3D classification O-EM epochs

Hi everyone,
I have few questions regarding the 3D classification parameters used to handle continuous flexibility:

I don’t find a definition of O-EM epochs, can someone please share it?

How does Number of O-EM epochs and Maximum number of F-EM iters influence the classification toward separation? is the default F-EM iters 25 iters too much not enough, what do people use? lower than 25 or higher?

I’m used to do 3D classification without alignment in Relion playing with 3 parameters: nb of classes, tau T, and nb iterations;

What are the most significant parameters in cryoSparc that can push the classification to separate particles into classes when you are dealing with a continuous flexibility of an enzyme?

Best Regards

Hi @Proteino,

3D classification in cryoSPARC uses an ‘online’ (i.e., mini-batch) expectation maximization (O-EM) procedure to alternate between computing most likely class assignments and computing the class volumes given those assignments. O-EM epochs refers to the amount of times the O-EM procedure will iterate through the entire dataset.

After the O-EM epochs are complete, 3D classification uses a number of ‘full’ EM iterations to do the same procedure but now with the entire dataset – this hopefully ‘cleans up’ any remaining class uncertainty and assigns particles to a single class. I believe you are referring to the maximum F-EM iteration parameter – this is there to ensure the procedure does stop at some point, but in general, F-EM should stop based on the automatic convergence criterion (default: less than 2% of the particles switched classes).

The 3D classification guide page goes over some of the salient parameters – a few that we often play with internally are number of classes, Target resolution, Force hard classification, and Per-particle scale. I’m sure others can chime in here with other salient parameters!

If your enzyme is relatively small, you may also want to take a look at some tips we put together for membrane proteins here. For continuous flexibility, 3D variability analysis and/or 3D Flexible refinement will also be useful to try if you haven’t already.



Hi Valentin, thank you for your feedback. I use 8A as a target resolution and +/- force hard classification which didn’t seem to make a big difference. I haven’t changed the Per-particle scale parameter.

I’m also using 3D variability analysis 5 modes and 8 A target resolution, displayed in a cluster of 20 or 100 classes, so far no luck yet. I’m dealing with 2 subunit enzyme (100KDa) bound to another protein DNA complex that is rigid (300 kDa)

In 3D classification and 3D variability I’m using a focused mask that englobes the enzyme.

Thanks again.

No problem! For posterity – I also want to note @CryoEM2 's comment here: 3D classification error - #5 by CryoEM2 – reducing Class Similarity can also be helpful!

1 Like