Robust RNA-seq normalization

Question

[rasmus.agren]

I have an experiment with significant technical issues. There are large differences in RIN values, as well as in QC metrics from the sequencing. Coverage is bad and varying, with samples ranging from 2 - 6 M uniquely mapping reads (human samples). Unfortunately the samples are few and expensive, and the experimental design requires that all samples have to be used. The purpose of the analysis is global differential expression analysis.

Which is the most robust normalization method to deal with this? I can assume that the amount of RNA per cell should be constant and I expect a relatively small number of DE genes. I don't expect any correlation between the biological question and the quality metrics, but I only have 24 samples and multiple conditions so that could occur by chance. I would typically use TMM normalization and then include RIN as a covariate in the DE testing. Previously I've used SVA for correction, but I'm worried about distorting the biological signal. Are there any better approaches to this?

[ATpoint]

Can you show some diagnostic plots? Do the samples really separate on a PCA based on RIN?

Answer 1

[Kristoffer Vitting-Seerup]

The best way of dealing with it is to add those factors to the design matrix when you make your differential expression analysis model