This results came from 200 classes, Initial classification uncertainty factor: 1, bathch size: 200, Use clamp-solvent to solve 2D classes: True.
I tested default options with 200 classes. But there was no difference.
Here are options what I changed in pre-processing.
Patch Motion correction
Maximum alignment resolution (A): 3
Patch CTF estimation (multi)
Amplitude Contrast (A) : 0.07
Use Template Picker
I picked my particles meticulously.
Extract from Micrographs with box size : 350, 384, 450, also binned to 256 from 450.
2.8~3.7 millions from 14000 micrographs
All results were similar regardless boxsizes.
Around 130~40 of 200 classes are looked like background with patterns. I could see smear parts of my particles in some classes.
In the 1st round of 2D classification, 70%~80% of overall particles are initally looked like that.
Is there anyone can give advices or validate my mistake?
Have you run another round of 2D on the particles from classes that are not obvious junk? You might just have a lot of junk diluting your protein and if so features can come out with removal, which might take a few rounds of 2D cleanup. I would choose all those classes that have possible particles and run another round of 2D. I would also bin more at this stage to save time, probably to 4 or 5 A.
I would not recommend using clamp solvent - it generally does not offer any benefit. You might try lowering the maximum alignment resolution (e.g. to 12 Å). I would also suggest in the initial round, with such a large particle stack, binning more aggressively (e.g. to ~4-8Å/pix) while cleaning up junk. Agree with Aaron re running multiple rounds of 2D. I would also try switching off Force/Max over poses shifts, and increasing the number of iterations to 40.
Thanks for all your suggestions.
My current 2D classification is not working for switching off Force/Max over poses shifts with 700k particles and 450 boxsize and 100 classes. I may need to reduce particle numbers and classes to reduce GPU memory usage or binning boxsize to smaller. The My data set is too big, doesn’t allow that option off (memory error). So after some optimizing trials, I will update results.
Thank you very much.
I agree with “optimizing trials”. Make a small selection of 100k particles at random (very easy to do with particle sets tool). Try a variety of parameters (I think extraction box size is a very important one here… you can see the images of single particle extracted occasionally throughout that job, does the particle take up ~75% of the box or is there tons of extra space around the tiny particles?) and then you can apply these parameters with confidence to your very large particle stack. With 100k particles the jobs should run in a matter of minutes and it will be very easy to find suitable options. My guess is your box is too large, remember it is multiplied by pixel size. Of course if your particle prefers thick ice it is probably large but this is something to check first.
The reason I decided bigger box is the calculation you mentioned before.
So If my molecule is 200A, aiming resolution is 3A, Defocus range is 2.5um, Box size is closed to 500. Also when I measured length ratio in screen, box size 384 showed 1(box):0.6(molecule) length ratio. Some 2D class touched edge of box. When I increased box size to 450. ratio was 1:0.5. All 2D class were well centered without touching edge. Am I wrong?
Thank you very much.