[3dem] RELION 5 Bayesian polishing repeatedly rejects ~50% of particles after successful refinement

[Takanori Nakane]

Hi,

How many optics groups do you have?

 > collection was performed with and without tilt

Did you perform MotionCorr in two jobs for tilted and non-tilted images?
If so, you have to join these two motion correction STAR files
(corrected_micrographs.star) before giving it to Polish.

Polish removes particles on movies absent in corrected_micrographs.star.

Best regards,

Takanori Nakane

On 2026/04/28 7:50, Erik Hartwick via 3dem wrote:
> Hello 3DEM community,
> 
> I am reaching out to see whether anyone has encountered a similar issue 
> with Bayesian polishing in RELION 5 via SBGrid. I have run polishing on 
> the same “system” for another project without issue; however, this 
> project seems to be problematic.
> 
> We are processing an SPA dataset in RELION 5.0.1. Refinement behaves 
> normally, but during Bayesian polishing a very large fraction of 
> particles is consistently removed (~50%). The job completes far enough 
> to report the issue, but polishing repeatedly rejects roughly half of 
> the particles. This was a large collection 17k images at 130kx (​ 
> ~0.9apix), 300kv, 100A target dimension, preferred orientation is 
> strong, collection was performed with and without tilt, and lastly Blush 
> is being used to help with the orientation bias.
> 
> What we are seeing:
> 
>   *
> 
>     upstream particle set refines normally at 1xbin box 256 pix
> 
>   *
> 
>     Bayesian polishing then removes about 133k particles from a starting
>     set of about 261k
> 
>   *
> 
>     this has happened repeatedly across multiple attempts
> 
>   *
> 
>     one representative message was:
> 
>     WARNING: 136430 particles in the optics group 1 were removed
> 
>   *
> 
>     we also saw:
> 
>     DEBUG: FC_WEIGHT didn't match
> 
>     although this same debug line also appeared in MotionCorr stderr, so
>     I am no longer sure it is directly relevant
> 
> What we have already tried:
> 
>   *
> 
>     reran MotionCorr from scratch
> 
>   *
> 
>     reran CTF estimation
> 
>   *
> 
>     reran micrograph selection
> 
>   *
> 
>     re-extracted particles
> 
>   *
> 
>     reran refinement
> 
>   *
> 
>     traced the lineage backward through the pipeline to look for an
>     earlier metadata issue
> 
> What we found:
> 
>   *
> 
>     the problem does not appear to be explained by missing micrograph-
>     path matches between the particle STAR and corrected_micrographs.star
> 
>   *
> 
>     refinement continues to work on these particle sets
> 
>   *
> 
>     the issue seems to become visible only at polishing
> 
>   *
> 
>     the problematic lineage appears to emerge after a 3D
>     classification / selection branch, where a parent stack of about ~1M
>     particles was narrowed to a subset of about ~250k particles
> 
>   *
> 
>     downstream versions of that selected subset continue to show the
>     polishing rejection problem
> 
> A few additional notes:
> 
>   *
> 
>     this is not normal behavior in our hands, and I have not encountered
>     it before with this workflow
> 
>   *
> 
>     MotionCorr does report some movies where RELION could not fit a
>     reasonable local motion model and used global trajectory only, but I
>     am not sure whether that is related
> 
>   *
> 
>     we also saw repeated warnings during extraction about some
>     micrographs having very few or no particles after selection, but
>     regrouping did not appear to explain the polishing rejection
> 
> My main questions are:
> 
>  1.
> 
>     Has anyone seen Bayesian polishing repeatedly remove such a large
>     fraction of particles while refinement still works normally?
> 
>  2.
> 
>     Are there known metadata inconsistencies or STAR-file issues that
>     can survive refinement but cause polishing to reject particles?
> 
>  3.
> 
>     Has anyone seen this associated with a specific 3D classification /
>     selection lineage?
> 
>  4.
> 
>     Are there RELION 5-specific polishing issues with re-extraction,
>     optics groups, or movie metadata that we should inspect more carefully?
> 
> Any suggestions on where to look next would be greatly appreciated. 
> Efforts to start with a new processing attempt have started but knowing 
> why this occurred would be great.
> 
> All the best,
> 
> Erik
> 
> 
> Erik Hartwick PhD
> Director BioKEM Core Research Facility
> Department of Biochemistry
> University of Colorado Boulder
> 
> _______________________________________________
> 3dem mailing list
> 3dem at ncmir.ucsd.edu
> https://mail.ncmir.ucsd.edu/mailman/listinfo/3dem