binding partner (usually a protein or nucleic acid) is available from
X-ray crystallography or nuclear magnetic resonance experiments.
While ligand-based screening does not require knowledge of
the binding target, it assumes that active molecules are likely to
share shape and chemical similarities with a known, biologically
active ligand. In short, ligand-based screening can be described as
a similarity search between a known ligand and the molecules in a
database (Fig.1).1.1 Using Machine
Learning to Identify
Functional Groups
Associated with
Biological Activity
To guide virtual screening, understand biological mechanisms, and
aid the design of more potent inhibitors or activators of molecular
processes, several different techniques have been developed to
analyze datasets of molecular descriptors and measured activity. A
common goal in quantitative structure–activity relationship
(QSAR) modeling includes the prediction of the in vitro or
in vivo activity of molecules given their features. Another common
goal is to gain insights into the importance of individual functional
groups for binding or chemical activity; such insights are invaluable
for the discovery and optimization of potent agonists or inhibitors.
More detailed discussions of QSAR can be found in Kubinyi et al.
[13] and Verma et al.[14].Fig. 1An illustration of the two broad categories of virtual screening: Structure-
based virtual screening involves docking into a binding site to maximize pro-
tein–ligand surface complementarity, and ligand-based virtual screening
involves evaluating small-molecule similarity with a known ligandInferring Activity Discriminants 309