Question: GSVA and Limma for differential expression of gene sets
gravatar for SB
14 months ago by
United States
SB0 wrote:

Dear Bioconductors,  

I recently used GSVA to get enrichment scores for three gene sets and four sample groups. From the enrichment scores, I am interested in determining if they are differentially expressed across several sample groups. I intend to do many sets of comparisons and thus I was planning to create separate contrast matrices for each set of comparisons. However, I am wondering if it is incorrect to break up these comparisons into separate matrices and if I instead should combine all comparisons for the study into one contrast matrix. Further, though I have used the Benjamini-Hochberg procedure for multiple comparisons, I think I have what is considered planned comparisons (i.e. only a few sensible comparisons which were decided before looking at the data). As such, is it incorrect to use the BH method or would you recommend using Bonferroni corretion? 



limma gsva • 395 views
ADD COMMENTlink modified 7 months ago • written 14 months ago by SB0

Does the gsva_es matrix that you enter to limma have only 3 rows?

ADD REPLYlink written 14 months ago by Gordon Smyth39k

Yes, three rows since I am only interested in examining enrichment of 3 gene sets in the above samples.​

ADD REPLYlink modified 7 months ago • written 14 months ago by SB0

With only three rows, there's hardly any empirical Bayes moderation for limma to do. So the results from limma will be nearly the same as if you did a linear model or anova analysis for each row separately. There's no possible harm in using limma however if you find it convenient.

ADD REPLYlink modified 14 months ago • written 14 months ago by Gordon Smyth39k
Answer: GSVA & Limma for DE of Gene Sets using NanoString Data
gravatar for Gordon Smyth
14 months ago by
Gordon Smyth39k
Walter and Eliza Hall Institute of Medical Research, Melbourne, Australia
Gordon Smyth39k wrote:

I have no experience with entering GSVA scores into limma, nor with Nanostring in general. I can make a couple of comments though.

It's fine to break up the contrasts into groups that are tested separately.

In limma, p-value adjustments are usually used to adjust for multiple testing over genes (or pathways in your case) rather than over contrasts. If you want to adjust for multiple contrasts as well as multiple pathways, then use method="global" in decideTests().

Using Bonferroni is never a good idea as the "holm" method does the same thing and is less conservative, see help("p.adjust"). I'm not clear why you don't want to use the BH method but, if you want to use family-wise type I error rate control instead of FDRs, then specify adjust.method="holm" to decideTests().

If you really only make a few pre-planned tests, then you might not need multiple testing adjustment at all -- in that case you might simply use the p-values.


ADD COMMENTlink modified 14 months ago • written 14 months ago by Gordon Smyth39k
Answer: GSVA & Limma for DE of Gene Sets using NanoString Data
gravatar for Robert Castelo
14 months ago by
Robert Castelo2.3k
Spain/Barcelona/Universitat Pompeu Fabra
Robert Castelo2.3k wrote:

hi, regarding the GSVA transformation, the parameters you have used seem fine to me. The rest of the question seems to be more related to general experimental design. I would try though to have both analysis, at feature level and gene-set level and try to have a sense whether you are using the appropriate gene set definitions. The fact that you are only using 3 gene sets might also require attention since the limma pipeline informs the differential expression statistics with the variability of all the features, which in this case are only three.



ADD COMMENTlink written 14 months ago by Robert Castelo2.3k
Please log in to add an answer.


Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
Powered by Biostar version 16.09
Traffic: 202 users visited in the last hour