Dear all,
El 10/09/2020 a las 16:02, Takanori Nakane escribió:
> Hi,
>
> This is a very important question but, as far as I know, there is no
> consensus yet.
>
> I myself is uncomfortable using "modified" maps for refinement.
> To see the danger, consider two extreme cases as thought experiments.
>
> 1. A hypothetical program that completely ignores the input
> and outputs our 1.22 Å apoferritin map. Of course this is nonsense,
> but the
> resolution evaluated by half-set FSC of the output would "improve".
> This is an example of introduction of wrong prior knowledge.
>
> 2. Another hypothetical program that internally builds an atomic model
> and outputs
> a map calculated from the model. The output would show wonderful
> density with holes
> in aromatics. The half-set FSC and FSC against (your manually created)
> model would both improve, but the new map would tell nothing new.
>
Following the argument that deepEMhancer is a nonlinear regressor,
y=f(x), your two thought experiments would correspond to extreme cases
in which f(x) does not depend on x (1), or depends on x through a kind
of a nearest neighbour database search (2).
> Of course, existing density modification programs try to avoid these
> issues
> but I don't know how well they are validated. Especially it is hard to
> evaluate
> how universal neural networks are.
> (e.g. what happens if the input is very different from the training
> cases?)
This is an obvious problem in regression, you can only trust your model
in those regions of x in which there were training samples. Otherwise,
it is an unsafe extrapolation (and neural networks are particularly bad
in this, despite all the regularization precautions). We have used a
large training dataset from a wide variety of proteins. However, we
cannot guarantee that this dataset has samples absolutely everywhere in
the input protein space.
>
> One possibility would be:
>
> 1. refine an atomic model against the density modified half map 1
Note that this is a different f(x) (B-factor boosting) which is not
necessarily the ground truth of "restoration" functions (actually we
showed that global B-factor correction is not always good, in J.L.
Vilas, J. Vargas, M. Martínez, E. Ramírez-Aportela, R. Melero, A.
Jiménez-Moreno, E. Garduño, P. Conesa, R. Marabini, D. Maluenda, J.M.
Carazo, C.O.S. Sorzano. Re-examining the spectra of macromolecules.
Current practice of spectral quasi B-factor flattening. J. Structural
Biology 209: 107447 (2020))
> 2. refine an atomic model against the original half map 1
> 3. calculate FSC between model 1 or 2 vs the ORIGINAL half map 2
>
This suggestion goes a bit along my suggestion in the previous mail of
using several maps to do the fitting and benefit from those areas in
which each restoration function, even the raw map, better preserves the
structural information.
> It the density modification procedure improved the atomic model,
> the FSC of model 1 vs the original half map 2 should be better than
> the FSC of model 2 vs the original half map 2.
I would disagree that fitting against any particular restoration
function (deepEMhancer, B-factor correction, local deblur, or any other)
should be systematically better for the whole protein. In my opinion,
that would locally depend on regions in which the restoration function
has performed a particularly good job, and certainly this is not
globally reflected in an omnibus measure like the global FSC.
>
> At least, one should deposit the original unfiltered half maps to EMDB
> in addition
> to post-processed and/or density modified maps.
I cannot agree more with this statement, and this was one of our
conclusions of the paper above already stated in its abstract.
Depositing the unfiltered half maps would enable the application of any
future new development that could be eventually better than whatever
restoration functions we have now.
Kind regards, Carlos Oscar
>
> Best regards,
>
> Takanori Nakane
>
> On 2020/09/10 13:57, Oosterheert, W. (Wout) wrote:
>> Dear all,
>>
>> I’ve been playing around with map postprocessing using DeepEMhancer,
>> and I’m quite impressed by the results, especially for datasets that
>> suffer from preferred-orientation.
>>
>> I recently obtained a 3.4 Å, gold-standard refinedmap using Relion,
>> and subjected the unfiltered half-maps to postprocessing using
>> DeepEMhancer. I then tried to refine my model against the
>> DeepEMhancer-postprocessed map using Phenix Real-space refine. The
>> model looks fine after refinement, but I noticed some strange values
>> in the validation statistics:
>>
>> * The model vs. data CC is only 0.73, whereas the CC is 0.84 when I
>> refine against the Relion-postprocessed map.
>> * The map to model FSC curve tails off very slowly, FSC = 0.5 is at
>> 3.2 Å, and FSC = 0.3 at 2.2 Å. Is this type of behaviour expected?
>>
>> In the bioRxiv manuscript
>> (https://www.biorxiv.org/content/10.1101/2020.06.12.148296v3),
>> manual model building using Coot is performed in the DeepEMhancer
>> maps, but refinements are not mentioned. So my question is:
>> can/should I run refinements against DeepEMhancer-postprocessed maps,
>> or should I only use these maps to guide model building in Coot?
>>
>> Thanks,
>>
>> Wout
>>
>> /Wout Oosterheert; PhD Candidate; Crystal and Structural Chemistry;
>> Bijvoet Center for Biomolecular Research; Utrecht University; The
>> Netherlands/
>>
>>
>> ------------------------------------------------------------------------------------------------------------------------------------------------------
>>
>>
>> To unsubscribe from the CCPEM list, click the following link:
>> https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCPEM&A=1
>>
>
> ########################################################################
>
> To unsubscribe from the CCPEM list, click the following link:
> https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCPEM&A=1
>
> This message was issued to members of www.jiscmail.ac.uk/CCPEM, a
> mailing list hosted by www.jiscmail.ac.uk, terms & conditions are
> available at https://www.jiscmail.ac.uk/policyandsecurity/
########################################################################
To unsubscribe from the CCPEM list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/WA-JISC.exe?SUBED1=CCPEM&A=1
This message was issued to members of www.jiscmail.ac.uk/CCPEM, a mailing list hosted by www.jiscmail.ac.uk, terms & conditions are available at https://www.jiscmail.ac.uk/policyandsecurity/
|