Select top differentially expressed genes based on both p-value and fold-change

Question

[GR]

I always selected top differentially expressed genes based on p-values (DESeq2 for differential-expr analysis). Just recently, someone suggested me to use both p-value and fold change to select top genes. I am little confused as p-values already take fold-change into consideration. Can someone shed some more light on what I am missing here?

[fuyanhu2015]

Hi， have you solved your problem? I am also confused about rankproduct. The reviewer asked me to consider both p-values and foldchange and use rankproduct to select top genes. According to reference, rankproduct is another method to find DEG. Could you please help me? Thanks.

[Joe]

You do not need to comment twice in the same thread. I'd suggest you ask a new question, since you sem to be more interested in rankproduct than the P-values or fold change specifically.

Answer 1

[Santosh Anand]

Fold change is a measure of the ratio of means of two populations (say control and treatment).

p value measures how much confidence you have in that ratio. The p value usually will take into account the difference in means and variances (or standard deviation) of the populations being compared, but not directly the fold-change. Of course, if the fold change is high => difference in mean is high => the gene has more chance to becomes significant. However, even if the difference in mean is small, and the variances in two groups are also small => you are confident that this difference is real and not coming just by chance => the gene could become significant. That is to say, p-value is not directly determined by fold change

[GR]

so what is the best way to select top genes? As I said, I was suggested to consider both p-values and fold-change and use rankproduct for this. Rankproduct is used for identification of DEGs. I am not sure how to apply it to foldchange and pvalue.

Any other ideas of combining pval and fc for seelcting top genes are welcome.

[Joe]

One typical approach would be to decide on a cut-off P-value you consider reasonable given your data and what you're asking. For instance, is P < 0.05 enough? Too many false positives? Go to P < 0.01 perhaps to reduce the size of your dataset?

Once you've defined a significance cutoff, then sort your DEGs that survive filtering by fold change, and take all of those above a fold change cutoff.Many people decide on log2 fold change cutoffs arbitrarily.

[Santosh Anand]

which, where, what rankproduct (Reference?)