Computational Approaches in Peptide and Protein Design: An Overview - Peptide and Protein Design for Biopharmaceutical Applications - page 27

number of significantly different folds in the PDB [68,69]. Therefore,
threading was developed to suggest plausible folds for given sequences
when direct sequence alignment with those in the PDB resulted in homol-
ogy of less than 30% (see also [70]). Threading was not expected to yield
an accurate 3D model of the target protein, but rather a rough model
(such as C a -trace) similar to a chosen 3D template. A suitable template
may be determined by various procedures, such as by establishing distant
evolutional relations between the sequence of the target protein and the
sequences available in the PDB [71]; by building templates out of frag-
ments possessing high sequence homology to some parts of the target
sequence only [72]; by involving statistical predictions of the regular
elements in the sequence, such as fragments of a-helices and b-strands
[73]; or by combining these procedures with one another. To evaluate the
quality of the templates with the suggested 3D models for the targets,
threading procedures use various sequence-structure fitness functions.
Threading is also often used as the first step in building much more
detailed 3D structures of proteins. After an approximate 3D model of
the protein chain has been obtained from threading, it is refined through
calculation of an appropriate scoring function (e.g. the statistically-
derived residue-residue potentials) and then in all-atom approximation,
adding backbone atoms and sampling different rotamers of the side
chains, as exemplified in one of the recent studies [74]. Recently, several
servers accessible on the Internet offered automated threading of protein
sequences, employing various threading techniques or their combinations
(e.g. [75]). The best predictions of 3D structures of proteins based on
automated servers are in the range of about 2-6 ˚ differences in the
atom-atom distance RMS values calculated between the predicted and
experimental structures (see assessment [76]). There have also been
recent attempts to derive 3D structures for sequences that proved difficult
for threading by developing specific neural network-based procedures for
machine learning [70].
De novo predictions of the protein structure became successful in the
last several years, mostly due to the development of fragment-based
assembly, the recently christened Rosetta algorithm [77]. The basic
idea was initiated after an analysis of six-residue fragments of proteins
in the PDB showed rather discrete clusters of conformers. Based on
the limitation in the statistics of the observed conformer vocabulary,
combinatorial assembly of hexamer fragments of known proteins present
in the targeted sequence was suggested as an approach to generate native-
like 3D structures for the target. Modern-day Rosetta starts with
selection of the sets of the backbone conformations represented in the