What's next after heterogeneous refinement?

open

#1

Hi all,

I used all the particles to do the homogeneous refinement but only got a ~10 A map. While I did the same with heterogeneous refinement, I could get a ~5 A map.

Right now, I’m confused about what to do next.

Do I need to import the 5 A map and the particles belong to that class and do a homogeneous refinement?


#2

Yes, you should try exactly that. You should also consider the following questions:

  • Does manual inspection of the map support those resolution numbers? (Never trust FSC alone).
  • Has all of the compositional or conformational heterogeneity been resolved? Should further rounds of 3D classification be performed, perhaps using focused masks or skipping alignment (Relion or cisTEM).
  • What does ResLog say about your dataset size and resolution limit?
  • Does the particle have intrinsic flexibility or diffuse densities (e.g. micelle) warranting non-uniform refinement? Does NU-refine significantly improve map quality over your final homogeneous refinement?

#3

Daniel, thanks for your reply. I have some further confusing:

  1. In CryoSPARC, is hetero-refinement similar to the 3D classification in the relion?
  2. After Hetero-refinement, how to have a sense of the total number of the particles of each class?
  3. 39%20PM
    In the output of the hetero-refinement, does the volume here represent the class with most of the total particles and highest resolution? And does particle_class_0_(34262) belong to that exact volume?

#4

Heterogeneous refinement is essentially the same as 3D classification in Relion, except with cryoSPARC’s innovations (e.g. branch-and-bound) and differences in implementation details.

The number of particles per class is listed in the outputs as you noted. The classes are annotated as “_class_0.” The number of classes is determined by the number of reference volumes you add, as in Relion, these can be be copies of the same volume or separate references.

The number of particles for a class might vary slightly between different screens in cryoSPARC, because of the membership probability threshold. CryoSPARC uses “soft” classification internally, where every particle has some probability for membership in every class. By default the threshold is 90%. This is also why the particle count for classes from ab-initio might disagree slightly between the summary screen and the job result screen.

In your case, it does look like all the particles might have ended up together in one class, in which case the improvement in resolution doesn’t make much sense.

What were your starting references? Can you check all the box sizes? Naturally the refinement will be limited by the Nyquist frequency given the box sized used in refinement. The default for heterogeneous refinement 128, while for homogeneous refinement it’s the size of the particles.


#5

I think what might be happening here is that your initial model is a little bit wrong. I have noticed that the homogeneous refinement is not good at tolerating slight inaccuracies in the initial model at the initial cutoff resolution (e.g. 30 Å). Relion on the other hand has a higher tolerance for slightly incorrect initial models, as they get adjusted at early iterations.

I think you have given the heterogeneous refinement job a single class which is why all of your particles are in a single class. It could be at a higher resolution because the heterogeneous refinement model has more capacity to change the model, as that is the purpose of the algorithm. This corrected model can then give a higher resolution map, as it can refine more accurately.

I would try putting the output map from that into a homogeneous refinement and see if it refines to higher resolution and gives a better looking map.


#6

Hi @donaldb, @DanielAsarnow, @ruiruigo,

This is an interesting discussion and you are definitely correct in the interpretation of the results. Heterogeneous refinement does use a different algorithm than homogeneous refinement (even with only 1 class) and is more similar to ab-initio reconstruction in it’s ability to “explore” the space of 3D structures. Homogeneous refinement makes a strong assumption that the initial model is correct at low resolutions (which is generally the case after multi-class abinitio or hetero refinement with >1 class).
I think the best question to ask is really whether or not the 5A map from hetero refinement actually has 5A reasonable features that the 10A homogeneous refinement does not. If so, there is definitely signal left on the table, and in some cases, if you add more final passes to the homogeneous refinement, the resolution may improve.