Hi Shintaro,
> As I understand half-sets are also randomized every time you run a
> refinement
No. Half sets are allocated only once in the first Refine3D job and stored
in the
rlnRandomSubset column of run_data.star. This column is carried along
in all subsequent steps (classification, subset selection, re-extraction,
3D refinement, CTF refinement, Bayesian polishing, etc).
Bayesian Polishing also separately refines half sets to avoid over-fitting.
Best regards,
Takanori Nakane
> Hi Shintaro,
>
> You are right: per-particle CTF estimation does have room for
> overfitting! Therefore, we perform this under strict separation of the
> half-sets (the 'gold-standard'). If you don't do that, then you may
> indeed get spurious improvements. The current code is safe. Good
> indications that things are going well are continuous changes in defocus
> over each micrograph, as can be judged from the logfile.pdf. The exact
> approach will be published in our upcoming relion-3 paper. That might
> need to be carried over the Summer though...
>
> Regarding the autopicking: it is just taking evenly distributed views
> from the 3D reference. If you are worried about this, you can always
> provide your own 2D references as templates. We haven't seen any
> problems with the 3D approach.
>
> HTH,
>
> Sjors
>
>
>
> On 08/01/2018 07:56 PM, Shintaro Aibara wrote:
>> Hi Sjors,
>>
>> Hope you're doing well, and thanks for the new release of Relion3!?
>>
>> I just thought I'd ask a few questions about some of the new features
>> mainly surrounding CTF refinement (and to a degree generally
>> per-particle CTF estimation).
>>
>> *Per particle CTF estimation - when is it okay to do it?*
>> Since we are going from 3 parameters per micrograph (U,V,Angle) to
>> possibly 2x(# of particles) + 1 or 3x(# of particles) parameters where
>> each CTF that needs to fit is presumably noisier than that of the
>> whole micrograph, do we have anything to keep potential over-fitting
>> in check? We seem to be introducing a tonne more fitting parameters
>> with less data.?
>>
>> I seem to get the impression that per-particle CTF/CTF refinement
>> always improves the reported resolution by about 0.2 A (for a 2.8-3.5
>> A map) and this worries me slightly. It is quite hard to tell whether
>> the map has been genuinely improved or not at this level. However, I
>> suppose if this is the trend, we must all start doing it to make sure
>> our reconstructions keep up with the perceived average resolution of
>> the community.
>>
>> *CTF refinement - any concerns for independent halfsets?*
>> I noticed that the way CTF refinement works has been described to be
>> comparisons against high resolution reference projections. Presumably
>> this would normally be the output of a refinement run, where two
>> halfmaps have been combined. Would this mean then that there is
>> potential to introduce correlation at high resolution between the
>> independent halfsets since the particles have "seen" the combined map?
>> As I understand half-sets are also randomized every time you run a
>> refinement and so even if CTF refinement was being done against
>> half-maps the potential problem for the independent half-sets to be
>> not quite independent still remains.
>>
>> An empirical observation I made was that my FSC of a map after CTF
>> [log in to unmask] was very close to the FSC of the refinement without CTF
>> refinement @0.143 (equaling about an improvement in 0.2 A or so, as
>> above). Perhaps just a coincidence, but was food for thought.?
>>
>> *3D autopicking - picking particles with views that may not exist?*
>> Are there any adverse effects from picking with views of a
>> reconstruction that may be severely preferred (usually I find this
>> manifests as a streaky mess). It just seems slightly concerning about
>> biasing the picking with views that may or may not actually be present
>> in the dataset. Are there any decision making within the algorithm to
>> decide whether a projection is used as a reference or not, or are all
>> projections used?
>>
>> Best wishes
>> Shintaro
>>
>>
>>
>> On Wed, Aug 1, 2018 at 6:17 PM, Sjors Scheres
>> <[log in to unmask] <mailto:[log in to unmask]>> wrote:
>>
>> Dear EM-ers,
>>
>> After multiple months of in-house testing at MRC-LMB and the
>> SciLifeLab, and several months of beta-testing by a few external
>> expert groups, we think that RELION-3.0 is now ready for a more
>> general round of beta-testing. To that purpose, you can download
>> it from:
>>
>> git clone https://bitbucket.org/scheres/relion-3.0_beta.git
>> <https://bitbucket.org/scheres/relion-3.0_beta.git>
>>
>> The attached betaGuide.pdf contains a list of new features, and
>> instructions on how to install and provide us with your feedback.
>> Please read it carefully. As always, please use the ccp-em email
>> list (and not a direct email) to ask questions on how to use
>> RELION. We are grateful for your detailed bug reports through the
>> issue tracker on bitbucket, which will help to make RELION better.
>>
>>
>> Have fun,
>>
>> Jasenko, Takanori, Bjorn, Dari, Erik & Sjors
>>
>> --
>> 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
>> <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
>> <https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1>
>>
>>
>>
>>
>> --
>> Yours Sincerely,
>> Shintaro Aibara
>
> --
> 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
>
########################################################################
To unsubscribe from the CCPEM list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
|