I’m having consistent issues with 2D classification, getting classes either with no features at all or with a blurred appearance (see pictures). This issue occurs regardless of GPU assignment (I noticed a fix recently was use of only 1 GPU), and with different parameter setups (though, granted, nothing too off-the-cuff). For information, I’m uploading a goodparticles.star file from WARP, which appears to work fine. This data is also a sample which has been collected before and previously I could get pretty high resolution 2D classes. Seems that this round of processing isn’t so smooth.
Any suggestions would be greatly appreciated!
Why do the 2D classes look like they have a square window? The first class (junk) looks OK, but the others have a distinct square shape…?
Try a tighter mask around the particles rather than just the windowing function? Increase particles per class used during 2D?
The bug regarding multiple GPUs was specific to CyoSPARC 4.2.0 without the emergency patch. If you’re running any other version of CryoSPARC, the bug doesn’t manifest.
Could you post your micrographs, or better, what your particle picks look like? It’s strange that it’s putting basically all of your particles in the first few classes when none of them seem to have any features.
Thanks for the feedback! Not sure about the square window - I actually have a 110 angstrom mask on the classification… so I would really not expect particles to look like this. I have around 700,000 particles total and asked for 250 classes, so would generally expect that to be enough per class…?
I am using 4.2.1 so I suppose multiple GPUs shouldn’t be a problem now.
Thanks for the reply! Sorry for delay. Particle stack examples are attached… I’d expect at least some bad 2D classes, but seem to get junk for each of the 250.
Hm. The square window effect has me a bit concerned; not sure why it is appearing. Might be worth checking the particle stack for corruption? The first class looks like believable junk… maybe select all the other classes and do another 2D run, and do a 2D run on the first class as well to confirm it’s mixed junk.
How much did you vary your settings?
Settings I’ve found seem to work pretty well for everything I’ve thrown at them are as follows:
Number of 2D classes: [adjust to taste]
Initial classification uncertainty factor: 5
Circular mask diameter (A): [adjust to taste]
Number of final full iterations: 3
Number of online-EM iterations: 25
Batchsize per class: 500 (sometimes use 1000 on 1M+ particles)
Enforce non-negativity: try both on and off
Use clamp-solvent to solve 2D classes: turn on if turning on non-negativity
Number of GPUs to parallelize: [system dependent]
Due to the higher uncertainly factor, it can take a while to show anything worthwhile, then all classes will look pretty good for a few iterations, before some get better and some become obviously full of junk.
When I see 2D classes like this, I will increase the number of iterations (try 40-50). It usually happens when force max over poses & shifts is OFF, instead of ON.