I have imaged and have maps of a membrane protein under 2 conditions (apo and +ligand). Each subunit of my protein has 2 domains (A and B). Under condition 1 the map is clearly symmetric but in condition 2, the domain A is symmetric between the 2 subunits and good resolution (3-3.5) but subunit B does not appear to be (low resolution and is only visible in one subunit of the dimer). May be in condition 2 there is a mixture of states (say one in which domain B is symmetric and another where it truly isn’t) that I have not been able to separate as yet. I was hoping to get advice on strategies to better classify/perform reconstruction on cryosparc. I have ~1.1x10^6 particles so I think I have room to play here. My ideas were:
(1) Ab initio reconstruction with multiple classes:
- test with different values of class similarity scores. (I have thus far only used default value 0.1. Has
anyone tried other values which gave better classification?)
- test maximum alignment resolution. (I have only ever used the default value of 12 but may be I can
use lower like 6? My current overall map resolution goes to around 3.9A currently)
(2) Heterogenous refinement with C2 enforced, then take each class and do homogenous/NU/local refinements with C1 to test whether C2 features are retained in the various classes.
(3) If I do symmetry expansion and with the expanded set of particles a local refinement, then only the symmetric domain is expected to get better, right? How could I resolve the unsymmetric domain B better? Subtract density of the symmetric domain and then classification (via ab initio recon and heterogenous ref) with particles of subtracted densities?
It would be great to get suggestions from more experienced users.
Hi schow, I think you mention some good strategies. I would try ab initio into multiple classes with increased maximum resolution. Perhaps try 4.5A, 5A, and 6A and view the results after a C1 homo refine. You could certainly decrease the starting resolution to ~12 or so.
I think C2 symmetry expansion followed by 3D classification without alignments in relion (C1) is a good strategy.
Also, have you tried hetero refinement without symmetry?
@user123 Yes I have tried hetero-refinement with C1 but that didn’t help. However I suddenly think if I set the max alignment resolution to ~8A or so might also be a viable option to try? (May be this will prevent the classification based on just the high resolution features of the map (symmetric domain) and might better classify based on the lower/heterogenous regions?)
Symmetry expansion followed by 3D classification without alignment on Relion is a great idea! Thank you.
How did your heterogeneous refinement results look? Are all the classes highly similar? Sorry, I can’t make useful recommendations on settings here. Perhaps you could also try fourier cropping such that nyquist is 4 to 6 A. I haven’t had great results separating distinct classes with hetero ref but my differences are very small.
I definitely get my best separation results with Class3D w/o alignments with a mask on the most dissimilar region. You can try putting a tight mask on the jumbled region of one protomer, and possibly play with including some of the nearby well-structured region.
You could also try symmetry expansion, then 3DVA with a mask on the jumbled region. Or try a global mask to identify areas undergoing conformational change then classify with a mask on that region in Class3D. Once you have separation, remove duplicates (symmetry unexpand) and refine in C1.
@user123 Heterogenous refinement with C2 enforced is working! But I’ve had to play around with # of classes + number of times I repeat each round of heterogenous refinement. I’ve tried 3, 4, 5 classes. In each case, I take the output volumes of each run of the refinement as input volumes for the next run. 3 runs with 4 classes has worked better than 3 runs with 3 classes. Currently on the 3rd with 5 classes. (Each time, I get 2 classes with distinct conformational differences, and after a homogenous refinement with C1, the maps stay C2. There are more conformations still I think that I am not sure if I will be able to resolve/push to high resolution with the current data but this is at least a very good start).
Very cool, glad you found a strategy that works! I wonder if instead of multiple rounds, you could increase the number of classes, or increase the batch size per class (maybe 3000 or 5000?) and/or increase the number of final full iterations.
I second everything said by @user123! I think especially if this membrane protein is small, and the asymmetric domain B is small, aligning all particles to a common reference and then trying local classification w/o alignment with a mask on the domain B (or whatever is the dissimilar region) would be a good strategy – this way you can avoid the issue where there isn’t enough signal to fit the euler angles, and hopefully the classification can pick out distinct conformations. You can also try a similar procedure using the cluster mode of 3DVA.
I wouldn’t recommend moving onto local refinement until you are happy with separating out the distinct states via local classification – local refinement really does assume that there is one structure in the images, just with two separate regions with limited degrees of freedom of motion between them.
A related question and feature request. A symmetric protein with C5 has a single asymmetric protein bound on the C5 axis resulting in pseudosymmetry. Local refinement with a single masked unit requires very large angles and shifts to align properly the minor differences. Could local refinement include an option to perform the large rotate and shift about the symmetry axis and then finer rotation and shift for subsequent alignments? That way all particles will align properly in C1.
Hi @MHB - If I am understanding correctly, I think what you are requesting is “symmetry-relaxed” refinement - essentially multimodal local search, where the region around all of the (pseudo)-symmetrically related orientations for each particle is searched. I am all in favor, and have requested the same in the past. I believe some implementation of it is on the roadmap.
In the meantime, the best way I have found to deal with this is 3D classification without alignments, after refinement in the higher symmetry (with or without symmetry expansion). If the 5 orientations of your asymmetric guest can be separated, they can then be rotated appropriately to match each other using Volume Alignment Tools, and then combined for local refinement in C1.
Yes correct, symmetry-relaxed refinment would be ideal. The assymetric “guest” is quite small and 3D classification has not yielded great results so far. Maybe will attempt volume alignment with several alighnments followed by local refine. thanks
The other option I would consider is to export the particle set after C5 refinement, and try refinement in Relion starting from local angles, with symmetry relaxation enabled (maybe with the use of the SIDESPLITTER plugin if it is a membrane protein).
For 3D classification in cryoSPARC, in similar cases we have had success increasing the batch size (up to 10000) and epochs (to 50), and trying with or without “force hard classification” switched on, as well as varying the initial learning rate and expected class similarity.