T Value equivalent in cryoSPARC?

3D Classification
1

Hello,

a common strategy for sorting particles in Relion is to use 3D classification without alignment and then vary the T parameter, often to extreme values like 200 or more. This can help sort bound cofactors, etc. but can also be useful to find a small subset of particles that refine to a higher resolution than the initial refinement.

In my experience, I can get cryoSPARC 3D classification to sort heterogeneity by using focus masks or changing the filtration parameters, but I can never get it to find a high res subset of particles. Also, the ā€œfailureā€ rate is higher.

What are some strategies for modifying the 3D classification parameters? What parameters would most closely mimic the T value that Relion uses?

Thanks!

-Rick

1 Like
2

Indirect answer, I do het refine with high res model, low res model, and junk class, set initial lowpass to 8, 90% of particles go to the good class and subsequent refinements are better.

1 Like
3

awesome, Iā€˜ll try that!

4

Not sure if there is a ā€œregularizationā€ equivalent parameter in cryoSPARC 3D classification (maybe someone at cryoSPARC can comment). But some strategies Iā€™ve found incredibly productive to identify higher res subsets, or to sort based on finer signal features (for which I used to use relion 3D with high T value; but find better separation with much less overfitting in cryoSPARC recently ):

  • increase class3D_max_final_iters = something high (25, sometimes I go up to 150 for minute signal differences)
  • class3D_target_res = under 4 angstrom or something
  • class3D_oem_batch_size = 20,000 (or higher)
  • class3D_online_em_lr_init = 1 (Iā€™ve read this can be tuned to achieve optimal results on case by case basis)

And finally, something sort of unconventional: I try a few different k-values, see which are productive in giving cleanest volumes. Then perform replicates (n=3-4) for each classification condition, merge and remove duplicates for the best class per replicate/condition. This seems to be pretty powerful in getting higher res - think it compensates for some class uncertainty inherent to the low SNR behavior of SPA, by retaining/rescuing good particles that may get spuriously classified into junk bins.

4 Likes
5

Amazing. Thank you so much for the detailed response. This is exactly what I was hoping for. It is very much appreciated. Iā€™m off to try these!

6

Also - convergence criteria to 0%, rms change check off, to force it to run through user defined iteration number

1 Like