RNA-seq Differential Gene Analysis

Question

[mlai2567]

Hi All,

I'm currently performing differential gene analysis on my RNA-seq data set. What is the common threshold utilized for analysis in terms of fold change. I've currently set the fold change threshold at 1.5. Is this stringent enough?

Thanks in advance!

[russhh]

I never apply a fold-change threshold for differential expression.

[russhh]

The elaboration being that I only ever apply p-value / FDR thresholds for differential expression; what matters is reproducibility not the estimated change. For biomarker discovery, I might apply an FC threshold.

[ahaswer]

Can you elaborate on that?

[Kevin Blighe]

I think that some just recommend to order genes by fold-change, as opposed to just setting a hard filter threshold. Any good and honest analyst will understand the limitations of setting a hard threshold, and also the limitations of not setting a threshold.

[ahaswer]

That's for sure. I was just bit surprised to hear that one should never apply FC threshold.

[Kevin Blighe]

Nobody implied that nobody should ever apply such a threshold. russhh just stated that he never uses it.

[Teo]

Hi mlai2567, can I ask you how did you perform the differential gene expression analysis? I have two dataset, control and stimolate, and I would like to know what is the fold change of all genes of a specific population (in my case the subpopulation is macrophages Adgre1+)

Answer 1

[noorpratap.singh]

It depends on the needs. Ideally first a FDR (mostly 0.05 but this can also change) is set and then the fold change is applied. Start with a lower threshold (say 0.5 or 1) and observe the total number of genes derived. Do an enrichment and see the key processes arrived at. Then increase the threshold (1.5,2) and again see the number of genes and the corresponding processes. The number and information of genes that you find suitable at a particular cutoffs defines the so called fold change. It depends on the data and the phenotype within which you want to find the differences and how heterogeneous and different they are. There is no fixed or a given fold change.

[Kevin Blighe]

Moved to an answer as this is, from my perspective at least, the one and only way to answer a question on this topic. In addition, mlai2567, you should clarify whether you are using log (base 2) fold changes - most likely you are. Keep in mind that, if using log (base 2), then a cut-off of 1.5 still equates to an un-logged difference of 2.83 fold change (2^1.5)