Replicates do not cluster together

Question

[rin]

Hi all.

I am analyzing some proteomics data in cell lines before and after the silencing of certain genes and at different time points. A PCA (see attached) shows no clustering of my samples that come from the same treatment. The samples were analyzed in the same batch, so batch effects cannot be the reason why I am not seeing what I would expect.

Would it be possible to go on with a differential expression analysis given that my replicates are not clustering together? Any suggested analysis that could help me identify if that would be possible?

https://ibb.co/QfjT1bB

Answer 1

[Friederike]

What do the different colors represent in the PCA plots?

Would it be possible to go on with a differential expression analysis given that my replicates are not clustering together?

Sure. There may not be results and, unless you find sources that may explain the variation that you do see, you may not be able to use those sources of noise as covariates, but you won't know that until you do it.

[rin]

The colors represent Condition + Timepoint.

Thank you for your comment. Indeed, I guess I just have to try!

[Kevin Blighe]

Looks like you are using PCAtools (my package)? If you use plotloadings(), you can see which genes are 'driving' the variation along each PC.

Also be wary of using scale = TRUE or scale = FALSE with PCAtools::pca(). I would prefer scale = FALSE.

[rin]

Yes, I am using PCAtools. (Since we are here, I have to say that even if I discovered it pretty recently, your package quickly became my go-to package for PCA. Thank you!)

Thanks for the tip. Would you mind elaborating a bit why scale = FALSE is preferable? Thank you!

[Kevin Blighe]

Hi, thanks for the comment regarding PCAtools. Regarding scaling, there was another recent discussion on this, here: C: Scale and Center [normalized] RNA-seq expression counts for PCA ?

Scaling is neither recommended by Michael Love (DESeq2 developer), although, I cannot find thee post where he mentions this.

In a nutshell, the PCA formula is fundamentally based on variation and covariation; by scaling, we 'disrupt' (break) the natural [true] variation that may exist in our data. A Full-Time statistician would obviously give a more technical answer.