Reviews in Computational Chemistry Volume 18

Application of Scoring Functions in Virtual Screening 63
Rank Ordering Sets of Related Ligands
For structure prediction, structures of protein–ligand complexes from the
PDB can serve as a common pool to test scoring functions. It is more difficult
to draw valid conclusions about the relative performance of scoring functions
to rank order sets of ligands with respect to their binding to the same target.
First, there are few published studies in which different functions have been
applied to the same data sets. Second, experimental data are often not measured
under the same conditions but collected from various literature references.
The latter practice can have especially dramatic effects when
inhibitory concentrations for 50% reduction of a biological effect (IC 50 data)
are used instead of Ki values.
On average, empirical scoring functions seem to lead to better correlations
between experimental and calculated binding energies than do force field
based approaches because the nonbonded interactions in the latter are usually
not optimized to reproduce individual intermolecular binding phenomena.
However, the only available calculated data for most published functions
are those for the complexes used in the derivation of the functions themselves.
Very promising results of rank ordering have also been obtained with the
knowledge-based functions DrugScore 93 and PMF. 86,88,182
The task of rank ordering small (ca. 10–100) sets of related ligands with
respect to a target can also be accomplished with methods that are computationally
more demanding than simple scoring functions. The most generally
applicable methods are probably force field scores augmented with electrostatic
desolvation and surface area terms. An example is the MM–PBSA
method that combines Poisson–Boltzmann electrostatics with AMBER MD
calculations. 183 This method has been applied to an increasing number of
studies, and it has led to promising results. 106–108,184 Poisson–Boltzmann
calculations have been performed on a variety of targets with many related
computational protocols. 102,138,139,185–188 Alternatively, extended linear
response protocols 112 can be used. The OWFEG grid method by Pearlman
has also shown promising results. 114
APPLICATION OF SCORING FUNCTIONS IN
VIRTUAL SCREENING
In recent years, virtual screening of large databases has emerged as the
central application of scoring functions. In the following sections, we describe
special requirements that scoring functions must fulfill for successful virtual
screening, and we indicate the level of accuracy that can nowadays be
expected from virtual screening.
As discussed in the introductory sections, the goal of virtual screening is
to use computational tools together with the known 3D structure of the target
to select a subset of compounds from chemical libraries for synthesis and