What's the key difference between Heterogenous refinement and 3D classification?

I was wondering what makes the difference between the heterogenous refinement and 3D classifications that are both used to sort out different conformations. Tried to go over the tutorial website but it did not explicitly state the main differences/ advantages/disadvantages of each approach.
Does heterogenous refinement mainly rely on alignment of 2D averages (relative to the input volumes?) to sort out different classes while the 3D classification looks at ‘3D volumes’?

Thanks in advance for your help.

1 Like

There is an older ab initio method (not in cryoSPARC) that uses 3D reconstructions based on 2D class averages, otherwise all 3D refinements methods align each particle according to origin and 3D rotation. The 2D classes are only used for particle sorting, usually by hand.

In general, there are three kinds of 3D classification. 1) We can perform classification, whilst also globally searching all possible poses for each particle, 2) we can perform classification whilst locally searching poses that are nearby a previously estimate pose, and 3) we can perform classification only, using just the previously known pose itself. In all cases, both the pose and class search components are performed by cross-correlation of test projections made using one or more reference volumes transformed according to one or more poses.

In cryoSPARC, “Heterogeneous Refinement” does 3D classification using global search; it uses branch and bound to rapidly estimate poses at each iteration similar to Homogeneous Refinement. Like 3D classification in other softwares, there is no independent half-set refinement (the so-called “gold standard” approach).

The “3D classification” job does 3D classification without any alignment search (like the --skip-align option to Relion).

Both approaches are a bit different than in Relion, and the CS jobs have more options for controlling initialization, especially when it comes to using multiple unique reference volumes (it can be done in Relion using the option to give a .star file with a list volumes as the initial reference, but it’s not as natural).

The relative advantage of the different search methods is mainly a factor of the molecular mass in the region of interest. If there is enough mass to reliably find or improve the alignments, then pose search is likely to be helpful. If not, then “skip-align” (this is a good search term) is probably necessary. The purpose of local search is to accelerate high-resolution pose refinement by focusing only in the most likely part of the pose space.

Hope that’s clear! :nerd_face: