Strategies for high resolution refinement

Hey tarek, nice improvement! I should not have assumed, and now I’m sure you’ll get that extra 0.05A.
I think the next best option is iterative bayesian polishing. Run relion_reconstruct on the half sets of particles from you NUR job, postprocess, then train on 5-10k particles, then polish and refine again in cryosparc. This usually gives me the best improvement.
You’re getting into the range where ewald curvature correction may help. Run through relion_reconstruct and postprocess. Run a parallel postprocess on the cryosparc half maps to detect improvement.
ALSO, you can try the new NU refinement in v3.0, which now has ctf optimization, but the new algorithm alone may give an improvement.

Hi Aaron,

at the moment I try to stay inside cryosparc, simply for the reason that I would need to do movie alignment again with relion. Do I get it right, that you would always run the refinement with cryosparc and only reconstruct/post-process with relion?
What improvements would you expect from polishing?

Ewald sphere correction for a particle diameter of ~ 100 A seems to me less relevant, however I can give it a try.
Using the latest NU refinement improved resolution from 2.1 to 2.05, using on-the-fly CTF correction yielded 2.07 A.

1 Like

Hi tarek,
To my knowledge bayesian polishing is the best option for the largest res improvement at this stage. It may work, but may not. I think it will work for such high res on the falcon.
As an example for a 100A particle of mine, K3 data. It goes to <2.5A with patch motion + local motion. Then process fresh with MC2, autopick same settings, cleaned up particles go to about the same res, slightly worse (don’t be discouraged at this stage). Then bayesian polish train, polish with those settings, and shiny particles get ~0.25A better. Second round of polishing adds another 0.06A.
You will need to re-run mc2 in relion and the whole process will likely take a few days but well worth it if you want to push res.

1 Like

Hi Aaron,

For my reconstructions, my boxsize is 456 (0.55 A/pix in super-resolution mode) but I have binned by 2 and so it is 228 and my map right now is ~3.8A. To try larger box sizes for particles, would you have suggestions for what increments of box sizes I could try?

Thank you

Hi schow,

456 is a weird box size. For computational reasons I would choose a different number with better divisor. Find inspiration here:

Estimation of an appropriate box size to sample the full CTF can be made here easily.


Hi schow, see Oliver’s post about optimal box size here. You can see it does depend on defocus. You can always test, say going up to 512 if your gpu can handle the refinements.

It sounds like your Nyquist is 2.2A, which is pretty far from your 3.8A nominal resolution. Is the particle close to the edges of the box? Less than 30A from the edge of the circle mask (85% of the box diameter by default)? Unless the bounds are very tight, further classification or focused processing, etc. are probably going to be more useful. (And it will be faster with the smaller box size, though you might change to 512px / 256px which is actually a bit faster).

Hi Daniel,

I re-extracted my particles with 512pix (bin/2) box size but the overall resolution and map quality did not change much. My particle is membrane protein and come to think of it, the edges of the micelle do indeed come close to (probably within 30A) of the circular mask (using the 85% default). May be I should try an even larger box size (~540/560?) or increase the circular mask dia to 90-95% box size and see if there are any improvements. Is this reasonable?

I have tried to reclassify/refine but the resolution usually worsens (probably due to loss of particles/heterogeneity issues)

Using a wider circle mask is easier - give it a shot. In practice, I have not seen a wider box improve resolution for any of my projects, even when my initial box was a bit tight. Unless you’ve collected only at defocus > 1.5 micron, you likely have lots of particles with 500 nm - 1 um defocus. These have less CTF delocalization and better CTF envelopes so probably contribute the most to your high resolution signal anyway.

If you really do think you need a much larger box, remember that you don’t need to do integer binning. I often use 540/360 or 432/288.

Most likely though, you are only limited by heterogeneity and/or alignment accuracy. You may need to combine your non-uniform refinement results with e.g. classification without alignment in cisTEM or Relion, or perform more extensive focused classification in those programs, or apply 3DVA or local refinement in cryoSPARC in order to improve your resolution further. With Nyquist of 2.2 A, you should not need to unbin until you reach better than 3 A.

Tried the wider circle mask - the resolution did not improve but the B-factor reduced from 98 to 94 (after Homogenous ref and Local ref with NU ref) although I’m not sure whether map quality is improved.

Increasing box size from my previous “weird” choice of 456/228 to 512/256 did lead to resolution improvement by 0.05A and map looks slightly better in a few places. (85% circular mask)

I tried to increase the box size further to 600/300 and the resolution worsened by ~0.3A and map quality was also worse (85% circular mask). Not sure why this happened. (larger box = more noise outside particle and worse SNR?)

To tackle heterogeneity on cryosparc, I have been mostly using a combination of heterogenous ref (take particle set and use multiple copies of the same reference map) and Ab initio reconstruction (choose the particle stack with best features for downstream refinement). Is this reasonable? (Haven’t tried to play with 3DVA as yet)

There’s nothing wrong with 456px, it’s just a bit slower than some other nearby sizes like 432 or 512, due to the way the FFT works. Small changes in B-factor and nominal FSC are to be expected. You will get changes of that magnitude just by running the same refinement multiple times - even if you use the same random seed. GPU computations are intrinsically non-deterministic due to floating point errors.

I agree about going to too of a box influencing refinement, especially if there are neighboring particles or a strong background (e.g. high salt).

Sometimes there is no solution other than classification with local search or without alignment, in which case you will have to try something other than cryoSPARC for now. Or perhaps you are now limited by your biochemistry/intrinsic dynamics. If you have > 150,000 particles I’m sure you can get a little further by focused classification, especially if there are some specific regions that are less well resolved.

For more sophisticated guidance we need better information about your sample, FSC curves, map slices etc.
Giving your pixel size you should easily achieve better than 3.8A if your sample/data allows.
Just for comparison: I mostly run refinement with bin2 (1.7 A/pix) until resolution reaches Nyquist (3.4A).
If you can not get there something else is limiting.


1 Like

My sample is a 250kDa membrane protein. Under my biochemical conditions it is quite stable but its not like apoferritin which readily gives ~Nyquist reconstructions. So I would be (pleasantly) shocked if I get that resolution with my protein.

In any case, I am not particularly short of particles. My refinements thus far which give ~4A maps use 500K+ particles and my protein is C2 symmetric. So I am cautiously optimistic that I might be able to push the resolution further. Also, while my overall res is ~4A, the core membrane domain is at around 3.3-3.6A. There are other peripheral domains which are res 5-6A. I have tried using masking on cryosparc (local resolution/particle subtraction) but I almost always get masking artifacts and I’m not sure how to resolve it but may be I have not tried hard enough.

Before moving to relion I wanted to see if there are other suggestions on classification in cryosparc but it seems that I might be running low on options on cryosparc at this point. Thank you @DanielAsarnow and @tarek

@schow Same here. With C4 implied, I’m also getting a 4.5A smooth FSC convergence but lots of really strong micelle densities which I’d like to believe artifacts due to masking. Not sure how CS is dealing with micelles !

Cryosparc is doing fine with micelles. You can find many published structures of membrane proteins solved with csparc, we have good experiences too.
If sample/data allows sub 3A reconstructions should routineously be possible.

@tarek Would you have some suggestions on masking methods? I usually keep the threshold between 0.1 and 0.2 and the near/far parameters either 3/6 or 6/12 (usually dynamic). Anything special you would recommend trying?

@schow I hardly change the default masking parameters during cryosparc refinements. I’m not a big fan of signal subtraction, therefore I can not comment on this.
For local refinement masks I choose a contourlevel in chimera that is covering the interesting regions generously, mostly after LP-filtering. The segger tool helps a lot cleaning artifacts.
With the CS volume tool in most cases dilation/softening of 2/2 is doing well.

1 Like

Not directly related to the use of cryosparc, but would be happy to hear if someone has good experience with the density modification (resolvecryoEM) of phenix.
I think the result is not too dissimilar to that of sharpening, or of using deepEMenhancer. I think the low resolution features will be largely gone.

I see that in this paper is used after NU refinement/local refinement, with quite a bit of gain in resolution (whatever that means in terms of real gain in quality of the map)
(by the way, quite interesting what they did with cryosparc on sorting 2D classes).

Many thanks

1 Like

I think it is a great tool to squeeze out a little extra information from your reconstructions to help modeling. Maybe I’m too conservative, but personally I wouldn’t take the results for deposition or use the determined FSC as resolution estimate.
Especially in less resolved regions (by flexibility, heterogeneity etc.) local resolution will drop making the “sharpening” suffer and artificial.
Also, sharpening of structures with very heterogeneous electron scattering like ribosomes (rna vs protein) seems not to work well.


I agree mostly with tarek, although I believe it’s fine to deposit the density modified map as long as that’s what you used to build/refine and you also deposit the associated half-maps used as input, and don’t report phenix-resolve resolution. I agree it can oversharpen, but it can also brings out really nice improvements. Like many things, it’s highly dependent on local resolution.