overdispersion and biological replicates

[shilez]

Hello,

I really appreciate the work of DESeq and EdgeR for taking the overdispersion into account for biological replicates. However assuming when analyzing tumor samples in a few cancer subgroups, do samples in the same subgroup count as "biological replicates" too? Particularly does the overdispersion model in DESeq statistically sound for a potential more variable samples within a subgroup?

[Simon Anders]

However assuming when analyzing tumor samples in a few cancer subgroups, do samples in the same subgroup count as "biological replicates" too?

Of course.

Particularly does the overdispersion model in DESeq statistically sound for a potential more variable samples within a subgroup?

Yes. (Please have a look at the development version, as we made quite some changes.)

This is not to say, however, that you are likely to find much if you compare cancer subgroups and have only a few samples per group. Precisely because the variation is so large you need very many samples to have enough power to find significant differences. Furthermore, pairing samples with normal tissue from the same patient (and using GLMs for the analysis) reduces the variability.

See also this thread: http://seqanswers.com/forums/archive...p/t-11133.html

[shilez]

Thanks a lot Simon!
My sample size is 10/subgroup without the normal tissue sequenced from the same individual.

My primary goal is gene expression analysis and transcript expression analysis.
I haven't decided on SNP detection, and fusion gene analysis. Do you think 50bp single read would not be powerful enough to get meaningful results?

[Simon Anders]

I also realized after posting that the thread I mention talks about SNPs, not gene expression, but the problems are similar.

50bp single-end is definitely sufficient to get good expression values.

I don't know that much about cancer transcriptomics, but 10 samples sounds like a quite small number to me, especially without matched normal tissue, unless the expression differences between your sub-types are quite pronounced. Better have a good look at the literature (also look at microarray studies, the power issues should be quite similar) to see what sample sizes gave good results in the past.