Deseq2 using 2x2 design

Question

[cassandraregine]

Hello,

I'm trying to use Deseq2 to find differences in microbiota composition in a 2x2 crossover study. I am not sure what model to use to detect differences between P (which is a placebo) and T (treatment). Each subject only has the two time-points (one with P and one with T). Below is my colData.

 Subject   Treatment    Phase   Sequence    Combined1   Combined2
  S1           T     One       BA         T_One   T_BA
  S1           P     Two       BA         P_Two   P_BA
  S2           P     One       AB         P_One   P_AB
  S2           T     Two       AB         T_Two   T_AB
  S3           T     One       BA         T_One   T_BA
  S3           P     Two       BA         P_Two   P_BA
  S4           T     One       BA         T_One   T_BA
  S4           P     Two       BA         P_Two   P_BA
  S5           P     One       AB         P_One   P_AB
  S5           T     Two       AB         T_Two   T_AB
  S6           P     One       AB         P_One   P_AB
  S6           T     Two       AB         T_Two   T_AB
  S7           T     One       BA         T_One   T_BA
  S7           P     Two       BA         P_Two   P_BA
  S8           P     One       AB         P_One   P_AB
  S8           T     Two       AB         T_Two   T_AB

I tried combining Treatment with Phase and Treatment with Sequence and then using the below regressions (switching out Combined1 and Combined2 leads to the same error message).

dds <- DESeqDataSetFromMatrix(countData = data,
                          colData = colData,
                          design=  ~ Subject + Combined1)

Whenever I do this, I just end up getting an error message of-

Please read the vignette section 'Model matrix not full rank':

I would like to test for the bacteria that are significantly different between the two treatments (P vs T), but I also want to control for sequence due to the crossover design.

What design seems appropriate?

Thank you.

[jv]

is subject S2 the only subject the received sequence AB? If so, I think that might be the issue since sequence AB is confounded with subject S2.

[cassandraregine]

They are not! I updated the table. I originally only included half. Thanks.

[jv]

That helps. I'm not a statistician, but the current design of ~Subject + Combined1 doesn't seem to me like a suitable way to test for treatment effect.

I googled "DESeq2 2x2 cross over design", there are some discussions on Bioconductor support and on Biostars, for example https://support.bioconductor.org/p/p132767/ and https://www.biostars.org/p/482861/ but I think your study design or question differs a bit these other discussions.

It seems to me that testing for T versus P while controlling for subject and sequence variation would use ~ Treatment + Subject + Sequence as an initial design.

[swbarnes2]

Each subject was only done with only two conditions. I don't think you can include subject in the design