[BioC] limma - Identifying interactions

James W. MacDonald jmacdon at med.umich.edu
Thu May 7 22:01:34 CEST 2009


Hi Suzanne,

Please don't take things off list - the archives are intended to be a 
resource for people to search.

Suzanne Szak wrote:
> 
> Dear Jim,
> 
> Thank you  very much for your response.   I've been spinning my wheels 
> on this ... and by no means am I a statistician, which makes it all the 
> more daunting.
> 
> Actually, I currently only have 3 contrasts represented in my 
> contrast.matrix:
> 
> 1)  "Combo   vs.   Control"  designated as    combo-control  .  Note 
> that in this column, I don't have a contribution from x or   y.
> 2)  "Drug X   vs.    Control"  designated as    x-control
> 2)  "Drug Y    vs.   Control"  designated as    y-control
> 
>  >contrast.matrix
> 
>         combo-control        x-control        y-control        
> x                        0                1                0
> y                        0                0                1
> combo                        1                0                0
> control                -1                -1                -1

You are right - my mistake. Which brings me to another point. Your 
design matrix is what I would call a 'factor effects' parameterization, 
in which everything is compared to an intercept, in this case the 
controls. So the coefficient for e.g., the x treatment is actually x - 
control. So the contrast matrix above is wrong, as you are subtracting 
the control twice from every sample, which as you have noticed is bad. 
You need to substitute a zero in the control row for every contrast.

> 
>  From here, I would typically just run:
> 
>  > fit <- lmFit(chips.norm, design)
>  > fit2 <- contrasts.fit(fit, contrast.matrix)
>  > fit2 <- eBayes(fit2)
> 
> Then I would impose filters on the lods score, fold change, etc. to 
> identify my genes of interest.  Using these matrices, I get ~80% of all 
> genes on the Affymetrix chip as being significant in my "Combo vs 
> Control" scenario .... no way!  That's why I used the phrase  "without 
> luck."
> 
> Anyway, from your response, it sounds like my contrast.matrix should 
> have a column like this:
> 
>  > contrast.matrix.try
>         combo-control
> x                        -1
> y                     -1
> combo               1
> control             1

Close, but not quite. This will give you 
(combo-control)-(x-control)-(y-control)+control = combo-x-y+2control

I think you want a zero in the control row, so you end up with

(combo-control)-(x-control)-(y-control) = combo-x-y+control

I would do it this way, as this will account for the possibility that 
there is a baseline expression of the gene (captured by the control), 
and the x, y, combo treatments just cause the expression to go up or 
down from this point. So combo is really combo+baseline, x is 
x+baseline, etc, so the above translates to

(combo+baseline)-(x+baseline)-(y+baseline)+baseline = combo-x-y

Or you could argue for

x        -.5
y        -.5
combo    1
control   0

which would be about the same, but you are averaging the x, y contribution.

Best,

Jim




> 
> But when I try this, again, about ~60% of the genes on the chip are 
> significant  (e.g. lods>0,  abs(fold change) > 2).
> 
> I know you must be very busy, and I'd really appreciate any time that 
> you might dedicate to helping me.  
> 
> Thanks,
> Suzanne
> 
> 
> 
> 
> *"James W. MacDonald" <jmacdon at med.umich.edu>*
> 
> 07-May-2009 10:22 AM
> 
> Message Size: *5.0 KB*
> 
> 	
> To
> 	Suzanne Szak <suzanne.szak at biogenidec.com>
> cc
> 	bioconductor at stat.math.ethz.ch
> Subject
> 	Re: [BioC] limma - Identifying interactions
> 
> 
> 	
> 
> 
> 
> 
> 
> Hi Suzanne,
> 
> Suzanne Szak wrote:
>  > Hi all,
>  >
>  > I'd like to use limma to identify a possible interaction between two 
> drugs
>  > (called "x" and "y") which would be reflected in gene expression.  That
>  > is, each drug has its own effect, but I think there might be synergy
>  > between the two drugs if cells are treated with both of them ("combo").
>  >
>  > What should the design matrix and contrast matrix look like?  I tried 
> the
>  > matrices below (as well as other variations) without any luck.  And, 
> given
>  > the correct design and contrast matrices, how do I interpret the results
>  > to get the answer I want?  (e.g. I want to find genes in which "combo" >
>  > "x" + "y".)
> 
> What do you mean by 'without any luck'? The fourth contrast below should
> give you what you want (Combo - x - y + control).
> 
> Best,
> 
> Jim
> 
> 
>  >
>  > Thanks much,
>  > Suzanne
>  >
>  >> design
>  >                         x       y       Combo           Control
>  > x.cel                   1       0       0               1
>  > x.cel                   1       0       0               1
>  > y.cel                   0       1       0               1
>  > y.cel                   0       1       0               1
>  > control.cel             0       0       0               1
>  > control.cel             0       0       0               1
>  > combo.cel               1       1       1               1
>  > combo.cel               1       1       1               1
>  >
>  >
>  >> contrast.matrix
>  >
>  >         combo-control   x-control       y-control
>  > x                       0               1               0
>  > y                       0               0               1
>  > combo                   1               0               0
>  > control         -1              -1              -1
>  >
>  >
>  >                  [[alternative HTML version deleted]]
>  >
>  > _______________________________________________
>  > Bioconductor mailing list
>  > Bioconductor at stat.math.ethz.ch
>  > https://stat.ethz.ch/mailman/listinfo/bioconductor
>  > Search the archives: 
> http://news.gmane.org/gmane.science.biology.informatics.conductor
> 
> -- 
> James W. MacDonald, M.S.
> Biostatistician
> Douglas Lab
> University of Michigan
> Department of Human Genetics
> 5912 Buhl
> 1241 E. Catherine St.
> Ann Arbor MI 48109-5618
> 734-615-7826
> 

-- 
James W. MacDonald, M.S.
Biostatistician
Douglas Lab
University of Michigan
Department of Human Genetics
5912 Buhl
1241 E. Catherine St.
Ann Arbor MI 48109-5618
734-615-7826



More information about the Bioconductor mailing list