[phenixbb] Fwd: [ccp4bb] Phenix; converting a map to structure factors

[Pavel Afonine]

Hi Mark,

> Would reciprocal space refinement improve maps - or would all the improvement come from model bias? i.e. would maps look nicer, but with no new information content?

cryo-EM map is your data, similarly to Fobs in case of X-ray. We never 
change Fobs during structure solution and refinement, likewise cryo-EM 
map should not be changed.

What happens when you use reciprocal space refinement programs to refine 
against "Fobs and phases" generated from cryo-EM map is that the program 
will calculate 2mFo-DFc and mFo-DFc maps using model phases combined 
with phases from the map. Obviously such calculated maps will be model 
biased. I would say this is not desirable. (Oh and don't forget - you 
need to use electron form-factors if you refine against the cryo-EM map 
in reciprocal space!).

> I guess the difference from X-ray is that while each HKL has contributions from all atoms in the ASU, for an EM map converted to structure factors this is not true. Also, EM maps have substantial local resolution differences, and I don't understand how this would impact SF calculation.

This brings another point.. If you calculate a box of reflections from 
the map, then conversion Map <> Structure Factors is lossless and 
biunique. The problem is that when we convert map to structure factors 
of certain resolution we essentially carve a sphere with reflections out 
of the box; this is because reciprocal-space refinement programs are 
designed to work with sphere of reflections not the box. This 
manipulation result in loss of information: Map1 -> sphere of 
reflections -> Map2 with Map1 being not the same as Map2.

> I know that some in the EM community, who are not used to hands on (i.e. coot/O etc) real space refinement, are disappointed by the small radius of convergence for automatic real space refinement that is not guided by a human being sitting at a graphics terminal.

One of the great things about real-space refinement is that it can be 
done locally. In particular this means optimization methods with larger 
convergence radius can be used: for example, one can systematically 
sample chi angles of a residue side chain to find the best fitting 
rotamer. Another example: data / restraints target optimal weights can 
be obtain in a matter of seconds, while obtaining optimal weight in 
reciprocal space refinement may be a very time consuming task. I can 
bring many more examples like these.

In real-space refinement one can use refinement target functions that 
are much faster to compute (such as sum of map values at atom centers) 
compared to reciprocal space LS or ML functions. This allows programs to 
try more things during refinement and eventually arrive at a better result.

We use all this in phenix.real_space_refine.

Of course a human being moving atoms by hand at a graphics is the 
optimization tool with perhaps largest convergence radius. I would say 
in terms of convergence radius real-space refinement sits in between 
reciprocal-space refinement and human being.

Pavel