Hi all,
We have been seing strange patterns with FSC curves in one of our datasets when doing NU-refine jobs, with the tight mask curve falling and then rising.
Previous threads here indicated that perhaps widening the dynamic mask far setting may help, but it didn’t here. We’ve also experimented with widening dynamic mask near, and both widening near and far at the same time or shrinking near and far at the same time, and still we see this pattern.
What might be causing this?
Thanks!
It looks like you have over fitting caused by a strong orientation bias towards the view on the right.
The map looks very strange. How do the 2D classes look? And the orientation distribution?
There appears to be no secondary structure.
As Oli says, what do the 2D classes look like? At <5A, you should be seeing clear alpha-helices at the least (with an option on “might be beta-sheet” in better regions)…
There definitely is a preferred orientation to top views, but enough side views present that we have been able to to get ~5A reconstructions. In the best result so far, we do see secondary structure, but as we continue to filter out particles to try to improve the refinement, this issue above becomes more and more prevalent.
The strategy that worked best for me in cases of strong orientation bias is minimizing 2D classifications. I like to start with a single 2D classification with very generous class selection then move straight into heterogeneous refinements using one “good” class and 2 or more junk classes depending on the size of the particle stack. Follow this job with a homogeneous refinement of the good class. After that, I repeat this process with the particles and volume from the “good” class and then continue doing this until it stops sorting particles as junk. You can play around with different reference models for the junk classes and get different results usually.
I’m not sure what’s gone into the pipeline to get to the result you’re showing but in my experience, per particle CTF, abberations, and motion correction on orientation biased data in cryosparc can lead to overfitting. I saw a marginal benefit from NU refinement but too many iterations of it can also produce weird results.
Hope this helps,
Ryan
