I’m curious to hear expert opinions on the best practices for RBMC, specifically when and how to implement in the processing pipeline.
Particle set 1: All “quality” particles after 2D/het refine (millions) yield a 2.5A structure with a well-resolved core and horribly flexible peripheral domain due to conformational/compositional variability.
Particle set 2: Further classification yields a 2.7A structure of 100k particles with slightly lower resolution core compared to 1 but a well-resolved peripheral domain of interest.
Would you expect the best results with A) particle set 1 for RBMC applied to final particles, B) same but providing mask around the core during RBMC C) particle set 2 for RBMC applied to final particles? And given that the peripheral domain is of particular interest, case D) particle set 2 with mask around the region of interest, albeit lower resolution.
Has anyone seen a benefit of having the upstream job “minimized over per particle scale”? I also routinely correct higher order aberrations and defocus prior and following RBMC.
My assumption is that highest resolution reference volume will allow the best global correction which can then be applied to any set of particles, so A or B? FWIW I routinely use A (which is the slowest option) and constant low-level input replacement to yield the ~0.2A improvement to any downstream particle set.
RBMC/polishing is close to the last thing I do. I’ll do it on cleaned (final or near final) stacks - with CryoSPARC all datasets can be worked on in a single run (“Number of reference volumes” allows multiple stacks/volumes to be analysed for motion/weighting/motion at once whilst remaining independent stacks) but with RELION I’ll run multiple times for each reconstruction.
I’d rather have flexible regions (which are usually the part we’re most interest in) resolve better than an extra 0.1 Ang on a region which is already high resolution.
As I commented in another thread, I don’t really see an appreciable difference between minimised scale and not, but I do try not to forget to enable it, so my sample size for “it doesn’t make a difference” is rather small…
Interesting, sort of the opposite of my expectation. yes it’s near final, yes you can do it for any number of particle sets/volumes as it’s per-particle. but I assumed, like Relion, that the fine details of high resolution are critical (“don’t even run polishing unless you’re at 2.5Å”), so there would be a benefit to having a higher resolution reference to drive better hyperparameters, which can then be applied either by low-level inputs or by RBMC starting with those parameters.
Good to hear your thoughts though, I’ll certainly try it empirically given it’s much faster to run on a near-final small subset of particles.
I also use RBMC close to the last step, because it takes less time with fewer particles, and I don’t need the improvement for particles I have discarded before the “final” set.
That said, as I understand it, per-particle motion correction should be stable as soon as you have a reconstruction good enough to match its projection to the experimental images, so probably in the ~3-4 Å resolution range (secondary structures clearly distinguishable, but no need to have side-chain resolution). So I think “don’t even run polishing unless you’re at 2.5Å” is not good advice.
I remember reading somewhere (possibly the RELION documentation) that a higher resolution 3D reference is beneficial for both CTF refinement and per-particle motion correction, therefore there is no preferred order in which to run these two jobs and one might even need to iterate between them until no further improvement is made.
It seems reasonable to me to fix the biggest problems first. Motion is generally the biggest problem, unless you know you have little motion and/or more of something else, like a very broad defocus range and/or thick ice, in which case refining per-particle defocus first might be better.
Thanks, this is helpful context. I also recall a dev’s suggestion that hyperparameters should suffice for similar but not common datasets, which matches this idea too. I’ll try iterating, running RBMC a second time after already seeing RBMC improvement in the model (with other corrections on), using the refined reference.
It’s a little less evident in CryoSPARC since there is patch CTF estimation, but in RELION I’ve noticed even narrow defocus ranges across a micrograph (<300 Ang range) can give large improvements in resolution (on untilted data) after re-refinement (2.7 → 2.3 Ang for one dataset or 4 → 3.4 Ang for another), so I’d usually work with defocus/astigmatism first, then magnification anisotropy, then beam tilt/trefoil if I’m below 4 Ang, then tetrafoil if below 3 Ang, further defocus as it’ll change slightly from higher order corrections, then polishing, then another round of defocus/astigmatism. If I’m pushing as hard as possible, I’ll try further beam tilt refinements. The exception are giant particles (200 nm+) where I’ll run CTF refinement at lower resolutions but watch very, very carefully what the output refinements estimate. Since RELION fits a resolution shell beyond the GS-FSC during polishing, I’m perhaps overly cautious about running polishing more than once and don’t do it.
I agree, I’ve never seen the advice “don’t run polishing worse than 2.5 Ang” but I don’t watch CCP-EM much so might well have missed it. RELION “?” guides for tetrafoil used to say “only recommended for data sets which extend beyond 2.5 Ang” but now says 3 Ang, perhaps that’s where it came from? Since tetrafoil has a tendency to go off into space if trying to fit at low resolutions, it’s one reason I never do global CTF refinement during NU refinement any more, but run alternating rounds and watch the output of the Global CTF job when refining tetrafoil alone.
yes I think that makes sense. I have not had almost any success finding a classification scheme that separates high resolution differences though. 3D class with resolution 4 does nothing as far as I can tell, het refine of low res and high res isotropic references does remove some low res particles that have the correct shape, but almost all of my classification is conformational/compositional and haven’t found great protocol for quality/details. Would love some ideas!
het refine of low res and high res isotropic references
You might consider reducing the learning rate - maybe down to 0.01 or so - so that the references aren’t updated so aggressively, and switching force hard classification on. That way your low res decoy will remain low res, rather than gaining high res features which might suck in good particles. You would need to carefully tune the resolution difference between the two volumes, though, so that the decoy doesn’t end up just zeroing out. Also might want a larger number of full iterations than the default - maybe 5-10? Test how many needed for convergence.
thanks for the advice! I’ll give it a try. that job does so far work really well, keeping low res as low res, removing only ~10% particles, and resulting 90% are higher resolution NU-refine than they were at input.
