molecules

Review
Computer-Aided Drug Design towards New Psychotropic and
Neurological Drugs
Georgia Dorahy 1,2 , Jake Zheng Chen 1,2 and Thomas Balle 1,2, *

1 Sydney Pharmacy School, Faculty of Medicine and Health, The University of Sydney,
Sydney, NSW 2006, Australia
2 Brain and Mind Centre, The University of Sydney, Camperdown, NSW 2050, Australia
* Correspondence: thomas.balle@sydney.edu.au

Abstract: Central nervous system (CNS) disorders are a therapeutic area in drug discovery where
demand for new treatments greatly exceeds approved treatment options. This is complicated by the
high failure rate in late-stage clinical trials, resulting in exorbitant costs associated with bringing
new CNS drugs to market. Computer-aided drug design (CADD) techniques minimise the time
and cost burdens associated with drug research and development by ensuring an advantageous
starting point for pre-clinical and clinical assessments. The key elements of CADD are divided into
ligand-based and structure-based methods. Ligand-based methods encompass techniques including
pharmacophore modelling and quantitative structure activity relationships (QSARs), which use the
relationship between biological activity and chemical structure to ascertain suitable lead molecules.
In contrast, structure-based methods use information about the binding site architecture from an
established protein structure to select suitable molecules for further investigation. In recent years,
deep learning techniques have been applied in drug design and present an exciting addition to
CADD workflows. Despite the difficulties associated with CNS drug discovery, advances towards
new pharmaceutical treatments continue to be made, and CADD has supported these findings. This
review explores various CADD techniques and discusses applications in CNS drug discovery from
2018 to November 2022.

Keywords: structure-based drug design; ligand-based drug design; artificial intelligence; docking;
Citation: Dorahy, G.; Chen, J.Z.; Balle, QSAR; pharmacophore; deep learning; molecular dynamics; Alzheimer’s disease; schizophrenia;
T. Computer-Aided Drug Design
neuropathic pain; neurological; psychotropic; virtual screening; computer-aided drug design
towards New Psychotropic and
Neurological Drugs. Molecules 2023,
28, 1324. https://doi.org/10.3390/
molecules28031324
1. Introduction
Academic Editors: Lee Wei Lim and
Mental disorders including neurological and psychiatric disorders represent an area of
Luca Aquili
medicine where there is a considerable unmet need for new and more advanced treatments.
Received: 15 December 2022 Diseases such as Alzheimer’s disease and Parkinson’s disease only have treatments avail-
Revised: 23 January 2023 able that provide symptomatic relief [1,2]. As these diseases progress, the efficacy of current
Accepted: 26 January 2023 therapies wanes and is no longer able to manage these conditions. Mental illnesses such as
Published: 30 January 2023 schizophrenia have several treatment options available; however, they are associated with
a plethora of adverse drug reactions that can severely affect a patient’s physical health due
to cardiometabolic syndrome [3]. For other neurological conditions, such as brain injuries,
there is no treatment or cure. Thus, designing drugs to prevent or halt neuronal death and
Copyright: © 2023 by the authors.
subsequent deficits is necessary and urgent.
Licensee MDPI, Basel, Switzerland.
The discovery of new drugs targeting mental disorders is associated with some of the
This article is an open access article
highest fail-rates in drug discovery, with 85% of drugs failing in phase II and III clinical
distributed under the terms and
trials [4]. This makes the development of central nervous system (CNS) drugs extremely
conditions of the Creative Commons
expensive, given their tendency to fail in later stage trials [5], with an estimated cost of
Attribution (CC BY) license (https://
upwards of $2 billion to bring a drug to market in 2019 [6]. Ensuring the best possible
creativecommons.org/licenses/by/
4.0/).
starting point for new discovery projects is imperative, and computer-aided drug design

Molecules 2023, 28, 1324. https://doi.org/10.3390/molecules28031324 https://www.mdpi.com/journal/molecules

Molecules2023,
Molecules 2023,28,
28,1324
x FOR PEER REVIEW 22 of
of 22
24

design (CADD)
(CADD) techniques techniques are important
are important in this These
in this context. context. These methods
methods are an attractive
are an attractive starting
starting point for new projects and have become one of
point for new projects and have become one of the mainstays in the early drug the mainstays in the early
discoverydrug
discovery
process process
given their given
reduced their reduced
time and labourtime and labour
intensity inintensity
comparison in comparison
to traditional to drug
tradi-
tional drug
design design and testing.
and laboratory laboratory testing.
Thus, CADD Thus, CADD
can help can help to the
to shorten shorten
timethe fromtime from
initial
initial research
research to bringing
to bringing a drug to a drug
market to market and alleviate
and alleviate the highthe high associated
associated costs. costs.
CADDisistypically
CADD typicallyclassified
classifiedinto
into ligand-based
ligand-based andand structure-based
structure-based methods
methods (Figure
(Figure 1).
Ligand-based methods work on the principle that the chemical structure of a drug isis
1). Ligand-based methods work on the principle that the chemical structure of a drug
relatedto
related toits
itsbiological
biologicalactivity.
activity.Thus,
Thus,with withaaseries
seriesofofknown
knownactive activeand
andinactive
inactiveligands
ligands
athand,
at hand,structure
structureactivity
activityrelationships
relationships (SARs)(SARs) cancan bebe derived
derived and and used
used to to predict
predict new new
andbetter
and bettermolecules.
molecules.The Themain
mainchallenge
challengein inligand-based
ligand-baseddrug drugdesign
designisishowhowto todescribe
describe
chemical
chemicalstructure.
structure. This
This can
canbebedone
doneto tovarious
variouslevels
levelsof ofsophistication
sophisticationranging
rangingfrom from2D 2D
to
to3D3Ddescriptors.
descriptors.Structure-based
Structure-baseddrug drugdesign
designmethods
methodsrely relyon onthe
theknowledge
knowledgeof ofthethe3D
3D
structure
structureof ofthe
thebiological
biologicaltarget.
target.This
Thisprovides
providesan aninsight
insightinto
intothethebinding
bindingsite sitearchitecture
architecture
which
whichmay may then be beutilised
utilisedtotoassess
assess if aifligand
a ligand would
would make make a suitable
a suitable lead molecule
lead molecule based
based
on theon the binding
binding site interactions.
site interactions. Both ligand-
Both ligand- and structure-based
and structure-based methods methods
rely on rely on
good-
goodness
ness of fitofforfit for small
small molecules
molecules to to select
select compoundsthat
compounds thatmaymaybe be best
best suited for for further
further
research. Molecules are
research. Molecules arescored
scoredaccording
accordingtoto features
features such
such as similarity,
as similarity, correlations
correlations to
to spe-
specific molecular properties, or binding energy, making them
cific molecular properties, or binding energy, making them ideal techniques for the earlyideal techniques for the
early
stages stages
of theofdrug
the drug discovery
discovery process.
process. AlthoughAlthough
thesethese techniques
techniques will be will be discussed
discussed sepa-
separately,
rately, it is important to note that they are often used in concert to yield more accurate
it is important to note that they are often used in concert to yield more accurate
results
resultsandandtotoreduce
reducethe thecomputational
computational burden
burden when
when chemical
chemical libraries being
libraries screened
being are
screened
large. The result
are large. is a complimentary
The result is a complimentary process towards
process new new
towards chemical entities.
chemical entities.

Figure1.1.AAdrug
Figure drugdesign
designworkflow
workflowincluding
includingthe
thestages
stagesof
ofCADD.
CADD.Both
Bothstructure-
structure-and
andligand-based
ligand-based
applications are outlined. A sample of chemical libraries and software applications
applications are outlined. A sample of chemical libraries and software applications usedused in the dif-
in the
ferent stages of the workflow are highlighted. It must be noted that these lists are not exhaustive
different stages of the workflow are highlighted. It must be noted that these lists are not exhaustive
and other libraries and software applications are available for use.
and other libraries and software applications are available for use.

Inthis
In thisreview,
review,wewesummarise
summarisethethe most
most common
common CADD
CADD methods,
methods, including
including homol-
homology
ogy modelling, molecular docking, molecular dynamics simulations,
modelling, molecular docking, molecular dynamics simulations, pharmacophore mod- pharmacophore
modelling,
elling, quantitative-structure-activity
quantitative-structure-activity relationship
relationship (QSAR)
(QSAR) methods,
methods, andmore
and the the more re-
recent
cent deep
deep learning
learning (DL) applications
(DL) applications that have
that have provenproven their efficacy
their efficacy in CADD.
in CADD. Further-
Furthermore,
more, we provide examples of their application in drug discovery projects for
we provide examples of their application in drug discovery projects for psychiatric and psychiatric
Molecules 2023, 28, 1324 3 of 22

neurological conditions. The selected examples are from 2018 to January 2023 and focus on
applications that are supported by experimental data to demonstrate the validity and value
of the in-silico methods applied. The examples cover a wide range of biological targets
including cannabinoid receptor 1, acetylcholinesterase and the α7 nicotinic acetylcholine
receptor. We demonstrate that CADD is an important tool in the discovery of CNS drugs,
especially during the early stages of the drug discovery pipeline.

2. Ligand Based Techniques

Ligand-based drug design (LBDD) uses data from a known ligand or set of ligands
to anticipate the activity of other small molecules based on the commonality with the
molecules being investigated. This works on the principle that compounds with a similar
chemical structure will exhibit similar binding properties.

2.1. 2D Based Methods

In 2D based methods, the chemical structure is represented as a set of descriptors,
which are then used to screen for molecules possessing similar properties. A range of
physicochemical properties and molecular features are calculated from a 2D molecular
graph [7]. These properties, including molecular weight and logP [8], are essential in
assessing the suitability of a molecule as a drug based on Lipinski’s rule of five [9]. By
extension, the physicochemical properties provide information about how likely a ligand
is to transverse the blood–brain barrier (BBB) [10], which is an integral consideration in
CNS drug development. Molecular features, such as topological indices, fragment counts
and substructure counts can also be ascertained from a 2D molecular graph. The molecular
and physicochemical properties provide the data needed to find new lead compounds
using techniques such as similarity searching [11,12]. A set of molecular descriptors will
be defined to screen for the ligands which best match and thus make suitable candidates
for further research. Although details of the 3D molecule are not considered with this
technique, it is an attractive approach for ultra-large chemical libraries due to the reduced
computational burden associated with screening for hit compounds.

2.2. Pharmacophore Modelling

A pharmacophore model is a collection of chemical features that defines the inter-
actions of a ligand with a biological target that elucidates a physiological response [13].
Pharmacophore modelling can also be used to differentiate whether a ligand would behave
as an agonist or antagonist. This technique is particularly useful where detailed informa-
tion of the structure of the target is not available. Furthermore, indirect information about
binding site features based on the molecular properties of known active molecules can be
inferred from pharmacophore modelling. Pharmacophore modelling was applied in search
for antagonists of the α7 nicotinic acetylcholine receptor (nAChR) prior to the 3D structure
of this protein being solved. A high potency lead, T761-0184, was uncovered, providing a
novel lead for the treatment of CNS disorders such as schizophrenia [14].
3D-pharmacophore models are comprised of molecular features, such as the steric
and electrostatic properties that allow the ligand to bind to the target receptor to produce
a pharmacological response. Functional groups represented as hydrogen bond donors or
acceptors, charges, hydrophobic and aromatic are identified in the pharmacophore [15]. In
essence, pharmacophoric modelling uses a training set of at least two ligands to generate a
3D map of the key molecular features (Figure 2), which is then validated using a test set of
other active molecules and decoys [16]. Decoys can be obtained from the Database of Useful
Decoys: Enhanced (DUD-E), which is a database of biologically inactive small molecules
with similar physicochemical properties [17]. This technique is particularly useful as a pre-
screening tool for the filtering of ultra large chemical libraries prior to further assessment
using more computationally expensive methods such as molecular docking. A similar
workflow was used in pursuit of human dopamine transporter (hDAT) antagonists for
new Parkinson’s disease therapies. A pharmacophore was also generated using 25 known
Molecules 2023, 28, x FOR PEER REVIEW 4 of 24
Molecules 2023, 28, 1324 4 of 22

antagonists for new Parkinson’s disease therapies. A pharmacophore was also generated
using 25ligands
active known and
active
50ligands
decoysandas a50 decoysset
training as and
a training set and 50
50 substrates substrates
and and for
2500 decoys 2500
the
decoys for the validation set. The top scoring 1000 ligands from the docking
validation set. The top scoring 1000 ligands from the docking studies of 300,000 ligands studies of
300,000 ligands then
then underwent underwent pharmacophore
pharmacophore screening, whichscreening, whichsix
uncovered uncovered
promisingsixmolecules.
promis-
ing molecules.
Three Three leads demonstrated
leads demonstrated a statistically asignificant
statistically significant
inhibition inhibition
of hDAT of hDAT
receptor uptakere-in
ceptor uptake
biological in biological testing.
testing.

Figure 2. A sample pharmacophore of two known P2X7 antagonists coloured in fuchsia and green.
Figure 2. A sample
The ligands have pharmacophore of two
been clustered such known
that P2X7
similar antagonists
molecular coloured
features in fuchsia
are aligned withand green.
each other
The ligands have been clustered such that similar molecular features are aligned with
and the pharmacophore. This sample pharmacophore constitutes five key features. Two aromatic each other
and the pharmacophore. This sample pharmacophore constitutes five key features. Two aromatic
groups are represented by the orange rings, two hydrogen bond acceptors by pink spheres and
groups are represented by the orange rings, two hydrogen bond acceptors by pink spheres and
hydrogen bond donors by light blue spheres. The grey spheres surrounding these features are known
hydrogen bond donors by light blue spheres. The grey spheres surrounding these features are
as exclusion
known volumes,
as exclusion whichwhich
volumes, mimic whatwhat
mimic the protein binding
the protein pocket
binding is expected
pocket to look
is expected like.like.
to look Thus,
query
Thus, ligands
query must must
ligands not enter
not these
enter regions to prevent
these regions steric clashes.
to prevent The collection
steric clashes. of these features
The collection of these is
what isisexpected
features to contribute
what is expected most to drug
to contribute mostreceptor
to drug interactions.
receptor interactions.

ToTodevelop
developa robust
a robust model
model that
that is is representative
representative ofof
allall known
known active
active molecules,
molecules, phar-
phar-
macophore generation must be an iterative process whereby models
macophore generation must be an iterative process whereby models are scored and re-are scored and refined
untiluntil
fined the most suitable
the most solution
suitable is ascertained.
solution Typically,
is ascertained. the scoring
Typically, and generation
the scoring of phar-
and generation
macophore models can be categorized into either overlay or root mean
of pharmacophore models can be categorized into either overlay or root mean square de- square derivative
(RMSD)
rivative scoring
(RMSD) [18]. [18].
scoring Overlay methods
Overlay methodsgenerate andand
generate score a hypothesis
score a hypothesisforfora apharma-
phar-
macophore by matching the radii of chemical features from the alignment of moleculesinin3D
cophore by matching the radii of chemical features from the alignment of molecules
3Dspace
space[19,20]. Alternatively,
[19,20]. Alternatively,RMSD
RMSD uses thethe
uses measured distances
measured between
distances a functional
between group
a functional
on a on
group query molecule
a query and and
molecule the pharmacophore
the pharmacophore model [21,22].
model Pharmacophore
[21,22]. Pharmacophore modelling
mod-
programs are reviewed in greater detail by Sanders et al., and Giordano et al. [18,23].
elling programs are reviewed in greater detail by Sanders et al. and Giordano et al. [18,23].
2.3. QSAR
2.3. QSAR
The concept of the quantitative structure–activity relationship (QSAR) is underpinned
by The concept ofbetween
the correlation the quantitative structure–activity
the physicochemical relationship
properties (QSAR)features
and topological is under- of a
pinned
molecule and the biological activity they exert on a target [24]. QSAR studies usefea-
by the correlation between the physicochemical properties and topological this
tures of a molecule and the biological activity they exert on a target [24].
relationship to filter and rank libraries of molecules and predict biological activity. TheseQSAR studies
use this relationship
predictions to filter
are made and rank
possible by thelibraries of molecules
use of statistical and predict
methods biological
to correlate activity.
molecular de-
These predictions are made possible by the use of statistical methods to correlate
scriptors to biological data such as binding affinity (KD ) or functional potency (EC50 or molecu-
larICdescriptors to biological data such as binding affinity (K D) or functional potency (EC50
50 ) values. Data may be obtained in-house (for example, using proprietary data), or, more
orcommonly,
IC50) values. Data may
chemical be obtained
databases (Figurein-house (fortoexample,
1) are used using
access data for proprietary data),
the training and or,
testing
more commonly,
of QSAR models.chemical databasesmodels
These predictive (Figureare1) are used touseful
especially accesswhen
data for the training
attempting and
to design
testing
a drugofwith QSAR models.
multiple Thesesuch
targets, predictive
as thosemodels
in the are especially
examples useful in
presented when attempting
chapter 6. QSAR
tomodelling
design a drug with
is also multiple
attractive fortargets, such as those
lead optimisation in the the
through examples presented
identification in chapter
of areas respon-
6.sible
QSAR for biological activity. A 3D-QSAR model was utilised for the lead optimisationofof
modelling is also attractive for lead optimisation through the identification
Molecules 2023, 28, 1324 5 of 22

phosphodiesterase 4 (PDE4) inhibitors that could be used for major depressive disorder [25].
Previous leads were optimised by the addition of hydrophobic and hydrogen bonding
groups that extended into other pockets of the active site. In vitro assessments indicated
that the new compounds had nanomolar IC50 values and demonstrated anti-inflammatory
properties in microglial cells.
Ideally, the molecules used in model generation should be split into approximately
80% for training and 20% for testing in addition to an external validation set [26,27]. It
is important that the molecules in the training set are as chemically dissimilar as possi-
ble [27] and the bioactivity data of these chemicals is distributed across the full range of
endpoints [28] to ensure validity of predictions and minimise biases. Given these molecules
cannot capture the full breadth of chemical space, it is essential to define an applicability
domain (AD). An AD stipulates the area of chemical space for which the model can make
predictions with good reliability [29]. This is one of the five guidelines outlined by the
Organisation for Economic Co-operation and Development (OECD) recommendations for
valid QSAR development [26].
The dimensionality of molecular descriptors in QSAR may range from zero-dimensional
to six-dimensional (0D to 6D). Increasing dimensions of chemical representation will in-
crease the level of detail about molecules in the QSAR model and, subsequently, the
complexity and computing power needed. The examples in this study are limited to 3D
methods, and therefore dimensions 4D–6D will not be discussed. All dimensions of QSAR
are explained and reviewed in detail by Manoj et al. [30]. Table 1 outlines the details of
molecular descriptors from 0D to 3D [31,32].

Table 1. Methods for the generation of molecular descriptors for QSAR modelling.

Dimension Definition
Only contains the molecular formula. Thus, the only information is
0D
the atom types and numbers of each.
Molecular properties that pertain to the entire chemical structure,
1D such as logP and pKa. It also includes substructural details of
molecular fragments.
Topologies are mathematically encoded to represent the connectivity
2D
of atoms using a 2D graph.
Details of the spatial arrangement of atoms and non-covalent
3D
interaction sites guided by 3D topologies.

Once a dataset has been curated and descriptors generated, an algorithm must be
selected to complete the regression task. Broadly speaking, the algorithms used in QSAR
can be categorised into linear and non-linear methods [33]. Popular linear regression tools
include partial least squares (PLSs) [34]. PLSs transform large, high dimension data such
as molecular descriptors into linear solutions to make predictions. Non-linear methods
include k-nearest neighbours (k-NNs) [35], support vector machines (SVMs) [36] and
random forest (RF) [37]. Both k-NN and SVMs use the distances between parameters in the
hyperplanes to determine solutions to the QSAR regression problems. In contrast, RF uses
a collection of decision trees to build one robust predictive model.
To ensure the QSAR model being used has robust predictive abilities, validation studies
must be undertaken. This is another recommendation from the OECD for the development
of a credible QSAR model [26]. It has been suggested that an external validation set
approximately 15 to 20% the size of the entire dataset should be used to ascertain a model’s
performance [38]. R2 (1–(RSS/TSS), where RSS is the residual sum of squares and TSS is
the total sum of squares) and q2 are the most popular parameters to check the goodness of
fit for QSAR models [39,40]. The q2 value is obtained by calculating R2 using leave-one-out
cross validation [27]. Ideally, R2 should be as close to 1 as possible for goodness of fit [27];
however, other studies have suggested that q2 > 0.5 and R2 > 0.6 is sufficient [38]. It is
Molecules 2023, 28, 1324 6 of 22

important to note that a high q2 does not always guarantee good external validity [27].
Alternatively, root mean squared error (RMSE) or MAE can be used, which in some cases
may be better indicators of predictive ability on experimental data [41].

3. Structure-Based Methods
Structure-based drug design (SBDD) is a branch of CADD and is utilised when the
3D-structure of a biological target is available. The insight into the composition of a ligand
binding site allows for the screening of compound libraries and the specific design of
molecules to fit optimally to the ligand binding site. Recent advances in crystallisation
techniques and technical advances in cryogenic electron microscopy (cryo-EM) means
that the pool of available 3D-structures of important drug targets is rapidly expanding.
Furthermore, the completion of the Human Genome Project and use of artificial intelligence-
based structure prediction tools such as AlphaFold2 [42] and RoseTTAFold [43] allow
for prediction of structures of proteins from sequence alone. In addition, the increase
in the speed of high-performance computing and the use of graphical procession units
(GPUs) means that the screening of ultra-large libraries of commercially available and
make-on-demand molecules is now possible [44]. These developments in SBDD permit
an expansion in the areas of chemical space that can be explored in the pursuit of new
drugs and may benefit the development of drugs for the treatment of psychological and
neurologic conditions.

3.1. Homology Modelling and Molecular Docking

The first step in structure-based drug design (SBDD) is selecting a target structure
and identifying the binding site of interest. The target structure can be determined experi-
mentally by nuclear magnetic resonance (NMR), X-ray crystallography or cryo-EM. The
Protein Data Bank (PDB) [45] is a repository for such experimentally determined structures.
Alternatively, where the structures of the target of interest are not available, they can be
modelled computationally using the structure of one or more evolutionary related pro-
teins [46,47]. A minimum similarity of 30% between the template and target sequences is
usually recommended [48]. The PDB contains nearly 200,000 experimentally determined
structures at the time of writing, all of which can be used as templates [45]. The sequences
of both the desired target and the template can be accessed via the Uniprot [49] and Swis-
sprot [50] databases and aligned using programs such as T-Coffee [51] or BLAST [52].
Tools such as MODELLER [53] and SWISS-MODEL [54] or a commercial software such as
Prime [55] and MOE [56] are used to perform the homology modelling process. In essence,
the template structure is used as a basis to guide predictions of protein folding by using
regions of similarity. Further details about the processes involved in homology modelling
have been reviewed recently [57,58]. Typically, many models based on different starting
seeds are generated and the most accurate model suitable for further studies is ascertained
using scoring functions such as the discrete optimized protein energy (DOPE) [59] or quali-
tative model energy analysis (QMEAN) [60]. Both scoring functions are statistical analyses
that assess the local and global energy potentials of a homology model to discern if the
structure is realistic. Homology models of the human dopamine transporter (hDAT) [61]
were an essential first step in ascertaining new leads towards treatments for Parkinson’s
disease. Similarly, homology modelling assisted in uncovering dual-target ligands for the
α7 nicotinic acetylcholine receptor (nAChR) and acetylcholinesterase (AChE) in the search
of improved therapies for Alzheimer’s disease [62].
Molecular docking is a SBDD technique that predicts the preferred orientation and
conformation of a ligand in a binding site. It can be used to screen databases of ligands,
provide scores for the predicted binding energy and rank the ligands to assist in finding
potential hit molecules [63,64]. In molecular docking studies, possible poses of the ligand
are explored by allowing the ligand to be flexible when binding into the protein, which
ensures the lowest possible energy state of each ligand–protein complex to be determined
when scoring and ranking the ligands [65]. Conformers are generated either a-priori, by
Molecules 2023, 28, 1324 7 of 22

systematically exploring all possible degrees of freedom with respect to ligand binding,
or by randomly modifying parameters, such as torsional angles, depending on the search
method chosen [63]. Once ligand conformations are generated, they are then docked
into a rigid protein binding site for evaluation and the scoring of binding interactions.
Alternatively, induced fit docking (IFD) protocols can be used, whereby the protein exhibits
a degree of flexibility. Whilst this can lead to more accurate predictions in terms of binding
potential, it is significantly more computationally expensive compared to standard docking
protocols [66]. Thus, it is typically used in post processing to ascertain a hit compound’s
binding mode. To assess the binding interactions of ligands with dual activity against
tumour necrosis factor receptor 1 (TNFR1) and inhibitor of nuclear factor kappa-β kinase
subunit β (IKKB) complex, an induced fit docking protocol was employed [67]. This
identified avanafil as the most promising lead compound for biological evaluation. Avanafil
demonstrated neuroprotective effects in a mouse model of neuroinflammation as well
as a reduction in the formation of amyloid-β plaques and inflammatory cytokines in
mouse brains.
The scoring of ligands determines which ligand pose is the most energetically favourable
and ranks the library of screened ligands to indicate which compounds are most likely to
be active and suitable for further research [68]. This is particularly important when large
chemical libraries are being screened. One study, which was investigating new inhibitors
against the voltage gated sodium channel Nav 1.7 towards new therapies for neuropathic
pain, used docking studies in BioSolveIT to refine a chemical library of 1.5 million ligands
down to nine leads with demonstrated efficacy in a mouse model of neuropathic pain [69].
An important secondary finding from the docking studies was that a new binding mecha-
nism to previously described sulfonamide Nav 1.7 inhibitors at the active site was noted
with the new ligand. Generally, scoring functions can be divided into three different cate-
gories: (1) empirical [70–72], (2) knowledge-based [73,74] and (3) force field-based scoring
functions [75]. The scores are generated either through accounting for the individual con-
tributions of energy terms, including hydrogen bonds, electrostatics and hydrophobicity,
or they are derived from statistical analysis of experimentally determined ligand–protein
complexes. Regardless of the scoring function, the primary goal is to filter through large
chemical libraries to find the ligands with the best properties for further research.

3.2. Molecular Dynamics Studies

Docking examines a single frame in time whereas, in reality, proteins and ligands are
flexible over time and present in a dynamic, complex cellular environment in contact with
water, membranes, and ions. Molecular dynamics (MD) can account for this as a technique
either before or after docking to study protein conformation, flexibility, stability and ligand
sampling [76]. Due to the ability to account for and simulate motion, MD intrinsically
simulates the dynamic protein system and therefore accounts for protein structural changes,
which is not possible in conventional docking studies, which are typically limited to rigid
protein structures or structures with localised movements [77,78]. The validation of docking
results was carried out during investigations for antagonists of LRRK2 for new treatments
of Parkinson’s disease [79]. The MD simulations provided insight on the likely binding
mode of a potent new ligand, LY2019-005.
MD aims to simulate the time-based change in atom positions, using Newton’s equa-
tion of motion (F = ma) to offer a relationship between the force and acceleration of the
atom dependent on atomic mass over time. These algorithms are implemented as force
fields, with common examples including CHARMM [80], GROMOS [81], AMBER [82] and
OPLS [83]. These force fields are implemented in molecular dynamics engines such as
GROMACS [84], AMBER [85], OpenMM [86] and Desmond [87], which can support one or
more of the aforementioned force fields. Recent developments of large-scale simulations
using coarse-grained (CG) force fields such as MARTINI [88] are gaining traction for large
membrane and organelle level studies. Furthermore, recent advances in GPU hardware
capabilities and algorithms such as CUDA [89] has resulted in computing performance
Molecules 2023, 28, 1324 8 of 22

increases of up to 100 times faster as compared to when calculations are performed on a

traditional central processing unit (CPU)-only workflow, enabling microsecond time-scale
simulations. However, often the binding of ligands involves energy barriers which need to
be overcome for ligand binding or conformational changes. Several methods have been
devised to overcome this issue, including metadynamics, replica exchange and alchemical
techniques including free energy perturbation (FEP). Details of these methods are beyond
the scope of this review, but they are nonetheless critical for applications in drug design
and discovery.
MD simulations require a starting point, typically experimental models based on X-ray
crystallography, cryo-EM or NMR spectroscopy. Computational models such as homology
modelling, described in Section 3.1, and deep learning methods such as AlphaFold2 [42]
may also be utilised as a starting model where the experimental structure is not available.
Where computational methods have been implemented, MD simulations may help to
ensure the validity of these models. Park et al. [90] employed MD simulations to investi-
gate the binding properties of an optimised lead compound with predicted antagonistic
effects against the G2019S mutant leucine-rich repeat kinase 2 (LRRK2) receptor. The
mutated LRRK2 receptor is known to contribute to Parkinson’s disease pathophysiology
through increased activity. The MD simulations indicated that the protein–antagonist
complex remained stable, and in vitro testing confirmed that this optimised lead had
nanomolar potency.
In MD simulation preparation, membrane proteins are placed within a membrane
model consisting of one or more lipid species such as phosphatidylcholine (POPC), phos-
phatidylethanolamine (POPE) and cholesterol. In an explicit water simulation system, indi-
vidual water molecules are placed to fill the simulation box and random water molecules
are removed and replaced with ions such as Na+ or Cl− . Simulation systems require several
steps of processing before they can be simulated for production runs, with this including
minimisation and equilibration, where the MD system is simulated for a short period to
reduce the overall system energy and to stabilise the system for production simulations.
Post-simulation analysis metrics include measures such as the root mean square deviation
(RMSD), a measure to determine the overall structural deviation from the initial starting
pose, and root mean square fluctuations (RMSF), a measure to determine the movement
of individual residues in the protein to determine the flexibility of the protein and the
ligand of interest. Further analysis including ligand-to-residue contact analysis, ligand
environment analysis and binding pose clustering is also performed to determine potential
ligand binding poses and intermediate binding sites.

4. ADMET Property Prediction

A key challenge in all drug discovery endeavours is ensuring that new leads have
acceptable pharmacokinetic parameters of absorption, distribution, metabolism and excre-
tion whilst also minimising toxic effects (ADMET). There exists further difficulty for CNS
pharmaceuticals that must cross the blood–brain barrier (BBB), as this is a complex process
that involves both the passive and active diffusion mechanisms [91]. The incorrect predic-
tion of ADMET properties may be an extremely costly error if not determined early in the
drug discovery pipeline, as unacceptable pharmacokinetic parameters have been reported
to account for 40% of failures in phase II clinical trials [92]. Thus, several methodologies to
predict favourable ADMET properties as part of CADD workflows exist. Much like QSAR
studies, the prediction of pharmacokinetic parameters utilises machine learning techniques
including SVM, k-NN and RF to ascertain the relationship between molecular descriptors
and ADMET properties. The relationship between molecular descriptors and in vitro data
points is used to predict the pharmacokinetic parameters that new leads would exhibit.
The applications of ADMET prediction in computational drug discovery is extensively
reviewed in [93,94]. Given that the focus of this review lies in the domain of neurological
and psychiatric drug discovery, the remainder of this chapter will discuss advances in the
computational prediction of BBB permeability.
Molecules 2023, 28, 1324 9 of 22

The prediction of BBB permeability by passive diffusion can largely be predicted from
physicochemical properties including lipophilicity, polarity and ionisation at physiological
pH. However, several mechanisms of active transport into and efflux out of the CNS
must also be accounted for in these predictions [95]. Additional data such as in vivo
BBB permeability has proven useful for these considerations [96]. The use of datasets
that encompass drug phenotypes including the clinical indication and CNS-related side
effects, in addition to physicochemical properties, to determine BBB permeability is a
similar approach [97]. This work was expanded upon by Miao et al., with a deep learning
algorithm which demonstrated a marked improvement of 97% accuracy, compared to the
86% accuracy of the initial SVM model [98]. The increased accuracy can be attributed to
the enhanced capacity of deep learning models to understand the abstract relationships
between parameters. Work has also been carried out in the development of QSAR models
that identify substrates of efflux proteins implicated in poor CNS uptake, such as the multi-
drug resistance protein 1 (MRP-1) [99] and breast cancer resistance protein (BCRP) [100].
More recent applications in this area make use of image recognition [101] and natural
language processing [102] advancements in deep learning to aid in BBB permeability
predictions, which have resulted in accuracies as high as 99%. The underlying principles of
these applications are discussed below.

5. The Rise of Deep Learning in Computer-Aided Drug Discovery

While deep learning (DL) is not a new methodology, with it having been applied to
language and image processing for several decades now, its applications to drug discovery
efforts have only emerged in the last few years [103]. This has been accelerated by the
use of GPUs to handle the computationally expensive calculations associated with deep
learning [104]. DL extends beyond traditional machine learning methods by using several
processing layers, known as neurons, to make predictions based on large collections of
multi-dimensional data [105]. Biological data for different CNS targets to train DL models
may be obtained from open-source databases such as ChEMBL [106], PubChem [107] and
MolData [108]. In-house datasets from experimental studies are also commonplace in the
pharmaceutical industry [109]. Whilst several types of deep learning architectures exist,
the most prominent applications in drug design and discovery are convolutional neural
networks (CNNs), recurrent neural networks (RNNs), long-short term memory (LSTM)
and multi-task learning (MTL) (Figure 3). Thus, these will be the primary focus of the
review. Detailed reviews on all types of deep learning and their underlying principles are
provided by Le Cun et al. [105] and Schmidhuber et al. [110].
The most widely used application of CNNs are in image recognition, as the architecture
mimics that of the visual cortex [111]. CNNs collect information about the presence or
absence of features in different locations of images using a feature map, which is derived
from convolutional layers. It is these feature maps that give CNNs the advantage in terms
of spatial awareness over other DL architectures. Next, a rectified linear activation function
(ReLU) is applied, which serves to account for interactions between variables and non-
linearities in the model. A pooling layer will summarise the information collected in the
aforementioned layers to prevent model overfitting. These steps will be repeated several
times, with at least 1 fully connected layer to link the hidden layers to the output [112]. In
drug design, CNNs are used to extract features from 2D or 3D molecular graphs, with them
demonstrating superior performance over other machine learning and deep neural network
(DNNs) methods which use molecular fingerprints [113]. The features extracted from the
molecular graphs can be used to predict pharmacokinetic properties [113] or to ascertain
the correct binding poses and binding affinities of ligand–protein complexes [114,115].
Graphical CNNs were employed, with images of molecule fingerprints as the input, in the
search for new AChE inhibitors for the treatment of Alzheimer’s disease. The deep learning
model outperformed three other machine learning methods: linear regression, random
forest and XGBoost. Two hit molecules were identified from a library of 2 million small
Molecules 2023, 28, 1324 10 of 22

Molecules 2023, 28, x FOR PEER REVIEW

molecules. 10 of 24
The resulting leads are able to traverse the BBB and outperformed galantamine
in vitro, thus demonstrating promise as potential drugs [116].

Figure 3.
Figure 3. Workflows
Workflows of
of LSTM
LSTM (a),
(a), RNNs
RNNs (b),
(b), multi-task
multi-task learning
learning (c)
(c) and
andCNNs
CNNs(d).
(d).

The most widely

Recurrent neural used
networks application
can processof CNNs are in image
sequential data by recognition, as the architec-
linking information from
turetime
one mimicspointthat of the
back to an visual
earliercortex
time[111].
point.CNNs collect information
This property means theyabout thea presence
require memory
or absence
buffer of features
to store in different
information locations states
from previous of images[117].using
Thesea feature map, mean
properties whichRNNs
is de-
rived
are from convolutional
excellent tools in language layers.processing.
It is these feature
Thesemaps neuralthat give CNNs
networks are the advantage
cyclic in nature,in
terms of spatial
meaning hiddenawareness
neurons receive over other DL architectures.
feedback from both the Next, a rectified
input and thelinear
memory activation
buffer.
Information
function (ReLU) about correct and
is applied, whichincorrect
serves to outputs
accountare forfed back to the
interactions respective
between neurons
variables and
which contributed
non-linearities in thetomodel.
that prediction
A poolingusing layer backpropagation.
will summarise the However,information RNN models
collected in
are limited in that itlayers
the aforementioned is difficult to learn
to prevent andoverfitting.
model store over Thesevery long
stepstimewill points givensev-
be repeated the
sequential
eral times, nature
with atof the1models.
least Long-short-term-memory
fully connected layer to link the hidden is an extension
layers to theof RNNs
outputwhich
[112].
can overcome
In drug design, this limitation
CNNs are used [118]. LSTM works
to extract featuresusing
from a similar
2D or 3D principle
molecular to RNNs
graphs,butwith
has
athem
memory cell which accumulates input information by connecting
demonstrating superior performance over other machine learning and deep neural to itself at future time
points.
networkIn(DNNs)
addition, the LSTM
methods whicharchitecture
use molecular contains leaky gated
fingerprints [113].neurons, whichextracted
The features learn to
decide whether to clear the stored information at later time points.
from the molecular graphs can be used to predict pharmacokinetic properties [113] or to The most significant
contribution
ascertain the of RNNsbinding
correct and LSTM poses to and
medicinal
binding chemistry
affinitiesare of their abilities tocomplexes
ligand–protein learn and
correctly
[114,115].predict
Graphical SMILES
CNNsstrings or other linear
were employed, withchemical
images of notation
molecule systems. This can
fingerprints as thebe
used in retrosynthesis to design new drugs with feasible synthesis
input, in the search for new AChE inhibitors for the treatment of Alzheimer’s disease. The routes [119,120]. In
addition, RNNs have been trained to interpret protein sequence
deep learning model outperformed three other machine learning methods: linear regres- data and used concurrently
with
sion, graphical
random forestCNNs andto XGBoost.
predict ligand–protein
Two hit molecules affinity in the
were DeepAffinity
identified from amodel
library[121].
of 2
Multi-task
million small molecules. The resulting leads are able to traverse the BBB and to
learning uses a collection of learning algorithms and analysis methods make
outper-
predictions for multiple
formed galantamine tasks.
in vitro, These
thus tasks are learnt
demonstrating promisein parallel whilst
as potential making
drugs [116].what is
learned by eachneural
Recurrent task available
networkstocan theprocess
overall sequential
model through data backpropagation
by linking information [122]. from
This
one time point back to an earlier time point. This property means they require a memory
buffer to store information from previous states [117]. These properties mean RNNs are
excellent tools in language processing. These neural networks are cyclic in nature,
Molecules 2023, 28, 1324 11 of 22

is particularly useful in models where the biological data being used in the models have
different values from several experiments for the same target. In addition, multi-task
learning has demonstrated capabilities in the prediction of drug activity against proteins
from the same class (GPCR and ion channels) [123–125].
In computational drug discovery, as demonstrated above, deep learning applications
are producing advances over more traditional techniques. Novel algorithms which com-
bine both CNNs and LSTM for de novo design, such as the RELATION model [126], show
promise in finding novel leads from an expansive chemical space. A similar application ex-
ists specifically for CNS drug design [127] which also accounts for the added complexity of
BBB permeability. The benefit of applying DL techniques to computational drug discovery
is the ability to better process the complexity of molecular descriptors and their interactions
with biological systems. Neural networks and other machine learning-based methods have
also been utilised to understand the relationships between genes, an individual’s environ-
ment and disease biomarkers for mental illnesses such as schizophrenia and depression
as well as Alzheimer’s disease [128–131]. These modelling techniques are aimed towards
precision medicine and enhanced disease understanding for improved therapeutics. In ad-
dition, DL architectures require large quantities of data for high performance and therefore
are well positioned to handle information from large databases. Although DL applications
in neurological and psychiatric drug discovery are still emerging, there is clear potential
for this application to enrich drug discovery efforts.

6. Applications to Neurological and Psychiatric Conditions

Computational methodologies have become an essential part of drug design. Recent
examples have demonstrated the potential of these techniques in terms of accelerating the
drug discovery pipeline and reducing the time and money spent on laboratory testing and
clinical trials. Table 2 provides an insight into how CADD techniques are being applied to
develop drugs targeting neurological and psychiatric conditions, and other examples are
given in the text below. From our search of the literature, Alzheimer’s disease was the most
prominent disease target, with several CADD studies using a diverse range of drug targets
being reported below. Given that over 55 million people globally suffer from Alzheimer’s
disease, with the World Health Organisation projecting this to more than double over the
next 30 years [132], the need for new treatments is imperative. The applications of CADD
studies for Alzheimer’s disease exemplifies this demand. Schizophrenia was the most
common psychiatric disease target in the review of the literature, with targets such as α7
nAChR selected to improve available treatment options for sufferers. It is important to note
that even though a number of the distinct therapeutic areas discussed below share similar
target proteins, these possess vastly different clinical phenotypes. The relationship between
shared disease genotypes with distinct clinical manifestations can be accounted for, in part,
by epigenetic factors [133] and is beyond the scope of this paper. This section will discuss
the therapeutic areas where CADD studies for CNS diseases have been reported, with
Table 2 providing further exemplars (presented below).

6.1. Alzheimer’s Disease

A multi-target activity 3D-QSAR model against acetylcholinesterase (AChE), serotonin
transporter (SERT), beta-secretase 1 (BACE1) and glycogen synthase kinase-3 (GSK3β) was
built to towards new therapeutics for Alzheimer’s disease [134]. The QSAR model was built
using IC50 data from ChEMBL [106], with both a multilinear regression and an artificial
neural network (ANN) model being used. The 2D structures of ligands were converted to
3D using OpenBabel, and molecular descriptors were generated using FQSARModel. Dur-
ing model validation, it was apparent that ANN models performed better, and these were
therefore selected to be used in virtual screening. Over 20,000 compounds from the ZINC
(biogenic) database were docked against all four proteins. After docking in both Glide and
Autodock, 57 compounds with drug properties and favourable ligand efficacy were then
screened against the QSAR models. The models indicated that five ligands held promise
Molecules 2023, 28, 1324 12 of 22

in terms of targeting at least three of the four proteins. One compound, ZINC4027357 (1,
Molecules 2023, 28, x FOR PEER REVIEW 13 of 24
Figure 4A), demonstrated the inhibition of both AChE and BACE1. None of the selected
hits had inhibitory properties against SERT or GSK3β within the selected potencies.

Figure
Figure 4. Chemical structures
4. Chemical structuresofoflead
leadcompounds
compounds with
with thethe highest
highest reported
reported in vitro
in vitro activity
activity from
from
exemplar papersdiscussed
exemplar papers discussedbelow.
below. TheThe molecules
molecules are proposed
are proposed to use
to be of be of use in therapeutics
in therapeutics for Alz- for
Alzheimer’s disease
heimer’s disease (A), (A), Parkinson’s
Parkinson’s disease
disease (B), neuropathic
(B), neuropathic pain
pain (C) (C)schizophrenia
and and schizophrenia
(D). (D).

A structure-based
6.2. Parkinson’s Disease application of the multi-target approach was also used to identify
lead compounds with dualstudy
A drug repurposing activity against
aimed AChE
to find and
new α7 nAChR.disease
Parkinson’s The ZINC15 database
treatments using[135],
consisting
associationsofbetween
over 7.5approved
million smalldrugsmolecules,
and proteinswasassociated
filtered towith
remove molecules
Parkinson’s with un-
disease.
favourable properties such as Lipinski’s violations, resulting in
The CNN model demonstrated superiority against other benchmark approaches (e.g., 3.8 million ligands being
selected for virtual screening in Glide targeting the human AChE
DTINet and deepDTnet), with an accuracy of 91.57%. In addition, the CNN model out-and an α7 nAChR homol-
ogy model, traditional
performed which wasmachine
built using MODELLER.
learning There
algorithms. The were
top 57
10 compounds
ranked compoundsshared between
from
both proteins, of which 16 were selected for in vitro assessment.
the unknown samples underwent molecular docking against the 5-hydroxytryptamine Compound Ymir-2 (2,
Figure
receptor 4A)2Apossessed
(5HTR2A)the most favourable
to ascertain favourable chemical profile
interactions and dual
between thesetarget activity
ligands and [136].
the
target proteins. Pimvanserin was used as a positive control, for which three of thewith
A deep learning approach based on a series of regression models that were built ten the
aim of predicting
ligands bindingbinding
had comparable free energy towards
energy, AChEthe
of which was produced. Ofinhibitor
topoisomerase the regression models,
topotecan
−1
a(4,
graphical CNN
Figure 4B) wasmodel had promising
the most [137].with an RMSE of (1.580 ± 0.137 kcal mol ). This
the best results,
model was selected
Another to screenapproach,
deep learning a datasetusing
of 2 million compounds,
deep neural networkofarchitecture
which 6 were identified
was built
as
to suitable
identify for docking with
piperine-like AutoDock
compounds andVina, MDagainst
drugs simulations using GROMACS
these targets. The modeland in vitro
demon-
assessment. Benzyl trifluoromethyl
strated an accuracy 87.5%. A total of ketone
57,423(3,compounds
Figure 4A)from outperformed
the ZINC galantamine,
and PubChem with
an IC50 value
databases of 0.33 µM
underwent against AChE.
a similarity search Permeability
based on piperine assessments
to find suggested these ligands
similar structures. In
may traverse
all, 101 the blood–brain
compounds were selectedbarrier [116]. investigation through docking in AutoDock
for further
4.0, of which 5 were suitable for MD simulations on the AMBER platofrm. The docking
and MD studies revealed that an additional ring in top performing compounds (5, Figure
Molecules 2023, 28, 1324 13 of 22

6.2. Parkinson’s Disease

A drug repurposing study aimed to find new Parkinson’s disease treatments using as-
sociations between approved drugs and proteins associated with Parkinson’s disease. The
CNN model demonstrated superiority against other benchmark approaches (e.g., DTINet
and deepDTnet), with an accuracy of 91.57%. In addition, the CNN model outperformed
traditional machine learning algorithms. The top 10 ranked compounds from the un-
known samples underwent molecular docking against the 5-hydroxytryptamine receptor
2A (5HTR2A) to ascertain favourable interactions between these ligands and the target
proteins. Pimvanserin was used as a positive control, for which three of the ten ligands had
comparable binding energy, of which the topoisomerase inhibitor topotecan (4, Figure 4B)
was the most promising [137].
Another deep learning approach, using deep neural network architecture was built
to identify piperine-like compounds and drugs against these targets. The model demon-
strated an accuracy 87.5%. A total of 57,423 compounds from the ZINC and PubChem
databases underwent a similarity search based on piperine to find similar structures. In all,
101 compounds were selected for further investigation through docking in AutoDock 4.0,
of which 5 were suitable for MD simulations on the AMBER platofrm. The docking and
MD studies revealed that an additional ring in top performing compounds (5, Figure 4B) is
likely to help to form hydrogen bonds in the active site, which leads to a greater potency
against Monoamine-oxidase A and B (MAO-A and MAO-B) [138].
A docking study of over 1.6 million small molecules was conducted against a homol-
ogy model of LRRK2, which was conducted using both the Glide and Prime modules of
Schrodinger’s Maestro software. In total, 28 high performing molecules were purchased
for biological evaluation. Two small molecules with novel features, namely LY2019-005 and
LY2019-006, were identified, with these also being able to pass the blood–brain barrier. MD
simulations were conducted to investigate the binding mode of these ligands. Both ligands
possessed nanomolar IC50 values for both the wild-type and G2019S mutant enzyme, with
LY2019-005 (6, Figure 4B) being the most potent. Given the neurotoxic potential of the
G2019S mutant of LRRK2, nanomolar IC50 is particularly important [79].

6.3. Neuropathic Pain

The biological data of 180 sigma-1 receptor (S1R) antagonists was curated from the
literature and split into training and test sets in a 4:1 ratio, with the randomization of this
process repeated 50 times to avoid bias. MOE software was used to generate 206 molecular
descriptors, which then underwent dimensionality reduction using principal component
analysis (PCA). An atom-based 3D-QSAR model was developed using the partial least
squares technique, with the final model having an RMSE of 0.29 and R2 of 0.92, both
of which are suggestive of a model with good predictive abilities. An energy-based
pharmacophore was also developed to supplement the 3D-QSAR model by providing
conformational information about the binding site. This was used in the virtual screening
studies to ensure the ligands were in the correct orientation for the binding site, developed
using Schrödinger’s Glide. After pre-filtering steps, 1935 FDA-approved drugs from the
DrugBank database were initially screened against the pharmacophore model, with the best
fitting conformer of each ligand then being screened in the 3D-QSAR model. Twelve of the
best performing ligands with no prior biological affinity data against S1R underwent further
assessment in vitro using a radio–ligand binding assay. Two drugs, phenyltoloxamine and
diphenhydramine, exhibited 66 and 70% inhibition at a concentration of 1 µM, respectively.
This also conferred with reports in the literature that diphenhydramine (7, Figure 4C) can
be used as an adjective analgesic [139].
A highlight on target sequence (HoTS) deep learning model was used to scan the
purinergic P2X3 protein sequence to identify new binding sites in search of treatments for
neuropathic pain [140]. Once the DL model had identified potential binding sites, MD
simulations in the CDOCKER software package were used to measure the volume of the
binding sites to ensure their feasibility. Four new binding regions were identified, for which
Molecules 2023, 28, 1324 14 of 22

a pharmacophore model was developed using the binding mode of a known antagonist and
its derivatives in the BIOVIA software. Over 97,000 compounds were screened against the
pharmacophore model. A total of 2346 ligands were then docked to assist in prioritisation
for in vitro assessment, of which 500 were selected for experimental validation. A total
of 16 compounds with novel structures and low micromolar IC50 values were identified.
Compound 8 (Figure 4C) was the most potent lead compound.

6.4. Schizophrenia
A 2D-QSAR model was developed using 159 inhibitors of the sigma 2 receptor (S2R)
Molecules 2023, 28, x FOR PEER REVIEW 15 of 24
reported in the literature. MOE software was used to generate molecular descriptors
of each ligand for QSAR model generation. Four algorithms were generated, namely
stepwise regression, Lasso, genetic algorithm (GA) and an algorithm, GreedGene, which
wasshape-based
developed by screening.
the authors.Ligands that shared
GreedGene hadsimilarities to siramesine
the best performance, R2 of 0.56,
andanligands
with with a
piperazine-containing scaffold or tetrahydroisoquinolinyl structures
and was selected for screening. A pharmacophore model was also generated using Glide were kept for in vitro
for testing. These screening.
use in virtual scaffolds are known
Over 2000for high
small S2R binding
molecules fromaffinity. A total of
the DrugBank 30 compounds
database were
possessed
screened this the
against scaffold
QSAR and were which
model, identified
hadas promising
a pKi cut-off leads.
of 5.5.Six molecules
A total of 823underwent
ligands
werebiological testing, against
then screened which revealed three FDA approved
the pharmacophore drugs, the
model, before nefazodone, cinacalcet and
best 120 underwent
pimozide,screening.
shape-based had nanomolar Ligands binding affinitysimilarities
that shared values, with nefazodoneand
to siramesine (9, ligands
Figure 4D) withbeing
a
piperazine-containing
the most potent of the scaffold
three or tetrahydroisoquinolinyl structures were kept for in vitro
[141].
testing. AThese scaffolds are was
pharmacophore known for highusing
generated S2R binding affinity.agonists
11 α7 nAChR A total of 30 compounds
from the literature.
possessed this scaffold and
The pharmacophore were identified
consisted as promising
of a hydrogen bondingleads.
region, Sixa molecules
hydrophobic underwent
centre and
biological testing, ionised
one positively which revealed
group. To three FDAthe
reduce approved
numberdrugs,
of falsenefazodone,
positives, acinacalcet
recursiveandparti-
pimozide,
tioning had nanomolar
model was alsobinding
used. Aaffinity
virtualvalues, withofnefazodone
screening the ChemDiv (9, Figure
database 4D)against
being thethese
mosttwopotent
modelsof the
wasthree [141]. After filtering to ensure no Lipinski parameter violations, 13
performed.
A pharmacophore
ligands were selectedwas for generated using 11 α7
in vitro assessment, 10 nAChR
of whichagonists from the literature.
had demonstrated inhibitory
Theeffects.
pharmacophore consisted of a hydrogen bonding region, a
T761-0184 was selected for further investigation due to its high potency.hydrophobic centreThis
andlig-
oneandpositively ionised group. To reduce the number of false positives,
underwent induced fit docking to a homology model of α7 nAChR to ascertain the a recursive parti-
tioning model
binding mode wasforalso used. Aoptimisations.
structural virtual screening of the
Of the ChemDivstructures,
51 optimised database against
B10 (10,these
Figure
two4D)models was subtype
exhibited performed. After filtering
selectivity to ensure
for α7 nAChR overno Lipinski
other nAChR parameter
subtypes.violations,
B10 was also
13 ligands were
one of the mostselected
promisingfor inligands,
vitro assessment,
with an IC10 of which had demonstrated inhibitory
50 value of 5.4 µM [14].
effects. T761-0184 was selected for further investigation due to its high potency. This ligand
underwent induced
Table 2. Further fit docking
examples to a homologydrug
of computer-aided model of α7 nAChR
discovery for new to ascertain the
neurological andbinding
psychiatric
mode for structural
treatments from theoptimisations.
literature. Of the 51 optimised structures, B10 (10, Figure 4D)
exhibited subtype selectivity for α7 nAChR over other nAChR subtypes. B10 was also one
Drug Target and of the most promising ligands, with an IC50 value of 5.4 µM [14].Structure
Study Significance Chemical Reference
Methodology
ATable
ligand-based
2. Furtherscreening
examples of method known as
computer-aided atom
drug discovery for new neurological and psychiatric
category
treatments extended
from theligand overlap score (xLOS)
literature.
was used to ascertain leads from a library of over
Drug Target and
900,000 small molecules. Study This method was chosen
Significance Chemical Structure Reference
Methodology
due to a lack of information about the structure
A ligand-based screening method known as atom category
andextended
binding pocket
ligand of TRPM4.
overlap score (xLOS) Three reference
was used to ascertain
Drug target: leads from a library of over 900,000 small molecules. This
compounds and the database compounds were
Transient receptor po- method was chosen due to a lack of information about the
converted into 3D structures using CORINA soft-
Drug sub family M4 re- structure and binding pocket of TRPM4. Three reference
target:
tential
Transient receptor potential sub ware. xLOS was
compounds then
and the used compounds
database to compare database
were converted into
ceptor
family M4 (TRPM4)
receptor (TRPM4) 3D structures using CORINA software. xLOS was then used to
Disease target:
ligands to the reference compounds, 9-phenan-
compare database ligands to the reference compounds,
Disease target:
Multiple sclerosis throl, glibenclamide
9-phenanthrol, and flufenamic
glibenclamide acid,
and flufenamic and
acid, and rank [142]
[142]
Multiple sclerosis them. A total of 214 of
of the
Software packages: rank them. A total 214topofmolecules were purchased for
the top molecules
Software packages:
CORINA biological evaluation. An additional round of xLOS screening
Methods: were purchased
on the for biological
Princeton database evaluation.
was performed Antop
using the ad-three
CORINA
xLOS hits from the first
ditional round ofround
xLOSof screening
biological evaluation. The biological
on the Prince-
Methods: evaluation was conducted on 247 ligands from the second
tonround
database was performed
of screening. using
The top scoring leadthe
had top three
potency at
xLOS
hitsapproximately
from the first 1 µMround of biological
IC50 , which is a marked evaluation.
improvement
Theover the initial reference compounds.
biological evaluation was conducted on 247
ligands from the second round of screening. The
top scoring lead had potency at approximately 1
μM IC50, which is a marked improvement over
the initial reference compounds.
Drug target: In silico ADMET assessments and docking stud-
 throl, glibenclamide and flufenamic acid, and [142]
Multiple sclerosis Study Significance Chemical Structure Reference
Methodology rank them. A total of 214 of the top molecules
Software packages:
Molinspiration were purchased
epilepsy withoutfor anybiological
adverse effects evaluation. on motor An ad- ac-
CORINA
Cheminformatics ditional round of xLOS screening on the Prince-
tivity.
Methods:
AutoDock 4 ton database was performed using the top three
xLOS
Molecules 2023, 28,
Molecules 1324
2023, 28, x FOR PEER REVIEW 15 of16
22 of 24
Methods: hits from the first round of biological evaluation.
Docking The biological evaluation was conducted on 247
Aligands
VS study fromwas theperformed
second round againstof screening.
the CB1 recep- The
Table
topusing 2.
scoring Cont.lead had potencysubset at approximately 1
Drug Target and tor a natural products of the ZINC12
Study Significance Chemical Structure Reference
Methodology
Drug Target and
μM
database.IC 50, which
Nearly is a marked
300,000 small improvement
molecules over
were
Study Significance Chemical Structure Reference
Methodology
Molinspiration the
filteredinitial
epilepsy and reference
docked.any
without compounds.
Theadversefilteringeffectsand dockingon motor us-ac-
Drug
Drug target:
Cheminformatics
target: Intivity.
ing silico
standard ADMET and assessments
extra precision andsettings
docking instud-
Glide
N-methyl-D-aspartate
N-methyl-D-aspartate
Cannabinoid receptor1 indicated
receptor ies revealed 32 three compounds
top-performing with
ligands, acceptable
of which 18
AutoDock 4
(NDMA) GluN1-GluN2A
In silico ADMET assessments and docking studies revealed
receptor
(CB1) (NDMA) three compounds
pharmacological
were selected for properties, with acceptable
further in vitro pharmacological
including
testingthe properties,
ability
through
Methods:
subunits
including the ability to traverse the BBB. These compounds
Disease
Diseasetarget:
GluN1-GluN2A target: subu- clustering
to traverse tothe BBB.
ensure These
structural compounds
diversity demon-
amongst
Docking demonstrated similar binding interactions to endogenous
[143]
Epilepsy [143]
nits packages:
Substance
Software abuse disor- hits. strated
Aligands
VS similar
Of study
thebut18 with
wasbinding
ligands,
improved
performed 7interactions
demonstrated
binding capacity.
against totheendoge-
more
The CB1leadthan
recep-
compounds resulted in a reduced number of seizures observed [144]
Disease target:
ders
Molinspiration Cheminformatics nous
50%tor ligands
displacement
using a but
natural within improved
competitive
products binding
subset
in a mouse model of epilepsy without any adverse effects on binding of thecapacity.
at ZINC12
AutoDock 4
Epilepsy packages:
Software
Methods:
The
10uM. lead
motor
database. compounds
Compound Nearly 16
activity. resulted
had the
300,000 small in amolecules
greatest reducedpotency num-
were as
Software
Glide
Docking packages: aber of
selectiveseizures inverse observed agonist.
filtered and docked. The filtering and docking us-in a mouse
Ligands model
with 80% of sim-
Methods:
Drug target: ilarity
ingA VS tostudy
compound
standard wasand performed 16 were
extra screened
precision
against the CB1 and as-
settings
receptor inusing
Glide a
Docking sessednatural for products
CB1 and subset CB2 of the ZINC12Two
activity. database.
ligands Nearly
Cannabinoid receptor 1 indicated 32 top-performing ligands, of which 18
300,000 small molecules were filtered and docked. The filtering
Drug(CB1)
target: were andidentified
were selected
docking using that
forstandardhad nanomolar
further in extra
and affinity
vitroprecision
testing to- in
through
settings
Cannabinoid target:1 (CB1) wards
Diseasereceptor GlideCB1.
clustering indicated This 32provided
to ensure top-performing
structuralkey ligands,
information
diversity of which for
18 were
amongst
Disease target: selected for further in vitro testing through clustering to ensure
Substance
Substance abuse disorders
further
abuse disor- hits. structural
Of the
structural 18 ligands,
diversity
optimization
amongst7 hits.
for
demonstrated inverse
Of the 18 ligands,
ago-
more than
Software nists targeting
7 demonstrated CB1.
more than 50% displacement in competitive
[144] [144]
ders packages: 50% displacement in competitive binding at
Glide binding at 10uM. Compound 16 had the greatest potency as a
Software packages: The10uM. role of caspase-1
Compound in febrile
16 Ligands
had thewith seizures
greatest was
potencyini- as
Methods: selective inverse agonist. 80% similarity to
Molecules
Glide
Docking 2023, 28, x FOR PEER tially
REVIEW assessed.
acompound
selective 16 Mice
inverse
were with the
agonist.
screened caspase-1
andLigands
assessed with
for gene
CB1 80%
and CB2sim- 17 of 24
activity.out
knocked Twodid ligandsnotwere developidentified that had
febrile nanomolar
seizures, and
Methods: ilarity to compound 16 were screened
affinity towards CB1. This provided key information for further
and as-
Docking their
sessed wild-type
structural CB1litter
foroptimization and mates CB2 hadagonists
activity.
for inverse an Twoincrease ligands
targeting inCB1.
Drug Target and caspase-1
were
The role prior
identified
of tothat
caspase-1 the in onset
febrile of a febrile
had nanomolar
seizures seizure.
was affinity
initially to-
assessed.
Drug Methodology
target: One Micemillion
with the Study gene
compounds
caspase-1 Significance
fromknockedtheinformation
ChemBridge
out did not develop Chemical Structure Reference
wards CB1. This provided key for
febrile seizures, and their wild-type litter mates had an increase
Caspase-1
Drug target:
database
further
More than
in caspase-1
were
structural
73,000 docked
prior to small
the
against
optimization
molecules
onset
thefor
of a febrile
active
inverse
from
seizure.
site
the ofmillion
ago-
Specs
One
ca-
Disease target:
Caspase-1 pase-1.
nists
natural The
targeting
compounds top
compounds from2000CB1. ligands
the ChemBridge
and PubChem from the
database extra
were docked
databases preci-
Disease
Febriletarget:
seizures sion against
docking the active
stage sitewereof capase-1.
filtered Thetotop 2000 ligands
ensure they from
Drug
Febrile target:
seizures The
were role precision
the docked
extra of caspase-1
against docking instage
PGK-1 febrile
in seizures
search
were filtered was
oftoagonists
ensure ini-they
Software
Software packages:
packages:
Phosphoglycerate ki- 2023,had
Molecules tially
28,
to hadsuitable
x FOR
protect assessed.
suitable
PEER drug
drug Mice
REVIEW
against properties.
properties.
brain with The
damage the The inremainder
caspase-1
remainder strokeweregene were for
clustered
patients. 17 of 24
Glide chemicalfor similarity usingsimilarity
the Tanimoto co-efficient. Fifty
[145]
[145]
Glide
nase-1 clustered
knocked out chemical
did was not develop using
febrile thethatTan-
seizures, and
AMBER 14 (PGK1) The initial
ligands library
were purchased filtered
for to
experimental confirm
validation ofthe
AMBER 14
Disease target:
Methods: imoto
their
small co-efficient.
wild-type
molecules
predicted bindingpossessed Fifty
litter
affinity.matesligands were
had an increase
Fourdrug-like
compounds purchased
properties.
had potent in
Methods:
Docking Drug Target
for and
experimental
inhibitory
caspase-1 effects
prior validation
onto caspase-1.
the onset of
When predicted
of a compared
febrile binding
to diazepam,
seizure.
Stroke
Molecular dynamics
The remaining 35,414 ligands Study underwent
the top compound, CZL80, showed a capacity to prevent the
Significance HTVS in Chemical Structure Reference
Docking Methodology
affinity. Four compounds had potent inhibitory
Drug target:
Software packages: Oneonsetmillion
LibDock of and
a second compounds
the remaining
episode of from
FS, top
with the
4% ChemBridge
were
diazepam docked
not being able
More than 73,000 small molecules from the Specs [146]
Molecular
Caspase-1
Discovery dynamics
Studio effects
database
using on
this.caspase-1.
to doextra were
CZL80
precision also When
docked reduced
(XP) compared
against
inthe the
risk
Glide. ofThe to highest
active
adult diaze-
site ofwhen
epilepsy ca-
administered natural
after compounds
an episode of and PubChem
febrile seizures. databases
Disease
LibDock target: pam,
pase-1. the
Drug target:ranked 20%were
top
The compound,
top
of ligands 2000
docked from
CZL80,
ligands
againstXP fromshowed
docking
PGK-1 the a
extra
in search were capac-
preci-
of agonists
ity More than 73,000
to prevent the small
onset molecules
of afiltered from the
second Specs natural
episode of FS,
Febrile
Glide
Drug target:
seizuresPhosphoglyceratesion
clustered docking
ki- to stage
toascertain
protect were
against
chemical
compounds and PubChem databases were docked against
brain to
damage
similarity ensure
in stroke they
amongst patients.
Software packages:
Canvas
Phosphoglycerate nase-1
kinase-1 (PGK1) with
(PGK1) had
hits. diazepam
A
PGK-1 suitable
total The
in search not
initial
of drug
19 being
library
properties.
of compounds
agonists able
to was to
protect The
from do
filtered this.
remainder
the
against CZL80
to different
confirm
brain that the
were
damage in
Disease target: Disease target:
also reduced
stroke small
patients. the Themolecules
risk of
initial possessed
adult
library epilepsy drug-like
was filtered when
tothe properties.
ad-
confirm that [145]
Glide
Methods: clustered
clusteres for chemical
were selected similarity
for experimental using valida-Tan-
Stroke Stroke the small molecules
The remaining possessed drug-like
35,414 ligands properties.
underwent TheHTVS in
AMBER
Docking
Software 14
packages:
ministered
imoto
tion. Two
remaining
after
co-efficient.
ligands, an7979989
35,414 ligands
episode
Fifty of febrile
ligands
and
underwent were
Z112553128,
HTVS
seizures.
inpurchased
LibDock wereand the
Software packages: LibDock and the remaining top 4% were docked
Discovery Studio
Methods: remaining
for
noted top 4% were
experimental
as potential docked
validation
PGK1 using
of
activators extra precision
predicted (XP) in
binding [146]
[146]
LibDock Discovery Studio
Glide. The using
highestextra precision
ranked (XP)
20% of ligands inas demon-
Glide.
from The
XP highest
docking
Docking
Glide affinity.
LibDock strated in Four
a ranked
were clustered compounds
Drosophilia
to ascertain hadfrom
20% ofoxidative
ligands
chemical potent
stress
similarity inhibitory
model.
XP docking
amongstwerehits. A
Molecular dynamics
Canvas Glide Aeffects on
total ofof
total 196.2caspase-1.
clustered
milliontofrom
compounds When
ascertain compared
chemical
the different
compounds to diaze-
similarity
clusteres
and wereamongst
fragments
Methods: Canvas selected forhits.
experimental
A total ofvalidation.
19 compoundsTwo ligands, 7979989
from the and
Docking
pam,
from the top
ZINC
Z112553128,
compound,
database
were notedwere
CZL80,
screened
as potential
showed
PGK1to searcha different
activators
capac-
for
as
Methods: clusteres were selected for experimental valida-
Drug target: ity to prevent
negative inthe
allosteric
demonstrated onset ofoxidative
modulators
a Drosophilia a second
(NAMs) episode
stress of FS,
of mGlu5.
model.
Docking tion. Two ligands, 7979989 and Z112553128, were
with
Metabotropic glutamate Initially, diazepam
A total docking
of 6.2 not
million
noted was
as
being
compounds
potential
able
benchmarkedandto
PGK1
dousing
this.from
fragments
activators
CZL80
asan ini-
ZINC
demon-
Drug target:
database
also were screened
reduced the risk toof
search
adult forepilepsy
negative allosteric
when ad-
receptor 5glutamate
Metabotropic (mGlu5)receptor tialmodulators
library of known
strated in a NAMs
Drosophilia and decoysstress
oxidative
(NAMs) of mGlu5. Initially, docking was
withmodel.
5 (mGlu5)
Disease target: ministered
structural Aafter an
similarities. episode
From of
this,febrile seizures.
benchmarked total
usingofan6.2 million
initial library of59
compounds leads
known andand
NAMs 59
fragments
and
Disease target:
Fragile
Fragile X syndrome
X syndrome fragments were identified for experimental vali- for
decoys from
with ZINC
structural database
similarities. were
From screened
this, 59 to
leadssearch
and
Depression Drug target: 59 fragments were identified
negative allosteric formodulators
experimental validation.
(NAMs) of mGlu5. [147]
[147]
Depression dation. Inassessments
In vitro vitro assessments
identified 11identified 11 identi-
identified molecules as
Software packages: Metabotropic glutamate Initially, docking was benchmarked using an ini-
Software packages: fiedNAMs. Compound
molecules as F1 demonstrated
NAMs. Compound the greatest
F1 level in terms
demon-
DOCK3.6 receptor 5 (mGlu5) tial library ofTanimoto
known NAMs and assessment
decoys with
of novelty in a pairwise co-efficient with
DOCK3.6
Methods: strated
Disease target: the greatest
other mGlu5 structural
ligands onlevel in terms
similarities.
the ChEMBL Fromofthis,
novelty
database. 59 inhad
F1leads
also aandthe
59
Docking
Methods: pairwise
greatest Tanimoto
Fragile X syndrome affinity,
fragmentswithco-efficient
an assessment
Ki ofidentified
were 0.43µM. with vali-
for experimental
[147]
Docking Depression other mGlu5 dation.
ligandsIn vitro assessments
on the ChEMBL identified
database. 11 identi-
F1
Software packages: fied molecules as NAMs.
also had the greatest affinity, with an Ki of Compound F1 demon-
DOCK3.6 strated the greatest level in terms of novelty in a
0.43μM.
Methods: pairwise Tanimoto co-efficient assessment with
Drug target: Docking
MD studiesother performedmGlu5 ligands on the ChEMBL
on a homology model database.
of F1
 Fragile X syndrome fragments were identified for experimental vali-
[147]
Depression dation. In vitro assessments identified 11 identi-
Software packages: fied molecules as NAMs. Compound F1 demon-
DOCK3.6 strated the greatest level in terms of novelty in a
Methods:
Molecules 2023, 28, 1324 pairwise Tanimoto co-efficient assessment with 16 of 22
Docking other mGlu5 ligands on the ChEMBL database. F1
also had the greatest affinity, with an Ki of
0.43μM.
Table 2. Cont.
Drug target:
MD studies performed on a homology model of
Excitatory amino
Drug Target and acid
Study Significance Chemical Structure Reference
Methodology the EAAT2 suggested the presence of an allosteric
transporter 2
Drug target: binding site. Five key residues from the allosteric
(EAAT2)
Excitatory amino acid site were identified as key binding residues
Disease
transporter 2 target:
(EAAT2) through
MD studiessite-directed
performed onand functional
a homology model mutagenesis
of the EAAT2
Stroke suggested the presence of an allosteric binding site. Five key
Disease target: studies. The virtual screening of 3 million small
residues from the allosteric site were identified as key binding
Brain trauma
Stroke
molecules was performed
residues through against
site-directed and this mutagenesis
functional pharmaco-
Brain trauma
Neurodegenerative dis- studies.After
phore. The virtual
virtual screening of 3 million
screening small molecules
and filtering for fa-was
Neurodegenerative disorders
orders
Software packages:
performed against this pharmacophore. After virtual screening
vourable ADMET
and filtering properties
for favourable ADMET and no Lipinski’s
properties and no vi-
Software packages:
MODELLER
Lipinski’s58violations,
[148][148]
Desmond olations, ligands58were ligands were selected
selected for docking
for docking
MODELLER against the EAAT2 homology model. The docking studies
Sybyl 8.1 Unity against
yielded the EAAT2ofhomology
10 molecules model.assessment.
interest for further The dockingA
GOLDDesmond
Methods: studies yielded 10 molecules of interest for
SciFinder search confirmed the novelty of these fur-
ligands.
Sybyl 8.1 Unity In vitro testing confirmed four compounds as NAMs, three as
Homology modelling ther assessment. A SciFinder search confirmed
PAMs and three as inactive against EAAT2. One of the top
GOLDdynamics
Molecular
the novelty of these GT949,
performing molecules, ligands. In vitro
possessed testing potency.
nanomolar con-
Methods:
Hybrid structure-based
pharmacophore firmed four compounds as NAMs, three as PAMs
Homology
Docking
modelling
and three as inactive against EAAT2. One of the
Molecular dynamics
top performing molecules, GT949, possessed na-
Hybrid structure-based
nomolar potency.
7. Conclusions
pharmacophore
Computational drug design is a powerful tool in the pursuit to discover new therapies
for neurological and psychiatric conditions. This literature review presents a distillation of
the computational approaches to drug design from multi-disciplinary research articles to
highlight the importance of CADD in finding new CNS drugs. The exemplars provided
demonstrate how CADD methodologies have been utilised to present a basis for future
drug development. It is apparent from the literature presented that progress towards novel
therapies in CNS drug discovery is being made, in particular for drug targets where there
are no therapeutics available. It is hoped that by creating novel medications towards these
new targets, the standard of patient care will improve by both increased treatment efficacy
and reducing side effects. In the context of CNS computational drug discovery, there
is a clear trend that the majority of research is focused on investigating new treatments
for neurodegenerative disorders such as Alzheimer’s and Parkinson’s disease. However,
psychiatric conditions such as schizophrenia and substance abuse disorder, as well as
neurological conditions such as brain injuries and neuropathic pain, are other research
foci, all of which present significant disease burdens for both the patient and society as
a whole. Although it is evident there is still a long way to go for new lead molecules to
become approved drugs, CADD is helping to hasten this process. As newer techniques
such as deep learning become more mainstream within academic drug design research, it
is expected that the efficiency with which studies can be carried out will greatly increase.

Author Contributions: Investigation, G.D; writing—original draft preparation, G.D.; writing—

review and editing, G.D., J.Z.C. and T.B.; supervision, T.B. All authors have read and agreed to
the published version of the manuscript.
Funding: G.D is funded by a Research Training Program (RTP) scholarship jointly funded by NSW
Health, the University of Sydney and Saniona A/S, and J.Z.C is funded by an Australian Government
RTP Scholarship. The authors are supported by the Australian Government’s National Collaborative
Research Infrastructure Strategy (NCRIS), with access to computational resources provided by the
National Computing Infrastructure (NCI) through the National Computational Merit Allocation
Scheme (NCMAS-2022-154). Furthermore, the authors acknowledge the computational resources
and technical assistance provided by the Sydney Informatics Hub, a Core Research Facility of the
University of Sydney.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Molecules 2023, 28, 1324 17 of 22

Data Availability Statement: Not applicable.

Conflicts of Interest: The authors declare no conflict of interest.

References
1. Weller, J.; Budson, A. Current understanding of Alzheimer’s disease diagnosis and treatment. F1000Research 2018, 7, F1000
Faculty Rev-1161. [CrossRef]
2. De Bie, R.M.; Clarke, C.E.; Espay, A.J.; Fox, S.H.; Lang, A.E. Initiation of pharmacological therapy in Parkinson’s disease: When,
why, and how. Lancet Neurol. 2020, 19, 452–461. [CrossRef] [PubMed]
3. De Hert, M.; Detraux, J.; Van Winkel, R.; Yu, W.; Correll, C.U. Metabolic and cardiovascular adverse effects associated with
antipsychotic drugs. Nat. Rev. Endocrinol. 2012, 8, 114–126. [CrossRef] [PubMed]
4. Harrison, R.K. Phase II and phase III failures: 2013–2015. Nat. Rev. Drug Discov. 2016, 15, 817–818. [CrossRef] [PubMed]
5. Brown, D.G.; Wobst, H.J. A Decade of FDA-Approved Drugs (2010–2019): Trends and Future Directions. J. Med. Chem. 2021, 64,
2312–2338. [CrossRef]
6. Congressional Budget Office. Research and Development in the Pharmaceutical Industry; Congressional Budget Office: Washington,
DC, USA, 2021.
7. Katritzky, A.R.; Kuanar, M.; Slavov, S.; Hall, C.D.; Karelson, M.; Kahn, I.; Dobchev, D.A. Quantitative Correlation of Physical and
Chemical Properties with Chemical Structure: Utility for Prediction. Chem. Rev. 2010, 110, 5714–5789. [CrossRef]
8. Leo, A.J. Calculating log Poct from structures. Chem. Rev. 1993, 93, 1281–1306. [CrossRef]
9. Lipinski, C.A.; Lombardo, F.; Dominy, B.W.; Feeney, P.J. Experimental and computational approaches to estimate solubility and
permeability in drug discovery and development settings. Adv. Drug Deliv. Rev. 1997, 23, 3–25. [CrossRef]
10. Clark, D.E. In silico prediction of blood–brain barrier permeation. Drug Discov. Today 2003, 8, 927–933. [CrossRef]
11. Kearsley, S.K.; Sallamack, S.; Fluder, E.M.; Andose, J.D.; Mosley, R.T.; Sheridan, R.P. Chemical similarity using physiochemical
property descriptors. J. Chem. Inf. Comput. Sci. 1996, 36, 118–127. [CrossRef]
12. Willett, P.; Barnard, J.M.; Downs, G.M. Chemical similarity searching. J. Chem. Inf. Comput. Sci. 1998, 38, 983–996. [CrossRef]
13. Wermuth, C.; Ganellin, C.; Lindberg, P.; Mitscher, L. Glossary of terms used in medicinal chemistry (IUPAC Recommendations
1998). Pure Appl. Chem. 1998, 70, 1129–1143. [CrossRef]
14. Zhang, H.; He, X.; Wang, X.; Yu, B.; Zhao, S.; Jiao, P.; Jin, H.; Liu, Z.; Wang, K.; Zhang, L.; et al. Design, synthesis and biological
activities of piperidine-spirooxadiazole derivatives as alpha7 nicotinic receptor antagonists. Eur. J. Med. Chem. 2020, 207, 112774.
[CrossRef]
15. Ghose, A.K.; Wendoloski, J.J. Pharmacophore modelling: Methods, experimental verification and applications. In 3D QSAR in
Drug Design; Springer: Berlin/Heidelberg, Germany, 2002; pp. 253–271.
16. Poptodorov, K.; Luu, T.; Hoffmann, R.D. Pharmacophore model generation software tools. Pharm. Pharm. Searches 2006, 32, 15–47.
17. Mysinger, M.M.; Carchia, M.; Irwin, J.J.; Shoichet, B.K. Directory of useful decoys, enhanced (DUD-E): Better ligands and decoys
for better benchmarking. J. Med. Chem. 2012, 55, 6582–6594. [CrossRef]
18. Sanders, M.P.; Barbosa, A.J.; Zarzycka, B.; Nicolaes, G.A.; Klomp, J.P.; De Vlieg, J.; Del Rio, A. Comparative analysis of
pharmacophore screening tools. J. Chem. Inf. Model. 2012, 52, 1607–1620. [CrossRef]
19. Wolber, G.; Langer, T. LigandScout: 3-D pharmacophores derived from protein-bound ligands and their use as virtual screening
filters. J. Chem. Inf. Model. 2005, 45, 160–169. [CrossRef]
20. Van Drie, J.H. Generation of three-dimensional pharmacophore models. Wiley Interdiscip. Rev. Comput. Mol. Sci. 2013, 3, 449–464.
[CrossRef]
21. Dixon, S.L.; Smondyrev, A.M.; Knoll, E.H.; Rao, S.N.; Shaw, D.E.; Friesner, R.A. PHASE: A new engine for pharmacophore
perception, 3D QSAR model development, and 3D database screening: 1. Methodology and preliminary results. J. Comput. Aided.
Mol. Des. 2006, 20, 647–671. [CrossRef]
22. Vuorinen, A.; Schuster, D. Methods for generating and applying pharmacophore models as virtual screening filters and for
bioactivity profiling. Methods 2015, 71, 113–134. [CrossRef]
23. Giordano, D.; Biancaniello, C.; Argenio, M.A.; Facchiano, A. Drug Design by Pharmacophore and Virtual Screening Approach.
Pharmaceuticals 2022, 15, 646. [CrossRef] [PubMed]
24. Hansch, C.; Maloney, P.P.; Fujita, T.; Muir, R.M. Correlation of Biological Activity of Phenoxyacetic Acids with Hammett
Substituent Constants and Partition Coefficients. Nature 1962, 194, 178–180. [CrossRef]
25. Huang, C.; Zhong, Q.P.; Tang, L.; Wang, H.T.; Xu, J.P.; Zhou, Z.Z. Discovery of 2-(3, 4-dialkoxyphenyl)-2-(substituted pyridazin-3-
yl) acetonitriles as phosphodiesterase 4 inhibitors with anti-neuroinflammation potential based on three-dimensional quantitative
structure–activity relationship study. Chem. Biol. Drug Des. 2019, 93, 484–502. [CrossRef] [PubMed]
26. OECD. Guidance Document on the Validation of (Quantitative) Structure-Activity Relationship [(Q)SAR] Models; OECD: Paris,
France, 2014.
27. Golbraikh, A.; Tropsha, A. Beware of q2! J. Mol. Graph. Model. 2002, 20, 269–276. [CrossRef]
28. Dearden, J.C.; Cronin, M.T.; Kaiser, K.L. How not to develop a quantitative structure-activity or structure-property relationship
(QSAR/QSPR). SAR QSAR Environ. Res. 2009, 20, 241–266. [CrossRef]
Molecules 2023, 28, 1324 18 of 22

29. Netzeva, T.I.; Worth, A.P.; Aldenberg, T.; Benigni, R.; Cronin, M.T.; Gramatica, P.; Jaworska, J.S.; Kahn, S.; Klopman, G.;
Marchant, C.A. Current status of methods for defining the applicability domain of (quantitative) structure-activity relationships:
The report and recommendations of ecvam workshop 52. Altern. Lab. Anim. 2005, 33, 155–173. [CrossRef]
30. Manoj, G.D.; Sanjay, N.H.; Firoz, A.K.K.; Devanand, B.S.; Jaiprakash, N.S. Recent advances in multidimensional QSAR (4D-6D): A
critical review. Mini Rev. Med. Chem. 2014, 14, 35–55.
31. Mauri, A.; Consonni, V.; Todeschini, R. Molecular Descriptors. In Handbook of Computational Chemistry; Leszczynski, J., Kaczmarek-
Kedziera, A., Puzyn, T.G., Papadopoulos, M., Reis, H., Shukla, K.M., Eds.; Springer International Publishing: Cham, Switzerland,
2017; pp. 2065–2093.
32. Testa, B.; Kier, L.B. The concept of molecular structure in structure–activity relationship studies and drug design. Med. Res. Rev.
1991, 11, 35–48. [CrossRef]
33. Patel, H.M.; Noolvi, M.N.; Sharma, P.; Jaiswal, V.; Bansal, S.; Lohan, S.; Kumar, S.S.; Abbot, V.; Dhiman, S.; Bhardwaj, V.
Quantitative structure–activity relationship (QSAR) studies as strategic approach in drug discovery. Med. Chem. Res. 2014, 23,
4991–5007. [CrossRef]
34. Wold, S.; Eriksson, L.; Clementi, S. Statistical validation of QSAR results. In Chemometric Methods in Molecular Design; VCH:
Weinheim, Germany, 1995; pp. 309–338.
35. Cover, T.; Hart, P. Nearest neighbor pattern classification. IEEE Trans. Inf. Theory 1967, 13, 21–27. [CrossRef]
36. Cortes, C.; Vapnik, V. Support-vector networks. Mach. Learn. 1995, 20, 273–297. [CrossRef]
37. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
38. Tropsha, A. Best Practices for QSAR Model Development, Validation, and Exploitation. Mol. Inform. 2010, 29, 476–488. [CrossRef]
39. Gramatica, P. Principles of QSAR models validation: Internal and external. QSAR Comb. Sci. 2007, 26, 694–701. [CrossRef]
40. Consonni, V.; Ballabio, D.; Todeschini, R. Evaluation of model predictive ability by external validation techniques. J. Chemom.
2010, 24, 194–201. [CrossRef]
41. Roy, K.; Das, R.N.; Ambure, P.; Aher, R.B. Be aware of error measures. Further studies on validation of predictive QSAR models.
Chemom. Intell. Lab. Syst. 2016, 152, 18–33. [CrossRef]
42. Jumper, J.; Evans, R.; Pritzel, A.; Green, T.; Figurnov, M.; Ronneberger, O.; Tunyasuvunakool, K.; Bates, R.; Žídek, A.; Potapenko, A.
Highly accurate protein structure prediction with AlphaFold. Nature 2021, 596, 583–589. [CrossRef]
43. Baek, M.; DiMaio, F.; Anishchenko, I.; Dauparas, J.; Ovchinnikov, S.; Lee, G.R.; Wang, J.; Cong, Q.; Kinch, L.N.; Schaeffer, R.D.; et al.
Accurate prediction of protein structures and interactions using a three-track neural network. Science 2021, 373, 871–876. [CrossRef]
44. Lyu, J.; Wang, S.; Balius, T.E.; Singh, I.; Levit, A.; Moroz, Y.S.; O’Meara, M.J.; Che, T.; Algaa, E.; Tolmachova, K.; et al. Ultra-large
library docking for discovering new chemotypes. Nature 2019, 566, 224–229. [CrossRef]
45. Sussman, J.L.; Lin, D.; Jiang, J.; Manning, N.O.; Prilusky, J.; Ritter, O.; Abola, E.E. Protein Data Bank (PDB): Database of three-
dimensional structural information of biological macromolecules. Acta Crystallogr. Sect. D Biol. Crystallogr. 1998, 54, 1078–1084.
[CrossRef]
46. Martí-Renom, M.A.; Stuart, A.C.; Fiser, A.; Sánchez, R.; Melo, F.; Sali, A. Comparative protein structure modeling of genes and
genomes. Annu. Rev. Biophys. Biomol. Struct. 2000, 29, 291–325. [CrossRef] [PubMed]
47. Blundell, T.L.; Sibanda, B.L.; Sternberg, M.J.; Thornton, J.M. Knowledge-based prediction of protein structures and the design of
novel molecules. Nature 1987, 326, 347–352. [CrossRef]
48. Pearson, W.R. An introduction to sequence similarity (“homology”) searching. Curr. Protoc. Bioinform. 2013, 42, 3-1. [CrossRef]
[PubMed]
49. Consortium, U. UniProt: A hub for protein information. Nucleic Acids Res. 2015, 43, D204–D212. [CrossRef] [PubMed]
50. Boeckmann, B.; Bairoch, A.; Apweiler, R.; Blatter, M.-C.; Estreicher, A.; Gasteiger, E.; Martin, M.J.; Michoud, K.; O’Donovan, C.;
Phan, I. The SWISS-PROT protein knowledgebase and its supplement TrEMBL in 2003. Nucleic Acids Res. 2003, 31, 365–370.
[CrossRef] [PubMed]
51. Notredame, C.; Higgins, D.G.; Heringa, J. T-Coffee: A novel method for fast and accurate multiple sequence alignment. J. Mol.
Biol. 2000, 302, 205–217. [CrossRef]
52. Ye, J.; McGinnis, S.; Madden, T.L. BLAST: Improvements for better sequence analysis. Nucleic Acids Res. 2006, 34, W6–W9.
[CrossRef]
53. Webb, B.; Sali, A. Comparative protein structure modeling using MODELLER. Curr. Protoc. Bioinform. 2016, 54, 5–6. [CrossRef]
54. Waterhouse, A.; Bertoni, M.; Bienert, S.; Studer, G.; Tauriello, G.; Gumienny, R.; Heer, F.T.; de Beer, T.A.P.; Rempfer, C.;
Bordoli, L.; et al. SWISS-MODEL: Homology modelling of protein structures and complexes. Nucleic Acids Res. 2018, 46,
W296–W303. [CrossRef]
55. Jacobson, M.P.; Pincus, D.L.; Rapp, C.S.; Day, T.J.; Honig, B.; Shaw, D.E.; Friesner, R.A. A hierarchical approach to all-atom protein
loop prediction. Proteins Struct. Funct. Bioinform. 2004, 55, 351–367. [CrossRef]
56. Molecular Operating Environment (MOE). Chemical Computing Group ULC; Molecular Operating Environment (MOE): Montreal,
QC, Canada, 2022.
57. Lohning, A.E.; Levonis, S.M.; Williams-Noonan, B.; Schweiker, S.S. A practical guide to molecular docking and homology
modelling for medicinal chemists. Curr. Top. Med. Chem. 2017, 17, 2023–2040. [CrossRef]
58. Pitman, M.R.; Menz, R.I. 2—Methods for Protein Homology Modelling. In Applied Mycology and Biotechnology; Berka, R.M.,
Singh, G.B., Arora, D.K.., Eds.; Elsevier: Amsterdam, The Netherlands, 2006; Volume 6, pp. 37–59.
Molecules 2023, 28, 1324 19 of 22

59. Shen, M.Y.; Sali, A. Statistical potential for assessment and prediction of protein structures. Protein Sci. 2006, 15, 2507–2524.
[CrossRef]
60. Benkert, P.; Tosatto, S.C.; Schomburg, D. QMEAN: A comprehensive scoring function for model quality assessment. Proteins 2008,
71, 261–277. [CrossRef]
61. Djikic, T.; Marti, Y.; Spyrakis, F.; Lau, T.; Benedetti, P.; Davey, G.; Schloss, P.; Yelekci, K. Human dopamine transporter: The first
implementation of a combined in silico/in vitro approach revealing the substrate and inhibitor specificities. J. Biomol. Struct. Dyn.
2019, 37, 291–306. [CrossRef]
62. Kowal, N.M.; Indurthi, D.C.; Ahring, P.K.; Chebib, M.; Olafsdottir, E.S.; Balle, T. Novel Approach for the Search for Chemical
Scaffolds with Dual Activity with Acetylcholinesterase and the alpha7 Nicotinic Acetylcholine Receptor-A Perspective for the
Treatment of Neurodegenerative Disorders. Molecules 2019, 24, 446. [CrossRef]
63. Kitchen, D.B.; Decornez, H.; Furr, J.R.; Bajorath, J. Docking and scoring in virtual screening for drug discovery: Methods and
applications. Nat. Rev. Drug Discov. 2004, 3, 935–949. [CrossRef] [PubMed]
64. Morris, G.M.; Lim-Wilby, M. Molecular docking. In Molecular Modeling of Proteins; Springer: Berlin/Heidelberg, Germany, 2008;
pp. 365–382.
65. Lorber, D.M.; Shoichet, B.K. Flexible ligand docking using conformational ensembles. Protein Sci. 1998, 7, 938–950. [CrossRef]
66. Sherman, W.; Beard, H.S.; Farid, R. Use of an induced fit receptor structure in virtual screening. Chem. Biol. Drug Des. 2006, 67,
83–84. [CrossRef]
67. Gurram, P.C.; Satarker, S.; Kumar, G.; Begum, F.; Mehta, C.; Nayak, U.; Mudgal, J.; Arora, D.; Nampoothiri, M. Avanafil mediated
dual inhibition of IKKβ and TNFR1 in an experimental paradigm of Alzheimer’s disease: In silico and in vivo approach. J. Biomol.
Struct. Dyn. 2022. [CrossRef]
68. Huang, S.Y.; Grinter, S.Z.; Zou, X. Scoring functions and their evaluation methods for protein-ligand docking: Recent advances
and future directions. Phys. Chem. Chem. Phys. 2010, 12, 12899–12908. [CrossRef] [PubMed]
69. Chandra, S.; Wang, Z.; Tao, X.; Chen, O.; Luo, X.; Ji, R.R.; Bortsov, A.V. Computer-aided Discovery of a New Nav1.7 Inhibitor for
Treatment of Pain and Itch. Anesthesiology 2020, 133, 611–627. [CrossRef]
70. Böhm, H.-J. Prediction of binding constants of protein ligands: A fast method for the prioritization of hits obtained from de novo
design or 3D database search programs. J. Comput. Aided Mol. Des. 1998, 12, 309. [CrossRef]
71. Gehlhaar, D.K.; Verkhivker, G.M.; Rejto, P.A.; Sherman, C.J.; Fogel, D.B.; Fogel, L.J.; Freer, S.T. Molecular recognition of the
inhibitor AG-1343 by HIV-1 protease: Conformationally flexible docking by evolutionary programming. Chem. Biol. 1995, 2,
317–324. [CrossRef]
72. Wang, R.; Lai, L.; Wang, S. Further development and validation of empirical scoring functions for structure-based binding affinity
prediction. J. Comput. Aided Mol. Des. 2002, 16, 11–26. [CrossRef]
73. Li, X.-L.; Hou, M.-L.; Wang, S.-L. A residual level potential of mean force based approach to predict protein-protein interaction
affinity. In Proceedings of the International Conference on Intelligent Computing, Changsha, China, 18–21 August 2010; Springer:
Berlin/Heidelberg, Germany, 2010; pp. 680–686.
74. Head, R.D.; Smythe, M.L.; Oprea, T.I.; Waller, C.L.; Green, S.M.; Marshall, G.R. VALIDATE: A new method for the receptor-based
prediction of binding affinities of novel ligands. J. Am. Chem. Soc. 1996, 118, 3959–3969. [CrossRef]
75. Yin, S.; Biedermannova, L.; Vondrasek, J.; Dokholyan, N.V. MedusaScore: An accurate force field-based scoring function for
virtual drug screening. J. Chem. Inf. Model. 2008, 48, 1656–1662. [CrossRef]
76. Hollingsworth, S.A.; Dror, R.O. Molecular dynamics simulation for all. Neuron 2018, 99, 1129–1143. [CrossRef]
77. Totrov, M.; Abagyan, R. Flexible ligand docking to multiple receptor conformations: A practical alternative. Curr. Opin. Struct.
Biol. 2008, 18, 178–184. [CrossRef]
78. Huang, S.-Y. Comprehensive assessment of flexible-ligand docking algorithms: Current effectiveness and challenges. Brief.
Bioinform. 2017, 19, 982–994. [CrossRef]
79. Tan, S.; Gong, X.; Liu, H.; Yao, X. Virtual Screening and Biological Activity Evaluation of New Potent Inhibitors Targeting LRRK2
Kinase Domain. ACS Chem. Neurosci. 2021, 12, 3214–3224. [CrossRef]
80. Brooks, B.R.; Bruccoleri, R.E.; Olafson, B.D.; States, D.J.; Swaminathan, S.; Karplus, M. CHARMM: A program for macromolecular
energy, minimization, and dynamics calculations. J. Comput. Chem. 1983, 4, 187–217. [CrossRef]
81. Van Gunsteren, W.F.; Billeter, S.; Eising, A.; Hünenberger, P.; Krüger, P.; Mark, A.; Scott, W.; Tironi, I. Biomolecular Simulation:
The GROMOS96 Manual and User Guide; Vdf Hochschulverlag AG an der ETH Zürich: Zürich, Switzerland, 1996; Volume 86,
pp. 1–1044.
82. Wang, J.; Wolf, R.M.; Caldwell, J.W.; Kollman, P.A.; Case, D.A. Development and testing of a general amber force field. J. Comput.
Chem. 2004, 25, 1157–1174. [CrossRef] [PubMed]
83. Jorgensen, W.L.; Maxwell, D.S.; Tirado-Rives, J. Development and Testing of the OPLS All-Atom Force Field on Conformational
Energetics and Properties of Organic Liquids. J. Am. Chem. Soc. 1996, 118, 11225–11236. [CrossRef]
84. Van Der Spoel, D.; Lindahl, E.; Hess, B.; Groenhof, G.; Mark, A.E.; Berendsen, H.J. GROMACS: Fast, flexible, and free. J. Comput.
Chem. 2005, 26, 1701–1718. [CrossRef]
85. Case, D.A.; Aktulga, H.M.; Belfon, K.; Ben-Shalom, I.; Brozell, S.R.; Cerutti, D.S.; Cheatham, T.E., III; Cruzeiro, V.W.D.;
Darden, T.A.; Duke, R.E. Amber 2021; University of California: San Francisco, CA, USA, 2021.
Molecules 2023, 28, 1324 20 of 22

86. Eastman, P.; Swails, J.; Chodera, J.D.; McGibbon, R.T.; Zhao, Y.; Beauchamp, K.A.; Wang, L.-P.; Simmonett, A.C.; Harrigan, M.P.;
Stern, C.D. OpenMM 7: Rapid development of high performance algorithms for molecular dynamics. PLoS Comput. Biol. 2017,
13, e1005659. [CrossRef]
87. Bowers, K.J.; Chow, E.; Xu, H.; Dror, R.O.; Eastwood, M.P.; Gregersen, B.A.; Klepeis, J.L.; Kolossvary, I.; Moraes, M.A.;
Sacerdoti, F.D.; et al. Scalable algorithms for molecular dynamics simulations on commodity clusters. In Proceedings of the 2006
ACM/IEEE Conference on Supercomputing, Tampa, FL, USA, 11–17 November 2006; p. 84-es.
88. Marrink, S.J.; Risselada, H.J.; Yefimov, S.; Tieleman, D.P.; De Vries, A.H. The MARTINI force field: Coarse grained model for
biomolecular simulations. J. Phys. Chem. B 2007, 111, 7812–7824. [CrossRef]
89. Nickolls, J.; Buck, I.; Garland, M.; Skadron, K. Scalable parallel programming with cuda: Is cuda the parallel programming model
that application developers have been waiting for? Queue 2008, 6, 40–53. [CrossRef]
90. Park, H.; Kim, T.; Kim, K.; Jang, A.; Hong, S. Structure-Based Virtual Screening and De Novo Design to Identify Submicromolar
Inhibitors of G2019S Mutant of Leucine-Rich Repeat Kinase 2. Int. J. Mol. Sci. 2022, 23, 12825. [CrossRef]
91. Abbott, N.J.; Patabendige, A.A.; Dolman, D.E.; Yusof, S.R.; Begley, D.J. Structure and function of the blood–brain barrier. Neurobiol.
Dis. 2010, 37, 13–25. [CrossRef]
92. Morgan, P.; Van Der Graaf, P.H.; Arrowsmith, J.; Feltner, D.E.; Drummond, K.S.; Wegner, C.D.; Street, S.D.A. Can the flow of
medicines be improved? Fundamental pharmacokinetic and pharmacological principles toward improving Phase II survival.
Drug Discov. Today 2012, 17, 419–424. [CrossRef]
93. Goller, A.H.; Kuhnke, L.; Montanari, F.; Bonin, A.; Schneckener, S.; Ter Laak, A.; Wichard, J.; Lobell, M.; Hillisch, A. Bayer’s
in silico ADMET platform: A journey of machine learning over the past two decades. Drug Discov. Today 2020, 25, 1702–1709.
[CrossRef]
94. Bhhatarai, B.; Walters, W.P.; Hop, C.E.C.A.; Lanza, G.; Ekins, S. Opportunities and challenges using artificial intelligence in
ADME/Tox. Nat. Mater. 2019, 18, 418–422. [CrossRef]
95. Vatansever, S.; Schlessinger, A.; Wacker, D.; Kaniskan, H.U.; Jin, J.; Zhou, M.M.; Zhang, B. Artificial intelligence and machine
learning-aided drug discovery in central nervous system diseases: State-of-the-arts and future directions. Med. Res. Rev. 2021, 41,
1427–1473. [CrossRef]
96. Faramarzi, S.; Kim, M.T.; Volpe, D.A.; Cross, K.P.; Chakravarti, S.; Stavitskaya, L. Development of QSAR models to predict
blood-brain barrier permeability. Front. Pharmacol. 2022, 13, 4486. [CrossRef]
97. Gao, Z.; Chen, Y.; Cai, X.; Xu, R. Predict drug permeability to blood-brain-barrier from clinical phenotypes: Drug side effects and
drug indications. Bioinformatics 2017, 33, 901–908. [CrossRef]
98. Miao, R.; Xia, L.Y.; Chen, H.H.; Huang, H.H.; Liang, Y. Improved Classification of Blood-Brain-Barrier Drugs Using Deep
Learning. Sci. Rep. 2019, 9, 8802. [CrossRef]
99. Lingineni, K.; Belekar, V.; Tangadpalliwar, S.R.; Garg, P. The role of multidrug resistance protein (MRP-1) as an active efflux
transporter on blood-brain barrier (BBB) permeability. Mol. Divers. 2017, 21, 355–365. [CrossRef]
100. Garg, P.; Dhakne, R.; Belekar, V. Role of breast cancer resistance protein (BCRP) as active efflux transporter on blood-brain barrier
(BBB) permeability. Mol. Divers. 2015, 19, 163–172. [CrossRef]
101. Tong, X.; Wang, D.; Ding, X.; Tan, X.; Ren, Q.; Chen, G.; Rong, Y.; Xu, T.; Huang, J.; Jiang, H.; et al. Blood-brain barrier penetration
prediction enhanced by uncertainty estimation. J. Cheminform. 2022, 14, 44. [CrossRef]
102. Tang, Q.; Nie, F.; Zhao, Q.; Chen, W. A merged molecular representation deep learning method for blood-brain barrier permeability
prediction. Brief Bioinform. 2022, 23, bbac357. [CrossRef]
103. Ma, J.; Sheridan, R.P.; Liaw, A.; Dahl, G.E.; Svetnik, V. Deep neural nets as a method for quantitative structure–activity relationships.
J. Chem. Inf. Model. 2015, 55, 263–274. [CrossRef] [PubMed]
104. Raina, R.; Madhavan, A.; Ng, A.Y. Large-scale deep unsupervised learning using graphics processors. In Proceedings of the 26th
Annual International Conference on Machine Learning, Montreal, QC, Canada, 14–18 June 2009; pp. 873–880.
105. LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [CrossRef] [PubMed]
106. Gaulton, A.; Bellis, L.J.; Bento, A.P.; Chambers, J.; Davies, M.; Hersey, A.; Light, Y.; McGlinchey, S.; Michalovich, D.; Al-Lazikani, B.
ChEMBL: A large-scale bioactivity database for drug discovery. Nucleic Acids Res. 2012, 40, D1100–D1107. [CrossRef] [PubMed]
107. Wang, Y.; Xiao, J.; Suzek, T.O.; Zhang, J.; Wang, J.; Bryant, S.H. PubChem: A public information system for analyzing bioactivities
of small molecules. Nucleic Acids Res. 2009, 37, W623–W633. [CrossRef]
108. Keshavarzi Arshadi, A.; Salem, M.; Firouzbakht, A.; Yuan, J.S. MolData, a molecular benchmark for disease and target based
machine learning. J. Cheminform. 2022, 14, 10. [CrossRef]
109. Smalley, E. AI-powered drug discovery captures pharma interest. Nat. Biotechnol. 2017, 35, 604–606. [CrossRef]
110. Schmidhuber, J. Deep learning in neural networks: An overview. Neural Netw. 2015, 61, 85–117. [CrossRef]
111. Cadieu, C.F.; Hong, H.; Yamins, D.L.; Pinto, N.; Ardila, D.; Solomon, E.A.; Majaj, N.J.; DiCarlo, J.J. Deep neural networks rival the
representation of primate IT cortex for core visual object recognition. PLoS Comput. Biol. 2014, 10, e1003963. [CrossRef]
112. Le Cun, Y.; Boser, B.; Denker, J.; Henderson, D.; Hubbard, W.; Jackel, L. Handwritten Digit Recognition Width a Backpropagation
Network. Advances in Neural Information Systems; Morgan Kaufman: San Mateo, CA, USA, 1990; Volume 2.
113. Grebner, C.; Matter, H.; Kofink, D.; Wenzel, J.; Schmidt, F.; Hessler, G. Application of Deep Neural Network Models in Drug
Discovery Programs. ChemMedChem 2021, 16, 3772–3786. [CrossRef]
Molecules 2023, 28, 1324 21 of 22

114. Francoeur, P.G.; Masuda, T.; Sunseri, J.; Jia, A.; Iovanisci, R.B.; Snyder, I.; Koes, D.R. Three-Dimensional Convolutional Neural
Networks and a Cross-Docked Data Set for Structure-Based Drug Design. J. Chem. Inf. Model. 2020, 60, 4200–4215. [CrossRef]
115. Lim, J.; Ryu, S.; Park, K.; Choe, Y.J.; Ham, J.; Kim, W.Y. Predicting Drug-Target Interaction Using a Novel Graph Neural Network
with 3D Structure-Embedded Graph Representation. J. Chem. Inf. Model. 2019, 59, 3981–3988. [CrossRef]
116. Nguyen, T.H.; Tran, P.T.; Pham, N.Q.A.; Hoang, V.H.; Hiep, D.M.; Ngo, S.T. Identifying Possible AChE Inhibitors from Drug-like
Molecules via Machine Learning and Experimental Studies. ACS Omega 2022, 7, 20673–20682. [CrossRef]
117. Rumelhart, D.E.; Hinton, G.E.; Williams, R.J. Learning representations by back-propagating errors. Nature 1986, 323, 533–536.
[CrossRef]
118. Hochreiter, S.; Schmidhuber, J. Long short-term memory. Neural Comput. 1997, 9, 1735–1780. [CrossRef]
119. Yasonik, J. Multiobjective de novo drug design with recurrent neural networks and nondominated sorting. J. Cheminform. 2020,
12, 14. [CrossRef]
120. Sanchez-Lengeling, B.; Aspuru-Guzik, A. Inverse molecular design using machine learning: Generative models for matter
engineering. Science 2018, 361, 360–365. [CrossRef]
121. Karimi, M.; Wu, D.; Wang, Z.; Shen, Y. DeepAffinity: Interpretable deep learning of compound-protein affinity through unified
recurrent and convolutional neural networks. Bioinformatics 2019, 35, 3329–3338. [CrossRef]
122. Caruana, R. Multitask Learning. In Learning to Learn; Thrun, S., Pratt, L., Eds.; Springer: Boston, MA, USA, 1998; pp. 95–133.
123. Simões, R.S.; Maltarollo, V.G.; Oliveira, P.R.; Honorio, K.M. Transfer and Multi-task Learning in QSAR Modeling: Advances and
Challenges. Front. Pharmacol. 2018, 9, 74. [CrossRef]
124. Zhao, Z.; Qin, J.; Gou, Z.; Zhang, Y.; Yang, Y. Multi-task learning models for predicting active compounds. J. Biomed. Inform. 2020,
108, 103484. [CrossRef]
125. Lee, K.; Kim, D. In-Silico Molecular Binding Prediction for Human Drug Targets Using Deep Neural Multi-Task Learning. Genes
2019, 10, 906. [CrossRef]
126. Wang, M.; Hsieh, C.Y.; Wang, J.; Wang, D.; Weng, G.; Shen, C.; Yao, X.; Bing, Z.; Li, H.; Cao, D.; et al. RELATION: A Deep
Generative Model for Structure-Based De Novo Drug Design. J. Med. Chem. 2022, 65, 9478–9492. [CrossRef] [PubMed]
127. Bung, N.; Krishnan, S.R.; Roy, A. An In Silico Explainable Multiparameter Optimization Approach for De Novo Drug Design
against Proteins from the Central Nervous System. J. Chem. Inf. Model. 2022, 62, 2685–2695. [CrossRef] [PubMed]
128. Maes, M. Precision Nomothetic Medicine in Depression Research: A New Depression Model, and New Endophenotype Classes
and Pathway Phenotypes, and A Digital Self. J. Pers. Med. 2022, 12, 403. [CrossRef] [PubMed]
129. Maes, M.; Anderson, G. False Dogmas in Schizophrenia Research: Toward the Reification of Pathway Phenotypes and Pathway
Classes. Front. Psychiatr. 2021, 12, 663985. [CrossRef]
130. Anastasio, T.J. Predicting the Potency of Anti-Alzheimer’s Drug Combinations Using Machine Learning. Processes 2021, 9, 264.
[CrossRef]
131. Tuladhar, A.; Moore, J.A.; Ismail, Z.; Forkert, N.D. Modeling Neurodegeneration in silico with Deep Learning. Front. Neuroinform.
2021, 15, 56. [CrossRef]
132. Gauthier, S.; Webster, C.; Servaes, S.; Morais, J.; Rosa-Neto, P. World Alzheimer Report 2022; Alzheimer’s Disease International:
London, UK, 2022.
133. Iida, M.; Iwata, M.; Yamanishi, Y. Network-based characterization of disease–disease relationships in terms of drugs and
therapeutic targets. Bioinformatics 2020, 36, i516–i524. [CrossRef]
134. Ivanova, L.; Karelson, M.; Dobchev, D.A. Multitarget Approach to Drug Candidates against Alzheimer’s Disease Related to
AChE, SERT, BACE1 and GSK3beta Protein Targets. Molecules 2020, 25, 1846. [CrossRef]
135. Sterling, T.; Irwin, J.J. ZINC 15–ligand discovery for everyone. J. Chem. Inf. Model. 2015, 55, 2324–2337. [CrossRef]
136. Oddsson, S.; Kowal, N.M.; Ahring, P.K.; Olafsdottir, E.S.; Balle, T. Structure-Based Discovery of Dual-Target Hits for Acetyl-
cholinesterase and the alpha7 Nicotinic Acetylcholine Receptors: In Silico Studies and In Vitro Confirmation. Molecules 2020,
25, 2872. [CrossRef]
137. Liu, J.; Peng, D.; Li, J.; Dai, Z.; Zou, X.; Li, Z. Identification of Potential Parkinson’s Disease Drugs Based on Multi-Source Data
Fusion and Convolutional Neural Network. Molecules 2022, 27, 4780. [CrossRef]
138. Khan, A.; Chandra Kaushik, A.; Ali, S.S.; Ahmad, N.; Wei, D.Q. Deep-learning-based target screening and similarity search for
the predicted inhibitors of the pathways in Parkinson’s disease. RSC Adv. 2019, 9, 10326–10339. [CrossRef]
139. Peng, Y.; Dong, H.; Welsh, W.J. Comprehensive 3D-QSAR Model Predicts Binding Affinity of Structurally Diverse Sigma 1
Receptor Ligands. J. Chem. Inf. Model. 2019, 59, 486–497. [CrossRef]
140. Kang, K.M.; Lee, I.; Nam, H.; Kim, Y.C. AI-based prediction of new binding site and virtual screening for the discovery of novel
P2X3 receptor antagonists. Eur. J. Med. Chem. 2022, 240, 114556. [CrossRef]
141. Yu, Y.; Dong, H.; Peng, Y.; Welsh, W.J. QSAR-Based Computational Approaches to Accelerate the Discovery of Sigma-2 Receptor
(S2R) Ligands as Therapeutic Drugs. Molecules 2021, 26, 5270. [CrossRef]
142. Ozhathil, L.C.; Delalande, C.; Bianchi, B.; Nemeth, G.; Kappel, S.; Thomet, U.; Ross-Kaschitza, D.; Simonin, C.; Rubin, M.;
Gertsch, J.; et al. Identification of potent and selective small molecule inhibitors of the cation channel TRPM4. Br. J. Pharmacol.
2018, 175, 2504–2519. [CrossRef]
Molecules 2023, 28, 1324 22 of 22

143. Coaviche-Yoval, A.; Trujillo-Ferrara, J.G.; Soriano-Ursua, M.A.; Andrade-Jorge, E.; Sanchez-Labastida, L.A.; Luna, H.; Tovar-
Miranda, R. In silico and in vivo neuropharmacological evaluation of two gamma-amino acid isomers derived from 2,3-
disubstituted benzofurans, as ligands of GluN1-GluN2A NMDA receptor. Amino Acids 2022, 54, 215–228. [CrossRef]
144. Pandey, P.; Roy, K.K.; Liu, H.; Ma, G.; Pettaway, S.; Alsharif, W.F.; Gadepalli, R.S.; Rimoldi, J.M.; McCurdy, C.R.; Cutler, S.J.; et al.
Structure-Based Identification of Potent Natural Product Chemotypes as Cannabinoid Receptor 1 Inverse Agonists. Molecules
2018, 23, 2630. [CrossRef]
145. Tang, Y.; Feng, B.; Wang, Y.; Sun, H.; You, Y.; Yu, J.; Chen, B.; Xu, C.; Ruan, Y.; Cui, S.; et al. Structure-based discovery of CZL80, a
caspase-1 inhibitor with therapeutic potential for febrile seizures and later enhanced epileptogenic susceptibility. Br. J. Pharmacol.
2020, 177, 3519–3534. [CrossRef]
146. Qiang, S.J.; Shi, Y.Q.; Wu, T.Y.; Wang, J.Q.; Chen, X.L.; Su, J.; Chen, X.P.; Li, J.Z.; Chen, Z.S. The Discovery of Novel PGK1
Activators as Apoptotic Inhibiting and Neuroprotective Agents. Front. Pharm. 2022, 13, 877706. [CrossRef]
147. Kampen, S.; Rodriguez, D.; Jorgensen, M.; Kruszyk-Kujawa, M.; Huang, X.; Collins, M., Jr.; Boyle, N.; Maurel, D.; Rudling, A.;
Lebon, G.; et al. Structure-Based Discovery of Negative Allosteric Modulators of the Metabotropic Glutamate Receptor 5. ACS
Chem. Biol. 2022, 17, 2744–2752. [CrossRef] [PubMed]
148. Kortagere, S.; Mortensen, O.V.; Xia, J.; Lester, W.; Fang, Y.; Srikanth, Y.; Salvino, J.M.; Fontana, A.C.K. Identification of Novel
Allosteric Modulators of Glutamate Transporter EAAT2. ACS Chem. Neurosci. 2018, 9, 522–534. [CrossRef] [PubMed]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.