Significance within conditions

Question

[Assa Yeroslaviz]

Hi all,

I am looking for a method to compare the significance of the expression difference of two different transcripts within a specific condition.

with that said, i will try to explain a bit more. We have a RNA-Seq experiment with three conditions. We are comparing each pair of conditions to find significant differentially regulated genes on gene (deseq2,cuffdiff) and transcript level (kallisto,cuffdiff,dexseq).

For some of the transcripts (isoforms of the same gene) we would like not only to know whether or not they are significant DE between the two conditions, but also if transcriptA is significantly higher/lower expressed than transcriptB.

About the data - we have 10 replicates for each of the three conditions.

For the first question I have done done the calculations with the mentioned tools. I am struggling to find a way to calculate the second problem. Is there a method to calculate such significances?

thanks for the help,

Assa

Answer 1

[Devon Ryan]

Comparisons between different transcripts are somewhat problematic, but you might try using the GC (and other) bias correction options in salmon and then put the TPMs through voom and limma (you'll need to log2 transform them). You'll need to munge the data into an appropriate format for the comparison (i.e., remove transcripts and take them on with a new sample label and a row name changed to be what you want to compare against).

[Assa Yeroslaviz]

thanks Devon for the fast reply.

I have already the TPMs from kallisto. Does it make sense to use them, or would it be better to run salmon with its own options? or for that matter any counts ( I can use voom to transform them into log-counts per millions)

by munging the data, do you mean taking the results of each "transcriptA.condition1" and "transcriptB.condition1" and consider them as two separate data sets for the comparison?

[Devon Ryan]

Yes, you'll need to remove transcripts so you have things like major and minor transcript columns. So something like

Treat1.sample1.major | ... | Treat1.sample10.major | Treat1.sample1.minor | ...

You only care about major vs. minor isoform differences. You don't need all of the treatments if you don't care about them, though you can leave them all in if it's useful.

Answer 2

[Assa Yeroslaviz]

Hi Devon,

I am not exactly sure how to proceed. I have ran salmon for the 30 samples I would like to analyze.

I will try to explain better what i am looking for. I have two conditions (treatment1 and treatment2; no control). I have two transcripts of Internest (TIT1a and TIT1b from gene TIT).

I would like to know if the expression changes of TIT1a.treatment1 vs. TIT1a.treatment2 is significantly higher/lower than the expression changes of the comparison TIT1b.treatment1 vs. TIT1b.treatment2.

the normal two-way comparison i have both on the gene-level (DESeq2, cuffdiff) as well as on the transcript level (kallisto, cuffdiff, salmon, dexseq). But these comparisons only gives me the differences for one transcript between the two conditions, not comparing the two transcript changes.

the data for my TITs looks like that

sample   treatment     TIT1    TIT2
1X          treat1  2505.73  373.75
2X          treat1  2687.13 324.901
3X          treat2  2381.98 325.676
4X          treat2  3026.95 533.415
...

by munging the data I will get

  treat1.1X.TIT1|treat1.2X.TIT1|treat2.3X.TIT1|treat2.4X.TIT1|treat1.1X.TIT2|treat1.2X.TIT2|treat2.3X.TIT2...
         2505.73        2687.13        2381.98        3026.95         373.75        324.901        325.676...

I am not sure how to analyse this data, as I have only one row as a possible entry.

Is my way of thinking correct, or do I miss here something?

thanks,

Assa

[Devon Ryan]

If you only have one row then just use glm.nb(). See here for examples. You're interested in the interaction coefficient.