How to apply a linear mixed model onto an RNA-Seq data?
0
0
Entering edit mode
Assa Yeroslaviz ★ 1.5k
@assa-yeroslaviz-1597
Last seen 11 weeks ago
Germany

Our data set contains 10 samples from three different conditions (We also have other timepoints, for now we are though only interested in what happens at the beginning). The metadata is as such:

sampleName  condition   time    urea    weight
C20 CTRL0h  0h  8.0 8.0
C22 CTRL0h  0h  8.0 8.0
C24 CTRL0h  0h  8.0 8.0
HP11    HP0h    0h  2.0 5.0
HP12    HP0h    0h  4.0 5.0
HP14    HP0h    0h  5.0 7.0
CR4Wo1  CR4W0h  0h  2.0 3.0
CR4Wo2  CR4W0h  0h  2.0 2.0
CR4Wo3  CR4W0h  0h  2.0 1.0
CR4Wo4  CR4W0h  0h  2.0 2.0


We would like to apply a linear (mixed) model to the data set to understand how the two factors urea and weight affects gene expression. But I'm not sure if this is possible at all here with the samples we have, as I don't have a complete set of combinations for the two factors i would like to analyze. Do I need to have more samples in able to do that?

I would appreciate any ideas/help as to how i can (if at all) apply such analysis to the data I have.

limma linear mixed model voom rna-seq • 508 views
0
Entering edit mode

If what you want to know is the proportion of variance (of gene expression) explained by the covariates, then you could try the variancePartition package. You can also use its' dream function, which is a linear-mixed model extension of limma-voom, if you want to do repeated measures differential expression analysis later on. Remember to use ddf = "Kenward-Roger" for the small-samples adjustment for your DE analysis.

From what experiments did you get your samples from? If you have minimal intra-group variations (like from mice studies or cell cultures), then perhaps you have (barely) enough number of samples.

0
Entering edit mode

Correct me if I’m wrong, but I think the duplicatecorrelation function in limma is equivalent to adding a random effect.

0
Entering edit mode

Yes indeed, however, limmaestimates the random effect globally (i.e. for all genes). Whereas dreamestimates the random effect separately for each gene. So, if you have a lot of inter-individual variation on the gene expression, IMHO dream will call less false positive results.