Hi all,
I am getting some weird results when I try to use local refinement. They just started happening and I am not sure why. Is this possibly an issue with overfitting?
Thanks!
Meranda
Hi Meranda,
Those plane wave artefacts in the volume are sometimes observed after the sharpening of the maps (see this post for some discussion). In the streamlog, refinements always plot the sharpened versions of the maps but the artefacts might not be present in the full unsharpened map, or the half-maps. For an iteration that has these artefacts, if you download the full unsharpened map, and one of the half maps, do you also see these plane-wave artefacts when you open either map in UCSF Chimera?
You can download these maps by going to the “Outputs” tab in the job panel, scrolling down to the “JXX.volume.map” (and “JXX.volume.map_half_A”) sections, clicking the “versions” drop down menu, and clicking the download icon at the right side of the iteration bar, for any of the iterations that show this problem. Below is a screenshot of the expanded drop down menus with the download buttons for an example job output.
Best,
Michael
Hi @mmclean,
Thanks for the quick reply! I checked on the unsharpened mask and one of the half-masks and the artifacts are present in the unsharpened map, but not in the half-map. The half-map does just look weird though… Like there are large chunks of volume missing.
Is it possible that my mask isn’t large enough? I am trying to refine a very small region of my particle so I used a very small mask.
Best,
Meranda
Hi Meranda,
Are you using signal-subtracted particles (from the Particle subtraction job) for local refinement, or are you using the particles directly from a consensus refinement? Subtracted particles will (ideally) only contain signal outside of the mask that was used for subtraction, thus if there are large chunks of volume missing that are within the mask used for local refinement, that likely indicates something went awry somewhere in the signal subtraction. In that case, you may want to try running a local refinement first using the raw unsubtracted particles from the consensus refinement, then run a second local refinement using the signal subtracted particles, and compare the resulting output maps to make sure that density within the mask looks similar and no chunks are missing – this can help validate how successful the signal subtraction was.
Per the mask size issue, generally local refinements need a reasonably sized mask in order to produce optimal results. A commonly cited metric is that the masked domain of the protein should be at least ~150 kDa in mass. The reason for this is that only signal within the mask is used to align particles to the volume, and with a diminishing amount of signal, alignments can get less accurate. The main protection against this in local refinement is the rotation and shift search extents, which constrain the alignments to lie within the specified magnitude around the original poses and shifts. If using a larger mask is not an option, the main way to prevent overfitting / poor alignments is to make the search extents very tight (e.g. a few degrees and 1-2 pixels of shift).
Best,
Michael
Hi @mmclean,
Thank you! I am not using signal-subtracted particles, I am using the particles from a NU-refinement job that I ran previously (and made the mask from). I’ll try to make the mask a bit bigger and also try to make the search extents tighter.
I also have not assigned a fulcrum point for this local-refinement. Could that in combination with the small mask (the one I have is maybe about 10 kDa so WAY small) be why I am getting these artifacts?
Best,
Meranda
Hey Meranda,
Yeah, that’s really small – that in combination with the fact that you aren’t using subtracted particles (and still are seeing chunks of volume gone) leads me to suspect that alignments are definitely going awry. Certainly try to see if you can make a mask covering more of the protein (ideally, the largest part of the protein that’s suspected to move as a rigid body, and that preserves tertiary structure). For the fulcrum, try setting it to the center of mass of the mask. Unfortunately there isn’t currently a way to do this in cryoSPARC automatically – you’ll have to use Chimera’s measure center command (see this post for some details from DanielAsarnow – also a generally useful thread for tips in using local refinement).
Best, Michael
1 Like
Thank you @mmclean for your help! I’ll give this a go and update the post so others can see how it turns out.