Resolution lost after non-uniform refinement processing the map imported from Relion

I imported the ~170 k particles, a full map with 1.98 Å obtained from post-processing, and a mask from Relion. And they were fed to Non-uniform refinement with C2 symmetry, but as a result, I only got a poor 3.8 Å in GSFSC as the plot attached below.

I want to ask
the input map of NU refinement is a full map or half map?
and should I adjust any other parameters in NU refinement or perform other refinements to improve my map?

Thank you in advance.

Input for NU refinement is a full map.

Does it look like a real 1.98A map in Relion? Is there a possibility it was overfit?

It might be that the mask was too tight. Maybe re-processing without a mask could help.

As jenchem said you should use the full map from relion. I would suggest start with ab initio without symmetry then run NU-R with that volume as input or with or without symmetry.

Hi @yuhx0914,
In addition to @jenchem and @user123’s comments, I just want to add that from the image you shared it seems that something is definitely wrong: it looks like the loose and tight masked FSC curves are worse than no-mask - which indicates at least an issue with masking and possibly many others. Can you confirm that the 1.98A map you have is really looking like a 2A map?

Another sanity check is that if you can import the from relion refinement (from the final iteration), cryoSPARC will also import the pose and shift alignments from the particles. Then you can take those particles and connect them to a “homogeneous reconstruction only” job that will perform a single pass reconstruction from the particles with fixed poses. This should yield nearly exactly the same 3D map as what you get from Relion’s refinement.

1 Like

Yes, the map has already been fitted to the model and refined, all are submitted. But can we run the post-processing without a mask in RELION?

Oh, sorry. Maybe you mean run the NU-refinement without mask, right?

Okay, thank you. I will try the ab intio first, and then NU-R to see the results.

Thanks for your kind reply.

Yes, the 1.98 Å map has been examined and submitted.

I tried feeding the same substracted particles to “homogeneous reconstruction only”, (because I did not find any in 3D refinement of Relion), and also, the same mask fed in Relion. Finally I got a volume with 2.11 Å attached below, which is much more similar than before.
Do you have other suggestions to improve the map in the next step.

When you look at the early iterations of NU refinement, do you see masking artifacts that appear around iteration 2 or 3? If so you may want to try setting the “dynamic mask start resolution” to a value that will prevent dynamic masking until after the 2nd iteration.

Hi @yuhx0914,

It would be great if you can try @Ablakely’s suggestion.

Also, this type of issue is an important test case for us. I will reach out to you by email about possibly confidentially sharing data with us.


1 Like

Thanks for @Ablakely 's advice.
But I did not run successfully after changing the parameters as the attached Fig. 1, where only several iterations could be completed.
And attached Fig. 2 and 3 are the results of iteration 2 and 3 with default settings.
Fig. 4 and 5 are the results of iteration 2 and 3 with the settings in Fig. 1.
Fig. 1

Fig. 2

Fig. 3

Fig. 4

Fig. 5

Thank you.

those images confirm a problem with your masking. One problem with your settings is that the difference between “near” and “far” should be >6 in my experience. This determines the width of the soft edge on the mask.
Also if your reconstruction exceeds 6A resolution after iteration 0 or iteration 1 maybe you need to set “dynamic mask start resolution” even lower (to 3 or 4 A ). Changing this setting shouldn’t really be necessary but it has a strong impact in my tests.

Edit: on second look at fig 4, it appears that the mask is not aligned with your map. This might be due to using “Align map to symmetry”. you can try the refinement without giving it the Relion mask as an input and that should avoid any mask/map alignment errors.


Thank you.
I followed your strategy to change the parameters.
The dynamic mask threshold is set as 1, “near” and “far” are set as 2 and 8, the “dynamic mask start resolution” is set as 6 Å because the resolution did not reach 6 Å after iteration 0 or iteration 1 before, with the input mask as blank.
This time I can complete the job this time with the resolution ends at 4.7 Å, it is still far away from ~2 Å.
The screenshots from the final iteration are shown below, but I feel strange that there is no shape of the map that could be recognized.
By the way, after the Local Refinement, I can obtain a map at 2.03 Å.


That is very strange. There is still a problem with the mask but its not clear whether that is the root issue or a symptom. Have you tried as jenchem suggested and simply try the refinement without an input mask, but leaving the other masking parameters the same? Since the (presumably) C1 local refinement works, but not the C2 NU refinement, you may also try C1 in NU refinement as a troubleshooting step.

Yes, I followed @jenchem 's suggestion. Perform the No-refinement without a mask.
But similarly, a weird map was given with almost no shape as the attached Fig. 1, 2, and 3 depicted.
I imposed a C2 symmetry in local refinement and got the 2.03 Å map above mentioned.
Fig. 1

Fig. 2

Fig. 3