Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed in the Laboratory

Article abstract of DOI:10.1016/j.chempr.2018.02.002

The Chematica program was used to autonomously design synthetic pathways to eight structurally diverse targets, including seven commercially valuable bioactive substances and one natural product. All of these computer-planned routes were successfully executed in the laboratory and offer significant yield improvements and cost savings over previous approaches, provide alternatives to patented routes, or produce targets that were not synthesized previously. Although computers have demonstrated the ability to challenge humans in various games of strategy, their use in the automated planning of organic syntheses remains unprecedented. As a result of the impact that such a tool could have on the synthetic community, the past half century has seen numerous attempts to create in silico chemical intelligence. However, there has not been a successful demonstration of a synthetic route designed by machine and then executed in the laboratory. Here, we describe an experiment where the software program Chematica designed syntheses leading to eight commercially valuable and/or medicinally relevant targets; in each case tested, Chematica significantly improved on previous approaches or identified efficient routes to targets for which previous synthetic attempts had failed. These results indicate that now and in the future, chemists can finally benefit from having an “in silico colleague” that constantly learns, never forgets, and will never retire. Multistep synthetic routes to eight structurally diverse and medicinally relevant targets were planned autonomously by the Chematica computer program, which combines expert chemical knowledge with network-search and artificial-intelligence algorithms. All of the proposed syntheses were successfully executed in the laboratory and offer substantial yield improvements and cost savings over previous approaches or provide the first documented route to a given target. These results provide the long-awaited validation of a computer program in practically relevant synthetic design.

Full text of DOI:10.1016/j.chempr.2018.02.002

Article
Efficient Syntheses of Diverse, Medicinally
Relevant Targets Planned by Computer and
Executed in the Laboratory
Tomasz Klucznik, Barbara
Mikulak-Klucznik, Michael P.
McCormack, ..., Milan Mrksich,
Sarah L.J. Trice, Bartosz A.
Grzybowski
milan.mrksich@northwestern.edu (M.M.)
sarah.trice@sial.com (S.L.J.T.)
nanogrzybowski@gmail.com (B.A.G.)
HIGHLIGHTS
Computer autonomously designs
chemical syntheses of medicinally
relevant molecules
The syntheses are successfully
executed in the laboratory
The machine-designed routes
improve on previous approaches
Multistep synthetic routes to eight structurally diverse and medicinally relevant
targets were planned autonomously by the Chematica computer program, which
combines expert chemical knowledge with network-search and artificial-
intelligence algorithms. All of the proposed syntheses were successfully executed
in the laboratory and offer substantial yield improvements and cost savings over
previous approaches or provide the first documented route to a given target.
These results provide the long-awaited validation of a computer program in
practically relevant synthetic design.
Klucznik et al., Chem 4, 1–11
March 8, 2018 ª 2018 The Authors. Published
by Elsevier Inc.
https://doi.org/10.1016/j.chempr.2018.02.002

Please cite this article in press as: Klucznik et al., Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed
in the Laboratory, Chem (2018), https://doi.org/10.1016/j.chempr.2018.02.002
Article
Efficient Syntheses of Diverse,
Medicinally Relevant Targets Planned
by Computer and Executed in the Laboratory
1
1
2
2
ꢀ 1
Tomasz Klucznik, Barbara Mikulak-Klucznik, Michael P. McCormack, Heather Lima, Sara Szymkuc,
Manishabrata Bhowmick,2 Karol Molga,1 Yubai Zhou,3 Lindsey Rickershauser,2 Ewa P. Gajewska,1
Alexei Toutchkine,2 Piotr Dittwald,1 Michał P. Startek,4 Gregory J. Kirkovits,2 Rafał Roszak,1
1
1
3,
2,
ꢀ
*
*
Ariel Adamski, Bianka Sieredzinska, Milan Mrksich, Sarah L.J. Trice,
and Bartosz A. Grzybowski1,5,6,
*
SUMMARY
The Bigger Picture
Although computers have
The Chematica program was used to autonomously design synthetic pathways
to eight structurally diverse targets, including seven commercially valuable
bioactive substances and one natural product. All of these computer-planned
routes were successfully executed in the laboratory and offer significant yield
improvements and cost savings over previous approaches, provide alternatives
to patented routes, or produce targets that were not synthesized previously.
demonstrated the ability to
challenge humans in various games
of strategy, their use in the
automated planning of organic
syntheses remains unprecedented.
As a result of the impact that such a
tool could have on the synthetic
community, the past half century
has seen numerous attempts to
create in silico chemical
INTRODUCTION
Teaching the computer to plan chemical syntheses has been one of the outstanding
challenges of modern-era organic chemistry. Despite decades of research and many
ingenious approaches,1–12 there have been no literature reports of complete syn-
thetic pathways designed by the computer and then successfully executed in the
laboratory.11 The inadequacy of computer programs reflected, among other factors,
their limited knowledge base of chemical transformations, their inability to navigate
enormous ‘‘trees’’ of synthetic possibilities in an intelligent fashion, and the lack of
higher-order logic prescribing how individual steps should be put together to pro-
duce elegant, or at least viable, pathways. Building on over a decade of research
on chemical networks,13–16 we have recently disclosed10 a de novo retrosynthetic
module within the Chematica platform (henceforth, simply Chematica) that unites
network theory, modern high-power computing, artificial intelligence, and expert
chemical knowledge to rapidly design synthetic pathways leading to arbitrary (i.e.,
previously made or never attempted) targets. Although Chematica has attracted
considerable interest,17,18 its predictions have not been verified experimentally until
now. Here, we describe the results of a systematic evaluation in which synthetic path-
ways leading to eight structurally diverse and medicinally relevant targets were first
designed by Chematica without any human supervision and subsequently executed
in the laboratory. All of these syntheses were successful and either improved on pre-
vious approaches or provided the first documented route to a given target.
intelligence. However, there has
not been a successful
demonstration of a synthetic route
designed by machine and then
executed in the laboratory. Here,
we describe an experiment where
the software program Chematica
designed syntheses leading to
eight commercially valuable and/or
medicinally relevant targets; in
each case tested, Chematica
significantly improved on previous
approaches or identified efficient
routes to targets for which previous
synthetic attempts had failed.
These results indicate that now and
in the future, chemists can finally
benefit from having an ‘‘in silico
colleague’’ that constantly learns,
never forgets, and will never retire.
Starting in 2005,13 we have published extensively on the algorithms and methods
enabling representation of synthetic pathways as the so-called bipartite graphs,
which can then be queried in the Chematica10,16 platform according to different
sets of criteria. In retrosynthesis, these criteria are rules describing various types of
Chem 4, 1–11, March 8, 2018 ª 2018 The Authors. Published by Elsevier Inc.
1
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Please cite this article in press as: Klucznik et al., Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed
in the Laboratory, Chem (2018), https://doi.org/10.1016/j.chempr.2018.02.002
reactions. The rules are coded by expert chemists and to ensure applicability to arbi-
trary targets, they cover not only popular and simple transformations but also the
advanced methodologies essential in the synthesis of complex targets. At the
core of each of Chematica’s ꢀ50,000 rules is a decision tree such as the one shown
in Figure 1A for double stereodifferentiating condensation of esters with aldehydes.
The various conditions within the tree specify the range of admissible and also
possible (i.e., not only those based on prior literature precedents) substituents or
atom types. Importantly, all rules account for stereo- and regioselectivity and also
for the ‘‘context’’ of the molecule; that is, for groups incompatible with the reaction
or those to be protected (for these and other aspects of rule application including
electronic and steric effects, see Supplemental Information, Sections S2–S5).
The reaction rules are only the basic ‘‘moves’’ from which the complete synthetic path-
ways (‘‘games’’) are to be constructed. Because the number of choices at each retrosyn-
thetic step10 is ꢀ100 (commensurate with the number of choices at each step in a chess
game), the number of possibilities within n steps scales as 100n. To search such an
enormous synthetic space (Figures 1B and 1C; Supplemental Information, Section S6),
intelligent algorithms are needed to truncate and revert from unpromising ‘‘branches’’
and channel the searches toward the most efficient and elegant sequences of steps.
Chematica avoids unpromising routes by using numerous heuristics prohibiting unlikely
structural motifs, penalizing reactions that are non-selective, or those that would have to
proceed through very strained intermediates (Supplemental Information, Sections S5.2
andS6).Thesearchesarethenguidedtowardthemostfeasiblesolutionsbytheso-called
scoring functions that evaluate (1) the sets of substrates made at each step and (2) the
sequences of reactions that were used to reach any particular set. Importantly, to enable
searches and scoring of substrate sets rather than individual molecules, the bipartite
reaction graphs are transformed into so-called hypergraphs (with ‘‘supernodes’’
combining several individual substance nodes; Supplemental Information, Sections
S6.1 and S6.2). The algorithm navigating these hypergraphs takes advantage of
higher-order logic (concatenating individual steps into ‘‘strategic sequences,’’ elimi-
nating sequences of steps in which highly reactive groups are dragged along, etc.; see
Supplemental Information, Section S7) and terminates when reaching commercially
available(currently, over200,000chemicalsfromtheSigma-Aldrichcatalog)orothersyn-
thetically popular substrates (ca. 7,000,000 molecules from literature and patents, each
with the value of its connectivity within the Network of Chemistry13–16). Finally, because
up to millions of viable pathways can be found for typical targets, dynamic linear pro-
gramming algorithms are used to retrieve pathways that are not only best scoring but
also significantly different from each other (Supplemental Information, Section S6.3). In
the pathways presented to the user, each substance can be further inspected via built-
in molecular mechanics tools, and each reaction comes with suggestions for reaction
conditions, literature citation(s) illustrating this type of chemistry, information on which
groups need to be protected (and with what protecting groups), examples of similar re-
actions reported in literature, and more (see Supplemental Information, Section S8).
1Institute of Organic Chemistry, Polish Academy
of Sciences, ul. Kasprzaka 44/52, Warsaw 01-224,
Poland
2MilliporeSigma, 400 Summit Drive, Burlington,
MA 01803, USA
3Department of Chemistry, Northwestern
University, 2145 Sheridan Road, Evanston,
IL 60201, USA
4Faculty of Mathematics, Informatics, and
Mechanics, University of Warsaw, 02-097 Warsaw,
Poland
5IBS Center for Soft and Living Matter and
Department of Chemistry, Ulsan National
Institute of Science and Technology, 50
UNIST-gil, Eonyang-eup, Ulju-gun,
Ulsan 689-798, South Korea
RESULTS AND DISCUSSION
Choice of Targets
The algorithms and methods described above were used to design the syntheses of
eight targets chosen as follows. The first six targets were provided by MilliporeSigma
(MS, formerly Sigma-Aldrich) and were all biologically active compounds of high
commercial value (>US$100/mg) for which previous (and in most cases numerous)
synthetic attempts at MS were very low yielding, not scalable, or altogether failed.
Here, the main objective for Chematica was to design routes improving over these
6Lead Contact
*Correspondence:
milan.mrksich@northwestern.edu (M.M.),
sarah.trice@sial.com (S.L.J.T.),
nanogrzybowski@gmail.com (B.A.G.)
https://doi.org/10.1016/j.chempr.2018.02.002
2
Chem 4, 1–11, March 8, 2018

Please cite this article in press as: Klucznik et al., Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed
in the Laboratory, Chem (2018), https://doi.org/10.1016/j.chempr.2018.02.002
Figure 1. Reaction Rules and Reaction Networks Underlying Chematica
(A) An example of a decision tree for one of Chematica’s ꢀ50,000 reactions rules (double
stereodifferentiating condensation of esters with aldehydes). The tree begins with a condition of
the reaction being intermolecular. To ensure face selectivity of the enolate, conditions for the
substituents at positions #8, #1, and #3 are considered. Conditions at positions #12, #2, #11, #14
follow and ensure proper face selectivity of the aldehyde. The last two conditions are common for
both substrates. The substrates should be acyclic because cyclic structures might distort the
aldehyde-titanium chelate conformation or face selectivity of the ester enolate. The other
requirement concerns the consonant selectivity at both substrates that ensures the desired
Chem 4, 1–11, March 8, 2018
3

Please cite this article in press as: Klucznik et al., Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed
in the Laboratory, Chem (2018), https://doi.org/10.1016/j.chempr.2018.02.002
Figure 1. Continued
diastereoselectivity. The mechanistic reasons for each condition as well as the translation of the
tree into SMILES notation are discussed in detail in the Supplemental Information, Section S3.4.
(B) The rules such as the one described in (A) are used to explore the graphs of synthetic
possibilities emanating from the target and growing with the number of search iterations. Each
node corresponds to a set of substrates. The image shown here is for the early stage of planning the
synthesis of the BRD7/9 inhibitor 8. In a typical planning task, the program constructs and analyzes
networks that are tens to thousands (sic!) of times larger than the one shown.
(C) A subgraph of (B) showing only the viable synthetic pathways terminating in commercially
available (red nodes) or known (green nodes) substrates. Once such feasible syntheses are found,
the program extracts them from its internal network representation and displays as in the actual
synthetic routes in Figures 2 and 3 (also see Movie S1).
previous attempts in terms of overall cost. The seventh target was a blockbuster anti-
arrythmic drug, dronedarone, from Sanofi-Aventis; this choice was made by the
Grzybowski group and was motivated by the fact that tens of patents have been
granted to protect dronedarone’s synthesis (see list in the Supplemental Informa-
tion, Section S16.1), raising the bar for the computer to find alternative pathways.
The eighth target was selected by the Mrksich group to validate Chematica in the
synthesis of a natural product, engelheptanoxide C, that was recently isolated but
not yet synthesized.19 Last but not least, the DARPA agency sponsoring most of
our effort within the ‘‘Make-it’’ program—whose aim is to automate both synthetic
planning and actual synthetic procedures—has been interested in whether a retro-
synthetic software such as Chematica could ‘‘empower’’ less experienced chemists
to perform synthetic work that would typically be carried out in classic synthetic
laboratories. Accordingly, the first four targets (Figure 2) were made by MS chemists,
whereas the last four (Figure 3) were synthesized by the Grzybowski and Mrksich
students who are not experienced in multistep organic synthesis.
Execution Criteria and Expected Deliverables
The syntheses were planned completely autonomously by Chematica (running on a
64-core machine) within 15–20 min for all targets with the exception of dronedarone
for which the search used an older and slower version of the software and was al-
lowed to continue for several hours. The top-scoring pathway for each target was
chosen provided that (1) it was significantly different from any previous approaches,
and (2) the starting materials were immediately available. If the best-scoring pathway
did not meet these criteria, the second-best was chosen (in three cases). No alter-
ations in the proposed routes were allowed save for straightforward adjustments
in reactions conditions (e.g., temperature, solvent, specific base, catalyst, etc.) for
the sake of optimization. For the MS targets, the additional criteria—reflecting the
constraints and demands of industrial reality—were to attempt each step no more
than three to five times and to deliver at least a few hundred milligrams (and in
most cases >1 g) of product within 8 weeks (no more than 70 hr of bench work)
with a final high-performance liquid chromatography purity above 98% and no
single impurity above 0.5% (except in the case of (S)-4-hydroxyduloxetine,
which was obtained in ꢀ95% purity). For the student-led projects, the time con-
straints were ca. 3–4 months with similar purity requirements.
Syntheses of Targets
The first target was the recently discovered and the first potent and selective inhibitor20
of the so-called bromodomain-containing proteins BRD7 and BRD9 implicated in can-
cer. Previous attempts at MS to synthesize this quinolone-fused enantiomerically pure
lactam (8 in Figure 2A) according to the reported, eight-step literature procedure20 re-
sulted in very low isolated yields (overall, ꢀ1%; see Supplemental Information, Section
S10) and required the use of flash column chromatography (FCC) in all but one of the
4
Chem 4, 1–11, March 8, 2018

Please cite this article in press as: Klucznik et al., Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed
in the Laboratory, Chem (2018), https://doi.org/10.1016/j.chempr.2018.02.002
Figure 2. Syntheses of the First Four Targets Performed by the MS Chemists
BRD 7/9 inhibitor (A), a-hydroxyetizolam (B), ATR kinase inhibitor (C), and inhibitor of human acute-
myeloid-leukemia cells (D). The images are screenshots of pathway ‘‘graphs’’ as displayed in
Chematica (see also Movie S1). The positions of the structures in the chemical schemes below
mirror, as much as possible, those in Chematica’s graphs. Red nodes, commercial chemicals (with
prices in US$/g); green nodes, known substances (with numerals denoting synthetic popularity);
violet nodes, unknown substances; blue halo, protection needed.
steps. Chematica proposed a novel and shorter route starting with a concise, three-
component aza-Henry reaction21,22 of aryl amine 1, aldehyde 2, and nitroalkane 3.
This reaction gave the acyclic adduct 4 and its diastereomer in 78% yield and with
Chem 4, 1–11, March 8, 2018
5

Please cite this article in press as: Klucznik et al., Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed
in the Laboratory, Chem (2018), https://doi.org/10.1016/j.chempr.2018.02.002
Figure 3. Syntheses of the Second Set of Four Targets Performed in the Grzybowski and Mrksich
Laboratories
For a Figure360 author presentation of Figure 3, see https://doi.org/10.1016/j.chempr.2018.02.
002mmc3.
(S)-4-hydroxyduloxetine (A), 5b/6b-hydroxylurasidone (B), dronedarone (D), and engelheptanoxide
C (D). Color coding of nodes is the same as in Figure 2.
82:18 distribution of anti:syn products. Intermediate 6 was then sulfonylated with 7 in
69% yield, and the desired enantiomer 8 was isolated by chiral supercritical fluid chro-
matography in 40% yield. Overall, Chematica’s route improved the yield by six to eight
times and required fewer (three versus five) FCC separations.
6
Chem 4, 1–11, March 8, 2018

Please cite this article in press as: Klucznik et al., Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed
in the Laboratory, Chem (2018), https://doi.org/10.1016/j.chempr.2018.02.002
The second target was a-hydroxyetizolam (14 in Figure 2B), which is one of two main
metabolites used as indicators of etizolam usage in humans.23 Although the synthe-
sis of benzodiazepines is well established, making hydroxy metabolites presents
more of a challenge and MS previously deemed 14 as ‘‘too risky’’ to prioritize.
Chematica’s proposal is unique in that it rests on a multicomponent Gewald reac-
tion. Although this method has been used to prepare similar hydroxylated thio-
phene intermediates, this particular side chain has not been reported.24 Chematica’s
route started with protected 3-hydroxybutanal 11 (need for protection, indicated in
the software’s plan by a blue halo) and nitrile 10 (purchased rather than made from
bromoketone 9). These substrates were allowed to react with sulfur under classic
Gewald conditions to afford the desired thiophene 12. The remaining steps followed
classic benzodiazepine transformations to give the desired a-hydroxyetizolam 14 in
3.2% overall yield (see Supplemental Information, Section S11).
The third target was a potent and selective ATR kinase inhibitor denoted as 21 in Fig-
ure 2C. The synthesis of this target has been described previously25 and involves
seven steps (see Supplemental Information, Section S12). Although the overall liter-
ature-reported yield is about 20%, yields that MS were able to achieve on numerous
attempts with the published route remained below 10%. In addition, the literature
route is not readily amenable to scale-up. In this light, Chematica’s short, four-
step solution shown in Figure 2C appeared attractive, such that intermediate 18
was prepared in two rather than five published steps (cf. Supplemental Information,
Section S12). This difference does not stem from using drastically different types of
chemistries along the route but rather from a choice of a different starting material
(as mentioned, Chematica "knows" some 200,000+ commercial substrates plus
ca. 7,000,000 known molecules), which avoids some unnecessary interconversion
of functional groups. When executed, Chematica’s route offered an overall, repro-
ducible, and gram-scalable yield of 20%–22%, and provided 30% time savings
(45 versus 62 hr for the published protocol).
The fourth target, marked as 29 in Figure 2D, has been shown26 to selectively inhibit
proliferation and induced differentiation of human acute-myeloid-leukemia cells. The
reported26 preparation of 29 is, despite optimization efforts at MS, low yielding
(1% overall; see Supplemental Information, Section S13), requires four chromato-
graphic separations, and is not readily amenable to scale-up for commercialization.
Chematica’s route was attractive in that it proposed Suzuki coupling before forming
the desired amide (which, in the literature pathway, was the lowest-yielding step,
10%). The route starting from the bromide 24 was chosen because it was commercially
available. Conditions for the requisite Suzuki-Miyaura coupling were optimized with a
KitAlysis Reaction Screening Kit.27 Although several of the screening conditions were
successful, the SPhos Pd G2 catalyst was chosen, and the desired biphenyl 26 could
be purified without chromatography by recrystallizing the hydrochloride salt of
the crude product. The hydrochloride salt was then converted to the amide 28 via
Schotten-Baumann acylation, followed by aqueous hydrolysis of 28 and neutralization
to afford 29 in 72% yield. In this step, the 2-Cl-pyridine 27 rather than the 2-OH deriv-
ative was used even though the latter could potentially avoid the hydrolysis step.
Although Chematica identified both solutions, it decided to use 2-Cl on the basis of
its higher synthetic popularity and significantly lower price. Overall, this approach al-
lowed for minimal chromatographic purification steps, gave product in gram quantities
with a significantly improved 60% yield, and offered time and cost savings of ꢀ50%.
Turning to the syntheses performed by students in the Grzybowski and Mrksich lab-
oratories, the fifth target was (S)-4-hydroxyduloxetine, 34 in Figure 3A, which is one
Chem 4, 1–11, March 8, 2018
7

Please cite this article in press as: Klucznik et al., Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed
in the Laboratory, Chem (2018), https://doi.org/10.1016/j.chempr.2018.02.002
of the main metabolites of duloxetine (the blockbuster drug Cymbalta). This
‘‘simple-looking’’ target was chosen because of its importance in clinical testing,
high market price (ꢀ$200/mg), and because MS had previously failed to reproduce
the only literature route28 involving a nucleophilic aromatic substitution between an
alcohol and O-protected-4-fluoronaphthol but published without basic experi-
mental details. Chematica suggested a much simpler route detecting no counterin-
dications for the key Mitsunobu reaction between 32 and 33. Indeed, this reaction
proceeded neatly in 58% yield and was followed by simultaneous deprotection of
Fmoc and OAc protecting groups in 84% yield. The overall yield of the entire
pathway was 20% (see Supplemental Information, Section S14 for further details).
The sixth target was 5b/6b-hydroxylurasidone (43, 430 in Figure 3B), which is the main
active metabolite of lurasidone, a US Food and Drug Administration-approved atyp-
ical antipsychotic drug for the treatment of schizophrenia and bipolar disorder.29
The objective for Chematica in this exercise was to design an efficient pathway
that would avoid the only known but patented route30 (see Supplemental Informa-
tion, Section S15). The path designed by Chematica starts from anhydride 35 and
diol 36. The former was converted into imide 37 in quantitative yield, whereas the
latter was protected (indicated by a blue halo in Chematica’s plan) with tert-butyldi-
methylsilyl chloride and then iodinated into 38 in 94% yield. The imide was alkylated
with the iodide to give 39 in 93% yield. This was followed by deprotection in 87%
yield, activation to iodide 40 (99%), N-alkylation of 1,2-benzisothiazole 41 to pro-
duce 42 (82%), and finally hydroxylation of the double bond. We note that in the
original patent,30 hydroxylation was performed early on, before the potentially frag-
ile 1,2-benzisothiazole moiety was installed. However, Chematica judged that this
moiety would not be affected during hydroxylation, as indeed was verified experi-
mentally and rewarded by an over 90% yield (versus 37% in the patented route).
All in all, the route proved easily scalable to multigram quantities and gave an overall
yield of ꢀ55%; that is, two and a half times higher than reported in the patented
route.
The ability to find routes significantly different from patented syntheses (see list of
46 patents in the Supplemental Information, Section S16.1) also motivated the syn-
thesis of the seventh target, Sanofi-Aventis’ dronedarone (51 in Figure 3C). Chema-
tica’s route began with a Sonogashira coupling between aryl iodide 44 and 1-hexyne
to give 45. The next, key step was the palladium-catalyzed carbonylative annulation
of alkyne 45, aryl iodide 46, and carbon monoxide.31,32 This three-component reac-
tion was not used in any previous syntheses of dronedarone, but we found it to pro-
ceed neatly in 76% yield to construct the entire central ring system of the target. Sub-
sequent steps were straightforward (see Supplemental Information, Section S16),
resulting in an overall pathway yield of 39.6%, which is comparable with the 41%
yield of the Sanofi-Aventis synthesis;33 the latter, however, starts from a nitrobenzo-
furan derivative (see Supplemental Information, Section S16.1), which is a more
advanced intermediate than the simple 2-iodo-4-nitrophenol starting material
Chematica selected (on the flipside of the coin, the iodide generated in the
Sonogashira reaction might be costly to dispose of, which could be problematic if
the synthesis was ever carried out on industrial scales).
Finally, the eighth target was engelheptanoxide C, 56, which is a natural product
recently isolated from stems of Engelhardia roxburghiana but not yet synthesized.19
The main virtue of Chematica’s pathway in Figure 3D is that the program was able to
construct an elegant and convergent route by using chemistries most appropriate for
this type of a scaffold. Specifically, a ‘‘modern’’ enantioselective allylation of a primary
8
Chem 4, 1–11, March 8, 2018

Please cite this article in press as: Klucznik et al., Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed
in the Laboratory, Chem (2018), https://doi.org/10.1016/j.chempr.2018.02.002
alcohol 53 according to Krische’s protocol34 ([Ir(cod)Cl]2, Cs2CO3, (R)-BINAP) pro-
ceeded in 65% yield and 93% ee and avoided a step that a ‘‘classic’’ aldehyde allylation
would entail (requiring oxidation and involving fragile alkylaldehyde). This step set the
first stereocenter. The remaining two stereocenters were created in one step via the
Prins cyclization of the tetrahydropyran ring.35,36 Depending on the conditions used,
this reaction gave either 45% yield and 72% ee or 30% yield and 88% ee (see
Supplemental Information, Section S17). In both cases, the synthesis of 56 was
completed by quantitative hydrogenation carried over Pd(OH)2/C.
Conclusion
In summary, after over a decade of laborious development, Chematica is finally
capable of designing novel and experimentally efficient syntheses of medicinally
and industrially relevant targets. Guided by its scoring functions promoting synthetic
brevity and penalizing any reactivity conflicts or non-selectivities, the program finds
solutions that might be hard to identify by a human user; for instance, four of the
described syntheses rely on hard-to-spot multicomponent reactions and in at least
one case (5b/6b-hydroxylurasidone), the program made a choice that a chemist
might consider ‘‘risky.’’ However, these approaches and choices are not a question
of Chematica’s luck but rather manifestation of its dexterity in analyzing and discon-
necting complex graphs coupled with its ability to consider simultaneously large
numbers of logically related criteria (of groups’ cross-reactivity, selectivity, etc.).
Looking forward, future development and wider dissemination of Chematica
are critically reliant on the expansion of the underlying computer infrastructure
(i.e., multiprocessor machines potentially linked into larger clusters) required
for the exploration of extremely large retrosynthetic trees. With such expansion
backed by MS and in close academic collaborations with several leading synthetic
groups, Chematica’s next and perhaps ultimate aim is to attack the syntheses of
very complex targets at the forefront of modern synthesis.
EXPERIMENTAL PROCEDURES
Full experimental procedures are provided in the Supplemental Information.
SUPPLEMENTAL INFORMATION
Supplemental Information includes algorithmic details of Chematica and experi-
mental details of the syntheses, 132 figures, 8 schemes, and 1 movie and can be
found with this article online at https://doi.org/10.1016/j.chempr.2018.02.002.
ACKNOWLEDGMENTS
S.S., K.M., E.P.G., P.D., Y.Z., M.M., and B.A.G. thank the US Defense Advanced
Research Projects Agency for generous support under the ‘‘Make-It’’ Award,
69461-CH-DRP #W911NF1610384. T.K., B.M.K., A.A., M.P.S., and B.S. gratefully
acknowledge support from the Symfonia Award (#2014/12/W/ST5/00592) from
the Polish National Science Center (NCN). R.R. was supported through the Merck
funds and also the FUGA postdoctoral program from NCN (#2016/20/S/ST5/
00361). B.A.G. also gratefully acknowledges personal support from the Institute
for Basic Science Korea, Project Code IBS-R020-D1. We would like to thank Profes-
sor Phil Baran for providing helpful comments during preparation of the manuscript.
AUTHOR CONTRIBUTIONS
S.S., K.M., E.P.G., P.D., M.P.S., R.R., and T.K. were key developers of Chematica.
M.B. synthesized the BRD7/9 inhibitor. H.L. synthesized a-hydroxyetizolam. A.T.
synthesized ATR kinase inhibitor. M.P.M. synthesized the anti-leukemia drug
Chem 4, 1–11, March 8, 2018
9

Please cite this article in press as: Klucznik et al., Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed
in the Laboratory, Chem (2018), https://doi.org/10.1016/j.chempr.2018.02.002
candidate and T.K. and B.M.-K. synthesized 5b/6b-hydroxylurasidone, dronedar-
one, and (S)-4-hydroxyduloxetine. A.A. and B.S. helped in the synthesis of
(S)-4-hydroxyduloxetine, and L.R. and G.J.K. provided helpful advice; B.S. also
helped prepare substrates for the synthesis of dronedarone. Y.Z. synthesized engel-
heptanoxide C under supervision of M.M. S.L.J.T. conceived and supervised the
technology evaluation project on behalf of MilliporeSigma. B.A.G. conceived
Chematica in graduate school and has directed its development ever since. All au-
thors contributed to the writing of the manuscript.
DECLARATION OF INTERESTS
After the successful experiments outlined in this manuscript, Chematica was ac-
quired by Merck KGaA, Darmstadt, Germany. For various access options, including
academic collaborations, please contact S.L.J.T. Part of the current research on
Chematica is sponsored by Merck KGaA.
Received: November 11, 2017
Revised: December 23, 2017
Accepted: February 1, 2018
Published: March 1, 2018
REFERENCES AND NOTES
ꢀ
1. Corey, E.J., and Wipke, W.T. (1969). Computer-
assisted design of complex organic syntheses.
Science 166, 178–192.
10. Szymkuc, S., Gajewska, E.P., Klucznik, T.,
14. Bishop, K.J.M., Klajn, R., and Grzybowski, B.A.
(2006). The core and most useful molecules in
organic chemistry. Angew. Chem. Int. Ed. 45,
5348–5354.
Molga, K., Dittwald, P., Startek, M., Bajczyk, M.,
and Grzybowski, B.A. (2016). Computer-
assisted synthetic planning: the end of the
beginning. Angew. Chem. Int. Ed. 55, 5904–
5937.
2. Gelernter, H.L., Sanders, A.F., Larsen, D.L.,
Agarwal, K.K., Boivie, R.H., Spritzer, G.A., and
Searleman, J.E. (1977). Empirical explorations
of SYNCHEM. Science 197, 1041–1049.
15. Grzybowski, B.A., Bishop, K.J.M., Kowalczyk,
B., and Wilmer, C.E. (2009). The ‘‘wired’’
universe of organic chemistry. Nat. Chem. 1,
31–36.
11. The literature on computational predictions
validated in the synthetic practice is very
scarce. In our own work,10 we showed that
Chematica was able to suggest several
synthetic pathways in which each individual
step had been previously confirmed in
experimental papers by other groups.
Bøgevig et al.9 used the ICSYNTH program
to suggest several synthetic pathways,
which, however, were not carried out
3. Hanessian, S., Franco, J., and Larouche, B.
(1990). The psychobiological basis of heuristic
synthesis planning – man, machine and the
Chiron approach. Pure Appl. Chem. 62, 1887–
1910.
16. Kowalik, M., Gothard, C.M., Drews, A.M.,
Gothard, N.A., Wieckiewicz, A., Fuller, P.E., and
Grzybowski, B.A. (2012). Parallel optimization
of synthetic pathways within the network of
organic chemistry. Angew. Chem. Int. Ed. 51,
7928–7932.
4. Hendrickson, J.B. (1977). Systematic synthesis
design. 6. Yield analysis and convergency.
J. Am. Chem. Soc. 99, 5439–5450.
17. Gunther, M. (2016) Revolutionary computer
program could change chemistry forever.
Chemistry World. https://www.chemistryworld.
com/news/software-could-revolutionise-
chemistry/1017236.article.
because the funding of the project was cut
(only one cyclization reaction leading to
oxaspiroketone was performed). One other
notable contribution was from Ivar Ugi, who
used his bond-electron matrix formalism to
propose and then execute several novel
pericyclic reactions and a novel
5. Bauer, J., Herges, R., Fontain, E., and Ugi, I.
(1985). IGOR and computer-assisted
innovation in chemistry. Chimia 39, 43–53.
6. Ugi, I., Bauer, J., Bley, K., Dengler, A., Dietz, A.,
Fontain, E., Gruber, B., Herges, R., Knauer, M.,
Reitsam, K., et al. (1993). Computer-assisted
solution of chemical problems – the historical
development and the present state of the art of
a new discipline of chemistry. Angew. Chem.
Int. Ed. 32, 201–227.
18. Lowe, D. (2016) The algorithms are coming.
http://blogs.sciencemag.org/pipeline/archi-
ves/2016/04/12/the-algorithms-are-coming.
rearrangement of a-aminoalkylboranes.6
More recently, Currie and Goodman12
used a combination of reaction-template-
matching and energy calculations to predict
as feasible (under acidic conditions) and then
execute experimentally (but under basic
conditions) a retro-Claisen rearrangement of
an intermediate hemiacetal in the multistep
synthesis of (ꢁ)-dolabriferol. The remaining
18 steps of the synthesis were planned
manually (i.e., not by the computer).
19. Wu, H.C., Cheng, M.J., Peng, C.F., Yang, S.C.,
Chang, H.S., Lin, C.H., Wang, C.J., and Chen,
I.S. (2012). Secondary metabolites from the
stems of Engelhardia roxburghiana and their
antitubercular activities. Phytochemistry 82,
118–127.
7. Hollering, R., Gasteiger, J., Steinhauer, L.,
¨
Schultz, K., and Herwig, A. (2000). Simulation of
organic reactions: from the degradation of
chemicals to combinatorial synthesis. J. Chem.
Inf. Comput. Sci. 40, 482–494.
20. Clark, P.G.K., Vieira, L.C.C., Tallant, C.,
Fedorov, O., Singleton, D.C., Rogers, C.M.,
Monteiro, O.P., Benett, J.M., Baronio, R.,
Mueller, S., et al. (2015). LP99: discovery and
synthesis of the first selective BRD7/9
bromodomain inhibitor. Angew. Chem. Int. Ed.
54, 6217–6221.
8. Ravitz, O. (2013). Data-driven computer aided
synthesis design. Drug Discov. Today Technol.
10, e443–e449.
12. Currie, R.H., and Goodman, J.M. (2012). In
silico inspired total synthesis of
(ꢁ)-Dolabriferol. Angew. Chem. Int. Ed. 51,
4695–4697.
9. Bøgevig, A., Federsel, H.-J., Huerta, F.,
Hutchings, M.G., Kraut, H., Langer, T., Low, P.,
¨
21. It is worth mentioning that the aza-Henry
reaction required a large excess (ꢀ20 equiv) of
methyl 4-nitrobutanoate, which contrasts with
conditions reported in literature for other
substrates.22 Also, in the presence of strong
acids or bases, 4 reverts back to starting amine
13. Fialkowski, M., Bishop, K.J., Chubukov, V.A.,
Campbell, C.J., and Grzybowski, B.A. (2005).
Architecture and evolution of organic
chemistry. Angew. Chem. Int. Ed. 44, 7263–
7269.
Oppawsky, C., Rein, T., and Saller, H. (2015).
Route design in the 21st century: the ICSYNTH
software tool as an idea generator for synthesis
prediction. Org. Process. Res. Dev. 19,
357–368.
10
Chem 4, 1–11, March 8, 2018

Please cite this article in press as: Klucznik et al., Efficient Syntheses of Diverse, Medicinally Relevant Targets Planned by Computer and Executed
in the Laboratory, Chem (2018), https://doi.org/10.1016/j.chempr.2018.02.002
1 by a retro aza-Henry reaction. A small amount
of the retro aza-Henry product was also
observed during LiOH-mediated hydrolysis of
4. Finally, we note that for this reaction,
Chematica provided conditions from Cruz-
Acosta et al.,22 where either enantioselective or
diastereoselective catalyst systems were
reported. Because Schreiner’s catalyst was
commercially available (i.e., the
Discovery of 4-{4-[(3R)-3-methylmorpholin-4-
yl]-6-[1-(methylsulfonyl)cyclopropyl]pyrimidin-
2-yl}-1H-indole (AZ20): a potent and selective
inhibitor of ATR protein kinase with
monotherapy in vivo antitumor activity. J. Med.
Chem. 56, 2125–2138.
31. Hu, Y., Zhang, Y., Yang, Z., and Fathi, R.J.
(2002). Palladium-catalyzed carbonylative
annulation of o-alkynylphenols: syntheses of
2-substituted-3-aroyl-benzo[b]furans. J. Org.
Chem. 67, 2365–2368.
32. Nakamura, M., Ilies, L., Otsubo, S., and
Nakamura, E. (2006). 2,3-Disubstituted
benzofuran and indole by copper-mediated
CꢁC bond extension reaction of
26. Getlik, M., Smil, D., Zepeda-Velazquez, C.,
Bolshan, Y., Poda, G., Wu, H., Dong, A.P.,
Kuznetsova, E., Marcellus, R., Senisterra, G.,
et al. (2016). Structure-based optimization of a
small molecule antagonist of the interaction
between WD repeat-containing protein 5
(WDR5) and mixed-lineage leukemia 1 (MLL1).
J. Med. Chem. 59, 2478–2496.
enantioselective catalyst didn’t need to be
synthesized in house), the diastereoselective
route was pursued.
3-zinciobenzoheterole. Org. Lett. 8, 2803–2805.
33. Gubin, J., Chatelain, P., Lucchetti, J., Rosseells,
G., Inion H. (1993). Alkylaminoalkyl derivatives
of benzofuran, benzothiophene, indole and
indolizine, process for their preparation and
compositions containing them. Patent US
5223510, filed July 26, 1991, and published
June 29, 1993.
22. Cruz-Acosta, F., de Armas, P., and Garcia-
Tellado, F. (2013). Water-compatible
hydrogen-bond activation: a scalable and
organocatalytic model for the stereoselective
multicomponent Aza-Henry reaction. Chem.
Eur. J. 19, 16550–16554.
27. Sigma-Aldrich. KitAlysis High-Throughput
Screening Kits. https://www.sigmaaldrich.com/
chemistry/chemical-synthesis/kitalysis-high-
throughput-screening-kits.html.
34. Kim, I.S., Ngai, M.-Y., and Krische, M.J. (2008).
Enantioselective iridium-catalyzed carbonyl
allylation from the alcohol or aldehyde
oxidation level via transfer hydrogenative
coupling of allyl acetate: departure from
chirally modified allyl metal reagents in
carbonyl addition. J. Am. Chem. Soc. 130,
14891–14899.
23. Miki, A., Tatsuno, M., Katagi, M., Nishikawa, M.,
and Tsuchihashi, H. (2002). Simultaneous
determination of eleven benzodiazepine
hypnotics and eleven relevant metabolites
in urine by column-switching liquid
28. Kuo, F., Gillespie, T.A., Kulanthaivel, P., Lanz,
R.J., Ma, T.W., Nelson, D.L., Threlkeld, P.G.,
Wheeler, W.J., Yi, P., and Zmijewski, M. (2004).
Synthesis and biological activity of some known
and putative duloxetine metabolites. Bioorg.
Med. Chem. Lett. 14, 3481–3486.
chromatography-mass spectrometry. J. Anal.
Toxicol. 26, 87–93.
24. Stransky, W., Weber, K., Walther, G., Harreus,
A., Stenzel, J.C., Muacevic, G., and Bechtel,
W.D. (1990). Thieno-1,4-diazepines. Patent US
4900729 A, filed October 11, 1988, and
published February 13, 1990.
29. Meyer, J.M., Loebel, A.D., and Schweizer, E.
(2009). Lurasidone: a new drug in development
for schizophrenia. Expert Opin. Investig. Drugs
18, 1715–1726.
35. Tadpetch, K., and Rychnovsky, S.D. (2008).
Rhenium(VII) catalysis of Prins cyclization
reactions. Org. Lett. 10, 4839–4842.
30. Kojima, A., Ono, Y., and Yoshigi, M. (1996).
New imide derivative. Patent JP 08333368, filed
June 9, 1995, and published December 17,
1996.
36. Crosby, S.R., Harding, J.R., King, C.D., Parker,
G.D., and Willis, C.L. (2002). Oxonia-Cope
rearrangement and side-chain exchange in the
Prins cyclization. Org. Lett. 4, 577–580.
25. Foote, K.M., Blades, K., Cronin, A., Fillery, S.,
Guichard, S.S., Hassal, L., Hickson, I., Jacq, X.,
Jewsbury, P.J., McGuire, T.M., et al. (2013).
Chem 4, 1–11, March 8, 2018
11