Motion corrected micrographs from Relion 3.1.3

Hi there,
We are analysing some data collected through EPU group AFIS strategy using a Falcon 4 detector (EERs). I have created the file and did Motion Correction on Relion 3.1.3. Then I imported the Motion Corrected .mrc micrographs to CS. I labelled the Relion’s Motion Corrected Micrographs’ directory as the movie path and used the *.mrc wild card. Then I did Patch CTF Estimation (multi) on CS and keep processing as usual. Is that a reasonable approach with this EPU_AFIS strategy?

During Inspection of Blob picks, I noticed the power threshold went only upto 140. Also, here are 2 examples of CTF estimation with CS and Relion (CTFFind), respectively:

CTF_CS_After motioncorr with relion


We are still using CS 3.3.1. Experts, please suggest.

Forhad, UQ.

Hi Forhad,

Yes, this is a reasonable strategy.

The display of the power spectrum on motion corrected mics is some kind of display bug, I reported it a while back but it doesn’t seem to affect the quality of the underlying CTF fits (just display of the PS). What do the 1D plots & estimated CTF fit resolutions look like?


Hi @olibclarke,

Thanks for your reply. This is one of the representative 1D plots:

This is the overview of CTF fit res. During Manually Curate Exposure job, I limited them by roughly 10 angs.

Any critical comment?


they look on the lower res side, but not inconsistent with the PS you showed from CTFFIND. I would select and double check the ones with estimated fits of >10Å, and see why they are failing.

Hi @olibclarke,

Here is an example of a mic with an estimated fit of 33.72 angs.

Hi @sdawood @olibclarke

Sorry for the repeated knock. Actually the problem persists with this dataset. The 3D volumes are showing a res of 5.68 on fsc plot, but the volume is super bloby and I can’t see any details. After I refined the 2D classes, res dropped. So, I suspect, I didn’t import corrected mics from Relion properly.
I checked few info from metadata to compare. Average defocus and CTF res fit is very close after CTF estimation with CS and relion (for example, ave. def is 11837 on CS, 11820 on relion; CTF fit is 7.82 on CS, 7.31 on relion, on a specific micrograph). The significant difference was in astigmatism. For this specific mic, it’s 119 on CS, but 365.54 on relion. Checked on a few, it was similar.

Please suggest what would be a better approach? This dataset was collected on a Glacios.

Thank you.

Looking at the one mic you posted, I would maybe try limiting the lower resolution used for CTF fitting - it looks like there is an issue with fitting the first 0 in the PS, but the fit is decent beyond that - so maybe try limiting res from 20-5 or 15-5

Hi Oli,
Thanks. I am trying that now. Just a quick thing. My max res was 4 A. Should be okay, right?

However, I was worried about the astigmatism, as the data was collected using beam shift. Relion creates a few other files per corrected movies, i.e. .eps, .log, .star. Should the info from those file needed to be incorporated into CS import? If yes, then what’s the best practice? PyEM or something? Please advise… I can’t sleep :sleepy:
Bunch of thanks.

No, you don’t need anything else, the micrographs are fine.

If you collected with image shift without hardware coma correction, you may want to split into image shift groups later in processing, but that shouldn’t affect the initial phases you are describing here.

Yes, max res of 4 should be fine, unless you have a lot of thick and/or crystalline ice, in which case steering clear of the water ring may give better results.

Hi @olibclarke

So I have tried to limit the lower res to both 15 and 20. 15 is yuk so let’s not talk about that. Here’s a comparison between 25 and 20. To my rookie eyes, there’s no significant difference, but I would love to get your opinion on this.

With limiting lower res to 20 Angs:

With limiting lower res to 25 Angs:

Agree there is minimal difference between these. I would just check whether changing the res limit “rescued” some of the mics that previously had a very low resolution estimate for CTF fit, e.g. the 33.72 Å one you posted above.

Hi @olibclarke
Apologies for the late - break, sickness, interruptions, conferences - many things to blame.

Lets get back to where we left. After limiting the lower resolution limit 20 A, the CTF fit resolutions were better:

Compared to 25 A:

However, that particular example of 33.72 CTF fit was now showing a fit of ~10, with a super-weird ps:

I don’t know what’s happening!

I kept processing with the exposures with CTS estimation done with lower limit of 25 A. With CS, when I’m limiting the resolution limit to very high (something around 3 A), I’m getting an overfitted fsc. My maps don’t show any secondary feature, it’s amplifying all the noises and look very spiky. So, based on what you’ve said, we must split them into image shift groups or something. I’m not sure if hardware coma correction was done as we colleceted this data from a different source, but I can check. Should I raise a different discussion with this issue? @sdawood


Please could you share the FSC and the preceding 2D slice plots?

Hi @wtempel

This is when I ran ab-initio with all the default parameters and 3 classes (I used around 352 k particles), and NU refine with all default parameters:


I ran another ab-initio with same particle stack and 2 classes; however, I changed the maximum resolution to 3 A and initial resolution to 6 A. Then ran NU refine with extra final pass - 5, maximum align res, initial low pass and GSFSC split resolution all to 3 A. I increased the batchsize SNR factor to 200. Resolution improved, but the map doesn’t show any secondary feature, quite spiky though.


I’m happy to share the map as well if we’re allowed to.

We had a high reslotion template for the same protein, which I used for refinement with the same particle stack. Resolution is still good on fsc, but the map quality didn’t improve.


@forhad - what do 2D classes look like for these particles?

Hi @olibclarke

Thanks for getting back. The 2d classes I mentioned above having ~352 k particles are as follows:
(I did a slight mistake here. I ran this with clamp solvent, but without non-negativity. Later, I did with both ticked on, but result wasn’t great)

Later, I refined further. These selected ones have ~100 k; however, map didn’t improve. In this case, I ran without clamp solvent and non-negativity. In all the cases, I normally started with 40/50 iterations and batchsize/class 200. But I increased iterations to 60 (with 3 final iterations) and batchsize/class 400 during next rounds where I link good classes for further classification.

One more thing @olibclarke. Do you think exposures with CTF estimation limiting to 20 A would have been better to use? Or, its about the optical groups? Haven’t heard back from the people who collected about hardware coma correction. Is there any way that we can detect from the data?


These look like rings that are picked a bit off center - does that track with what you see on the micrographs? Or are they more C-like? If they are actually rings, then improving picking may yield better results (denoising the mics may help to identify whether this is the case)

Yeah, it does. They are basically kind of continuous oligo, which doesn’t form a complete ring. The nature of the curvature you can see on the 2D slices above.

Truly so. So I focused to pick the central core of the oligo based on a diameter of 200 A I used from the previous model as a starting guide.

Does that mean I denoise and train using topaz? I have tried, but courtesy of my “Beginner individuality”, I failed to follow the CS tutorial on topaz picking (none of colleagues is familiar as well unfortunately). Could you suggest any tutorial on topaz picking? I will try it next.

Thanks @olibclarke for your continuous support.


Hi @forhad,

If they are truly continuous, with variable curvature, then you might try treating them more like filaments - picking segments all along the “C”. This could help in the event that variable curvature & conformational flexibility are precluding high resolution reconstruction.

Re denoising, I was suggesting this to better visualise your particles - not necessarily as a precursor to Topaz picking, though you could certainly use it this way. As a brief tutorial for Topaz, I would suggest sorting your mics by the number of particles remaining after classification (in the previous round of picking), then selecting maybe the top 100-500 to use for training Topaz - enough to get ~2000 good particles (the important thing is the quality of the initial picks used for training, so it is worth taking time to inspect and curate these manually). Then run Topaz train, and extract from a different set of mics (not the ones used for training) to test the quality of your Topaz model and optimize extraction parameters. Does that make sense? Where were you running into trouble with the tutorial?


I had 2 problems with that, and a question.
P1: If I look into small segments, they are really small. This protein as a monomer is less than 20 kDa (and of course, this is why I took 200 A box to pick initially as that would ideally cover 8 to 12 subunits). This is a representative micrograph, (You can clearly see “C” from one angle, ladder from another. Sometimes you feel they are ring, but that’s actually 2 "C"s sitting together).

P2: I have tried filament tracer, but due to that curvature, 2D classes were yuk.

Now, my question is, can any job in CS actually do this? Outside CS, e2boxer I know would be an option which is arduous. I don’t know if I can make anything that would autopick along my oligos. I have more than 4000 movies, so manual picking is near impossible.
Besides, with what I have now, should I just try a helical refinement instead of homogeneous/NU?

Thanks for the suggestion on Topaz.

Should that be enough for training?

Well I tried to train Topaz with 20 good micrographs, then used Topaz extract job using that topaz_model. Then extracted and ran 2D. It could extract only ~200k ptcls and 2D classes were really bad.
I was then trying to denoise using the process described here, but kinda got lost :dizzy_face:

This is I reckon I have been failing to do. I am trying to develop in this regard.

Thanks for suggestions @olibclarke