Great explanation! Thanks!
Have to clarify that when I mention foldchange 2, it is not log2.
For differentiated expression genes, if taking high confident hits for
downstream validation and analysis, it should be more consistent and
reliable. However, I think false positive rate should increase for
hits close to the cutoff line, and I am wondering how much this is
caused/affected by the noise level of the method/data. Is any
literature discussing about this? Say how the noise of data affects
cutoff and false positive rate. If so, can cite them here? Thanks!
From: email@example.com on behalf of Simon Anders
Sent: Tue 3/27/2012 10:06 AM
Subject: Re: [BioC] Up & Downregulated genes using DESeq
On 03/27/2012 03:43 PM, Sunny Yu Liu wrote:
> Because no matter how fancy the statistics is, we have to find
> having biological significance. For example, in most time, even have
> p<0.000000000001, still get no biological effects.
> For most gene expression change, people always use fold change 2 as
> cutoff for microarray or qPCR. As for RNAseq, since the method is
> more sensitive, I guess it must lose some specificity, so I think it
> need a higher cutoff number than 2.
I guess, this needs some clarification, before we confuse newcomers to
RNA-Seq too much,
A cut-off of 2 on a log2 scale means a fold change of at least
four-fold. This is a lot, and there is plenty of cases where a weaker
signal is biological meaningful. Depending on the experimental setup,
fold changes of +/-20% (.26 on a log2 scale) or even much less can be
In most RNA-Seq experiments, a p value cut-off to control false
rate at some sensible value, say 5% or 10%, will not allow any genes
with a log2 fold change below a certain value to be called
and in my experience, this value nearly always make further fold-
This belief that fold-change cut-offs are important may stem from the
proliferation of incorrect analysis methods in the RNA-Seq literature.
In very many papers, analysis methods based on Fisher's test or on a
likelihood ratio test based on a Poisson distribution are used. In
there are even several reviews which suggest such approaches.
Apart from the fact that these tests simply inadmissible (see e.g.
Baggerly et al., Bioinformatics 19 (2003) 1477), they cause a peculiar
pattern: They give all genes with expression strength above a few tens
of thousand reads absurdly low p values even in case of very weak fold
changes, so that nearly all strongly expressed genes are significant.
I've seen in several posts on this and other mailing lists the advice
use a very small p value threshold (adjusted p value < .001 or the
like), combined with a fold change cut-off, to rectify the situation.
With a correct test, it is rare that you have significant genes with
weak fold change that you have to doubt their biological relevance.
I see, however, at least one case where a fold-change cut-off is
It is unavoidable that the decision boundary that separated
from non-significant fold-changes goes down with increasing count
values. Hence, any list of "hits" will be enriched for strong genes.
this is problematic for downstream analysis, one may opt, as a crude
workable remedy, to decide on a fold change cut-off and consider all
genes below this as not significant _and_ omit from the universe of
enrichment tests all genes with a count value below the count required
for this fold change to becone significant (i.e., considering these
genes as essentially "not testable" rather than "not significant").
Bioconductor mailing list
Search the archives:
[[alternative HTML version deleted]]