Page 117 - Macromolecular Crystallography

106 MACROMOLECULAR CRYS TALLOGRAPHY

(Schwarzenbacher et al., 2004). Hence the need to communication). Therefore, bytruncatingthemodel
examine critically the multialignement and, in some judiciously, keeping only the conserved core and
cases, to modify it manually. Luckily, it is now pos- removing all the variable loops, one might indeed
sible to couple the multialignement refining process have a useful model.
to the building of models, that is process simulta- So altogether, it seems that the trend of con-
neously 1D and 3D information, and this can be servatively using only models of high sequence
done on the fly. This allows visualization directly identity is now changing, as homology modelling
on the 3D level the effect of a modification of the techniques and low-sequence identity structural
alignment and makes it possible to avoid mean- homolog detection methods are being refined, stim-
ingless alignments (ViTo, (Catherinot and Labesse, ulated by the increase in both the number of
2004)). However, it demands human intervention sequences and the number of structures, and also
and cannot be made automatic. It should be stressed because automatic protocols allow for the testing
that homology modelling techniques can handle of many different models. Indeed, given the num-
effectively the case where several structural models ber of possibilities to generate plausible models,
(templates) are available. clearly there is room for an automatic method trying
different things in turn, and then ranking the differ-
ent solutions. This is precisely what has been done
7.8.3 Detecting low-homology models (using
recently in a suite of programs such as the one called
fold-recognition algorithms)
CaspR (Claude et al., 2004), showing very promis-
Recently, as became apparent during the last Crit- ing results. This is described in more details in the
ical Assessment of Techniques for Protein Struc- following section.
ture Prediction (CASP5) competition (CASP5, 2003),
methods to detect structural homology with vir-
tually no sequence homology have become more
reliable and convincing. They are often based on 7.9 The integrated molecular
so-called metaservers, which address requests to replacement method: comparison
web-based servers of various sorts (secondary pre- of automatic protocols
diction methods, threading methods, etc.) and then 7.9.1 CaspR (Claude et al., 2004)
issue some sort of a consensus score more reliably
than any of the separate methods used (Douguet and CaspR is a combination of well-established, stand-
Labesse, 2001). alone software tools; for a general flowchart of
One might wonder, then, if it would be possible to the program see Fig. 7.3. First it reads an MTZ file
use such methods to pick up remote homologs and (CCP4), and extracts from it the unit cell parameters
use them in MR problems (Jones, 2001). Obviously, and space group number. Then it runs T-Coffee
the obtained models will contain a lot of errors, so (Notredame et al., 2000) and 3D-Coffee (O’Sullivan
why use them? In particular, homology modelling et al., 2004) to get the best possible alignment of
programs rely heavily on the original backbone the template and the target sequences; in doing
coordinates of the template; as there is a well-known so, it identifies variables regions that are likely to
exponential law relating the lack of sequence iden- be non-conserved in the target structure. Then it
tity between two proteins and the RMSD of their runs Modeller (Sali and Blundell, 1993) to construct
coordinates (Chothia and Lesk, 1986), this is indeed 10 different models. All these different models are
worrisome. So at first sight, it might appear that this then subjected to AMoRe (Navaza, 2001) molecu-
kind of model would be useless. However, recent lar replacement protocols, with all sorts of different
re-examination of the same data show that, if one modifications: either truncated from the unreliable
filters out the outliers in the paired atoms of the regions or not, either as poly-Ala or not. Then CNS
structural alignment, the relation between RMSD (Brünger et al., 1998) is used to subject the models
and lack of sequence identity is no longer exponen- with the best MR scores to a round of simulated
tial but simply linear (Martin and Labesse, personal annealing protocol in internal coordinates space.