Which statistical test: GSEA on duplicated RNA-Seq data

[JesperGrud]

Hi everybody

I'm struggling a bit trying to do GSEA of RNA-Seq data. I've ended up settling with a package known as GAGE (Generally Applicable Gene-set Enrichment). The main reason is that this algorithm is the only one I've been able to find that does not require 10+ biological replicates.

The algorithms employed by GAGE is targetted toward microarray data and as such there are some adjustments that are necessary prior to analysis. Basically, I need to do a transformation to make the data homoscedastic and I need to do length bias correction.

I'm then asking GAGE to do a paired comparison between treatment and control and for each pair this will give me some enrichment score. What I'm struggling with is what sort of method I should ask GAGE to use for statistical testing, as I have only two replicates.

The options are:

1. Two-sample T-test (either parametric or rank-based)
2. One-sample z-test
3. K-S test

I think the correct test to use is the rank-based two sample t-test, but it would be nice if someone with more statistical knowledge could comment on my workflow.

[Xi Wang]

Hi, If you haven't solved this out, I'd like to suggest you try our newly developed R package SeqGSEA, which is available at http://bioconductor.org/packages/rel...l/SeqGSEA.html.

It can integrate differential expression and differential splicing together for GSEA. By using negative binomial to model read counts, SeqGSEA can correctly capture technical and biological variance in RNA-Seq data. It doesn't require a large number of biological replicates, but I'd like to know how many you have.

Cheers
Xi

[JesperGrud]

Hi.

In the end i used a rank-based Wilcoxon-test. This might yield some false negative in the sense that p-value scoring is too conservative, if the data are in fact normally distributed. However, since im only interested in the say top 5 pathways from a biological point of view, it does not matter that much.

Our current setup is to have just 2 replicates from cell culture experiments. Our tests show that 2 replicates gives us sufficient data to detect most differential expression using DESeq.

I havnt read the details concerning your package, but the questions that spring to mind is if it requires paired end reads? and if the differential splicing analysis can be omitted?

[Xi Wang]

Hi.

In the end i used a rank-based Wilcoxon-test. This might yield some false negative in the sense that p-value scoring is too conservative, if the data are in fact normally distributed. However, since im only interested in the say top 5 pathways from a biological point of view, it does not matter that much.

Our current setup is to have just 2 replicates from cell culture experiments. Our tests show that 2 replicates gives us sufficient data to detect most differential expression using DESeq.

I havnt read the details concerning your package, but the questions that spring to mind is if it requires paired end reads? and if the differential splicing analysis can be omitted?

Thanks for the info.

I am afraid 2 replicates are too few for SeqGSEA, as its based on sample label permutation for statistical significance.

Regarding your questions, SeqGSEA doesn't require PE reads. It only takes read-count data. SeqGSEA can work with DE-only GSEA.

Hope this info would help with your future data analysis and other researchers. Cheers.