How can we compare EdgeR to limma-voom?

Question

[Barista]

Hi all,

I did all my RNA-seq analysis for 500 patients with the EdgeR workflow as described in the manual. Now I read an article (F1000 research) where they state that EdgeR is mostly preferred in low-count situations, where limma-voom is recommended in large scale datasets. I thought using EdgeR was fine for my experiment, but this article has me worried. Would limma-voom be preferred in my case or is EdgeR equally fine? Maybe good to know, I have no problems with computational times when using EdgeR.

Many thanks in advance!

Answer 1

[bigomics.team]

Low count and large scale are different things. Low count means that average counts are low so in those cases Poisson/neg-binomial models are supposed to work better. Large scale mostly means more samples, in your case 500 samples is considered quite a lot. Still you can have "large scale with low counts". If you have no computation problems with EdgeR, then I would stick to that.

[Barista]

Yes I noticed the difference, I do not know if I can regard my experiment of low count, the libraries vary between 2M and 10M reads. For normal RNA-seq this is quite small, but because of QuantSeq the number of reads is lower than in traditional RNA-seq experiments.

In general, if computation is no problem, would EdgeR be a suitable option, also if you have a large scale, high count situation? In other words, is the problem of EdgeR mostly that it is more computational intensive or are there other reasons why one should choose limma over EdgeR?

[ATpoint]

I do not think that you have to worry. For such a large sample size limma-voom should be fine. In fact benchmarking papers usually rank DESeq2, edgeR and limma-voom highly for RNA-seq, see e.g. this blog post from the DESeq2 author: https://mikelove.wordpress.com/2016/09/28/deseq2-or-edger/

He uses limma-voom himself when sample size is large, mainly due to the speed benefits. But if you want to use edgeR that is (imho) perfectly fine and no reviewer would probably complain about this. Remember, RNA-seq (or statistics) becomes easier when sample sizes are large and difficult if small, this is why these special approaches have been developed.

[bigomics.team]

I beg to differ. You need to differentiate between "large number of reads" and "large number of samples". See my comment above.