Fine 2D classes of nanodiscs, but no reasonable desnity downstream

Hi everyone, I’m processing a dataset of ~120 kDa complex embedded into lipid nanodiscs, and I can’t get a reasonable density after 2D classes. They look like this (500k particles):

whereas usually they look like this (also 500k particles):

Particles from the the first image produced no reasonable reconstruction, while the second got ~3.1 A structure.

I tried doing “3D-classification” (ab initio with 5 classes and hetero-refinement afterwards), but it didn’t yield any density that would look like a nanodisc to go into the second round of ab initio:

Any suggestions of what might be wrong here?
My guess was that there’s something wrong with the micrographs themselves – it’s eer format that I never worked with, and also 2D classes are “faded” compared to the normal ones – so maybe some issues with gain / header / whatever?

if you’re lucky this a simple pixel size / box size issue. EER import will have a so-called “EER upsampling factor” that will by default cut your pixel size in half (Tutorial: EER File Support - CryoSPARC Guide), some reasons here (Movie Compression — RELION documentation) so you might want to double check your extraction box size (that looks quite large in the top classes; you definitely capture some adjacent particles as well, as visible in some of your 2D classes). You could try re-extracting with 50/65% of the box size, sort out dimers/trimers via 2D classification or try particle windowing during ab initio / homo / hetero refinement.
Nothing wrong with EER, if anything, it should be more accurate compared to movie fractions.
Good luck!

do you have any recommendations on how to use that? The mouse-over help in GUI and documentation don’t really explain what that means, unfortunately. I’ll try just reducing them from default 0.85/0.95 to some other values, but maybe there’s a rationale behind them.

It looks like you have top views in the second dataset (which works) but not in the first (which doesn’t).

I have seen similar failure modes of ab initio, where all views are assigned as side views, giving a cylindrical model.

If they are related samples, you might try just initializing heterogeneous refinement starting from maps generated from the sample that works (the one with top views)?


1 Like

yes, they are related (different batches but same sample). But, I didn’t quite understand what you’re suggesting – I skip the “Ab initio” part and initialize “Hetero refinement” from a) high-resolution maps from a different grid b) any random low-resolution blob that I had during this processing, am I right?

correct, start heterorefinement using a map from your other sample (filtered to an appropriate resolution and resampled to an appropriate box/pixel size), plus decoy volumes


so, if my initial box size is 256 px (from other project) with 1.022 A pixel, and current is 128 px with 0.93*2=1.86 A pixel, I resample to 256*1.022/1.86 ~= 140 px, and then crop to 128 px. Is that what you meant?

(sorry if it’s too much detail, I still find re-sampling confusing and try to double-check with someone more experienced here)

Just crop/resample to match the box sizes (in Å).

So in this case, your initial box size is 256*1.022=261.6 Å

Your other box is 2*0.93*128=212.48Å.

So first, resample back to bin1, then crop to 261.6/0.93~=280px - that should get you close enough. From there you can bin as you like, so long as the box size in Å matches you should be fine.


1 Like