Entering edit mode
2.2 years ago
melatoninixo
▴
10
Hi All, I am trying to compare 2 transcriptomic profiles of samples treated with different drugs, in order to see whether they have the same influence on the cell's transcriptomes after treatment. Is a correlation metric like spearman rank correlation suitable for comparing between 2 transcriptomic profiles of 2 samples with different experimental conditions quantitatively? Is it possible to use the log2FC obtained from DEG analysis for it? I have tried sourcing for a paper documenting such a comparison but to no avail.
Any input is greatly appreciated. Thank you!
With only two things to compare, the problem will be that you have no way to evaluate the resulting number. If the correlation comes back as 0.68, is that high? the drugs are having the same effect on both samples? If it were 0.56 would you be able to say that it's low and the drugs are not having the same effect? If you have several conditions, some of which are known to have similar or different effects, then you might be able to "calibrate" your correlations to establish relatedness, but even then comparison of entire transcriptomic profiles can be a tough sell. Every technology has different amounts of noise across the board, so comparing two microarray data sets will be different than comparing two RNA-Seq data sets in terms of expectation around correlation numbers.
On the other hand, you might be able to ask more specific questions, such as using a Venn diagram to see if each experiment is showing enrichment or depletion of similar gene sets. Or if you know anything about the biology you could use something like the camera method (Wu & Smith, 2012), implemented in limma and edgeR through the geneSetTest() function, to ask if sets of genes are similarly differentially expressed in each data set. This would allow you to argue that the drugs are inducing the same biological response if your gene sets show similar behavior between transcriptomes. Any functional analysis you do on one data set (GO, KEGG, GSEA, etc.) should show some similarity in the other data set if the response is the same. Other than that, if you have other data sets, you could potentially use a heat map to illustrate or query how your profiles behave relative to other sorts of conditions (i.e. do the cluster together), though it sounds like you don't have other data sets.
Thank you for your informative response! I have seen such a comparison in a paper here: https://www.cell.com/action/showPdf?pii=S0092-8674%2800%2900015-5 , which linked a method comparing the log10(expression ratio) between deletion mutants. I am not sure what the expression ratio represents (what normalization is done etc).. and what correlation metric they used.
I will definitely try the geneSetTest() function! Thank you!
Yes, thanks for the reference, that Hughes paper was quite impressive at the time and is based on microarray data. It's a perfect example of what I was saying above - if you have hundreds of experiments to compare, then correlation can become meaningful, especially when done in a controlled manner on a single platform. However, microarrays are generally noisy enough that any particular batch of experiments at the time could have correlations that fluctuate quite a bit. The same is true for RNA-Seq experiments.The correlation can fluctuate quite a bit in response to all kinds of factors having little to do with the conditions being compared.
ohh, would you say that the statistical approach I have mentioned is something that is feasible? or should I consider another correlation metric and another normalization approach?