Vous êtes sur la page 1sur 8

426 Chem. Pharm. Bull. 67, 426–432 (2019) Vol. 67, No.

Regular Article

Three-Dimensional Classification Structure–Activity Relationship


Analysis Using Convolutional Neural Network
Hirotomo Moriwaki,a,† Yu-Shi Tian,a Norihito Kawashita,b and Tatsuya Takagi*,a
a
 Graduate School of Pharmaceutical Sciences, Osaka University; 1–6 Yamadaoka, Suita, Osaka 565–0871,
b
Japan: and  Faculty of Sciences and Engineering, Kindai University; 3–4–1 Kowakae, Higashiosaka, Osaka
577–8502, Japan.
Received September 28, 2018; accepted January 28, 2019

Quantitative structure–activity relationship (QSAR) techniques, especially those that possess three-di-
mensional attributes, such as the comparative molecular field analysis (CoMFA), are frequently used in mod-
ern-day drug design and other related research domains. However, the requirement for accurate alignment of
compounds in CoMFA increases the difficulties encountered in its use. This has led to the development of several
techniques—such as VolSurf, Grid-independent descriptors (GRIND), and Anchor-GRIND—which do not
require such an alignment. We propose a technique to construct the prediction model that uses mo-lecular
interaction field grid potentials as inputs to convolutional neural network. The proposed model has been found
to demonstrate higher accuracy compared to the conventional descriptor-based QSAR models as well as
Anchor-GRIND techniques. In addition, the method is target independent, and is capable of provid-ing useful
information regarding the importance of individual atoms constituting the compounds contained in the chemical
dataset used in the proposed analysis. In view of these advantages, the proposed technique is expected to find
wide applications in future drug-design operations.
Key words activity prediction; molecular interaction field; deep learning; three-dimensional quantitative
structure–activity relationship

Introduction of descriptors. In contrast, 3D-QSAR models are constructed


Accurate prediction of biological and chemical properties, using potential energy values calculated by force fields from the
such as physiological activity, toxicity, and so on, through use 3D structure of compounds. Such models are, therefore, able to
of the chemical structure of compounds are expected to curb avoid problems concerning the explanatory capabil-ity of
high costs of unnecessary biological and physical property descriptors, which are encountered by descriptor-based QSAR
evaluations. Quantitative structure–activity relation-ship methods.
(QSAR) -based approaches serve to accurately predict The comparative molecular field analysis (CoMFA)1) tech-
physiological activities. Modern-day QSAR techniques may be nique employs a typical 3D-QSAR model that adheres to the
classified into two categories—descriptor- based QSAR and following sequence of operation.
three-dimensional QSAR (3D-QSAR). In conventional 1. All compounds in the dataset are aligned using coordi-
descriptor-based QSAR methods, a number of descriptors nates of atoms.
representing the physical and chemical characteristics, such as 2. A grid is generated around each compound.
molecular weight, log P values, and number of hydrogen bond 3. Virtual atoms, called probes, are placed at each grid point,
donors or acceptors, are first calculated from the structure of and the net interaction energy between the probe and
compounds contained in the dataset. Subsequently, a quantita- compound (termed as the molecular interaction field
tive model is created by using these descriptors as features and (MIF)) is calculated.
effects, such as physiological activity, toxicity, and so on, as 4. Using partial least square (PLS) regression or other suit-
response values. Despite the fact that this method is simple, able regression schemes, a prediction model is created
effective, and widely accepted for use, it must be remembered from the calculated MIF values calculated for each com-
that it suffers an inherent limitation of being dependent on the pound constituting the dataset.
explanatory capabilities of descriptors calculated in the first By allocating various probes that possess different van der
step. For example, models created using two-dimensional (2D) Waals’ parameters and partial atomic charge at each grid point,
descriptors are difficult to recognize differences in response useful information concerning drug design—such as
variables arisen from differences in molecular chirality. This identification of a site where an increase or decrease in the value
occurs on account of molecular chirality, which is not ac- of response variables has been observed depend-ing on the
counted for by most 2D descriptors. Also, the descriptors are presence of a hydrogen-bond donor, acceptor, or hydrophobic
merely specific properties of the compounds. As such, they may group—may be obtained. CoMFA, therefore, is considered and
not be able to cover all factors needed to explain the response employed as an important technique in medi-cal and drug-
values. This explains the low explanatory capability related research applications. The disadvantage in its use,
however, is the necessity of strictly superimposing compounds
† for the purpose of alignment, as mentioned in the first step of
 Present address: RIKEN Center for Biosystems Dynamics Research; 1–7–
22 Suehiro-cho, Tsurumi-ku, Yokohama 230–0045, Japan. the above sequential procedure. This align-
* To whom correspondence should be addressed. e-mail: ttakagi@phs.osaka-u.ac.jp
© 2019 The Pharmaceutical Society of Japan
Vol. 67, No. 5 (2019) Chem. Pharm. Bull. 427

ment, in turn, governs the prediction accuracy of the CoMFA


technique. In most cases, such an alignment requires a series of
compounds possessing similar, if not identical, skeletons (e.g.
compounds having a steroid skeleton). Moreover, in order to
achieve accurate alignment, an advanced chemical
understanding is required for determining equivalent atoms
amongst various chemical compounds contained in the data-set.
Manual manipulation is also required during the train-ing and
prediction steps. To overcome this shortcoming of the CoMFA
technique and reduce, if not eliminate, the need for manual
manipulation, alternate QSAR methods, such as VolSurf,2)
Grid-independent descriptors (GRIND), 3) and Anchor-
GRIND,4) have been proposed. However, these meth-ods
generate Grid-independent descriptors from MIFs, thereby
making them susceptible to the limitations of descriptor-based
QSAR techniques. Recently, the deep or multilayered neural
network has been reported to demonstrate state-of-the-art
performance in the fields of image5) and speech recognition,6)
translation,7) and so on, and has also been underlined as a ma-
chine-learning technique. For image-recognition applications, Fig. 1. Six Examples of Dataset Compounds
the convolutional neural network (CNN) is mainly employed To show the contents and varieties of the dataset, six respective examples are shown
and has been a topic of active research.8–10) CNN falls within here. These six compounds are selected from a clustering using k-means method
(number of clusters is six). The inhibitory constant (Ki) values along with the classes
the deep- neural-network spectrum and comprises a layer that (active/inactive) are also shown.
performs the convolution operation. This study reports a tech-
nique to construct the prediction model that uses MIF grid po- Table 1. Numbers of Active and Inactive Compounds in Training and Test
tentials as inputs to CNN. The proposed technique is capable of Data Sets
constructing models without the need for human interven-tion,
and demonstrates higher prediction accuracy compared to Active Inactive Total
previously reported MIF- based techniques and descriptor-based
Training 191 99 290
models. Moreover, the proposed technique demonstrates
Test 88 57 145
comprehension of the important aspects of activity contributed
Total 279 156 435
by each atom within compounds belonging to a particular
dataset, thereby indicating its potential for wide application in
drug-design operations. argument, all compounds contained in the dataset employed in
this study were neutralized as per recommendations. Sub-
Experimental sequently, hydrogen atoms were added, and the 3D structure
Dataset A dataset comprising a series of published in-hibitors was embedded with 100 conformers set as the upper limit.
of Factor Xa of the benzamidine family,4) reported by Fontaine Structural optimization by the Merck molecular force field
et al., was used for analysis in this study. This series of (MMFF 94)12) was then performed for each structure using
inhibitors contains 435 compounds with their respective RDKit-2017_03_3.13) Lastly, a secondary structural optimiza-
inhibitory activities varying between 0.005 nM to 410 µM in tion calculation was performed for all conformers by employ-
inhibitory constant (K i). The separation of active and inactive ing the PM7 method14) through use of the MOPAC 2016 soft-
compounds according to the magnitude of inhibitory constant ware package.15) Amongst the conformers for each compound,
values was reported,4) and here we follow the results. Thus, the structure demonstrating the most stable energy in the PM7
there are 156 inactive (Ki > 1 µM) and 279 active (K i < 10 nM) analysis was adopted for further analysis.
compounds in the dataset. Six examples are shown in Fig. 1. To Data Augmentation In the learning phase of deep neural
construct prediction model, the dataset was randomly split into networks, data augmentation, which increases data for the
training (290 compounds) and test sets (145 compounds), which purpose of preventing over fitting, is often performed. Data
is also as the same as previously reported. 4) Thus, we can rotation, clipping, translation, lateral inversion, scaling, etc. are
compare directly the results of this study and the previous study. well-known techniques of data augmentation. Unlike 2D
Numbers of active and inactive compounds in training and test images, some of these operations may cause a change in MIF
data sets are summarized in Table 1. In their study, 4) Fontaine et attributes. For example, lateral inversion brings about a change
al. reported the Anchor-GRIND model demon-strating a in the enantiomer while scaling alters atomic proper-ties, such
prediction accuracy of 88%. as van der Waals’ radii. In this study, in order to avoid such
Structure Preparation The structure of each compound in changes in the shape of individual molecules, data rotation was
the dataset was imported in the SMILES format. It is generally applied exclusively. Each structure of the data-set was
difficult to determine whether a compound car-ries a charge or multiplied by a random rotation matrix M (Eq. 1). In this way,
not on account of the inability to accurately examine 100 such rotated structures were generated for each compound
experimental conditions and surroundings in which individual (Fig. 2) resulting in 29000, and 14500 structures, respectively,
compounds interact. In such cases, it is recom-mended to in the training and test datasets.
neutralize these compounds.11) In line with this
428 Chem. Pharm. Bull. Vol. 67, No. 5 (2019)

to ∞. Grid-potential scaling was, therefore, performed prior to


initiation of the learning exercise. N-clipping (Eq. 4) or
hyperbolic tangent (tanh) (Eq. 5) represent the initial scaling
methods adopted in this study. Subsequently, an additional
scaling (Eq. 6) was performed using the mean and standard
deviation of the entire training set to adjust the average to 0 and
variance to 1.

 l ( x > l)

y = x (l ≥ x ≥ −l) (4)

Fig. 2. Overview of Data Rotation − l ( − l > x)
An initial structure of an example molecule is shown on the top. After multiply-ing y =tanh(l × x) (5)
the coordinates of its atoms by a random rotation matrix, a new set of coordi-nates
showing a rotated structure (e.g., the structure shown on the lower left) can be
obtained. By changing the rotation matrix 100 times, 100 rotated structures of the x−μ
molecule can be obtained.
y= σ (6)

As an illustration of this scaling, potential energies of the grid


 3×3 
Q , R =QR  generated for the compound depicted in Fig. 3A and in-cluded
 2  in the training dataset are shown in Fig. 3C.
d =diag( R)
d
M=Q×| d|
where QR(x) implies QR decomposition of matrix x; Nr × c is an
r × c matrix of random numbers based on the standard normal
distribution, and diag(x) represents diagonal elements of matrix
x.
Grid Generation The grid used in the analysis was made
large enough to contain all compounds comprising the data-set.
In order to determine the grid size, principal component analysis
(PCA) was performed on each structure in the train-ing dataset.
Lengths of compounds in the direction of the first principal
component were obtained and that of the largest compound was
determined to be 17 Å. The size of one voxel was set to 1 × 1 × 1
Å3, and one side of the grid was set to measure 32 Å in length,
which was obviously larger than 17 Å. The center of mass of
each compound was superposed on the center of the grid. To
calculate grid potential, the van der Waals’ term (Evdwij) (Eq. 2)
and electrostatic term (EQij) (Eq.
3) of MMFF 94, which incorporates 94 types of atoms, were
employed. Terms, ε i and R*,i in E vd were calculated from van
der Waals’ parameters
j
for
j
atom
wij types i and j. Also, partial

atomic charges (qi and qj in EQ ) are calculated from the atom


and connected bonds. In this ijstudy, oxygen atoms in water
molecules—with an atom type identified as 70 and possess-ing a
partial atomic charge of −0.86—were used as probing atoms.
Equations 2 and 3 below give relationships to calculate the grid
potentials.

*7 
 1.07Rij*  7  1.12Rij −2
Evdwij =εij    

   7 7
* *
 Rij + 0.07Rij   Rij +0.12Rij
qi q j
EQij =332.0716 D ( Rij +δ)
Scaling Grid potentials calculated in the above step can-not
directly be used as inputs to CNN because the computa-tional
grid-potential domain covers the entire range from −∞
(1) Learning The network architecture employed in this study is depicted in
Fig. 4, and was constructed with reference to the Visual Geometry Group
(VGG) model10) using mxnet-0.10.0.16)
Weight initialization was performed in a manner similar to the
one previously reported by Glorot and Bengio17) Batch size and
epoch were set to 100 and 500,
respectively. The softmax cross-entropy
error function was employed to account for
losses along with use of the Adam
optimization algorithm.18) Other hyper
parameters were determined using the
method described in the next section.
Tuned hyper parameters, in Fig.
4, are indicated in italics. The input MIF
was first prepared using the ‘preparation
type’ and ‘preparation range’ hyper
parameters. Convolutional layers were then
added a number of times equal to the value
of the layers parameter. The density layers
are then repeated a number of times equal
to the value of the ‘densely connected
layers’ parameter. Finally, the soft-max
cross-entropy loss function is applied.
Hyper-Parameter Tuning In order to
obtain optimum values for hyper
parameters, 29 compounds, corresponding
to 10% of the training set, were randomly
selected as a valida-tion set and a random
search operation was performed. Each
hyper parameter listed in Table 2 was
randomly chosen, and subsequent learning
was performed on the training set. This
procedure was performed for 100 different
hyper-parameter patterns, and the
parameter demonstrating highest accuracy
was selected for the validation test.
Prediction and Model Evaluation One
prediction was obtained from each of the
hundred rotated structures cor-responding
to each compound, and the final prediction
result was derived on the basis of majority
vote with reference to the 10-crop testing.9)
Figure 5 depicts a schematic of this op-
(2)
eration, which is hereinafter referred to as
100-rotate testing (100-RT). The prediction
accuracy of the model was evaluated by the
test set.

(3) Inspection of Important Atoms An


analysis of the im-portance of each atom in
activity prediction of the proposed model is
expected to have an influence on drug
design. Since most classical force fields,
including the MMFF 94, formulate the
total energy by summing up the energies of
all atoms, importance of each individual
atom could be determined by
Vol. 67, No. 5 (2019) Chem. Pharm. Bull. 429

Fig. 3. Overview of Grid Potential Generation


(A) An example compound contained in the training dataset. (B) Schematic diagram of corresponding grid containing the compound. Probes are put on each lattice point and
interaction energy between the compound and probes is calculated. (C) Heat maps of interaction energy sliced along the z-axis in (B) are shown. The interaction energy measured
by the probe are depicted on a color scale, wherein negative energies are depicted in blue, positive energies in red, and the transition between them is indicated in white. (Color
figure can be accessed in the online version.)

Fig. 4. CNN Architecture


Each MIFs generated in “Scaling” is fed to input layer. Firstly, a 3 × 3 × 3 convolutional layer, a batch normalization layer, and a rectified linear unit (ReLU) layer are applied
twice sequentially. Secondly, a dense layer, a dropout layer and a ReLU layer are applied sequentially. Lastly, a dense layer and a softmax layer are applied and predicted value is
obtained.

estimating the prediction probability of a compound when a QSAR with 2D Descriptors For the purpose of com-parison,
certain atom is not present. In the proposed study, each atom random forest (RF)19) and logistic regression (LR)20) models
constituting the compound to be inspected was omitted one at a using 2D descriptors exclusively were constructed to form a
time, and the corresponding grid potential was calculated. The baseline case. The Mordred -0.6.0 package21) was used for
100-RT operation was then performed for each omitted atom. descriptor calculation while the scikit-learn-0.19.022) pack-age
The decline of the predictive probability between the initial and was used for model construction. Hyper parameters were once
atom-omitted structures indicates the importance of the omitted again determined through a random search of 100 points in the
atom. Finally, all the atoms demonstrating high importance were ranges listed in Table 3. For evaluating hyper param-eters,
highlighted according to the calculated atomic importance. validation accuracy determined using the 10-fold cross-
Figure 6 shows an illustration of the above process. validation (10-CV) technique was used.
430 Chem. Pharm. Bull. Vol. 67, No. 5 (2019)

Table 2. Tuned Hyper Parameters and Their Corresponding Range Results and Discussion
The proposed method was developed to investigate the
Hyper parameter Range feasibility of CNN—to which MIF potential grids serve as in-
Convolutional layers 4,6 puts—as a useful 3D-QSAR tool in contemporary drug design
Initialchannel 2 1.0 ,5.0 and related medical research. To demonstrate the applicability
Densely connected layers 0,3 of the proposed method, this section compares its prediction
Densely connected layer width 2 7.0 ,11.0 capability against that of conventional descriptor-based models
Preparation type tanh or clip employing 2D descriptors and the previously reported Anchor-
Preparation range  0.05,1.0

prepare type =tanh GRIND model. Subsequently, the contribution of hyper pa-

 0.05,100.0 prepare type =clip rameters involved in CNN is discussed. Lastly, the usefulness of

Ul, u represents the value sampled from the uniformly distributed random number in
the proposed analysis procedure to highlight the importance of
the range [l, u) (continuous distribution if l and u are decimal numbers, and discrete individual atoms of a compound in activity prediction is
distribution if they are integers). discussed.

Fig. 5. Overview of the 100-RT Operation


For each compound, 100-time rotations create 100 conformations, the MIF values of each conformation are calculated. For the training data, these MIF values are input to the
CNN architecture to train the model; for the test data, as shown here, after inputting these MIF values into the model, each conformation achieves a probability of being active. The
average probability of 100 conformations is calculated and if the value is larger than 0.5, the input compound is finally judged as an active one, otherwise as an inactive one, by
this model. (Color figure can be accessed in the online version.)

Fig. 6. Overview of Inspection Process for Important Atoms


Each atom within a compound is omitted one each time. The average probability of being active of these artificial atom-omitted structures are predicted using 100-RT method
aforementioned. Important atoms shown in red are defined as large decline of the average probability between the initial and atom-omitted compounds. (Color figure can be
accessed in the online version.)
Vol. 67, No. 5 (2019) Chem. Pharm. Bull. 431

Table 3. Tuned Hyper Parameters and Their Corresponding Range in 2D Descriptor Models

Model Hyper parameter Description Range


Common impute How to complete missing values (mean, median, most frequent value)
RF class_weight Class weight (none, balanced)
criterion Criterion (gini, entropy)
estimators Number of decision trees 2 5.0 ,11.0

max_depth Max depth of each decision tree 1,50

LR C Cost 2 0.0 ,10.0

penalty Penalty (L1, L2)

Table 4. Comparison of Result of the Learning Exercise Performed on


the Proposed and Baseline 2D Models

Test
Validation
Model accuracy Balanced
Accuracy
(%) accuracy F1 score
(%)
(%)
Proposed 96.4 91.7 91.0 0.933
RF 94.1 85.5 83.7 0.885
LR 94.5 90.3 89.6 0.921

Table 5. Three Co-crystallographic Structures of Compounds with Re-


ceptor Factor Xa

Ligand ID PDB ID Activity (FXa)/nM*

1 1FJS 0.11
2 1NFY 1.3
3 1LQD 9
* All the three ligands are in active class.

Comparison with Other Methods Table 4 lists hyper Fig. 7. Inspection of Importance of Individual Atoms in a Compound
parameters of the proposed, RF, and LR models along with The atomic importance of ligands 1, 2, and 3 are shown here. Circle indicates
corresponding accuracies of their respective validation (hold out important atoms according to the calculation of atomic importance in a scale that varies
validation of the propped method and 10-CV technique for RF from colorless to red. A rescaled mapping of atomic importance of ligand 3 is also
shown as 3 here. (Color figure can be accessed in the online version.)
and LR methods) and test sets.
As seen in Table 4, the proposed method demonstrates a test-
set accuracy of 91.7%, which is higher compared to the Anchor- proposed model, is depicted in Fig. 7.
GRIND (88%) and baseline models employing 2D de-scriptors Each molecule is highlighted based on the significance of its
(85.5% for RF and 90.3% for LR). Balanced accuracy and F1 constituent atoms, which is expressed by means of a color scale
score of the proposed model are also higher than those of 2D that varies from colorless to red with increase in sig-nificance.
descriptor models. By directly performing the predic-tion for all Since the activity of Ligand 3 is close to the active/ inactive
14500 rotated structures in the test set without invoking the 100- threshold and it is difficult to distinguish between active and
RT operation, the accuracy was found to be 90.2%, thereby inactive, most atoms in Ligand 3 are highlighted. Thus, the
suggesting that the 100-RT operation serves to achieve an rescaled image (3 ) is also shown.
improvement of roughly 1.5 percentage points in prediction The amidine substructure of benzimidazole in Ligands 1 and
accuracy. 3 was highlighted to be important but was not in Ligand 2 was
Explanation of the Model Explanation of QSAR models is not. To identify reason of this discrepancy, X-ray co-crystal
generally difficult. For the purpose of drug design, however, an structures of Ligands 1–3 (PDBID: 1FJS, 1NFY, and 1LQD) are
explanation of the constructed model is necessary. As pre- superimposed and compared (Fig. 8). Ligands 1–3 are colored
viously mentioned, the proposed model is capable of providing in red, green, and cyan, respectively. Residues demonstrating
information regarding the importance of individual atoms in a major interaction with compounds are also shown. Compared
compound by comparing prediction probability obtained via with Ligands 1 and 3, Ligand 2 is clearly showing different
omission of individual atoms from the compound one-at-a-time. binding mode in the co-crystal structure. For Ligand 2, instead
For the dataset considered in this study, three co- of benzamidine, chlorobenzene is posi-tioned in the active site
crystallographic structures of compounds with receptor Factor and one of nitrogen of benzamidine in Ligand 2 forms hydrogen
Xa were reported to the Protein Data Bank (PDB) (Table 5). All bond with Glu97. In Fig. 7, chloro group in Ligand 2 is strongly
the three ligands are in active class. An illustration of the atomic highlighted and one of nitrogen of benzamidine is highlighted.
importance of these compounds, as determined by the The results consistent with
432 Chem. Pharm. Bull. Vol. 67, No. 5 (2019)

as an occluding unit in model interpretation, but the more


descriptive result may be able to obtain using other occluding
units such like functional groups. Lastly, all compounds in this
dataset contain benzimidazole. Although most compounds in the
dataset have a large non -benzimidazole moiety, strictly
speaking, they are congeneric compounds. By increasing the
varieties of skeletons in the dataset, a more robust model may be
constructed. Resolution of these limitations may enhance the
worthiness of this method.

Acknowledgments This work was partially supported by the


Hirose International Scholarship Foundation for fellowship to
Fig. 8. Co-crystal Structures of Ligands with Factor Xa
Yu-Shi Tian.
X-Ray co-crystal structures of Ligands 1–3 (PDBID: 1FJS, 1NFY, and 1LQD) are
superimposed. Ligands 1–3 are colored in red, green, and cyan, respectively. Residues
Conflict of Interest The authors declare no conflict of
demonstrating major interaction with compounds are also shown. (Color figure can be interest.
accessed in the online version.)

References
X-ray crystal structures. The difference in activities of respec- 1) Cramer R. D., Patterson D. E., Bunce J. D., J. Am. Chem. Soc., 110,
tive molecules was assumed to be attributed to the presence or 5959–5967 (1988).
absence of these atoms. The inspection of important atoms in 2) Cruciani G., Pastor M., Guba W., Eur. J. Pharm. Sci., 11 (Suppl. 2),
low-activity compounds discussed in this paper is independent S29–S39 (2000).
3) Pastor M., Cruciani G., McLay I., Pickett S., Clementi S., J. Med.
of the receptor-molecule structure. The proposed technique can,
Chem., 43, 3233–3243 (2000).
therefore, be expected to find applications in drug design 4) Fontaine F., Pastor M., Zamora I., Sanz F., J. Med. Chem., 48, 2687–
procedures, especially for cases wherein the structure of target 2694 (2005).
molecules is unknown. 5) He K., Zhang X., Ren S., Sun J. ArXiv151203385 Cs (2015).
6) Amodei D., Anubhai R., Battenberg E., Case C., Casper J., Catan-zaro
Conclusion B., Chen J., Chrzanowski M., Coates A., Diamos G., Elsen E., Engel
This study proposes a 3D-QSAR technique that uses MIF J., Fan L., Fougner C., Han T., Hannun A., Jun B., LeGresley P., Lin
grid potentials as input to CNN and reports an analysis to L., Narang S., Ng A., Ozair S., Prenger R., Raiman J., Sath-eesh S.,
investigate its effectiveness in comparison to conventional de- Seetapun D., Sengupta S., Wang Y., Wang Z., Wang C.,
scriptor-based 2D- QSAR methods. The proposed method does Xiao B., Yogatama D., Zhan J., Zhu Z. ArXiv151202595 Cs (2015).
not require human manipulation, such as compound alignment 7) Wu Y., Schuster M., Chen Z., Le Q. V., Norouzi M., Macherey W.,
Krikun M., Cao Y., Gao Q., Macherey K., Klingner J., Shah A.,
or designation of anchor atoms required in the case of CoMFA
Johnson M., Liu X., Kaiser Ł., Gouws S., Kato Y., Kudo T., Kazawa
and Anchor-GRIND, respectively. Furthermore, the proposed
H., Stevens K., Kurian G., Patil N., Wang W., Young C., Smith J.,
technique demonstrates higher prediction accuracies compared Riesa J., Rudnick A., Vinyals O., Corrado G., Hughes M., Dean J.
to previously reported 3D-QSAR techniques and models con- ArXiv160908144 Cs (2016).
structed using 2D descriptors. Since the model is constructed by 8) Gehring J., Auli M., Grangier D., Yarats D., Dauphin Y. N.
exclusively using a force field that is well-tested for physical ArXiv170503122 Cs (2017).
and chemical properties, it is expected that problems concern- 9) Krizhevsky A., Sutskever I., Hinton G. E., Proc. 25th Int. Conf.
ing explanatory capabilities of the model are much reduced in Neural Inf. Process. Syst., 1, 1097–1105 (2012).
this method compared to conventional descriptor-based QSAR 10) Simonyan K., Zisserman A. ArXiv14091556 Cs (2014).
techniques. This is because the proposed technique does not 11) Fourches D., Muratov E., Tropsha A., J. Chem. Inf. Model., 50, 1189–
1204 (2010).
involve use of descriptors that extract only specific proper-ties,
12) Halgren T. A., J. Comput. Chem., 17, 490–519 (1996).
such as molecular weight, log P values, and the number of
13) Landrum G., RDKit: Open-source cheminformatics.: ‹http://www.
hydrogen-bond donor or acceptor. Additionally, this study rdkit.org›.
proposes a method to analyze the constructed model. The 14) Stewart J. J. P., J. Mol. Model., 19, 1–32 (2013).
proposed method of analysis not only helps to confirm which 15) J. P. Stewart J. MOPAC2016.: ‹http://openmopac.net/›.
part of the structure contains compounds that may be adjudged 16) Chen T., Li M., Li Y., Lin M., Wang N., Wang M., Xiao T., Xu B.,
as being active or inactive, based on the presence or absence of Zhang C., Zhang Z. ArXiv151201274 Cs (2015).
particular atomic species, but also provides information 17) Glorot X., Bengio Y., Proc. Thirteen. Int. Conf. Artif. Intell. Stat., 9,
regarding the most important atoms responsible for binding to 249–256 (2010).
the receptor. The authors, therefore, believe that their proposed 18) Kingma D. P., Ba J. L. ArXiv14126980 Cs (2014).
19) Breiman L., Mach. Learn., 45, 5–32 (2001).
method could find wide acceptability in drug-design applica-
20) Cox D. R., J. R. Stat. Soc., B, 20, 215–242 (1958).
tions. However, this method includes several limitations. First-
21) Moriwaki H., Tian Y.-S., Kawashita N., Takagi T., J. Cheminformat-
ly, this study is designed for constructing a model of a series of ics, 10, 4 (2018).
compounds with targets or co-crystal structures unknown. 22) Pedregosa F., Varoquaux G., Gramfort A., Michel V., Thirion B.,
Although this method can be applied to a data set of co-crystal Grisel O., Blondel M., Prettenhofer P., Weiss R., Dubourg V., Van-
structures solved, in those cases, other approaches using crys-tal derplas J., Passos A., Cournapeau D., Brucher M., Perrot M., Duch-
structure are recommended for effectiveness. Secondly, to esnay E., J. Mach. Learn. Res., 12, 2825–2830 (2011).
simplify the procedure, in the current study atoms are used