لخّصلي

Ligand-based drug design (LBDD) approaches are applied
when the three-dimensional structure of the target is unknown
or when it is not fully reliable (i.e., homology models of the
target are based on poor sequence identity) [1]. In this case, the
experimental activity data of active compounds are generally
used to construct pharmacophore models. These models include
an ensemble of steric and electronic features (usually indicated
as descriptors) that are likely to be those mainly responsible for
activity (Fig. 8). Typical descriptors are physicochemical
features such as molecular weight, geometry, surface accessible
area, aromaticity index, electronegativity, polarizability, and
solvation properties. In general, a proper set of descriptors
should cover a broad chemical diversity space. In this way, the
more relevant descriptors for building predictive
pharmacophores are likely to be recognized.
These are then used to help identify new active compounds, for
example, in a virtual library of drug-like molecules. The same
principles are valid for molecular similarity methods, where
similitudes in geometrical or physicochemical properties among
active compounds are used to identify new active compounds
from among the many that can be virtually screened. A
similarity coefficient, such as the Tanimoto or Euclidean
coefficients, is then used to identify whether a set of new
compounds is likely to interact with a given target [5].
Geometrical and electronic descriptors are used to build a
quantitative structure–activity relationship (QSAR), which
results in a mathematical model able to predict the biological
activity of new compounds [5]. For example, activity data of
active molecules can be used to extract mathematical models
for early prediction of activity or, more often, metabolic
properties that could generate toxicity problems. Con- ventional
3D QSAR models consider the ensemble of conformations,
orientations, tautomers,
Fig.8 An example of apharmacophoremodelof(R)-roscovitine,which
tightly bindsCDK5[23].Crucialinteractingresidues of CDK5 are
shown in green (hydrophobic residues), blue (H-bond acceptors),
and magenta (H-bond donors). The regions of the ligand that match
those types of interaction are indicated with the same color
stereoisomers, and protonation states of the initial ligand set. More
exotic multidimensional 4D or 5D QSAR models have also been
developed, which take into account all energy contributions of
ligand binding, including solvation energy and conformational
entropy. Usually, linear regression and principal component analysis
can be combined for a partial least square analysis that directly
relates the reduced set of descriptors to the biological activity.
However, a nonlinear relation is often present between the
biological activity and the descriptors used.
Therefore, nonlinear regression models using machine- learning
algorithms have recently been introduced in QSAR. Of these,
artificial neural networks algo- rithms are used to discover the
relationship between descriptors and biological activity via an
iterative process [21].
Absorption, distribution, metabolism, excretion, and toxicity
(ADMET) are the key parameters that can be optimized with QSAR
approaches to get a drug candidate with a proper drug-like
pharmacokinetic (PK) profile. ADMET and PK properties are
critical to the success of any drug discovery program. A poor drug-
like PK profile can result in the rapid metabolism and elimination of
the compound from the body.
A drug-like compound should be characterized by certain
physicochemical properties, which are loosely recapitulated in
Lipinski’s rule of five. This says that an optimal orally bioavailable
drug should have a molecular weight less than 500, less than 5 H-
bond donor sites, and less than 10 H-bond acceptor sites, while the
log of the octanol/water partition coefficient (logP), a measure of
hydrophobicity, should be below 5. Despite its practical utility for
quickly evaluating drug-likeness, it is nowadays clear that this rule
can only estimate the compound’s probable success by highlighting
potentially problematic aspects of its physicochemical and structural
profile [22]. In fact, many approved drugs, such as large macrolide
antibiotics (MW 1000), differ in one or more descriptors from
Lipinski’s rule. Other important features can heavily affect the PK
profile. For ideal absorption and distribution, an aqueous solubility
above 10 6 M is advised (logS >
6).
Transport properties, such as membrane permeability and
brain/blood partitioning (log BB), for allowing blood–brain barrier
penetration must also be taken into account in certain drug
discovery projects. Additionally, interactions with influx/efflux
transporter proteins or metabolic enzymes can greatly affect the
final therapeutic effect and should be considered in the early phases
of a drug discovery program. That is, for a given molecular
structure, properties such as solubility, membrane permeability,
partition coefficients, blood–brain barrier penetration, plasma
protein binding, and metabolite formation are computed and
compared against a database of known drugs. Computational tools
and methods for predicting ADMET properties and producing
predictive statistical models are thus valuable. These statistical
models are trained on experimental ADMET data obtained from
several tested compounds.
These computational tools are therefore crucial for selecting,
prioritizing, and optimizing promising lead compounds.