PCA / MDS analysis useful if only small number of genes (e.g. 1-20) diff. expressed?

[Alex234]

All in the title really - if a condition only resulted in a change of expression of only a small number of genes (e.g. 1-10), would you expect to see that on a PCA / MDS plot?


Thanks

Alex

[dpryan]

That rather depends on the rest of the data. It's usually a waste of time to use PCA plots for anything other than finding outliers/similar QC. If your DE genes have a large fold-change and the rest of the data has low variability, then samples will probably cluster by condition on the PCA plot. Otherwise, you're unlikely to see clustering by condition. A lot of people over-interpret PCA plots simply because they don't know how they work (there's currently a discussion about this on the Bioconductor email list, in fact).

[Alex234]

Thanks dpryan - when you say 'large fold-change', what sort of magnitude do you mean?

My data has a high degree of variability, I think because some of it comes from different strains, and because it is from stem cell lines, which are supposed to have a great deal of transcriptional heterogeneity.

If PCA / MDS are only relevant QC in some cases, which would you say are the QCs that are important for every experiment?


Thanks

Alex

[dpryan]

Thanks dpryan - when you say 'large fold-change', what sort of magnitude do you mean?

There's no a priori answer to that, that's sort of the thing with PCA.

My data has a high degree of variability, I think because some of it comes from different strains, and because it is from stem cell lines, which are supposed to have a great deal of transcriptional heterogeneity.

Yeah, strains alone will often separate things via PCA.

If PCA / MDS are only relevant QC in some cases, which would you say are the QCs that are important for every experiment?

I also find hierarchical clustering useful, though it's generally similar to the PCA plot. In addition, the normal plots from FastQC and various plots depending on how and what kind of statistics you do downstream (e.g., an MA plot, dispersion vs. mean, etc.).