[BioC] Using limma with contrast matrix ,replicate spots, donor effects

[Peter Wilkinson]

At 12:17 AM 1/22/2005, Gordon Smyth wrote:
>Having within-array replicate spots on your arrays makes no difference at 
>all to the design and contrast matrices. (With one exception, which is 
>that you can't fit a random block effect in limma and estimate a duplicate 
>spot correlation at the same time.) Is there something which has caused 
>you to become concerned about this?

I was originally going to subtract out t_0 from my t_6, t_24, and t_72 as 
my experiments are against a universal, and express all my ratios relative 
to t_0.

Ann Muller pointed out to me that the issue of a randomized block. Now 
since I am not from a strong statistical blood-line (I am more of a 
programmer and biochemist, than a stats person), I now need to go read up 
on randomized block designs because I don't know much about them. So this 
is how the randomized block thing came about, not that I knew anything 
about randomized block designs.

So I guess in my case I do have replicate spots and yes it seems that I 
could apply a randomized block in my case, but as you pointed out limma 
does not support this.

I had not realized how to get the duplicate spot correlations done _at the 
same time_ as calculating the contrasts, I was looking at example 14.1 and 
got confused. I read through all the function descriptions and found that I 
could include from the start the 'ndups' with:

RG$printer <- getLayout(RG$genes, guessdups=TRUE)  which took takes care of 
the dupes for me.

I am ok with this issue now. I have spent more time with the documentation 
in general and I think I am getting a better handle on how limma works. I 
will have to practice with some basic statistical examples to get used to 
interpreting the statistics and knowing what models to apply.


>I suggest you try accommodating the donor effect simply by including a set 
>of coefs for the donor effects in your design matrix. You form the design 
>matrix as you would for an additive two-way anova with donor as one of the 
>two factors. Comparisons between infections, infect types, and infect 
>times will then be in effect made _within_ donor.

I will try this.

Thanks

Peter


>Gordon
>
>>Date: Thu, 20 Jan 2005 10:48:21 -0500
>>From: Pita <pwilkinson_m at xbioinformatics.org>
>>Subject: [BioC] Using limma with contrast matrix ,replicate spots,
>>         donor effects
>>To: bioconductor <bioconductor at stat.math.ethz.ch>
>>
>>This question is because I am misunderstanding how certain things fit
>>together in Limma. There is no example like this in the documentation, and
>>I am trying to figure out how to do this based on examples section 10.5
>>and  14.1.
>>
>>sorry for the lengthy post, this is a complicated one, but it might be an
>>interesting case example for some of you.
>>
>>A simplified version of my experiment follows
>>
>>Background:
>>
>>Blood from 8 separate donors have been collected and undergone a cell sort.
>>The sorted cells that we are interested in were divided and infected with
>>HIV according to the following table (the letters do not mean the literal
>>HIV subtype in this case, I have just simplified it to A,B,C and
>>N=non-infected.).
>>
>>Filename        Cy3     Cy5     Donor
>>1               Ref     N_0     1
>>2               Ref     N_6     1
>>3               Ref     N_24    1
>>4               Ref     N_74    1
>>5               Ref     A_0     1
>>6               Ref     A_6     1
>>7               Ref     A_24    1
>>8               Ref     A_74    1
>>9               Ref     B_0     1
>>10              Ref     B_6     1
>>11              Ref     B_24    1
>>12              Ref     B_74    1
>>13              Ref     C_0     1
>>14              Ref     C_6     1
>>15              Ref     C_24    1
>>16              Ref     C_72    1
>>...for 7 more donors
>>
>>-  I have a series of 2 channel array hybridizations against a common 
>>reference
>>- the array used uses DUPLICATE spots (spacially spotted in pairs).
>>-  N is non-infected(this exp its HIV),
>>-  A,B,C are three different infection types
>>-  0,6,24 are the times that the cells were harvested and RNA isolated.
>>-  A_0 is infected at time 0 which is different from non-infected 0 (N_0)
>>in that A_0 is after 2 hours of incubation with the virus.
>>- Total of 8 donors
>>
>>The question I have is how to deal with the ' donor effect' using Limma.
>>First case (1): I could assume that my donor variability is much less than
>>the variability in the treatments and just plow ahead(probably worth
>>trying).  In the second case (2), the problem being that there can be quite
>>the donor variability so I am thinking that what might be better is if I
>>subtract the 0 time point for each infection type WITHIN each donor from
>>all the others so that all expression values are relative to 0:
>>
>>For
>>example   Donor1  N_72-N_0,  N_24-N_0,  N_6-N_0,        A_72-A_0,
>>A_24-A_0,  A_6-A_0,    etc
>>                 Donor1
>>N_72-N_0,  N_24-N_0,  N_6-N_0,        A_72-A_0,  A_24-A_0,  A_6-A_0,    etc
>>
>>
>>I would like to compare the difference between each donor for the
>>non-infected N to characterize the donor variability so that I understand
>>it and I would like to compare the infection types for each time point in
>>the 2 different ways (cases). My ultimate goal it to compare the infection
>>types at each time point against each other while reducing the noise due to
>>donor variability.
>>
>>There are 2 things i need to know how to do
>>
>>How do I combine creating the contrast matrix and use it with calculating
>>duplicate spot correlation in 14.1,  for case 1?
>>How do I create a contrast matrix to account for normalising against time 0
>>as in case (2) and then combine that with the duplicate spot correlation?
>>
>>
>>lastly, are there in fact other proven methods for dealing with donor
>>variability ?
>>
>>Thanks for any insight.
>>
>>Peter W.
>
>
>