[BioC] of limma and superfluous arrays

[Gordon Smyth]

Dear Yannick,

 From a statistical point of view, you should include in your limma 
analysis any arrays you have which will share the same genewise 
variances as the arrays involved in your contrasts.

How do you know which arrays share the same genewise variances? In 
practice, this means you should include arrays with very comparable 
RNA samples (same tissue, similar treatment), same probe set, 
collected and hybridised at the same time, i.e., arrays which really 
are part of the same greater experiment. Arrays more different than 
that should not be included.

Best wishes
Gordon

>Date: Tue, 29 Jan 2008 22:35:41 +0100
>From: Yannick Wurm <Yannick.Wurm at unil.ch>
>Subject: [BioC] of limma and superfluous arrays
>To: bioconductor at stat.math.ethz.ch
>
>Dear List,
>
>I'm starting to do limma analyses on a small timecourse loop design
>with 2-color cDNA chips as follows:
>         0h vs 6h
>         6h vs 24h
>         24h vs 0h
>Four biological replicates -> and then four biological replicates dye
>balanced <-
>
>My targets file begins like this (only the first two sets of three
>listed):
>         US22502600_F82_S01.gpr  A_0h    A_24h
>         US22502600_F65_S01.gpr  A_24h   A_6h
>         US22502600_F153_S01.gpr A_6h    A_0h
>         US22502600_F85_S01.gpr  F_0h    F_6h
>         US22502600_F60_S01.gpr  F_24h   F_0h
>         US22502600_F72_S01.gpr  F_6h    F_24h
>         ... with eight such sets of three.
>
>But then I also have some chips -> against our labs "standard"
>reference RNA:
>         US22502600_F67_S01.gpr  A_24h   Ref
>         US22502600_F83_S01.gpr  F_24h   Ref
>         ... and six more
>
>For my limma analysis, I have three options:
>         *a*: use only the minimal number of chips (ie each loop of three,
>and nothing to connect the loops). In this case, limma is unable to
>estimate one parameter in each small loop (eg the 6h timepoint). I
>can ask how many genes are differentially expressed between 24h and 0h:
>                 >design.noref = modelMatrix(targets.noref, ref="A_0h")
>                 >fit.noref = lmFit(MA.noref.p, design.noref)
>                 >cont.matrix= makeContrasts(T24_T0 = 
> (A_24h+C_24h+F_24h+K_24h+N_24h
>+Q_24h+R_24h+T_24h -C_0h-F_0h-K_0h-N_0h-Q_0h-R_0h-T_0h)/8,
>levels=design.noref)
>                 >fit.noref2= contrasts.fit(fit.noref, cont.matrix)
>                 >fit.noref2=eBayes(fit.noref2)
>                 >summary(topTable(fit.noref2,n=10000)$adj.P.Val<=0.05)
>
>         ---> I get 3668 differentially expressed spots.
>
>         *b*: provide my "24h" vs Ref chips as well
>                 using ref="Ref" in my design and
>                 > cont.matrix= makeContrasts(T24_T0 = 
> (A_24h+C_24h+F_24h+K_24h+N_24h
>+Q_24h+R_24h+T_24h -A_0h-C_0h-F_0h-K_0h-N_0h-Q_0h-R_0h-T_0h)/8,
>levels=design)
>
>         ---> I get 3796 differentially expressed spots.
>
>
>         *c*: use those in *b*, as well as eight additional chips done in
>parallel, that are XXX vs Ref. The XXX samples don't connect to
>anything other than Ref (they're superfluous).
>
>         ---> I get 3583 differentially expressed spots.
>
>Searching the archives, several posts mentioned that providing more
>chips gives limma a better estimation of variance. Thus it makes
>sense to provide more. And doing so finds more differentially
>expressed genes in *b* than in *a*.
>But so would it be defendable to input all the chips I did in that
>batch to limma? All the chips I've ever done?
>
>And then I get a smaller number of differentially expressed spots in
>*c* than in *b*. Which surprises me, because using more chips should
>make my estimation of variance more precise. Comparing *b* with *c*
>leads me to conclude that the chips I've added to the analysis in *c*
>are funky because they increase estimates of variance, or that the
>chips in *b* show artificially low variance.
>
>Does this make sense?
>Obviously, in this analysis my numbers of differentially expressed
>genes are quite similar in these three cases, and 5% more or less
>significant spots probably won't make a difference. But it would be
>good to know what is most valid for future analyses as well.
>
>
>Thanks and regards,
>
>yannick
>
>
>
>--------------------------------------------
>           yannick . wurm @ unil . ch
>Ant Genomics, Ecology & Evolution @ Lausanne
>    http://www.unil.ch/dee/page28685_fr.html