Re: [ccp4bb]: beast

[Randy Read <rjr27@cam.ac.uk>]

***  For details on how to be removed from this list visit the  ***
***          CCP4 home page http://www.ccp4.ac.uk         ***

On Thursday 07 November 2002 18:48, Daniele de Sanctis wrote:
> ***  For details on how to be removed from this list visit the  ***
> ***          CCP4 home page http://www.ccp4.ac.uk         ***
>
> i'm trying to use beast to make a molecular replacement with a polyala
> model, and i need a few advices to understand the output: which is a
> good value for the LLG? and what is consider a cluster and a member?

I can't really tell you what is a good value for the LLG.  First, to put it 
on a scale that could be compared among different problems, it would have to 
be LLG/reflection.  Second, the LLG to expect depends on what is being 
tested, and how precisely you'd expect to be able to explain the data.  To be 
more specific, if you're carrying out a rotation search the probability 
distribution is broader than for a translation search because of the extra 
uncertainty introduced by unknown relative phase angles of the contributions 
of symmetry-related molecules.  In a rotation search, higher symmetry leads 
to flatter probability distributions (more contributions with unknown 
relative phases) and lower LLG scores.  In any molecular replacement search, 
model incompleteness and model error also lead to lower LLG values.  What we 
really need, to tell what is a good LLG score, is a statistical analysis of 
the probability distributions for LLG scores for correct solutions and 
incorrect solutions, but this is a complicated problem.

The bottom line is that, at the moment, the best we have to go on is some 
general rules of thumb.  First, the LLG should be positive, because it 
measures how much better we can predict the data than if we use the Wilson 
distribution (i.e. our structure contains atoms in random positions); if it's 
negative, our model is worse than random.  (If that happens in a search, 
increase the assumed RMS error or think again about model completeness.)  
Second, there should be good signal to noise between the highest LLG and the 
mean LLG for the whole search.  Values greater than 5 sigma are a good sign, 
but the best score may be around 4 sigma for difficult rotation searches.  
Third, there should be good contrast between the top (hopefully correct) 
solution and the next highest solution.  A difference greater than 1 sigma 
gives me confidence, but even 0.5 sigma is reasonable.  Third, the LLG should 
increase as you make your model of the crystal structure more detailed, i.e. 
as you go from a rotation to a translation search, or as you add more 
molecules.

In difficult cases, the correct orientation may not give the highest LLG 
score, but in anything I've been able to solve the correct orientation was 
near the top.  (Of course, that may reflect on how slow Beast is, because I 
can never afford the computing time to carry out a full translation search 
for, say, 100 or 1000 orientations!)  So, unless the answer from a rotation 
search is unambiguous, it's a good idea to carry out a translation search 
with several orientations.

On the specific point of using a poly-Ala model:  I tend not to trim the 
model down, on the principle that the side-chain atoms will be close to 
something in the true structure, even if the side chains are different.  But 
if you do trim the model, then you should remember that you are making your 
model less complete, and adjust the completeness parameter accordingly.

The clustering algorithm is basically a simple peak-picking procedure.  The 
LLG scores are sorted, then the solution with the top LLG score represents 
the first cluster.  We go down the list, and anything that is sufficiently 
similar to something higher in the list becomes a member of the same cluster. 
Anything that isn't similar to something higher in the list represents a new 
cluster.

I hope that answers your questions.

-- 

Randy J. Read
Department of Haematology, University of Cambridge
Cambridge Institute for Medical Research      Tel: + 44 1223 336500
Wellcome Trust/MRC Building                   Fax: + 44 1223 336827
Hills Road                                    E-mail: rjr27@cam.ac.uk
Cambridge CB2 2XY, U.K.                       www-structmed.cimr.cam.ac.uk