Thank you! You’re absolutely right. My apologies. Here some example pictures, together with the orientation diagnostic. I put three views of the map at two different threshold levels. The nominal resolution is 3.12Å.
Best
Grand. No apology necessary. 
Those 2D classes look beautiful! But not what I expect from a membrane protein - no detergent belt?
Immediately that map looks severely anisotropic so I think you’re on the right path. The orientation distribution (relative signal/Fourier sampling) indicates a moderate preferred orientation also.
When you say you rebalanced the particles, what job did you use? Take the output from the NU-refine and use “Rebalance Orientations” rather than rebalance 2D, perhaps try a cutoff of 90% (to get the best result you might need to play with this, but I’ve usually got good results from around there) then try that cleaned set with a new NU refine. Then you can experiment with the dumped particles to see whether any more can be recovered via further 2D/NU-refine work.
Also looks like you’re leaving a lot of potential particles on the cutting room floor (so to speak) if the left micrograph is all the particles you extracted from that micrograph. There might be some more orientation variation hiding in the denser areas if the aggregation can be cleaned out.
Thanks! I also had the impression that those 2D class were good.
The protein is embedded into nanodiscs and I also expected the detergent/disc belt around the protein. For other proteins I can clearly see it, for this one not. Dunno what to think about this. In some 2D class there’s a disc shape without any other feauture, so I think they are empty nanodiscs. I’ve tried to include those views into the 3D reconstruction, but they always ended up in a separate class without features.
I did exactly what you mentioned: rebalance orientation with 75% cutoff and NU-refine again. The map I showed comes from this new NU-refinement. I will try with the dumped particles and see what happens. In the meantime if you have any other suggestion I’m here to listen well 
No, those are only particles examples from the micrographs. There are much more picked from each.
Thanks a lot!
The absence of a clear nanodisc density is surprising. Do you have other biochemical evidence that nanodisc incorporation was successful? E.g. co-elution of the MSP, clear MW shift on gel filtration or mass photometry?
Agree with @rbs_sci re diagnosis - clearly anisotropic. One other consideration is symmetry. Did you apply symmetry? From your 2D classes it looks like there is at least a C2 axis present - if you haven’t tried taking advantage of this then it may be worth doing.
The absence of the nanodisc puzzled me, too. However, I’m sure of the nanodisc incorporation. I have evidence from gel filtration and sds-page that shows the MSP. Besides, I’ve already incorporated the same protein into other nanodiscs and in that case the nanodisc density was very clear. This time I just changed the lipid composition. It might be interesting to know what happened. I just need to solve the structure
Yes, it should have C2 symmetry, which I did apply (actually I tried with both C1 and C2) either in Ab initio or refinement, or both.
Any idea on how to solve the anisotropy besides going to square one and freeze the grids differently?
Many thanks to all of you!
make a 2D selection completely devoid of “front” views and ab initio/refine it. then do het refine with that map, this anisotropic map, and a few junk classes. if it succeeds in generating a good class, a preferred view class, and separating junk, then repeat until desired.
you could pick 10x more particles than shown above, I assume this is after cleanup?
take 1 of your rare views, 2D and 2Dselect it down to 1000 pristine particles, run topaz train with “expected particles” ~100. Do the same for another rare view. extract and 2D, you should get more rares to add to your full particle set.
Sometimes, 3D classification without alignments will allow you to identify a class that is more isotropic.
Also, classification with Blush regularization in relion has been known to help identify classes with more isotropic features (e.g. Data-driven regularization lowers the size barrier of cryo-EM structure determination | Nature Methods & x.com).
You could also try spIsoNet (combined with refinement in relion): https://www.biorxiv.org/content/biorxiv/early/2024/04/14/2024.04.11.588921.full.pdf
Or CryoPROS: https://www.biorxiv.org/content/10.1101/2023.09.26.559492v1.full.pdf
But ultimately I think you will be better off trying to improve the sample & mitigate preferred orientation at the sample preparation (or possibly data collection) stages…
1 Like
Sorry for the slow response, but I just resurfaced from a vacation.
Thank you all for your very helpful suggestions!
I will try them all and I will post as soon as I have news.
Thanks! Trying right now.
Any suggestion on how to successfully run Topaz? I tried several times, but always picked much worse. Anything to change besides the “expected particles”? Leave everything else as default?
Best
-
The no.1 most important thing for topaz (or any NN-picker) is to provide good training data - well-centered picks on real particles, with minimal (ideally zero) false positives. 1-3k particles should do the trick, either picked manually or curated from a previous round of classification.
-
In Topaz extract, remember to adjust the extraction radius - this controls the minimal distance between adjacent picks. It is specified in downsampled pixels, so you will need to take that into account.
-
For downsampling during training, you want to choose your downsampled pixel size such that the entire particle fits within ~72 pixels.
-
In Topaz extract, you may need to tweak the extraction radius to get optimal results.
Adding: if you give expected total particles per micro, it will pick that many and include the dominant views.
If you give number of expected rares (~10x less) then it will pick only those.
Can topaz pick a few times with low numbers using slightly different “clean” inputs as oli described, then remove duplicates.
Maybe this is very stupid…but, how may I tell if a view is rare?
From the anisotropic map can we select the particles with rare views? Or from the 2Ds?
Given that my micrographs are crowded, would make sense re-extract the pristine particles with a box size equal to the mask used for 2D classification before running Topaz training?
A view is rare if it is not one of the over-represented ones from your 2D. Could also utilize this: Orientation Sorting Tools - #7 by robertstass?
From your 2D/3D above, the view that is most represented (at least in 3D alignment) is looking at the triangle. your “side” and “top” would look better if more particles contributed there. so in your 2d just take class 1,3,4,(maybe 5 seems most rare)7,8,(9 also very rare “top”) etc, and disregard 2,6,10.
Strangely, from the 2D above you have great distribution. So it could be a problem that your reference is anisotropic and therefore your particles fit best to the preferred view, exacerbating the effect. Or more likely, the 2D you showed is not the same particle set that yields the 3D shown.
If you take those 2D shown and run ab initio 1 class, 2 classes, 3 classes etc and look for any single good reference, I would expect that NU-refine (or het refine of multiple copies) of all particles in the shown 2d set would look great.
as for topaz - no need to re-extract pristine particles. it doesn’t “see” the 2D classes, it “sees” the particles which includes the background noise of near particles.
1 Like
I used this tool and it’s amazing. I was able to select some views I didn’t see before. They are apparently very rare, so I’m trying Topaz with those.
However, I think Topaz is not training well (see exemplar pic). What do you think could be the problem?
This is what I did:
- Used the ViewSelect tool to extract the most rare particles in different subsets (5 subsets, see pic).
- Used Curate Exposure to map back those particles to their micrographs
- Used Curate Exposure to pool together the micrographs corresponding to the 5 subsets
- Used those pooled micrographs (about 5k) to train Topaz. I run five separate instances of Topaz, one with each subset using the following parameters:
- 100 as expected particles
- auto downsampling based on 100Å diameter (downsampling used was 9) OR manual
downsampling to 4. But the results are comparable.
- All other parameters are default.
Maybe the number of particles needs to be reduces since for those rare views I was able to pick 2 or 3 particles per image on only about 1000 micrographs per rare view out of 30k?
Maybe there is no need to downsample since my pixel size is 0.83Å, therefore the whole particle will fit within 85Å which is close to what Oli said?
Sorry for the long post!
Why do you think this is not training well? How do the output picks from Topaz extract look?
Topaz extract is ongoing. However, Test-Average Precision reached a maximum of 0.05. The example given in the Topaz guide shows a value of 0.75 or so. Am I wrong?
Don’t worry about the absolute value of the test average precision - it is hard to compare between datasets (see elsewhere on the forum or the topaz github for an explanation of why). You just want to see it increasing, then plateauing. I would pay more attention to whether the output picks look sensible
EDIT: E.g. see here: Understanding topaz output - #2 by rposert
and here: Topaz Train: low value of test average-precision over epochs - #2 by alexjamesnoble
agreed, looks like good training and should be capable to pick more rares and bolster the dataset with “hidden” particles that were in your data but difficult to pick/identify. to alleviate your original problem of over-representation.
1 Like
On a different note - you might consider using rebalance orientations (& perhaps BLUSH regularization in relion) as described here: x.com
I have tried iterating single class ab initio & rebalance orientations (discarding paricles with worst alignments3d/error), and in some cases have seen substantial improvement in the quality (isotropy) of the initial model.
1 Like