Relion Autorefine resolves better than Frealign?

Submitted by tarek on Fri, 2017-06-02 11:20

Forums

Frealign

Dear all,

I am trying to squeeze out any fourier pixel possible from a sample dataset and therefore ran the same data with different tools.
I just realized that the resolution achieved by Relion2 is notably better than what I get from frealign.
To avoid any difference in normalization, internal weighting etc. I directly reconstructed the volume using parameters determined by relion with frealign_calc_reconstructions: 4.59A
after 20 iterations of local refinement (mode 1) I end up with 4.92A.

Did anyone experience something similar before?
Any ideas what is going on?

@Niko
How can I specify in frealign the search radius, increment and shift if I already have well aligned data?
Using Mode 1 I couldn't find ways for tuning...

Best,
Tarek

niko Fri, 2017-06-02 17:19

You may have to adjust the resolution limit to something lower to optimize the refinement. If the resolution limit is chosen too high refinement can degrade the resolution. Please also check the actual density since the FSC is not a reliable resolution measure.

The only option to limit the parameter changes is to turn off parameters using the parameter mask. There is no way to limit the search radius.

tarek Tue, 2017-06-06 03:51

@rdrighetto

Indeed sometimes to improve a map you should *NOT* refine all parameters simultaneously in MODE 1, especially if they are already close to optimal. Try playing with parameter_mask, for example, try refining only the Euler angles or only the x,y shifts for a few iterations, or other combinations among them;

I will try this, thanks for your suggestion!


Also, if you are using FFILT=T make sure you have the correct value specified for MW (mol_mass);

check


Other parameters that you may try tuning to really squeeze out higher-resolution information are:
-PBC (lower values tend to improve the quality of the reconstruction, but there is usually a sweet spot beyond which it doesn't improve or gets worse)

what range do you try usually? For a previous refinement I was tuning between 2-10 but nothing really improved. That sample was way better anyway, maybe already at it's limit with 3.3


-thresh_reconst (PBC does a better job, but this may also help)
-FBOOST (beware of the refinement resolution limit!)
-Provide a mask for refinement
-3D classification

check


Also for a fair comparison it's important to always use the same reconstruction parameters between your initial map and what you get after refinement, but you seem to be already doing that;

yes

@niko


You may have to adjust the resolution limit to something lower to optimize the refinement. If the resolution limit is chosen too high refinement can degrade the resolution.

I was using 6 A, maybe that was too close. Do you have a rule of thumb or glp number? According to relion resolution is 3.8A (post-processed).


Please also check the actual density since the FSC is not a reliable resolution measure.

true


The only option to limit the parameter changes is to turn off parameters using the parameter mask. There is no way to limit the search radius.

coming from spider this is, unfortunately, a drawback.
How would you solve, e.g. realignment of particle shift due to different movie correction methods or change in data binning?
What is the maximum shift Mode 1 would allow?

Thanks again for your helpful comments.

rdrighetto Mon, 2017-06-12 09:01

what range do you try usually? For a previous refinement I was tuning between 2-10 but nothing really improved. That sample was way better anyway, maybe already at it's limit with 3.3

For me, PBC values between 2 and 10 usually work well.

For limiting the search radius in the refinements:

I have a modified version of Frealign v9.11 that allows imposing restraints on the alignment parameters, thus preventing them from changing "too much" from one cycle to the next. This is an experimental feature, but if you want to give it a try please contact me: ricardo.righetto-at-unibas.ch

rdrighetto Sat, 2017-06-03 09:28

Indeed sometimes to improve a map you should *NOT* refine all parameters simultaneously in MODE 1, especially if they are already close to optimal. Try playing with parameter_mask, for example, try refining only the Euler angles or only the x,y shifts for a few iterations, or other combinations among them;

Also, if you are using FFILT=T make sure you have the correct value specified for MW (mol_mass);
Other parameters that you may try tuning to really squeeze out higher-resolution information are:

-PBC (lower values tend to improve the quality of the reconstruction, but there is usually a sweet spot beyond which it doesn't improve or gets worse)
-thresh_reconst (PBC does a better job, but this may also help)
-FBOOST (beware of the refinement resolution limit!)
-Provide a mask for refinement
-3D classification
-Using different sets of movie frames for refinement and reconstruction (raw_images_ref and raw_images_rec)

Also for a fair comparison it's important to always use the same reconstruction parameters between your initial map and what you get after refinement, but you seem to be already doing that;

But sometimes the map simply cannot be improved, or only marginally :-(