I have a conceptual question that I was hoping someone could answer.

Can I say that microRNA A is expressed x-fold greater than microRNA B directly from the TCGA miRseq data? Can I do this after normalizing the data? Does it matter if I use RSEM or RPKM values. It seems to me that it should be legitimate in any case since microRNAs are approximately the same length, but maybe I am overlooking something.

For example, I am following a paper published in Nature Communications entitled "Identification of a pan-cancer oncogenic microRNA superfamily anchored by a central core seed motif". The authors download the data and collapse isoform reads to a single read count using the reads. They say they used the reads per million microRNAs mapped, which establishes each microRNA read count as a fraction of the total microRNA population. The authors then do upper quartile normalization which they say is important because a subset of microRNAs (miR-143 in particular) contributes so significantly to the total read count. In the text, the authors appear to use the resulting values to do a direct comparison between microRNAs.

I definitely want the collapsed isoforms, and I think it makes sense to do the normalization. However, I would like to say that a particular microRNA is expressed x-fold higher than another. Can I do this from the collapsed and normalized data?

If this has already been answered, I apologize. I could not find it. Thanks.


I would be very careful in making such a claim from sequencing data. There are well known biases in sequencing due to things like GC composition and hexamer composition. Because of this, it's possible to compare relative levels of some species across treatments but tricky to do so within a given sample or samples. One way around this is to perform qPCR using absolute quantification on a few samples and then compare that to sequencing results. Granted, you're not going to have the same samples as TCGA or exactly identical library prep, but it'll be a reasonable starting point. You could then use that to calibrate how reliable the relative levels might be within the TCGA dataset.

  • $\begingroup$ Thanks Devon. If there were no sequencing biases, do you think this with be a legitimate claim with isoform-collapsed and normalized data. I have qPCR data on cell lines (though I used relative quantification and not absolute quantification), so I wanted to compare clinical data to that. As an aside, I have been comparing the fold-change from microRNA A versus small RNA RNU6B to the fold-change from microRNA B versus small RNA RNU6B to determine fold change of microRNA to microRNA B. I am wondering if your suggestion to use absolute quantification suggests that you don't think this is ok? $\endgroup$ – user14631 Feb 24 '15 at 21:56

Small RNA sequencing is good for a preliminary screen because as with most high throughput experiments, the sample size is less and therefore inter-sample comparisons may not be accurate.

For miRNA expression calculation, I use mirdeep2's quantifier script with a slight modification. The script basically aligns the reads to known pre-miRNAs and finds if they are aligning to the annotated mature miRNA region (some window of partial mapping into non-mature region is allowed and can be set to zero as well). It uses bowtie for alignment. What I generally do is, instead of running bowtie in the -v mode (as set in mirdeep), I run it in the -n alignment mode by making a small modification in the script. The -n mode lets you define a seed region, and number of mismatches in seed and non-seed regions. I set seed length to 10 with 0 seed mismatches and around 2 non-seed mismatches.

To calculate reads per million (RPM) I normalize it with number of unique mappings to genome, with same alignment parameters (except --norc). As per this study an RPM of 100 is biologically relevant.

Can I say that microRNA A is expressed x-fold greater than microRNA B?

Yes, if the read counts are clearly different. EM and other likelihood models calculate a confidence interval for read counts or RP(K)M and if two RNAs do not have overlapping intervals then they can be said to be expressed in different levels. I haven't used likelihood models for miRNAs though, but I guess it is fine because:

  1. The search space is highly reduced.
  2. Search is stringent.
  3. Isoforms are collapsed.

This effectively eliminates the possibility of dubious reads. You can set an RPM difference of 100 to call two miRNAs as differentially expressed, as 100 is the lower limit of biological significance (as previously mentioned). Most relevant miRNAs have comparable GC content; so there may not be such a great sequencing bias (moreover the read itself is quite small).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.