0
$\begingroup$

I'm using DESeq2 Package in R for RNA-SEQ Analysis. Using Count Matrix. DE analysis is between two groups for eg: like Normal vs Tumor.

To select for the significant differential expression of genes in RNA-seq do I need to consider both FDR and log2FC (or) only FDR is fine?

When I consider the cutoff FDR < 0.05 & log2fc > 1/log2fc < 1, among 51539 I see 137 ~ Up-regulated and 540 ~ down-regulated genes

When I consider the cutoff only FDR < 0.05, among 51539 I see 6984 ~ Up-regulated and 7115 ~ down-regulated genes

And from these DE genes I wanted to filter lncRNAs based on lncRNA annotation.

So, which one is the priority for selecting the significant DEGs?

$\endgroup$

1 Answer 1

2
$\begingroup$

If you think I misinterpret your question, please tell me. I'll edit my answer.

To me your question is:

Should I use FDR or log-fold in my analysis?

Log-fold gives you the fold-change between the two conditions. You shouldn't use the magnitude of your LF to decide which gene is statistically differentiated because:

  • It's just a number and thus no probability distribution, no p-value, no confidence interval, no null hypothesis and no inference.
  • It's sensitive to lowly-expressed genes where the variability is high and thus your LF value might not be accurate.
  • You don't "borrow" information from other similar abundant genes.
  • What's the point of using DESeq2 if you just want to use log-fold? The point of DESeq2 is to estimate dispersion for your negative binomial model (because you have counting data).

You should use the FDR column. The FDR column gives you adjusted p-value (q-value) for each gene. Compare each q-value with your significance level.

Note: FDR and log-fold are two very different thing

$\endgroup$
7
  • $\begingroup$ Yes, What you said is right. But in some papers for RNA SEQ analysis I see that they used both q-value and log2FC to get significant genes. Andin some other they used only q-value to get significant genes. So, I'm confused in this part. $\endgroup$ Mar 1, 2017 at 11:10
  • $\begingroup$ @user3351523 Do you have a paper? I'll take a read and edit my answer. $\endgroup$
    – SmallChess
    Mar 1, 2017 at 11:13
  • 1
    $\begingroup$ Significantly differential expression based on FDR < 1% [nature.com/articles/srep37821] See in the Results: Paragraph "Global changes of the lincRNA species in breast cancer" (Last second line) $\endgroup$ Mar 1, 2017 at 11:26
  • 1
    $\begingroup$ Significantly Differential expression based on two criteria |log2FC| >1 and qvalue < 0.05 [nature.com/articles/ncomms14421] See in the Methods: Differential expression analysis (second paragragh) $\endgroup$ Mar 1, 2017 at 11:30
  • 1
    $\begingroup$ @SmallChess you certainly need to look at both p-value and fold change. p-value is the significance of the test of different means. The fold change in DESeq is the effect size. In a biological context, it is the effect size that matters. However, if you see a large effect size on a test with no significance, you can't trust that fold change value. Thus you need both and should always report both. This is true regardless of how critical authors and journal reviewers are. See: stats.stackexchange.com/a/238326/141304 $\endgroup$
    – abalter
    Aug 18, 2019 at 8:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.