Datasets that benefit from 3D classification without alignment

Features and Functionality
#1

Hi everyone,

We’re in the process of improving 3D classification in cryoSPARC. Initially, we are working on implementing 3D classification without alignment. As part of our internal testing, we’re looking for datasets where cryoSPARC’s heterogeneous refinement has not performed well, but where 3D classification without alignment has proven to be a useful step.

We would appreciate any pointers! Please feel free to also DM me if you’d like to (confidentially) share data.

Thanks!
Valentin

14 Likes
#2

Awesome. Hope this long awaited feature will finally be available in cryosparc.
If not planned yet, I would emphasize to include the option to use a mask for focused classification.

6 Likes
#3

Hi @vperetroukhin this would benefit many datasets!! Focused classification without alignments is an essential tool. There are many possible test datasets, but EMPIAR-10256 might be a good one (paper: https://www.pnas.org/content/116/18/8869) as it is a case including both pseudosymmetry and partial binding of a small, relatively flexible protein binding partner, where focused classification without alignments was used.

4 Likes
#4

I would say this is a great place to start:

Extensive focussed classification without alignment, and with different T values to yield a consensus reconstruction that would have been impossible (or at least improbable) with conventional 3D classification.

Additionally, this dataset might be useful:

Focussed classification without alignment of symmetry-expanded icosahedral particles.

Hope this helps!

4 Likes
#5

Thank you @olibclarke, @JackBravo, @tarek for the feedback and suggestions!

#6

@vperetroukhin One point I want to bring up regarding improving 3D classification in cryoSPARC is the issue of forming donuts instead of closed ‘spherical’ particles for icosahedral capsids. We were never able to resolve this (except by enforcing I symmetry which is of course not ideal for 3D classification). For more than 20 icosahedral capsid samples, none of the suggestions made in this forum prevented donut formation. Curiously, Relion 3D class with standard parameters seems to have no issues here and always gives us nicely closed capsids.

Best,
twg

2 Likes
#7

In general 3D classification in cryosparc should be improved. The benefit of non-alignment 3D classification might be specific to some datasets, but would be great to have this option in cryosparc too to finally eliminate the need to export particles and try this out in relion. Here is a paper where they explicitly indicate that non-alignment 3D classification did not work better than standard 3D classification (for GPCR-G-protein complexes):
https://www.nature.com/articles/s41467-021-24650-3#Sec13
“The particle stack was then further cleaned through a 3D classification with 5 classes and small local searches (same search steps as in the last iteration of the previous 3D auto-refine), which in our experience works better than no-alignment 3D classification for sorting-out low-quality particles.”. .

1 Like
#8

yes, the ideal would be to have a multi-class implementation of local refinement, where searches could be set arbitrarily small or zero

1 Like
#9

I am using 3d classification skipping alignment especially for data sets with preferred orientation, as I see splitting based on views when using alignment.

Also, I often tried to use heterogenous refinement to re-assign particles after identifying sub-classes, as I am interested in performing some sort of statistics on particle populations. In my experience, it was quite too inaccurate for this though. Usually, i found redundant classes down the road e.g. using 3D variability again.

I think it would be pretty cool to combine the potential to use heterogenous templates with the skipping alignment feature and masks. As pointed out, it would be even better, if you could set only small local searches.

1 Like
#10

every dataset with symmetry and asymmetrically associated ligand.

#11

Thanks again everyone for the suggestions!

FYI: A new 3D Classification (BETA) job is now included in cryoSPARC v3.3. The tutorial includes some results on the EMPIAR-10256 data (thank you for the suggestion @olibclarke!). We’ll try and get to more of these datasets as we improve the job in future releases. Looking forward to hearing feedback :slight_smile:.

Valentin

5 Likes
#12

Hi Valentin,

Its great to see that cryosparc can now do this! I’m curious: does this new 3D classification without alignments work with helical/filament structures? I’m curious as I have a helical map without helical symmetry imposed (due to the way I picked and extracted) and I’m hoping that maybe I can get away with using this new feature to see if I can find better aligned particles to improve the map.

Also can local refinements be used as inputs?

Thanks!

Thomas

#13

Hi @TMcCorvie,

Great questions – 3D Classification won’t impose symmetry internally, but you can do symmetry expansion on the particles beforehand to have the same effect. (If you have a helical map without symmetry imposed, you wouldn’t need to worry about this though). Local refinement inputs should work as well!

Best,
Michael

2 Likes
#14

This looks great - can’t wait to try it out! If 100-class 3D classifications become common, might be worth building in some kind of tool for classifying/clustering the classes! :laughing: - maybe like the “Rebalance 2D classes” tool but for 3D…?

Would also be great eventually to have a focused classification tool with alignment… kind of like a combo of heterogeneous and local refinement, with arbitrarily small orientational searches. This would allow for maximum flexibility, and sometimes works better in cases where classification without alignments doesn’t give optimal results.

1 Like