CuffDiff and Significant Splicing events

[Starr_Hazard]

Folks,

We are looking at RNASeq data from rat brain.
The experimental protocol is to compare sham (or S: saline treated) versus experimental (or C: chemical treated).
Several pairs of animals were tested.


When I group all the data from groups 1, 2 & 3 the number of genes with significantly different expression ( ie with "yes" in *.diff from cuffdiff output) can be a few score of genes. However, when I look a the C vs S for a single pair I can see several thousand genes, isoforms splicing variants. Why the huge decrease? I tell the experimentalists that this may be due to the multiple test corrections but they are not convinced.


When I examine pairs separately ( ie group 1 C vs S, group 2 C vs S and group 3 C vs S) I get a number of significant genes and isoforms that show common expression patterns in all groups.
However the number of splicing events shows almost no instances where the same event is found in all three groups. Why is that?


Starr

[endether]

Hi Starr,

Firstly, I don't think it is a result of multiple test correction, because it is only related to your number of tests, which should be the total number of candidate genes/splicing events.

I assume you have multiple groups within each 'test groups' (S and C respectively). It might be the case that the variance among your groups is so big that the S vs C variance is less significant. For example, if there are 3 groups (or individuals?) in S and C, one feature (gene/splicing event ) has following values S1 = 1, S2 = 10, S3 = 5 and C1 = 10, C2=1, C3=2. If you look at S1 vs C1, S2 vs C2, and S2 vs C3, you might draw the false conclusion that it is deferentially expressed when comparing S and C. However, if you take them together and compare (S1, S2, S3) vs (C1, C2, C3), you cannot safely say that they are different because it is totally possible that the difference between S1 and C1 is a result of system variance (noise or individual difference) instead of biological variance (your treatment).

Of course, if you have multiple individuals in each group and you believe there are difference between groups, you might want to stick with the group-wise test.

Best,
Zheng

[glados]

I had this problem as well. When I increased the sample size, the number if sig. genes decreased drastically in cuffdiff. However I got the opposite result in DESeq, much more sig. genes with more replicates.

[DZhang]

Hi Starr,

What you observed supports the notion that technical or biological replicates should be utilized to reduce the false positives in differential expression analysis. However, you need to keep in mind that maybe there are some real differences among these three pairs so when individually compared they have many hits but grouped together they are either canceled out or get buried. What I recommend is to ask the bench scientists 1) are the pairs matched, age-wise, sex-wise, etc.? 2) were there any difference among these three pairs in terms of how the experiments were conducted; 3) you may have to ask them to perform validation experiments (say, qPCR) on a limited scale.

Best regards,
Douglas

https://www.contigexpress.com