Yes, a crystallized protein will always be in a single conformation, or you won't be able to see it because the electron density map will be an average of all possible conformations, and therefore meaningless.
Single particle gives you the opportunity to see different conformations, but only if the data is discrete. If there's a continuous amount of conformations (think a molecular motor that's rotating) you would need nearly infinite data to resolve a nearly infinite number of conformations. If the data is less than continuous, you can image enough particles to see all the different conformations by constructing multiple models in parallel and using 3D angular searching to bin them by what conformation they are in. This is a computationally exhausting process, however.
I hate to be that guy, but your first statement is technically incorrect: you can have discrete alternate conformations superimposed in the electron density (usually 2-3 is the most that can be resolved), and you can have different conformations of multiple copies of the molecule. (I've personally worked with both, although the differences in the second case were small.) That's not even counting ensemble-based approaches to modeling the crystal structure, although that's arguably just a different way of representing the uncertainty.
All that aside, crystallization certainly biases it towards specific conformations, which single-particle EM does not.
Didn't realize that. My experience with xray crystallography is limited. Obviously there's some variance, but I always assumed it would simply be unresolvable/disordered in that case.
Single particle gives you the opportunity to see different conformations, but only if the data is discrete. If there's a continuous amount of conformations (think a molecular motor that's rotating) you would need nearly infinite data to resolve a nearly infinite number of conformations. If the data is less than continuous, you can image enough particles to see all the different conformations by constructing multiple models in parallel and using 3D angular searching to bin them by what conformation they are in. This is a computationally exhausting process, however.