This is an open access article published under an ACS AuthorChoice License, which permits

copying and redistribution of the article or any adaptations for non-commercial purposes.

Article

pubs.acs.org/jcim

De Novo Design of Drug-Like Molecules by a Fragment-Based

Molecular Evolutionary Approach
Kentaro Kawai,*,† Naoya Nagata,† and Yoshimasa Takahashi‡
†
Central Research Laboratories, Kaken Pharmaceutical Co. Ltd., 14, Shinomiya Minamikawara-cho, Yamashina, Kyoto 607-8042,
Japan
‡
Department of Computer Science and Engineering, Toyohashi University of Technology, 1-1, Hibarigaoka, Tempaku-cho,
Toyohashi 441-8580, Japan
*
S Supporting Information

ABSTRACT: This paper describes a similarity-driven simple

Downloaded from pubs.acs.org by 106.193.130.203 on 09/17/18. For personal use only.

evolutionary approach to producing candidate molecules of

new drugs. The aim of the method is to explore the candidates
that are structurally similar to the reference molecule and yet
somewhat different in not only peripheral chains but also their
scaffolds. The method employs a known active molecule of our
interest as a reference molecule which is used to navigate a
huge chemical space. The reference molecule is also used to
J. Chem. Inf. Model. 2014.54:49-56.

obtain seed fragments. An initial set of individual structures is

prepared with the seed fragments and additional fragments
using several connection rules. The fragment library is
preferably prepared from a collection of known molecules related to the target of the reference molecule. Every fragment of
the library can be used for fragment-based mutation. All the fragments are categorized into three classes; rings, linkers, and side
chains. New individuals are produced by the crossover and the fragment-based mutation with the fragment library. Computer
experiments with our own fragment library prepared from GPCR SARfari verified the feasibility of our approach to drug
discovery.

■ INTRODUCTION
In the drug discovery process, various characteristics, including
respectively. The general advantage of ligand-based approaches
is their wide range of applicability, because those approaches
biological activities and ADMET properties, must be can be used in the case of that the three-dimensional (3D)
considered simultaneously. Medicinal chemists synthesize structure of the target is not available.
congeneric series of compounds to clarify the structure− Evolutionary algorithms are actively used for the compu-
activity relationship (SAR) of their own hits or lead series, and terized molecular design. They are based on concepts derived
they use the SAR knowledge to optimize the compounds in from biological evolution, including reproduction, mutation,
further synthesis. In addition, drug discovery entails compli- crossover, and selection. The algorithms are widely used to
cated tasks of optimization with respect to ADMET properties. solve various drug discovery problems, such as parameter
Medicinal chemists are often required to change scaffolds optimization of QSAR/QSPR models28 and 3D-ligand align-
further by so-called core hopping to address scaffold-dependent ment29 as well as the compound design described above. New
issues.1−3 One of the good examples is GSK’s B-Raf inhibitor molecules are designed by repeatedly applying the evolutionary
program.4,5 They changed the molecular frameworks (scaf- operations to existing molecules. Among these operations,
folds) during both of lead generation and lead optimization mutation and crossover are vital for generating new chemical
stages to discover a compound for clinical trial. structures. Mutation methods are roughly classified into the
Together with medicinal chemistry, computational chemistry fragment-based mutation and the atom-based mutation. In our
plays an important part in the discovery of new drugs. previous study, the atom-based method was used for the
Computational molecular design has been an active research mutation, in which an atom is modified into another atom to
area over the past decades. Many computerized structural explore the chemical space. The method often resulted in a lot
design approaches have been developed, which utilize protein- of unfavorable structures that contained invalid hetero−hetero
structures and/or ligand-structures. 6−27 For example, atom bonds such as O−O and N−F.20 One of the approaches
LEGEND,6 LUDI,7 SPROUT,8 LEA3D,9 LigBuilder,10,11 and to avoiding this problem is to use exclusion rules of the
SYNOPSIS12 use protein structures; whereas TOPAS,13 substructures reported by Huang et al.18 An alternative method
CoG,14 and Flux15,16 use the structures of known ligands.
The former methods are referred to as the structure-base design Received: July 17, 2013
and the latter methods to as the ligand-based design, Published: December 28, 2013

© 2013 American Chemical Society 49 dx.doi.org/10.1021/ci400418c | J. Chem. Inf. Model. 2014, 54, 49−56
Journal of Chemical Information and Modeling Article

is to use a fragment-based mutation instead of the atom-based prepared using the seed fragments and additional fragments
mutation. We also considered that the use of fragments is a from a fragment library. The procedure of the present
good way to generate chemically feasible structures when those evolutionary approach is summarized as follows:
fragments are derived from known molecules. (1) Input a reference structure.
Evolutionary algorithms use fitness score to select the (2) Generate seed fragments.
surviving structures. For example, Molecule Evoluator17 uses (3) Make an initial set of individual structures.
medicinal chemists’ knowledge and Flux uses a similarity index (4) Generate offsprings for the next generation by mutation
(Tanimoto coefficient or Euclidean distance) as the fitness and crossover.
function. In addition, building structures with chemical (5) Evaluate fitness of the structures and select some of them
feasibility is an important point for de novo design. Frag- to survive.
ment-based approaches were favorably used for the purpose. (6) Steps 4 and 5 are repeated until that the alternation of
For example, NovoFLAP19 uses fragments with thirty-two generation reaches to the specified number.
chemical transformation operators to generate structures
obeying valence rules. Flux builds molecules by a RECAP- In the following, we will use the term “fragment” to mean a
based rule (11 reaction schemes) to connecting fragments. building block. In our approach, a molecule can be built by
These approaches are reasonable to generate feasible structures. connecting the fragments at their connection points specified in
However, we considered that a more simplified approach with advance. The connection points of a fragment are identified
fewer connection rules has an advantage in use. from the original molecule as shown in Figure 2.
In this paper, we propose a similarity-driven fragment-based
evolutionary approach to producing drug-like molecules for
drug discovery. Aim of the method is to explore the candidates
that are similar to a reference molecule and yet somewhat
different in not only the side chains but also their scaffolds.
Chemical feasibility of the candidates is also considered.

■ METHODS
Outline of Evolutionary Algorithm. The basic idea of the
present method is a similarity-driven simple evolutionary
approach (Figure 1). The method employs an existing active
molecule of our interest. It is used as a reference molecule to
navigate a chemical space to be explored. The reference
molecule is also used to obtain seed fragments for making the
initial population. The initial set of individual structures is
Figure 2. Example of generation of seed fragments and making a initial
structure. A connectable point is indicated by an asterisk.

Seed Fragments. In the present work, the seed fragments

are automatically prepared by fragmentation of the reference
molecule. Then, they are employed for making the initial
structures with additional fragments that were randomly
selected from a fragment library (Figure 2). The detail of
making a structure is described in a later section (Making
Structure and Fragment Connection Rules).
Preparation of Fragment Library. In this work, we
prepared our own fragment library. Every fragment of the
library is preferably obtained from a collection of known
compounds that are related to the target of interest. Three
different types of fragments (ring, linker, and side-chain) were
defined as building blocks and used to make a molecule.
Walters et al. reported that 70% of the compounds in the
Journal of Medicinal Chemistry are composed of a relatively
small number of building blocks (e.g., only 37 rings, 53 linkers,
and 16 side-chains).30 Thus, we thought that a wide variety of
new molecular frameworks could be designed by the
combination of the building blocks such kind.
For making a molecule from the fragments, we explicitly
treated their possible connection points by labeling with
dummy atoms. For example, ortho-substituted benzene and
para-substituted benzene are treated as different fragments (see
ring fragments in Figure 3). Every type of monocyclic rings,
fused rings, or spiro rings is defined as ring fragments. It should
Figure 1. General scheme of the molecular evolutionary algorithm. be noted that biphenyl is not a ring fragment but a combination
50 dx.doi.org/10.1021/ci400418c | J. Chem. Inf. Model. 2014, 54, 49−56
Journal of Chemical Information and Modeling Article

fragment library. Even in a simple estimation, the number of

possible structures that consist of one ring, one linker, and one
side-chain exceeded 108. When discriminating with respect to
connection points and using multiple rings, multiple linkers,
and multiple side-chains, a very large number of structures can
be generated by using the present fragment library.
Making Structure and Fragment Connection Rules.
Individual structures of each population are produced from the
fragments according to several connection rules. We defined
three allowable connections. They are ring−ring connection,
ring−linker connection, and ring−side chain connection. On
the other hand, three unallowable connections were defined.
They are linker−linker connection, side chain−side chain
connection, and linker−side chain connection. Figure 4
illustrates the connection rules employed in this work.
For making a molecule, two fragments can be connected
according to the rules only if they have the same bond
multiplicity at both of the connection points. If there are other
connection points that are free for the connection, then implicit
hydrogen atoms are attached to all the connection points to
Figure 3. Preparation of fragment library. An example of complete the structure.
fragmentation to obtain the fragments from a chemical structure. A Mutation. Mutation is a key operation for generating
connection point is described with a dummy atom (labeled by an
asterisk).
individuals for the next generation. In mutation, a parent
molecule is randomly selected from existing molecules, and one
of the following operations is then applied against the parent
of two ring fragments because two phenyl rings do not share molecule.
any atom. A chain or an atom between rings is defined as a • add a fragment
linker fragment, which has two or more attachment points. • remove a fragment
Every acyclic fragment with a single connection point is defined
as a side-chain fragment. • replace a fragment
In this work, Bioactive molecules in GPCR SARfari 2.031 of In the operation add a fragment, a base fragment was
ChEMBL32 (141 990 molecules, 947 914 assays) were used to randomly chosen from fragments that constituted the parent
prepare the fragment library. First, compounds with biological molecule. Then, a new fragment to be introduced was selected
activities indicated by IC50, Ki, Kd, log IC50, log Ki, log Kd, from the fragment library. If the base fragment type is ring, then
pIC50, pKi, or pKd values were retrieved. The subsequent a new ring, linker, or side-chain fragment is selected from the
removal of large molecules (heavy atoms ≥50), duplicated library and linked with a single or a double bond that obeying
molecules, and metal-containing molecules such as ferrocenes valence at the attachment point (Figure 5a). If the base
yielded 97 084 unique molecules with 313 980 assays. Then, fragment is a linker or a side-chain, then a ring fragment was
fragments were extracted from the molecules and duplicated selected and added to the base fragment. In the case of remove a
fragments were removed by comparing canonical SMILES f ragment, a terminal fragment was randomly removed from the
generated by OpenBabel.33 Again, phenyl rings with one, two, parent molecule (Figure 5b). In replace a f ragment, a base
or three connection points in different configurations are fragment was randomly selected and then replaced with a new
treated as different fragments, respectively. Frequent fragments fragment selected from the fragment library. To achieve this
with molecular weights ≤300 were extracted, and finally, 1527 operation, the fragment types and bond orders at the
rings, 605 linkers, and 471 side-chains were prepared as a attachment point must be consistent (Figure 5c).

Figure 4. Fragment connection rules for generating molecules with chemical validity. Allowd connections (a−c) and prohibited connections (d−f).

51 dx.doi.org/10.1021/ci400418c | J. Chem. Inf. Model. 2014, 54, 49−56

Journal of Chemical Information and Modeling Article

Selection. Population of the next generation (new parent

molecules) was selected from both the parent and the offspring.
A tournament method was used to determine the surviving
individuals. First, two individuals were exclusively chosen from
the population of parent and offspring in a random manner.
Then, the individual with the higher fitness value was selected,
and another individual was discarded. This process was applied
to all of the molecules, and finally, half of the population was
selected as surviving individuals.
Molecular Evolution Experiment. Computer experiments
for exploring GPCR ligands were carried out to verify the
feasibility of the present approach. The fragment library for
mutation was prepared using the fragments obtained from
GPCR SARfari. We carried out molecular evolution experi-
ments for different targets (hAA2A and r5HT1A) from GPCR-
SARfari database. They are among largest classes in terms of
the number of the compound entries. The reference structures
for each target were randomly chosen from its own class of
compounds in the database. The number of structures to be
generated in each generation was set to 100. This means that in
every generation, 100 offspring molecules were generated from
100 parent molecules. Among the 200 molecules, 100
molecules were selected as surviving molecules. With respect
to the operators, a mutation rate of 0.8 and crossover rate of 0.2
were used for the evolution parameters. This means that each
individual member of the parent population produces an
Figure 5. Examples of molecular mutations (a−c) and a crossover offspring with a mutation rate of 80% and a crossover rate of
operation (d). Connection position is indicated by an asterisk. 20%. The number of generations to be evolved was set to 500,
and the individuals that survived in the final generation were
Crossover. The crossover operation generates new two referred as the “designed molecules”. The values of the
child molecules by exchanging a fragment set derived from a evolutionary parameters were determined based on our
parent molecule with the other fragment sets derived from preliminary study. The result showed that higher fitness was
another parent molecule. Two parent molecules were randomly obtained when mutation rate was set between 0.7 and 0.8. With
selected, and a crossover point was then randomly selected respect to the number of generations, the fitness reached its
from each parent molecule to define a set of fragments to be plateau within 500 generations. Every computer experiment
was repeated 10 times for each reference structure.

■
crossed over. To achieve crossover, both the fragment type and
bond order at the two crossover points must be the same
(Figure 5d). RESULT AND DISCUSSION
Fitness Evaluation. Our aim of the current work is to Change of the Mean Fitness. First, a computational trial
explore the candidate structures that are similar to a reference of the molecular evolution was carried out to explore the
molecule and yet somewhat different in the scaffolds. For that
reason, the Tanimoto coefficient was used as a fitness function
to evaluate the molecular similarity to the reference molecule.
Surviving compounds were then selected in accordance with
the fitness scores (details are to be described in the Selection
subsection). These evolutionary processes were repeated for
the number of generations specified in advance to yield the
designed structures as an output.
In this study, a particular descriptor (referred to as
topological-fragment-descriptor; TFD) was employed to profile
chemical structures. The TFD was calculated in a manner
similar to that of topological-fragment-spectra method.34,35 For
the first, we enumerated all possible structural fragments that
have the specified number (six atoms in this work) or less
connected atoms excepting hydrogen atoms. Each fragment was Figure 6. Change of the mean fitness for reference molecule 1. The
red shows the total average of the ten trials. The blue line shows that
characterized by its constituent atoms based on atomic type,
for the best molecule, and the green for the worst molecule.
hybridization, and whether the atom contained at least one
(aromatic or aliphatic) ring. Then, each characterized fragment
was hashed into a single integer. The occurrence of individual candidate structures for hAA2A with a reference molecule 1. To
fragments with the same characteristic value was then counted examine the performance of the current approach, the change
to generate a numerical vector. Every chemical structure was of the mean fitness was measured.
described as a multidimensional numerical pattern vector by The mean fitness of each population was calculated for the
means of the TFD method. 100 surviving molecules. As mentioned above, every experi-
52 dx.doi.org/10.1021/ci400418c | J. Chem. Inf. Model. 2014, 54, 49−56
Journal of Chemical Information and Modeling Article

structures of 0.11 to the final molecules of 0.84 for the

reference molecule 1.
We tried another computational experiment to explore the
candidate structures for the other target, r5HT1A. The mean
fitness curve is similar to the above case was observed for the
reference molecule 2 (Figure 7). Fitness of the best molecules
and worst molecules are also plotted in Figures 6 and 7. The
fitness of the best molecules in each generation also increased
as the number of generations increased. These results suggest
that our evolutionary approach was successful for exploring the
candidate molecules in a huge chemical space. For the
computational trial, creating 500 generations of an evolutionary
Figure 7. Change of the mean fitness for the reference molecule 2. experiment with one CPU core required approximately 75 min.
The red shows the total average of the ten trials. The blue line shows Ligand Design for hAA2A. Figure 8 shows part of the
that for the best molecule, and the green for the worst molecule. results of the hAA2A ligand design, which were obtained by the
current evolutionary approach. We investigated the individual
scaffolds of the designed molecules in terms of chemical graph
(CG) expression for nonterminal vertex graph (NTG).36 An
ment was repeated 10 times for each reference structure. Figure NTG is defined as a graph that has no terminal vertex and no
6 shows the total average of them obtained from those trials. isolated vertex. In the design experiment, 100 molecules with
The mean fitness favorably increased as the number of 11 NTG scaffolds including molecules 3−6 are designed using
generations increased, and the fitness reached its plateau within the seed fragments derived from the reference molecule 1.
500 generations. For example, fitness increased from initial Among them, many molecules (44 molecules including 3 and

Figure 8. Result of hAA2A ligand design. All of the designed molecules are taken from the 5th trial. The NTG (NTG/CG) of a molecule is specified
by bold atoms and bold bonds.

53 dx.doi.org/10.1021/ci400418c | J. Chem. Inf. Model. 2014, 54, 49−56

Journal of Chemical Information and Modeling Article

Figure 9. Successful examples of r5HT1A ligand design. Structures with known NTG (NTG/CG) were successfully designed from seed fragments.
The NTG moiety is specified by bold atoms and bold bonds.

Table 1. Result of Scaffold Analysis Using NTG (NTG/CG) was designed from a reference of 15, and compound 17 that
a shares same NTG was found as a known active molecule. These
GPCR SARfari designed
design examples show the applicability of our proposed
active NTG- method.
target moleculesb NTGsb moleculesb NTGsb sharedc
Chemical Feasibility of Designed Molecules. In Figure
hAA2A 2214 811 2975 610 6
8, molecules 2−5 are the designed molecules with the highest
r5HT1A 3195 1260 4063 865 10
a
fitness, the lowest fitness, and in-between molecules collected
Five reference molecules were used for the design. bDuplicates were from the fifth trial (run 5). The reference molecule used in this
excluded. cNumber of NTG/CGs that are shared in both of the entries case was 1. Chemical feasibility (or chemical validity) of the
of GPCR SARfari and the designed molecules.
designed molecules was examined because the candidate
structures should not include unfavorable structures such as
4) shared the same NTG of reference 1, demonstrating the invalid heterohetero atom bonds that often appeared in our
evolutionary direction of the similarity-based approach. The previous work. In this work, we introduced a fragment library
notable points of the design are the following: (1) NTG for the mutation operation to avoid the problem. The
scaffolds of the designed molecule 5 and a known hAA2A connection rules for the fragments defined in the present
ligand 7 are same. (2) NTG scaffolds of 5 and the reference method may also play an important role to improve the
molecule 1 are different. The structural difference between 5
performance. The designed molecules are highly similar to the
and 7 is only three methyl groups on the furan rings and the
reference molecule. The matter is obvious from the visual
amide linker. As mentioned, a new molecule with similar but
inspection of Figure 8 as well. In particular, the scaffolds of
different scaffold could be successfully designed from simple
seed fragments. It should also be noted that molecules with compounds 3 and 4 (shown by bold atoms and bold bonds) are
lower fitness are to be worthy of remark. For example, the the same as the scaffold of the reference molecule.
fitness of the molecule 5 is 0.73. Scaffold Variation of Designed Molecules. We
Ligand Design for r5HT1A. Figure 9 shows part of the investigated scaffold variation of the designed molecules
results of the r5HT1A ligand design. The designed molecules obtained from the molecular evolution experiment. Again, the
for r5HT1A were compared with the active molecules of GPCR chemical graph (CG) representation of the nonterminal vertex
SARfari as well. Some successful examples are shown in Figure graph (NTG) was used to define the scaffold. As shown in
9. When 2 was used as a reference, molecule 8 was obtained as Figure 8, the scaffold of the designed molecule 3 is the same as
one of the designed molecules. NTG scaffolds of 2 and 8 are that of the reference molecule 1, but the scaffold of the
different from each other. Compound 9 was identified from molecule 5 is different from 1. The number of unique
GPCR SARfari as a similar molecule of 8. The structural molecules and unique scaffolds are summarized in Table 1.
difference between 8 and 9 was only a methoxy group of the The results clearly show that a large number of unique
phenyl ring. Compound 10 was also designed from the molecules with a variety of the scaffolds were produced by the
reference 2. Compound 11 was identified from the database current molecular evolutions. The ratios of the number of
that has the same NTG of 10. Compound 13 was designed unique molecules to the number of unique scaffolds were 4.88
from a reference of 12, and we were able to find 14 that for hAA2A and 4.70 for r5HT1A, respectively. This means that
perfectly matched a molecule in the database. Compound 16 the designed molecules that shared the same scaffold are less
54 dx.doi.org/10.1021/ci400418c | J. Chem. Inf. Model. 2014, 54, 49−56
Journal of Chemical Information and Modeling Article

than five in average. For a comparison, the number of unique Notes

NTGs that appeared in the active molecules in GPCR SARfari The authors declare no competing financial interest.
is shown in Table 1, too.
The number of NTG scaffolds that shared in both of the
designed molecules and those of GPCR SARfari are also
■ ACKNOWLEDGMENTS
We thank Dr. Teruki Honma (RIKEN) for his helpful
summarized in Table 1. It shows that, for the case of hAA2A, 6 comments and discussions. We also thank the reviewers for
of the 811 NTGs appeared in GPCR SARfari’s molecules were their helpful comments to improve the manuscript.

■
successfully designed by our approach without any special
consideration. In other words, six known (validated) scaffolds ABBREVIATIONS
and 604 new scaffolds were produced during the current
molecular evolution for the ligand design. For the case of QSAR, quantitative structure−activity relationship; QSPR,
r5HT1A, 10 known scaffolds and 855 new scaffolds were quantitative structure−property relationship; GPCR, G pro-
produced during the molecular evolution with the reference tein-coupled receptor; r5HT1A, rat 5-hydroxytryptamine
receptor 1A; hAA2A, human adenosine receptor A2a

■
molecules.
Comparison with Other Methods. The performance of
the current approach was compared with other methods. We REFERENCES
compared with two recent works, NovoFLAP19 and Flux,15 (1) Jenkins, J. L.; Glick, M.; Davies, J. W. A 3D similarity method for
because they reported the chemical structures of both of scaffold hopping from known drugs or natural ligands to new
reference molecules and designed molecules. Here, we focused chemotypes. J. Med. Chem. 2004, 47, 6144−6159.
on molecular similarity and medicinal chemistry viewpoint. (2) Oyarzabal, J.; Howe, T.; Alcazar, J.; Andrés, J. I.; Alvarez, R. M.;
Dautzenberg, F.; Iturrino, L.; Martínez, S.; Van der Linden, I. Novel
First, the study was performed using the reference molecules of approach for chemotype hopping based on annotated databases of
CP99994 (s1) and ICI (s6).19 Then, another study was chemically feasible fragments and a prospective case study: new
performed using the reference molecules of Gleevec (s11) and melanin concentrating hormone antagonists. J. Med. Chem. 2009, 52,
a Factor Xa inhibitor (s14).15 The designed molecules with the 2076−2089.
highest fitness are shown in Supporting Information Figures S1 (3) Bahl, A.; Barton, P.; Bowers, K.; Caffrey, M. V.; Denton, R.;
and S2. When CP99994 (s1) was used as a reference, s2 was Gilmour, P.; Hawley, S.; Linannen, T.; Luckhurst, C. A.; Mochel, T.;
designed as the best molecule and s3 was designed as the Perry, M. W.; Riley, R. J.; Roe, E.; Springthorpe, B.; Stein, L.;
second best molecule. The designed molecule s2 is very similar Webborn, P. Scaffold-hopping with zwitterionic CCR3 antagonists:
to s1; the difference is only the substitution position of the identification and optimization of a series with good potency and
methoxy group. The difference between s1 and s3 was the size pharmacokinetics leading to the discovery of AZ12436092. Bioorg.
Med. Chem. Lett. 2012, 22, 6694−6699.
of the central heteroring, in which such a design is not shown in (4) Stellwagen, J. C.; Adjabeng, G. M.; Arnone, M. R.; Dickerson, S.
the literature. This type of designed molecule is medicinally H.; Han, C.; Hornberger, K. R.; King, A. J.; Mook, R. A., Jr.; Petrov, K.
relevant because of an empirical knowledge that reducing the G.; Rheault, T. R.; Rominger, C. M.; Rossanese, O. W.; Smitheman, K.
ring size may improve metabolic stability.37 In the case of ICI N.; Waterson, A. G.; Uehling, D. E. Development of potent B-
(s6), candidate molecules which have new molecular frame- RafV600E inhibitors containing an arylsulfonamide headgroup. Bioorg.
works (s7, s8) were produced by connecting the known Med. Chem. Lett. 2011, 21, 4436−4440.
fragments in novel ways. Although it is difficult to strictly (5) Rheault, T. R.; Stellwagen, J. C.; Adjabeng, G. M.; Hornberger, K.
compare the performance or the quality of different methods, R.; Petrov, K. G.; Waterson, A. G.; Dickerson, S. H.; Mook, R. A.;
the result shows at least that the similar and medicinally Laquerre, S. G.; King, A. J.; Rossanese, O. W.; Arnone, M. R.;
relevant analogues were successfully designed by our method. Smitherman, K. N.; Kane-Carson, L. S.; Han, C.; Moorthy, G. S.;

■
Moss, K. G; Uehling, D. E. Discovery of dabrafenib: a selective
inhibitor of Raf kinases with antitumor activity against B-Raf-driven
CONCLUSIONS tumors. ACS Med. Chem. Lett. 2013, 4, 358−362.
We reported a similarity-driven simple evolutionary approach (6) Nishibata, Y.; Itai, A. Automatic creation of drug candidate
to producing candidate molecules for drug design and structures based on receptor structure. Starting point for artificial lead
discovery. The method makes it possible to produce candidate generation. Tetrahedron 1991, 47, 8885−8990.
molecules that are similar to the reference molecule and yet (7) Bohm, H. J. The computer program LUDI: A new method for
the de novo design of enzyme inhibitors. J. Comput.-Aided Mol. Des.
somewhat different in not only side chains but also their 1992, 6, 61−78.
scaffolds. And it is also expected that those candidate structures (8) Gillet, V.; Johnson, A. P.; Mata, P.; Sike, S.; Williams, P.
are chemically feasible. The method was implemented on a SPROUT: A program for structure generation. J. Comput.-Aided Mol.
software tool and validated with the computer experiments for Des. 1993, 7, 127−153.
the GPCR-related ligand design using our own fragment library (9) Douguet, D.; Munier-Lehmann, H.; Labesse, G.; Pochet, S.
prepared from GPCR SARfari. LEA3D: a computer-aided ligand design for structure-based drug

■
design. J. Med. Chem. 2005, 48, 2457−2468.
ASSOCIATED CONTENT (10) Wang, R.; Gao, Y.; Lai, L. A multi-purpose program for
structure-based drug design. J. Mol. Model. 2000, 6, 498−516.
*
S Supporting Information (11) Yuan, Y.; Pei, J.; Lai, L. LigBuilder 2: A practical de novo drug
Results of comparison study (Figures S1 and S2). This material design approach. J. Chem. Inf. Model. 2011, 51, 1083−1091.
is available free of charge via the Internet at http://pubs.acs.org. (12) Vinkers, H. M.; de Jonge, M. R.; Daeyaert, F. F.; Heeres, J.;

■
Koymans, L. M.; van Lenthe, J. H.; Lewi, P. J.; Timmerman, H.; Van
AUTHOR INFORMATION Aken, K.; Janssen, P. A. SYNOPSIS: synthesize and optimize system in
silico. J. Med. Chem. 2003, 46, 2765−2773.
Corresponding Author (13) Schneider, G.; Lee, M. L.; Stahl, M.; Schneider, P. De novo
*Tel.: +81-75-594-0787. Fax: +81-75-594-0790. E-mail: kawai_ design of molecular architectures by evolutionary assembly of drug-
kentaro@kaken.co.jp. derived building blocks. J. Comput.-Aided Mol. Des. 2000, 14, 487−494.

55 dx.doi.org/10.1021/ci400418c | J. Chem. Inf. Model. 2014, 54, 49−56

Journal of Chemical Information and Modeling Article

(14) Brown, N.; McKay, B.; Gilardoni, F.; Gasteiger, J. A graph-based (35) Kawai, K.; Fujishima, S.; Takahashi, Y. Predictive activity
genetic algorithm and its application to the multiobjective evolution of profiling of drugs by topological-fragment-spectra-based support
median molecules. J. Chem. Inf. Comput. Sci. 2004, 44, 1079−1087. vector machines. J. Chem. Inf. Model. 2008, 48, 1152−1160.
(15) Fechner, U.; Schneider, G. Flux (1): A virtual synthesis scheme (36) Takahashi, Y. Chemical data mining based on non-terminal
for fragment-based de novo design. J. Chem. Inf. Model. 2006, 46, 699− vertex graph. In 2004 IEEE International Conference on Systems, Man
707. and Cybernetics, The Hague, Oct 10−13; IEEE: Piscataway, NJ, 2004;
(16) Fechner, U.; Schneider, G. Flux (2): Comparison of molecular Vol. 5, pp 4583−4587.
mutation and crossover operators for ligand-based de novo design. J. (37) St. Jean, D. J., Jr.; Fotsch, C. Mitigating heterocycle metabolism
Chem. Inf. Model. 2007, 47, 656−667. in drug discovery. J. Med. Chem. 2012, 55, 6002−6020.
(17) Lameijer, E. W.; Kok, J. N.; Bäck, T.; Ijzerman, A. P. The
molecule evoluator. An interactive evolutionary algorithm for the
design of drug-like molecules. J. Chem. Inf. Model. 2006, 46, 545−552.
(18) Huang, Q.; Li, L. L.; Yang, S. Y. PhDD: a new pharmacophore-
based de novo design method of drug-like molecules combined with
assessment of synthetic accessibility. J. Mol. Graph. Model. 2010, 28,
775−787.
(19) Damewood, J. R., Jr.; Lerman, C. L.; Masek, B. B. NovoFLAP: A
ligand-based de novo design approach for the generation of
medicinally relevant ideas. J. Chem. Inf. Model. 2010, 50, 1296−1303.
(20) Kawai, K.; Yoshimaru, K.; Takahashi, Y. Generation of target-
selective drug candidate structures using molecular evolutionary
algorithm with SVM classifiers. J. Comput. Chem. Jpn. 2011, 10, 79−87.
(21) Gillet, V.; Johnson, A. P.; Mata, P.; Sike, S.; Williams, P.
SPROUT: A program for structure generation. J. Comput.-Aided Mol.
Des. 1993, 7, 127−153.
(22) Eisen, M. B.; Wiley, D. C.; Karplus, M.; Hubbard, R. E. HOOK:
A program for finding novel molecular architectures that satisfy the
chemical and steric requirements of a macromolecule binding site.
Proteins 1994, 19, 199−221.
(23) Bohacek, R. S.; McMartin, C. Multiple highly diverse structures
complementary to enzyme binding sites: Results of extensive
application of de novo design method incorporating combinatorial
growth. J. Am. Chem. Soc. 1994, 116, 5560−5571.
(24) Pearlman, D. A.; Murcko, M. A. CONCERTS: Dynamic
connection of fragments as an approach to de novo ligand design. J.
Med. Chem. 1996, 39, 1651−1663.
(25) Beccari, A. R.; Cavazzoni, C.; Beato, C.; Costantino, G. LiGen:
A High Performance Workflow for Chemistry Driven de Novo Design.
J. Chem. Inf. Model. 2013, 53, 1518−1527.
(26) Schneider, G.; Fechner, U. Computer-based de novo design of
drug-like molecules. Nat. Rev. Drug Discov. 2005, 4, 649−663.
(27) Dey, F.; Caflisch, A. Fragment-based de novo ligand design by
multiobjective evolutionary optimization. J. Chem. Inf. Model. 2008, 48,
679−690.
(28) Wang, J.; Krudy, G.; Xie, X. Q.; Wu, C.; Holland, G. Genetic
Algorithm-Optimized QSPR Models for Bioavailability, Protein
Binding, and Urinary Excretion. J. Chem. Inf. Model. 2006, 46,
2674−2683.
(29) Jones, G.; Gao, Y.; Sage, C. R. Elucidating molecular overlays
from pairwise alignments using a genetic algorithm. J. Chem. Inf. Model.
2009, 49, 1847−1855.
(30) Walters, W. P.; Green, J.; Weiss, J. R.; Murcko, M. A. What do
medicinal chemists actually make? A 50-year retrospective. J. Med.
Chem. 2011, 54, 6405−6416.
(31) GPCR SARfari. https://www.ebi.ac.uk/chembl/sarfari/
gpcrsarfari (accessed Mar 8, 2013).
(32) Gaulton, A.; Bellis, L. J.; Bento, A. P.; Chambers, J.; Davies, M.;
Hersey, A.; Light, Y.; McGlinchey, S.; Michalovich, D.; Al-Lazikani, B.;
Overington, J. P. ChEMBL: a large-scale bioactivity database for drug
discovery. Nucleic Acids Res. 2012, 40, D1100−1107.
(33) O’Boyle, N. M.; Banck, M.; James, C. A.; Morley, C.;
Vandermeersch, T.; Hutchison, G. R. Open Babel: An open chemical
toolbox. J. Cheminf. 2011, 3, 33.
(34) Takahashi, Y.; Ohoka, H.; Ishiyama, Y. Structural Similarity
Analysis Based on Topological Fragment Spectra. In Advances in
Molecular Similarity; Carbo, R., Mezey, P., Eds.; JAI Press: Greenwich,
CT, 1998; Vol. 2, pp 93−104.

56 dx.doi.org/10.1021/ci400418c | J. Chem. Inf. Model. 2014, 54, 49−56