Reviews in Computational Chemistry Volume 18

64 The Use of Scoring Functions in Drug Discovery Applications
biological testing. This subset typically consists of ca. 100–2000 compounds
selected from libraries containing 100,000–500,000 compounds. Therefore,
it is essential that the computational process including the scoring function
is fast enough to handle several thousand compounds in a short period of
time. Consequently, only the fastest scoring functions are currently used for
this purpose. Speed is especially important for those scoring functions used
as objective functions during the docking calculations, since they are evaluated
several hundred to a thousand or so times during the docking process of a
single compound. 18
Following a successful virtual screening run, the selected subset of compounds
contains a significantly enhanced number of active compounds as compared
to a random selection. A key parameter to measure the performance of
docking and scoring methods is the so-called ‘‘enrichment factor.’’ It is simply
the ratio of active compounds in the subset selected by docking divided by the
number of active compounds in a randomly chosen subset of equal size. In
practice, enrichment factors are far from the ideal case, where all active compounds
are placed on the top ranks of a prioritized list. Insufficiencies of current
scoring functions, as discussed in the previous section, are partly
responsible for moderate enrichment rates. Another major reason is the fact
that the receptor is still treated as a rigid object in the computational protocols
being used. To generate correct binding modes of different molecules, it is
necessary to predict induced fit phenomena. Unfortunately, predicting protein
flexibility remains extremely difficult and computationally expensive. 189–196
Seeding Experiments
Enrichment factors can be calculated only when experimental data are
available for the full library. But only a few libraries containing experimental
data that have been measured under uniform conditions for all members are
available to the public. Several authors have therefore tested the predictive
ability of docking and scoring tools by compiling an arbitrarily selected set
of diverse, drug-like compounds and then adding to it a number of known
active compounds. This ‘‘seeded’’ library is then subjected to the virtual
screen, and, for the purpose of evaluation, it is assumed that the added active
compounds are the only true actives in the library. Several such experiments
have been published. An example is a study performed at Merck with the
docking program FLOG. 59 A library consisting of 10,000 compounds including
inhibitors of various types of proteases and HIV protease was docked into
the active site of HIV protease. This resulted in excellent enrichment of the
HIV protease inhibitors: all inhibitors but one were among the top 500 library
members. However, inhibitors of other proteases were also considerably
enriched. 197
Seeding experiments allow for comparisons of different scoring functions
with respect to their performance for different targets. Seeding experiments