I am processing a map that has a very rigid inner region surrounded by a very flexible outer region so I am running a Local Refinement using a mask surrounding the rigid region based on a NU-refinement and I am trying to optimize parameters. Since the GSFC is going to be a much higher value (lower resolution) for the overall structure than the region within the mask, I’m thinking the most important thing would be to manually override the maximum align resolution to 2x the pixel size but I guess it depends on this, is the local refinement using the GSFC from the NU-refinement or is it calculating it as it goes based on only the region within the mask?
Also wondering if you have pros and cons of heterogenous versus non uniform refinement in this situation, or if it is inconsequential because of the applied filters?
The default is to use the FSC 0.5 resolution in the masked region as the alignment limit. Generally speaking, higher alignment resolutions than that will lead to overfitting.
I can’t advise you strongly enough NOT to set the alignment limit to the Nyquist frequency as you have suggested. Overfitting to noise (likely all you have at Nyquist) doesn’t simply mean the resolution won’t improve, it actually means refining incorrect parameters that will make your structure worse.
You can set a limit manually instead, but you should use a conservative value that is somewhat lower resolution than the true resolution of the (masked) map. This is a good idea if for example the FSC is not being estimated well due to the tight mask. Usually the direction is the opposite though - you would see that the FSC resolution is spuriously high, and then keep back the alignment limit to a lower resolution.
For your 2nd question - heterogeneous refinement simultaneously classifies and does global alignment searches without a half-map separation. Thus it can be prone to overfitting. You might use heterogeneous refinement to try and separate different sets of particles that had relatively consistent conformations at the flexible portion of the structure, and then use those classes for their own homogeneous or non-uniform refinements.
Non-uniform refinement is similar to homogeneous refinement, except that instead of smoothing the entire map with a single parameter (to regularize the refinement and prevent overfitting), the map is locally smoothed with different unique parameters learned by comparison of “quarter-maps.” This approach works well when the resolution varies across the structure - in your case I would expect non-uniform refinement to give you the best consensus refinement. Then you can use the alignments from the NU-ref to start local refinement of different regions, or focused classification (using 3D classification in cryoSPARC or Relion), or 3D variability analysis, to try and figure out what’s going on in the flexible regions.