spIsoNet for subtomogram averaging

[Ruchao Peng]

Hi Yuntao,

Thank you so much for sharing this amazing program! I got nice results from the tutorial data. Now I am trying to apply that to my STA dataset that suffered from preferred orientation problem. This dataset contains about 1500 subtomograms and produces an average at ~20A following standard relion pipeline, but obviously distorted due to orientation problem. I have a high-resolution SPA map of this object, so I applied a lowpass fileter to 60A to use as the reference. By default, spisonet only starts the processing when resolution goes beyond 15A. I manually changed it to 35A to force it to work during STA. But I am not sure how reliable is this process for moderate resolution structures like most STA projects? I noticed that after spIsoNet processing, the unmasked FSC is obviously improved and goes much closer to masked FSC. I assume this is a combined result of denoising and the effect of mask during refinement. As far as I understand, this process may also introduce some model bias as we retain more high-resolution signal to help particle alignment. Please correct me if I am wrong.

For sure, the output structure after applying misalignment correction looks more reasonable, but I am not sure how confident I can be for the result. What't the best practice for the application for STA workflow, particularly for moderate and low resolution structures?

Any suggestions are highly appreciated!

Thanks,

Ruchao

[YUNTAO LIU]

Hi Ruchao,

I am happy to see first hand experience using spIsoNet for subtomogram averaging, which we have not tested thoroughly. Note for tomography folks, the environment variable "ISONET_START_RESOLUTION" probably needs to be changed.

In principle, the algorithm works for any resolution range. The reconstruction is based solely on the data you have, not from any other source (i.e. pdb or emdb). Thus, the performance only relies on how much information content you have in your map and data. I think low resolution maps have less information content, so that it is difficult for spIsoNet to recover information. But on the other hand, if the diameter of the map is large enough, e.g. when you have a viral capsid, or the entire microtubule or ribosome, the information could still be sufficient for spIsoNet to learn with low resolution.

I also observed that the refinement resolution is higher when using spIsoNet. I think what you said "introduce some model bias as we retain more high-resolution signals to help particle alignment" is not a model bias problem, but rather an overfitting problem, opposite to model bias. The high resolution signals indeed could be caused by the overfit of the user's own data (https://www.nature.com/articles/nmeth.2115). The misalignment correction is embedded inside RELION, and uses its internal gold standard FSC to regularize the reconstruction. This is pretty much what we can do with overfitting as far as I know in the entire field. I believe the high resolution information that comes from spIsoNet is as reliable in the current practice of cryoEM, but I would like to know other people's opinions on this.

 

--
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/0ab5d9c3-0c2b-4057-968c-4577434fa93fn%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Best Regards,

Yuntao Liu,  Postdoc.

California NanoSystem Institute

University of California Los Angeles

[Ruchao Peng]

Hi Yuntao,

Thanks a lot for the prompt feedback! That all makes sense! I would also like to hear the experience of other folks.

cheers,

Ruchao

Ruchao Peng, Ph.D.

Postdoc

Dept. of Biochemistry & Biophysics

Perelman School of Medicine

University of Pennsylvania

YUNTAO LIU <yun...@g.ucla.edu> 于2024年5月16日周四 16:32写道：