Advice for 3DVA of glycoprotein bending and breathing


I have a consensus refinement of a glycoprotein to 4.34A global resolution and used it as input for 3DVA with a lowpass filter of 8A and 3 modes.

The motions I observed are large glycan motions, flexing at a hinge point between two subdomains, and breathing within each subdomain. I find this very interesting and would really like to observe it by reconstructing the individual states using NU-refine. Since it is continuous, 3DVA clustering doesn’t seem to be an option for creating particle subsets for downstream refinement. I have seen this post ( Continuous vs Discrete motion in 3DVA - 3D Variability Analysis - cryoSPARC Discuss) but am still not clear on how I can further analyze the motions I observe. Would local refinement with a mask around each subdomain be a good idea? Or could ab initio and hetero refine be used to separate these differences?
Due to the small size (140kDa) and high flexibility, I am not sure if local refinement will be possible after particle subtraction. I would really appreciate any advice!

Output from 3DVA:

component 0 ‘simple’ display movie (purple) and component 2 ‘simple’ display movie (green):
P7 J173 mode0
P7J173 mode2

Output from clustering with 5 clusters:

I have since performed NU-refinement on the 5 clusters shown above and obtained maps of around 4-5A for each (there’s still likely heterogeneity due to glycosylation and each cluster is only of around 20-30k particles which would decrease SNR and limit resolution).
When I morph between maps in Chimera I observe the motions shown by 3DVA although a bit less pronounced.
Is this an appropriate way of analyzing my data? Can I now build models into these maps (consensus plus 5 clusters)? The consensus has 120k particles and better features in some regions likely due to higher SNR.

The motions may be less pronounced following separation into clusters because the cluster output does do not give ideal separation for the individual modes of variability, as it is working in 3D coordinate space rather than 1D. This will also reduce the number of particles in each reconstruction.
Since you know your data are good to at least 5A, you could try modelling/refining into the ‘simple’ output maps filtered to that resolution. I would just avoid reporting GSFSC resolution for those frames, but rather indicate that a manually-set resolution cuttof was used.

Thanks for the reply @Ablakely!

I never thought that it would be acceptable to build models into the ‘simple’ mode maps but appreciate the suggestion. In that case, would you recommend using the extremes of each mode for model building? Would such maps only be used for display in a publication and not deposited?

I have since repeated the consensus refine with a tighter mask (it had to be very wide in early jobs to avoid FSC artifacts) and saw slight improvements in resolution so thought that 3DVA should also be repeated using the improved alignments (and a dilated mask for 3DVA). This helped indeed and the motions are more pronounced than before. I again performed clustering and refined those particle stacks. Although the 5 clusters are not distinct in the interactive display, I see a range of helix positions and overall resolution of around 4.5-5A for each. Maybe I should then just build models into the clusters instead?

Frame0.mrc and frame19.mrc from component 0 respectively compare very well to the refined clusters 4 and 2.

grey mesh=frame 0, blue surface=cluster4:

yellow mesh=frame 19, green surface=cluster2:

refined cluster 2 (green) vs cluster 4 (blue):

I suppose it makes sense that clusters 2 and 4 show the greatest change in component 0 since they are at the extremes of x on the clustering plot:

You could build into the cluster output as well. If you are limited by particle numbers then it isn’t ideal though, as you can see in your last image cluster 0, 1 and 3 are essentially the same as far as component 0 is concerned, and those particles are split into 3 clusters rather than 1. Since the variability modes are uncorrelated I can’t think of a good reason why one would want to cluster in 3 dimensions, but perhaps I am thinking about this incorrectly.
On the other hand, if you are happy with the resolution/ particle number per cluster, having real refinements done on the clusters is an advantage for reporting or depositing the structures when you publish (without having to do your own python scripting to get 1-dimensional clusters).

Having more particles didn’t give me much higher resolution (the consensus was at 4.1A global resolution and visually only slightly better than the 4-5A clusters) . Probably because more particles equaled more heterogeneity and thus again, decreased resolution. Looks like the small size, high glycosylation and flexibility ultimately limits the achievable resolution in this case. The defocus range used in data collection was -1.8um to -3.0um so perhaps that is also limiting? At present my box size is 256pix for a particle of ~180A at 1.06apix so maybe an increased box size could help with resolution?

I can try to postprocess these maps by local resolution filtering and then build into the consensus map (where some parts of the active site core are better resolved) and the clusters (to describe the motions).