Question

limma contrast matrix interaction effect

0

Entering edit mode

steven wink ▴ 90

@steven-wink-5440

Last seen 5.5 years ago

Dear list, I was hoping for some input since I have difficulty figuring out if I obtain a valid DEG for my biological question. By following the limmaUsersGuide manual I came up with a modification that might be suitable for my analysis. I have 4 treatments: control, A, B and AB so AB is a combined treatment of A and B After defining my design matrix I set up contrasts: SynergisticEffect = (AB - control ) - ((A - control) + (B - control)) Would this contrast give me a DEG where I want to test if the combined treatment has more effect than the sum of individual treatments? Would I first have to test for main effects before stating significant interaction effects? Does this in effect mean I should remove genes from DEG obtained from interaction test that are also in DEG of main effects (ie A-control or B-control and also AB-control)? Any help on making this more clear for me is greatly appreciated! Best regards, Steven [[alternative HTML version deleted]]

• 2.3k views

ADD COMMENT • link updated 10.7 years ago by Gordon Smyth 51k • written 10.7 years ago by steven wink ▴ 90

score 1 · Answer 1 · 2014-03-05

Hello, The correct contrast is AB - A - B + Control. After you get results, you can use interaction.plot and plot the expression of the top few genes, to see that there is an interaction. Genes which are affected by the interaction of two treatments will have lines that are not parallel. -------------------------------------- Dario Strbenac PhD Student University of Sydney Camperdown NSW 2050 Australia

score 1 · Answer 2 · 2014-03-07

1

Entering edit mode

Gordon Smyth 51k

@gordon-smyth

Last seen 56 minutes ago

WEHI, Melbourne, Australia

Dear Steven, > Date: Wed, 5 Mar 2014 13:42:34 +0100 > From: steven wink <hardervidertsie at="" gmail.com=""> > To: <bioconductor at="" stat.math.ethz.ch=""> > Subject: [BioC] limma contrast matrix interaction effect > > Dear list, > > I was hoping for some input since I have difficulty figuring out if I > obtain a valid DEG for my biological question. > > By following the limmaUsersGuide manual I came up with a modification that > might be suitable for my analysis. > > I have 4 treatments: control, A, B and AB > so AB is a combined treatment of A and B > After defining my design matrix I set up contrasts: > > SynergisticEffect = (AB - control ) - ((A - control) + (B - control)) > > Would this contrast give me a DEG where I want to test if the combined > treatment has more effect than the sum of individual treatments? It would be correct if you changed the "+" to "-". > Would I first have to test for main effects before stating significant > interaction effects? Does this in effect mean I should remove genes from > DEG obtained from interaction test that are also in DEG of main effects > (ie A-control or B-control and also AB-control)? No, there is no need whatsoever to do this. Best wishes Gordon > Any help on making this more clear for me is greatly appreciated! > > Best regards, > Steven ______________________________________________________________________ The information in this email is confidential and intend...{{dropped:4}}

ADD COMMENT • link 10.7 years ago Gordon Smyth 51k

0

Entering edit mode

Did you notice the double brackets? SynergisticEffect = (AB - control ) -* (*(A - control) + (B - control)*)* If I change the "+" to a "-" I get AB + C - A + B The model becomes aver.gene.expr=â1*X_AB + â2*X_C - â3*X_A + â4*X_B + error, with â the average gene expression of each group and X_xx the respective rows of the contrast matrix. I don't see any suitable way to test any hypothesis here I am interested in. I would like to find significant genes where abs(log2_FC(AB)) >> abs( log2_FC(A) + log2_FC(B) ), where FC is fold change as compared to control (C) To simplify I should write: AB + C - A - B instead of (AB - control ) - * (*(A - control) + (B - control)*)* if H0: â1 + â2 = â3 + â4 is rejected I have my genes (I believe) 2014-03-07 0:20 GMT+01:00 Gordon K Smyth <smyth@wehi.edu.au>: > Dear Steven, > > Date: Wed, 5 Mar 2014 13:42:34 +0100 >> From: steven wink <hardervidertsie@gmail.com> >> To: <bioconductor@stat.math.ethz.ch> >> Subject: [BioC] limma contrast matrix interaction effect >> >> Dear list, >> >> I was hoping for some input since I have difficulty figuring out if I >> obtain a valid DEG for my biological question. >> >> By following the limmaUsersGuide manual I came up with a modification that >> might be suitable for my analysis. >> >> I have 4 treatments: control, A, B and AB >> so AB is a combined treatment of A and B >> After defining my design matrix I set up contrasts: >> >> SynergisticEffect = (AB - control ) - ((A - control) + (B - control)) >> >> Would this contrast give me a DEG where I want to test if the combined >> treatment has more effect than the sum of individual treatments? >> > > It would be correct if you changed the "+" to "-". > > Would I first have to test for main effects before stating significant >> interaction effects? Does this in effect mean I should remove genes from >> DEG obtained from interaction test that are also in DEG of main effects (ie >> A-control or B-control and also AB-control)? >> > > No, there is no need whatsoever to do this. > > Best wishes > Gordon > > Any help on making this more clear for me is greatly appreciated! >> >> Best regards, >> Steven >> > > ______________________________________________________________________ > The information in this email is confidential and inte...{{dropped:10}}

ADD REPLY • link 10.7 years ago steven wink ▴ 90

0

Entering edit mode

On Fri, 7 Mar 2014, steven wink wrote: > Did you notice the double brackets? SynergisticEffect = (AB - control ) -* > (*(A - control) + (B - control)*)* Yes, I missed the double brackets. Your contrast was fine. > To simplify I should write: AB + C - A - B instead of (AB - control ) Yes, as Dario has already suggested. It makes no difference. Any algebraic equation that is equivalent will give the same result in limma. > I would like to find significant genes where abs(log2_FC(AB)) >> abs( > log2_FC(A) + log2_FC(B) ), where FC is fold change as compared to > control (C) I don't think that this is really what you mean. Surely you would want to find genes for which the AB effect was in a different direction to the A and B effects, even it was not greater in absolute value. The contrast does this. If what you state is really is what you mean, then this is not something you can formally test in limma (or an any other statistics program afaik). Gordon ______________________________________________________________________ The information in this email is confidential and intend...{{dropped:4}}

ADD REPLY • link 10.7 years ago Gordon Smyth 51k

0

Entering edit mode

Thank you for your help Gordon and Dario, And I agree with you on the last statement, did not realize that yet. Best regards Steven WInk 2014-03-08 7:30 GMT+01:00 Gordon K Smyth <smyth@wehi.edu.au>: > On Fri, 7 Mar 2014, steven wink wrote: > > Did you notice the double brackets? SynergisticEffect = (AB - control ) >> -* >> (*(A - control) + (B - control)*)* >> > > Yes, I missed the double brackets. Your contrast was fine. > > > To simplify I should write: AB + C - A - B instead of (AB - control ) >> > > Yes, as Dario has already suggested. It makes no difference. Any > algebraic equation that is equivalent will give the same result in limma. > > > I would like to find significant genes where abs(log2_FC(AB)) >> abs( >> log2_FC(A) + log2_FC(B) ), where FC is fold change as compared to control >> (C) >> > > I don't think that this is really what you mean. Surely you would want to > find genes for which the AB effect was in a different direction to the A > and B effects, even it was not greater in absolute value. The contrast > does this. If what you state is really is what you mean, then this is not > something you can formally test in limma (or an any other statistics > program afaik). > > Gordon > > > ______________________________________________________________________ > The information in this email is confidential and inte...{{dropped:10}}

ADD REPLY • link 10.7 years ago steven wink ▴ 90