Limma DEG analysis experiment with partial paired design + batch effect
1
3
Entering edit mode
@englishserver-15152
Last seen 2.6 years ago
Iran

Hi! While re-analyzing GSE65106 from GEO, I have constructed a table containing a subset of samples, as follows:

> a # see below for dput(a)
         accession   cond donor batch pair
A1.r1   GSM1587362    ASD   AA1    1a    1
A1.r2   GSM1587363    ASD   AA1    2a    1
A2.r1   GSM1587364    ASD   AA2    1a    2
A2.r2   GSM1587365    ASD   AA2    2a    2
A3.r1   GSM1587366    ASD   AA3    1a    3
A3.r2   GSM1587367    ASD   AA3    2a    3
SC1.r1  GSM1587368 Normal   AN1    3a    1
SC1.r2  GSM1587369 Normal   AN1    3a    1
SC2.r1  GSM1587370 Normal   AN2    3a    2
SC2.r2  GSM1587371 Normal   AN2    3b    2
SC3.r1  GSM1587372 Normal   AN3    3a    3
SC3.r2  GSM1587373 Normal   AN3    3b    3
ASC1.r1 GSM1587374 Normal   NN1    4a    1
ASC1.r2 GSM1587375 Normal   NN1    4a    1
ASC2.r1 GSM1587376 Normal   NN2    4a    2
ASC2.r2 GSM1587377 Normal   NN3    4a    2

Here, r1/r2 are the replications of experiment on the same subject,

Rows named A1\A2 are autistic (diseased) subjects

Rows named SC1\SC2 are sibling controls; the number denoting to which A(diseased subject) they are paired,

Rows named ASC1\2 are age- and sex-matched controls to ADF.

I want to calculate DEGs between ASD & Normal conditions:

    a=structure(list(accession = structure(1:16, .Names = c("ADF.AA1.1", 
"ADF.AA1.2", "ADF.AA2.1", "ADF.AA2.2", "ADF.AA3.1", "ADF.AA3.2", 
"USCf.AN1.1", "USCf.AN1.2", "USCf.AN2.1", "USCf.AN2.2", "USCf.AN3.1", 
"USCf.AN3.2", "ASCF.NN1.1", "ASCF.NN1.2", "ASCF.NN2.1", "ASCF.NN2.2"
), .Label = c("GSM1587362", "GSM1587363", "GSM1587364", "GSM1587365", 
"GSM1587366", "GSM1587367", "GSM1587368", "GSM1587369", "GSM1587370", 
"GSM1587371", "GSM1587372", "GSM1587373", "GSM1587374", "GSM1587375", 
"GSM1587376", "GSM1587377"), class = "factor"), cond = structure(c(ADF.AA1.1 = 1L, 
ADF.AA1.2 = 1L, ADF.AA2.1 = 1L, ADF.AA2.2 = 1L, ADF.AA3.1 = 1L, 
ADF.AA3.2 = 1L, USCf.AN1.1 = 2L, USCf.AN1.2 = 2L, USCf.AN2.1 = 2L, 
USCf.AN2.2 = 2L, USCf.AN3.1 = 2L, USCf.AN3.2 = 2L, ASCF.NN1.1 = 2L, 
ASCF.NN1.2 = 2L, ASCF.NN2.1 = 2L, ASCF.NN2.2 = 2L), .Label = c("ASD", 
"Normal"), class = "factor"), donor = structure(c(ADF.AA1.1 = 1L, 
ADF.AA1.2 = 1L, ADF.AA2.1 = 2L, ADF.AA2.2 = 2L, ADF.AA3.1 = 3L, 
ADF.AA3.2 = 3L, USCf.AN1.1 = 4L, USCf.AN1.2 = 4L, USCf.AN2.1 = 5L, 
USCf.AN2.2 = 5L, USCf.AN3.1 = 6L, USCf.AN3.2 = 6L, ASCF.NN1.1 = 7L, 
ASCF.NN1.2 = 7L, ASCF.NN2.1 = 8L, ASCF.NN2.2 = 9L), .Label = c("AA1", 
"AA2", "AA3", "AN1", "AN2", "AN3", "NN1", "NN2", "NN3"), class = "factor"), 
    batch = structure(c(ADF.AA1.1 = 1L, ADF.AA1.2 = 2L, ADF.AA2.1 = 1L, 
    ADF.AA2.2 = 2L, ADF.AA3.1 = 1L, ADF.AA3.2 = 2L, USCf.AN1.1 = 3L, 
    USCf.AN1.2 = 3L, USCf.AN2.1 = 3L, USCf.AN2.2 = 4L, USCf.AN3.1 = 3L, 
    USCf.AN3.2 = 4L, ASCF.NN1.1 = 5L, ASCF.NN1.2 = 5L, ASCF.NN2.1 = 5L, 
    ASCF.NN2.2 = 5L), .Label = c("1a", "2a", "3a", "3b", "4a"
    ), class = "factor"), pair = structure(c(1L, 1L, 2L, 2L, 
    3L, 3L, 1L, 1L, 2L, 2L, 3L, 3L, 1L, 1L, 2L, 2L), .Label = c("1", 
    "2", "3"), class = "factor")), class = "data.frame", row.names = c("A1.r1", 
"A1.r2", "A2.r1", "A2.r2", "A3.r1", "A3.r2", "SC1.r1", "SC1.r2", 
"SC2.r1", "SC2.r2", "SC3.r1", "SC3.r2", "ASC1.r1", "ASC1.r2", 
"ASC2.r1", "ASC2.r2"))


eset= matrix (rnorm(80),5)*10# test dataset
    library (limma)
    mm = model.matrix (~0+cond+donor+batch+pair,a)
    mc= makeContrasts ('condASD-condNormal', levels=mm) 
    lf= lmFit (eset, mm )
    cf=contrasts.fit (lf, mc)
    eb=eBayes (cf)
    tt=topTable(eb)
    tt

I wondered: 0) Is it a (biologically) common/accepted practice to treat unaffected siblings and age-and sex-matched controls equally?

1) Whether what I coded correctly answer the problem?

2) What is the "practical" meaning of following warning: (I am very uncomfortable with "pair2 pair3" being part of the message! )

Coefficients not estimable: donorNN3 batch3b batch4a pair2 pair3

Warning message:

Partial NA coefficients for 5 probe(s)

3) Is it passable to neglect certain factors - here, pair for example- and proceed with the "less accurate" analysis?

limma paired batch • 1.9k views
ADD COMMENT
1
Entering edit mode

For the benefit of other readers, this is what OP's data.frame looks like:

> a
         accession   cond donor batch pair
A1.r1   GSM1587362    ASD   AA1    1a    1
A1.r2   GSM1587363    ASD   AA1    2a    1
A2.r1   GSM1587364    ASD   AA2    1a    2
A2.r2   GSM1587365    ASD   AA2    2a    2
A3.r1   GSM1587366    ASD   AA3    1a    3
A3.r2   GSM1587367    ASD   AA3    2a    3
SC1.r1  GSM1587368 Normal   AN1    3a    1
SC1.r2  GSM1587369 Normal   AN1    3a    1
SC2.r1  GSM1587370 Normal   AN2    3a    2
SC2.r2  GSM1587371 Normal   AN2    3b    2
SC3.r1  GSM1587372 Normal   AN3    3a    3
SC3.r2  GSM1587373 Normal   AN3    3b    3
ASC1.r1 GSM1587374 Normal   NN1    4a    1
ASC1.r2 GSM1587375 Normal   NN1    4a    1
ASC2.r1 GSM1587376 Normal   NN2    4a    2
ASC2.r2 GSM1587377 Normal   NN3    4a    2
ADD REPLY
0
Entering edit mode

Thank you. I updated the original question according to your post. PS. I also edited the title.

ADD REPLY
2
Entering edit mode
@gordon-smyth
Last seen 3 hours ago
WEHI, Melbourne, Australia

You are correct to be concerned about the results. The model you have fitted is heavily over-parametrized and will not give meaningful results.

It is not meaningful to put factors into a linear model that are confounded with each other. Here cond is completely confounded with both donor and batch. On top of that, pair is founded with donor and donor is almost completely confounded with batch.

In your case it is not just "passable" to remove factors but compulsory. The issue is not accuracy but rather the need to estimate quantities that have a meaning.

One simple way to analyse the data would be to use design <- model.matrix(~cond) in conjunction with duplicateCorrelation to estimate the within-donor correlation. That doesn't model all the complexity of the data, in particular it doesn't allow for which normals are siblings of which affected, but at least it gives an overall comparison between ASD and Normal and it does allow for the fact that the replicate samples from the same donor are correlated.

Note that you can't adjust for batch because it is completely confounded.

You ask whether it is biologically acceptable to treat age-sex-matched controls the same as sibling controls. I would think not, but it is up to you. The matched controls couldn't be expected to be as similar to the corresponding affected as the siblings are, but there might be some pairing. You are the biologist and this is your analysis, so it is your job to make biological judgements rather than mine.

ADD COMMENT
0
Entering edit mode

Thank you @Gordon Smyth for your analytic response. In geneal, how should one check whether factors are confounded? On googling I ended up with chisq.test(). However,

chisq.test(a$batch, a$donor) #p-value = 0.2867

chisq.test(a$pair , a$donor) #p-value = 0.01

chisq.test(a$cond , a$batch) #p-value = 0.003019

chisq.test(a$cond,  a$donor) #p-value = 0.04238

Sorry if the question seems out of place or naive -I have a biology background- I fail to see why 'donor is almost completely confounded with batch' I wondered whether it is a typo or I am missing something, especially that factors have more than two levels.

ADD REPLY
0
Entering edit mode

limma has done the checking for you. When it says that certain coefficients are not estimable, it means that these coefficients are confounded with other factors in the model.

Confounding is not an esoteric mathematical thing, but a scientific principle that you can see yourself just by looking at the factors. Considering cond and donor. There are three ASD donors and 6 Normal donors. There is no overlap between the ASD donors and the Normal donors -- they are different people. In other words, knowing which donor a sample comes from already tells you what the condition is. That is what I mean by cond being completely confounded with donor.

When you include donor in the model you are allowing all the donors to be different, i.e., for gene expression to depend on the donor, and you are trying to estimate an expression level for each donor. When you include cond in the model you are allowing the ASD samples to be different from the Normal samples and you are trying to estimate an expression level for each condition. But there is an ambiguity here. Comparing ASD to Normal is the same as comparing AA1, AA2 and AA3 to AN1, AN2, AN3, NN1, NN2 and NN3. If you put cond and donor into the model at the same time, then you are trying to estimate the same difference twice, which just can't be done.

From a scientific point of view, you are trying to test for differences between ASD and Normal. You are not trying to test for differences between these donors specifically. Hence cond is a factor to be included in the model but donor is an experimental unit that represents biological replication. The donors represent a sample of possible subjects rather than an experimental condition and therefore should not be in the model.

ADD REPLY
0
Entering edit mode

Great response, and thank you again for your detailed response. I think I got the main ideas.

ADD REPLY

Login before adding your answer.

Traffic: 341 users visited in the last hour
Help About
FAQ
Access RSS
API
Stats

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6