Just wanted to remark again (couldn’t find the original thread) that it would be nice for the manual resolution limit in heterogeneous refinement to be taken as the max (worse) of the max FSC over classes and the limit, so that the refinement will start at low resolution (from the initial low pass resolution) and progress to the limit rather than going to the limit directly.
I was reminded of this because of having success with a small complex providing a limit (8 Å) - but I am still a little worried about over fitting when jumping from 20 Å to 8 Å directly.
2 Likes
Hey @DanielAsarnow,
Unless I’m misunderstanding, this should be how the current hetero refinement job works. The alignment radius is set at each iteration as: the numerical minimum of {the ‘best’ (i.e., largest) FSC=0.143 radius among the classes, manual maximum limit}.
Perhaps the ‘jumping’ you’re describing occurs because there is a class that immediately has a resolution better than 8A?
Valentin
Doesn’t that mean that if in the first iteration the FSC resolution is say, 16Å but the manual refinement limit is 8Å, that it will immediately proceed to 8Å?
It seems like it would be better to use the /worse/ of the max over classes and the manual limit, to prevent overfitting in early iterations?
I.e. advance res limit normally, but when you get to high res, stop at the manual limit.
Isn’t this controlled by the “Use max FSC over classes for filtering” parameter, which defaults to on?
I.e. if this is on, it will filter all classes at the highest res limit of any single class, whereas if it is off, it will (I guess) use the per-class FSC resolution values, up to the max resolution limit (or maybe the latter just affects resolution to use for alignment, not filtering)?
It doesn’t just proceed straight to the max alignment resolution in my experience
Reviving this older thread!
Doesn’t that mean that if in the first iteration the FSC resolution is say, 16Å but the manual refinement limit is 8Å, that it will immediately proceed to 8Å?
Sorry I may have worded the above in a confusing way. It is the ‘worst of’ {best class res, manual limit}. We take the minimum over wave numbers (i.e., measuring radius away from Fourier origin) rather than resolution directly. So in this case, the 16A best class resolution will have a smaller wavenumber than the 8A limit and we’ll use that.
Isn’t this controlled by the “Use max FSC over classes for filtering” parameter, which defaults to on?
I.e. if this is on, it will filter all classes at the highest res limit of any single class, whereas if it is off, it will (I guess) use the per-class FSC resolution values, up to the max resolution limit (or maybe the latter just affects resolution to use for alignment, not filtering)?
This parameters affects the filtering of the class volumes (which do indeed get per-class FSC filtering applied if it is off), but not the alignment of the particles.
Hope that clears things up!
2 Likes