Combining on-chip synthesis of a focused combinatorial library with computational target prediction reveals imidazopyridine GPCR ligands

Article abstract of DOI:10.1002/anie.201307786

Using the example of the Ugi three-component reaction we report a fast and efficient microfluidic-assisted entry into the imidazopyridine scaffold, where building block prioritization was coupled to a new computational method for predicting ligand-target

Full text of DOI:10.1002/anie.201307786

.
Angewandte
Communications
DOI: 10.1002/anie.201307786
Combinatorial Chemistry
Hot Paper
Combining On-Chip Synthesis of a Focused Combinatorial Library
with Computational Target Prediction Reveals Imidazopyridine GPCR
Ligands**
Michael Reutlinger, Tiago Rodrigues, Petra Schneider, and Gisbert Schneider*
Abstract: Using the example of the Ugi three-component
reaction we report a fast and efficient microfluidic-assisted
entry into the imidazopyridine scaffold, where building block
prioritization was coupled to a new computational method for
predicting ligand–target associations. We identified an innova-
tive GPCR-modulating combinatorial chemotype featuring
ligand-efficient adenosine A1/2B and adrenergic a1A/B receptor
antagonists. Our results suggest the tight integration of micro-
fluidics-assisted synthesis with computer-based target predic-
tion as a viable approach to rapidly generate bioactivity-
focused combinatorial compound libraries with high success
rates.
targets.[5–8] While entry into this che-
motype through an Ugi three-com-
ponent
reaction
has
been
reported,[5–7] these methods do not
allow for the quick assembly of
combinatorial libraries and scaling
up. Therefore, our initial efforts
focused on developing a robust and scalable process in flow
using a continuous synthesis system equipped with low-
pressure, pulsation-free syringe pumps. The setup included
a 3-2-way solenoid valve to allow for automated cycles of
building block filling and dispensing. The amine and benzal-
dehyde components were dissolved in ethanol, together with
perchloric acid, while the isocyanide component was pumped
independently. The concentrations of the stock solutions were
adjusted to afford the desired final concentrations in the
microreactor. A borosilicate DeanFlow chip with a total
volume of 5 mL and a zig-zag mixing zone was used as the
primary reactor (Figure 1A). Alternatively, we used a Kom-
T
he fast pace of drug-discovery programs is supported by
high-throughput screening campaigns to identify new chem-
ical entities, where the underlying compound collections used
for screening benefit from combinatorial libraries with lead-
and drug-like properties.[1] While numerous synthesis proto-
cols are available, a reliable assessment of potential macro-
molecular targets of these compounds is desirable for the
compilation of bioactivity-focused combinatorial libraries.
For the Ugi four- and three-component reactions,[2] which
have shown robustness in producing both model compounds
and drug candidates,[3,4] we report a fast and efficient micro-
fluidic-assisted entry into the imidazopyridine scaffold, cou-
pled to a new computational prediction method for “deor-
phaning” ligand–target associations. We identified an inno-
vative GPCR-modulating combinatorial chemotype featuring
ligand-efficient adenosine A1/2B and adrenergic a1A/B receptor
antagonists (GPCR = G-protein coupled receptor). Our
results suggest that the tight integration of microfluidics-
assisted synthesis with computer-based target prediction is
a viable approach to rapidly generate bioactivity-focused
combinatorial compound libraries with high success rates.
Imidazopyridines may be considered to be a privileged
scaffold given their diverse range of macromolecular drug
Figure 1. DeanFlow (A) and KombiMix (B) microreactor chips.
biMix chip with a reaction volume of 13 mL (Figure 1B). The
protocol was then scripted with Cetoni Qmix Elements
software to automate all steps, including the washing of the
microfluidic channels.
In an initial screening of the reaction conditions we
performed sequential and automated syntheses of compound
1. Conversion rates were derived from 1H NMR spectra
(Figure 2A,B). In the first instance we investigated optimal
flow rates and reaction temperatures, using 10 mol% of
catalyst and a final concentration of each building block equal
to 0.3m, as described previously.[6] Generally, the reactions
gave better results at lower temperatures (30 and 708C) than
at 170 and 2008C. Additionally, we conducted control
reactions in glassware at room temperature and 308C for
two hours, and measured conversions of 73% and 80%,
respectively. The results pinpoint the usefulness of a micro-
reactor, both for improving conversions and drastically
shortening reaction times. Reactions carried out under
higher flow rates (30 and 60 mLsꢀ1) gave poorer results than
[*] M. Reutlinger, Dr. T. Rodrigues, Dr. P. Schneider,
Prof. Dr. G. Schneider
Swiss Federal Institute of Technology (ETH)
Department of Chemistry and Applied Biosciences
Wolfgang-Pauli-Strasse 10, 8093 Zurich (Switzerland)
E-mail: gisbert.schneider@pharma.ethz.ch
Dr. P. Schneider, Prof. Dr. G. Schneider
inSili.com GmbH
Segantinisteig 3, 8049 Zurich (Switzerland)
[**] This research was financially supported by a research grant from the
OPO Foundation, Zurich. GPCR=G-protein coupled receptor.
Supporting information for this article is available on the WWW
under http://dx.doi.org/10.1002/anie.201307786.
582
ꢀ 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Angew. Chem. Int. Ed. 2014, 53, 582 –585

Angewandte
Chemie
Figure 2. Synthesis of imidazopyridines in flow: A) Screening of optimal flow rate and temperature (T), at constant catalyst loading (10 mol%)
and building block (BB) concentration (0.3m); B) Screening of optimal catalyst loading and building block concentration, at fixed flow rate
(15 mLsꢀ1) and temperature (70 and 1008C); C) Focused library synthesized in the present study and yields of isolated products.
their 3.75 mLsꢀ1 and 7.5 mLsꢀ1 counterparts, possibly due to
shorter residence times in the reactor chip. We observed the
highest conversion at intermediate temperatures (70 and
1008C). Interestingly, at 708C the reaction appears to tolerate
a wide range of flow rates, while at 1008C a rate of 15 mLsꢀ1 is
preferable.
Having determined the best binary combination of
temperature and flow rate, we screened for the ideal catalyst
loading and final concentration of building blocks and found
these to be 10 mol% and 0.3m, respectively (Figure 2B).
Comparable conversion rates were obtained in a microwave
procedure[6] and the setup described herein (94% vs. 93%,
respectively). Of note, these results were obtained using
a lower reaction temperature in the flow system (1008C in
flow vs. 1708C in the microwave reactor) and shorter reaction
times (0.3 s in flow vs. 15 min in the microwave reactor).
Finally, the optimized reaction conditions were compared in
the DeanFlow and KombiMix microreactors. While the
conversion of 1 in the DeanFlow chip was 93%, an 88%
conversion was observed in the KombiMix chip.
the computer model predicts pAffinity values for each target,
which goes beyond related computational tools.[10] Further-
more, to ensure meaningful, nontrivial, and high-value
predictions we calculated the Mahalanobis distance (MD)
of the predicted values to the predictions made for a large
collection of randomly selected molecules. Here, we consid-
ered only drug targets for which we obtained pAffinity > 5.5
and MD > 0.5 standard deviations. With these mildly restric-
tive criteria we predicted an average of 18 targets per
compound. Basically due to the low pAffinity bound, this
number exceeds other theoretical considerations and exper-
imental findings reporting between two and ten targets per
drug, depending on the target class.[11] We obtained an
average of four targets per imidazopyridine compound with
the more conservative boundaries pAffinity > 6 and MD > 1.
Keeping the permissive estimate we selected a total of 41
targets with high pAffinity predictions for further study. For
these targets the model yielded favorable cross-validated
accuracies of Q2 = 0.68 ꢁ 0.10, MAE = 0.65 ꢁ 0.11 and
BEDROC = 0.67 ꢁ 0.15 (all values mean ꢁ standard devia-
tion).[12]
With these results in hand we synthesized a small focused
library of imidazopyridines 1–12 (Figure 2C) using the Dean-
Flow reactor chip, and predicted potential biological targets
with Gaussian process regression models, which were con-
structed from 469 drug targets that are annotated in the
ChEMBL database (version 14).[9] Given a query compound,
We finally selected five targets based on majority
predictions for the whole library, potential pharmaceutical
interest, and assay availability. pAffinity values were in the
micromolar range (Table 1); even though the predicted
variance was high. This observation emphasizes the potential
Angew. Chem. Int. Ed. 2014, 53, 582 –585
ꢀ 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
www.angewandte.org
583

.
Angewandte
Communications
Table 1: Summary of results for selected compounds 1, 3, 5, 7, 8, 9, and 12.
discovery, such as on-chip synthesis
and computational target predic-
tion, may advance hit and lead
identification in chemical biology
and molecular medicine. In light of
recent advances in lab-on-a-chip
technologies,[14] one could even
envisage a fully automated hit-find-
ing automaton that integrates com-
putational target prediction and
building block selection for the
microfluidic-assisted synthesis and
testing of candidate compounds.
Target
Predicted Mahalanobis
Experimental
LE[a]
LLE[b]
SILE[c]
pAffinity
distance
pKi or % binding
1
3
5
7
8
9
a1A[d]/PDE10A[e]
a1B
5.7/5.7
6.2
5.8/6.5
6.1
5.7
6.4/5.8
6.0
0.7/0.8
2.4
0.7/2.4
2.0
3.2
2.6/1.7
3.3
<4/<4
5.6
–
0.33
–
3.46
–
3.04
[f]
[g]
a1A/A2B
a1B
5.4/5.2
5.7
0.30/0.29 3.07/2.86 2.87/2.76
0.40
–
–
–
1.74
3.23
>80%[i]
<4/<4
>80%[i]
–
–
–
–
–
–
[h]
A1
A2B/PDE10A
12 A1
[a] Ligand efficiency. [b] Lipophilic ligand efficiency. [c] Size-independent ligand efficiency. [d] Adrenergic
a1A receptor. [e] Phosphodiesterase 10A. [f] Adrenergic a1B receptor. [g] Adenosine A2B receptor.
[h] Adenosine A1 receptor. [i] Radioligand assay; activity values are averaged from two measurements.
novelty of the scaffold compared to known ligands in the
ChEMBL database. In fact, to the best of our knowledge,
Experimental Section
Computations. For training the Gaussian process models[15] we used
the ChEMBL database (version 14) containing 1213242 distinct
compounds with 10129256 bioactivities for 9003 targets.[9] Protein
targets with fewer than 200 annotated human bioactivities were
excluded. All activity end-points were standardized to pAffinity =
ꢀlog10(activity). The final affinity data set consisted of 209293
compounds with 431313 bioactivities for 469 human targets. Post-
processing was conducted using Python (http://www.python.org) and
Knime v.2.6.0.[16] Molecular structures were standardized using the
“wash” function in MOE 2012.10 (The Chemical Computing Group
Inc., Montreal, Canada); logP(o/w) was calculated with MOE. Two
different molecular descriptors were calculated for each compound:
topological pharmacophores (CATS2, 0–9 bonds, type-sensitive
scaling),[6] and an ECFP-like topological circular fingerprint
(Morgan fingerprint, radius = 4, 2048 bit; RDKit: http://www.rdkit.
org).[17] Predictive models were implemented using Matlab R2012b
(The MathWorks Inc., Natick, USA) and the GPML toolbox v3.1
(http://www.gaussianprocess.org). We assessed prediction quality by
tenfold stratified cross-validation (cross-validated squared correla-
tion coefficient, Q2; mean absolute error, MAE). The Boltzmann-
enhanced discrimination of ROC (BEDROC; a = 56, top 3%
contribute 80% to the score) was used to quantify the early
enrichment performance.[18] We used the lower confidence-bound
pAffinity estimate throughout this study: prediction ¼ m ꢀ s2 ,
imidazopyridines with this framework have not been reported
as adenosine or adrenergic receptor ligands.[20]
Having predicted potential macromolecular targets for all
synthesized compounds, we tested those compounds for
which we had obtained robust pAffinity predictions. For one
of the prominent targets, phosphoinositide 3-kinase, activity
had previously been reported for the underlying imidazopyr-
idine scaffold,[6] which corroborated the prediction. As
a proof-of-concept, we then explored a range of predicted
GPCR targets aiming at the discovery of a new activity island
in chemical space. In radioligand displacement assays probing
the direct ligand–receptor binding and in cell-based func-
tional activity assays, 71% of the compounds were found to
be active as predicted (Table 1). More specifically, com-
pounds 3 and 7 presented antagonistic Ki values of 2–3 mm,
respectively, against the adrenergic a1B receptor, while com-
pound 5 showed similar low micromolar antagonistic potency
against the adrenergic a1A and adenosine A2B receptors.
Compounds 8 and 12 turned out to be potent direct ligands of
the A1 receptor (84% and 89% binding at 100 mm, respec-
tively), but were inactive in the functional cell-based assay.
Additional tests will be required to determine selectivity
profiles in a full GPCR panel screen.
Several quality indices have been suggested to guide hit
prioritization in drug discovery.[13] Accordingly, our com-
pounds fully qualify as lead structure candidates (Table 1).
For example, compound 7 is a scarcely decorated, yet highly
ligand-efficient chemical entity (LE = 0.40; SILE = 3.23)
which might justify development as an adrenergic a1A
receptor antagonist. On the other hand, although less efficient
than 7, compound 3 presents a better balance between affinity
and computed logP(o/w) (LLE = 3.46 vs. 1.74). Most impor-
tantly, the leads presented herein are dissimilar to their
nearest neighbors from the training data (structural similarity
Tanimoto = 0.16–0.30, Table S1) and would likely not have
been selected using straightforward substructure-based sim-
ilarity searching.
*
*
where m* is the modelꢀs predictive mean and s2 the predictive
*
variance. To distinguish from random predictions we calculated the
Mahalanobis distance of an activity prediction: MD(prediction) =
(predictionꢀmr)/sr, where mr and sr are the mean and standard
deviation of a randomized predictive distribution. The background
consisted of 50000 randomly selected molecules from ChemDB.[19]
Synthesis. Stock solutions of building blocks were prepared in
ethanol. The amine and aldehyde components were premixed, and
perchloric acid was added. Two independent syringe pumps delivered
the amine/benzaldehyde/perchloric acid solution and the isocyanide
solution at suitable flow rates. The reaction chamber containing the
microchip was heated at different temperatures and the crude product
was collected in a vial. The crude mixtures were purified by
preparative HPLC (acetonitrile/H2O + 0.1% formic acid in each
solvent) using a gradient of 30–95% or 5–50% acetonitrile over
16 min. Microfluidics hardware and the Qmix Elements software were
obtained from Cetoni (Korbussen, Germany). Microwave synthesis
was performed in a Biotage Initiator (Uppsala, Sweden) in 1–2 mL
vials, as described.[6]
Altogether, our chemistry-driven approach to the design
of a target-focused combinatorial library, in an expeditious
and efficient manner, led to the identification of a molecular
framework targeting four GPCRs. The results highlight the
imidazopyridine scaffold as a privileged motif and demon-
strate how the integration of emerging technologies in drug
Testing. Activity determinations were performed by Cerep (Le
Bois l’EvÞque, 86600 Celle l’Evescault, France) on a fee-for-service
basis. For details see the Supporting Information.
Received: September 4, 2013
Published online: November 26, 2013
584
www.angewandte.org
ꢀ 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Angew. Chem. Int. Ed. 2014, 53, 582 –585

Angewandte
Chemie
EP1974729 A1, 2008; f) B. E. Evans, K. E. Rittle, M. G. Bock,
Keywords: combinatorial chemistry · computer chemistry ·
drug design · microfluidics · multicomponent reactions
.
R. M. DiPardo, R. M. Freidinger, W. L. Whitter, G. F. Lundell,
D. F. Veber, P. S. Anderson, R. S. L. Chang, V. J. Lotti, D. J.
Cerino, T. B. Chen, P. J. Kling, K. A. Kunkel, J. P. Springer, J.
Hirshfield, J. Med. Chem. 1988, 31, 2235 – 2246.
[1] a) R. Macarron, M. N. Banks, D. Bojanic, D. J. Burns, D. A.
Cirovic, T. Garyantes, D. V. S. Green, R. P. Hertzberg, W. P.
Janzen, J. W. Paslay, U. Schopfer, G. S. Sittampalam, Nat. Rev.
Drug Discovery 2011, 10, 188 – 195; b) M. C. Bryan, C. D. Hein,
H. Gao, X. Xia, H. Eastwood, B. A. Bruenner, S. W. Louie, E. M.
Doherty, ACS Comb. Sci. 2013, 15, 503 – 511.
[2] I. Ugi, Angew. Chem. 1962, 74, 9 – 22; Angew. Chem. Int. Ed.
Engl. 1962, 1, 8 – 21.
[3] B. Beck, S. Srivastava, K. Khoury, E. Herdtweck, A. Dçmling,
Mol. Diversity 2010, 14, 479 – 491.
[4] C. Kalinski, H. Lemoine, J. Schmidt, C. Burdack, J. Kolb, M.
Umkehrer, G. Ross, Synthesis 2008, 4007 – 4011.
[5] H. Bienaymꢁ, K. Bouzid, Angew. Chem. 1998, 110, 2349 – 2352;
Angew. Chem. Int. Ed. 1998, 37, 2234 – 2237.
[6] M. Reutlinger, C. P. Koch, D. Reker, N. Todoroff, P. Schneider, T.
Rodrigues, G. Schneider, Mol. Inf. 2013, 32, 133 – 138.
[7] M. Hieke, C. B. Rçdl, J. M. Wisniewska, E. Buscatꢂ, H. Stark, M.
Schubert-Zsilavecz, D. Steinhilber, B. Hofmann, E. Proschak,
Bioorg. Med. Chem. Lett. 2012, 22, 1969 – 1975.
[8] a) S. Meister, D. M. Plouffe, K. L. Kuhen, G. M. C. Bonamy, T.
Wu, S. W. Barnes, S. E. Bopp, R. Borboa, A. T. Bright, J. Che, S.
Cohen, N. V. Dharia, K. Gagaring, M. Gettayacamin, P. Gordon,
T. Groessl, N. Kato, M. C. S. Lee, C. W. McNamara, D. A.
Fidock, A. Nagle, T.-g. Nam, W. Richmond, J. Roland, M.
Rottmann, B. Zhou, P. Froissard, R. J. Glynne, D. Mazier, A.
Chatterjee, T. T. Diagana, E. A. Winzeler, Science 2011, 334,
1372 – 1377; b) T. Wu, A. Nagle, K. Kuhen, K. Gagaring, R.
Borboa, C. Francek, Z. Chen, D. Plouffe, A. Goh, S. B.
Lakshminarayana, J. Wu, H. Q. Ang, P. Zeng, M. L. Kang, W.
Tan, M. Tan, N. Ye, X. Lin, C. Caldwell, J. Ek, S. Skolnik, F. Liu,
J. Wang, J. Chang, C. Li, T. Hollenbeck, T. Tuntland, J. Isbell, C.
Fischli, R. Brun, M. Rottmann, V. Dartois, T. Keller, T. Diagana,
E. Winzeler, R. Glynne, D. C. Tully, A. K. Chatterjee, J. Med.
Chem. 2011, 54, 5116 – 5130; c) A. T. Baviskar, C. Madaan, R.
Preet, P. Mohapatra, V. Jain, A. Agarwal, S. K. Guchhait, C. N.
Kundu, U. C. Banerjee, P. V. Bharatam, J. Med. Chem. 2011, 54,
5013 – 5030; d) J. M. Wisniewska, C. B. Rçdl, A. S. Kahnt, E.
Buscato, S. Ulrich, Y. Tanrikulu, J. Achenbach, F. Rçrsch, S.
Grçsch, G. Schneider, J. Cinatl, Jr., E. Proschak, D. Steinhilber,
B. Hofmann, Biochem. Pharmacol. 2012, 83, 228 – 240; e) S.
Nordhoff, H. Deppe, U. Abel, A. Feurer, I. Ott, G. Metz,
[9] A. Gaulton, L. J. Bellis, A. P. Bento, J. Chambers, M. Davies, A.
Hersey, Y. Light, S. McGlinchey, D. Michalovich, B. Al-Lazikani,
J. P. Overington, Nucleic Acids Res. 2011, 40, D1100 – D1107.
[10] a) E. Lounkine, M. J. Keiser, S. Whitebread, D. Mikhailov, J.
Hamon, J. L. Jenkins, P. Lavan, E. Weber, A. K. Doak, S. Cꢃtꢁ,
B. K. Shoichet, L. Urban, Nature 2012, 486, 361 – 367; b) J.
Besnard, G. F. Ruda, V. Setola, K. Abecassis, R. M. Rodriguiz,
X. P. Huang, S. Norval, M. F. Sassano, A. I. Shin, L. A. Webster,
F. R. Simeons, L. Stojanovski, A. Prat, N. G. Seidah, D. B.
Constam, G. R. Bickerton, K. D. Read, W. C. Wetsel, I. H.
Gilbert, B. L. Roth, A. L. Hopkins, Nature 2012, 492, 215 – 220.
[11] a) A. A. Antolꢄn, J. Mestres, Therapeutic Targets, Wiley, New
York, pp. 309 – 326; b) Y. Hu, J. Bajorath, AAPS J. 2013, 15,
808 – 815.
[12] a) A. Tropsha, Mol. Inf. 2010, 29, 476 – 488; b) A. Nicholls, J.
Comput.-Aided Mol. Des. 2008, 22, 239 – 255; c) K. Roy, I. Mitra,
Comb. Chem. High Throughput Screening 2011, 14, 450 – 474.
[13] a) M. M. Hann, G. M. Keserꢅ, Nat. Rev. Drug Discovery 2012,
11, 355 – 365; b) G. R. Bickerton, G. V. Paolini, J. Besnard, S.
Muresan, A. L. Hopkins, Nat. Chem. 2012, 4, 90 – 98.
[14] a) P. N. Nge, C. I. Rogers, A. T. Woolley, Chem. Rev. 2013, 113,
2550 – 2583; b) D. Lombardi, P. S. Dittrich, Expert Opin. Drug
Discovery 2010, 5, 1081 – 1094; c) T. Tsukahara, K. Mawatari, T.
Kitamori, Chem. Soc. Rev. 2010, 39, 1000 – 1013.
[15] a) C. E. Rasmussen, C. K. I. Williams, Gaussian Processes for
Machine Learning, The MIT Press, Cambridge, 2006; b) M.
Rupp, T. Schroeter, R. Steri, H. Zettl, E. Proschak, K. Hansen,
O. Rau, O. Schwarz, L. Mꢅller-Kuhrt, M. Schubert-Zsilavecz,
K. R. Mꢅller, G. Schneider, ChemMedChem 2010, 5, 191 – 194.
[16] M. R. Berthold, N. Cebron, F. Dill, T. Gabriel, T. Koetter, T.
Meinl, P. Ohl, C. Sieb, K. Thiel, B. Wiswedel, Studies in
Classification, Data Analysis, and Knowledge Organization
(GfKL), Springer, Berlin, 2007, pp. 319 – 326.
[17] D. Rogers, M. Hahn, J. Chem. Inf. Model. 2010, 50, 742 – 754.
[18] J. F. Truchon, C. I. Bayly, J. Chem. Inf. Model. 2007, 47, 488 – 508.
[19] J. H. Chen, E. Linstead, S. J. Swamidass, D. Wang, P. Baldi,
Bioinformatics 2007, 23, 2348 – 2351.
[20] Chemical Abstracts Service, SciFinder, https://scifinder.cas.org/,
Columbus, OH, U.S.A.
Angew. Chem. Int. Ed. 2014, 53, 582 –585
ꢀ 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
www.angewandte.org
585