Threshold of removing genes with low counts ?

Question

[hafiz.talhamalik]

What should be the threshold of removing genes with low counts in RNA-seq data ? i have removed all genes with sum of counts 0, but there are many genes whose sum of counts are 1 , 2 ,3 etc. should i remove those ? or removing them will effect my results ?

[VHahaut]

If your question is about differential expression analysis this should help you answering the question. Else try to be more precise on what you want to do.

https://support.bioconductor.org/p/63234/

It depends which differential expression algorithm you are using. DESeq2 seems to deal with this problematic for you (EdgeR too I suppose?) but if I remember well you'll have to it manually with limma (this need to be confirmed).

[hafiz.talhamalik]

i am using edgeR. i will also use DESeq2 for the comparative analysis, but first i need this thing solved for edgeR

[VHahaut]

So if you are using edgeR I advice you to read this post:

https://www.biostars.org/p/93553

[hafiz.talhamalik]

Thanx alot for your kind help

Answer 1

[firatuyulur]

I do not think there is a certain threshold. When it comes to drawing a plot out of counts, the zero lines do make your computer suffer as there will be a couple of thousand 0 lines that I remove them directly. While analyzing, you will use differential expression, and there will be a p-value / p.adj-value where a count going from 0 to 1 across control and treatment samples will be far less significant than a count going from 100 to 500. My point is, either you remove them manually or not, your end point will be filtering based on the significance of the expression change and most of the small numbers will be filtered out.

[VHahaut]

I don't think your answer is entirely correct. It is true to say that 0-1 counts won't be significant with programs such as DESeq/EdgeR. But on the other hand, if you leave them in the analysis, you will increase the adjusted pvalues leading to potentially less significant results on the 100-500 read genes (when you adjust for multi-testing, the more tests you have the more stringent it is). So filtering can impact your final results.

Answer 2

[Agustin Gonzalez-Vicente]

This may be of some interest.

Filtering and collapsing data