What is wrong with my corrections

48 views
Skip to first unread message

William Nicolas

unread,
Aug 12, 2024, 2:06:32 PM8/12/24
to spIsoNet
Hey Yuntao,

I am running spIsonet tests on some of our data that presents strong anisotropy. I wrote a script to make it easier and screenshoted the spisonet command I use (with limit_res set to 2.5 here but I used 3.5 yesterday. I am currently running tests at 2.5.


When I compare final corrected maps with limi_res at 3.5 (I used relion post-process -lowpass 3.5), I don't really see a correction. What I see is actually that the corrected maps seemed blurred, or low pass filtered. I have the sense that something is up with the filtering settings I used.

I tried Relion LP filters of -1 and 2.5 (a little below max rez of this dataset) and will be checking out outcome later today.

I also attached presentation with some slides showing the corrected structure that just seems blurred to me.

Am I using it correctly?

Thank you,

Will
for_yuntao.pptx
spisonet_script.png

YUNTAO LIU

unread,
Aug 12, 2024, 7:18:56 PM8/12/24
to William Nicolas, spIsoNet
Hi William,

This is interesting, from the FFT you showed, the corrected information is actually populated in the previous empty area. Also the denoising seems to work as you can see in the 3dmod slice view. This means the spIsoNet actually did something. 

If you are concerned that the spIsoNet is a low pass filter, you should use the spIsoNet output (do not use the xx_filtered one) for post processing with relion with --lowpass at estimated 2.5A and with and different B factors. 

The --limit_res 3.5 A in spIsoNet means that the resolution lower than 3.5A is modified by spIsoNet, and the information beyond 3.5A spIsoNet will keep the original. Again, you may not want to use xx_filtered.mrc as it cuts off resolution at 3.5A resolution. 

This is my reasoning for setting --limit_res in spIsoNet: you may have seen the FFT view in 3DMOD, the reliable information (I believe) has specific patchy patterns, this pattern sometimes becomes pure noise beyond a certain resolution, and (I believe) this is actually over estimated resolution or overfitting whatever. I reason (although with limited test cases) that spIsoNet will not learn anything at an overestimated resolution, --limit_res to a lower resolution rather than the estimated FSC=0.143 can help network learning.  But if the structure is reliable, limiting to FSC=0.143 or even without limitation should work well.   

--
You received this message because you are subscribed to the Google Groups "spIsoNet" group.
To unsubscribe from this group and stop receiving emails from it, send an email to spisonet+u...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/spisonet/f3a21f9d-2a6a-4bc0-af4e-d686b919bef6n%40googlegroups.com.


--
Best Regards,
Yuntao Liu,  Postdoc.

California NanoSystem Institute
University of California Los Angeles

William Nicolas

unread,
Aug 13, 2024, 1:16:56 PM8/13/24
to spIsoNet
Hello Yuntao,

Yes the FFT shows filling but when I compare volumes, it seems like a low pass filtered was applied.

I do not use the corrected_xxx_filtered.mrc ones. I use the corrected_xxx.mrc and feed them to relion_postprocessing.py.

I ran a test yesterday where the --limit_res was set to 2.5A and I tried --low_pass from relion with -1 (uses gold standard FSC if I understand), 2.5 and 3.5. I do see an attenuation of this "low pass" visual effect when I decrease the number.

>This is my reasoning for setting --limit_res in spIsoNet: you may have seen the FFT view in 3DMOD, the reliable information (I believe) has specific patchy patterns, this pattern sometimes becomes pure noise beyond a certain resolution, and (I >believe) this is actually over estimated resolution or overfitting whatever. I reason (although with limited test cases) that spIsoNet will not learn anything at an overestimated resolution, --limit_res to a lower resolution rather than the estimated >FSC=0.143 can help network learning.  But if the structure is reliable, limiting to FSC=0.143 or even without limitation should work well. 

This is useful. Maybe increasing limit reZ to 3A then would be good since cryosparc estimates reZ at 2.4A.

Meanwhile I am currently running the misalignment correction on the particle stack that led to the anisotropic structure. The results you show in the paper are quite impressive and want to see where we get with our anisotropic structures. I will then run anisotropic correction after.

Thank you,

William Nicolas

unread,
Aug 13, 2024, 5:42:46 PM8/13/24
to spIsoNet
Hey,

I am trying to run misalignment correction on our HPC. I submit job with qsub using Relion GUI, set the --external_reconstruct. Relion runs but when it's time to call the relion_wrapper.py, I get the attached error.
Screenshot 2024-08-13 at 2.34.39 PM.png
Reply all
Reply to author
Forward
0 new messages