deseq2 analysis, accounting for pen effect
1
0
Entering edit mode
bjarke ▴ 10
@bjarke-9149
Last seen 2.4 years ago
Denmark

I am going to analyse micro biome data with the design listed below.

It is an animal trial with pens (SampleGroup) where all the animals in a given pen receive the same treatment. I want to obtain the effect of treatment but need to block out the effect of pen (SampleGroup). How can I achieve this with deseq2 (or other packages) ? So far I have addressed the problem by using pens as the experimental unit. I.e. averaging over all animals in a pen with the use of phyloseq. But are there other and better ways to handle the situation ?

Best regards,

Bjarke

 

 

 

 SampleNo Treatment SampleGroup
1    FT0024        T1         G41
2    FT0027        T1         G41
3    FT0026        T1         G41
4    FT0025        T1         G41
5    FT0028        T1         G41
6    FT0033        T1         G44
7    FT0032        T1         G44
8    FT0031        T1         G44
9    FT0029        T1         G44
10   FT0035        T4         G46
11   FT0001        T4         G46
12   FT0037        T4         G46
13   FT0002        T4         G46
14   FT0038        T4         G46
15   FT0034        T4         G46
16   FT0036        T4         G46
17   FT0004        T4         G49
18   FT0006        T4         G49
19   FT0039        T4         G49
20   FT0042        T4         G49
21   FT0040        T4         G49
22   FT0005        T4         G49
23   FT0041        T4         G49
24   FT0003        T4         G49
25   FT0008        T4         G62
26   FT0010        T4         G62
27   FT0044        T4         G62
28   FT0009        T4         G62
29   FT0007        T4         G62
30   FT0046        T4         G62
31   FT0047        T4         G62
32   FT0045        T4         G62
33   FT0043        T4         G62
34   FT0011        T4         G62
35   FT0048        T1         G63
36   FT0013        T1         G63
37   FT0014        T1         G63
38   FT0012        T1         G63
39   FT0015        T1         G67
40   FT0017        T1         G67
41   FT0018        T1         G67
42   FT0016        T1         G67
43   FT0019        T1         G67
44   FT0023        T4         G68
45   FT0021        T4         G68
46   FT0020        T4         G68
47   FT0030        T4         G68
48   FT0022        T4         G68

deseq2 microbiome • 1.5k views
ADD COMMENT
0
Entering edit mode
@steve-lianoglou-2771
Last seen 12 months ago
United States

Unless I'm missing something, this design formula should do the trick: ~ SampleGroup + Treatment

You can then monkey with the Treatment covariates to assess the effects of interest.

If you follow along using the DESeq2 vignette in section 1.6 "Multi-factor designs", you can treat your SampleGroup and Treatment variables just like the type and condition variables in the "pasilla" dataset -- in that dataset, you want to control for "type" (single or paired end) and test on "condition".

So, assessing the logFC of T1/T2 while controlling for "pen" would look something like this:

results(dds, contrast=c('Treatment', 'T1', 'T2'))

 

ADD COMMENT
0
Entering edit mode

Thanks for the fast answer Steve. The scenario you describe in the vignette is different than my example. In my pens (types in the vignette) I have only one treatment (condition). When I do as you describe  I get this:

Error in DESeqDataSet(se, design = design, ignoreRank) :
  the model matrix is not full rank, so the model cannot be fit as specified.
  one or more variables or interaction terms in the design formula
  are linear combinations of the others and must be removed

 

ADD REPLY
0
Entering edit mode

The problem is that the experimental design has confounded sample group and treatment. The sample groups associated with treatment T1 do not appear in treatment T4 and vice versa, so you cannot control for sample group and simultaneously estimate the treatment effect.

I recommend in cases like this that people use the duplicateCorrelation() approach within limma + voom, which let's you inform the analysis that certain samples are correlated, in the case that you cannot actually control for these using fixed effect terms due to the confounded design.

ADD REPLY
0
Entering edit mode

Indeed -- yikes, sorry. My brain crisscrossed what my eyes were seeing. Thanks for that.

ADD REPLY
0
Entering edit mode

Many thanks Michael. Really useful ! I will use limma.

 

 

ADD REPLY

Login before adding your answer.

Traffic: 676 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