How are the per particle scale factors refined during 3D-classification, when selecting “optimal” scale factors?
Are they refined for just the masked region, or the entire unmasked particle?
If the former I guess for small masks it might be better to use input scale factors (or constant scale), hence the query.
The optimal scales are refined with respect to the consensus volume + (potentially auto-generated) solvent mask. We do not use the focus mask as we’ve found scale optimization with smaller masks to be unstable.
Thanks Valentin, that is very helpful! Perhaps it would be useful to show a histogram of the per particle scale factors in the log, like for NU-refine etc? I can look at the ones refined by 3D-class only by doing a subsequent reconstruction. When I do this I see that a bunch of scale factors have refined to zero - is this expected?
Hey @olibclarke, if you set the per-particle scales to ‘input’ or (leave at) ‘optimal’ this histogram should also be output near the beginning of the event log, before the initial structure slices. Can you double check that they’re not there?
Ah I see it now - sorry! Is it odd to see so many scale factors refining to 0? I don’t recall seeing this before during refinement?
Ah no we haven’t really seen this either – is this a ‘large’ high(er) SNR particle? This could be a bunch of junk that’s being ‘thrown out’ via the per-particle refinement.
Yes it’s a large particle - pretty clean dataset so there shouldn’t be too much junk there…
This is where it would be particularly useful to have a particle curation tool, so we could select the particles with anomalous scale factors and either map them back to mics or run 2D…
Definitely on our roadmap! For now – you could take a look at using the new cryosparc-tools to visualize particles (taken from the output of that NU refine job) that have
alignments3D/alpha below some small threshold.