Spherical 2D classes

Hey I am processing a dataset of a 140Ax140A particles that look like that:

The particles, look fine, I think. However if I run blob picker and subsequent 2D class averaging I end up with spherical class averages

The parameters I used were:
number of 2d classes: 100
Force max over poses/shifts:off
batchsize per class: 200
number of online iterations: 20
full iteration: 1.
Any tips, how I can get reasonable looking class averages?
Thank you

If you have Force/max off, it will take a lot longer to converge - usually 40, sometimes up to 80 iterations. Have you tried with Force/max on (the default) to compare?

yes, initially I did and they looked weird. I found the tip with switching off force max, hoping to get better classes

I will try to run averages with more iterations. So I increase the number of online iterations to 40 or 80 right?

Yes that’s right, although there are also other things you could try with force/max on - increasing batch size and increasing the number of final full iterations (to e.g. 10) can help

keep in mind 99% of your particles are in the early classes. the spheres are close to nothing. I bet if you take the first classes, use a smaller box, maybe 1000 per batchsize, maybe 40 classes instead of 100, you could pull out something. it’s also possible that what you picked/2D are not protein particles with defined physical shape. take class4 with 367k particles only, run 200 classes of those.

Thanks for your advise! I run 2D with force max over poses on; 10 full iterations, 20 online iterations, 10000 batch size. Now they look like that :

I think some are looking okay. Parallel to that, I picked roughly 2000 particles manually and averaged that, which looks like that :

Might be an option to use those classes after manual picking as templates for template picker?

Looks much better! I would pick these and do another round.

Looking at your micrograph, and the first two artefactual classes with the squiggly linear features, I would also double check gain correction, as removing those features may help you.


I ran a few rounds of 2D class averaging and ended up with the following classes, which I think look nice. It should be a membrane protein in DDM micelle boudn to a soluble protein.

Should I use that now for a round of template picking or rather try abinitio?

Was this a his tag purification by any chance? Expressed in E. coli? The top class looks very familiar, it reminds me a lot of complex 4, which is His rich and can co purify on IMAC…

In any case, I would def try a quick round of ab initio and see what you get, and I would also try optimizing your picking (either via template picking, or preferably training a neural network model, e.g. for Topaz).


(For comparison, here is a complex 4 side view in nanodisc, rotated to facilitate comparison):


[omitted], I can see the similarity. I will work on optimizing the picking. According to Masspec our proteins of interests are the most abundant. I hope that is the same in cryo. Thanks for all you tips. I will keep you updated!

I would also try ab initio with selected classes - e.g. to me these two look more like what you describe (membrane protein with bound soluble protein):