EOG Electro-olfactogram
GPCR G protein-coupled receptor
QSAR Quantitative structure–activity relationship
SBS Sequential backward selection
SFS Sequential feature selection
VS Virtual screening
ZINC12 Zinc Is Not Commercial database, version 12
1 Introduction
In this chapter, we apply machine learning to analyze the results
from a virtual screening (VS) project for discovering inhibitors of
GPCR signaling in a vertebrate, to infer the importance of func-
tional groups for their biological activity. Computer-based ligand
screening, also known as ligand-based screening, is frequently used
in pharmaceutical discovery because it performs robustly in identi-
fying active molecules from the top-scoring set and does not
require the availability of an atomic structure of the protein target
[1, 2]. Further, it has been shown that ligand-based virtual screen-
ing is capable of exploring different active scaffolds, making it a
valuable alternative to structure-based methods such as molecular
docking, even when atomic structures of the protein target are
known [3, 4].
However, scientists typically focus on the most active handful
of compounds and test their closest analogs while not making use
of the activity data available from all the tested compounds to
identify correlations between their chemical groups and activity
values. Part of this may be due to the need to establish spatial
correspondences between chemical groups in compounds contain-
ing different molecular scaffolds (e.g., comparing substituents on a
steroid ring system versus a purine nucleotide). This problem has
been circumvented in the protocols presented here by considering
all molecules as fully flexible 3D structures and determining their
optimal overlay based on the volumes and partial charges of the
atoms, followed by comparing the chemical identities of neighbor-
ing atoms and small organic groups such as –NH 2. We will use the
term “functional groups” to refer to single or small groups of atoms
that are being compared between molecules. This flexible overlay
procedure provides a rational and quantitative way of comparing
chemical groups between compounds.
The most prominent approaches in the computer-aided discov-
ery of biologically active molecules arestructure-basedscreening
[5–8] andligand-basedscreening [1, 2, 9, 10] as well as hybrids
thereof [11, 12]. Traditionally, structure-based screening is
restricted to applications where an experimentally determined,
high-resolution three-dimensional (3D) structure of the ligand’s308 Sebastian Raschka et al.