Experimental support for Relion's Bayesian polishing in csparc2star.py

Import and Export
32

Hi @DanielAsarnow

Sure, I’ll give that a go.

In the mean time, I’ve tried to dig a bit into the discrepancy. I took the pre-shiny particle images from S3b, flipped them in Y (relion_image_handler --flipY), replaced the original stacks and refined against the original reference, in order to simulate the necessary shifts/transformations required if particles were sourced from Y-flipped mics.

Averaged across a few thousand particle images, the differences in rlnAngleRot, rlnAngleTilt, rlnAnglePsi, rlnOriginXAngst and rlnOriginYAngst, compared to right-side up particle images seemed best described by…

  • phi+180, 180-theta, -psi, XAngst (as is), -YAngst

Note, it’s similar to the transformation suggested by @sunch here, sans Y-shift inversion. That additional element is incorporated in --flipy. That one difference is perhaps why @olibclarke didn’t quite have success with just the Euler angle modifications at the time.

Scenario 3: CS to RELION
  1. c) tif movies Patch Motion-corrected in CS.
    → pre-shiny particle coordinates from S1 Y-inverted and used to extract from csMicrographs (Recenter using aligned shifts disabled).
    → Particles refined against original reference (3.1-3.2Å).
    → csparc2star.py --movie to generate csMicrograph and motion star files.
    → csparc2star.py particle conversion for polishing.
    → application of the above .star file modifications using awk.
    → half_maps and mask_fsc_auto used as is for postprocessing.
    → shiny particles now yield comparable, if not identical results to S3a.

I’ve also been wondering about the slight difference in polishing outcomes when sourcing from RC/MC2 vs PM in CS.

The FSC is improved vs un-polished particles in both cases, but RC/MC2 has been giving slightly better FSCs. I wondered earlier if it was stochastic, but I’m beginning to think that it’s due to the fact that RC/MC2 and PM do not share the same reference frame–PM has no reference frame. Meaning if particles are extracted from csMicrographs, the particles won’t be exactly where they should be when Relion goes back to the movies. That is, unless csparc2star.py corrects for that?

Perhaps another reason to stick to RC/MC2, for consistency.

Cheers,
Yang