Frealign refinement and multiple refinement

Submitted by huy on Fri, 2017-04-07 15:35

Forums

Frealign

Dear all,

I'm very new to Frealign. I get Frealign running. The result looks wonderful. However, there are a lot of questions to be asked:
1. If I understand correctly, Frealign FSC is not Gold Standard FSC. How to prevent overfitting apart from setting the refinement high resolution cut off lower than the current res estimated by FSC? Even if we don't have overfitting issue, what should we say if ever publishing the data (it is not gold standard FSC)?
2. I based my res_high_refinement roughly at the resolution at FSC at 0.5 of the previous iteration. And continue refining until no further improvement. Is that a good way to do?
3. I refined without using any Expert parameters such as thresh_reconst. How should I decide on this threshold?
4. So far, I refined without using a mask. Would a mask will improve my resolution without overfitting? How tight should I put the mask?
5. I tried to do a multiref refinement using the 2 references that I provided. The problem is that frealign always reconstructs 2 references randomly rather than my provided references. I guess I don't understand yet the naming convention in Frealign.

The way I do it is:

A_3_r1.par # Parameter file contains all the particles. I have all the particles aligned to ref 1 intially.
A_3_r1.mrcs # First ref
A_3_r2.mrcs # second ref

Best regards,
Huy

niko Fri, 2017-04-07 20:21

1) You should always describe in your publication what you have done. If you limited your resolution during refinement to prevent overfitting it higher resolution, this should be described in your methods section, including the resolution limit.
2) Setting the resolution limit to the value of the FSC at 0.5 in the previous round may work but also often leads to FSC curves that do not extend much beyond the resolution limit, i.e. the curves drop quickly to below 0.143 beyond the limit. Ideally, the final resolution of the reconstruction is much higher than the set limit used during refinement. If the slope of your FSC curve increases significantly beyond the resolution limit limit, this is a bad sign.
3) The expert parameters can be used to optimize an already very good reconstruction. The best values can be determined by trying new values and recalculating the reconstruction.
4) A mask should not cut into the density of the reconstructed particle. Masking can help suppress noise outside the particle density and therefore help in the refinement and reduce overfitting.
5) Please read the 2016 publication on Frealign in Methods Enzymology for naming conventions. Frealign does not use the "mrcs" extension; all MRC files should have the "mrc" extension. Normally, Frealign would generate multiple classes from randomly selected particles for each class and then refine each class to arrive at "pure" classes. However, if you would like to provide pre-existing 3D reconstructions representing different classes you have to duplicate the parameter file for each class. In your case, if you originally have A_3_r1.par and A_3_r1.mrc, copy A_3_r1.par to A_4_r1.par and A_4_r2.par and copy your two pre-existing reconstructions to A_4_r1.mrc and A_4_r2.mrc. Then start refinement with cycle 5. You have to make sure that the pre-existing reconstructions are aligned with your reconstruction from cycle 3 (A_3_r1.mrc). If they are not aligned, you will have to run at least one cycle of global search against your pre-existing reconstructions. It may be easier to just let Frealign generate starting reconstructions with randomly selected particles and refine these.

huy Mon, 2017-04-10 14:28

Thanks a lot, Niko.

1) Thanks. I found that in my case at the same resolution (not the same samples but very similar samples) 6 Angstrom map from Frealign (FSC 0.143) looks quite better than the map from Relion at 6 Angstrom.

2) So I should make the resolution limit at a region of the FSC before the slope of the FSC increases significantly.

3) I got my structure at ~4.0 Angstrom. Which one would you recommend to refine first? From fitting the atomic structure in, I feel that my pixel size can be 1.5-2% off from the value from the scope, should I refine this first? Any example of refining beam tilt that improves the structure well?

4) I used the soft mask. Here's what I put in the mparameters:

mask_file mask512_min.mrc
mask_edge 0

After running after a while, I got this error

.....
Cycle 17: refining particles 255375 to 270396, class 1 Mon Apr 10 00:53:02 EDT 2017
Cycle 17: refining particles 270397 to 285418, class 1 Mon Apr 10 00:53:02 EDT 2017
Cycle 17: refining particles 285419 to 300436, class 1 Mon Apr 10 00:53:02 EDT 2017
Job 14811 crashed Mon Apr 10 01:03:49 EDT 2017
Logfile /mnt/zurich/data0/20170118/newwf_frealign/bin1_hires/scratch/Sample_mult_refine_n_r1.log_240353_255374
Final lines:
CC for particle 241138 is 0.00344744 score is 7.02
CC for particle 241139 is 0.02867202 score is 7.07
CC for particle 241140 is 0.03470552 score is 6.96
CC for particle 241141 is 0.01454961 score is 5.69
CC for particle 241142 is 0.03785849 score is 6.69
CC for particle 241143 is 0.03605067 score is 5.77
CC for particle 241144 is 0.01290948 score is 6.68
CC for particle 241145 is 0.01999316 score is 6.52
CC for particle 241146 is 0.03892912 score is 5.96
CC for particle 241147 is 0.03334661 score is 5.82
Terminating...

Any suggestion why it is terminating?

5) I got it run with 2 refs. Thanks a lot. The reason why my run got the .mrcs because I named the stack with .mrcs and somehow Frealign automatically takes the extension of the stack for map files.

Thanks a lot again, Niko.

niko Mon, 2017-04-10 14:53

I would try different PBC values first, then maybe also different particle thresholds to see which values give the best reconstructions. For the results, please always look at the map as the FSC is unreliable (in all software packages, not just Frealign). For a soft mask, use an edge width that is larger than zero.