Hi Neha,

The eigenvalues associated with each eigenvectors can give you the percentage. So suppose you have 10 eigenvectors, the eigenvalue of first divided by the sum of all eigenvalues multiplied by 100 will give you the percent contribution of the first eigen vector towards the motion of the macromolecule. For cumulative eigenvalues, just calculate the cumulative sum of eigenvalues such that the total sum is 100%.

Hope that helps!

Soumya

Hi Neha,

The eigenvalues associated with each eigenvectors can give you the percentage. So suppose you have 10 eigenvectors, the eigenvalue of first divided by the sum of all eigenvalues multiplied by 100 will give you the percent contribution of the first eigen vector towards the motion of the macromolecule. For cumulative eigenvalues, just calculate the cumulative sum of eigenvalues such that the total sum is 100%.

Hope that helps!

Soumya

As an addendum to Soumyas answer (which is perfectly right: just normalize the eigenvalues by the sum and you have the percentage):

I'm still not sure what the physical meaning of the eigenvalues actually is. People interpret it as "what part of the total fluctuations is covered by some eigenvector" but to me it is not clear whether this might be an over-interpretation. It is true in terms of the variance, but what physical meaning does a high variance have (besides just pure wiggling)?

And I doubt that the "motion" described by PCA is a physical motion. It is perfectly fine to discriminate states and to state how the states differ from another. Yet, the "motion along a PC" is not necessarily related to the physical motion the system performs when crossing some barriers.

Two more points:

This can be applied to both gromcas xtc files as plain ASCII files containing e.g. dihedral angles or distances and renders PCA easier in my opinion

Dear Soumya and Matthias, I absolutely agree with you. Thank you for pointing out link to FastPCA. Cheers