I am working on a flexible protein (~128 kDa). I did Topaz picking and performed one round of 2D classification to remove obvious junk. I used the remaining particles to do Het Refinement with 3 models of paralog structures and 2 junk volumes (obtained after killing ab initio). The particles obtained after Het Refinement were used in an NU Refinement job, with a good volume from the previous Het Refinement as its input model. This gave a 5.5 Å map, but the map lacks those features.
I am doing the processing at bin 4 due to space constraints (pixel size at bin 4 is 2.6 Å/pixel). In parallel, I did multiple rounds of Het Refinement, but the map quality did not improve. Also, the cFAR value is always less than 0.5.
The central two core domains are stable, whereas the extended helical domains are quite flexible. Each of them has one domain attached via a linker.
I am a newbie to processing, so any advice on how to improve the map quality would be greatly appreciated.
Rather than using references from homologous proteins, why not try to do ab initio to get an initial reference? It’s usually much better to generate your initial reference this way as maps generated from pdbs will induce bias and there may be meaningful differences between your protein and these other related proteins.
Thank you for the suggestions. I tried it, and the overall domain architecture of the protein was present, but the secondary structure elements were not well resolved. Do you think performing local refinements on each individual domain would help improve the refinement of these regions?
That’s pretty good if ab-initio is able to produce a map that corresponds to what you expect. That thing looks like a wobbling chain of domains, and depending how much they move, it might be impossible to solve just like this. Not a case for newbies, so make sure you’ve got the motivation to insist and learn. First thing is to start with a real lot of particles, ideally way over 1 mi. Can you afford that ?
1 million particles should be plenty. We often will do a few het refines with one good class and several decoy volumes. Once most of the particles go into the single good class, we would then do HAIR ab initio with 3-4 classes according to Oli’s preprint, selecting only the best particles from each job. For such wiggly proteins, you may only end up with 30-40k particles in the end but they should refine really well.
Yes, please try HAIR as jcoleman suggests then post the outcomes. We can take from that later. In case high resolution is important: in the meantime, think about things you may try on the bench to decrease flexibility (add ligands, cryoEM scaffolds, inducing back folding if possible…). This will depend on the biological questions you are asking, and what are the important bits. I am not saying the current dataset won’t give you anything useful, but it might take a long time to process.
Thank you so much for recommending this protocol! I followed it with my bin 2 particles, and I was able to get a map at around 6 Å resolution. The domains also look quite well resolved for a 6 Å map. I’m planning to unbin the particles and run a NU refinement, followed by local refinement, to see if I can further improve the resolution of the core.
Thank you again for your guidance and for pointing me in the right direction, it was really helpful! I’ll let you know how the next steps go.
Thank you for the suggestions!. I’ll also think about possible approaches to reduce the flexibility on the bench, while continuing to explore the heterogeneity in the current dataset.
Hi, sorry for the very late reply. We were facing a space crunch, so I couldn’t run my jobs. I currently have around 0.2M particles, and after HR-HAIR followed by NU refinement, I obtained a 4.7 Å map. I then performed local refinement on the core after subtracting the peripheral domains, which improved the resolution to 4.1 Å.
I also performed one round of 3D classification with all 0.2M particles, but I ended up with very similar-looking classes. I’m planning to try 3DVA as well, and I’m also trying CryoDRGN to further investigate the flexibility.
3D classification in cryosparc may not work perfectly with the default setting from what you described.
I often will increase the batch size per class to 8k and use the mode ‘input’ instead of ‘simple.’ I will then supply many different volumes, some low res and some random density. This I find will produce better separation of particles.
There are many other things you can do also. Do you have relion?
Also change the filter resolution. I often set this higher than the resolution of my reconstruction. This also controls the box size, so if it’s too low then not much will happen.
I tried cleaning up the particle stack using 3D classification and obtained a class with ~88K particles. I then performed NU refinement, but the resolution remains more or less the same. However, the cFAR value has improved to 0.6, while the B-factor is still around 300. I am planning to try global and local CTF refinement next, with the hope of improving the B-factor and, consequently, the map quality.
I also tried local refinement, where the reported resolution improved numerically, but I couldn’t really see a corresponding improvement in the map density. Do you have any suggestions on what else I could try to improve the reconstruction?
I also tried 3D refinement in RELION, but it gave me a ~9 Å map, so I dropped that approach for now.
I also ran 3DVA on ~0.2M particles, but I couldn’t observe any obvious heterogeneity or large-scale flexibility. This may be a basic question, but I am a little confused: if there is no apparent flexibility and the dataset does not seem to have much preferred orientation, why might the particles still not be refining well in NU refinement? Could it be primarily due to the relatively low signal-to-noise ratio and the size of the particle (~120 kDa)?
Thank you so much for taking the time and putting in so much effort to help me with this. I really appreciate your guidance.
From the shape you show I doubt there is no flexibility. What might be happening is that 3DVA is not able to show that flexibility; 3DVA works best with clean sets of around 1 mi ptcls or more, and with filtered resolutions of 8 - 12 angstroms, starting from an initial alignement that needs to be already relatively good. What you can try is to play with local refinement to align one region, then put a focusing mask on the other region(s) for 3DVA. I’d give it a filter to 12 or even 20 angstroms in your case. I am not surprised that lowering to 88 k ptcls does not improve resolution, this is typical of flexible proteins because as you align one part, the other gets misaligned, so you end up with an average alignement that is not perfect for anything. You might have reached the limit for this set, but you can keep trying local refinements for individual subdomains, if helpful at all.
Forgot to say that re-extracting applying shifts after local refinement often helps the downstream processes. And at this point I’m usually in downsampled mode, apix 3.5 or so, to gain time.
