Drugs by numbers: Reaction-driven de novo design of potent and selective anticancer leads

Article abstract of DOI:10.1002/anie.201206897

A potent and selective inhibitor of the anticancer target Polo-like kinase 1 was found by computer-based molecular design. This type II kinase inhibitor was synthesized as suggested by the design software DOGS and exhibited significant antiproliferative effects against HeLa cells without affecting nontransformed cells. The study provides a proof-of-concept for reaction-based de novo design as a leading tool for drug discovery. Copyright

Full text of DOI:10.1002/anie.201206897

.
Angewandte
Communications
DOI: 10.1002/anie.201206897
Computer-Based Drug Design
Drugs by Numbers: Reaction-Driven De Novo Design of Potent and
Selective Anticancer Leads**
Birgit Spꢀnkuch, Sarah Keppner, Lisa Lange, Tiago Rodrigues, Heiko Zettl, Christian P. Koch,
Michael Reutlinger, Markus Hartenfeller, Petra Schneider, and Gisbert Schneider*
In light of recent analyses of drug development,[1,2] new lead
compounds with well-defined pharmacological activity pro-
files are urgently sought.[3,4] Computer-based de novo design
has been suggested as a method of choice to meet this need, as
it generates innovative molecular scaffolds and chemotypes
by accessing virtually infinite chemical space.[5–7] Here we
present the successful application of fully automated chemis-
try-driven de novo design to discover an innovative low-
molecular-weight inhibitor that selectively blocks inactive
human Polo-like kinase 1 (hPlk1) with nanomolar potency.
This potential anticancer compound was generated by the
algorithm “from scratch” and synthesized following the exact
reaction scheme suggested by our software. It reduced cancer-
cell proliferation without affecting the vitality of nontrans-
formed cells, and exhibited no inhibitory effects against
a panel of activated kinases. The computationally designed
compound is a derivative of the antidepressant fluoxetine, for
which we observed a similar but weaker cellular response
profile. This study provides proof-of-concept for de novo
design as a leading tool for generating novel chemotypes in
the absence of a structural model of the target protein and
with minimal experimental effort.
active conformation of the kinase, type II inhibitors block the
inactive kinase, in which the enzymeꢀs activation loop is in the
so-called “DFG-out” conformation.[10,11] We have recently
identified compound 1 as a strong (IC50 = 0.2 nm) inhibitor of
inactive hPlk1.[12] With compound 1 as a design template, we
employed our newly developed software DOGS (Design Of
Genuine Structures) for generating alternative chemotypes
that mimic the pharmacophoric features of the template
molecule but contain structurally distinct scaffolds.[13,14] The
software constructs molecules by applying a set of 83 chemical
reactions to a stock of over 25000 readily available molecular
building blocks.[15] Compounds are prepared in silico through
iterative application of motivated synthetic reactions.[16] In
each step along the virtual construction path, conserving the
pharmacophoric features and maintaining the structural
similarity of the growing ligand candidate to a template
structure (here compound 1) guide compound prioritiza-
tion.[17] Among the best designs, we identified compound 4 as
a chemotype extending the structural diversity of known
kinase inhibitors (Figure 1A). Surprisingly, this chemically
attractive de novo designed compound is a structural ana-
logue of fluoxetine, a well-known antidepressant.[18,19] This
observation points to functional similarities between the two
compounds, which motivated us to investigate compound 4 in
more detail.
The serine–threonine kinase hPlk1 plays a central role in
cell cycle control and is a target for the development of novel
cancer therapeutics.[8,9] While type I inhibitors bind to an
The DOGS software produced a total of 218 compounds
from 100 preferred starting fragments. The designs are
computed to have druglike properties and are synthetically
plausible (mean Æ s: molecular weight = 457 Æ 59 Da, lip-
ophilicity (SlogP) = 4.5 Æ 1.1, aqueous solubility (logS) =
À5.9 Æ 1.3, synthetic plausibility (rsynth) = 0.7 Æ 0.3). Hop-
kinsꢀ quantitative estimate of druglikeness (QED)[20] for the
set of designed compounds (QED = 0.5 Æ 0.2) is in agreement
with the average value of 0.49 obtained for approved drugs.[21]
In total, 57 different molecule scaffolds were generated, with
57% of the designs containing one of the ten most frequent
scaffolds (Figure 1B). This broad scaffold diversity reflects
the permissive pharmacophoric similarity measure that was
applied during the design process. Among the best ranking
designs we observed several branched structure motifs
although the backbone of the template 1 is linear. In
a preliminary study we had synthesized a de novo designed
compound with a linear scaffold from this series, which
exhibited pronounced inhibitory activity against inactive
hPlk1.[14] Here, we focused on the branched compound 4.
We computed a fitness landscape representing a probabil-
istic model of the structure–activity relationships of known
kinase inhibitors (Figure 1C). The landscape represents
a visualization of the distribution of 12647 bioactive com-
[*] Dr. T. Rodrigues, Dr. H. Zettl, C. P. Koch, M. Reutlinger,
Dr. M. Hartenfeller, Dr. P. Schneider, Prof. Dr. G. Schneider
Departement Chemie und Angewandte Biowissenschaften
Eidgençssische Technische Hochschule (ETH)
Wolfgang-Pauli-Strasse 10, 8093 Zꢀrich (Switzerland)
E-mail: gisbert.schneider@pharma.ethz.ch
Homepage: http://www.modlab.ethz.ch
Priv.-Doz. Dr. B. Spꢁnkuch, Dr. S. Keppner, L. Lange
Universitꢁtsfrauenklinik, Molekulare Onkologie und Gynꢁkologie
Eberhard Karls Universitꢁt
Calwerstrasse 7, 72076 Tꢀbingen (Germany)
[**] We thank Dr. M. Bieler for computing QED values and Sarah Haller
for technical support. Dr. T. Geppert performed the computational
docking experiment. Dr. M. Rupp contributed the ISOAK similarity
function to DOGS. This research was financially supported by the
Deutsche Krebshilfe (grant no. 108651), the Wilhelm-Sander-
Stiftung (grant no. 2009.024.1), the Messer-Stiftung (to B.S.), the
Swiss National Science Foundation (grant no. 205321-134783), and
the Deutsche Forschungsgemeinschaft (grant no. FOR1406TP4) (to
G.S.). B.S. is grateful to Dr. D. Wallwiener for providing working
facilities. G.S. is grateful to the Chemical Computing Group, Inc.
(Montreal, Canada) for a research license of MOE.
Supporting information for this article (including full experimental
details) is available on the WWW under http://dx.doi.org/10.1002/
anie.201206897.
4676
ꢀ 2013 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Angew. Chem. Int. Ed. 2013, 52, 4676 –4681

Angewandte
Chemie
between known and
designed molecules
further prompted us
to select it as a prefera-
ble
candidate
for
detailed investigation.
We also docked
compound 4 and fluox-
etine into the active
site of hPlk1 to obtain
a
preliminary struc-
tural model of the
enzyme–ligand com-
plex. Automated flexi-
ble docking into the
ATP binding cavity of
active hPlk1 (Protein
Data Bank,[28] PDB
ID:
model of inactive
2ou7)[29]
and
a
hPlk1, which we had
generated from an
inactive DFG-out con-
formation of Aurora A
kinase
(PDB
ID:
2c6e),[14,30] suggests the
preferred binding of
compound
inactive
4
to the
enzyme
(GOLD
software,
ASP score[31,32] = 40 vs.
47; greater positive
values indicate poten-
tially stronger bind-
ing). Fluoxetine dock-
ing yielded overall
lower scores for the
predicted ligand–pro-
tein complexes (active
hPlk1: 29; inactive
hPlk1: 37) but exhib-
ited the same trend.
These modeling results
Figure 1. A) De novo design and synthesis of compound 4: Reagents and conditions: 1) 2-pyridinecarboxaldehyde,
1,2-dichloroethane (DCE), NaBH(OAc)3, RT; 2) 4-trifluoromethylphenol, PPh3, diethyl azodicarboxylate (DEAD),
THF, 08C. The designed compound is a derivative of fluoxetine. B) Scaffold diversity: The 10 most frequently
generated scaffold structures among the computer-designed compounds (n: number of molecules). C) Artificial
fitness landscape: The plot presents the distribution of known kinase inhibitors (black dots) and the de novo
designed compounds (red dots) in the pharmaceutically relevant chemical space spanned by 12647 diverse druglike
bioactive compounds (COBRA database). Overall compound density is indicated by a color gradient from blue (few
data points) to orange (many data points). For visualization all molecules were represented by topological
pharmacophoric features (CATS descriptor) and projected onto two dimensions by stochastic neighbor embedding
(SNE).
pounds affecting 980 macromolecular targets,[22] together with
our 218 de novo designed molecules. For landscape modeling
we represented all molecules by a topological pharmacophore
descriptor (CATS)[23] and projected the resulting 210-dimen-
sional real-valued fingerprints onto the plane by stochastic
neighbor embedding.[24] We employed this nonlinear projec-
tion method to preserve and focus on the local neighborhood
behavior[25] and reduce the risk of projecting artifacts from
outliers and globally non-Gaussian distribution of data.[26,27]
Apparently, the majority of the designed compounds are
located in an area of chemical space that is adjacent to known
kinase inhibitors. This finding suggests that these two sets of
compounds possess partially overlapping but not identical
pharmacophoric features and the designed molecular entities
were generated beyond the structural diversity of known
kinase inhibitors. The location of compound 4 at the border
suggested that both compound 4 and fluoxetine might
actually bind and prefer inactive over active hPlk1 confor-
mations.
The designed compound was readily amenable to chem-
ical synthesis; we followed the two-step synthetic route
suggested by the software without further optimization
(Figure 1A). The racemic product 4 was obtained after
purification, through sequential reductive amination of 3-
amino-1-phenylpropan-1-ol (2) yielding the intermediate 3,
which underwent Mitsunobu reaction to form the ether
linkage.[33] Hereby we validate the potential of the reaction-
based generation of compounds for medicinal chemistry.
We initially analyzed compounds 1, 4, and fluoxetine for
their inhibitory activity against 48 active kinases. At a con-
centration of 10 mm none of the compounds significantly
interfered with active hPlk1 or any of the other human
Angew. Chem. Int. Ed. 2013, 52, 4676 –4681
ꢀ 2013 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
www.angewandte.org 4677

.
Angewandte
Communications
kinases tested (not shown). It is of special note that the de
novo designed compound 4 did not block the activity of
Aurora kinase A (STK6), which interplays with hPlk1 effects
in vivo.[34]
Subsequently, we performed kinase assays using
synchronized HeLa cells to analyze whether compound 4
inhibits hPlk1 kinase in its inactive form. Immunoprecipitated
hPlk1 was subjected to kinase assays with increasing concen-
trations of compound 4 and with casein as the substrate. We
observed significantly reduced hPlk1 activity to levels of 40–
60% for all concentrations of compound 4 (Figure 2A, 1 nm:
55%, p = 0.0031 (two-way ANOVA with Bonferroni correc-
tion) 10 nm: 60%, p = 0.0006; 100 nm: 46%, p = 0.007; 1 mm:
54%, p = 0.001; 33 mm: 41%, p = 0.01). We also determined
hPlk1 activity after incubation of immunoprecipitated hPlk1
with fluoxetine, which exhibited significant inhibition albeit
only at 100-fold higher concentrations of 10 and 100 mm
(Figure 2B). At a concentration of 10 nm, fluoxetine reduced
hPlk1 kinase activity to 80% (p = 0.08), 100 nm to 90% (p =
0.38), 1 mm to 71% (p = 0.09), 10 mm to 61% (p = 0.03), and
100 mm to 37% (p = 0.0003). These findings indicate that
fluoxetine is indeed a moderate inhibitor of inactive hPlk1,
and this might actually account for some of the reported
antiproliferative effects of fluoxetine on cancer cells.[35,36]
As an additional test for target selectivity, we performed
an inhibition assay with compound 4 and inactive Aurora A
kinase. No reduction of kinase activity was observed up to
a ligand concentration of 33 mm (Figure 2C). Thus, because
the obtained novel kinase inhibitor exhibits nanomolar
affinity to the inactive state of hPlk1 and favorable estimated
ligand efficiency[37] (DG/number of heavy atoms ꢀ 0.66), it
represents a promising starting point for optimization.
To study the effect of compound 4 and fluoxetine on cell
cycle progression, we performed FACScan analysis with
HeLa cells after double-thymidine block in the presence of
either one of these compounds. When we monitored the cell
cycle over 48 h we observed strongly delayed progression
throughout all phases (Figure S1A in the Supporting Infor-
mation). Four hours after release, one-third of the control
cells were in the G0/G1 phase, one-third in the S phase, and
one-third in the G2/M phase, while cells treated with 10 mm
compound 4 were still mainly in the S phase; when treated
with 33 mm compound 4 almost all cells were locked in the S
phase (Figure S1A, first panel). After six hours, the control
cells reached the G2/M phase, but cells treated with 33 mm
compound 4 were still in the S phase (Figure S1A, second
panel). After eight hours, the cells treated with 10 mm
compound 4 displayed approximately the same cell cycle
distribution as the control cells, while cells treated with 33 mm
compound 4 were still delayed in the S phase (Figure S1A,
third panel). Ten to fourteen hours after release, the control
cells and the cells treated with 10 mm compound 4 all reached
normal cell cycle distribution (Figure S1A, fourth to sixth
panel), and cells that had been treated with 33 mm compound
4 completed S phase and arrived in G2/M phase. Within 24–
48 h after release, the control cells had completed a normal
cell cycle, but cells treated with compound 4 progressed from
G2/M arrest (after 14 h) to apoptosis (Figure S1A, seventh to
bottom panel). Treatment with fluoxetine led to comparable
Figure 2. Inhibition of inactive hPlk1 and Aurora A kinase. Kinase
assays using immunoprecipitated hPlk1 from HeLa cells after incuba-
tion with compound 4 (A) and fluoxetine (B), or compound 4 together
with immunoprecipitated human Aurora A kinase from HeLa cells (C).
The autoradiograms show representative assays, and the bar graphs
show means of three independent experiments measuring hPlk1 or
Aurora A kinase activity after immunoprecipitation and incubation with
the two compounds, respectively, followed by subsequent kinase assay
using casein as a substrate.
4678 www.angewandte.org
ꢀ 2013 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Angew. Chem. Int. Ed. 2013, 52, 4676 –4681

Angewandte
Chemie
Figure 3. A–D) Cell proliferation and viability. Effect of compound 4 (A) or fluoxetine (B) on cell proliferation and wash-out grow-out assay in
HeLa cells, and effect of compound 4 and fluoxetine in wash-out grow-out assays in hTERT-RPE1 cells (C, D). Cell proliferation of HeLa cells was
analysed 24 to 72 h after treatment. Control cells were incubated with culture medium alone. Percentage of surviving cells is given as percentage
of the number of control cells 72 h after incubation with the drugs. All experiments: n=3. Additionally, HeLa cells (A, B) and hTERT-RPE1 cells
(C, D) were analysed using wash-out grow-out assays after treatment with compound 4 (A, C) or fluoxetine (B, D). Cells were treated with
respective compound for 72 h, harvested and one fifth (HeLa cells) or one third (hTERT-RPE1 cells), respectively, was re-seeded in new six-well-
plates with fresh medium without drugs. (E–H) Induction of cellular apoptosis. Western blot analysis of Parp cleavage after treatment of HeLa
cells with compound 4 (E) or fluoxetine (F). To determine the full-length Parp protein and the cleavage product in apoptotic cells, Western blot
analyses targeting Parp were performed in HeLa cells 48 h after incubation. Caspase 3/7 activation was monitored 24 h (G) and 48 h (H) after
treatment with compound 4. Luminescence is given as relative RLU levels (n=3, mean Æs). Control cells were incubated with normal culture
medium.
Angew. Chem. Int. Ed. 2013, 52, 4676 –4681
ꢀ 2013 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
www.angewandte.org 4679

.
Angewandte
Communications
effects as compound 4 concerning cell cycle progression of
HeLa cells over the period of 48 h after release from the
double thymidine block (Figure S1B). Cells showed delayed
progression in cell cycle and, as observed for compound 4,
they exited from G2/M phase into apoptosis at later points in
time.
another 72 h period. Both compounds induced slightly
impaired recovery, yet all cells were viable and able to
proliferate reseeding into fresh medium.
To analyze the induction of apoptosis, we first performed
Western blot analyses against Parp in HeLa cells. We were
able to detect an increasing amount of cleavage products of
85 kDa, accompanied by a decrease of the full-length protein
of 116 kDa, after treatment with compound 4, starting at
concentrations of 100 nm up to 33 mm, where completely
degraded full-length protein was observed (Figure 3E). After
incubation with fluoxetine the full-length protein was com-
pletely cleaved first at a drug concentration of 33 mm, but at
lower drug concentrations we observed only a slight increase
of cleaved Parp fragment compared to control cells (Fig-
ure 3F). To confirm the induction of apoptosis after treatment
with compound 4 we performed Caspase 3/7 assays after
incubation times of 24 and 48 h (Figure 3G,H); we found
concentration- and time-dependent Caspase 3/7 activation
indicative of apoptosis.
The most relevant outcome of this study: de novo
designed compound 4 proved to be a synthetically feasible,
druglike hPlk1 inhibitor with high in vitro potency and
selectivity, antiproliferative activity against cancer cells, and
with minimal effect on immortalized nontransformed cells.
The results confirm our concept of reaction-driven, template-
based de novo design as a premier methodology for the rapid
identification of novel bioactive molecules exhibiting
a desired biological activity spectrum. This approach may
well provide an opportunity to jumpstart stalled drug-discov-
ery projects.
To determine the impact of compound 4 and fluoxetine on
proliferation we counted cells in the period 24–72 h after
treatment. Compound 4 induced concentration- and time-
dependent reduction of HeLa cell proliferation with an EC50
value of 4 Æ 1 mm (Figure 3A, upper panel). After 72 h the
reduction was significant for concentrations of 100 nm and
higher (100 nm: 81%, p = 0.03; 1 mm: 60%, p = 0.02; 10 mm:
35%, p = 0.01; 33 mm: 0%, p = 0.002). To analyze whether the
cells were able to recover from the impairment induced by the
inhibitors, we determined cell proliferation a further 72 h
after reseeding the cells in new six-well-plates with fresh
culture medium without drugs (wash-out grow-out assay).
Cells were counted after the first 72 h period, then reseeded
and cell numbers were determined after another 72 h period.
Incubation with fresh medium without compound 4 led to
recovery for the cells that had been first treated with lower
concentrations (10 nm–1 mm), and the recovery was compa-
rable to that of the untreated control (Figure 3A, lower
panel). The increase in cell numbers was approximately 400–
500% compared to the number of the reseeded cells in the
72 h period. In contrast, higher concentrations resulted in
a reduction of recovery for the cells that had been treated
with 10 mm of compound 4 (42%). The cells treated first with
higher concentrations (16 mm and 22 mm) did not recover;
after the additional 72 h period with fresh medium without
drug the relative cell proliferation was 0%. Those cells
treated with 28 and 33 mm compound 4, respectively, were
completely dead at the time of reseeding, thus leading to 0%
cells after the additional 72 h with fresh medium without drug.
Compared to compound 4, fluoxetine induced a weaker
reduction of HeLa cell proliferation over the period of 24 to
72 h, but it was still significant for concentrations exceeding
10 mm (Figure 3B, upper panel). The reduction of cell
proliferation was concentration- and time-dependent with
an EC50 value of 14 Æ 2 mm [10 mm: reduction to 64% (p =
0.02); 33 mm: 1% (p = 0.0001); 66 mm: 0% (p < 0.0001); and
100 mm: 0% (p < 0.0001)]. Wash-out grow-out assays gave
comparable results, with recovery after fluoxetine treatment
similar to those of control cells for concentrations up to 1 mm
(430–570% compared to the reseeded cell number), but
higher concentrations led to impaired recovery (10 mm:
295%) (Figure 3B, lower panel). After treatment with 66 or
100 mm fluoxetine, the cells were completely dead within the
first 72 h, leading to 0% recovery after an additional 72 h with
fresh medium without drug, as observed for compound 4.
We also investigated the impact of compound 4 and
fluoxetine on the recovery of hTERT-RPE1 cells (Fig-
ure 3C,D) to determine their capacity as potential anticancer
drugs. For that reason, hTERT-RPE1 cells were treated with
100 nm and 10 mm compound 4, and with 10 and 33 mm
fluoxetine, respectively. As described for HeLa cells,
hTERT-RPE1 cells were counted after the first 72 h period,
then reseeded and the cell numbers were determined after
Received: August 25, 2012
Published online: November 20, 2012
Keywords: drug design · enzyme inhibitors ·
.
fragment-based design · Polo-like kinase · virtual synthesis
[1] M. E. Bunnage, Nat. Chem. Biol. 2011, 7, 335 – 339.
[2] K. I. Kaitin, J. A. DiMasi, Clin. Pharmacol. Ther. 2011, 89, 183 –
188.
[3] F. Pammolli, L. Magazzini, M. Riccaboni, Nat. Rev. Drug
Discovery 2011, 10, 428 – 438.
[4] Y. L. Bennani, Drug Discovery Today 2011, 16, 779 – 792.
[5] G. Schneider, U. Fechner, Nat. Rev. Drug Discovery 2005, 4,
649 – 663.
[6] M. Hartenfeller, G. Schneider, WIREs Comp. Mol. Sci. 2011, 1,
742 – 759.
[7] G. Schneider, J. Comput. Aided Mol. Des. 2012, 26, 115 – 120.
[8] C. McInnes, M. D. Wyatt, Drug Discovery Today 2011, 16, 619 –
625.
[9] K. Strebhardt, Nat. Rev. Drug Discovery 2010, 9, 643 – 660.
[10] Y. Liu, N. S. Gray, Nat. Chem. Biol. 2006, 2, 358 – 364.
[11] I. Akritopoulou-Zanze, P. J. Hajduk, Drug Discovery Today
2009, 14, 291 – 297.
[12] S. Keppner, E. Proschak, G. Schneider, B. Spꢁnkuch, Chem-
MedChem 2009, 4, 1806 – 1809.
[13] M. Hartenfeller, H. Zettl, M. Walter, M. Rupp, F. Reisen, E.
Proschak, S. Weggen, H. Stark, G. Schneider, PLoS Comput.
Biol. 2012, 8, e1002380.
[14] G. Schneider, T. Geppert, M. Hartenfeller, F. Reisen, A.
Klenner, M. Reutlinger, V. Hꢁhnke, J. A. Hiss, H. Zettl, S.
4680 www.angewandte.org
ꢀ 2013 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Angew. Chem. Int. Ed. 2013, 52, 4676 –4681

Angewandte
Chemie
Keppner, B. Spꢁnkuch, P. Schneider, Future Med. Chem. 2011, 3,
415 – 424.
[26] M. Reutlinger, G. Schneider, J. Mol. Graphics Modell. 2012, 34,
108 – 117.
[27] M. Rupp, P. Schneider, G. Schneider, J. Comput. Chem. 2009, 30,
2285 – 2296.
[15] M. Hartenfeller, G. Schneider, Methods Mol. Biol. 2011, 672,
299 – 323.
[28] H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat,
H. Weissig, I. N. Shindyalov, P. E. Bourne, Nucleic Acids Res.
2000, 28, 235 – 242.
[16] M. Hartenfeller, M. Eberle, P. Meier, C. Nieto-Oberhuber, K.-H.
Altmann, G. Schneider, E. Jacoby, S. Renner, J. Chem. Inf.
Model. 2012, 52, 1167 – 1178.
[29] M. Kothe, D. Kohls, S. Low, R. Coli, A. C. Cheng, S. L. Jacques,
T. L. Johnson, C. Lewis, C. Loh, J. Nonomiya, A. L. Sheils, K. A.
Verdries, T. A. Wynn, C. Kuhn, Y. H. Ding, Biochemistry 2007,
46, 5960 – 5971.
[30] N. M. Heron, Bioorg. Med. Chem. Lett. 2006, 16, 1320 – 1323.
[31] G. Jones, P. Willett, R. C. Glen, A. R. Leach, R. Taylor, J. Mol.
Biol. 1997, 267, 727 – 748.
[32] G. Jones, P. Willett, R. C. Glen, J. Mol. Biol. 1995, 245, 43 – 53.
[33] T. Y. But, P. H. Toy, Chem. Asian J. 2007, 2, 1340 – 1355.
[34] S. M. Lens, E. E. Voest, R. H. Medema, Nat. Rev. Cancer 2010,
10, 825 – 841.
[35] A. Krishnan, R. Hariharan, S. A. Nair, M. R. Pillai, Biochem.
Pharmacol. 2008, 75, 1924 – 1934.
[17] G. Schneider, M. Hartenfeller, M. Reutlinger, Y. Tanrikulu, E.
Proschak, P. Schneider, Trends Biotechnol. 2009, 27, 18 – 26.
[18] D. T. Wong, J. S. Horng, F. P. Bymaster, K. L. Hauser, B. B.
Molloy, Life Sci. 1974, 15, 471 – 479.
[19] D. T. Wong, F. P. Bymaster, E. A. Engleman, Life Sci. 1995, 57,
411 – 441.
[20] G. R. Bickerton, G. V. Paolini, J. Besnard, S. Muresan, A. L.
Hopkins, Nat. Chem. 2012, 4, 90 – 98.
[21] P. Leeson, Nature 2012, 481, 455 – 456.
[22] P. Schneider, G. Schneider, QSAR Comb. Sci. 2003, 22, 713 – 718.
[23] G. Schneider, W. Neidhart, T. Giller, G. Schmid, Angew. Chem.
1999, 111, 3068 – 3070; Angew. Chem. Int. Ed. 1999, 38, 2894 –
2896.
[24] G. Hinton, S. Roweis, Adv. Neural Inf. Proc. Syst. (NIPS), 2002,
15, 857 – 864.
[36] A. Stepulak, W. Rzeski, M. Sifringer, K. Brocke, A. Gratopp, K.
Kupisz, L. Turski, C. Ikonomidou, Cancer Biol. Ther. 2008, 7,
1685 – 1693.
[37] A. L. Hopkins, C. R. Groom, A. Alex, Drug Discovery Today
2004, 9, 430 – 431.
[25] D. Horvath, C. Jeandenans, J. Chem. Inf. Comput. Sci. 2003, 43,
680 – 690.
Angew. Chem. Int. Ed. 2013, 52, 4676 –4681
ꢀ 2013 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
www.angewandte.org
4681