|
Hi Basil,
Thank you for your insightful questions. Although it's not clear from your
message, I assume you left some density in the partixles? I.e. you
subtracted a map from which a part was masked away (which was thus not
subtracted)? If that is indeed the case, how much was the approximate mass
of the remaining density?
For gamma-secretase, we only had 30-40 kDa of density left in the
subtracted particles. That is not enough to align particles on, and that's
why we chose not to do any further alignment: we only used the subtracted
particles in a 3D classification without alignment (using 8 classes, and
later also using 6 and 10 classes). Then, when the correct classes were
identified, we went back to the original, non-subtracted particles, to use
those in a normal 3D auto-refinement. The latter can be done conveniently,
by swapping the header labels, rlnImageName and elnImageOriginalName.
To answer your questions:
1) no: haven't encountered them
2) no need to: as relion reconstructs and projects with intact
grey-scales, projections of the map that was reconstructed from the data
will be (on average) on the correct greyscale.
3) to preserve the greyscale from the experimental particles, this should
be the map BEFORE postprocessing, e.g. Refine3D/run1_class0001.mrc (or
Class3D/run2_it025_class003.mrc)
4) no. this should also work at lower resolutions.
Finally: looking at raw low-pass filtered particles like you sent me is
very difficult to assess whether subtraction was too much or not.
Sometimes doing a 2D classification with the subtracted particles works
better. Or if the remaining density is too small for that, you could use
the orientations and classes from a 2D classification with the original
particles to calculate 2D class average densities for the subtracted ones.
Using --refdim 2 in relion_reconstruct, you can calculate fully
CTF-corrected 2D averages from particles with orientations in a STAR file.
HTH,
Sjors
> Dear Sjors,
>
> Me and several other lab members have read your recent manuscript on
> signal subtraction from particle images with great interest. I have been
> following the procedure outlined in the manuscript, using a data.star file
> from a completed global refinement run and the map.mrc from that same run.
> However, unexpectedly, the resolutions of classification and refinement
> runs using the signal-subtracted particles have been much worse than using
> the un-subtracted images. The large extent of resolution decline suggests
> that there is some issue.
>
> Upon closer inspection, I am under the impression that Relion may actually
> subtract a bit too much density, leaving dark areas in the images (darker
> than background: see attached image with strongly low-pass filtered
> particles to reveal the sites where subtraction happened). I would assume
> that because I used matching data.star and map.mrc files, there should not
> be any errors in the generation of the appropriate projections (in terms
> of angles and shifts), but that there may be an issue with the grey value
> scaling of the map. So, I have a few questions and would much appreciate
> to hear your thoughts on them:
>
> 1) Have you ever encountered such issues, and is there a reliable way to
> circumvent them?
>
> 2) Are the projections grey-value scaled before subtraction to match the
> signal of the individual particles images?
>
> 3) What type of volume should be used for subtraction - the direct output
> from the refinement procedure (unsharpened, filtered by Relion without
> masking), or a post-processed (sharpened) and filtered map? I have tried
> both of these options with very similar results.
>
> 4) Is there a requirement for a high-resolution structure for the signal
> subtraction procedure to work?
>
> Thank you very much & best regards,
>
> Basil
>
>
>
>
>
>
--
Sjors Scheres
MRC Laboratory of Molecular Biology
Francis Crick Avenue, Cambridge Biomedical Campus
Cambridge CB2 0QH, U.K.
tel: +44 (0)1223 267061
http://www2.mrc-lmb.cam.ac.uk/groups/scheres
|
