Homology Modeling: Generating Structural Models to Understand Protein Function and Mechanism - Computational Modeling of Biological Systems - page 110

compared structures that fall within a given distance cut-off. Usually, an average
of GDT at 1, 2, 4, and 8 A is used to measure structure quality, and is denoted as
GDT TS. Other variations include corrections to eliminate size-dependence [ 52 ]
and also to include a negative term for incorrectly positioned amino acids (resulting
in non-native contacts) [ 53 ]. There are various servers that have been developed to
predict GDT TS of a structural model in the absence of an experimental structure
with which to compare.
These quality assessment programs either assess a structural model by itself
(single model) [ 54 - 56 ], or in the context of a large reference set of struc-
tural models generated for a given sequence using different methods (consensus)
[ 57 - 59 ]. The single model methods can compare various structural parameters
such as secondary structure and solvent accessibility (whether a given residue is
buried or exposed) that are predicted for a given sequence with the corresponding
structural parameters of the given model. The scoring functions used in single-
model quality assessment programs also include knowledge-based potentials like
a sequence-dependent torsion term (usually in the context of three consecutive
residues), distance and cut-off based residue-residue interaction potentials and all-
atom interaction potentials. Consensus quality assessment programs rely on the idea
that if a diverse set of methods were used in generating many structural models for a
given query, models that incorporate the best of all methods will be the ones closest
to the experimental structure. How do the consensus methods select models based on
these criteria? Using a large reference set of models for a given sequence (obtained
from the various structure prediction servers), they determine the average distance
of a given model to all other models. The distance measure used in most cases is
GDT TS, although specific servers apply various modifications to the distance to
obtain better predictive power. Through several rounds of CASP, it is apparent that
the models with least average distance to the rest of the reference set feature the
best GDT TS when compared to the experimental structure. We have to emphasize
here that there are many mathematical formulations used in modifying the simple
average distance to obtain better predictions, but these formulations may not always
have a strong physical basis. Interestingly, weighting the average distances with
single-model score yields very good prediction of GDT TS of a model with respect
to the experimental structure [ 60 ]. Thus, based on the experience gained from CASP
competitions, the ideal strategy for constructing a homology-based structural model
would entail generating several models using heterogeneous methods, which can
also include human intervention during model building and the incorporation of
known experimental constraints. Once a handful of models are obtained, one can
use the quality assessment programs to obtain a prediction of how close the best
model will be to the experimental structure. The quality score will in turn determine
the use to which a structural model can be put (discussed below).
Apart from model accuracy, an important criterion for model quality is the
stereochemical quality of the given model. The stereochemical quality here broadly
defines the acceptable quality of the covalent geometry and the core-packing of a
given structural model. The covalent geometry of a structural model is assessed
by comparing all its bond lengths, bond-angles, and torsions to standard values.