Whats the issue with blurry 2D classes--bad ice?

2D Classification
1

Hi all,
I recorded 2k movies at 0.85A/px of a protein on gold-grids (UltrAUFoil), 3 images per hole.
The micrographs look fine to me, I see particles of correct size, no aggregation, concentration not too high/low (though could be a bit higher).

J1009_non_doseweighted_aligned_micrograph_for_004653757562706084320_foilhole_539474_data_546224_546226_20240601_221228_fractions
J1009_non_doseweighted_aligned_micrograph_for_004653757562706084320_foilhole_539474_data_546224_546226_20240601_221228_fractions512×512 191 KB

J1009_non_doseweighted_aligned_micrograph_for_000277454966741032796_foilhole_539473_data_546224_546226_20240601_221105_fractions
J1009_non_doseweighted_aligned_micrograph_for_000277454966741032796_foilhole_539473_data_546224_546226_20240601_221105_fractions512×512 185 KB

However, the best 2D classes I got so far are the following. Blob picker didnt work at all, so I used template picking with averages from another set (same protein, same TEM parameters) which worked a bit better.

Templates (I know they dont look good, thats why I made another session)

J655_2d_classes_for_iteration_60(1)
J655_2d_classes_for_iteration_60(1)1920×578 94.5 KB

Best(!) 2D classes from 2k movies

J1057_2d_classes_for_iteration_40
J1057_2d_classes_for_iteration_401920×963 165 KB

Whats the issue? Is the ice still too thick (7sec blotting, Atlas looks fine to me)? Or is it that there is such a variation in motion in the mics (see Patch motion corr results)
J1009_patch_motion_for_004653757562706084320_foilhole_539474_data_546224_546226_20240601_221228_fractions
J1009_patch_motion_for_000277454966741032796_foilhole_539473_data_546224_546226_20240601_221105_fractions

2

If that’s the final iteration output, that is an astonishingly small number of particles from 2000 micrographs. How many particles are you extracting and classifying?

I’d not recommend higher concentration, there is already some visible overlap. You have transmission and can see particles so ice thickness or motion impact should not be so severe as to cause those 2D classes as it’s not a smeared mess.

First 2D classes looked OK, just not many particles per class. Given how many particles appear to be present in the example micrographs, I’d expect the full dataset to be more like 1M+ particles, not <16,000… why are so many particles missing? How many are you picking per mic?

3

Thank you for the reply. In the first attempts I used all picked particles from blob picker, which result in even worse classes. So I used template picker and vie “inspect particles” used just the ones with the highest scores (above 0.4). This got me 14k particles from 2003 micrographs.

Here are three sets I recorded, each giving totally different 2D classes. The best looking mics (at least to me, by eye, 3rd set) give the worst classes

differentsets
differentsets2394×1626 989 KB

4

What are the relative ice thickness measures for these three sets? Do you see a strong water ring in the power spectra?

5

The few mics where I see the ice ring I reject

6

right - but relative ice thickness measurement? I am wondering if you have some thick amorphous ice

7

The first data set I have to run again, as it had better results with ctffind, which doesnt give rel. ice thickness.
but here are the other two

ice-thickness
ice-thickness1921×2134 490 KB