Unexpected out of plane tilt distribution in early cycles

Dear cryoSPARC team

I think something is going wrong in the first cycle of my helical refinement (it could be a bug), but I could not figure out why. Here are the details:

As a quality control, I should first say that:
(1) the averaged power spectrum of particles go to ~4Å with sharp layer lines;
(2) 4Å reconstruction can be obtained using Spider/Relion, and side chains can be clearly seen with good quality. So this is a do-able project.
(3) The tube is a wide, doubled-wall tube.

However, in cryoSPARC, no matter what symmetry I tried, the out-of-plane tilt always goes crazy at cycle 1. Here is what the reference looks like

And here is the out-of-plane tilt

This is quite strange, because at low resolution the out-of-plane tilt is expected to always look reasonable. Any thoughts on what could be the cause for this?

I should also add that (1) not applying symmetry until 6Å (2) not searching helical symmetries until 4Å (3) limiting the search range of out-of-plane tilt to 5 degrees have been tried, but none of them resolved this issue.

1 Like

Hi @Jerry,

Thanks for the post. The tilt distribution definitely looks a bit strange, as we’d expect something roughly gaussian looking – I have a few questions to help diagnose:

  • Does the refinement converge when run with the same parameters as in the plots above? Does the final tilt distribution look healthier, or is it similar to the initial one?
  • With other softwares like Spider or RELION, how are you initializing the structure?
  • Using the averaged power spectrum, do you have a set of symmetry candidates that you’re enforcing during the refinement?
    • If so, what was the extraction box distance and what is the helical symmetry order enforced? You can see the latter by checking the refinement, which should print out “Refining with helical symmetry order of __”; this is the number of times each image is used for backprojection.

It may be worth trying to initialize a cryoSPARC helical refinement with the initial volumes you found to be successful elsewhere. In my experience initialization is the trickiest part, as many helices do not have much lower frequency information to enable SGD-based reconstruction methods to work out of the box, and sometimes refinements struggle with this too if just initialized with hollow cylinders. (Without a volume input, helical refinement essentially uses hollow cylinders as initializers by reconstructing the filament with known in-plane rotation but scrambled azimuthal orientations).

If you are confident that the correct symmetry value is in the set of candidate symmetry parameters, then it may be worth increasing the helical symmetry order via the Maximum symmetry order to apply during reconstruction parameter. In my experience, larger values help the algorithm converge better by applying a stronger symmetry constraint during reconstruction.

Let me know if these tips help!

Best,
Michael

1 Like

Thanks Michael for the responses, to answer your questions first.

Does the refinement converge when run with the same parameters as in the plots above? Does the final tilt distribution look healthier, or is it similar to the initial one?
Nope, the refinement won’t converge to high resolution no matter what symmetry was used. Actually, without using an initial volume reference, all of them were stuck to ~7Å resolution. The final out-of-plane tilt distribution looks the same as the initial bad one.

With other softwares like Spider or RELION, how are you initializing the structure?
For SPIDER IHRSR approach, initially, we turned off the out-of-plane tilt search for a few cycles. The helical symmetry is applied at the end of every cycle anyway. So it would generate a low-resolution volume with the given symmetry, but without any impacts from the out-of-plane tilt searches. This may help prevent out-of-plane tilt falling apart I guess?

  • Using the averaged power spectrum, do you have a set of symmetry candidates that you’re enforcing during the refinement?
    • If so, what was the extraction box distance and what is the helical symmetry order enforced? You can see the latter by checking the refinement, which should print out “Refining with helical symmetry order of __”; this is the number of times each image is used for backprojection.

Yes, I have a long list of possible helical symmetries. The box shift is ~50Å (so I can have a smaller dataset to test them first), and the maximum possible helical rise is ~8Å. So there shouldn’t be any problem with that.

Thanks for the suggestion, I tested it but it didn’t really solve this issue.