RNA-seq and normalization numbers

[marcora]

Hi guys,
I have trancriptome data from Illumina and am using CLC Genomic workbench for data analysis. I dont know or not familiar with other programs for transcriptome analysis. the data are from 1 sample of control cells and 1 sample of treated cells (no replicate for each sample) and I am looking for differently express genes.

If I was a reviewer I would doubt any conclusion coming from an experiment with no biological replicates. Anyhow, DESeq allows for such design, you may wanna consider it. I am not familiar with CLC Genomic workbench.

[sikidiri]

Method to categorize mRNA-seq data based upon expression value

Hello All,
I have a pre processed mRNA-seq data for hg19 genome, in which for each gene the RPKM value is calculated. The value ranges from 99960 to zero. I have just one sample.
I want to categorise these genes into highly, medium and weakly expressed genes.
What could be the best way to do it?
Your suggestions would be highly appreciated.
Thanks a lot.

[steven]

Hello All,
I have a pre processed mRNA-seq data for hg19 genome, in which for each gene the RPKM value is calculated. The value ranges from 99960 to zero. I have just one sample.
I want to categorise these genes into highly, medium and weakly expressed genes.
What could be the best way to do it?
Your suggestions would be highly appreciated.
Thanks a lot.

As RPKM is a normalized expression measurement so you can in theory directly compare values between genes within the same sample -keeping in mind a couple of reported biases like the size of the gene, GC content, etc.

I would first sort the values and use percentiles ("tiers") to define categories with a similar population and inspect the threshold values.
You may also want to consider absolute thresholds (like "RPKM<1", "1<RPKM<10" and "10<RPKM") but I do not know if there are "standards" for such values and I actually doubt that it is in practice reasonable to use values obtained from different protocols/conditions/software/etc..

[sikidiri]

Hello Steven,
Thanks for your answer. But how to decide about the threshold to make these categories based upon expression is my main problem. Do you think any statistical tests would help me. Any paper/example would help me understand this better.
Thanks again.

[dpryan]

It's been a while since I've done it, but if you google "cluster optimal group number" you can find methods for gap calculations and other things to find an optimal cluster number. I recall there being R packages for a lot of this stuff, such as the cluster package.

[carmeyeii]

Thats correct- the procedure RCJ suggests will give you an estimate of the actual tag count for each transcript.

Is this to say that if one sums up the (FPKM * length in kb * reads mapped in millions ) of each transcript in a gene, one would obtain the total *estimated* read count for that gene?

But this has to be done individually for each transcript and then grouped into a gene, right?


Carmen

[lpachter]

If I don't care about isoforms or think that my coverage is too low to distinguish isoforms anyway, I expect to get optimal power by simply summing everything up.

Please see Figure 1 of http://www.nature.com/nbt/journal/va...0.html#/figure

Cuffdiff is, as I understand it, designed to deal with such issues, while our approach ignores them. I expect that DESeq, in compensation for being unsuitable to detect differences in isoform proportions as in your example, achieves much better detection power for differences in total expression (per gene, summing over isoforms), especially at very low counts.

Please see Figure 3 of
http://www.nature.com/nbt/journal/va...0.html#/figure

As I am not clear on how biological noise is taken into account by cuffdiff I cannot be fully sure whether this expectation will hold (and I'm quite curious to learn more about cuffdiff once your paper is out).

Please see
http://www.nature.com/nbt/journal/va...0.html#/figure

[jparsons]

Please see Figure 1 of http://www.nature.com/nbt/journal/va...0.html#/figure

It's important to note the limitations of raw-count methods, but has anyone done anything to check to see if any of the isoform-detection algorithms can actually discriminate between isoforms well enough to assign those counts properly? I've seen simulated data that showed RSEM incapable of reproducing 'truth' half the time with even simple 2-isoform mixes.

Cufflinks' model in figure 2 has three times more counts than figure 1 and doesn't differentiate anywhere near as cleanly between isoforms. Surely maximum-likelihood count assignment can be incorrect, too, given ambiguous reads? Looking at the supplementals, however, I'm inclined to accept that it may be incorrect less often than raw counts when dealing with real data.