Can we show by manually analysis why Cuffdiff and edgeR gives different results?

[sindrle]

So my supervisor asked me to do the following:

1) Theoretically - Whats the difference in the "algorithms"? He wants an explanation all the way from read level.

2) In practice - How to show why edgeR gives a positive signal for gen X, while Cuffdiff (or DESeq2) does not. He want raw data from Tophat2 and manual analysis in a table..

I dont quite know how to solve this..

Anyone have some tips?

[mastal]

You can start with 1) Theoretical.

Read the manual/publications/vignettes for edgeR and Cuffdiff, and they will explain what statistical methods each uses.

[sindrle]

You can start with 1) Theoretical.

Read the manual/publications/vignettes for edgeR and Cuffdiff, and they will explain what statistical methods each uses.

Y, thats the easiest...

What about 2)?

[sphil]

You should be able to answer 2) after identifying differences using 1)... (imho)

[keithforest]

my approach

The analysis that convinced my lab to switch from cuffdiff to edgeR/DESeq (a few years ago, so the results may be different now with updated tools):

1. Do differential analysis using cuffdiff and edgeR
2. Identify a set of loci where there is a strong disagreement in differential expression between the two approaches.
3. Use a genome browser to inspect the read mapping at loci where there is a strong disagreement

Using this approach, you could show your supervisor these cases of disagreement, to illustrate how the tools are working.

In my experience , most of the disagreement was due to cuffdiff false positives, where small changes in read mapping lead to large changes in isoform assignment.

[dpryan]

You should be able to answer 2) after identifying differences using 1)... (imho)

I guess the difficulty is how one includes the information from all of the other genes in a table, since there's information sharing...

[sindrle]

I guess the difficulty is how one includes the information from all of the other genes in a table, since there's information sharing...

Yes. The problem exactly.

Thanks for input all!

[sindrle]

The analysis that convinced my lab to switch from cuffdiff to edgeR/DESeq (a few years ago, so the results may be different now with updated tools):

1. Do differential analysis using cuffdiff and edgeR
2. Identify a set of loci where there is a strong disagreement in differential expression between the two approaches.
3. Use a genome browser to inspect the read mapping at loci where there is a strong disagreement

Using this approach, you could show your supervisor these cases of disagreement, to illustrate how the tools are working.

In my experience , most of the disagreement was due to cuffdiff false positives, where small changes in read mapping lead to large changes in isoform assignment.

Thats a good one! Ill do that.

[sindrle]

Hi again!
Heres some conclusions and questions after the first glance:

1. Genes called by all; edgeR and DESeq2 and Cuffdiff, are really consistent between the algorithms. Almost identical fold change etc. So whats so special about these genes?

2. edgeR and Cuffdiff share more genes together, but the variation in fold change is up to 45%, they do however agree on up- or down-regulated. But why does not DESeq2 call these genes?

3. edgeR and DESeq2 call genes thats not called by Cuffdiff, how can I find out why? How can I visualise the isoforms, if thats what makes Cuffdiff calculate differently?

4. DEseq2 and Cuffdiff does NOT share any genes which are not also found by edgeR! It makes me wonder why DESeq2 is more conservative..

Finaly,

1. Unique called genes by Cuffdiff OR edgeR OR DESEq2; Why are these genes only called by one algorithm?