Advice on refining this structure?


I am trying to solve the structure of a protein complex between a 75kDa protein and a 35kDa protein. I am still unsure about the stoichiometry of this complex, as some of my SEC-MALS-SAXS suggests it is not 1:1, but the particles on my EM grid look like it could only be a 1:1 interaction.

Under the assumption that my complex is indeed 1:1, these are the class averages from my data acquired on a Titan Krios 300keV scope:

These particles look a lot like my 75kDa protein, apart from the fact that the center of the particle looks like it scatters strongly (higher contrast). If you look at the crystal structure of the 75kDa protein below, you’ll see that the region of this protein with the fewest atoms/electrons is in the center:

On the right is the map_sharp from my homogeneous refinement. I see, I THINK, density in the middle in the map that doesn’t exist in the PDB. I am hoping this is my 35kDa protein ‘wrapped’ around the center, but I can’t seem to resolve this map to higher resolution to start to see features in that density (I think because I am limited in the viewing angles of my particles… most seem to be from top-down).

I wrote a script quickly yesterday that will take a PDB and voxelize it, then collapse the voxelized cube into a 2D projection to see what images of this protein should look like if it was on a grid whilst not in complex, and I don’t see any density in the middle, as suspected:


  • Is it reasonable to suspect my partner is binding in the center and this map is worth pursuing futher?
  • What steps can I take to improve the resolution of my map to start seeing features in this central density?
  • Could it be advantageous to acquire a tilt series on this grid to help improve resolution?

I am new to CryoEM, sorry for my noobishness… any help is much appreciated.

I would say it is reasonable to suspect your protein might be binding in the middle, as you suspect.

Have you tried aligning the calculated density from your pdb (e.g. created using molmap in Chimera) with your initial reconstruction? It may help to filter your initial map to lower resolution, to see less ordered parts more clearly.

By the way, to obtain projections in the future, you can just import your calculated density into cryosparc and then use the Create Templates job, which will generate evenly spaced projections of your map, which can be useful for comparison with your 2D classes.

1 Like


It might also be worthwhile to perform two sets of 2D classification one with a large mask to capture the long dimension and another with a smaller mask to capture the end-on views. Similar to the processing of alcohol dehydrogenase in this paper ( You also might find it useful to turn off “Force Max over poses/shifts”



Thanks @olibclarke for your input – by aligning calculated density, do you mean just manually superimposing in a molecular viewer? Or do you mean actually using the .mrc of my 75kDa protein in a processing job as a way of guiding things?


I tried your way of obtaining projections – I made an .mrc file from the crystal structure, imported it into CryoSPARC, but when I tried to create templates, there was an error when it tried to broadcast arrays between my volume’s cell and the one created by CryoSPARC:

[CPU: 211.0 MB]  Traceback (most recent call last):
  File "cryosparc_worker/cryosparc_compute/", line 84, in
  File "cryosparc_worker/cryosparc_compute/jobs/create_templates/", line 47, in
  File "/home/nick/CryoSPARC/install/cryosparc_worker/cryosparc_compute/", line 152, in zeropad
    new_X[a:b,a:b,a:b] = X
ValueError: could not broadcast input array from shape (60,61,74) into shape (74,74,74)

Is there some way to tell molmap to create a cubic cell? I think CryoSPARC just takes the largest dimension and creates a cubic array from that.

Thank you @kyestachowski as well. I tried the following:

  • Re-curating my exposures
  • Re-picking particles (750)… anything that looked like continuous, isolated density. Big and small things, in the hopes that some would be ‘face-on’ orientations. Class averaging these, and using them as templates for the template picker which then picked 850,000 particles
  • Class averaging the new particles with 80 classes, 50 iterations, ICUF of 2.0, Batchsize per class of 200, and Force max over poses/shifts set to False.

I obtained the following:

The ~50 classes below were basically all junk. I selected the first 17 classes and did an Ab initio reconstruction (I tried it twice, with 1 class and 3 classes to absorb junk). Both outputted similar maps. I then did a homogeneous refinement on the single class and got basically the same thing as before:


Green = calculated volume from crystal structure with cartoon alignment
Blue = my map with arbitrary alignment

My particle looks so different from the crystal structure, almost as though it is not big enough. But, it has central density that the crystal structure does not have.

Are these all artifacts from just not having enough orientations? Is maybe a tilt series the only way to go? You suggested I use different size masks, but I sadly don’t even see any ‘head-on’ smaller orientations like you describe :frowning:

Thanks again

PS. I am going to try the ‘Rebalance 2D Classes (BETA)’ job now, because I read in the docs if you observe smearing out along a particular axis due to preferential orientation, balancing the classes may help…

Yes, I meant aligning the calculated density to your map in Chimers, e.g. using fitmap

You may find running 2D classification with a smaller mask diameter may resolve some of those lower (third row) classes into clearer “along axis” views. I would also suggest trying picking using a neural network picker such as Topaz or crYOLO - these can be better for picking small, low contrast views.


The larger component seems to be 2-fold symmetric in 2D averages. Is it a true symmetry, is it a dimer?
I wonder if you have a case of a close pseudo-symmetry, which leads to averaging of the second component in two distinct orientations. I have not encountered this in cryo-EM, but symmetry misalignment is a common source of nasty twinning problems in crystallography, we had a share of those. Perhaps it could help to 3D-classify in relion omitting re-alignment?


Hi - I think you are also working with a small number of particles, I think with this size having more particles will help you.