Good 2D classes, poor 3D reconstruction

[xaoshi]

Hi everyone,

I’ve got a soluble protein complex (45kDa+55kDa) for cryo-EM study.
Movies: 0.58A, 300V, 40e/A2; Particle Extraction: 256pix->128pix

1st 2D Classification

1st Select 2D:get rid of most dummy particles (7,695,260 down to 1,123,196)

2nd 2D Classification, same settings

Could see secondary structures such as a-helices.
2nd Select 2D: (1,123,196 down to 350,114)

Ab-Initio (3 classes; Particles: 2nd select 2D, 350,114)

1st Hetero Refinement (Volume: classes0&1 from Ab-Initio job; Particles: 1st select 2D, 1,123,196; Force hard classification)

2nd Hetero Refinement (Volume: classes0&1 from Ab-Initio job; Particles: class0 from 1st Hetero Refinement, 608,026; Force hard classification)

3rd Hetero Refinement (Volume X2: class0 from 2nd Hetero Refinement; Particles: class0 from 2nd Hetero Refinement, 436,434)

The resolution of class0 could get to 4.1A.

Homo Refine (Volume&Particles: classe0 from 3rd Hetero Refinement, 237,255; Minimize over per-particle scale)

The resolution got worse, even worse than 3rd Hetero Refinement.
The maps are shown below:
_______________Homo Refine_________________________classe0 from 3rd Hetero Refine

Is there anything I could try to get a better 3D reconstruction?

[ggcc]

1. The box seems a bit small - maybe you could try with larger box size (around 2.5x your particle size in each direction)?

2. If you are absolutely sure that the complex is in a single state, you could try ab initio with only one class to force reconstruction with multiple views. There’s a chance that each orientation is being isolated to each class.

3. How does the ab initio look? The orientation distribution plot for hetero refinement looks like particles are not being properly assigned in the first place

4. Lastly, I suggest to remove duplicate particles - FSC curve is not dropping to near-zero.

[CryoEM2]

your het refine are selecting for preferred orientation. It’s non obvious how to do it, but you can trick het refine using different references to put ALL proper particles to one class and discard others, sometimes discarding excess dominant views too.

As a sidequest, would suggest to run 3D classification job on a large subset of particles, changing filter resolution parameter to a few different scales (5Å, 8Å, 12Å) and of course 0 class similarity. For some reason that job can output beautiful medium-resolution models where the particles have extreme anisotropy, but the references don’t. Use any of those classes that look good as inputs for the het refine.

[Das]

Dear Xaosi,
I would suggest

1. Check the size of your protein first before assigning box size. What you can do is take the reconstructed map and measure the dimensions in chimerax. say for example its 150 A, add 150 more to it, and divide the total by pixel size. Then the allowed box size can be found [here].(How big should my particle box be? – The Jiang Lab). Although your 2D looks good there are streaks, which means you are over-aligning the particles, so in this case you can keep alignment resolution 5 instead of 3.
2. If pref orientation is the problem, then without “select 2D” take everything from 2d classification to ab initio and make 5 classes. Keep the resolution between 14-8 A.
3. Take the best class for non-uniform refinement and use it for hetero refinement next.
4. Take the NU-refined volume and the particles and feed it into hetero refinement with 4 junk decoys from ab initio and repeat until you see any improvement in your main class volume.
   This should take care of the pref-orientation problem.
5. Finally do the NU-refinement.
   all the best

[xaoshi]

@ggcc @CryoEM2 @Das Thanks for all your advice. The particle size is ~90A. Following the 1st Select 2D, I removed duplicates and re-extracted 450k particles (400pix->200pix) from curated exposures (around half).
2D Classification.

To get a better reference for Hetero Refine, I tried the following with all 450k particles:

ab initio(5 classes, Initial resolution: 14A, Maximum resolution: 8A)

The last three maps are likely representatives of three orientations, which was seen in previous ab initio job.

In summary:

1. it seems that the orientation preference persists (mostly side views, third map);
2. different orientations sorted by ab initio job could not align well in 3D reconstruction

I’m gonna try graphene-coated grids and collect a small data set for a test.

[olibclarke]

If you are seeing ab initio separate orientations rather than true 3D heterogeneity, you might try switching off (together) “center structures in real space” and “enforce non-negativity”. I still don’t understand why this works, but sometimes it mitigates artefacts like this

[CryoEM2]

you could take particles from last three maps, pool them to one, and run ab initio asking for 1 class. Or try NU-refine 3 jobs each with all particles but using one of the available maps.

Yes, new data can always fast-track success, but seems like you don’t have extreme bias, just need to try a couple tricks to get the data processing to work.