I am trying to determine the structure of a small protein using cryoEM. I used cryoSPARC to process the data collected from Titan Krios and I ended up with a 3.8 angstroms structure after non-uniform refinement. Note that homogeneous refinement gives me a ~6 angstroms structure. I then exported the 2D classification data to RELION to further process it and improve the resolution but I have only been able to get up to ~5 angstroms structure. If RELION gives a better 3D classification that is supposed to improve the resolution, why am I not getting a high resolution structure of even up to 3 angstroms to replicate the result from non-uniform refinement in RELION? I have tried processing the data from scratch in RELION but I am only able to get a 5 angstroms structure.
I would caution to somewhat ignore the reported resolutions and instead inspect the maps for quality and relevant features (though you may already be doing this and seeing that they correlate well with reported resolution). As you suggest, you can take any particles from Cryosparc to Relion and vice-versa, and I’ve found that each program generally yields almost identical structure quality given identical particle sets. So the big difference is how they classify particles - separate the junk, classify the conformational and compositional heterogeneity, etc. And for this small protein there may be major differences in CTF/motion particle preprocessing and post processing.
Take your best NU-refinement of a lot of particles, export to Relion, run the Classification that you think is better, then import the STAR file back to cryosparc for the NU-refine again. This keeps all perceived benefits of cryosparc preprocessing and refinement, while utilizing perceived improved 3D class.
Both programs excel at all job types, but understanding how to use them to maximize the result takes some practice for both, maybe especially for cryosparc.
Thank you for your the suggestion. I will try exporting to RELION, processing it and re-importing to cryoSPARC to see if that will help.
Absolutely agree with @CryoEM2. I have a few edge cases where either RELION or CryoSPARC does particularly well on a given dataset (1.24 Ă… in RELION vs. 1.5 Ă… in CryoSPARC for one dataset and 2 Ă… in RELION vs. 1.5 Ă… in CryoSPARC for another are the standout ones for me), but generally the final results will be fairly similar.
@DanielAsarnow summarises some of the differences which cause RELION and CryoSPARC to give different results in his PyEM wiki articles, although some points have changed slightly since original publication, e.g. full passes during 2D, which is no longer true with VDAM.
Another point is that NU refinement and “standard” refinement in RELION are quite different due to the different marginalisation strategies used. CryoSPARC’s closest equivalent to 3D refinement in RELION is homogeneous refinement, where you’re reporting reasonably well aligned (in terms of resolution) results. CryoSPARC also optimises (tightens) the final post-processing mask, which boosts reported resolution while RELION does not (and recommends a very soft mask for postprocessing) which will have an impact.
If using RELION 5, try blush regularisation. It’s the closest equivalent to NU refine.
You could also try increasing the tau2 “fudge” factor (it makes a huge difference for amyloids, although I’ve not tested it for small proteins it might help similarly) but be very careful if trying that, RELION warns you extensively about it in the logs for very good reasons (beware overfitting!).