log2(FC) +/-inf. What do you do with this genes?

[mpalegre]

Hi verybody. I have analyze my RNA-Seq Data using TopHat2/CufDiff. I´m not sure how filter this results (output CuffDiff). Because the genes with +/- inf have 0 reads in any condiction but the FPKM in the other sample is inconsistents. this means if you visualizes these genes in IGV i.e., there is a sample with 0 reads and the other one has many very few reads though the FPKM is high?????? Anyone can i help me.

Thanks in advance

regards

[mbblack]

In my opinion, you cannot say anything about those genes for which you had zero counts. All you know is that you failed to detect those genes or transcripts in that sample. A failure to detect cannot be interpreted as a true absence of those features in a given sample. Thus you simply have no data at all for those features, and thus cannot say anything at all (that is biologically meaningful) about those features.

You should be filtering out those zero count features from your analysis - you cannot base a conclusion or an interpretation on the absence of data.

If you had done qPCR on these same samples, and some genes simply failed to amplify at all, would you have included those data in your analysis? A read count of zero is the same thing, IMO.

[mpalegre]

Thanks for your opinion

[sdriscoll]

That's interesting. Often we do knock-out experiments in my lab and we end up seeing 0 reads for those genes (depending on the knock-out approach of course). Would you still say that a 0 count has no confidence in that case?

[blancha]

Actually, if a gene had a count of 0, I would conclude that the expression level of the gene is too low to be detected at that sequencing depth.
This is useful information that should not be ignored.

Just set a minimum average FPKM threshold across the conditions being compared.
If the FPKM values are low across all the samples of the conditions being compared, you cannot conclude anything. However, if the FPKM values are low in only one of the conditions, you definitely want to keep this information.

Once you've set your average FPKM threshold across the conditions being compared, you'll find the results to be much more consistent. The higher the FPKM values, the more reliable the results will be and the less variation there will be between the replicates.

I hope this make things clearer. Your question is not easy to understand. If you are from Sevilla in Spain, you could always ask your question in Spanish.

[mbblack]

That's interesting. Often we do knock-out experiments in my lab and we end up seeing 0 reads for those genes (depending on the knock-out approach of course). Would you still say that a 0 count has no confidence in that case?

If you have independently confirmed that your knockout truly has eliminated expression of the gene in your system, then of course that is a different situation. But I would not take the count of zero in an RNA-seq experiment as the evidence that a knockout was truly effective.

[GenoMax]

If you are from Sevilla in Spain, you could always ask your question in Spanish.

But then the rest of us would not understand the question

[mbblack]

Actually, if a gene had a count of 0, I would conclude that the expression level of the gene is too low to be detected at that sequencing depth.
This is useful information that should not be ignored.

Or, that feature is truly not expressed in that sample or tissue or under those conditions. But you have no way of distinguishing between the two situations. And you certainly cannot say anything quantitatively about relative expression in one case when you have nothing at all to compare to in the other.

[blancha]

If you have independently confirmed that your knockout truly has eliminated expression of the gene in your system, then of course that is a different situation. But I would not take the count of zero in an RNA-seq experiment as the evidence that a knockout was truly effective.

Why not?
I do.

There is one caveat, which is always present in RNA-Seq.
If the gene is expressed at a low level, it may only be detected at a higher sequencing depth.

Or, that feature is truly not expressed in that sample or tissue or under those conditions. But you have no way of distinguishing between the two situations. And you certainly cannot say anything quantitatively about relative expression in one case when you have nothing at all to compare to in the other.

Perhaps we are just disagreeing on the wording.
If I have 0 FPKM in one condition and a high FPKM in another condition, I would conclude that the gene is not expressed or very lowly expressed in one condition and very highly expressed in the other condition.
This is useful information that should not be discarded.

When the FPKM values are very low or nil in both conditions, nothing can be concluded on the relative expression levels in the 2 conditions

Anyway, I should get back to working, and spend less time on this forum.

[mbblack]

Why not?
I do.

There is one caveat, which is always present in RNA-Seq.
If the gene is expressed at a low level, it may only be detected at a higher sequencing depth.

I just don't know anyone who would be comfortable using lack of detection in an RNA-seq experiment as confirmation of a successful knockout. The gold standard still seems to clearly be demonstrating the absence of the protein, not a failure to detect expression of the transcript.

I know we've had knockouts that showed no evidence of expression with qPCR and digital PCR, but the actual protein was in fact detectable in small amounts by western.

[sdriscoll]

I just don't know anyone who would be comfortable using lack of detection in an RNA-seq experiment as confirmation of a successful knockout.

I see what you're saying. Yes, if a sample is taken out of context of the experiment (controls and replicate KO samples) then a count of zero is not conclusive. However when put into context with the other samples if that count of 0 is consistent and the count is very high in controls then one might conclude the gene was AT LEAST knocked down significantly. I agree that zero counts are always ambiguous since, as mentioned above, with additional sequencing depth it's possible we would get some hits eventually.