Should PCA be recaculated for each sub-sample when assessing clustering robustness?

[microbialman]

Hi,

More of a theoretical question than a code issue (so apologies if this isn't the appropriate place for it).

Looking at the function used for subsampling in chooseR, it appears that when calculating clusters on subsets of cells that the PCA reduction calculated from the total set of cells is used. Therefore, some information from cells not in the sub cluster will be influencing the overall clustering (as cells outside of the subset will still determine distances in PCA space).

I'm not sure if it would make a large difference to inter-parameter comparisons but my intuition is that it would indicate clustering is more stable between subsamples than it would be if PCAs were recalculated on each subset?

Hope that makes sense, and that I haven't misunderstood the pipeline!

Cheers,
Matt