Biology Reference
In-Depth Information
trial, research groups apply their prediction methods to sequences for
which the experimental structure is about to be determined; the accuracy
of these blind predictions is then assessed independently once the struc-
tures are made available.
8,38,59,60
There are also web servers LiveBench
61
and EVA
62
that assess protein structure prediction servers on an auto-
mated and continuous basis, using sequences from the PDB, before their
structures are released as modeling targets.
Retrospective assessment of the average accuracy of individual mod-
eling methods by projects such as CASP and EVA is valuable for the
development of modeling techniques, but unfortunately does not allow
drawing any conclusions about the accuracy of a specific model as the
correct answer is unknown in a real-life situation. Since the usefulness of
predictions crucially depends on their accuracy, a means of reliably pre-
dicting the likely accuracy of a protein structure model in the absence
of its known 3D structure is an important problem in protein structure
prediction. Accurate estimates of the errors in a model are an essential
component of any predictive method — protein structure prediction is
not an exception.
Different scoring schemes have been developed to determine
whether or not a model has the correct fold, to discriminate between
native and near-native states, to select the most near-native model in a
set of decoys, and to provide quantitative estimates for the coordinate
error of the predicted amino acids. A variety of methods have been
applied to address these tasks, such as physics-based energies, knowl-
edge-based potentials,
23
combined scoring functions, and clustering
approaches. Combined scoring functions integrate several different
scores, aiming to extract the most informative features from each of the
individual input scores.
63
Clustering approaches use consensus infor-
mation from an ensemble of protein structure models provided by
different methods.
64
Some structural aspects of a protein model can be verified using
methods based on fold recognition methods. These methods rely on
empirical pseudo-conformational energy potentials derived from the
pairwise interactions observed in well-defined protein structures. These
terms are summed over all residues in a model and result in a more (more
negative) or less (more positive) favorable energy. These methods can
