Multi-objective molecular de novo design by adaptive fragment prioritization

Article abstract of DOI:10.1002/anie.201310864

We present the development and application of a computational molecular de novo design method for obtaining bioactive compounds with desired on- and off-target binding. The approach translates the nature-inspired concept of ant colony optimization to combinatorial building block selection. By relying on publicly available structure-activity data, we developed a predictive quantitative polypharmacology model for 640 human drug targets. By taking reductive amination as an example of a privileged reaction, we obtained novel subtype-selective and multitarget-modulating dopamine D₄ antagonists, as well as ligands selective for the sigma-1 receptor with accurately predicted affinities. The nanomolar potencies of the hits obtained, their high ligand efficiencies, and an overall success rate of 90 % demonstrate that this ligand-based computer-aided molecular design method may guide target-focused combinatorial chemistry. Finding the best way: A molecular design method inspired by ant colony optimization generated novel, highly potent, druglike ligands for the sigma-1 and dopamine D₄ receptors. The computational approach is readily applicable to combinatorial chemistry and delivers focused compound libraries profiled for desired target-panel activities.

Full text of DOI:10.1002/anie.201310864

.
Angewandte
Communications
DOI: 10.1002/anie.201310864
Combinatorial Chemistry
Multi-Objective Molecular De Novo Design by Adaptive Fragment
Prioritization**
Michael Reutlinger, Tiago Rodrigues, Petra Schneider, and Gisbert Schneider*
Abstract: We present the development and application of
a computational molecular de novo design method for
obtaining bioactive compounds with desired on- and off-
target binding. The approach translates the nature-inspired
concept of ant colony optimization to combinatorial building
block selection. By relying on publicly available structure–
activity data, we developed a predictive quantitative polyphar-
macology model for 640 human drug targets. By taking
reductive amination as an example of a privileged reaction, we
obtained novel subtype-selective and multitarget-modulating
dopamine D4 antagonists, as well as ligands selective for the
sigma-1 receptor with accurately predicted affinities. The
nanomolar potencies of the hits obtained, their high ligand
efficiencies, and an overall success rate of 90% demonstrate
that this ligand-based computer-aided molecular design
method may guide target-focused combinatorial chemistry.
a nature-inspired optimization principle to chemistry-driven
molecular design.
For a proof-of-concept we focused on the reductive
amination reaction as a scheme for combinatorial synthesis.
By automated structure optimization, MAntA generated
small compound libraries with lead-like qualities, high hit
rates, and nanomolar activities. It implements a new design
strategy that is applicable to all kinds of chemistry-driven
computational methods,[8] and neither requires prior knowl-
edge about the bioactivity of scaffold classes nor is limited to
privileged scaffolds. In a retrospective study, ant colony
optimization turned out to perform better or on par with
other optimization methods.[7b] Here, we pioneer the concept
of polypharmacology-based molecular de novo design using
combinatorial chemistry. We demonstrate that both target-
selective, and multitarget-modulating members of large
combinatorial compound libraries are rapidly identified with-
out the need for full library enumeration and synthesis.
The molecular design method requires 1) a scheme for
compound synthesis, 2) a method for predicting the affinity of
the virtual products, and 3) a technique for optimizing the
building blocks. For our concept study, we chose the reductive
amination reaction and aldehydes/ketones and amines as
building blocks. We applied MAntA to the products of single-
step reductive amination starting from commercially avail-
able building blocks. The reaction products have a high
likelihood of possessing desirable druglike features, as
visualized in Figure 1, which presents a map of the known
bioactivity space. Virtual reaction products (green dots)
cluster in a densely populated area, and the reductive
amination may be regarded as a preferred reaction for drug
discovery.
For affinity prediction we trained individual Gaussian
process (GP) regression models[9] for 640 human targets
annotated in ChEMBL (v14),[10] based on 279866 compounds
with 569725 measured bioactivities. Molecules were repre-
sented by topological pharmacophore (“CATS2”)[11] and
substructure (circular Morgan fingerprints)[12] descriptors.
The choice of GP regression was motivated by extensive
comparison to other modeling techniques using the same
training data, where the GP approach performed best
(Tables S2 and S3 in the Supporting Information). In addition,
GP models compute a data-density-dependent confidence
estimate, which we combined with the quantitative bioactivity
prediction (pAffinity) to obtain a single robust prediction
score for each compound.
T
raditional combinatorial chemistry aims at the generation
of large diverse compound arrays for bioactivity screening.[1]
It has been realized that multiple “adaptive” synthesis-and-
test cycles using smaller, focused compound libraries might be
better suited, faster, and more economical to find lead-like
bioactive compounds.[2,3] Computational molecular design
methods offer the additional advantage of generating bioac-
tive compounds while considering multiple objectives in
parallel,[4] and combinatorial libraries with desired properties
can be obtained by relying on chemistry-oriented computa-
tional molecular design.[5,6] Though potentially appealing,
these methods have rarely been prospectively applied. Here,
we present the comprehensive application of a computational
concept for designing combinatorial libraries that exhibit an
accurately predicted bioactivity profile. We show that the
molecular ant algorithm (MAntA)[7] effectively transfers
[*] M. Reutlinger, Dr. T. Rodrigues, Dr. P. Schneider,
Prof. Dr. G. Schneider
Eidgençssische Technische Hochschule (ETH)
Departement Chemie und Angewandte Biowissenschaften
Vladimir-Prelog-Weg 4, 8093 Zꢀrich (Switzerland)
E-mail: gisbert.schneider@pharma.ethz.ch
Dr. P. Schneider, Prof. Dr. G. Schneider
inSili.com GmbH
Segantinisteig 3, 8049 Zꢀrich (Switzerland)
[**] We thank Dr. Michael Bieler for computing QED values. This
research was financially supported by a grant from the OPO
Foundation, Zꢀrich.
Equipped with this quantitative affinity prediction model,
MAntA performs an adaptive search for optimal combina-
tions of building blocks for the given reaction scheme
(Figure 2). The search space consists of all possible substrates
Supporting information for this article (details on the computa-
tional methods, synthesis, analytics, and binding assays) is
available on the WWW under http://dx.doi.org/10.1002/anie.
201310864.
4244
ꢀ 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Angew. Chem. Int. Ed. 2014, 53, 4244 –4248

Angewandte
Chemie
dopamine D4 receptors. The choice of D4 receptor was also
made to allow for a direct comparison to a recent publication
by Hopkins and co-workers.[5] In our study, the task was to
select a small number of preferred products from a total of
approximately 20 million. First, we discarded all designed
molecules with undesired structural motifs,[13] and poor
predicted absorption, distribution, metabolism, and excretion
properties (“negative design”).[14] From the 3529 remaining
molecules we selected candidates to match different criteria
(“positive design”):
1) Potent and selective (sigma-1) or multitarget-modulating
(dopamine D4) ligands
2) Target-subtype-selective ligands
3) Exploratory molecules, lying outside the training domain
as expressed by Morgan fingerprint Tanimoto similarities
< 0.20
4) Inactive compounds that are nearest neighbors to known
high-affinity ligands in ChEMBL bioactivity space.
Figure 1. The distribution of 5000 virtual products (green dots) gen-
erated by reductive amination in a druglike chemical space. The two-
dimensional landscape was calculated from the density of 10000
druglike molecules sampled from the ChEMBL database. The intensity
of the gray color indicates the density of known bioactive substances
(white: sparsely populated; black: highest local density). The com-
pounds were represented by topological pharmacophores (“CATS2”
descriptor)[11] and projected to the plane (x’,x’’) by stochastic neighbor
embedding (SNE), which led to a local-neighborhood preserving map
of chemical space. The axes represent nonlinear combinations of the
original molecular descriptors. The image was generated with
LiSARD.[20]
For the sigma-1 receptor, we selected compounds 1–3
according to the high-affinity criterion. Molecules 4 and 5
were designed as receptor-selective ligands (Figure 3a). In
fact, the MAntA designs were experimentally validated for
their specific goals with accurately predicted pKi values
(Table 1). Compounds 1–3 exhibited Ki values of 1.1–2.2 nm,
and designs 4 and 5 displayed more than 2500-fold selectivity
for the sigma-1 receptor over the d, k, and m opioid receptors.
It should be noted that 1–4 are equipotent to their nearest-
neighbor counterparts from ChEMBL, despite being struc-
turally dissimilar (Tanimoto similarity ꢀ 0.45), thereby
endorsing the exploratory potential of MAntA. Furthermore,
the low molecular weights of 1–5, coupled to low nanomolar
Ki values, endow these compounds with high ligand efficiency.
Additionally, we synthesized and tested compounds 6–8 as
scaffold hops from known ChEMBL chemical space (struc-
tural Tanimoto similarity to nearest neighbors ꢀ 0.20), with-
out critical loss of affinity (sigma-1: Ki = 10–210 nm, DpKi
ꢀ 0.5; Table 1, Figure 3b) with the exception of compound
8. Furthermore, compounds 1–8 contain scaffolds that were
not present in the training data used for model building
(Table S4). Apparently, the low structural resemblance to
known small molecules did not considerably affect the
algorithmꢀs performance. Finally, compound 9 was designed
and validated as a low-affinity sigma-1 ligand (Ki > 2500 nm)
despite having a highly potent nearest neighbor (Ki ꢀ 6 nm,
Tanimoto similarity = 0.45, Table 1), pinpointing the adaptive
design capabilities of MAntA that go beyond structural
similarity analysis. Altogether, the experimental results are in
agreement with the landscape projection of the preferred
sigma-1 activity islands (Figure 3c; individual target land-
scapes are shown in Figure S9). Furthermore, as an off-target
for the synthesized compounds, MAntA predicted moderate
histamine H3 receptor affinities, which were partly confirmed
experimentally.
Figure 2. Selection of molecular building blocks by combinatorial ant
colony optimization (MAntA). The arrows represent artificial ant paths
for this two-component combinatorial library. The widths of the paths
correspond to pseudo-probabilities (“pheromone concentrations”) that
influence the choice made by the ants and thereby determine the
actual product spectrum. The pheromone concentrations are adaptive
and subject to “evaporation”.
labeled with pseudo-probabilities (“pheromones”), according
to their contributions to the computed predictive score.
Individual ants traverse the search space following phero-
mone trails and assemble virtual products. These are scored
and the pseudo-probabilities on their respective molecular
building blocks are adjusted accordingly. Over simulation
time, high-scoring combinations of building blocks emerge
from the ant colonyꢀs optimal path-finding capability.
Next, we designed antagonists for the dopamine D4
receptor. D4 receptors are especially implicated in attention-
deficit hyperactivity disorder, mood disorders, and Parkin-
sonꢀs disease, among other neuropsychiatric illnesses.[15] From
the top 1600 prioritized small molecules with predicted pKi >
We employed MAntA for the multi-objective design of
novel ligands for high-profile macromolecular drug targets
that are involved in neuropsychiatric disorders—sigma-1 and
Angew. Chem. Int. Ed. 2014, 53, 4244 –4248
ꢀ 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
www.angewandte.org
4245

.
Angewandte
Communications
Table 1: The designed molecules and their nearest neighbors from the ChEMBL training data with the predicted and experimentally determined
binding affinities of compounds 1–16.
MAntA designs
Nearest neighbors (training data)
ID
Structure
Predicted
pAffinity
pKi
LE[a]
Structure
ChEMBL
Structural
similarity[c]
pKi
ID[a]
1
2
9.7
9.4
9.0
8.9
0.63
0.65
143089
0.70
0.44
9.4
9.0
112124
154397
3
9.3
8.7
0.58
0.44
8.7
4
5
6
9.0
8.1
7.9
8.8
7.9
7.2
0.61
0.55
0.50
154397
154397
111909
0.47
0.34
0.21
8.7
8.7
7.8
7
8
8.1
8.0
8.1
6.7
0.42
0.32
179530
544748
0.21
0.20
7.7
8.7[d]
9
4.1
9.4
n.d.
8.7
n.d.
20976
0.45
0.70
8.2
9.6
10
0.44
379602
11
9.0
8.3
0.50
285577
0.57
8.1
12
13
7.9
7.8
8.0
7.9
0.53
0.46
210405
345552
0.41
0.43
9.0
8.4
14
7.9
6.6
0.33
305061
0.23
5.4
15
16
7.3
2.5
7.6
0.41
n.d.
143027
129931
0.27
0.40
7.7
8.5
n.d.
[a] Ligand efficiency (LE=À1.4ꢁpKi/number of heavy atoms). [b] ChEMBL IDs are given without the “CHEMBL” prefix. [c] Tanimoto similarity index
(Morgan fingerprints with radius=3). [d] pIC50 value. n.d.: not determined.
7 for the D4 receptor, we selected compounds 10 and 11 as
high-affinity ligands. Although 10 (Ki = 2.0 nm) features an
already known scaffold,[16] 11 represents a notably different
entity that is more ligand efficient than its ChEMBL nearest
neighbor (Tanimoto similarity ꢀ 0.6, Table 1). While the
selected ligands were primarily designed as high-affinity D4
receptor antagonists, a polypharmacology profile was not
precluded. Accordingly, a promiscuous binding profile to the
dopamine D1–5 and 5-HT1A receptors was predicted. Subse-
quent binding tests confirmed the multitarget-modulating
profiles of 10 and 11 in agreement with the MAntA-predicted
bioactivity spectra and landscape projections (Table 1, Fig-
ure 3b, preferred design zones Figure 3d; individual target
landscapes are shown in Figure S10). Conversely, compounds
4246 www.angewandte.org
ꢀ 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Angew. Chem. Int. Ed. 2014, 53, 4244 –4248

Angewandte
Chemie
polypharmacology profile of 14 and 15, which extend the
known chemical diversity of D4 receptor antagonists, is also in
agreement with the pKi predictions. Remarkably, 14 is one log
unit more potent against the D4 receptor than the closest
related reference antagonist, which together with the screen-
ing results of the designed inactive 16, demonstrates the
successful application of MAntA to dopamine receptors.
Evidently, considerably extended experimental GPCR panel
activities will be required for the further hit-to-lead pro-
gression of the MAntA designs.
With regard to the polygenic nature of most major
diseases of the central nervous system and the individual
variability of their genetic basis, new drugs with selected
polypharmacological activities are desirable.[19] The results of
this study suggest a feasible solution for the combinatorial
design of new chemical entities with affinity profiles and
properties that exceed the average druglike-ness for approved
drugs (quantitative estimate of druglike-ness, QED = 0.72 Æ
0.10 vs. 0.49).[3] The automated molecular design method
should be broadly applicable to other classes of drug targets
and chemical reactions, provided reliable structure–activity
data are available for constructing affinity prediction models.
The actual computational design process is fast (within
minutes on a desktop computer), so that focused combina-
torial library design and synthesis can be realized within a day
of work. A particular advantage of MAntA compared to
many other approaches, for example, the meticulous work of
Besnard et al. on adaptive drug design,[5] lies in the simulta-
neous generation of both potent structural analogues and
innovative scaffold-hops from known reference compounds.
Together with rapid computation, low-cost synthesis, and
readily accessible chemical structures, the concept of adaptive
building block and fragment prioritization might become
widely applicable.
Figure 3. Comparison of the Gaussian process prediction for sigma-
1 (a) and dopamine D4 (b) receptors, together with the observed
experimental results. Predictions are variance-corrected pAffinity
values. Experimental data are expressed as %inhibition (competitive
radioligand binding assays). Colors are linearly interpolated from the
predicted pAffinity intervals [4, 9] (a) and [4, 8] (b), and %inhibition
interval [20, 100] for the experimental data. The lower panels presents
LiSARD[20] multitarget selectivity landscapes for the sigma-1 receptor
(c) and the dopamine D4 receptor (d) as the respective on-target.
Received: December 15, 2013
Published online: March 12, 2014
Keywords: computer-assisted drug design · GPCR ·
.
machine learning · polypharmacology · reductive amination
[1] A. Long, Curr. Protoc. Pharmacol. 2012, 9, 9 – 16.
[2] G. Schneider, M. Hartenfeller, M. Reutlinger, Y. Tanrikulu, E.
Proschak, P. Schneider, Trends Biotechnol. 2009, 27, 18 – 26.
[3] a) M. M. Hann, T. I. Oprea, Curr. Opin. Chem. Biol. 2004, 8,
255 – 263; b) G. R. Bickerton, G. V. Paolini, J. Besnard, S.
Muresan, A. L. Hopkins, Nat. Chem. 2012, 4, 90 – 98.
[4] a) V. J. Gillet, Methods Mol. Biol. 2004, 275, 335 – 354; b) C. A.
Nicolaou, N. Brown, C. S. Pattichis, Curr. Opin. Drug Discovery
Dev. 2007, 10, 316 – 324; c) G. Fang, M. Xue, M. Su, D. Hu, X. Li,
B. Xiong, L. Ma, T. Meng, Y. Chen, J. Li, J. Li, J. Shen, Bioorg.
Med. Chem. Lett. 2012, 22, 4540 – 4555.
[5] 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.
[6] M. Reutlinger, T. Rodrigues, P. Schneider, G. Schneider, Angew.
Chem. 2014, 126, 593 – 596; Angew. Chem. Int. Ed. 2014, 53, 582 –
585.
12 and 13 were designed to bind selectively to the D4 receptor.
Selective D4 antagonists are equally relevant in clinics, as they
can prevent stress-induced cognitive dysfunction without
extrapyramidal motor symptoms or neuroendocrine side
effects.[17] Weak binding affinities of 12 and 13 were predicted
for the off-targets in the assay panel. Structural simplicity and
low nanomolar affinities (Ki = 10–12 nm) distinguish these
compounds. Their selectivity for the D4 receptor is partic-
ularly significant, given the high structural similarity to
promiscuous molecules 10 and 11. Of note, 1,4-disubstituted
aromatic piperazines have previously been recognized as
predominant in promiscuous biogenic amine G-protein-
coupled receptor (GPCR) ligands.[18] The opposing target-
engagement profiles for the arylpiperazines 10 and 11 as well
as 12 and 13 confirm effective building-block selections. The
Angew. Chem. Int. Ed. 2014, 53, 4244 –4248
ꢀ 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
www.angewandte.org 4247

.
Angewandte
Communications
[7] a) J. A. Hiss, M. Hartenfeller, G. Schneider, Curr. Pharm. Des.
2010, 16, 1656 – 1665; b) J. A. Hiss, M. Reutlinger, C. P. Koch,
A. M. Perna, P. Schneider, T. Rodrigues, S. Haller, G. Folkers, L.
Weber, R. B. Baleeiro, P. Walden, P. Wrede, G. Schneider, Future
Med. Chem. 2014, 6, 267 – 280.
[8] a) M. Hartenfeller, H. Zettl, M. Walter, M. Rupp, F. Reisen, E.
Proschak, S. Weggen, H. Stark, G. Schneider, PLoS Comput.
Biol. 2012, 8, e1002380; b) M. Hartenfeller, G. Schneider,
WIREs Comput. Mol. Sci. 2011, 1, 742 – 759; c) De Novo
Molecular Design (Ed.: G. Schneider), Wiley-VCH, Weinheim,
2013.
[12] D. Rogers, M. Hahn, J. Chem. Inf. Model. 2010, 50, 742 – 754.
[13] S. Saubern, R. Guha, J. B. Baell, Mol. Inf. 2011, 30, 847 – 850.
[14] D. Lagorce, O. Sperandio, H. Galons, M. A. Miteva, B. O.
Villoutreix, BMC Bioinf. 2008, 9, 396.
[15] F. I. Tarazi, R. J. Baldessarini, Mol. Psychiatry 1999, 4, 529 – 538.
[16] S. Lçber, H. Hꢁbner, P. Gmeiner, Bioorg. Med. Chem. Lett. 2006,
16, 2955 – 2959.
[17] A. F. Arnsten, B. Murphy, K. Merchant, Neuropsychopharma-
cology 2000, 23, 405 – 410.
[18] S. Lçber, H. Hꢁbner, N. Tschammer, P. Gmeiner, Trends
Pharmacol. Sci. 2011, 32, 148 – 157.
[9] C. E. Rasmussen, C. K. I. Williams, Gaussian Processes for
Machine Learning, MIT, Cambridge, 2006.
[19] P. Giusti-Rodrꢂguez, P. F. Sullivan, J. Clin. Invest. 2013, 123,
4557 – 4563.
[10] 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. 2012, 40, D1100 – D1107.
[11] M. Reutlinger, C. P. Koch, D. Reker, N. Todoroff, P. Schneider, T.
Rodrigues, G. Schneider, Mol. Inf. 2013, 32, 133 – 138.
[20] M. Reutlinger, W. Guba, R. E. Martin, A. I. Alanine, T.
Hoffmann, A. Klenner, J. A. Hiss, P. Schneider, G. Schneider,
Angew. Chem. 2011, 123, 11837 – 11840; Angew. Chem. Int. Ed.
2011, 50, 11633 – 11636.
4248 www.angewandte.org
ꢀ 2014 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Angew. Chem. Int. Ed. 2014, 53, 4244 –4248