Across-the-World Automated Optimization and Continuous-Flow Synthesis of Pharmaceutical Agents Operating Through a Cloud-Based Server

Article abstract of DOI:10.1002/anie.201809080

The power of the Cloud has been harnessed for pharmaceutical compound production with remote servers based in Tokyo, Japan being left to autonomously find optimal synthesis conditions for three active pharmaceutical ingredients (APIs) in laboratories in Cambridge, UK. A researcher located in Los Angeles, USA controlled the entire process via an internet connection. The constituent synthetic steps for Tramadol, Lidocaine, and Bupropion were thus optimized with minimal intervention from operators within hours, yielding conditions satisfying customizable evaluation functions for all examples.

Full text of DOI:10.1002/anie.201809080

A Journal of the Gesellschaft Deutscher Chemiker
Angewandte
International Edition Chemie
www.angewandte.org
Accepted Article
Title: Across-the-world automated optimization and continuous flow
synthesis of pharmaceutical agents operating through a cloud-
based server
Authors: Daniel Ewert Fitzpatrick, Timothé Maujean, Amanda Evans,
and Steven Victor Ley
This manuscript has been accepted after peer review and appears as an
Accepted Article online prior to editing, proofing, and formal publication
of the final Version of Record (VoR). This work is currently citable by
using the Digital Object Identifier (DOI) given below. The VoR will be
published online in Early View as soon as possible and may be different
to this Accepted Article as a result of editing. Readers should obtain
the VoR from the journal website shown below when it is published
to ensure accuracy of information. The authors are responsible for the
content of this Accepted Article.
To be cited as: Angew. Chem. Int. Ed. 10.1002/anie.201809080
Angew. Chem. 10.1002/ange.201809080
Link to VoR: http://dx.doi.org/10.1002/anie.201809080
http://dx.doi.org/10.1002/ange.201809080

Angewandte Chemie International Edition
10.1002/anie.201809080
COMMUNICATION
Across-the-world automated optimization and continuous flow
synthesis of pharmaceutical agents operating through a cloud-
based server
[a]
[b]
[c]
[a]
Daniel E. Fitzpatrick, Timothé Maujean, Amanda Evans and Steven V. Ley*
Abstract: The power of the Cloud has been harnessed for
pharmaceutical compound production with remote servers based in
Tokyo, Japan being left to autonomously find optimal synthesis
conditions for three active pharmaceutical ingredients (APIs) in
laboratories in Cambridge, UK. A researcher located in Los Angeles,
USA controlled the entire process via an internet connection. The
constituent synthetic steps for Tramadol, Lidocaine and Bupropion
were thus optimized with minimal intervention from operators within
hours, yielding conditions satisfying customizable evaluation
functions for all examples.
Here we demonstrate the ability to work across international
borders and time domains by harnessing the Cloud through
servers operating in Japan to produce three active
pharmaceutical ingredients (APIs). Processes are managed from
Los Angeles, USA, with equipment and chemicals located in
Cambridge, UK. The autonomous nature of this set up helps with
the efficient use of equipment located at remote sites to avoid
redundancy. The system[12] can be rapidly reconfigured to
accommodate new reaction combinations and safe shutdown
sequences, and is securely protected through appropriate
firewalls and other IT security apparatus (refer to SI). Compound
delivery at distant sites in this fashion has consequences for
The demands of modern day synthesis often go beyond the task
of simply assembling a particular target molecule and include
additional evaluation criteria whereby cost, efficiency, robustness
and sustainability can also be key factors.[1] Furthermore, this
leads to the need for the discovery of greater and more diverse
reactivity patterns together with improved optimization tools[2] and
other enabling technologies[3] to facilitate levels of automation.
Deeper reaction understanding, data acquisition and mining with
machine learning algorithms are fueling opportunities in artificial
intelligence (AI) and machine intuition for example.[4]
[13]
future access to medicines across the world
with wider
applications also being possible through enhanced collaboration,
including expanded access to specialized knowledge and
equipment.
Further, with continued modern developments in inexpensive
microcontrollers[14–17] and computers, such as the Raspberry Pi, a
standardization of reaction protocols can be established through
the use of an automated control system. Such a process would
[18]
integrate with the future Internet of Chemical Things, potentially
improving reproducibility and data collection for experimentalists,
feeding deep learning algorithms of the future.
New protocols are necessary for maximizing resource-
accelerated synthesis,[5] which is an area where flow chemistry
methods and continuous processing can demonstrate distinct
advantages.[6,7] Automation of repetitive and trivial reaction
sequences expedites development, leading to more efficient
processing paradigms, particularly when integrated with
numerous unit operations involving in-line work-up and reaction
telescoping techniques.
To demonstrate the utility of the approach, tramadol (3),
lidocaine (8) and bupropion (12) were studied as representative
agents.
Tramadol
We began our study with the preparation of (±)-tramadol (3). The
most common synthetic pathway to this compound follows two
steps: the formation of amine 1 via a Mannich condensation;
followed by a Grignard addition to yield the final product 3 as a
mixture of diastereomers. For this first example, we focused solely
on the second addition step as this has been the subject of
continued investigation.[19,20]
Unlike previous elegant studies of multi-step flow synthesis of
natural products[8] and active pharmaceutical ingredients
(
APIs),[9,10] the work reported here breaks new ground in that there
is a move away from using pre-optimized reaction combinations
towards self-optimized processing[11] through information
feedback as reaction monitoring.
A straightforward equipment layout was constructed (Figure
1a), consisting of two reagent supply lines, a 20 mL reactor coil
and FlowIR unit to provide spectroscopic performance feedback.
Additional valves and solvent reservoirs were included with the
supply lines to facilitate a reactor flushing step to minimize risk of
disruption arising from solid aggregation (refer to SI for more
information).
[
a]
Dr. D.E. Fitzpatrick, Prof. S.V. Ley
Department of Chemistry
University of Cambridge
Lensfield Road, Cambridge, UK.
E-mail: svl1000@cam.ac.uk
T. Maujean
[
b]
c]
Département de Chimie
Ecole Normale Supérieure Paris Saclay
The Los Angeles-based operator configured the system
settings to optimize for three parameters: temperature (between
94235 Cachan Cedex, France.
30 °C – 70 °C); residence time (5 min – 20 min); and equivalents
[
Prof. A. Evans
of Grignard 2 to ketone 1 (0.5 eq. to 1.6 eq.). For this reaction the
control system did not just optimize for conversion, but also for
material throughput and starting material consumption (Figure 1b).
The primary feedback parameter, namely conversion, was
established by comparing the ratio of the IR absorption peaks
corresponding to ketone 1 and target product 3 (Figure 1c). The
Department of Chemistry & Biochemistry
California State University Fullerton
800 N. State College Blvd.
Fullerton, CA, 92831 USA
Supporting information for this article is given via a link at the end of
the document.
This article is protected by copyright. All rights reserved.

Angewandte Chemie International Edition
10.1002/anie.201809080
COMMUNICATION
complex method[21] was used to drive the optimization process for
this reaction, and in all subsequent examples.
Figure 2. (a) The modern synthesis route to lidocaine (8) follows a two step
process; (b) Equipment layout used for the self-optimization of the first step to
lidocaine (8); (c) four term evaluation function used to optimize the first step,
where τ is residence time, p is product IR absorbance (chloroacetamide 6), s is
starting material IR absorbance (amine 4), x is equivalents of acid chloride 5
supplied and T is reactor temperature.
Overall yields obtained when conducting this synthesis under
flow conditions vary greatly. Raston et al. reported an overall yield
of 85 % for the segmented-telescoping[23] of both steps, where no
intermediate purification was performed between reaction steps
Figure 1. (a) Equipment layout for the three-dimensional self-optimization of
tramadol (3). Feedback from an inline infrared spectrometer (IR) was used by
the control system to drive optimization; (b) the evaluation function for the
optimization of tramadol (3), where τ is residence time, p is product IR
absorbance (compound 3), s is starting material IR absorbance (compound 1)
and x is the equivalents of Grignard reagent 2 supplied to the reactor; (c)
but differences in reaction flow rates prevented continuous
telescoping.[
24]
This yield reduced to 15 % when modified
conditions for a fully telescoped process were implemented. More
recently, Jamison et al. reported an isolated yield of 90 % of
lidocaine (8) for a fully telescoped process.[9] Each of the fully
continuous processes could produce lidocaine (8) at rates
averaging 22.5 g day-1 and 16.2 g day , respectively. We set out
to test whether an automated self-optimizing approach would
produce improved results notably with minimal researcher
intervention.
Overlaid IR spectra of target product
wavenumbers monitored for each.
3 and starting material 1, and
-1
The system autonomously conducted nine experimental trials
over the space of three hours (refer to SI), which identified
conditions that gave 86 % yield (NMR) during one hour of steady
state operation. In this case, operating at 41 °C with a residence
time of 10.9 min and with 1.6 equivalents of Grignard reagent
gave the most favorable result, equivalent to a production rate of
We began these efforts with the acylation of 2,6-
dimethylaniline (4) with chloroacetyl chloride (5) to give
intermediate 6. Equipment layout resembled that used previously
for tramadol (Figure 2b). As in this first example, three parameters
were selected for optimization: temperature (40 °C – 130 °C);
residence time (5 min – 25 min); and equivalents of acid chloride
1,651 drug doses per day.
The volumetric yield[22] for this process of 0.172 g mL-1 h-1
represented a significant improvement over a comparable flow
-1
process optimized previously by Rencurosi et al. (0.045 g mL
h ).
5
to amine 4 (0.8 eq. – 2.5 eq.). In addition to the throughput,
-1 [20]
conversion and consumption terms in the evaluation function, we
also included an energy term to encourage the unsupervised
system to consider the energy impact of the reaction process
Lidocaine
In the next example lidocaine (8), a local anesthetic, was
produced under full self-optimization conditions via a two-step
synthesis process (Figure 2a). Producing this compound in flow
provides a number of benefits, including improved handling of
particularly hazardous reagents such as acid chloride 5 and
increased thermal control.
(
Figure 2c).
Within 2 hours 40 minutes, the system identified conditions
(105 °C, 5.0 min, 1.98 eq. of 5; refer to SI) that gave an isolated
yield of 87 %. The system was then left to run at steady state for
2.5 hours to produce 39.7 g (201 mmol) of intermediate 6, a
portion of which was used for the next optimization. Thus, within
one working day, the control system had moved from unoptimized
conditions to a set up that produced almost 40 g of material – with
minimal intervention from researchers.
This article is protected by copyright. All rights reserved.

Angewandte Chemie International Edition
10.1002/anie.201809080
COMMUNICATION
-
1
The second step consisted of an amine alkylation, where
intermediate 6 was reacted with diethylamine (7) to give lidocaine
8). Equipment layout was similar to the first stage (Figure 3a),
with differences arising only in the composition of feedstock
solutions and reservoirs. We wished to test how our version of
Complex implementation would react to a scenario where
chemists specifically chose to maximize the production rate of the
target compound. Thus, the evaluation function contained
throughput and conversion terms (Figure 3b). As for the previous
examples, temperature (70 °C – 130 °C), residence time (5 min –
shift in characteristic peaks between the ketone 9 (1216 cm ) and
-
1
brominated product 10 (1300 cm ). Here, the system manipulated
temperature (30 °C – 80 °C), residence time (5 min – 20 min) and
stoichiometry (0.95 eq. – 2.0 eq. of bromine to starting material 9)
to drive optimization. Within three hours, the system performed
nine experiments identifying conditions which gave 95 % yield
(44 °C, 9.7 min and 0.95 eq. bromine), corresponding to 8.1 g
hr-1 of the target material.
(
30 min) and stoichiometry (1.0 eq. – 4.0 eq. of amine 7 to
chloroacetamide 6) were optimized.
Figure 4. (a) Equipment layout for the three-dimension self-optimization of the
first step to bupropion (12); (b) the evaluation function consisted of three terms,
where τ is residence time, p is product IR absorbance (bromide 10), s is starting
material IR absorbance (ketone 9) and x is equivalents of bromine supplied to
the reactor.
Figure 3. (a) Equipment layout for the optimization of the amine alkylation
reaction to form lidocaine (8); (b) the evaluation function for the amine alkylation
reaction to lidocaine (8) consisted of throughput and conversion terms only (τ is
residence time, p is product IR absorbance, s is starting material 6 IR
absorbance).
With these results in hand, the system was then reconfigured
to optimize the second and final step which consisted of the amine
alkylation of intermediate 10 with tert-butylamine. Initial attempts
to perform this reaction in DCM yielded poor results, which were
improved through the use of NMP (Figure 5a).
This optimization process spanned three hours, during which
the control system conducted ten iterations (refer to SI),
identifying an optimum at a reactor temperature of 99 °C,
residence time of 17.8 minutes and 3.9 eq. of amine 7 to
chloroacetamide 6. This optimum gave 98 % isolated yield. A
large amount of material remained in the feed reservoirs and so
the system was left operating under these optimized conditions
for four hours, allowing 15.7 g of lidocaine (8) to be isolated
following purification.
We were therefore very pleased that the system had identified
optimal conditions for the two-step synthesis (85 % overall yield)
within two working days.
Figure 5. (a) Equipment layout for the optimization of the amine alkylation
reaction to form bupropion (12); (b) single-term evaluation function for the final
step to bupropion (12), where p is product IR absorbance and s is starting
material IR absorbance (bromide 10).
Bupropion
In the final example, we wished to demonstrate how
effectively and rapidly the system would perform when moving
from unexplored conditions to a fully telescoped, optimized
process for a two-step synthesis. Bupropion (12), a smoking
cessation aid and anti-depressant, presented an excellent
opportunity to showcase the platform’s capabilities.
As for the first step, temperature (30 °C – 90 °C), residence
time (5 min – 30 min) and stoichiometry (0.95 eq. – 3.0 eq. of tert-
butylamine (11) to bromide 10) were chosen as parameters to
drive optimization. For this case, we wished to observe how the
system would react when only conversion was optimized (Figure
The first step of the synthetic route to bupropion (12)
consisted of the α-bromination of 3’-chloropropiophenone (9) to
yield the intermediate bromide 10. Given the corrosive nature of
bromine solution when in contact with stainless steel equipment,
an inert peristaltic pump replaced an HPLC pump used for
previous optimizations (Figure 4a). The system incorporated IR
feedback in the evaluation function (Figure 4b), monitoring the
5
b). Within 4.5 hours, the system optimized to 80 % yield (90 °C,
0 min, 3.0 eq.) having performed 11 experiments. These
2
conditions allowed for the production of 0.72 g hr-1 of bupropion.
The final challenge was to telescope both synthetic steps
together, enabling the continuous production of bupropion.
Although conditions for the first and last segments had been
This article is protected by copyright. All rights reserved.

Angewandte Chemie International Edition
10.1002/anie.201809080
COMMUNICATION
identified through the self-optimization processes, suitable
workup actions were still required to process the crude reaction
mixture from the first step into a form that was compatible with the
second. To achieve this, crude bromination reaction mixture was
mixed vigorously with an aqueous sodium bisulfite stream. The
organic phase, containing the bromide intermediate, was then
mixed with an NMP stream and directed into a thin-film
evaporation column (refer to SI) where 87 % of DCM was
removed (molar basis). The resulting NMP-enriched stream was
passed to the second reactive step.
The telescoped process consisted of four unit operations
which could be separated into three segments (Figure 6a): α-
bromination; inter-stage workup; and the final amine alkylation
transformation.
Two modifications were made to the reaction conditions to
maximize efficiency of the telescoped process. Firstly, the ratio of
bromine to ketone 9 in the first step was raised to 1.0 to push the
reaction towards quantitative yield. Secondly, the concentration of
amine 11 in the feed reservoir was set to 1.928 M to align
optimized reaction conditions with the concentration of bromide
10 leaving the evaporation unit.
The control system was configured to follow three distinct
stages, corresponding to process start-up, steady-state operation
and process shutdown (refer to SI). During start-up, IR feedback
was used to detect the presence of the bromide intermediate in
the workup stream and thus to trigger the start of the second
reaction step. This process enabled staggering of the operations,
minimizing material wastage and supervisory oversight required
from operators.
Figure 6b shows the data collected from the infrared
spectrometer during all phases of the telescoped process. After
an initial period of transient behavior, the first reaction and workup
sequences reached steady state after approximately 0.6 hours. It
was interesting to observe how process disturbances propagated
through the system during 8 hours of steady state operation,
giving an indication of the sensitivity that inline detectors can bring
when monitoring multistage sequences.
For example, at approximately 6.9 hours the NMP reservoir
emptied. Although it was refilled within 3 minutes, during this time
the feed into the evaporation unit consisted only of intermediate
10 in DCM. Thus, a disproportionate amount of this compound
was added to the NMP solution that had yet to be removed from
the column. This led to the 8 minute concentration fluctuation
shown by the rise and fall in infrared absorbance as the system
returned to steady state.
Figure 6. (a) The telescoped synthesis of bupropion (12) consisted of two
reaction steps and two downstream processing steps, and involved 7 items of
equipment; (b) data collected from the FlowIR during steady state operation
process, which include the effects of a process disturbance at approx. 6.9 hours.
Conclusions
We were pleased with the results obtained for this target
compound: the control system had facilitated the rapid transition
from an unexplored route to an optimized and telescoped process,
capable of producing bupropion at an average rate of 2.88 g h ,
within four working days.
In summary, we have demonstrated that the use of an
automated control system, unencumbered by location or time
domain, has the capability to greatly assist with drug development
and synthesis, liberating researchers to spend time on more
productive pursuits and assisting with ideas linked to delocalized
manufacturing. This proof-of-concept approach was applied to
previously-reported synthetic routes of three API targets, in which
a colleague in the US remotely initiated, monitored and controlled
self-optimization reactions conducted using equipment in the UK,
-1
This article is protected by copyright. All rights reserved.

Angewandte Chemie International Edition
10.1002/anie.201809080
COMMUNICATION
[
[
[
[
4]
5]
6]
7]
C. W. Coley, W. H. Green, K. F. Jensen, Acc. Chem. Res. 2018, 51,
via servers in Japan. In all cases, optimized conditions were found
within hours and, in the case of bupropion, it was possible to
devise a fully optimized and telescoped system within four
working days from the initial idea and synthesis plan.
1281–1289.
S. Newton, C. F. Carter, C. M. Pearson, L. de C Alves, H. Lange, P.
Thansandote, S. V. Ley, Angew. Chem. Int. Ed. 2014, 53, 4915–4920.
S. V. Ley, D. E. Fitzpatrick, R. J. Ingham, R. M. Myers, Angew. Chem.
Int. Ed. 2015, 54, 3449–3464.
S. V Ley, D. E. Fitzpatrick, R. M. Myers, C. Battilocchio, R. J. Ingham,
Angew. Chem. Int. Ed. 2015, 54, 10122–10136.
Harnessing the Cloud for such reactions presents exciting
opportunities to accelerate synthetic optimization, share
standardized reaction procedures across the world, contribute to
machine learning algorithms of the future and facilitate distributed
use of equipment. We have demonstrated that synthesis does not
need to be trapped in typical confined environments but can be
opened up to promote research collaborations, maximize
resources and establish reliable robust synthesis protocols
beyond today’s practices using conventional methods.
[8]
J. C. Pastre, D. L. Browne, S. V. Ley, Chem. Soc. Rev. 2013, 42, 8849-
8869.
[
[
9]
A. Adamo, R. L. Beingessner, M. Behnam, J. Chen, T. F. Jamison, K.
F. Jensen, J.-C. M. Monbaliu, A. S. Myerson, E. M. Revalor, D. R.
Snead, T. Stelzer, N. Weeranoppanant, S. Y. Wong, P. Zhang, Science
2016, 352, 61–67.
10] For selected examples, see: a) M. Baumann, I. R. Baxendale, Beilstein
J. Org. Chem. 2015, 11, 1194–1219; b) L. Kupracz, A. Kirschning, Adv.
Synth. Catal. 2013, 35, 3375-3380; c) J. Hartwig, A. Kirschning, Chem.
Eur. J. 2016, 22, 3044-3052.
[11] D. C. Fabry, E. Sugiono, M. Rueping, Isr. J. Chem. 2014, 54, 341–350.
[
[
[
12] D. E. Fitzpatrick, C. Battilocchio, S. V. Ley, Org. Process Res. Dev.
015, 20, 386–394.
13] J. L. Howard, C. Schotten, D. L. Browne, React. Chem. Eng. 2017, 2,
81–287.
14] P. L. Urban, Angew. Chem. Int. Ed. 2018, Accepted Manuscript, DOI:
0.1002/anie.201803878.
2
2
Experimental Section
1
Refer to Supporting Information.
[15] R. J. Ingham, C. Battilocchio, J. M. Hawkins, S. V. Ley, Beilstein J. Org.
Chem. 2014, 10, 641–652.
[
[
[
[
[
16] M. O’Brien, P. Koos, D. L. Browne, S. V. Ley, Org. Biomol. Chem.
012, 10, 7031–7036.
17] S. V. Ley, R. J. Ingham, M. O’Brien, D. L. Browne, Beilstein J. Org.
Chem. 2013, 9, 1051–1072.
18] S. V. Ley, D. E. Fitzpatrick, R. J. Ingham, N. Nikbin, Beilstein Magazine
2015, DOI: 10.3762/bmag.2.
19] S. Kobayashi, K. Shibukawa, Y. Miyaguchi, A. Masuyama, Asian J.
Org. Chem. 2016, 5, 636–645.
2
Acknowledgements
The authors gratefully acknowledge financial support from the Woolf
Fisher Trust (DEF), EPSRC (SVL; grant codes EP/K009494/1,
EP/K039520/1 and EP/M004120/1) and H2020-FETOPEN-2016-2017
programme of European commission (DEF, SVL; grant agreement
number: 737266-ONE FLOW).
20] E. Riva, S. Gagliardi, M. Martinelli, D. Passarella, D. Vigo, A.
Rencurosi, Tetrahedron 2010, 66, 3242–3247.
[21] R. F. Kazmierczak Jr., SSRN Electronic Journal. 1996, DOI:
0.2139/ssrn.15071.
1
[
22] Defined here as the mass of product produced per unit reactor volume
per unit time.
Keywords: self-optimization • flow chemistry • continuous
[
23] Segmented-telescoping refers to procedures where materials are
collected and held in reservoirs after steps are carried out
independently. Only after the upstream reaction has completed is
material fed into the next step.
synthesis
[
1]
D. E. Fitzpatrick, C. Battilocchio, S. V. Ley, ACS Cent. Sci. 2016, 2,
31–138.
1
[24] J. Britton, J. M. Chalker, C. L. Raston, Chem. Eur. J. 2015, 21, 10660–
[
[
2]
3]
V. Sans, L. Cronin, Chem. Soc. Rev. 2016, 45, 2032–2043.
D. E. Fitzpatrick, S. V. Ley, Tetrahedron 2018, 74, 3087–3100.
10665.
This article is protected by copyright. All rights reserved.

Angewandte Chemie International Edition
10.1002/anie.201809080
COMMUNICATION
This article is protected by copyright. All rights reserved.