Virtual Screening of C. Sativa Constituents for the Identification of Selective Ligands for Cannabinoid Receptor 2 - MDPI
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
International Journal of
Molecular Sciences
Article
Virtual Screening of C. Sativa Constituents for the
Identification of Selective Ligands for Cannabinoid
Receptor 2
Mikołaj Mizera 1 , Dorota Latek 2 and Judyta Cielecka-Piontek 1, *
1 Department of Pharmacognosy, Poznan University of Medical Sciences, 60-781 Poznań, Poland;
mikolajmizera@gmail.com
2 Faculty of Chemistry, University of Warsaw, 02-093 Warsaw, Poland; dlatek@chem.uw.edu.pl
* Correspondence: jpiontek@ump.edu.pl; Tel.: +48-854-67-10
Received: 28 April 2020; Accepted: 24 July 2020; Published: 26 July 2020
Abstract: The selective targeting of the cannabinoid receptor 2 (CB2) is crucial for the development of
peripheral system-acting cannabinoid analgesics. This work aimed at computer-assisted identification
of prospective CB2-selective compounds among the constituents of Cannabis Sativa. The molecular
structures and corresponding binding affinities to CB1 and CB2 receptors were collected from ChEMBL.
The molecular structures of Cannabis Sativa constituents were collected from a phytochemical database.
The collected records were curated and applied for the development of quantitative structure-activity
relationship (QSAR) models with a machine learning approach. The validated models predicted the
affinities of Cannabis Sativa constituents. Four structures of CB2 were acquired from the Protein Data
Bank (PDB) and the discriminatory ability of CB2-selective ligands and two sets of decoys were tested.
We succeeded in developing the QSAR model by achieving Q2 5-CV > 0.62. The QSAR models helped
to identify three prospective CB2-selective molecules that are dissimilar to already tested compounds.
In a complementary structure-based virtual screening study that used available PDB structures of
CB2, the agonist-bound, Cryogenic Electron Microscopy structure of CB2 showed the best statistical
performance in discriminating between CB2-active and non-active ligands. The same structure also
performed best in discriminating between CB2-selective ligands from non-selective ligands.
Keywords: QSAR; endocannabinoid system; Cannabis Sativa
1. Introduction
The medical effect of Cannabis sp. bioactive ingredients is the subject of extensive research [1–3].
It has been discovered that Cannabis sp. contains over 500 various compounds, with the cannabinoid
group itself having about 110 molecules [4]. Phytocannabinoids present in Cannabis sp. have slight
differences in their chemical structures (types: cannabidiol, cannabichromene, cannabitriol,
cannabicycol, cannabinodiol, cannabinol, cannabielsoin, cannabigerol, ∆9 -tetrahydrocannabinol,
∆8 -tetrahydrocannabinate) and are not selective. The most well-known phytocannabinoids are THC
(∆8 -tetrahydrocannabinol) and CBD (cannabidiol). Research to date reports that both THC and CBD
have an affinity for various types of endocannabinoid system receptors [5–7]. For example, cannabidiol
is a non-competitive CB1 antagonist, CB2 inverse agonist, GPR55 and GPR18 antagonist, Peroxisome
proliferator-activated receptor (PPAR-γ) agonist, α1, α3 glycine agonist, TRPM8 antagonist, and an
agonist and antagonist of receptors of various types of serotonin 1A receptors (5-HT1A) [8–11].
The pharmacological effects of the interaction of Active pharmaceutical ingredients (APIs) with CB1
and CB2 receptors are the most investigated. The differences in the density of CB1 and CB2 receptor
distribution also determine the activities within the central nervous system [12]. The pharmacological
Int. J. Mol. Sci. 2020, 21, 5308; doi:10.3390/ijms21155308 www.mdpi.com/journal/ijms
Int. J. Mol. Sci. 2020, 21, 5308 2 of 14
effect of cannabinoids results from modification of the signaling pathway of two G protein-coupled
receptors (GPCRs): cannabinoid receptors 1 and 2 (CB1 and CB2). Activation of the CB1 receptor,
which is expressed mainly in the central nervous system, is responsible for their psychotropic
action, while CB2 receptors are found mainly in the immune system [13]. This distinct distribution
of CB2 receptors in human tissues suggests that selective cannabinoids are promising APIs with
anti-inflammatory, analgesic, and anti-neuroinflammatory actions [14]. The discovery of new selective
cannabinoids can be streamlined with the application of in silico methods such as structure-based
virtual screening (VS) [15–17] or ligand-based (quantitative structure-activity relationship (QSAR))
VS [18].
Machine learning-based QSAR models can be successfully applied in pharmaceutical research
related to drug discovery [19], drug formulation [20], and pharmaceutical analysis [21,22].
The applicability of QSAR modeling to predict the selectivity of CB receptors has also been reported in
the literature. The CB2-selective activity of 29 benzimidazole and benzothiophene derivatives was
investigated with comparative molecular field analysis (CoMFA) and comparative molecular similarity
indices analysis (CoMSIA) 3D-QSAR models [23] and showed an external predictive performance of R2
> 0.9. In the recent 4D-QSAR study involving molecular dynamics (MD), the modeling was performed
on 29 structurally similar CB2 receptor inverse agonists [24]. The 4D-QSAR approach based on partial
least squares (PLS) and multiple linear regression (MLR) resulted in Q2 = 0.719 and Q2 = 0.761 for the
PLS and MLR models, respectively. The selectivity of 29 arylpyrazole derivatives was investigated
with the application of 3D-QSAR/CoMFA analyses [25]. The QSAR model helped to identify the causes
of CB1 selectivity by producing counter maps of affinity for both CB receptor subtypes. Nevertheless,
the narrow applicability of these models due to similar scaffolds used as training examples may
be a significant limitation when they are used in the prediction of novel chemotypes. In contrast,
Floresta et al. conducted a 3D-QSAR study on a diverse dataset containing 312 molecules with reported
experimental CB1 affinities and 187 molecules with reported CB2 affinities [26]. These models showed
a predictive performance of Q2 = 0.62 and Q2 = 0.72 for the CB1 and CB2 QSAR models, respectively.
The diversity of molecular structures in the training set also allowed Floresta et al. to perform VS
for novel chemical scaffolds. The idea of the QSAR application for the identification of bioactive
compounds in plant extract was also used by Labib et al. [27]. This VS study aimed to predict the
activities of CB1 and opioid receptors among compounds isolated from Pinus roxburghii bark extract.
The model of Labib et al. explained the synergistic anti-inflammatory action of Pinus roxburghii and
provided information on the activity of several bioactive molecules identified in the extract.
Our study aimed to predict CB2-selectivity for molecules identified in Cannabis Sativa by using a
validated QSAR model. This goal was achieved by the execution of several steps: the collection of
diverse ligands’ structures with experimental data describing CB1 and CB2 affinity; data curation;
conducting QSAR study for CB1 and CB2; and the prediction of the CB1 and CB2 affinity of
phytochemicals in the Cannabis Sativa. So far, the lack of crystal structures in cannabinoid receptors has
been a major obstacle in searching for new, selective cannabinoids. However, there have been attempts
to predict binding modes of well-known actives of CB1 and CB2 using homology models of these
receptors and molecular dynamics [17,28,29]. Recent advances in structural studies of cannabinoid
receptors 6KPC [30], 6KPF [30], 6PT0 [31], and 5ZTY [32] provided new data that supplemented drug
discovery studies [30]. In this study, we tested the applicability of available experimental structures of
CB2, solved in both active and inactive conformations, in structure-based VS to search for novel or
more CB2-selective ligands.
2. Results
2.1. Study Design
Data on the experimentally evaluated affinity of diverse compounds to CB1 and CB2 receptors
and data on structures identified in Cannabis Sativa with unknown CB1/CB2 affinity were collected
Int. J. Mol. Sci. 2020, 21, 5308 3 of 14
from publicly accessible online resources. Next, the data with known affinity values were curated and
subsequently used for the development and validation of CB1 and CB2 QSAR models. The validated
models were used for the prediction of the CB1 and CB2 affinity of Cannabis Sativa ingredients and
the prediction of prospective CB2-selective Cannabis Sativa ingredients. The dataset collected for the
QSAR study was used to evaluate the statistical characteristics of the discriminatory ability of CB1 and
CB2Int.crystal
J. Mol. Sci.structures inPEER
2020, 21, x FOR theREVIEW
docking study. 3 of 14
2.2.validated models were used for the prediction of the CB1 and CB2 affinity of Cannabis Sativa
Data Curation
ingredients and the prediction of prospective CB2-selective Cannabis Sativa ingredients. The dataset
Datasets
collected of ligands
for the withwas
QSAR study experimental affinities
used to evaluate for CB1
the statistical and CB2 were
characteristics of thejoined with corresponding
discriminatory
ability of CB1
records andmetadata
assay CB2 crystal from
structures in the docking
ChEMBL. study.
The resulting dataset was composed of six columns
containing: the structure of the molecule in simplified molecular-input line-entry system (SMILES)
2.2. Data Curation
format, a standard type of reported value (i.e., Ki, IC50, etc.), reported value, the mathematical relation
Datasets ofvalue
of the reported ligands towith
theexperimental
experimentally affinities for CB1 and
measured CB2 were
value, the ID joined with corresponding
number of the source document
records of assay metadata from ChEMBL. The resulting dataset was composed of six columns
with reported study, and the confidence score. The initial dataset described above was subjected
containing: the structure of the molecule in simplified molecular-input line-entry system (SMILES)
to the removal
format, of potentially
a standard unreliable
type of reported value data (Figure
(i.e., Ki, IC50, 1).
etc.),The initialvalue,
reported datasetthe included
mathematical 14,126 records
forrelation
CB1 and 13,506 records for CB2 (Figure 1.1). Records without reported
of the reported value to the experimentally measured value, the ID number of the source SMILEs, which included
metal complexes
document and polymers
with reported study, andwere excludedscore.
the confidence fromThe theinitial
dataset (Figure
dataset 1.2).above
described Forwas the remaining
subjected
records, thetorespective
the removalmolecular
of potentially unreliablewere
structures data (Figure 1). The initial
standardized anddataset included 14,126
2D coordinates were generated.
Werecords
assumed for CB1
that and
both13,506 records which
the assays, for CB2were(Figure 1.1). Records
carried out onwithout
a targetreported
proteinSMILEs, which
or a homologous protein
included metal complexes and polymers were excluded from the dataset (Figure 1.2). For the
were reliable. Records meeting these criteria were annotated in ChEMBL with a confidence score
remaining records, the respective molecular structures were standardized and 2D coordinates were
of 9generated.
and 8, respectively
We assumed (Figurethat both1.3). Only activities
the assays, which were measured
carried out in on
large and consistent
a target protein or a assays were
picked from the dataset (Figure 1.4).That is, an assay was considered
homologous protein were reliable. Records meeting these criteria were annotated in ChEMBL with as suitable for aselection if it
contained
confidence a score
consistent
of 9 andgroup of 10 or
8, respectively more1.3).
(Figure molecules, all with
Only activities the affinity
measured measured.
in large and consistent These large
assays
assays wereidentified
were picked from the dataset
based on the (Figure
document 1.4).That is, an assay
ID reported was considered
in ChEMBL. In theas next
suitable
step,forrecords with
selection if it contained a consistent group of 10 or more molecules, all with the affinity measured.
a reported affinity in any standard value type related to Ki were kept (Figure 1.5) and subsequently
These large assays were identified based on the document ID reported in ChEMBL. In the next step,
converted to pKi. Duplicates analysis (Figure 1.6) was conducted by comparing records InChIKeys.
records with a reported affinity in any standard value type related to Ki were kept (Figure 1.5) and
Duplicates
subsequentlywere mergedtoifpKi.
converted theDuplicates
standardanalysis
deviation of duplicate
(Figure measurements
1.6) was conducted wasrecords
by comparing lower than 10% of
theInChIKeys.
entire range of measurements.
Duplicates were merged if Asthea standard
result, only one record
deviation was measurements
of duplicate kept, with thewas pKi value averaged.
lower
than 10%
Mordred of the entireand
descriptors range of measurements.
Morgan fingerprints As were
a result, only one record
computed for eachwascompound
kept, with the
in pKi
the dataset and
thenvalue averaged. Mordred
concatenated to createdescriptors
a featureandvector.
MorganRecords
fingerprintswithwere computedfeature
duplicate for eachvectors
compound were removed
in the dataset and then concatenated to create a feature vector. Records with duplicate feature vectors
(Figure 1.7). These included, e.g., compounds differing only by chiral hydrogen atoms and those that
were removed (Figure 1.7). These included, e.g., compounds differing only by chiral hydrogen atoms
could
and not
thosebe differentiated
that by the used
could not be differentiated by descriptors and fingerprints.
the used descriptors and fingerprints.
Figure 1. Data records that passed through the subsequent data curation steps: (1) data acquisition,
Figure 1. Data records that passed through the subsequent data curation steps: (1) data acquisition,
(2) removing records with no SMILES included, (3) removing records with Confidence Score < 8, (4)
(2)keeping
removing
only records with large
records from no SMILES included,
assays, (5) (3) removing
keeping only records
records with with
included Confidence
Ki values, (6) Score
Int. J. Mol. Sci. 2020, 21, 5308 4 of 14
2.3.
Int. Model
J. Mol. Sci. Description
2020, 21, x FOR PEER REVIEW 4 of 14
The independent CB1 and CB2 QSAR models were created according to the architecture presented
The independent CB1 and CB2 QSAR models were created according to the architecture
in Figure 2. The input feature vectors for the machine learning algorithm were molecular descriptors
presented in Figure 2. The input feature vectors for the machine learning algorithm were molecular
and fingerprints of compounds from the training set (Figure 2.1). The feature vectors along with
descriptors and fingerprints of compounds from the training set (Figure 2.1). The feature vectors
experimental CB1 and CB2 pKi values were used as a training set for the CB1 and CB2 QSAR models
along with experimental CB1 and CB2 pKi values were used as a training set for the CB1 and CB2
(Figure 2.2).
QSAR models (Figure 2.2).
Figure
Figure2.2.Machine
Machinelearning-based
learning-basedquantitative
quantitativestructure-activity
structure-activityrelationship
relationship(QSAR)
(QSAR)model.
model.
The
Thevalue
valuepredicted
predictedinineach
eachleaf
leafofofdecision
decisiontrees
treesininaagradient
gradientboosting
boosting(GB)(GB)ensemble
ensemblewas wasused
used
for the embedding of a descriptor space (Figure 2.3). The embedded representation
for the embedding of a descriptor space (Figure 2.3). The embedded representation of the descriptorof the descriptor
space
spacewas
wasused
usedtotocreate
createaaseparate
separatek-nearest
k-nearestneighbor
neighbor(kNN)(kNN)model
modelfor foreach
eachreceptor
receptor(Figure
(Figure2.4).
2.4).
The
ThekNN
kNNalgorithm
algorithmwaswasmodified
modifiedtotopredict
predictan anaverage
averagepKi pKivalue
valueofofaacompound
compoundonly onlyififall
allnearest
nearest
neighbors
neighborsofofthe query
the querymolecular
molecular structure
structurewere within
were withina given distance.
a given This This
distance. maximal distance
maximal was
distance
used as anasapplicability
was used domain
an applicability domainthreshold related
threshold to the
related confidence
to the confidence of of
prediction
prediction(Figure
(Figure2.5).
2.5).
Different
Different thresholds weretested
thresholds were testedtotoassess
assess their
their influence
influence on statistical
on the the statistical characteristics
characteristics of kNN ofmodels.
kNN
models. We selected
We selected 20 thresholds
20 thresholds as percentiles
as percentiles 5 to a100
5 to 100 with with
step a step
equal to 5%equal to distribution
of the 5% of the distribution
of maximal
ofEuclidean
maximal distances
Euclideanwithin
distances
kNN within
clusterskNN clustersfor
obtained obtained for the
the training set.training set.
2.4.Model
2.4. ModelValidation
Validation
InInFigure
Figure 3,
3,we
wepresent
present thethe
dependence of cross-validated
dependence Q2 on Q
of cross-validated the applicability
2 on domain threshold.
the applicability domain
The predictive
threshold. The performance showed consistent
predictive performance showedbehavior, declining
consistent with an
behavior, increasedwith
declining threshold. The best
an increased
statistical The
threshold. characteristics Q >characteristics
2
best statistical 0.8 for the CB1Qand
2 CB2for
> 0.8 models
the CB1 was observed
and for thewas
CB2 models lowest thresholds
observed for
(QSAR
0.6, as best practices
suggested [33]. of QSAR best practices [33].
in studies
An autocorrelation plot (Figure 4) shows good agreement between out-of-sample predictions and
experimental values for the majority of compounds, and importantly, for the CB2 selectivity ranges of
pKi (pKi < 6 for CB1 and pKi > 6.5 for CB2). The majority of data points presented in Figure 4 are
distributed between 6 and 8. Because machine learning techniques used in this study are interpolating
techniques, we expected our model to overestimate pKi predictions at lower values and underestimate
predictions at large ones.
Figure 3. Q2—applicability domain threshold dependence for CB1 and CB2 k-nearest neighbor (kNN)
models. Each point on the curves represents a Q2 calculated for external predictions for a cumulative
fraction of molecular structures in the training dataset within a step of 5%.
In Figure 3, we present the dependence of cross-validated Q2 on the applicability domain
threshold. The predictive performance showed consistent behavior, declining with an increased
threshold. The best statistical characteristics Q2 > 0.8 for the CB1 and CB2 models was observed for
the lowest thresholds ( 0.6, as suggested in studies of QSAR best practices [33].
Int. J. Mol. Sci. 2020, 21, x FOR PEER REVIEW 5 of 14
Int. J. Mol. Sci. 2020, 21, x FOR PEER REVIEW 5 of 14
An autocorrelation plot (Figure 4) shows good agreement between out-of-sample predictions
and An
experimental valuesplot
autocorrelation for (Figure
the majority of compounds,
4) shows good agreement and importantly, for the CB2predictions
between out-of-sample selectivity
ranges
and of pKi (pKi values
experimental < 6 for CB1
for theandmajority
pKi > 6.5of forcompounds,
CB2). The majority of data points
and importantly, for presented in Figure
the CB2 selectivity
4 are distributed between
ranges of pKi (pKi < 6 for CB1 6 and 8. Because machine learning techniques used in this
pKi > 6.5 for CB2). The majority of data points presented in Figure study are
4interpolating
areFigure
Figure techniques,
distributed2 between we6 expected
Q2—applicability
3. Q
3. —applicability and
domain
domain our model
8. threshold
Because
threshold to overestimate
machine
dependence
dependence learning
forCB1
for CB1andpKi
and predictions
techniques
CB2
CB2 usedat
k-nearest
k-nearest lower
in values
study and
this (kNN)
neighbor
neighbor (kNN) are
underestimate
interpolating
models. predictions
techniques,
Each point on weat
the large
expected
curves ones. our
representsmodel
a Q2to overestimate
calculated for pKi
externalpredictions
predictions at
for lower
a values
cumulative and
models. Each point on the curves represents a Q calculated for external predictions for a cumulative
2
underestimate
fraction of
fraction predictions
of molecular
molecular at largein
structures
structures inones.
the
the training
training dataset
dataset within
within aa step
step of
of 5%.
5%.
Figure 4. Autocorrelation of experimental and predicted pKi values for the CB1 model (blue) and the
CB2 model
Figure
Figure (orange).
Autocorrelation
4. Autocorrelation
4. of experimental
of experimental and
and predicted
predicted pKi
pKi values
values for
for the
the CB1
CB1 model
model (blue)
(blue) and
and the
the
CB2 model
CB2 model (orange).
(orange).
2.5. Virtual Screening of Cannabis Sativa Phytochemicals
2.5. Virtual
2.5. Virtual Screening
Screening of
of Cannabis
Cannabis Sativa
Sativa Phytochemicals
Phytochemicals
Sixty-eight Cannabis Sativa phytochemicals with a molecular weight between 250D and 500D
Sixty-eight
wereSixty-eight
subjected to Cannabis
VS using
Cannabis Sativa phytochemicals
our validated
Sativa CB1 and
phytochemicals with
CB2aaQSAR
with molecular weight
models.
molecular betweenthe
On average,
weight between 250D
250D and 500D
500D
selectivity
and of
were
were subjected
compounds
subjected to
from VS using
Cannabis
to VS our validated
Sativa
using our was less
validated CB1 and
CB1than CB2 QSAR
for the
and CB2 QSAR models.
ChEMBL-derivedOn average,
training
models. On average, the selectivity
the set. of
Still, we
selectivity of
compounds
observed
compounds fromCannabis
from
a relatively highSativa
Cannabis waswaslessless
autocorrelation
Sativa than for the
ofthan
the forChEMBL-derived
predicted CB1 and CB2
the ChEMBL-derivedtraining set. of
activity Still,
training we observed
Cannabis
set. Sativa
Still, we
a relatively
observed high
phytochemicals autocorrelation
(see Figure
a relatively of the predicted
5) in comparison
high autocorrelation of toCB1
the
the and CB2
ChEMBLCB1
predicted activity of
training Cannabis
set. The
and CB2 Sativa
average
activity phytochemicals
absoluteSativa
of Cannabis dpKi
(see Figure
between CB25) and
phytochemicals in (see
comparison
CB1 toin
was 0.69
Figure 5) the
and ChEMBL training
1.26 for molecules
comparison set. The average
in Cannabis
to the ChEMBL absolute
Sativa
training and
set. The dpKi
molecules
averagebetween
in the CB2
absolute and
training
dpKi
CB1respectively.
set, wasCB2
between 0.69 and
andCB1
1.26was
for molecules
0.69 and 1.26in Cannabis Sativainand
for molecules molecules
Cannabis Sativainand
the molecules
training set,in respectively.
the training
set, respectively.
Figure 5. The correlation
correlationbetween
betweenaffinity
affinityagainst
againstCB1
CB1and CB2
and forfor
CB2 compounds in Cannabis
compounds Sativa
in Cannabis and
sativa
compounds
Figure reported
and compounds in ChEMBL.
reported between
5. The correlation in ChEMBL.
affinity against CB1 and CB2 for compounds in Cannabis sativa
and compounds reported in ChEMBL.
Despite relatively low average predicted selectivity of compounds in Cannabis Sativa, structures
C1–C3 showing
Despite dpKi >low
relatively 1 and a Tanimoto
average predictedcoefficient
selectivity(TC) < 0.5 were in
of compounds identified
Cannabis(Table
Sativa,1).structures
Low TC
values indicated
C1–C3 a significant
showing dpKi > 1 andstructural
a Tanimotodifference in the
coefficient identified
(TC) molecular
< 0.5 were structures
identified in Cannabis
(Table 1). Low TC
Sativa compared
values indicated atosignificant
the respective most similar
structural molecular
difference structuresmolecular
in the identified in the training set. in Cannabis
structures
Int. J. Mol. Sci. 2020, 21, 5308 6 of 14
Despite relatively low average predicted selectivity of compounds in Cannabis Sativa, structures
C1–C3 showing dpKi > 1 and a Tanimoto coefficient (TC) < 0.5 were identified (Table 1). Low TC
Int. Int.
values J. Mol.
Sci. Sci.
indicated 2020, x21, x FOR
axxsignificant PEER REVIEW
structural 6 of6614
difference in the identified molecular structures of 14
Cannabis
6in
J.Int.
Mol. 2020, 21, FOR PEER REVIEW
Int. J.J.Int.
Int. Mol.
Mol.J.J.Mol.
Mol.
Sci.
Sci.
Sci.2020,
Sci.
2020,
2020,
2020,
21,
21,
21,
21,
FOR
FOR xxFOR
FOR
PEER
PEER
PEER
PEER REVIEW
REVIEW
REVIEW
REVIEW 66 of
of 14
14 of14
of 14
Sativa compared to the respective most similar molecular structures in the training set.
Table
Table 1. Compounds
1. Compounds fromfrom Cannabis
Cannabis Sativa
Sativa thatthat were
were predicted
predicted as CB2-selective
as CB2-selective andandareare
alsoalso dissimilar
dissimilar
Table
Table Table
Table 1.Compounds
1. Compounds
1. Compounds
1. Compounds from from
fromfrom Cannabis
Cannabis
Cannabis
Cannabis Sativa
SativaSativa
Sativa
that that
that that
were
were were
were predicted
predicted
predicted
predicted asCB2-selective
as CB2-selective
as CB2-selective
as CB2-selective and
and
and are
and are arealso
are
also
also also dissimilar
dissimilar
dissimilar
dissimilar
to
to the the ChEMBL
ChEMBL training
training set.
set. set.
to the
the
Table
to to
to1. the
the
ChEMBLChEMBL
ChEMBL
Compounds
ChEMBL training
training
training
training set. set.
from
set. Cannabis Sativa that were predicted as CB2-selective and are also dissimilar
Predicted
Predicted TaniTani
to the ChEMBL training
Structure
Structure set. Predicted
Predicted
Predicted
Predicted Similar
Similar Molecule
Molecule for for ChEMBL
ChEMBL Tani
TaniTani
Tani
Structure
Structure
Structure
Structure Value
Value Similar
Similar
Similar Molecule
Molecule
Molecule for for
for
ChEMBL
Similar Molecule for ChEMBL ChEMBL
ChEMBL moto
moto
Value
Value Value
Value moto
moto moto
moto
Structure Predicted Value Similar Molecule for ChEMBL CoefCoef
Tanimoto
CB1CB1 CB2CB2 CB1CB1 CB2CB2 Coef Coef
Coef
CB1CB1
CB1 CB1CB2 CB2
CB2CB2 Structure
Structure CB1
CB1 CB1 CB2
CB1 CB2 CB2Coef
CB2 ficie
ficie
Coefficient
pKipKi
CB1 pKi
CB2
pKi Structure
Structure
Structure
Structure pKi
CB1
pKi CB2
pKi pKi ficie
ficie ficie
ficie
pKipKi
pKi pKi
pKi pKi
pKi pKi Structure pKi pKi pKi
pKi pKi
pKi pKi nt nt
pKi pKi pKi pKi pKi nt nt ntnt
5.465.46 6.876.87
5.466.87 6.87 4.644.64
4.64 N/A
N/A 0.280.28
N/AN/A0.28 0.28
C1 C1
C1 5.465.46
5.46
5.46 6.87 6.87
6.87 4.64
4.64 4.64
4.64 N/A
N/A
N/A 0.28 0.28
0.28
C1C1C1
C1 (4-hydroxy-6-methoxyspiro[1,2-
(4-hydroxy-6-methoxyspiro[1,2-
(4-hydroxy-6-methoxyspiro[1,2-
(4-hydroxy-6-methoxyspiro[1,2-
(4-hydroxy-6-methoxyspiro[1,2-
(4-hydroxy-6-methoxyspiro
(4-hydroxy-6-methoxyspiro[1,2- CHEMBL1201151
CHEMBL1201151
dihydroindene-3,4’-cyclohexane]-1’-yl)
dihydroindene-3,4’-cyclohexane]-1’-yl) CHEMBL1201151
CHEMBL1201151
CHEMBL1201151
CHEMBL1201151
CHEMBL1201151
dihydroindene-3,4’-cyclohexane]-1’-yl)
[1,2-dihydroindene-3,4’-cyclohexane]
dihydroindene-3,4’-cyclohexane]-1’-yl)
dihydroindene-3,4’-cyclohexane]-1’-yl)
dihydroindene-3,4’-cyclohexane]-1’-yl)
acetate
acetate
-1’-yl) acetate
acetate
acetate
acetate
acetate
C2 C2 5.575.57
5.57 6.596.59
6.59
5.576.59 6.59 6.956.95 8.03
6.95 8.03 0.180.18
8.03
6.95 8.03 8.030.18 0.18
0.18
C2
C2 C2
C2 5.57
5.57 5.57 6.59 6.59 6.95
6.95 6.95 8.03 8.03 0.18 0.18
C2
2-[(2R,4aR,8R,8aR)-8-hydroxy-4a,8-
2-[(2R,4aR,8R,8aR)-8-hydroxy-4a,8-
2-[(2R,4aR,8R,8aR)-8-hydroxy-4a,
2-[(2R,4aR,8R,8aR)-8-hydroxy-4a,8-
2-[(2R,4aR,8R,8aR)-8-hydroxy-4a,8-
2-[(2R,4aR,8R,8aR)-8-hydroxy-4a,8-
2-[(2R,4aR,8R,8aR)-8-hydroxy-4a,8-
dimethyl-1,2,3,4,5,6,7,8a-
dimethyl-1,2,3,4,5,6,7,8a-
8-dimethyl-1,2,3,4,5,6,7, CHEMBL256753
CHEMBL256753
CHEMBL256753
dimethyl-1,2,3,4,5,6,7,8a-
dimethyl-1,2,3,4,5,6,7,8a-
dimethyl-1,2,3,4,5,6,7,8a-
dimethyl-1,2,3,4,5,6,7,8a- CHEMBL256753
CHEMBL256753
CHEMBL256753
CHEMBL256753
octahydronaphthalen-2-yl]prop-2-enoic
octahydronaphthalen-2-yl]prop-2-enoic
8a-octahydronaphthalen-2-yl]
octahydronaphthalen-2-yl]prop-2-enoic
octahydronaphthalen-2-yl]prop-2-enoic
octahydronaphthalen-2-yl]prop-2-enoic
octahydronaphthalen-2-yl]prop-2-enoic
acidacid
prop-2-enoic acid
acid
acidacid
acid
6.416.41 7.29
6.417.29
7.29
7.29
7.29 7.337.33 6.04
7.33
7.33 6.04
6.04 0.410.41
6.046.040.41 0.41
0.41
6.41 6.41
6.41 7.29 7.29 7.33 7.33
7.33 6.04 6.04 0.41 0.41
C3 C3
C3
C3 C3
C3
6-carboxy-2-methyl-2-(4-methylpent-
6-carboxy-2-methyl-2-(4-methylpent-3-
6-carboxy-2-methyl-2-(4-methylpent-3-
6-carboxy-2-methyl-2-(4-methylpent-3-
6-carboxy-2-methyl-2-(4-methylpent-3-
6-carboxy-2-methyl-2-(4-methylpent-3-
6-carboxy-2-methyl-2-(4-methylpent-3-
3-enyl)-7-pentylchromen-5-olate
enyl)-7-pentylchromen-5-olate
enyl)-7-pentylchromen-5-olate
enyl)-7-pentylchromen-5-olate
enyl)-7-pentylchromen-5-olate
enyl)-7-pentylchromen-5-olate
enyl)-7-pentylchromen-5-olate
2.6.2.6.
2.6. CB2 CB2 Structure-Based
Structure-Based Virtual
Virtual Screening
Screening Results
Results
2.6.CB2
2.6. 2.6.
2.6.
CB2
CB2
Structure-Based
CB2
CB2
Structure-Based
Structure-Based
Virtual
Structure-Based
Structure-Based Screening
Virtual
Virtual
Virtual
Virtual Screening
Screening
Screening
Screening
Results
Results
Results
Results
Results
WeWe We performed VS using four PDB structures of the CB2 receptor (PDB id: 6KPF, 6KPC, 6PT0,
Weperformed
performed VS using
VS usingfour PDB
four PDBstructures
structures of the ofofCB2CB2
the CB2receptor
receptor (PDB (PDB id: id:6KPF,
id: 6KPF,
6KPF, 6KPC, 6KPC, 6PT0, 6PT0, 6PT0,
We Weperformed
We performed
performed
performed VS
VSVSusing
VSusing
usingusing four
four
four four
PDB
PDB PDB
PDB structures
structures
structures
structures of the
of of
the the
CB2the CB2 CB2
receptor
receptor receptor
receptor (PDB
(PDB (PDB id:(PDB
id: 6KPF,
6KPF, id:
6KPC,
6KPC,6KPF,
6KPC, 6PT0,
6PT0, 6KPC,
6PT0,
andand 5ZTY)
and
and
5ZTY)
5ZTY)
5ZTY) and and
two
and
and
two
two
two
compounds’
compounds’
compounds’
compounds’
libraries.
libraries.
libraries.
libraries. The The
The library
librarylibrary for for
the
for thefirst
the first compound
compound
first compound was wasderived
was derived
derived from from
fromthethe
the from
and5ZTY)
and
and 5ZTY)
5ZTY) and
and
and two
two
two compounds’
compounds’
compounds’ libraries. TheThe
libraries.
libraries. The library
The
librarylibrary for the
library
for for first
the the
first
for first
the compound
compound
compound first waswas
compound
was derived
derived derived from
was
from from
the the
derived
the
curated,
curated,
curated, ChEMBL
ChEMBLChEMBL training
training
training dataset
dataset
dataset prepared
prepared
prepared for for
our
for our QSAR
our QSAR
QSAR modelsmodels
models and and
andit itit included
included
included CB2-selective
CB2-selective
CB2-selective
curated,
curated,
curated,
the curated, ChEMBL
ChEMBL ChEMBL
ChEMBL trainingtraining
training dataset
dataset
training dataset prepared
prepared
prepared
dataset for
prepared for
for ourourfor our
QSAR
QSAR QSAR
our models
models
QSAR models andand
and
models it it
it included
included
and included
it6KPF CB2-selective
CB2-selective
CB2-selective
included CB2-selective
ligands
ligands
ligands expanded
expanded
expanded with with
with CB2-non-selective
CB2-non-selective
CB2-non-selective ligands.
ligands.
ligands. ForFor thisthis
For compound
compound
this compound library,
library, library, 6KPF
6KPF andandand5ZTY 5ZTY
5ZTY
ligands
ligands
ligands expanded
expanded
expanded with
with with CB2-non-selective
CB2-non-selective
CB2-non-selective ligands.
ligands.
ligands. For
For For
this
this this compound
compound
compound library,
library, library, 6KPF
6KPF 6KPF and
and and5ZTY
5ZTY 5ZTY
ligands
structures expanded
structures achieved
achieved with
the the CB2-non-selective
highest
highest area area underunder the ligands.
the receiver
receiver For this
operating
operating compound
characteristic
characteristic library,
curve curve (ROC 6KPF
(ROC AUC) AUC)and 5ZTY
structures
structures
structures
structures achieved
achieved
achieved
achieved the the
the the
highest
highesthighest
highest area area
areaarea under
under under
underthe the
the the receiver
receiver
receiver
receiver operating
operating
operating
operating characteristic
characteristic
characteristic
characteristic curve
curve curve
curve(ROC
(ROC (ROC
(ROC AUC)
AUC) AUC)
AUC)
structures
value
value valuein the achieved
in the
discrimination
in discrimination
the the
discrimination
discrimination highest between
between area
between under
CB2-selective
CB2-selective
CB2-selectivethe receiver
ligandsligands
ligands versusoperating
versus
versus characteristic
CB2-non-selective
CB2-non-selective
CB2-non-selective ones onescurve(see
(see(see
ones Table(ROC
Table AUC)
Table
value
value value
in the
in in
the the discrimination
discrimination between
between
between CB2-selective
CB2-selective
CB2-selective ligands
ligands ligands versus
versus versus CB2-non-selective
CB2-non-selective
CB2-non-selective onesones
ones (see(see
(see Table
Table Table
value 2).
2). 2).
The
2). The
inThe the
The antagonist-bound,
discrimination
antagonist-bound,
antagonist-bound, 5ZTY5ZTY
between
5ZTY structure
structure
structure slightly
CB2-selective
slightly
slightly outperformed
ligands versus
outperformed
outperformed 6KPF 6KPF
6KPF (agonist-bound),
CB2-non-selective
(agonist-bound),
(agonist-bound), as
ones observed
as observed (see Table 2).
as observed
observed
2). The
2). The antagonist-bound,
antagonist-bound, 5ZTY 5ZTY structure
structure slightly
slightly outperformed
outperformed 6KPF 6KPF (agonist-bound),
(agonist-bound), as
as observed
observed
The inin aaantagonist-bound,
detailed
in antagonist-bound,
a detailed
detailed analysis
analysis
analysis of
5ZTY
5ZTY
of
ROC
of ROC structure
curves
curves
structure
ROC curves
slightly
obtained
obtained
slightly
obtained
outperformed
in in
our
outperformed
in ourenrichment
our
6KPF
enrichment
enrichment 6KPF
(agonist-bound),
study study
study (Figure
(Figure
(agonist-bound),
(Figure 6).
as
6).
The
6). The second
second
as
The observed
second in a
in aaindetailed
in a detailed
detailed analysis
analysis
analysis of ROC
of of ROC
ROC curves
curves curves obtained
obtained
obtained in our
in in our
our enrichment
enrichment
enrichment study
study study (Figure
(Figure (Figure 6). The
6). 6). The
The second
second second
compound’s
compound’s
compound’s library
library
library used used
usedin in
our
in our
our structure-based
structure-based
structure-based VS VS included
included
VS included the the
same
the same
same CB2-selective
CB2-selective
CB2-selective ligands
ligandsligands derived
derived
derived
detailed
compound’s
compound’s analysis
compound’s library
library of
libraryROC
usedused
used curves
in our
in in our
our obtained
structure-based
structure-based
structure-based in our enrichment
VS included
VS included
VS included the same
the study
the same
same (Figure
CB2-selective
CB2-selective
CB2-selective 6). The
ligands
ligands second
ligands derived
derived compound’s
derived
fromfrom thethe
from the ChEMBL
ChEMBLChEMBL training
training
training dataset,
dataset,
dataset, named named
named here herewith
here with
withthethe term
the term
term “actives”,
“actives”,
“actives”, andand and general
generalgeneral diverse
diverse
diverse decoys
decoysdecoys
from
library
from from
the
used
the the
ChEMBL
ChEMBL ChEMBL
in our training
training
training dataset,
dataset,
structure-based
dataset, named
named named
VS here here
included
here withwith
with the
the the the
term
term term
same “actives”,
“actives”,
CB2-selective
“actives”, andand
and general
general general
ligandsdiverse
diverse diversedecoys
derived
decoys decoys from the
thatthatwere
that were
were generated
generated
generated with with
DUD-E
with DUD-E
DUD-E [34],[34],
named
[34], named
named here hereas
here as “non-actives”.
“non-actives”.
as “non-actives”. In In
VSIn VS against
against
VS against this this
second
this second
second library,
library,
library,
thatthat
ChEMBL
that were
were were generated
generated
training
generated withwith
dataset,
with DUD-E
DUD-E DUD-E
named [34], [34],
[34],here named
named named
with here
here here
the asterm
as as “non-actives”.
“non-actives”.
“actives”,
“non-actives”. Inand
In VS
VS Inagainst
VS
against against
general this this
thisdiverse
second
second second library,
library,
decoys
library, that were
6KPF6KPF
6KPFandand 5ZTY
and 5ZTY
5ZTY structures
structures
structures performed
performed
performed similarly,
similarly,
similarly, discriminating
discriminating
discriminating CB2-actives
CB2-actives
CB2-actives from from
from non-actives
non-actives
non-actives at
at theat the
the
6KPF
6KPF 6KPFand
and and
5ZTY
5ZTY 5ZTY structures
structures
structures performed
performed
performed similarly,
similarly,
similarly, discriminating
discriminating
discriminating CB2-actives
CB2-actives
CB2-actives from
from from non-actives
non-actives
non-actives at
at the
theat the
generated
level ofwith
0.8
andandDUD-E 0.79, [34], named
respectively here
(see(see as
Table 2, “non-actives”.
2, In 6KPF
VS6KPF against this second library, 6KPF and
level
levellevel
level
of 0.8
level
of 0.8
of 0.8ofand
of 0.8
0.8
and
0.79,
and
and
0.79,
respectively
0.79,
0.79,0.79, respectively
respectively
respectively
respectively (see(see
(see
Table
(see
Table
Table Table
Table
2, ROC
2, ROC2,ROC
ROC
2, ROC
ROC AUC
AUC
AUC
AUC
AUC
AUC
values).
values).
values).
values).
values).
values). 6KPF
6KPF 6KPF
6KPF slightly
slightly
slightly
slightly
slightly
slightly
outperformed
outperformed
outperformed
outperformed
outperformed
outperformed
5ZTY
5ZTY
5ZTY
5ZTY
5ZTY
5ZTY
5ZTY structures
according
according
according to to
the performed
the
ROC
toROC ROC
the ROCROC AUC AUC
AUC
similarly,
characteristics
characteristics
characteristics
discriminating
and and6KPC
and 6KPC
6KPC performed CB2-actives
performed
performed the the
worst, from
worst,
the worst,
worst, however, non-actives
however,
however, the the
valuesat
values
the values
values
the
were level
were
were
of 0.8
according
according
according to the
to to
the the
ROC AUC
AUC AUC characteristics
characteristics
characteristics andand
and 6KPC
6KPC 6KPC performed
performed
performed the
the worst,
the worst, however,
however, however, the
the values
the values were
were were
andveryvery
0.79,close
very closewith
close with
respectively no
with no explicit
(see
explicit
noexplicit
explicit difference
Table
difference 2,
difference ROC between
between AUC
between activeactive
values).
activeand and inactive
6KPF
inactive
and inactive CB2 CB2
slightly structures.
outperformed
structures.
CB2 structures. Our Our
Our results
results 5ZTY
results showed
showed according
showed to
veryvery
very close
close close
withwith
with no no
no explicit
explicit difference
difference
difference between
betweenbetween active
active active
andand
and inactive
inactive
inactive CB2CB2
CB2 structures.
structures.
structures. OurOur
Our results
resultsresults
showed
showed showed
thethatthat
ROC even even
AUC slight
slight differences
differences
characteristics between
between andthe the the
6KPC crystal
crystal structures
structures
performed of
the of
the the
same
worst, same receptor
receptor
however, have have
the an an
impactimpact on on
the the
that
that that
that
eveneven
even even
slight
slight slight
slight differences
differences
differences
differences between
betweenbetween
between the the crystal
the
crystal
crystal crystal structures
structures
structures
structures of
of the
the of
of the same
the
same
same same
receptor
receptorreceptor
receptor have
have anvalues
have
have
an an impact
an
impact
impact impactwere
on
on onvery
on
the
the the close
the
VSVS
with no
VS
VS
results.
results.
explicit
results.
results.
A
A similar
AA similar
difference
similar
similar
conclusion
conclusionbetween
conclusion
conclusion waswaswas
was
derived
derived
activederived
derived from
and from
from
from
aa recent
a recent
inactive a recent
recent study
CB2 study on on
structures.
study
study on
on
glucagon
glucagon
glucagon
glucagon Our receptors,
receptors,
results
receptors,
receptors, wherewhere
showed
where
where an an
that
an even
VS results.
VS results.
ensemble
A similar
A similar conclusion
conclusion was derived
was derived from a
from a recent recent study
study on glucagon
onsimulations
glucagon receptors, receptors, where an an
an
where crystal
ensemble
ensemble
ensemble of ofof
receptor
receptor
of receptor
receptor
conformations
conformations
conformations
conformations
generated
generated
generated
generated in in short
in
in
short
short
short MDMD MD
MD simulations
simulations
simulations
outperformed
outperformed
outperformed
outperformed crystal
crystal
crystal
slight
ensemble
ensembledifferences of between
receptor
of receptor the
conformations crystal structures
generated in of the
short same
conformations generated in short MD simulations outperformed crystalMD receptor
simulations have an
outperformed impact on the
crystal VS results.
structures
structures
structures in VSin
in VS [35,36].
[35,36].
VS [35,36].
structures
structures
Astructures in
similar conclusion VSin VS [35,36].
[35,36].
in VS [35,36].was derived from a recent study on glucagon receptors, where an ensemble
of receptor conformations generated in short MD simulations outperformed crystal structures in
VS [35,36].
Int. J. Mol. Sci. 2020, 21, 5308 7 of 14
Int. J. Mol. Sci. 2020, 21, x FOR PEER REVIEW 7 of 14
Table 2. Results of the enrichment study for CB2-receptor structures.
Table 2. Results of the enrichment study for CB2-receptor structures.
CB2-Non-Selective Decoys General Decoys
Metric
Metric 6KPF
CB2-Non-selective
6KPC 6PT0
Decoys
5ZTY 6KPF 6KPC
General Decoys 5ZTY
6PT0
6KPF 6KPC 6PT0 5ZTY 6KPF 6KPC 6PT0 5ZTY
EF 2% 0 0 0 1.7 3.3 0 0 3.3
EF 2%
EF 5% 0.67 0 0 0 0 0 1.3 1.7 4.7 3.3 1.3 0 1.3 0 3.3 3.3
EF10%
EF 5% 1.7 0.67 0 0 0.33 0 1 1.3 4.0 4.7 2 1.3 1.7 1.3 3 3.3
ROC
EF 10% 1.7 0 0.33 1 4.0 0.73 2 0.77 1.7 0.79 3
0.6 0.53 0.53 0.6 0.8
AUC
ROC AUC 0.6 0.53 0.53 0.6 0.8 0.73 0.77 0.79
A
B
Figure 6.6. The
Thereceiver operating
receiver characteristic
operating (ROC)
characteristic curves
(ROC) obtained
curves in theinenrichment
obtained study study
the enrichment using
libraries including (A) CB2-selective and CB2-non-selective ligands and (B) CB2-actives and CB2-non-
using libraries including (A) CB2-selective and CB2-non-selective ligands and (B) CB2-actives and
active ligands. ligands.
CB2-non-active
3. Discussion
3. Discussion
The
The main
main goal of this
goal of this study
study was
was to to predict
predict CB2-selective
CB2-selective compounds
compounds among among thethe phytochemicals
phytochemicals
that Cannabis Sativa.
that constitute Cannabis Sativa. We achieved this goal through the following steps: collecting aa large
constitute We achieved this goal through the following steps: collecting large
set
set of experimental data from ChEMBL for the training set, data curation, model development and
of experimental data from ChEMBL for the training set, data curation, model development and
validation,
validation, andandpredicting
predictingthe selectivity
the selectivityof phytochemicals
of phytochemicals fromfromCannabis Sativa.Sativa.
Cannabis In parallel, we tested
In parallel, we
the applicability
tested of recently
the applicability of released
recently PDB structures
released PDB of CB2 in structure-based
structures VS by conducting
of CB2 in structure-based VS an by
enrichment study. The dataset collected by our group was curated, which resulted
conducting an enrichment study. The dataset collected by our group was curated, which resulted in in the removal
of
themore
removalthanof90%
moreof than
unreliable
90% ofand inconsistent
unreliable data. Despite
and inconsistent theDespite
data. removalthe of removal
this largeoffraction of
this large
data, theof
fraction final dataset
data, included
the final dataset1958 records1958
included for CB1 andfor
records 2616 records
CB1 for CB2.
and 2616 Notably,
records such
for CB2. a large
Notably,
dataset is sufficient for employing the machine learning approach. Much smaller
such a large dataset is sufficient for employing the machine learning approach. Much smaller datasets datasets have been
successfully used to develop QSAR models for CB1/CB2 studies [23,24]. Nevertheless,
have been successfully used to develop QSAR models for CB1/CB2 studies [23,24]. Nevertheless, focusing on the
dataset
focusingdiversity
on the in developing
dataset our in
diversity QSAR model also
developing ourcreated
QSAR amodelchallengealsofor predicting
created the activity
a challenge for
of outlier compounds. We solved this problem by determining an applicability
predicting the activity of outlier compounds. We solved this problem by determining an applicability domain for each
compound
domain forindividually.
each compound Embedding the descriptor
individually. Embedding space
theofdescriptor
base GB modelsspace made
of baseit GB
possible to assess
models made
the prediction confidence based on Euclidean distances of k-nearest neighbors.
it possible to assess the prediction confidence based on Euclidean distances of k-nearest neighbors. The GB algorithm
The GB algorithm was also successfully applied by Ancuceanu et al. [37] for cytotoxicity prediction.Int. J. Mol. Sci. 2020, 21, 5308 8 of 14
was also successfully applied by Ancuceanu et al. [37] for cytotoxicity prediction. The performance
of GB in the modeling of various QSARs was compared to other ensemble methods in a study by
Kwon et al. [38].
Our approach was validated with the five-fold cross-validation protocol with resulting statistical
characteristics of Q2 > 0.6 for the entire dataset and Q2 > 0.8 for the most confident predictions.
Regarding small subsets with the highest confidence predictions, our model outperformed much more
complex 3D and 4D QSAR models created for equally small sets of similar compounds [23,24]. We used
our validated QSAR models for the prediction of the CB1 and CB2 affinities of phytochemicals from
Cannabis Sativa. As a result, we identified three compounds C1–C3 (see Table 1) dissimilar to the training
dataset with a TC < 0.3 compared to the respective most similar structure, and predicted pKi difference
CB2 vs. CB1 (dpKi) ≥ 1. The common name of C1 is cannabispirol, and it is a compound isolated
from Japanese domestic Cannabis Sativa [39]. C1 has been experimentally evaluated for antimicrobial
activity [40] and was also tested in a study on targeting multidrug resistant mouse lymphoma cells [41].
Ilicic acid (C2) showed activity in G2/M cell cycle arrest of tumor cells [42]. Cannabichromente (C3) is
a precursor for biosynthesis of various cannabinoids [43].
As for structure-based VS using PDB structures of CB2, we observed that these structures were able
to discriminate not only CB2-actives from non-active ligands (DUD-E decoys) but also CB2-selective
from non-selective ligands (the ChEMBL-derived dataset). The best results were obtained for 5ZTY
and 6KPF structures, regardless of the activation state.
In this study, we identified potential selective cannabinoids among the constituents of Cannabis sativa.
Our QSAR model identified three compounds of potential interest with significant dissimilarity to
compounds already evaluated experimentally and reported in ChEMBL. The statistical characteristics
of the developed QSAR models suggest a high probability of successful experimental validation,
which we hope will attract attention from experimental groups interested in searching for CB2-selective
ligands of natural origin.
4. Materials and Methods
4.1. Data Collection
Chemical structures from the ChEMBL database [44] with experimentally measured affinities for
the CB1 and CB2 receptors were used as training data. The datasets were composed of 14,126 records
for CB1 and 13,506 records for CB2. Compounds constituting the phytochemical profile of Cannabis
Sativa were acquired from the Collective Molecular Activities of Useful Plants Database (CMAUP) [45].
The ChEMBL data was exported and downloaded in comma-separated values (CSV) format, while the
CMAUP data was downloaded as an Structure Data Format (SDF) file.
4.2. Data Curation
The method of data curation followed a modified approach of Fourches et al. [46,47]. We modified
the curation method by the addition of extra steps that were relevant for curating data acquired
from ChEMBL accompanied by additional metadata specific to this database. Namely, we grouped
experimental values provided their source studies were the same, and then removed groups with less
than 10 records. We also assessed data reliability according to the source study confidence score as
provided by ChEMBL. We then removed records with a confidence score lower than 8. A final dataset
consisted of records with InChIKeys as molecular structure identifiers with associated standardized
2D structures of compounds and pKi values. RDKit was used for standardization and InChIKeys
calculation [48]. The curated CB2 and CB1 datasets are available on the website [49].
4.3. Descriptor Calculation
The Mordred python library [50] was used to calculate 1613 2D standard descriptors and 213
3D descriptors. 3D descriptors were used to include information about chirality in ligand structures.Int. J. Mol. Sci. 2020, 21, 5308 9 of 14
Descriptors with variance less than 0.05 or containing invalid values were removed. The descriptors
were concatenated with 1024-bit Morgan fingerprints with a radius equal to 3.
4.4. Machine Learning
A gradient boosting (GB) algorithm implemented in the Light Gradient Boosting Machines
library [51] was used to create base models. The GB algorithm and its parameters were selected in
an inner 5-fold cross-validation protocol based on grid search. Other algorithms including linear
regression, partial least squares, and support vector machines from the scikit-learn library [52] were
also tested. GB models for CB1 and CB2 were decision tree ensembles. The output of each tree in a
boosted ensemble was used to create embedding of a descriptor space. A kNN algorithm with k = 3 was
used to determine the applicability domain of the model based on the embedding. The applicability
domain was defined by a threshold, which is a maximum allowed distance between a query molecular
structure and nearest neighbors. Molecular structures with the Euclidean distance to nearest neighbors
greater than a given threshold were considered as out of the applicability domain. Python code for the
execution of trained models is provided in the Supplementary Materials.
4.5. Model Validation
Models were validated using a 5-fold cross-validation protocol. Out-of-sample predictions were
used to calculate Q2 according to the following equation:
P 2
n Yi − Ŷi
Q2 = 1 − P 2 (1)
n Yi − Yi
Here, n is the number of samples in the dataset, Yi is an experimental pKi of the ith sample,
Ŷi is a predicted value of pKi of the ith sample, and Yi is a mean pKi value. We evaluated Q2 for
20 different applicability domain thresholds and considered Q2 as a confidence measure for a model
using a given threshold.
4.6. Prediction of CB2-Selectivity of Cannabis Sativa Ingredients
A curated dataset of Cannabis Sativa constituents reported in the CMAUP database [45] was tested
against our QSAR CB1 and CB2 models. In such a way, we searched for selective molecules with
dpKi ≥ 1, where dpKi was defined as dpKi = pKiCB2 − pKiCB1 . For each predicted selective compound,
a Tanimoto coefficient (TC) was calculated to assess the similarity of predicted compounds to ones
already tested and used to exclude hits that were not novel with regard to training data. TC was
computed using 1024-bit Morgan fingerprints with a radius equal to 3.
4.7. Structure-Based Virtual Screening
CB2 receptor structures (PDB id: 6KPC [30], 6KPF [30], 6PT0 [31], and 5ZTY [32]) were derived
from the Protein Data Bank (PDB) [53]. This set included agonist-bound (6KPC, 6KPF, 6PT0) and
antagonist-bound (5ZTY) structures corresponding to different activation states of the CB2 receptor
(active and inactive, respectively). The CB2 structures used in this study were very similar to each other
(average pairwise RMSD = 6.41, with standard deviation = 3.40), yet differing in TMH6 conformation
(5ZTY and 6KPC vs. others) bending while interacting with a G protein complex. In all these
structures ligands were located in the orthosteric binding site of CB2, although allosteric modulation
has also been observed for these receptors [54]. The four PDB structures have nearly identical binding
sites except for a few residues: W258 (W6.48), S285, and to a lesser extent: F87 (F2.57), F91 (F2.61),
H95 (H2.65), F183 (EC2) (see Figure 7). W258 and F117 (F3.36) residues form a well-known Trp-Phe
toggle switch changing its position on the receptor activation [17,30]. Interestingly, the position ofInt. J. Mol. Sci. 2020, 21, 5308 10 of 14
W258 is slightly different not only in comparing the active vs. inactive conformations but also in three
active conformations, e.g., moved away from the ligand (6PT0 vs. two others).
Int. J. Mol. Sci. 2020, 21, x FOR PEER REVIEW 10 of 14
Figure
Figure 7.
7. Comparison
Comparison of of CB2
CB2 structures
structures available
available in
in the
the Protein
Protein Data
Data Bank
Bank (PDB).
(PDB). Here,
Here, an
an inactive
inactive
CB2
CB2 conformation
conformation (5ZTY)
(5ZTY) is
is shown
shown in in green
green and
and is
is the
the best
best performing
performing in
in virtual
virtual screening
screening (VS),
(VS),
active conformation (6KPF) is shown in yellow, and other CB2 active structures are shown
active conformation (6KPF) is shown in yellow, and other CB2 active structures are shown in grey. in grey.
Residues
Residues are
are labeled
labeled according to the
according to the 5ZTY
5ZTY numbering.
numbering.
Receptor
Receptor structures
structures were processed
processed withwith the
the Protein
Protein Preparation
Preparation tooltool 2017-4,
2017-4, Schrodinger
Schrodinger LLC,
LLC,
New York, NY,NY, USA
USA [55]. Ligands in PDB structures were used to determine the position of docking
grids spanning over the whole CB2 active site. The The prepared
prepared CB2 CB2 structures
structures were used for screening
against a CB2 actives
actives library
library that
that included
included CB2-selective
CB2-selective and and CB2-non-selective
CB2-non-selective compounds to assess
the ability of these CB2 structures to discriminate between between selective
selective and
and CB2-non-selective
CB2-non-selective ligands.
ligands.
This compound
compound library
librarywas
wasextracted
extractedfromfrom our ChEMBL training set (see Materials and
our ChEMBL training set (see Materials and Methods) and Methods)
and included
included compounds
compounds with pKiwith 7 for CB2-selective,
CB2 CB2-selective, while otherswhile
wereothers were
considered
considered CB2-non-selective.
CB2-non-selective. The second
The second round of VS wasround of VS was
performed performed
to test to what extentto test to what extent
experimental CB2
experimental CB2 structures were able to discriminate between CB2-actives
structures were able to discriminate between CB2-actives and non-actives. Here, we used the and non-actives. Here,
we used thelibrary
compounds compounds library
that was that was
generated usinggenerated
DUD-E [34] using DUD-E
with [34] with ChEMBL-derived
ChEMBL-derived CB2-
CB2-selective ligands
selective ligands
(see above) used (see
hereabove)
as CB2used hereHere,
actives. as CB2 weactives.
discardHere, we discard CB2-non-selective
CB2-non-selective ligands and we ligands
did not
and we them
include did not include
in the them in
compound the compound
library. library. VSwith
VS was performed wasGlide
performed
2017-4,with Glide 2017-4,
Schrodinger LLC,
Schrodinger
New York, NY, LLC,USA New[56]
York, NY,
that USA [56]
followed thethat followed
ligand the ligand
preparation withpreparation
Ligprep [57].withWeLigprep [57].
computed
We computed enrichment factors (EF) and areas under the curve (AUC) for
enrichment factors (EF) and areas under the curve (AUC) for receiver-operator curves (ROC) usingreceiver-operator curves
(ROC)
Maestro using Maestro
2017-4, 2017-4, LLC,
Schrodinger Schrodinger
New York,LLC, New
NY, USA.York, NY, USA .
Conclusions
5. Conclusions
We would like to emphasize that our study aimed to perform an in silico study to identify
prospective CB2-selective compounds from C. Sativa that could be confirmed confirmed by experimental
experimental studies
studies
in the future. This
This study
study shows
shows the
the first
first robust
robust CB1/CB2
CB1/CB2 QSARQSAR models
models that
that are reproducible and
applicable to a large variety of chemical scaffolds. Our Our developed
developed model
model was
was trained
trained on
on a large and
diverse set
setofofcompounds
compounds that surpassed
that surpassedrecent studies
recent [26]. We
studies [26].based our method
We based on machine
our method learning,
on machine
learning, which utilized data from thousands of experimental studies on CB1/CB2 activities. been
which utilized data from thousands of experimental studies on CB1/CB2 activities. Our study was Our
conducted
study was without making common
been conducted withoutmistakes
makingsuch as, e.g.,
common not conducting
mistakes such as,data curation
e.g., or its improper
not conducting data
curation or its improper execution, lack of a validation protocol, applying inappropriate statistical
characteristics for estimation of predictive performance, and finally, lack of applicability domain
determination [58–60]. We also showed that the applicability domain estimation based on a gradient
boosted model latent space allows the accurate prediction of the model confidence. What is more, we
validated the estimation of the model confidence interval, which allows users to make conclusionsYou can also read
