Question

Adjust for base line gene expression in limma model

0

Entering edit mode

S • 0

@0b4d0d5b

Last seen 3 months ago

Norway

Hi,

I was wondering is there a way to include baseline gene expression as covariate in model matrix formula for limma? I have microarray data from clinical trial patients in a control and an intervention group and I want to adjust for baseline gene expression changes in the linear model along with other covariates. Does the lmFit function of limma allow for this? If not, would it be advisable to just run linear model separately for each gene and use the baseline gene expression in that?

Code should be placed in three backticks as shown below


# include your problematic code here with any corresponding output 
# please also include the results of running the following in an R session 

sessionInfo( )

limma Microarray • 2.2k views

ADD COMMENT • link updated 3.0 years ago by Gordon Smyth 52k • written 3.0 years ago by S • 0

1

Entering edit mode

Can elaborate a bit more about what you mean by "baseline gene expression"?

Typically when I think of that term, I think of having something like two samples from the same person/subject before/after intervention, and you can include subject_id id as a term in the linear model to account for differences in individual "baseline expression"

Or you may expect to see differences in baseline expression between, say, males and females, and you want to adjust for that before doing the differential expression test. In this case you would add sex as a covariate.

Is that something like what you're thinking of?

ADD REPLY • link 3.0 years ago Steve Lianoglou ★ 13k

0

Entering edit mode

Hi,

Thank you for your insights. I actually meant gene expression measurement pre and post intervention which in this case are two different dietary patterns. Sorry for the confusion. I am interested in finding genes that are significantly changed in the intervention group diet compared to the control group diet.

ADD REPLY • link 3.0 years ago S • 0

score 1 · Answer 1 · 2021-12-16

1

Entering edit mode

Gordon Smyth 52k

@gordon-smyth

Last seen 28 minutes ago

WEHI, Melbourne, Australia

Do you have a repeated measures experiment or a simple comparison with independent subject? In repeated measures experiments, we apply multiple treatments to the same patients. When we analyse such an experiment, we have to allow for the fact that different patients have gene expression levels that start at different levels, even before the treatment is applied. Hence we have to include patient as a term in the linear model. I have described this in the past as "adjusting for baseline differences between the patients".

However baseline expression does not change and it shouldn't be treated as a covariate. Indeed the whole point of the term "baseline" is to refer to something that does not change. In your case, you appear to have a two group experiment with different subjects in each group. If that is correct, then the concepts of "baseline expression" or adjusting for patient effects do not apply to your experiment.

ADD COMMENT • link 3.0 years ago Gordon Smyth 52k

0

Entering edit mode

Hi,

Thank you again for your sights. What I have is gene expression data from a dietary intervention study of two groups of individuals fed two different diets where their gene expression were measured pre and post intervention. I am trying to compare the two groups to see see which genes are significantly changed in the group fed the experimental diet vs the control diet.

ADD REPLY • link 3.0 years ago S • 0

0

Entering edit mode

OK, it appears from your new comment that you have independent diet groups but also pre and post treatments on the same individuals, i.e., you have a sort of multi-level design. There are a number of ways such data could be analysed. Usually I would estimate the post vs pre treatment effect separately for each diet and also the difference between the two post vs pre effects (i.e., the diet by treatment interaction). Patients would enter the model as an additive factor. There would be no need to adjust for covariates. See Using edgeR on pre-post treatment control design or Limma when to use zero intercept for ways to created the design matrix for experiments like yours.

To get definitive advice you should explain your experimental design more completely, ideally by giving the complete targets frame.

ADD REPLY • link 3.0 years ago Gordon Smyth 52k