De Novo Design of Bioactive Small Molecules by Artificial Intelligence

Article abstract of DOI:10.1002/minf.201700153

Generative artificial intelligence offers a fresh view on molecular design. We present the first-time prospective application of a deep learning model for designing new druglike compounds with desired activities. For this purpose, we trained a recurrent neural network to capture the constitution of a large set of known bioactive compounds represented as SMILES strings. By transfer learning, this general model was fine-tuned on recognizing retinoid X and peroxisome proliferator-activated receptor agonists. We synthesized five top-ranking compounds designed by the generative model. Four of the compounds revealed nanomolar to low-micromolar receptor modulatory activity in cell-based assays. Apparently, the computational model intrinsically captured relevant chemical and biological knowledge without the need for explicit rules. The results of this study advocate generative artificial intelligence for prospective de novo molecular design, and demonstrate the potential of these methods for future medicinal chemistry.

Full text of DOI:10.1002/minf.201700153

Communication
www.molinf.com
DOI: 10.1002/minf.201700153
1
2
3
4
5
6
7
8
De Novo Design of Bioactive Small Molecules by Artificial
Intelligence
Daniel Merk,[a] Lukas Friedrich,[a] Francesca Grisoni,[a, b] and Gisbert Schneider*[a]
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
Abstract: Generative artificial intelligence offers a fresh view generative model. Four of the compounds revealed nano-
on molecular design. We present the first-time prospective molar to low-micromolar receptor modulatory activity in
application of a deep learning model for designing new cell-based assays. Apparently, the computational model
druglike compounds with desired activities. For this intrinsically captured relevant chemical and biological
purpose, we trained a recurrent neural network to capture knowledge without the need for explicit rules. The results of
the constitution of a large set of known bioactive com- this study advocate generative artificial intelligence for
pounds represented as SMILES strings. By transfer learning, prospective de novo molecular design, and demonstrate
this general model was fine-tuned on recognizing retinoid X the potential of these methods for future medicinal
and peroxisome proliferator-activated receptor agonists. We chemistry.
synthesized five top-ranking compounds designed by the
Keywords: Automation · drug discovery · machine learning · medicinal chemistry · nuclear receptor
24 Computational de novo design aims to generate new
promise to deliver actually synthesizable bioactive de novo
designs.
25 chemical entities with desired properties.[1] There are several
26 such methodologies, largely differing in the process of
The computational approach consisted of two basic
steps. First, we developed a generic model that learned the
constitution of druglike molecules from a large unfocussed
compound set. In a second step, we fine-tuned this generic
model on more specific molecular features from a small
target-focused library of actives. For the generic model, we
utilized a recently published deep recurrent neural network
(RNN) with long short-term memory (LSTM) cells,[6] which
had been trained on SMILES representations of 541,555
bioactive compounds (KD, Ki, IC/EC50 values <1 mM) ex-
tracted from the ChEMBL22[7] compound database.[5] Then,
we fine-tuned the model by transfer learning to enable the
de novo generation of target-specific ligands. For this
purpose, we used 25 fatty acid mimetics[8] with known
27 chemical structure generation and the scoring methods
28 employed.[2,3] Recently, an innovative concept of de novo
29 molecular design has been proposed that relies on
30 generative artificial intelligence (AI). It bears promise as a
31 way of learning from known bioactive compounds and
32 autonomously designs novel compounds with inherited
33 bioactivity and synthesizability (Figure 1).[4,5] Importantly,
34 these generative methods are expected to produce chemi-
35 cally correct structures without the need for explicitly
36 including building block libraries or rules for their fusion
37 and chemical transformation. However, until now, gener-
38 ative AI has only been applied to retrospective de novo
39 design by reproducing known bioactive ligands or generat-
40 ing predicted actives. In this first prospective study, we
41 apply generative AI to see if this approach lives up to its
[a] Dr. D. Merk, L. Friedrich, Dr. F. Grisoni, Prof. Dr. G. Schneider
Department of Chemistry and Applied Biosciences
Swiss Federal Institute of Technology (ETH)
Vladimir-Prelog-Weg 4, CH-8093 Zurich, Switzerland
E-mail: gisbert.schneider@pharma.ethz.ch
[b] Dr. F. Grisoni
42
43
44
45
46
47
48
49
50
51
52
Department of Earth and Environmental Sciences
University of Milano-Bicocca
P.za della Scienza, 1, IT-20126 Milan, Italy
Supporting information for this article is available on the WWW
under https://doi.org/10.1002/minf.201700153
© 2018 The Authors. Published by Wiley-VCH Verlag GmbH & Co.
KGaA. This is an open access article under the terms of the
Creative Commons Attribution Non-Commercial License, which
permits use, distribution and reproduction in any medium,
provided the original work is properly cited and is not used for
commercial purposes.
53
54
55
56
Figure 1. Concept of generative artificial intelligence (AI). A model
of the training data (e.g., molecular structures) is obtained that can
be used to emit new instances (new chemical entities) within the
training domain by sampling.
© 2018 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Mol. Inf. 2018, 37, 1700153
(1 of 5) 1700153

Communication
www.molinf.com
1
2
3
4
5
6
7
8
9
agonistic activity on retinoid X receptors (RXR)[9] and/or
peroxisome proliferator-activated receptors (PPAR).[10] From
the resulting fine-tuned AI model, we sampled 1000 SMILES
strings, applying fragment growing from the minimalist
start fragment “ÀCOOH”.
and building block availability. These five chemical entities
were not present in the ChEMBL,[7] PubChem,[12] Sure-
ChEMBL,[13] Reaxys[14] and SciFinder[15] databases, indicating
their novelty.
Compounds 1–5 were prepared over two to four steps
(Scheme 1). Amide coupling of 5-bromothiophene-2-car-
boxylic acid (6) and methyl 3-aminobenzoate (7), using
EDC/4-DMAP to 8, followed by Suzuki reaction with
benzeneboronic acid (9) to 10 and alkaline ester hydrolysis
afforded compound 1. Compound 2 was available from 4-
bromo-3-trifluoromethylbenzoic acid (11) and 2-hydroxy-5-
fluorobenzeneboronic acid (12), forming 13 in a Suzuki
reaction followed by Williamson ether synthesis to 14 with
excess 4-fluorobenzyl bromide (15) and subsequent hydro-
lysis of the resulting ester. For the preparation of compound
3, 4-bromosalicylic acid (16) was esterified (17) with
methanol and reacted with bromocyclopentane (18) to
form ether 19. Suzuki reaction of 19 with 3-hydroxybenze-
neboronic acid (20) to 21 and alkaline ester hydrolysis
yielded 3. Compound 4 was obtained from 3-(4-amino-
phenyl)propionic acid (22) by esterification (23), amide
coupling with 2-(4-chlorophenyl)benzoic acid (24) to 25
using EDC/4-DMAP, and alkaline ester hydrolysis. Suzuki
reaction of 4-formylphenylboronic acid (26) and 5-bromo-
resorcinol (27) to 28 followed by Knoevenagel condensation
in Doebner modification with malonic acid afforded com-
pound 5.
We then characterized designs 1–5 in hybrid reporter
gene assays for their agonistic effects on nuclear receptors
RXRa/b/g and PPARa/g/d in HEK293T cells.[16] These in vitro
tests involved constitutively expressed hybrid receptors
composed of the ligand binding domain of the respective
human nuclear receptor and the DNA-binding domain of
the nuclear receptor Gal4 from yeast. Gal4 responsive firefly
luciferase served as reporter gene, and constitutively ex-
pressed Renilla luciferase was used for normalization of
transfection efficiency and toxicity control.
The in vitro characterization of 1–5 revealed agonistic
activity on PPAR and RXR subtypes (Table 1). Four of the
compounds were active, and for each receptor studied, we
identified at least one agonist. Designs 1 and 2 turned out
as dual agonists of RXRs and PPARg, whereas 3 and 4 each
activated two PPAR subtypes but were inactive on RXRs.
Only design 5 showed neither RXR nor PPAR transactivation
activity. EC50 values of 1–4 ranged between double-digit
nanomolar for RXR agonist 1, despite moderate trans-
activation efficacy, and double-digit micromolar for design 4
on PPARd. Design 2 revealed micromolar potency on RXRs
but markedly higher transactivation efficacy than 1. With
regard to PPARg, design 2 showed micromolar agonistic
activity with equivalent efficacy as the reference agonist
pioglitazone. Design 3 behaved as a micromolar super-
agonist on PPARg, with about 2.5-fold greater transactiva-
tion efficacy than pioglitazone. 4 turned out as the least
potent design and showed partial agonistic activity on both
PPARg and PPARd.
The generated set included 93% valid and 90% unique
SMILES entries, all of which contained a carboxylic acid
function by default. None of the computer-generated
chemical structures was identical to compounds from the
10 training sets. Importantly, the newly generated molecules
11 populate the chemical space of the training data, residing
12 within the RXR/PPAR region of the fine-tuning set (Figure 2).
13 These observations corroborate the ability of the generative
14 AI model to produce novel chemical entities within the
15 training data domain.
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
Figure 2. Chemical space analysis by multi-dimensional scaling.
Compounds were represented by Morgan substructure fingerprints
(radius=0–4 bonds, length=1024 bit), and similarity was defined
by the Jaccard-Tanimoto index. Colored dots represent the training
data (light grey), fine-tuning set (green), known RXR (orange) and
PPAR (blue) agonists, sampled molecules (dark grey), and the
selected de novo designs 1–5 (red). Compounds 1, 2, 3 and 5
populate the same area as the known RXR and PPAR agonists, while
4 is similar to PPAR agonist but remote from known RXR actives.
Following this preliminary analysis, we computationally
47 ranked the de novo designs according to their potential
48 modulatory effects on RXRs and PPARs. For this purpose, we
49 employed a target prediction method (SPiDER),[11] and
50 molecular shape and partial charge descriptors to deter-
51 mine the similarity of the designed compounds to known
52 bioactive ligands. The individual screening lists were
53 merged, obtaining a final set of 49 high-scoring designs
54 (Supplementary Information). For proof-of-concept, we
55 selected five compounds (1–5, Scheme 1) from this list for
56 synthesis, taking into account their individual in silico ranks
© 2018 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Mol. Inf. 2018, 37, 1700153
(2 of 5) 1700153

Communication
www.molinf.com
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
Scheme 1. Synthesis of designs 1–5. Reagents & conditions: (a) H2NÀC6H4ÀCOOH (7), EDC, 4-DMAP, THF, reflux, 4 h; (b) C6H5ÀB(OH)2 (9),
Pd(PPh3)4, Cs2CO3, dioxane, 1008C, 16 h; (c) KOH, MeOH/THF/H2O, mw, 708C, 30 min; (d) HO-C6H3FÀB(OH)2 (12), Pd(PPh3)4, Cs2CO3, toluene/
EtOH, 1008C, 20 h; (e) F-C6H4-CH2-Br (15), K2CO3, DMF, mw, 1008C, 120 min; (f) MeOH, H2SO4cc, reflux, 4 h; (g) C5H9Br (18), K2CO3, DMF, mw,
1008C, 6 h; (h) HO-C6H4-B(OH)2 (20), Pd(PPh3)4, Cs2CO3, toluene/EtOH, 1008C, 16 h; (i) C6H4Cl-C6H4-COOH (24), EDC, 4-DMAP, CHCl3, relux, 12 h;
(j) C6H3Br(OH)2 (27), Pd(PPh3)4, Cs2CO3, dioxane/DMF, reflux, 4 h; (k) malonic acid, pyridine/piperidine, mw, 1008C, 30 min.
Table 1. In vitro activity of designs 1–5 on RXRs and PPARs (EC50 values Æ SEM [mM]; n=2 (when inactive) or 4 (when active) independent
experiments in duplicates; inactive, no statistically significant reporter transactivation at a compound concentration of 30 mM).
Compound no.
RXRa
RXRb
RXRg
PPARa
PPARg
PPARd
1
2
3
4
0.13Æ0.01
13.0Æ0.1
inactive
inactive
inactive
1.1Æ0.3
9Æ2
0.06Æ0.02
8.0Æ0.7
inactive
inactive
2.3Æ0.2
2.8Æ0.3
10.1Æ0.3
9Æ3
inactive
inactive
inactive
14Æ2
inactive
inactive
inactive
0.024Æ0.004
inactive
4.0Æ1.0
inactive
inactive
0.025Æ0.002
inactive
5
inactive
0.006Æ0.002
inactive
0.6Æ0.1
inactive
0.5Æ0.1
reference agonistsa)
0.033Æ0.002
a)
Reference agonists, literature data: bexarotene[17] for RXRs, GW7647[18] for PPARa, pioglitazone[19] for PPARg, L165,041[19] for PPARd
© 2018 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Mol. Inf. 2018, 37, 1700153
(3 of 5) 1700153

Communication
www.molinf.com
1
2
3
4
5
6
7
8
9
To exclude unspecific effects, we repeated the in vitro
assays in the absence of a hybrid receptor for every active
molecule, using a concentration at or above its EC80 value.
This time, only the reporter gene and control luciferase,
but no hybrid receptor, were transfected. Designs 1–4
caused no observable reporter transactivation without a
hybrid receptor, confirming that their activity was actually
mediated via RXRs and PPARs, respectively (Supplemen-
tary Information).
References
[1] G. Schneider, Nat. Rev. Drug Discov. 2018, doi: nrd.2017.232.
[2] P. Schneider, G. Schneider, J. Med. Chem. 2016, 59, 4077–4086.
[3] a) G. Schneider, U. Fechner, U. , Nat. Rev. Drug Discov. 2005, 4,
649–663; b) M. Hartenfeller, G. Schneider, Methods Mol. Biol.,
2011, 672, 299–323.
[4] a) M. Olivecrona, T. Blaschke, O. Engkvist, H. Chen, J. Chemin-
form. 2017, 9, 48; b) T. Blaschke, M. Olivecrona, O. Engkvist, J.
Bajorath, H. Chen, Mol. Inf. 2018, 37, 1700123; c) T. Miyao, K.
Funatsu, Mol. Inf. 2017, 36, 1700030.
[5] A. Gupta, A. T. Mꢀller, B. J. H. Huisman, J. A. Fuchs, P. Schneider,
G. Schneider, Mol. Inf. 2018, 37, 1700111.
[6] S. Hochreiter, J. Schmidhuber, Neural Comput. 1997, 9, 1735–
1780.
[7] A. P. Bento, A. Gaulton, A. Hersey, L. J. Bellis, J. Chambers, M.
Davies, F. A. Krꢀger, Y. Light, L. Mak, S. McGlinchey, M. Nowotka,
G. Papadatos, R. Santos, J. P. Overington, Nucleic Acids Res.
2014, 42, 1083–1090; https://www.ebi.ac.uk/chembl/.
[8] E. Proschak, P. Heitel, L. Kalinowsky, D. Merk, J. Med. Chem.
2017, 60, 5235–5266.
[9] P. Germain, P. Chambon, G. Eichele, R. M. Evans, M. A. Lazar, M.
Leid, A. R. De Lera, R. Lotan, D. J. Mangelsdorf, H. Gronemeyer,
Pharmacol. Rev. 2006, 58, 760–772.
10
These results experimentally validate the applicability of
11 generative AI to prospective de novo molecule design. The
12 computational approach led to the discovery of new
13 agonists of therapeutically relevant nuclear receptors. The
14 bioactive designs 1–4 possess considerable potency, as well
15 as diverse selectivity profiles on RXRs and PPARs, and may
16 serve as starting points for hit-to-lead expansion. All of the
17 selected compounds were easily prepared from commer-
18 cially available building blocks, suggesting that their
19 chemical synthesizability was intrinsically learned by the
20 computer model. The results also suggest that a proper
21 choice of compound libraries for model fine-tuning by
22 transfer learning enables application-tailored AI support for
23 de novo design. This particular concept might even be
24 suitable for concerted multi-target drug design. By provid-
25 ing rapid knowledge-driven access to innovative small
26 molecules, generative AI bears potential for medicinal
27 chemistry and chemical biology.
28
29
30 Conflict of Interest
31
32 G. S. declares a potential financial conflict of interest in his
33 role as life-science industry consultant and cofounder of
34 inSili.com GmbH, Zurich.
35
36
[10] L. Michalik, J. Auwerx, J. P. Berger, V. K. Chatterjee, C. K. Glass,
F. J. Gonzalez, P. A. Grimaldi, T. Kadowaki, M. A. Lazar, S.
O’Rahilly, C. N. Palmer, J. Plutzky, J. K. Reddy, B. M. Spiegelman,
B. Staels, W. Wahli, Pharmacol. Rev. 2006, 58, 726–741.
[11] D. Reker, T. Rodrigues, P. Schneider, G. Schneider, Proc. Natl.
Acad. Sci. USA 2014, 111, 4067–4072.
[12] S. Kim, P. A. Thiessen, E. E. Bolton, J. Chen, G. Fu, A. Gindulyte,
L. Han, J. He, S. He, B. A. Shoemaker, J. Wang, B. Yu, J. Zhang,
S. H. Bryant, Nucleic Acids Res. 2016, 44, D1202–D1213; https://
pubchem.ncbi.nlm.nih.gov/.
[13] G. Papadatos, M. Davies, N. Dedman, J. Chambers, A. Gaulton,
J. Siddle, R. Koks, S. A. Irvine, J. Pettersson, N. Goncharoff, A.
Hersey, J. P. Overington, Nucleic Acids Res. 2016, 44, D1220–
D1228; https://www.surechembl.org/.
[14] Reaxys, Elsevier 2017; https://www.reaxys.com/.
[15] SciFinder, Chemical Abstracts Service 2017; https://scifinder.ca-
s.org/.
[16] J. Schmidt, M. Rotter, T. Weiser, S. Wittmann, L. Weizel, A. Kaiser,
J. Heering, T. Goebel, C. Angioni, M. Wurglics, A. Paulke, G.
Geisslinger, A. Kahnt, D. Steinhilber, E. Proschak, D. Merk, J.
Med. Chem. 2017, 60, 7703–7724.
[17] M. Boehm, L. Zhang, B. Badea, S. White, D. Mais, E. Berger, C. Suto,
M. Goldman, R. Heyman, J. Med. Chem. 1994, 37, 2930–2941.
[18] P. Brown, L. Stuart, K. Hurley, M. Lewis, D. Winegar, J. Wilson, W.
Wilkison, O. Ittoop, T. Willson, Bioorg. Med. Chem. Lett. 2001, 11,
1225–1227.
37 Acknowledgements
38
39 We thank P. Schneider for compiling the subsets of the
40 ChEMBL database and A. T. Mu¨ller for technical support.
41 This research was financially supported by the Swiss
42 National Science Foundation (grant no. IZSEZ0_177477). D.
43 M. was supported by an ETH Zurich Postdoctoral Fellowship
44 (grant no. 16-2 FEL-07).
45
46
47
48
49
50
51
52
53
54
55
56
[19] T. Willson, P. Brown, D. Sternbach, B. Henke, J. Med. Chem.
2000, 43, 527–550.
Received: December 13, 2017
Accepted: December 20, 2017
Published online on &&&, &&&&
© 2018 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim
Mol. Inf. 2018, 37, 1700153
(4 of 5) 1700153

1
2
COMMUNICATION
3
4
5
Dr. D. Merk, L. Friedrich, Dr. F. Grisoni,
Prof. Dr. G. Schneider*
6
1 – 5
7
8
9
De Novo Design of Bioactive Small
Molecules by Artificial Intelligence
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56