FPKM confidence interval - cufflinks

[Boel]

Theoretical question about the estimation of expression level (FPKM) in cufflinks-cuffcompare. In the tmap files produced by the Cufflinks suite of programs a confidence interval (actually, a probability interval?) is given. I've assumed that this is the confidence interval for the estimation of the FPKM. The cufflinks webpage give a reference to a paper by H. Jiang and W.H. Wong where the analytical steps towards determining an expression level are discussed (I might have to read this paper a couple of times more..).

I've been pondering about the best way to use the uncertainty in the expression levels when determining differential expression. This together with biological replicates should show the "real" variation of expression levels better than only using the point estimates from each sample (when using few biological replicates as most of us seem to be doing). Or are these confidence intervals only an indication of how uncertain the expression level is - not something that can be interpreted as an indication of the variation in expression levels?

Has anybody else had any success in using the confidence intervals?

Sorry for a confused thread, any suggestions would be highly appreciated!

Thanks,
Boel

[lpachter]

The uncertainty in FPKM estimates that is reflected in the Cufflinks confidence intervals is unrelated to the variability one may observe from biological replicates when performing RNA-Seq. Rather, it is a measure of the inherent uncertainty resulting from the random sequencing experiment and from the ambiguity in assigning certain reads to transcripts.

[Boel]

Thanks, thats what I suspected.. Still, one would want to use this uncertainty in the assessment of differential expression etc, right?

[zee]

I posted this in another thread with no response so I'll try asking this group. How do we get the actual read counts per gene/isoform from Cufflinks. Is there a way it could be back-calculated from FPKM?

[Boel]

copy paste

Multiplying the average depth of coverage by the transcript length will give you the *estimated* number of reads assigned to each transcript. Note that Cufflinks' statistical model probabilistically assigns reads to individual isoforms, because when isoforms share exons, reads from those exons can't be unambiguously assigned to one particular isoform.

from this thread:

http://seqanswers.com/forums/showthread.php?t=3961&highlight=cufflinks+read+count

If you are interested in the statistical model its derived from the models explained in this article:
Statistical inferences for isoform expression in RNA-Seq. by Jiang H, Wong WH. PMID: 19244387

[Wolfgang Huber]

Thanks, thats what I suspected.. Still, one would want to use this uncertainty in the assessment of differential expression etc, right?

The answer will depend on what you mean by "differential expression", and in many cases probably be "no". Typically, people want to make statements about differential expression that take into account both technical and biological variability, and biological variability is (at least for abundant transcripts) greater than the technical noise, thus the FPKM confidence interval will be many times smaller than what would be a useful overall confidence interval for differential expression inference.

Best wishes,
Wolfgang

[lpachter]

To clarify again, the Cufflinks confidence intervals have nothing to do with "technical noise" (e.g. lane effects). They reflect the uncertainty in assigning reads to transcripts- an inherent aspect of the RNA-Seq experiment based on relatively short reads.

As far as measuring differential expression, it is essential to use the confidence intervals that Cufflinks provides. Of course, as Wolfgang says, they provide only a lower bound on the true variability (because they omit uncertainty due to "noise" and also due to biological variability). With additional data (replicates), that information can be used to expand the intervals. But in the absence of it, the Cufflinks confidence intervals are better than not some heuristics that have currently been applied (e.g. 2-fold change).

[Boel]

How to estimate the standard deviation of the FPKM values?

To clarify again, the Cufflinks confidence intervals have nothing to do with "technical noise" (e.g. lane effects). They reflect the uncertainty in assigning reads to transcripts- an inherent aspect of the RNA-Seq experiment based on relatively short reads.

As far as measuring differential expression, it is essential to use the confidence intervals that Cufflinks provides. Of course, as Wolfgang says, they provide only a lower bound on the true variability (because they omit uncertainty due to "noise" and also due to biological variability). With additional data (replicates), that information can be used to expand the intervals. But in the absence of it, the Cufflinks confidence intervals are better than not some heuristics that have currently been applied (e.g. 2-fold change).

This is exactly what I am trying to do (use FPKM CI to expand intervals), but I am unsure about how to estimate the SD of the FPKM estimates. Are the CI calculated according to the Jiang & Wong paper, where the dist of FPKM is approximated by a normal-distribution?
And, how would one do this for transcripts with low expression, where the CI starts at 0 and not symmetrical?