FPKM or counts?

[cnyh]

I have trouble figuring out whether to use FPKM values for measuring expression, or to stick with raw counts. I've heard that we might have too few biological replicates from each tissue, to be able to rely on the FPKM values. Does anyone else have experience with this issue? How do I decide which to use?

[bruce01]

You say you want to measure expression. But are you looking for differential expression between samples, or absolute expression to see what is most highly expressed?

FPKM is a measure of absolute abundance of a gene and can therefore be used to compare expression between genes.

Counts are relative. For differential expression analysis you are not looking between genes, but within them across replicates to see if a gene is more highly expressed in a condition or treatment.

If you are doing DE analysis using DESeq or edgeR for example, use counts. To look at sets of genes which may be co-expressed, for example, then FPKMs may be of interest.

[cnyh]

Thank you for your quick response!
- It is indeed differential expression between different samples that we are interested in.
- We have primarily been using cuffdiff for the purposes of DE.

How relevant the issue of having only 3 biological replicates for each sample, in deciding whether to choose counts or FPKM?

[iris_aurelia]

I usually stick to the counts. Using the counts you do know exactly how many reads are mapped to a gene, which I prefer. I always normalize the counts for its library size in order to compare the counts across samples.
Cufflinks does correct for gene length, but I don't think there is a need to correct for gene length when only comparing genes between samples.

In order to get the differentially expressed genes I usually use the Voom method which is in the Limma/edgeR package. This method takes raw genecounts as an input and does normalize the data within the voom method.

[bruce01]

Not sure the number of replicates is relevant at all in using FPKM or counts.

Personally I have never been happy with cuff* analysis. It seems very conservative. I like to get something out of my DE analysis, but then someone may criticise that attitude!

I would try using count data in edgeR if you can use R. The manual is pretty helpful and there are many tutorials on line. 3 replicates per condition is ok, the issue is you won't have too much confidence in the results, unless you use cell culture with a very well defined response to treatment(?)

@iris_aurelia I agree with not needing to correct gene length: the comparison is within the gene, not between them.

[cnyh]

I see. And I assume by that one can avoid the stringency issues that cuffdiff has with calculating q-values? That would be very promising. Thank you very much.

In that case I have a related question, but I don't know what the proper protocol is with asking separate questions within a single thread. Maybe I can link to it here: http://seqanswers.com/forums/showthread.php?t=28117

[pengchy]

The question still exists: what's the difference between Cuffdiff counts and HTseq counts? And why the counts of cuffdiff have decimal?

[cnyh]

The reason for that is that cuffdiff additionally normalizes the counts values.

[pengchy]

Originally Posted by cnyh

The reason for that is that cuffdiff additionally normalizes the counts values.

If cuffdiff has normalized the counts value, it will be not need to normalize again when we using edgeR/DESeq to detect DEG with the output of cuffdiff counts?

[cnyh]

That's what I understand, but I'm no expert in this.

[carmeyeii]

@Cynh @Pengchy

Cufflinks output should not be used for DeSeq/EdgeR. These use raw counts, which you can get after aligning with TopHat using HTSeqCount or similar programs.

Check this previous thread; there are a couple others referring to this issue.

http://seqanswers.com/forums/showthread.php?t=5793