We designed an experiment to explore the potential role of carbon dioxide on algae physiology using RNA-Seq. We analyse the differential gene expression using DESeq2 but now we are interested into analyse a few genes in detail.

My PI say to make the statistics over the ration response (expression_treatment/expression_control - 1) and my question is the following: Can I use TPM to calculate this ratio and make following statistics? I read in the documentation of DESeq2 that TPM we can't use this normalisation method to differential expression and my my other question is Why shouldn't?

  • $\begingroup$ What would in your view be the advantage of this custom solution over using the statistically valid fold changes from DESeq2, at best even the shrunken fold changes from the lfcShrink function? I personally see none other that you create uncertainty as plain FCs without proper statistics are not reliable. $\endgroup$ – ATpoint Mar 16 at 12:10
  • $\begingroup$ cause my advisor ask to re-do the statistical analysis using the ratio response (rr) so i need the rr for each sample. I think this doesnt make sense cause the result will be or should be the same that using DESeq2 $\endgroup$ – Manuel Sánchez Mendoza Mar 16 at 16:47
  • $\begingroup$ I googled "ratio response rna-seq", no hits, you may suggest to your advisor to simply not do that maybe? $\endgroup$ – ATpoint Mar 16 at 18:19
  • 1
    $\begingroup$ Why not just approximately compute this from the shrunken log fold-change from DESeq2? It's (2^lfc)-1. $\endgroup$ – Devon Ryan Mar 16 at 20:49
  • $\begingroup$ Finally I'm using the fold-change calculated by DESeq2 and I'm not sure if I have transformed the data correctly. I took the two columns log2FoldChange and lfcSE (both are log-transformed variables) so I went back to the non-transformed variable as follow: FC=2^log2FoldChange and SE=lfcSE*2^log2FoldChange. I saw this approximation in this post. IS that correct? I have obtained a very broad range of error $\endgroup$ – Manuel Sánchez Mendoza Mar 22 at 17:40

This is a common question, and the answer is while you can calculate this, it's not statistically robust. You're likely to arrive at false conclusions.

Why is this? Because while TPM is something you can calculate from a single sample, your question involves multiple replicates and those replicates vary between each other. Counts in units of TPM are missing information about the variation between samples. This extra information affects fold change estimates, and is why differential analysis of sequencing data is complicated.

See this answer to a related question, and this often-cited blog post for more information.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.