I am working on a protein complex with a Fab fragment attached, the whole particle is ~100kDa, in a T shape, with a major diameter of ~100A and a minor diameter of ~80A. It is collected with a 300kV Krios microscope, equipped with a K3 camera. The pixel size is 0.65 A/pix.
In the beginning, I tested with the blob picker and extracted the particles (576 pix to 288 pix) with bin 2. I got some reasonable 2D classes (despite that it’s hard to see the secondary structure). I ran an ab initio reconstruction with 3 classes and got one ab initio structure that looks good to me. However, when I proceeded to homogeneous refinement, the resolution didn’t improve, and I could not tell any secondary structure from this. When I tried to align it with the crystal structure of a component protein, the map didn’t fit well. I then looked back to my picking and my 2D classes, where I had the following questions:
Does my micrograph look good in terms of having many particles?
Does my current 2D classes look good to proceed or do I need to spend more time on the picking?
As a new user, I can only post one picture in my post. I will post the micrograph pictures in the follow-up reply post. I am now trying on TopaZ picking, I have tried template picker with my ‘good’ 2D classes but it seems like just randomly picking everything on the micrograph (it’s picking more particles than the blob picking!).
Any suggestions or discussion on this will be useful and appreciated.
Thanks,
Amy.
Use the 2D class to run template picking definitely. You could benefit from smaller box maybe 75% at least in early stages (can return to large box later). Hard to tell what’s going on in micrograph without scale. on the side, 2D select your favorite one 2D class then “inspect particle picks” and you will see the relevant picks on the micrograph to determine if they are prevalent relative to noise.
Agree 100% with @CryoEM2 . @Amy you can check this link for some useful insights: CryoSPARC Workflow Discussion: Comparing Different Processing Orders - #3 by zhe
From experience, flexible proteins demand some adaptation to the standard protocol of particle selection. Try to focus on 3D instead of 2D (hetero refinements, 3DVA, 3DFlex, 3D classification). Of course optimising the picking is important as it is expected to inflate the initial set, and if you want to reach reasonable resolution, you’ll probably need a lot of particles.
Thanks so much. I have a look at the 2D classes in ‘inspect particle picks’, and it seems like on each micrograph, the number of picked particles is low: the maximum picks is 25 particles per micrograph.
The diameter is 100A for each green circle. The micrograph is 20A low pass filtered.
It’s hard for me to recognize the particles, but I can see some reasonable shapes. This gives me worry that I may have excluded some particles during 2D classification rounds, or that I used a too-low concentration for the sample preparation.
What is your opinion on this? I am re-extracting with a smaller box size and let’s see what that will give me.
Many thank,
Amy.
this is normal, maybe even better than normal for Blot Picker. Agree there are believable shapes, this makes me think there are hundreds of “particles” per micrograph, and the template (even better is topaz) picking will find them no problem and more importantly will pick them right in the center and improve the 2D class. Agree for something this small the 2D class should be used mostly to 1) remove non-particles and egregious contrast outliers 2) characterize what you have and whether your picker is working well 3) identify orientation populations and find new templates for template picking to improve orientation distribution. Likely 2D class should not be used to filter the particle set to a tiny, high quality set. Use heterogeneous refinement for that. Ab initio with 10 particles in one case, and ab initio with as many as possible particles and from several orientations if possible, 2D selecting for best. Run het refine with 4x first junk class and 1x second hopefully particle-like class. you can iterate this process with the full or ~full dataset as you improve the good model and use it as reference. This is the best way imo to identify and maintain all the good particles and rare views. The het refine can do this and well, but might require a lot of particles and a lot of attempts with varying reference volumes. When you are happy with quantity/quality of the one good class, can proceed to NU-refine and 3D class/3D var etc.
And yes you can probably use 2-4x more concentration at the next sample prep.
Have you tried Micrograph Denoiser (with a model trained for your data)? If not I would give it a go - may make micrograph interpretation easier, and blob picker more accurate.
Thank you so much, that’s very useful. I am now trying the TopaZ training, one question coming to my mind is: do you use a small subset of particles & micrographs to do the training and then use the model to pick the whole dataset? Or I can just use my current good 2D classes (which are picked from the whole micrograph) to train and pick again on the whole micrographs?
Mank thanks,
Amy
Hi Oli,
Thank you for your suggestion. I am now trying to use the micrograph denoiser, but one problem I am facing is that the denoised micrograph can not be displayed in the ‘inspect pick’. I also tried ‘Curate Exposure Manually’ to see the denoised micrograph but again the job runs with no display of the micrograph. Do you have any idea why this is happening?
Thanks,
Amy.
Hi Amy, there is a dropdown menu near the micrograph that allows you to toggle whether to display denoised micrographs - it is not super obvious, but now that you know it is there hopefully you’ll be able to find it?
Based on the posts in the community, seeing an increase in the test-average-precision over epochs curve is good; however, the value is very low, ~0.042, indicating that the model is only picking very few of my ‘label’ particles, which is concerning. This is the model trained with my good 2D classes.
I then perform 2D classification with a small subset and use the 2D classes from there to train again. This is the new model:
In my opinion, the first model looks better than the second, but neither of them is very convincing for me to take them for picking.
May I ask for your opinion?
Agreed first is better, I typically use 1000-5000 great particles to train if possible after stringent 2D cleaning. I’ve had no problems with precisions as low as 0.0012, so as long as the precision improves over epochs then I wouldn’t be worried. The proof is in the pudding, if you topaz extract then extract then 2D all topaz picked particles, is it better than the blob picker? Can then select the best of those and iterate to build a model that truly knows what you’re looking for. Beware, regular selection and picking of the “best” single view will drive anisotropy in the dataset. As new classes appear or if you get orientation distribution in 3D het refinements, would train a few models on unique orientations.
Is it possible to do with the input I gave it? I just drag the denoised micrograph from the denoised job into the ‘curate exposure manually’.
Many thanks,
Amy
I’ve seen it before, seems to happen randomly* though and is absolutely infuriating when it does crop up.
* i.e., I thought I knew what was triggering it (doing denoise post-Exposure Curation) but I have a dataset from last week which that workflow does not trigger the issue. So I’m back to scratching my head again about it!
