although this can be 
>better achieved without introducing technical replication.

The sad thing (about me) is that I advised the researchers on the 
experimental design! I definitely agree now that technical dye-swaps are 
probably a waste of arrays. This was my first time handling spotted data, 
and I didn't appreciate all the intricacies that are involved; I had seen 
that limma had methods to handle duplicate spots and dye-swap technical 
reps, but I didn't realize that they could not be used simultaneously until 
starting to work with duplicateCorrelation and the ndups & block options 
within lmFit. I don't think this warning was in the vignette anywhere - 
perhaps a short sentence could be added to the technical replication section?

Cheers,
Jenny



>Cheers
>Gordon
>
>At 04:27 AM 22/02/2006, Jenny Drnevich wrote:
>>Hi Gordon,
>>
>>I didn't know a nested design would be handled the same as duplicate 
>>spots, since duplicate spots are technical replicates but multiple 
>>offspring are independent replicates. I guess when I have some free time 
>>I'll look into the math of how the block and correlation are used in 
>>lmFit... Unfortunately, this solution doesn't help me in this case 
>>because there are also both duplicate spots and technical replicates of 
>>arrays! If duplicateCorrelation can only be used once, I was going to 
>>average the duplicate spots, use duplicateCorrelation for the dye-swapped 
>>tech reps, fit a coefficient for each dam, and then extract the 
>>difference between sets of dams as a contrast. I know this will treat dam 
>>as a fixed effect, rather than as a random effect, but I'm not sure if 
>>there's a better way to do it.
>>
>>Cheers,
>>Jenny
>>
>>At 01:37 AM 2/21/2006, Gordon K Smyth wrote:
>>>Hi Jenny,
>>>
>>>This design is qualitatively the same as the "duplicate spot" situation, 
>>>where the treatment is
>>>applied at the array level but the measurements are made on multiple 
>>>spots per array.  In your
>>>case, treatments are applied to dams but measurements are made on 
>>>multiple offspring.
>>>
>>>Hence you can use the duplicateCorrelation() function in limma with dam 
>>>as the block.
>>>
>>>Best wishes
>>>Gordon
>>>
>>>On Tue, February 21, 2006 6:03 am, Jenny Drnevich wrote:
>>> > Hello,
>>> >
>>> > I was wondering if there was any (easy) way to handle a nested design in
>>> > limma. I looked in the Bioconductor archives, but the only references to
>>> > nested designs weren't really nested - one was just a factorial 
>>> design, and
>>> > the other was a repeated measurement design, which could be done in limma
>>> > as a blocking variable. In this experiment design, the treatments 
>>> (infected
>>> > and control) were made on the dams, but the effects were measured on
>>> > multiple offspring per dam; hence dam is nested within treatment. In SAS
>>> > terminology (forgive me...), the model would look like this:
>>> > log2_expression = treatment + dam(treatment) , with dam as a random
>>> > variable. The test statistic for treatment should now be formed using the
>>> > variance due to dam(treatment) and not the error variance. Can limma be
>>> > made to handle this sort of design?
>>> >
>>> > Thanks,
>>> > Jenny
>>> >
>>> > Jenny Drnevich, Ph.D.
>>> >
>>> > Functional Genomics Bioinformatics Specialist
>>> > W.M. Keck Center for Comparative and Functional Genomics
>>> > Roy J. Carver Biotechnology Center
>>> > University of Illinois, Urbana-Champaign
>>> >
>>> > 330 ERML
>>> > 1201 W. Gregory Dr.
>>> > Urbana, IL 61801
>>> > USA
>>> >
>>> > ph: 217-244-7355
>>> > fax: 217-265-5066
>>> > e-mail: drnevich at uiuc.edu
>>
>>Jenny Drnevich, Ph.D.
>>
>>Functional Genomics Bioinformatics Specialist
>>W.M. Keck Center for Comparative and Functional Genomics
>>Roy J. Carver Biotechnology Center
>>University of Illinois, Urbana-Champaign
>>
>>330 ERML
>>1201 W. Gregory Dr.
>>Urbana, IL 61801
>>USA
>>
>>ph: 217-244-7355
>>fax: 217-265-5066
>>e-mail: drnevich at uiuc.edu