[BioC] edgeR: Using ratios (translational efficiencies) as input

[Gordon K Smyth]

Dear Alvaro,

Well, honestly I don't know.  My naive and literal reading of the original 
email was that the interest was in "differential efficiencies between 
groups".  In your example, the ratio of ribosomal-bound RNA to normal mRNA 
is identical for the two groups, hence my naive interpretation is that 
there no evidence for "differential efficiencies" between life stages 1 
and 2.

Now that I see your interpretation of the data, I see that one could test 
for "differential efficiency" simply by using an interaction test in 
edgeR.

However I have no experience of this type of analysis, and I don't know 
what is scientifically sensible.  Making good plots is always a good way 
to go, but send suggestions to the original poster.  It's his problem, not 
mine!

Best wishes
Gordon

--------------- original message --------------
[BioC] edgeR: Using ratios (translational efficiencies) as input
Alvaro J. Gonzlez alvaro.gonzalez4 at gmail.com
Mon Apr 29 15:34:47 CEST 2013

But Gordon,

Isn't it the case that if you feed logs of ratios into limma you're 
automatically losing the statistical significance of those ratios, as well 
as the absolute expression in each condition, which can be relevant?

For instance, define "t" as one of Gowthaman's transcripts. As far as I 
understand, he has four RNAseq libraries measuring the activity of this 
transcript:

1) Transcripts from normal mRNA:
     1.1) Life stage 1, his transcript gets t_1.1 = 4 reads.
     1.2) Life stage 2, his transcript gets t_1.2 = 2 reads.
2) Transcripts from ribosome-bound RNAs:
     2.1) Life stage 1, his transcript gets t_2.1 = 100 reads.
     2.2) Life stage 2, his transcript gets t_2.2 = 50 reads.

Let's say edgeR being applied to the two 1) conditions produces:

log2(t_1.1/t_1.2) = log2(4/2) = 1 with adjP = 0.5, meaning, it seems like
the transcript was differentially overexpressed in life stage 1, but with
no statistical significance, so we're not really sure.

Then you do the same with the two 2) conditions:

log2(t_2.1/t_2.2) = log2(100/50) = 1 with adjP < 0.01, so you really
believe the transcript was overexpressed in life stage 1.

Now you feed those two logFCs into limma (1 and 1), and of course, you get 
nothing out, in terms of differential behavior. But the reality is that 
there was a huge change between normal and ribosomal RNAs which was 
diluted by the use of the ratios.

What do you think?

My suggestion, just to start, would be to produce a scatter plot of
logFC(normal RNA) vs logFC(ribosomal RNA), and to encode adjP values in
both axes: say for instance by using colors in the x-axis (red is
significant, green is not), and using dot shapes in the y-axis (star is
significant, dot is not).

This plot should show you those transcripts in which interesting stuff is
going on.

Regards,

- Alvaro.

> Dear Gowthaman,
>
> I'm not quite sure what translational efficiencies are.  Do you have a 
> different efficiency value for each gene and each RNA sample?  If you 
> do, why not take logs of the ratios (offsetting counts by 1/2 or 1 to 
> avoid zeros) and feed them into limma?
>
> Best wishes
> Gordon
>
>>
>> Hi Everyone,
>> I have been using edgeR for the last couple years with great success.
>> Thanks very much. Now I have slightly unconventional dataset to try. We
>> have two groups to compare (life stages) each with three replicates. 
>> But,
>> for each sample in each group, we made two different RNAseq libraries.
>> 1)  one from fragmented mRNA (classical RNAseq) and
>> 2)  another from Ribosome-bound RNA fragments. This library would
>> indicate how much of the RNA is actively being translated.
>>
>> I have used edgeR to analyse data from each of this separately (data 
>> from classical RNAseq or Ribosome-bound). So this let us study the 
>> differentially transcribed genes or differentially translated genes. 
>> And got really nice results.
>>
>> The next step is to compare the translational efficiencies between 
>> them. In each sample the ratio between read counts of Ribosome bound 
>> mRNAs and fragmented mRNA would give us the translational efficience of 
>> that gene. We can generate these efficiences (ratios) for each of the 
>> three replicates in each group. Can I feed this data to edgeR to find 
>> out which genes have 'differential efficiencies' between groups?
>>
>> I understand, edgeR insists on NOT normalizing the read counts and all 
>> the further statistics depends on the total library size count. By, 
>> using ratios, i completely throw edgeR off. But, i am not sure what is 
>> the best alternate to this?
>>
>> Any ideas?
>>
>> Much thanks in advance,
>> Gowthaman

______________________________________________________________________
The information in this email is confidential and intend...{{dropped:4}}