Non uniform refinement options meaning?


I’m confused about the meaning of two of these options.

Specifically, what is the difference between “Resolution to stop subsampling” and “Local processing start resolution”?

Given that subsampling is an integral part of the local processing, how are these two different, and why is the default resolution to stop subsampling (5Å) so much higher than the default local processing start res (8Å)?

The reason I ask is because some of my non uniform refinement runs are giving much poorer density maps with the same inputs in recent versions of cryosparc v2, and I’m trying to hunt down the change that might have caused it.


1 Like

Are these options explained anywhere @apunjani? Or maybe there is a preprint on the way describing the method?? :grin:

Hi, sorry to step into this discussion. I am getting encouraging results with NU refinement using the default options, could anybody comment on how which parameters one could change to possibly improve it even further. I am not providing any mask, so also comments in this direction are very welcome.

Thanks a lot for your help

Hi @marino-j,

The typical parameters that can be adjusted are:

  • Local processing start resolution: sometimes better to start at a coarser resolution than default
  • Step size: for computational efficiency at early iterations we downsample the halfmaps by this factor when performing the non-uniform local filtering, but setting it back to 1 can sometimes be slightly better (generally not much)
  • FSC threshold: default is 0.5, it can be increased to 0.7 or 0.9 if you are still seeing any “streaking” at the edges of a micelle etc, or if you change the next parameter. In the version of NU-refine that will be released shortly, this parameter is no longer used
  • Adaptive Window Factor: this controls the effective size of “regions” that are considered to have similar properties during NU-refinement (this is described in mathematical detail in the biorxiv paper). Making this value larger will mean that larger regions are considered similar, making it harder for the algorithm to spot fast-changing boundaries like protein–micelle or micelle–solvent or protein–solvent, but making it larger will make the filtering more stable and can help if you are seeing “streaking” artefacts. Generally it’s not recommended to make it larger. Making it smaller (e.g. as small as 3) can improve the level of detail with which NU-refine sees boundaries, but this comes at the cost of more noise in each region, so if you do make this value smaller, you should probably increase the FSC threshold (previous parameter). Note that in the upcoming version of NU-refinement this is the only parameter to adjust.

@apunjani Thanks for this note. Very helpful combined with Figure 2 of the preprint!

Will the AWF also be exposed in local refinement?

Hi @DanielAsarnow, yes the plan is that local refinement will be updated as well with the new version of non-uniform refinement, with the same parameters

Thanks a lot, Ali. looking forward to try these suggestions.

All the best