Meaning of classification statistics?


I was wondering what the “step ratio”, “R” and “S” parameters signify during ab initio heterogeneous reconstruction? The R and S parameters seem to more or less correlate with the quality/resolution of the 3D class, as far as I can tell - is this correct?


Hey Oliver,

ESS stands for Effective Sample Size. The values for R (rotations) and S (shifts) are the effective number of poses and in-plane shifts respectively that have significant probability for the images in that class. So a high ESS R means each image on average has high probability across many poses (so poor certainty about the correct pose). This is often the case if one class ends up being spherical/globular without well defined features.
Step ratio is an internal metric relating to the step sizes used in each iteration of the SGD optimization algorithm. It’s not readily interpretable in terms of what it means for results, but it’s the amount that each structure is changing relative to the amount of “stuff” (electron density) in each class. So large values like 0.12 mean that the structure of class 2 is changing by around 12% in each iteration.


Aha, that is a very helpful explanation! Thanks Ali


Follow up q - if the step ratio for a particular class remains high, am I right in thinking that further sub-classification of that group of particles may be worthwhile?

That’s a reasonable course of action, though I can’t confirm that there’s a strong correlation between step ratio and subclasses present. Generally we just try sub-classification of any good classes :slight_smile: