Computational Approaches in Peptide and Protein Design: An Overview - Peptide and Protein Design for Biopharmaceutical Applications - page 19

ligand-binding complexes (the training sets). Components of these func-
tions may include terms roughly related to free energy of binding,
systems of hydrogen bonding and hydrophobic/hydrophilic interactions
between ligand and receptor, values of solvent-accessible atomic sur-
faces and so on. All terms are usually combined into one linear equation,
and coefficients for each term are calibrated to fit the data on the
training set. A typical empirical scoring function is the one pioneered
by B ¨ hm more than a decade ago [38,39]. Empirical scoring functions
are convenient for the virtual screening of a large number of ligand
compounds that are structurally similar to the relatively small number
of compounds in the training set, but dependence on the given training
set remains a limitation.
Knowledge-based scoring functions use structural data on protein-
ligand complexes instead of the experimental data on binding
energies. The data are analysed in detail to derive simple distance-
dependent potential functions for interactions between specific
atomic groups of ligands and receptors by weighting the probability
of observed experimental distances between the atomic groups as
energies of interaction according to an inverse Boltzmann relation-
ship. In this respect, the knowledge-based scoring functions are, in
fact, statistical atom-atom potentials, which rely on the 'training
sets' of the available high-resolution data on protein-ligand com-
plexes. For instance, the prototypic knowledge-based scoring func-
tion, a potential of mean force initially suggested by analysing the
697 protein-ligand complexes available in 1999 [40], was recently
updated with the data on 7152 protein-ligand complexes available in
the PDB in 2005 [41].
The three types of scoring function are often combined in practical
tasks for peptide and protein design, especially in virtual screening,
into a united procedure of consensus screening, where ligand orienta-
tions with the high scores predicted by different scoring functions are
pooled to compensate for errors due to biases in each single function
(e.g. [34,42]). Many existing scoring functions are incorporated into
available molecular modelling packages and are widely used in drug
design [33,43,44]. However, the search for an optimal scoring func-
tion related to protein-ligand complexes is far from complete.
Examples of the recently described scoring functions are new statisti-
cally-derived atom-atom potentials [45], knowledge-based potentials
accounting for protein flexibility [46,47], a scoring function based on
properties of residues at the ligand-receptor interface [48] and several
others (e.g. [49-51]).