PCA trasformation on Z scored data

Question

[camillab.]

HI,

I have converted different bulk-RNAseq into Z scores and I want to compare them. Usually,for the PCA I log transform +1 pseudo count (log2 n + 1) and then I set center=TRUE and scale=FALSE. If I want to compute PCA for the genes shared across the different datasets (that have been already converted into Z-scores) do I have to log transform and set scale=F? Since the Zscored are standardized scores, I would not log transform and keep the scale=F. Is it correct?

thank you

Camilla

[shoujun.gu]

It depends on your downstream analysis. And you should think twice whether convert to z score respectively or on combined datasets.

[Kevin Blighe]

Indeed.

Answer 1

[Kevin Blighe]

Hey again,

If your data is already on Z-scale, I would set center = FALSE and scale = FALSE, and I would not apply any additional log transformation.

As you are looking at data across datasets, even on Z-scale there will likely be some batch effect(s).

Note that, in, e.g., metabolomics studies, we typically log[e] and then Z-scale the normalised data.

Kevin

[camillab.]

thank you! so for the scale, I guess my idea/hypothesis is correct but why would you set center= FALSE ? centering does not make the variable means centered at zero and how this would affect/ interfere the PCA if performed on Zscores?

[Kevin Blighe]

I had assumed that the data is already centered from the initial transformation. You could check via mean(), hist(), and summary() (?)