Optimal size of Fourier spectrum and aliasing

Submitted by superdex75 on Fri, 2016-04-01 14:41

Forums

ctffind & ctftilt

Optimal size of Fourier spectrum and aliasing

Hi Alexis,

in CTFFIND4 (Version - 4.0.17) what is the optimal size for the Fourier spectrum? I did 5 runs:

run01 fitting 30 - 4.5 unbinned 0.855 A/pixel 512 Thon rings with good fit up to : 4.9 Angstroms
run02 fitting 30 - 4.5 unbinned 0.855 A/pixel 1024 Thon rings with good fit up to : 5.4 Angstroms
run03 fitting 30 - 4.5 2 times binned 1.710 A/pixel 256 Thon rings with good fit up to : 5.0 Angstroms
run04 fitting 30 - 4.5 2 times binned 1.710 A/pixel 512 Thon rings with good fit up to : 5.4 Angstroms
run05 fitting 30 - 4.5 2 times binned 1.710 A/pixel 1024 Thon rings with good fit up to : 16.8 Angstroms

It looks to me that if the Fourier sampling is too small there is aliasing at high frequencies (e.g. run01 and run03), but with oversampling as in run05 the reported correlation is worse and the green experimental peaks took strange at low resolution. With aliasing it appears that at high frequency the maxima of the green and orange curve match, is this an artifact?.

Is there a formula to calucalte the frequency where alisaing ocurrs given the microscipe parameters, defocus and pixel size?

The obtained defocus values are pretty much the same, but I was just curious to know.

All the best, Simon

Axel Fri, 2016-04-01 15:28

Hi Simon,

I noticed the same thing, and started doing multiple runs of the fitting and using the best value as determined by the Thon rings.
I find that the failure rate of fitting is higher when I use large boz sizes with CTFFIND4, but that this also most often gives the best value.

It may be that there is not an "optimal" value for all micrographs, and thus an ad hoc approach is best.

I would be interested to hear what you all think about this.

Axel

Alexis Fri, 2016-04-01 17:07

Hi Simon,

I have also noticed cases where the defocus is large enough that aliasing seems to impacts the quality of the fit. Therefore, I typically run ctffind on resampled micrographs, with a pixel size of 1.5 to 2 A. I usually use a box size of 512, though I think if the defocus goes much beyond 2.5 microns, I would probably consider going to a bigger box.

I don't understand why run05 gave apparently bad results. It could be that the background subtraction didn't work well in this case. I must admit I haven't spent much time testing ctffind with a box size of 1024. Sometimes, the resolution of the fit is mis-estimated for artifactual reasons.

I do not have an easy analytical way of computing the frequency at which aliasing is significant. In the next version of ctffind4, I implemented a numerical way of estimating this and this next version will output a warning to the user when aliasing is "detected" within the fitting range.

You could look at Pawel Penczek's paper on the topic:
Penczek, P. a et al., 2014. CTER-rapid estimation of CTF parameters with error assessment. Ultramicroscopy, 140, pp.9–19. Available at: http://www.ncbi.nlm.nih.gov/pubmed/24562077

I'm also happy to share a preview version of ctffind 4.1 in private if you're interested in being an alpha/beta tester. If so, just email me.

Hope this helps, though I'm aware I haven't really given any definitive answers!

Alexis