Resolution worse in homogeneous refinement than in heterogeneous refinement

Hey all,

I work on a ~100kDa tetramer and I have a curious problem in which resolution and map get worse when I perform the new homogeneous refinement of a particular class after a round of heterogeneous refinement. Has anyone else encountered this problem and can give me some hints as to why this is and which parameters or tricks I can use to solve this issue?

Heterogeneous refine:

Homogeneous refine, all default parameters:

The hiccup in the FSC curve in homogeneous refine gets a little bit better if I change the mask threshold to 0.1 but I can’t say it improves the map much. I’m playing around now with dynamic mask near and far parameters. I read in a forum entry that an extra couple of final passes can help (I tried 1 extra pass once which made things worse. Maybe I’ll try 5 as suggested in the forum entry)

I suspect that I have a D2 symmetry. If I use this symmetry in homogeneous refinement, the map still looks worse than the heterogeneous refinement.

Any tips are highly appreciated!
Thank you,

Hi Claudia, when you use D2 symmetry, have you checked the map is aligned on the putative symmetry axis? The primary axis should be identical to Z (third projection/slice in cryoSPARC plots).

Hej Daniel,

I’m sorry, I’m not 100% sure I understand what you mean (I’m not a symmetry buff :sweat_smile:). If I understand correctly, D2 symmetry has 3 axes (not sure which one you call the primary axis to be honest since they are perpendicular) and the symmetry axes in my reconstruction do align with the xyz axes in CS (and also in Chimera, if I display a box around the map).

How does this help with reconstruction?

Thank you,

Dn symmetry groups comprise a primary Cn axis, and a secondary, perpendicular C2 axis. The secondary axis can be any axis, but for any Dn where 360/n isn’t a multiple/factor of 90, there is no third axis. D2 is one of those special cases, so you should get a C2 appearance on each axis like you say - I just used the general terminology for D groups out of habit.

If the reference isn’t aligned to the symmetry axes, then you will usually end up with a crazy looking (but symmetric!) density at the end of refinement. It’s a good thing to check looks right when you are imposing symmetry.

Hej Daniel,

thanks a bunch for explaining, that makes a lot of sense. So yea, I do believe that the reference is aligned okay. Problem too is that I also get worse looking maps in my homogeneous refinement if I don’t impose symmetry (in fact, the FSC curves from above are with a C1 homogeneous refinement).

I’ve been pondering if it has something to do that my particles don’t have many distinct features and it’s a small protein? Also, I realised that in the default heterogeneous refinements, the particles are downsampled. So I started a Het refine job with a full box size (still running XD…)

I also lowered the initial low pass of the reference to 20 A instead of 30 A, which tends to look better. However, I’m worried about Einstein noising my map into something that it is not.

Any thoughts on this?

Thank you,

The model bias can only be from resolutions present in the model. As long as you aren’t interpreting 20-A resolved features, it’s fine.

I recommend you also try refinement with the smaller box size. It could be that the higher SNR in the binned images is helping find correct low-resolution alignments.

Other things to try are 1) ab-initio with a higher res limit, say 6A, sometimes this works to medium resolution for a better start model and 2) further rounds of 2D classification using 60 online EM iterations and a batch size per class of 200 or 400. Perhaps there is still junk in there…

Hi @ClaudiaKielkopf,

I would second all of @DanielAsarnow’s suggestions.

One major difference between heterogeneous and homogeneous refinement is that heterogeous refinement does not use the Gold-standard split, i.e. all particles together are seen during processing. In some cases where signal is already limited, as with small particles, the slight boost from being able to see all particles can mean that heterogeneous refinement reports a higher resolution. Unfortunately this is not a “true” resolution that you can trust. If you are seeing better structure in the map from heterogeneous refinement, this could also be due to the difference in sharpening - hetero refine does not estimate B-factors but just blindly applies a -150 B-factor when outputting maps for display. So you could also try to use the sharpening tools to separately re-sharpen your homogeneous refinement map with a higher B-factor and see if those same features actually are present in the homogeneous refinement.

Hi Claudia,

Great to meet you in Lorne and congrats on a fab talk :slight_smile:

In cases like this I have sometimes found that ab initio actually gives better results than either heterogeneous or homogeneous refinement. In this case I’ll first run a single class ab initio (in this case I would try both in C1 and with D2 symmetry enforced), but to quite high resolution - say with an initial res of 9 and a final res of 4 - and then take the result of that and perform a local refinement.

The other thing, if heterogeneous refinement is giving a better map in C1, I would also try enforcing D2 symmetry during hetero and see if that gives a further improvement.

I would also try running a symmetry expansion followed by 3D variability analysis, to identify any conformational flexibility that might be causing you problems.

And finally, if you still have plenty of particles, I would continue trying to clean them up by either extensive 2D or 3D classification to find a subset that behaves better - but I expect you are doing this anyway.


1 Like

Hej all,

thank you Ali and Oli for your help and suggestions, I’ll be going back to more analysis this week after the Lorne conference and will definitely try your strategies. I’ll report on my progress :slight_smile:


PS: Oli, likewise! Have fun at ACMM and I hope to meet you again in future!

Oh - and one other strategy that helps sometimes in tricky cases:

Run a homogeneous or non uniform refinement, but limit the max alignment resolution to quite low resolution - say 10 Å, but I would test different values (say 8-15Å). Then, run a local refinement starting from these orientations and using the fsc or fsc_auto mask from the preceding refinement as the mask (in your case you would need to run a symm expansion before the local).

The idea is to get just approximately correct orientations nailed down in the first round, and then let the local refinement take you to high res. I’ve had some good results for this for pseudosymmetric things.


Hey Oli,

thank you once more for your suggestions!

One mystery is already solved here, I was looking at unsharpened maps the whole time, I only realised that after Ali’s comment on manual sharpening (big :woman_facepalming: moment :sweat_smile:) No wonder the maps didn’t look very detailed… just putting this out here in the category of “mistakes you only make once”.

So the auto-sharpened map from homogeneous refine looks a lot better which makes more sense with the resolution estimate. But this face palm moment also made me aware of the fact that sharpening in a map with a big resolution spread is tricky. I checked the local resolution, it spans from ca 7 to say 14 according to the histogram, probably courtesy of a glycoprotein. So I’m looking at doing more non-uniform refinement, even though the FSC curves usually look dreadful with the default parameters… will work on that too!