Local fine sampling around selected solutions

[amjjbonvin]

A feature request from a user: Would it be possible to perform a local fine search around a specific solution. Say finer rotational and translational sampling around the starting configuration - no longer full search

[mtrellet]

As a first solution, I would just suggest to trim the density map around the solution found...

[schaarj]

trimming would probably be the easiest but that's difficult on the web server part to detect this and allow more fine-grained sampling
adding a refine flag to powerfit might be an alternative but that would come with some additional parameters I guess (like max rotation and translation)

[schaarj]

also there are probably other programs capable of "optimising" an initial placement provided by powerfit

[amjjbonvin]

This is already something they are doing. The idea would be to take initial solutions and then don’t do the full search anymore, but sample at a finer grid spacing only limited translations (e.g. +-5A and +-20 degrees, but then is small steps of 0.5A and 1 degree)

…

On 16 Feb 2017, at 12:06, Mikael Trellet ***@***.***> wrote: As a first solution, I would just suggest to trim the density map around the solution found...

[schaarj]

that's pretty much what we discussed, so either implement it in powerfit or add a third party post-processing step option (maybe in chimera)

[mtrellet]

That's the second solution we thought of indeed, together with tunable parameters, as Jorg suggested.
@latrocinia Would it involve major changes in the code? We have the feeling that it could be quite straightforward but we might be wrong..!?

[amjjbonvin]

But the idea is to do a systematic limited search. Chimera can minimize, but does not sample systematically. Probably for Gydo to comment on this

[amjjbonvin]

Probably no FFTs anymore…

[latrocinia]

I guess the best approach is to trim the map, and resample it to the requested translational stepsize. To sample rotational space more densely use the naive approach of incremental euler angles. The map and rotations can be given as input to a PowerFit (class) instance. It's not too difficult to implement, most of the tools are already there, they just have to be placed in the correct order.

[amjjbonvin]

They are already trimming the map - but would then like to only sample say +/- 5A translation of the current model (in small steps, e.g. 0.5A) and +/- 10 degrees rotations of the current model (say in 1 degree steps). Might not make much sense…

On 16 Feb 2017, at 16:20, Gydo van Zundert ***@***.***> wrote: I guess the best approach is to trim the map, and resample it to the requested translational stepsize. To sample rotational space more densely use the naive approach of incremental euler angles. The map and rotations can be given as input to a PowerFit (class) instance. It's not too difficult to implement, most of the tools are already there, they just have to be placed in the correct order. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#2 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AFx-oM1RyCmx6_tD7s7jGwkXAZ1L4c8zks5rdGkkgaJpZM4MC3sN>.

================================== Alexandre Bonvin Talingweide 91, 3403 AC IJsseltstein, NL Tel: +31-30-7525883 Email: alexandre.bonvin@gmail.com ==================================

[latrocinia]

Ok, but this you can achieve with what I described, or am I missing something? By resampling the map to e.g. 0.5A voxelspacing you sample all the 0.5A translations, and you sample all rotations that are formed by the cartesian product of the three euler angles within a 10 degree range with an increment of 1 degree.

The other solution is performing it without FFT: Move the protein and give it a rotation and calculate the cross-correlation directly. Just feed it the correct set of translations and rotations one wants to sample.

[amjjbonvin]

You are the guru - so most likely your solution will work. Just that I don’t see what it means in practice :-)

[latrocinia]

:))