I've been handed RNA-seq data with a lot of associated covariates. This data has been put through the DESeq2 package and as a result normalised the data. One of the transcripts of interest still has non-normal distribution for two genotypes. This is because many of the normalised counts are now clustered around a value of 4.

QQplot of transcript expression against a genotype

This is making it tricky to perform statistical tests because in order to perform parametric tests (in this case I want to perform a ANCOVA or multiple linear regression to include the covariates) the data must be normally distributed.

I also don't want to use a non-parametric test because I can't include all the covariates.

Should I just perform parametric tests ignoring this violation or do I have to just perform non-parametric on all my data for a fair comparrison? Or is there another alternative option I am missing?

Thanks in advance

  • $\begingroup$ I personally find it hard to follow here. Please clarify: What is "a value of 4"? Is this referring to read counts? What sequencing platform is this? Is this standard RNA-seq or some kind of custom protocol? What do these plots show? Is this Z-score, and if so based on which values? You asked a question on RNA-seq before which indicated that you are new to these kinds of data. It was suggested there to follow standard pipelines rather than coming up with costom solutions. Did you do that? $\endgroup$ – ATpoint Nov 24 '20 at 16:01
  • $\begingroup$ The values are normalised counts. Platform is Illumina. This data was put through the DESeq2 package. $\endgroup$ – Connor Nov 25 '20 at 6:43
  • $\begingroup$ Your values of 4 look like they might have been vst transformed. This has a tendency to make non-expressed genes have a value around 4. The vst is in log-space, which you can then see in your plots as they aren't linearly distributed $\endgroup$ – SPPearce Nov 26 '20 at 8:22
  • $\begingroup$ why don't you use the glm model in DESeq2? $\endgroup$ – StupidWolf Nov 26 '20 at 16:11

TLDR- just do the analysis in DESeq2 (it can handle multiple covariates)

If I've interpreted your question correctly there seem to be a few confusions here.

First off, for a multiple linear regression aka ANCOVA, one of the assumptions are the residuals are normally distributed - however this is almost never an issue in practice. The distribution of the data is not so important, you make the assumption that the data has come from a population which is continuous and normally distributed that's why you've chosen to do an ANCOVA. But, RNAseq produces count data - not continuous and bounded at zero. You need to be looking at using a Poisson GLM, or negative binomial GLM as done by DESeq2 (lookup overdispersion)

The normalisation done by DESeq is to account for different library sizes,it doesn't transform the data to make it conform to a normal distribution. You'd need to do some form of data transformation to do that, but don't it's unnecessary.

I'd suggest reading the DESeq vignettes and following their standard pipeline.

Hope that is of some help.

  • $\begingroup$ Will the poisson GLM or Negative GLM account for the covariates as well? $\endgroup$ – Connor Nov 26 '20 at 6:36
  • $\begingroup$ You can use the negative binomial GLM with covariates yes $\endgroup$ – SPPearce Nov 26 '20 at 8:22

Your assessment is perfectly correct. The easiest approach is to perform a transformation to normalise your data and then you are good to go. DeSeq2 has a standard transformation (I think it is called deseq2) and you can check your results via a Q-Q plot and you are looking for an approximately linear relationship.

There are a multitude of transformations, but the transformation proceedure has been heavily investigated and standardised in RNAseq

If normalisation by deseq2 has failed (Q-Q) you simply examine an alternative transformation. However, do not transformation a transformation. If you cannot normalise the data then you are stuck with non-parametric stats in my personal opinion.

However, mixed models can work if you trap the underlying distribution and in a regression will result in a very low residual.

  • $\begingroup$ Hi Michael, this data has already been normalised by deseq2 so that's why im unsure how to proceed now to analyse it as the qqplots are not normal :( $\endgroup$ – Connor Nov 26 '20 at 6:33
  • $\begingroup$ Commented on above @Connor $\endgroup$ – M__ Nov 26 '20 at 7:45

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.