Chemistry Reference
In-Depth Information
affinity prediction scoring of experimental data. This correlation method is an
arithmetic average over all the complexes taking into consideration the number of
tested complexes (N), the experimentally determined binding energies as well as
the calculated scores for all the complexes.
The prediction of binding affinity is more challenging due in part to the origin of
the experimental values (different experimental conditions, research groups,
inherent experimental errors).
It is also possible to evaluate scoring from the capability of selecting and
accurately ranking hits from databases. The accumulated rate of active hits found
above a percentage of the ranked database (binders and inactive ligands) yields
the enrichment. Maintaining a fixed percentage, a higher enrichment indicates a
better scoring function.
Using AUC is another procedure for evaluation scoring in virtual database
screening,
i.e.
using the area under the receiver operating characteristic curve
(ROC) [551]. In cases when the number of active binders is comparable to the
inactive ligands, this method could be more appropriate. Due to inherent
limitations, most scoring functions should, but do not perform well for all the
procedures previously discussed. It is noteworthy that good binding affinity
prediction is not equivalent to good database ranking.
Construction of appropriate test/training sets is also important for evaluating
scoring. High quality/resolution structures are prioritized. Whenever possible
diverse protein type and a wide range of binding affinities should be used. Bond-
covalent binding and drug-like ligands should be prioritized. It is important to
avoid training/test sets overlaps. Publicly available benchmarks (CSAR/,
(
http://csardock.org
), DUD (
http://dud.docking.org
) and others are invaluable.
In 2013 it was reported a comparison of neural-network scoring functions and the
state of the art applications to common library screening. Neural networks are
computer models expected to mimic the microscopic architecture and
organization of the brain. There is
in silico
simulation of biological neurons and
connections. Data encoded on the neurodes and signals are propagated through the
