I am working with a small complex (~100kDa) with a low SNR particles. Aside from the low contrast, the complex shows flexibility. Old posts helped me with the 2D/3D classifications and got a subset of particles that look promising. My main issue now is to get the best resolution from this subset. I can create ab-initio models with distinguishable helices but when it comes to heterogeneous refinement and downstream NU-refinements the maps appear almost featureless with apparent resolutions of 5-6A. I can only see improvements with 3D masked classification and subsequent NU-refinement which renders clearly more defined maps yet the resolutions bump to 7-8A. My ab-initio maps are created in the 7-9A range as recommended across many posts and tutorials. I would like to know if someone has seen this behavior and explain why it happens. Also, I would appreciate any advice on how to deal with this low SNR small and flexible complex and its masking to get final maps.
I’ve had this issue a couple of times, both with particles showing substantial flexibility. If you run a local resolution job on the featureless NU-refined map with a high-resolution FSC cutoff, do you also see that the highest resolution estimates are found on the edge of your complex? In one of my cases, it seemed to be due to a mix of bad initial alignments (due to flexibility) and a polluted particle stack. In my case, it could be solved by sticking to homogenous refinement and limiting the alignment resolution to lower and lower resolution, until it did not overfit anymore. Then, at that resolution I did 3D classification to sort out the junk and repeated the homogenous refinement at higher resolution to slowly increase the resolution. At some point, the alignments and particle stack were good enough that a NU-refinement no longer just fitted to the particle envelope but instead on internal features, and after that it behaved normally.
However, it sounds like you already reached a good point with the 7-8 Å NU maps showing features, so it might be worth just doing some rounds of local refinement and 3D classification to see if you can improve your resolution, right? My feeling is that when you are working with small, flexible proteins, invest the time in inspecting your map instead of relying on the FSC until you have a very clean stack and a nice reconstruction.
Also, in my experience, if you are using dynamic masking, setting the threshold higher than 0.2 (eg. 0.3) and compensating by a larger padding can help to place less emphasis on the flexible, blurry parts.
Thanks for all the tips, Martin. According to masked 3D class, some classes reach 4.2A resolution but the maps do not even look like that. I have some questions:
I think the alignment issue is a fact that I have not solved completely. My 2D classes display flexible poor regions. Is there any setting I can adjust to improve them? I did a lot of tries and at the end what better worked was turning off poses/shifts, number of iterations 30-40 and batch size 200-400 (depending on how good the particle stack is). Along with this, which parameter did you modulate when you say “limiting the alignment resolution to lower and lower resolution”? I have not seen this strategy before. What would be a good low resolution range to try?
In my 3D classes, one of the main issues is that usually classes are split evenly. I somehow overcame this by setting class similarity to 0.8-0.9. Still particles on average can be assigned to 3 classes (out of 5).
Many thanks again!
Regarding 2D classification, I don’t see a lot you can do to improve them - increasing the batch size for small proteins definitely works wonders, I always just set the batch size to 500 pr. class. But fundamentally, if your particles are flexible, you’ll never get the whole particle well-resolved at the same time. If you can just tell protein from junk to some degree from 2D classes, that’s great, although be careful not to filter too harshly in 2D, because for small proteins it’s likely that you’ll have some projections without strong features, which will look like junk in 2D. For limiting the resolution, that is just one of the settings for both 2D classification and refinement jobs (Maximum align resolution (A)). I would run a homogenous refinement job and then look back through the iterations to see when it started overfitting and limit it to the last resolution before that.
Regarding the classification, it becomes a weird issue with flexible proteins because only a fraction of the particles will actually have agreeing conformations and so usually you would get better results by aiming for fewer particles in your classes than would usually be recommended. For my case, it worked best by setting the number of classes such that on average it would be less than 10k particles pr. class, but that was most likely specific for my particle. But I definitely think you’ll get better results by using more classes in ab inition and 3D classification.
Hopefully, these suggestions will help a bit!
Thanks for the advice. I took my initial particle stack and started with ab-initio using low resolutions followed by heterogeneous refinements at low resolutions as well. It is segregating populations pretty well. Now I am working on the homo refinement. Helices and internal protein elements are now visible, with overfitting, but I am adjusting the resolution from which I can go to 3D classes and lower resolutions.
Many thanks for the help.