[BioC] Limma-Contrasts-Question

Jenny Drnevich drnevich at illinois.edu
Thu Mar 18 14:38:14 CET 2010

Hi Biju,

I suggest you consult a statistics textbook or a 
local statistician to get some more information 
on 2x2 factorial designs, which is what your 
experiment has. In particular, contrast 3 is 
known as the interaction term, whereas (A1+B1)/2 
- (A2+B2)/2 is the main effect of condition 1 vs. 
condition 2 AVERAGED over the levels (A & B) of 
the second factor.  The reason why you want to 
use the interaction term instead of the main 
effect term is because the values for A2, B1 and 
B2 could be very similar, but if A1 is much 
larger, the average of A1 and B1 could be enough 
to give you a significant main effect term, even 
though the response in not really the same. In a 
typical 2x2 ANOVA, if the interaction term is 
significant, then you ignore the results for the 
main effects because they may be misleading...


At 07:37 AM 3/18/2010, Biju Joseph wrote:
>Thanks Jenny, Your suggestion answers my question.
>I make the following Contrasts as you suggest
>1. A1-A2
>2. B1-B2
>3. (A1-A2)-(B1-B2)
>And then the genes that are significant in contrasts 1 and 2 but not in
>contrast 3 would be considered as those that have a significant response in
>A and B. This is perfect and answers my question as well.
>But what is still not clear to me is what happens when I use the contrast
>(A1+B1)/2 - (A2+B2)/2.
>Why do I get a much larger list of genes called as significantly DE when I
>use this contrast compared to the method "significant in contrast 1 and 2
>but not in contrast 3".
>Institut für Hygiene und Mikrobiologie
>Universität Würzburg
>Josef-Schneider-Str. 2, Gebäude E1
>97080 Würzburg
>Email: bjoseph at hygiene.uni-wuerzburg.de
>Tel.:    0931 201 46708
>Fax:    0931 201 46445
>-----Ursprüngliche Nachricht-----
>Von: Jenny Drnevich [mailto:drnevich at illinois.edu]
>Gesendet: Dienstag, 16. März 2010 15:04
>An: Biju Joseph; Gordon K Smyth
>Cc: Bioconductor mailing list
>Betreff: Re: [BioC] Limma-Contrasts-Question
>Hi Biju,
>One thing to consider is that the question you're asking - "which
>genes have the SAME response in A and B?" is not what a statistical
>test is designed to measure. Instead, the null hypothesis is that the
>responses are the same, and only if there is enough evidence of a
>different between the responses will the statistical test become
>significant. One possibility would be to do the 3 following contrasts:
>1) A1-A2
>2) B1-B2
>3) (A1-A2) - (B1-B2)
>The third one tests whether the response in A is the same as the
>response in B. You could do a Venn Diagram on these three contrasts,
>and a those genes that are significant in 1) and 2) but not
>significant in 3) could be considered genes that have the a
>significant response in A and the "same" (i.e., not significantly
>different) significant response in B.
>Note that genes could be significant for 3), even if they change in
>the same direction, if they change by differing amounts (2-fold up
>versus 20-fold up). Whether you want to call this the "same" response
>depends on your research questions...
>At 02:17 AM 3/15/2010, Biju Joseph wrote:
> >Thanks Gordon for your answer.
> >
> >In my question, I was referring to the DE expressed genes in the same
> >direction.
> >
> >What I am actually unclear about is the following.
> >
> >Lets say, I generate topTables for
> >1. A1-A2
> >2. B1-B2
> >
> >Now using these 2 individual tables, I could pull out the genes common
> >in either direction using lets say Access.
> >
> >Is this method valid and safe?
> >How comparable should this manually generated common response between
> >strains A and B in conditions 1 and 2 be comparable to the topTable
> >generated using (A1+B1)/2 - (A2+B2)/2. Of course I think that these 2
> >methods should generate more or less the same results at least with
> >respect to numbers of DE genes.
> >
> >What FC is better to report for the common response? One from using
> >the (A1+B1)/2 - (A2+B2)/2 contrast or the FC from the individual
> >topTables.
> >
> >Best
> >Biju
> >
> >
> >Quoting Gordon K Smyth <smyth at wehi.EDU.AU>:
> >
> >>Dear Biju,
> >>
> >>Using the contrast (A1+B1)/2 - (A2+B2)/2 will find genes which
> >>response in the same direction, and perhaps by about the same fold
> >>change, in both A and B.
> >>
> >>Doing separate tests for A1-A2 and B1-B2, does not require genes to
> >>be changing in the same direction in A and B.
> >>
> >>Best wishes
> >>Gordon
> >>
> >>>Date: Thu, 11 Mar 2010 14:49:35 +0100
> >>>From: "Biju Joseph" <bjoseph at hygiene.uni-wuerzburg.de>
> >>>To: <bioconductor at stat.math.ethz.ch>
> >>>Subject: [BioC] Limma-Contrasts-Question
> >>>Message-ID: <000001cac121$b0cbd3b0$12637b10$@uni-wuerzburg.de>
> >>>Content-Type: text/plain;       charset="iso-8859-1"
> >>>
> >>>Dear all
> >>>
> >>>Sorry if this is a trivial question,
> >>>My experiment design is the following
> >>>2 strains (A & B) subjected to 4 conditions each (1,2,3,4) compared using
> >>>common reference design.
> >>>We are interested in various contrasts between the 8 samples.
> >>>Using limma - I was able to generate topTables for the required contrasts
> >>>eg:
> >>>A1-B1, A2-B2, A3-B3, A4-B4
> >>>A1-A2, A2-A3, B1-B2, B2-B3 and so on
> >>>
> >>>A comparison for example the topTable A1-A2 and B1-B2, would represent
> >>>common response in A and B from condition 1 and condition 2.
> >>>
> >>>Using this manual comparison of the 2 topTables, I saw that around 400
> >>>are commonly differentially regulated in strain A and strain B in the
> >>>conditions 1 and 2.
> >>>
> >>>Now when I include the following contrast in my model in limma
> >>>
> >>>(A1+B1)/2 - (A2+B2)/2 which in my understanding also generates the common
> >>>response between condition 1 and 2 in the 2 strains A and B.
> >>>
> >>>The topTable generated using this contrast shows only 10 genes to be
> >>>commonly differentially regulated between condition 1 and 2.
> >>>
> >>>Would be great if someone could explain this discrepancy to me and about
> >>>which method is safer to compare(comparison of the 2 individual toptables
> >>>the toptable generated using make.contrasts).
> >>>
> >>>Best
> >>>Biju
> >>______________________________________________________________________
> >>The information in this email is confidential and inte...{{dropped:10}}
> >
> >_______________________________________________
> >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
>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
>ph: 217-244-7355
>fax: 217-265-5066
>e-mail: drnevich at illinois.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

ph: 217-244-7355
fax: 217-265-5066
e-mail: drnevich at illinois.edu 

More information about the Bioconductor mailing list