I'm a student and not yet rather familiar with Modeller. I've faced
some difficulties when modeling long loops. My target sequences are 35
residues long and have just slight differences (approximately five
residues) from the template. The only restraint to the models is a
disulfide bond connecting the first and the last residues. Several
known structures (completely different) are used as templates; the
result of the modeling each time are structures similar to the one
used as template. The structures are supposed to be variable: nothing
is known about their preferable conformation. Also there is no
suggestion of elements with definite secondary structure.
1) whether it is possible somehow to obtain a set of possible
conformations (not only those similar to the initial ones) by making
changes in initial conditions or with the help of any other means?
Modeller builds models that look like the template by construction, so
the only way to get models that don't look like the template is to
misalign things. For example, if you have two domains with a 'hinge'
between them, you could align your sequence to each domain separately
(using two copies of the template) and align the hinge region to gaps in
the templates. Then, since you have no inter-residue distances in your
model, the two templates will hinge freely. Whether this is applicable
to your problem depends on the structures, of course.
2) can the tools for loop modeling be applied here, because the length
of the loop seems to be excessive. Taking into account the low
accuracy of the models obtained is it then possible using some
assessment method (DOPE for example) to pick out more reliable models?
Loop modeling is not reliable for loops longer than about 12 residues.
Beyond that, you have essentially the same problems as the protein
folding guys.
3) does the DOPE assessment method have any limitations on the minimal
length of the structure and how does its accuracy change with length
if it does?
It's just an atomistic pair potential, so there's no upper or lower
limit on system size. But I can't tell you whether it gives accurate
scores for very small structures, because the only way to test that is
to look at the performance compared to decoys and native structures, and
I doubt we had any very small structures in our benchmark sets (you'd
have to check the publications to be sure).