Gene set enrichment analysis with logFC and PValue

Question

[Vasu]

Hi,

I have RNA-Seq data with 100 samples and 30k genes. Samples as columns and genes as rows. Tumor vs Normal.

After filtering step I see around 19k genes were used for differential analysis. With differential analysis cutoff FC > 2 and FDR < 0.05, there are about 1000 differential expressed genes.

I'm going to use foldchange and Pvalue for ranking the genes and input for GSEA

For Gene set enrichment analysis do I need to use only those 1000 differential expressed genes or do I need to use those 19k genes as input?

thanq

[ruhuiwen0608]

Hi, I have one more problem when I use this formula: signed fold change * -log10pvalue, as some of the pvalue = 0. So -log10pvalue returns infinite, which can not be processed further in GSEA. So what should I do with those pvalue = 0? Is it suitable if I simply replace the inf to some extremely large number in the rnk file? Thank you in advance!

[ATpoint]

Yes, for example .Machine$double.xmin in R.

Answer 1

[h.mon]

You have to use the 19k genes, but how you will do so depends on the enrichment method you are using. For GSEA, you have to use the ranked vector of all 19k genes.

[Vasu]

I'm actually using this for GSEA. Do you think ranking genes based on FC and Pvalue like below is right?

x <- read.table("DE_genes.txt",sep = "\t",header = T)
head(x)
x$fcsign <- sign(x$logFC)
x$logP=-log10(x$p_value)
x$metric= x$logP/x$fcsign
y<-x[,c("Gene", "metric")]
head(y)
write.table(y,file="DE_genes.rnk",quote=F,sep="\t",row.names=F)

I will use that DE_genes.rnk as input for GSEA. Could you please tell me something about this. thanq

[h.mon]

This seems fine, it is the same metric as used at Gene Set Enrichment Analysis (GSEA) explained.

[1769mkc]

if i do it only with p value or adjusted p value would it be logical?

[sangram_keshari]

Yes it will be logical, but it's just that fold change for few genes might be very less to be called as differentially expressed.

[bhanu.chandra1]

Hi h.mon, In the hyperlink provided, it is written as "signed fold change * -log10pvalue" In the above-mentioned comment, the following is used: x$logP/x$fcsign

Are both of them are similar?

[ruhuiwen0608]

Hi, I tried this function in R: x$fcsign <- sign(x$logFC), and it only returns you 1 or -1. So it would not change the result between "x$logP/x$fcsign" and "signed fold change * -log10pvalue".

[ruhuiwen0608]

I found a similar question here and some of the response could be useful to my own question: Many DESeq2 P values are 0 thus preventing generation of a rank list for GSEA

[opplatek]

In DESeq2, stat is very similar to this metric. A simple plot of DESeq2 stat vs your metric: .

X-axis is stat from DESeq2 Bioconductor R package and y-axis is the metric calculated as you described. Some more details can be found here and in DESeq2 vignette.

Answer 2

[sangram_keshari]

Hi, Basically, if you want to check on which biological state you gene set belongs its ideal to check some of differentially expressed top up-regulated and down-regulated genes. That number depends upon you. I saw people taking top 100 also and other times the whole set of up and down-regulated genes.