2D classes much better in Live than classic?


After initial processing in Cryosparc Live, I am performing some manual processing in normal Cryosparc. Strangely, my 2D classes worked much better in Live than in normal CS, even using the same particle stack as input. I’ve tried to copy the conditions for the Live 2D classification (number of classes, circular mask) but I still am not able to reproduce the results from Live.

Has anyone encountered this before?


Live (usually) starts with a small subset of the particles and works on them to convergence, then adds the rest of the stack. As such, if you’ve got nice classes from the early particles you can find a lot of junk hiding in the “good” classes sometimes. What does a round of 2D look like from the good classes from Live?

1 Like

Live sets batchsize to 500 automatically, if I recall correctly. Try regular 2D but increase batchsize and guarantee better results with trade off of runtime.

Thanks for the suggestions. Ive tried both subsequent 2D on the good classes from live as well as increasing the batch size to 500, but I still am not able to get anywhere near the results from Live. I don’t think the quality of the data declined during collection. Using a curated subset of movies doesn’t seem to help either.

Can you try to select the nice looking particle and run another round of 2D classifications?
I sometimes have similar observations and often after 2-3 rounds of 2D classifications, I will get the same nice classes as live output.

Similar observations here, possibly useful discussion:

Thanks for the suggestions and for linking this in with the existing discussion. I found that re-extracting the particles from Live using refined shifts gave me satisfactory results in subsequent 2D.

Hi @Ablakely,

Thanks for posting this! We have a couple questions:

  1. Can you provide some more information on your sample? Is it a membrane/soluble protein? Is there a nucleic acid component in the complex? What is the MW?
  2. What version of CS and CS-live are you currently using?
  3. Would you be interested in sharing this dataset with us for some internal testing? If so, please send me a DM and I will get you information for transferring the data to us.

All the best,