Transcriptional noise

[chadn737]

In a lot of my RNA-seq data there are hundreds to thousands of genes that are represented by only one or two reads. Is this data biologically relevant or is it merely the result of transcriptional noise in the system? If it is transcriptional noise, how do you determine that and how do you go about determining a value you to filter out this noise?

[dariober]

Hello,

In case you haven't come across it already, I point out this paper which suggests it is transcriptional noise:

Most Dark Matter Transcripts Are Associated With Known Genes
http://dx.doi.org/10.1371%2Fjournal.pbio.1000371

Dark Matter Transcripts: Sound and Fury, Signifying Nothing?
http://dx.doi.org/10.1371%2Fjournal.pbio.1000370

I'd like also to hear some comments about your questions...

All the best
Dario

[steven]

Old debate. Here is another point of view:
The majority of total nuclear-encoded non-ribosomal RNA in a human cell is 'dark matter' un-annotated RNA.
Also, don't forget to take a look at http://pervasivetranscription.com/

[steven]

BTW, an alternative question: how do you know you have been sequencing enough?

[chadn737]

I have read those papers actually and I don't think they address really what I am talking about. What I meant by transcriptional noise is not reads mapping to unannotated regions, but rather reads that map to annotated genes. For example, in one of our data sets from a couple years ago, we had two biological replicates each at ~7 million reads. In one sample, only one read is mapped back to Gene A, while in the second replicate there is zero. For Gene B, two reads map in one sample, while one in the second sample.

If only one read is mapping to a gene, is that biologically relevant or is that simply noise and if it is noise, how do you filter that out? Some studies have shown that up to ~90% of transcription initiation events by RNA PolII is simply noise and not really relevant. http://www.nature.com/nsmb/journal/v...b0207-103.html

Now I know that 7 million is not that deep and we have new data that is far deeper, but still, there is this issue. Furthermore, how deep is deep enough? Is it 10 million reads? 20 million? 50 million? I think the deeper you go, the more genes that are going to be represented. Is this biologically relevant or not? One paper tried to estimate the level of noise by looking at reads mapping to regions that have no known genes in a 10Kb region.

http://www.nature.com/ng/journal/v42/n12/full/ng.703.html

I have a lot of caveats about this approach, but it seems like a first step in an issue that has not been addressed.

[steven]

A few comments:

1) With so few reads and so few samples, no statistical significance can be reached if you compare the genes or the conditions. That does not mean that a gene is "not (more) expressed", but only that your experiment does not allow you to prove it.

2) We both raised the same question, which i think is the key here: how deep is deep enough? I am looking forward to a general and empirical method to define this.

3) More generally, I believe that the concept itself of "noise" is deeply wrong in molecular biology, because by definition it is impossible to prove that an event is not functional. You can never prove that a transcript is not present, you cannot prove that a gene is not expressed, you cannot prove that a mutation has no consequence. You can only attempt to prove the opposite and fail, therefore concluding that "no significant difference could be observed". Molecular biology is an experimental science, with new mechanisms being discovered every year, that could not be previously discovered. I remember when alternative splicing events were first considered as exceptional mistakes

[Richard Finney]

Certainly, there does appear to be "background noise" in RNAs from a genome. The wet lab process may induce bogus, extremely low level expression for some truly "off" genes. It's also possible that a cell doesn't need a whole lot of a particular protein to do it's thing and that a very low level is all you're going to get for that gene. I'd love to see alternate protocols and technologies used to see if the same evidence for low level transcription appears. If so, then maybe it's signal, not noise.

If you have many case and control samples and you see the low level in one set but not the other, you may have your long sought statistical significance.

[steven]

I agree that there is some "background noise" in all experimental results, but in the cell no one knows where the threshold is -if any.

If a transcript does exist, I do not think it makes sense to call "noise" the transcriptional activity that produced it. There is no way to prove that a transcript has no function, and there is no way to prove that pervasive transcription is "random noise".
But i understand it is comfortable to think that way: if you do not understand something, just call it trash! It's the so convenient "junk DNA" philosophy..

What you call noise may be music for the cell

[Richard Finney]

I'd have to agree. It's probably real. I've stared at too many < 1X coverage for a gene in one sample and seen very similar < 1X on other samples, same study. If it is "noise", it's pretty consistent. I'm okay with getting proved wrong, though.

[amitm]

A few comments:
.
.
3) More generally, I believe that the concept itself of "noise" is deeply wrong in molecular biology, because by definition it is impossible to prove that an event is not functional. You can never prove that a transcript is not present, you cannot prove that a gene is not expressed, you cannot prove that a mutation has no consequence. You can only attempt to prove the opposite and fail, therefore concluding that "no significant difference could be observed". Molecular biology is an experimental science, with new mechanisms being discovered every year, that could not be previously discovered. I remember when alternative splicing events were first considered as exceptional mistakes

hello steven,
thats an absolutely unbiased view! Amazing, congrats And a very apt citation too.