Dear Teige,
Great question!
The measuring of overfitting as proposed by Richard is for demonstrating
that a new, or modified refinement method does not suffer from
overfitting. In this test, phase-randomised images behaved the same as
'empty' noise particles. I don't expect carbon noise to behave differently
from ice-only noise. Of course, when an exrta layer of carbon is present,
the SNR in the images becomes lower, and the risk of overfitting (if your
refinement algorithm isn't safe) becomes larger.
Now, the same method of phase-randomisation is also used to measure how
much the masking of the map, which aims to remove noise in the solvent
region, leads to undesirable convolution effects on the FSC curve, when
calculating the final resolution in the post-processing step of relion.
Although the resulting, corrected FSC curve is probably not perfect (e.g.
often artifacts near the resolution used for phase randomisation are
observed), the resolution indicated by where the corrected FSC curve
passes through 0.143 (please, please let no one be dragged into another
discussion on this cutoff) has been observed to correlate generally very
well with visible features in the map. Also here one should always be
careful. For example, we have seen artificially high FSC curves in cases
of strong 'rotational smearing' of the particles due to extremely broad
posterior probability functions of the orientations. To guard against
that, the refinement programs warns about broad probability functions.
HTH,
Sjors
> Dear CCPEMers,
>
> How much can a carbon background inflate overfitting-corrected FSC
> curves?
> Has anyone compared gold-standard FSC where phases are randomised, with
> gold-standard FSC using adjacent “non-particle” boxes - with the
> presence of continuous carbon background?
> Is the only ready-available method for this to manually pick a large set
> of adjacent featureless co-ordinates?
> Fig 2 of Chen et al 2013:
> https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3834153/ compares both
> methods of substituting signal+noise for noise, and reasons that phase
> randomisation is equally good and much easier to do. But the example is
> for a sample in plain ice, no patterned carbon background.
> Is there a good reason to believe that overfitted noise due to carbon is a
> small component of overfitted noise?
>
> All the best,
> Teige
> [cid:62554E71-8DCE-4E28-BAFF-6F13A8160012]
>
>
>
> The Francis Crick Institute Limited is a registered charity in England and
> Wales no. 1140062 and a company registered in England and Wales no.
> 06885462, with its registered office at 1 Midland Road London NW1 1AT
>
> ########################################################################
>
> To unsubscribe from the CCPEM list, click the following link:
> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
>
--
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
########################################################################
To unsubscribe from the CCPEM list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
|