Entering edit mode
3.3 years ago
panconchoclo
▴
10
I want to normalize metabolomics 'counts' data by Variance Stabilizing Transformation. In the meantime I already did that by VST function of using DESeq package and this improve in my opinion the clustering between replicates on a PCA plot which make sense for me. However I just wonder if I need actually use some modification on the function to avoid include library size normalization and what is different regarding the old VSN function of Limma package. Maybe I should be use that instead the implemented VST on DESeq? I mean; Is DESEq using a wrapper of VSN Limma for permom the VST or doing something else to normalize by library size? Any comment regarding my approach will be appreciated guys,
Is there any reason why you decided to use the variance stabilising transformation? From my experience, most log2 transform the peak areas (after QC), followed by a further Z-scale transformation. Admittedly, this is not ideal. Metabolomics suffers from a lack of reproducibility, and exhibits high variability, as you are probably aware.
I would post this over at BioC to get a response from Mike Love (the DESeq2 maintainer/developer). Don't forget to post a couple of plots to show how your data are distributed and give some background information on the nature of these data, e.g. if they follow a certain distribution and how replicates compare to each other.
Hi guys thank you for your reply.
Kevin I decided by VST because there are Heteroscedasticity. The behaviour of Means vs SD after and before VST transformation in my data is showed.
In the plot 1 you will see that there are some metabolites with a lot of variance. After VST the differences are slower (see scales of Y axis)
before VST normalization(raw data):
after VST:
If I do the PCA of raw data my replicates are not all of them together. If I do VST before PCA the behavioir in this sense is better. I mean my replicates are closer between them.
I would go with the suggestion of ATpoint and ask this on the Bioconductor forum. Michael Love will provide an answer or comment.
Thank you Kevin I will do that. To avoid repeat exactly the same question there I will dig more in detail on these packages.