Processing Geuvadis expression data - log of TPM results in sex based separation

[Al-Murphy]

I have been rerunning the processing of the RPKM -> log normalised TPMs for the Geuvadis expression data and I have noted that if the log transformation of the TPM values is not performed, you get mixing between sexes in the batches after regressing out the 10 PCs:

However with the log transformation, there is clear separation:

Did you find the same in your work? Do you think this will affect analysis?

To note, there is no separation in the TPM counts before regressing out the PCs for either the log or no log TPM values:

I think my approach matches yours as my other graphs (variance explained in PCs, correlation with normalised RPKM, mean variance trend of TPM) all match those in your notebooks.

I also tested dropping the number of PCs to regress out and found that lower numbers (6 or less) lead to better mixing in sex but greater separation in the labs/ancestries. This is all based on UMAP representations however which are known to be misleading.