I divided the protein with know structure (bovine rhodopsin receptor)
into three part: N-terminus region, transmembrane region and C-terminus
region without change their initial coordinates, and I try to use the
three parts to reconstruct the original structure of the protein.
...
However, the generated model was different from the original one,
looking closely, I found the difference mainly occurred in the
N-terminus and C-terminus regions, their coordinates have changed
greatly when compared with the original structure.
Modeller will try to build models that resemble the template whenever it
is aligned with your target sequence, but these restraints are on
internal coordinates, not Cartesian coordinates, so there could well be
rotation or translation of your three segments. In particular, by
breaking the sequence into three, you have prevented Modeller from using
the information present in your original structure about the distances
between these three segments.