I read about FSC and understood that to refine the 3D map, the input data gets randomly split in half, two 3D models get generated, and the correlation between their respective 3D Fourier transform gets depicted. The frequency where the correlation falls below a certain threshold, this is the (local) resolution.
Two questions:
How does this knowledge help to increase resolution in the final volume?
Why do you need an input structure? As I understood two new ones get generated and their FFT compared.
this knowledge helps to qualify structures by quantifying their resolution. it is a measurement, not a step towards map improvement.
in cryosparc when you provide input structure, you are also providing many other related files including coordinate locations, particles identities and their angular assignments and CTF values, masks used, and importantly, two half-maps that were automatically derived during the refinement job. the half maps are derived from half the particles ideally, split at random, and these are compared.
To address the title, strictly speaking, as you correctly state, to calculate and FSC, two (half) maps are required. Given the body text; technically, FSC does not require an input structure. Refinement requires an input structure (see point 2 below).
It doesn’t. It’s a metric used for evaluation. It’s not always a great one (posts on this forum and others can show how going only by FSC resolution can be wildly misleading) but it’s one of many tools available, and a fairly simple, quick (sanity) check.
To align random projections, and result in a structure, you need something to align against. “Projection matching” is a classic method of doing this sort of thing, even outside of EM. This can be (effectively) anything - even a simple sphere (which is usually the basis of initial model generation, that or random projection of some 2D views into a 3D volume). From there, slowly over many cycles, each 2D view can be compared against projections of the 3D and orientations optimised until the different projections of the 3D are as close as possible to the 2D views.
There’s so much literature on this it can be a little overwhelming.
