Assessing Validity of PCA results
Issue #949
new
I was looking through the archive and one of the inquiries featured some PCA projections which looked strange (Issue #784). I’m wondering if there are any functionalities in bio3d that allow you to gauge whether or not you can trust the modes that are extracted from PCA, or to test whether you have an ensemble that represents sufficient sampling as to not skew the PCA results. Possibly an RMSIP score calculation? I’m wondering because I got some weird PCA results myself (see attached) and I’m trying to figure out if it’s my trajectory or not.
Hi,
To my knowledge, no such function can detect unusual PCA results. The best way, I think, is still to inspect manually. For example, I would generate collective motions along suspicious PCs and see what happens:
mktrj(pc, pc=1, file="pc1.pdb")
where, the first argument ‘pc’ is the returned value of calling
pca()
.You can then open the ‘pc1.pdb’ file using VMD and play the movie. If a portion of the structure moves a lot, that may explain the “drifting” in the conformer plot. In contrast, if the entire structure moves wildly, it is most likely that conformations are not superimposed properly.
Hope it may help.