RNA-seq, three comparisons - stick to one linear model, or use separate models?

Question

[noodlejackson]

Hi all, I have RNA-seq data from tumour samples divided into three groups:

1. Primary tumour tissue from non-recurrent patients
2. Primary tumour tissue from recurrent patients
3. Recurrent tumour tissue (some matched to #2)

I’m running differential expression analyses for three comparisons:

1. Primary tissue of non-recurrent vs primary tissue of recurrent patients
2. Primary tissue vs recurrent tumour tissue (all samples)
3. Primary tissue vs recurrent tumour tissue in recurrent patients with matched samples

For #2 and #3, I’m using mixed effect models to account for repeated measures.

For #1, would it be better to run a separate linear model, using only primary samples from non-recurrent and recurrent patients? My reasoning is that a simpler model avoids extra parameters and could improve sensitivity, since the groups are independent and unmatched.

Or should I keep everything in a unified model for consistency?

Thanks for any input!

Answer 1

[ATpoint]

There is no scrict rule for this. Generally, if variability between groups is somewhat in the same ballpark, e.g. approxiated by PCA, then keep them together. That assumes that these are from the same batch, or batch is balanced. After all, if it's in the same model then using all samples improves the mean-variance trend estimation which is a critical basis for the entire analysis. Since Primary tissue seems to be used in all contrasts I would keep them together most likely. SInce you have repeated measures you might use limma-voom or limma-trend which can model that via duplicateCorrelation and generally is very flexible for any design.

[noodlejackson]

Thank you @ATpoint