Augmenting Adaptive Machine Learning with Kinetic Modeling for Reaction Optimization

Article abstract of DOI:10.1021/acs.joc.1c01038

We combine random sampling and active machine learning (ML) to optimize the synthesis of isomacroin, executing only 3% of all possible Friedl?nder reactions. Employing kinetic modeling, we augment machine intuition by extracting mechanistic knowledge and

Full text of DOI:10.1021/acs.joc.1c01038

pubs.acs.org/joc
Note
Augmenting Adaptive Machine Learning with Kinetic Modeling for
Reaction Optimization
A. Filipa Almeida, Filipe A. P. Ataíde,* Rui M. S. Loureiro, Rui Moreira, and Tiago Rodrigues*
Cite This: https://doi.org/10.1021/acs.joc.1c01038
Read Online
sı
*
Metrics & More
Article Recommendations
Supporting Information
ACCESS
ABSTRACT: We combine random sampling and active machine
learning (ML) to optimize the synthesis of isomacroin, executing
̈
only 3% of all possible Friedlander reactions. Employing kinetic
modeling, we augment machine intuition by extracting mechanistic
knowledge and verify that a global optimum was obtained with
ML. Our study contributes evidence on the potential of multiscale
approaches to expedite the access to chemical matter, further
democratizing organic chemistry in a data-motivated fashion.
kinase (PDGFRα) modulation while using self-organizing
ynthetic organic chemistry toward high value bioactive
Sentities and materials is key in modern molecular maps for target deorphanization.11,12 The tractability of 1 as a
prototype for medicinal chemistry was demonstrated by the
development of potent PDGFR13,14 and IκB kinase15,16
inhibitors based on the imidazolyl quinoline scaffolda
hinge binding motif (Figure 1). However, synthetic access to
medicine, but often delivers suboptimal processes and
insufficient amounts of chemical matter for advanced func-
tional profiling.1−3 Indeed, the lack of screening materials for
biological investigations may curb or delay the identification of
potentially life-changing therapies. Natural products (NPs)
have long been exploited as therapeutics or as a source of
inspiration for molecular design due to the biological
prevalidation of their frameworks as protein-binding motifs.4−6
In fact, ca. 33% of approved small-molecule drugs are either
NPs or NP-derived compounds, which highlights their value in
translational science.7
Fragment-like NPs usually provide simpler and synthetically
more accessible architectures that can also be readily adopted
for myriad discovery chemistry applications.8,9 For example, we
had unveiled that isomacroin (1, Figure 1)a fragment-like
NP from Macrorungia longistrobus10presented the blueprints
for efficient platelet-derived growth factor receptor alpha
̈
1 through Friedlander quinoline synthesis proved challenging
in our hands. Poor (15%) yield17 was obtained, which limited
screening efforts on a wider scale.
With that problem in mind, we have recently developed
LabMate.MLa self-evolving machine-learning (ML) routine
for digitalizing reaction optimizations under low data
regimeswith the goal of providing an interpretable, general-
izable, and nonexhaustive search space exploration alternative
to full/fractional factorial “design-of-experiments”.18 The
method leverages information in a small set of random
reactions for initialization, and adaptive random forests
powered by a bespoke experiment selection policy to iteratively
drive the optimization process. In a previous report, and in
alignment with related methods,19−21 we had shown that
LabMate.ML not only efficiently modeled real-valued and
categorical variables but, more importantly, was competitive
with expert intuition in competition studies. Taking advantage
of its decision process, interpretable outputs could also be
extracted to augment domain intuition.18
Notwithstanding the prior success of LabMate.ML, daily
practice in process chemistry optimization more frequently
Special Issue: Enabling Techniques for Organic Syn-
thesis
Received: May 3, 2021
Figure 1. Structure of isomacroin, 1, and related molecules as human
kinase inhibitors. The framework of 1 is highlighted in the NP-
derived, bioactive small molecules.
https://doi.org/10.1021/acs.joc.1c01038
J. Org. Chem. XXXX, XXX, XXX−XXX
© XXXX American Chemical Society
A

The Journal of Organic Chemistry
pubs.acs.org/joc
Note
exploits kinetic modeling to motivate reaction conditions lead
to productive syntheses.22,23 Ultimately, this approach allows a
data-informed elucidation of reaction mechanisms and tailored
protocol design.22,24 Herein, we build on our prior ML
workflow and provide a preliminary proof of concept for
cascading active learning and kinetic modeling as an approach
toward fast reaction optimization, including the identification
of a global optimum. Using 1 as the model compound, our ML
forest routinea supervised learning algorithm that harnesses
the “wisdom of the crowds” concept by aggregating individual
predictions from uncorrelated decision trees. Because each tree
analyzes a fraction of data, only weak predictions are
individually expected as output. However, averaging those
over n trees allows for a significantly more realistic prediction
and uncertainty estimation. Together, these characteristics
endorse the popularity of random forests in the chemical
sciences17,31,32 and make them robust to common pitfalls in
ML research, such as outliers, noise, and overfitting.33,34
To initialize the optimization campaign, 20 random
reactions were performed. Their outcomes (target variable/
yield) were collected via a calibration curve that was built from
the area under the curve for the required product peak at
different concentration values (HPLC−UV/vis−MS traces).
Interestingly, one such random reaction corresponded to
literature conditions,11 but afforded a higher yield in this
reassessment (76% here vs 15% literature). The obtained yield
was still used as benchmark, yet highlights the previously
discussed variability in chemical data that can determine the
predictive power of statistical learning.35 All random reaction
conditions and their yields (0−76%, Figure 2b) were employed
as an initial training set, and a stratified 10-fold cross-validation
routine, wherein analyses are repeated with 90% of data used
for training and 10% as internal test set, was implemented for a
preliminary assessment of the model utility. By harnessing a
previously validated18 “greedy” approach for experiment
prioritization, i.e., selecting reactions with high-predicted
yield and low uncertainty, our recommender tool rapidly
(over five iterations) converged to an optimum (Figure 2b).
Relative to the literature conditions, the ML routine elected to
substitute the base (KOH to t-BuOK) and its molar equivalent
amount (from 1.25 to 1.50 molar equiv) to afford 1 in 88%
yield (cf. Table S1). Our understanding is that such
modifications are reasonable and among the top variables a
skilled chemist would also probe, which supports a correct
formalization of chemical intuition by the computational tool.
Additionally, shuffling of target (Y) variables resulted in less
predictive control models. This evidenced that meaningful,
true patterns in the data structure had been disrupted in the Y-
randomization process (cf. Table S3). We deemed this control
necessary due to the high likelihood of ML heuristics
exploiting artifacts and memorizing rather than learning
data.36−39 This is especially critical when dealing with low
data in high dimensional space, as in the present case. Finally,
we studied if different baseline algorithms, e.g., linear, lasso,
and ridge regression, could provide similar or better statistical
models relative to our random forests while enforcing
simplicity in the decision process. Our method proved more
accurate and efficient in navigating the training data, as
assessed through RMSE and MAE values (cf. Table S2).
Overall, the results attest to the appropriateness and robustness
of adaptive random forests for assisting chemists in reaction
optimization problems.
̈
tool rapidly exploited a discretized Friedlander reaction space.
The generated insights were subsequently augmented with
kinetic modeling to ascertain a reactivity pathway and
corresponding rate/equilibrium constants, which to the best
of our knowledge have not been determined.25 We also
confirmed that the ML routine had rapidly converged to a
global optimum (88%) with minimal experimental effort.
Overall, our multiscale, human-in-loop approach may prove
transferable to other chemistries and enable the swift access to
high value chemical matter by democratizing organic synthesis.
To delve into the reactivity space and identify an optimal
synthesis protocol toward 1, we initially surveyed the literature
and collected preferable conditions that could be experimen-
tally probed26−30 (Figure 2a). These included different
Figure 2. Optimization of reaction conditions toward isomacroin, 1.
(a) Six reaction variables were probed in the optimization of a
̈
Friedlander quinoline synthesis. Both real-valued and categorical
features were considered, affording a search space of 756 possible
reactions. Categorical variables were one-hot encoded for modeling.
Real value variables were employed for modeling without any
transformation. (b) Optimization routine, following a user-defined
number of random reactions for initialization (orange) and an
additional five ML-suggested reactions until reasonable convergence
was obtained (blue).
solvents and their volume, bases and molar amount, temper-
ature, and reaction time. The reaction variables were encoded
as real values (e.g., volume) or strings as in the case of
categorical (e.g., base) variables.
For machine interpretation and modeling, categorical
variables were then one-hot encoded, denoting either the
presence (“1”) or absence (“0”) of a specified entity.
Collectively, 756 reactions were digitalized in high dimensional
space, which in this case constitutes more combinations than
are reasonably feasible to screen. This search space was
subjected to iterative investigation by our adaptive random
It is now established that different ML algorithms can afford
expert level predictions in numerous tasks relevant to the
chemical sciences.40−42 However, it is still usual to harness ML
as black boxes, wherein no insight into the decision process is
provided. This can not only hinder the adoption of potentially
useful technologies but also bypass on new knowledge that
might otherwise remain hidden to perception.43 To shed light
onto the key variables for our statistical model and infer on
their directionality, i.e., desired or undesired, we extracted
B
https://doi.org/10.1021/acs.joc.1c01038
J. Org. Chem. XXXX, XXX, XXX−XXX

The Journal of Organic Chemistry
pubs.acs.org/joc
Note
importance values from the regressor (Figure 3a) and
employed SHapley Additive ExPlanations (SHAP; Figure 3b)
values for the 2-aminobenzaldehyde starting material (0.94,
1.0, 1.45, 2.0), t-BuOK (1.25, 1.5, 2.2, 2.9), and temperature
values (40, 53, and 78 °C; cf. Table S4). In these experiments,
reaction mixture aliquots were collected at specified time-
points and all relevant species were quantified through
HPLC−UV/vis−MS (e.g., Figure 4a). This allowed us
Figure 3. ML interpretation on a global scale. (a) The importance of
each feature in the decision process of the adaptive random forest
model is provided over five iterations, until reasonable target variable
(yield/output) convergence was obtained. Data shows dynamic
feature importance fluctuations as more data becomes available in
the active learning process. Categorical variables are decoupled in the
plot. (b) SHAP44 was employed to infer on the directionality of each
feature for the model/reaction output/yield. Low reaction volumes,
higher temperature, and longer reaction times are preferred, together
with using EtOH as solvent and NaOH or t-BuOK as base. (c)
Visualization of the search space with UMAP (Uniform Manifold
Approximation and Projection),45 which preserves the local
neighborhood structure in the data set. Data shows specific islets of
reactivity and preferred regions leading to the formation of
isomacroin.
Figure 4. Kinetic modeling as an orthogonal approach to augment
machine-learning outputs. (a) Exemplary, overlaid HPLC−UV traces
(0.5 and 6 h) showing the formation of isomacroin over time. (b)
Exemplary plot showing the evolution of species concentration over
time and fitted curve. Kinetic parameters were extracted from a series
of similar experiments.
monitoring the concentration increase/decrease of each
starting material, product, and byproduct over 360 min of
reaction (e.g., Figure 4b). While reaction intermediates may
provide an additional layer of information and confidence in
downstream data inference, such species were not quantifiable
through our analytical method. We thus assumed their
transient nature and did not consider them for modeling.
as an orthogonal, model-agnostic approach. SHAP fits linear
models to provide both global (model) and local (individual
prediction) data interpretations.39
̈
A kinetic model for the Friedlander quinoline synthesis was
then developed based on differential equations that describe
the reaction rate as a function of concentration change over
time for all above-mentioned species. This valuable informa-
tion was modeled in DynoChem (Scale-up Systems46) by
minimizing the sum of squares between the experimental data
and model predictions using the gradient-based Levenberg−
Marquardt algorithm.
̈
In this Friedlander quinoline synthesis, small reaction
volumes were crucial, as was refluxing for a long period of
time (Figure 3a,b). Indeed, without encoding any explicit
chemical knowledge, our ML routine was able to autono-
mously formalize unwritten rules of intuition, which corrob-
orates its utility. These interpretations support that reaction
optimizations can effectively be configured as data mining
problems. Furthermore, with manifold learning (Figure 3c) we
confirmed that all reactions cocluster in specific islets of
reactivity that were exploited by our algorithm, and that each
cluster can afford 1 in moderate yields. This result provides a
readily visualizable interpretation of the pursued selection
policy. Moreover, it is possible to observe that the iterative
search chiefly revolved around the best performing random
reactions in the initialization step, without undermining the
identification of an optimum (e.g., iteration 2).
The fitted model led us to establish the reaction mechanism
toward 1, in line with the literature25 (Figure 5). To fully
profile the transformation, we extracted the first- and second-
order rates (k), equilibrium (Keq) constants, and activation
energy (Ea) values at a 95% confidence interval level (cf. Table
With this result in hand, we next wondered if the initial
search space discretization step had limited the ability to
identify a global optimum and, therefore, if ML served the
purpose of a swift yet coarse-grained interrogation of the
̈
Friedlander reactivity space with additional room for improve-
ment. To that end, we set up a series of experiments to model
the reaction kinetics, identify a global optimum, and establish a
mechanism. Specifically, we explored different molar equivalent
̈
Figure 5. Friedlander chemistry mechanism according to kinetic
modeling with DynoChem.
C
https://doi.org/10.1021/acs.joc.1c01038
J. Org. Chem. XXXX, XXX, XXX−XXX

The Journal of Organic Chemistry
pubs.acs.org/joc
Note
doublet of doublets, ddt = doublet of doublet of triplets, and m =
multiplet. A high-performance liquid chromatography (HPLC)
(Waters Corp., Milford, MA) system was used to quantify the
isomacroin, as product, and the samples were collected from the
chemical reactions. The chromatographic analysis was performed in
Agilent Eclipse XDB-C18 column (150 mm × 4.6 mm i.d. 3.5 μm) at
room temperature. The product was detected by ultraviolet (UV)
absorbance detection at 254 nm. The analyses were performed under
an appropriate gradient of ammonium acetate/acetonitrile (90:10) at
a pH of 8.5 and acetonitrile/ammonium acetate (90:10) and a total
flow rate of 1 mL/min. The injection volume was 5 μL, and the total
run time was 20.1 min. The ultraperformance liquid chromatography
(UPLC) analyses were performed on an ACQUITY UPLC system
equipped with a photodiode array (PDA) detector (Waters Corp.,
Milford, MA) coupled to a mass single-quadrupole detector (QDa
Waters). This system was used to identify the starting raw materials,
product, and byproducts, and the samples were collected from the
chemical reactions. All compounds were monitored at 254 nm by
PDA detector. The single ion recording (SIR) method was established
on a single quadruple mass detector. The QDa conditions were set as
follows: a cone voltage of 15 V, a capillary voltage of 0.8 kV, and a
source temperature of 600 °C. The data were acquired under the SIR
mode. The column, injection volume, flow rate, and solvent managers
were already described previously.
Synthesis of 1-(1-Methyl-1H-imidazol-5-yl)ethan-1-one. To a
flask containing 1-methyl-1H-imidazole (0.485 mL, 0.50 g, 6.0 mmol)
and tetrahydrofuran (2.87 mL) at −78 °C was added 1.6 M n-BuLi in
hexane (4.0 mL, 6.5 mmol) followed by stirring at −78 °C for 40 min.
Then chlorotrimethylsilane (0.8 mL, 6.3 mmol) was added slowly,
and the mixture was stirred at −78 °C for 1 h. The 1.6 M n-BuLi in
hexane (4.0 mL, 6.5 mmol) was added, the cooling bath was removed,
and stirring was continued until the temperature reached 10 °C. The
mixture was recooled to −78 °C, and a solution of N,N-
dimethylacetamide (0.463 mL, 5.0 mmol) was added. The cooling
bath was removed, and the stirring was continued for 40 min at room
temperature. The reaction was quenched with a few drops of
methanol (1.0 mL), and brine was added. The organic layer was
separated, and the aqueous layer was extracted with dichloromethane.
The combined organic phases were dried (sodium sulfate anhydrous),
filtered, and concentrated under reduced pressure. The crude was
filtered through a short silica plug and eluting with ethyl acetate (75
mL). The solvent was evaporated to afford the required compound.
Colorless oil, 93% (0.703 g), Rf 0.68 (1/1 v/v ethyl acetate/heptane).
1H NMR (300 MHz, (CD3)2CO): δ 7.28−7.24 (m, 1H), 7.00 (d, J =
1.0 Hz, 1H), 3.92 (s, 3H), 2.48 (s, 3H). 13C{1H} NMR (75 MHz,
(CD3)2CO): δ 189.7, 128.6, 127.4, 35.2, 26.1. As described in the
literature.47
S5). The reaction rate predictions suggest that formation of 1
occurs rapidly, with the dehydration step being rate-limiting.
The results also indicate that consumption of 2-amino-
benzaldehyde occurs through an aldol condensation toward
intermediate 2. Moreover, 1 is amenable to slow degradation
relative to the elimination step when excess 2-amino-
benzaldehyde and t-BuOK are employed. Domain knowledge
informs that formation of a Schiff base is not viable under basic
conditions. As an adversarial control to the proposed
mechanism, we fitted a model assuming the formation of an
imine as first step. The results are in line with established
intuition and assert the formation of 2 (kimine = 0.01 M−1 s−1 vs
0.05 M−1 s−1), thus providing an additional layer of confidence
on our kinetic model.
On a broader perspective, we confirmed the ML predictions
through an orthogonal means, as both methods in our
investigation converged to identical reaction protocols and
predicted yields. More specifically, kinetic modeling suggests
that a global optimum, with minimization of the byproduct, is
achieved at 78 °C and utilizing 2 and 1.05 molar equiv of 2-
aminobenzaldehyde and t-BuOK, respectively, over 593 min of
reaction (Figure S15). There is, however, value in coalescing
both approaches into a streamlined workflow. The kinetic
model augmented the ML-derived intuition by elucidating the
reaction mechanism and rate constants. It also informed the
impact of the byproduct for the development of an optimized
synthesis protocol, albeit requiring time-consuming computa-
tion (ca. 5−10 min for ML vs 220−250 min for kinetic
model). Together, this substantiates the power of pattern
identification through ML. It also provides a motivation for
employing computationally expensive methods when added
value is warranted.
In conclusion, we implemented a cascaded workflow
comprising adaptive learning and kinetic modeling to facilitate
̈
the access to 1 via Friedlander chemistry and afford physical
chemistry insights. The whole optimization processincluding
featurization, random selection, hyperparameter tuning, and
ML selectionallowed the prioritization of experiments and
identification of an optimum while executing only a minute
amount (3%) of all possible reactions. This is relevant because
ML can equally work as an optimizer or fine-tuner of
experiments, using poor or good yields as starting points,
respectively. Further, the process also allowed establishing a
mechanistic path to the transformation. Our study provides
proof of concept for a viable integration of well-established
concepts in process chemistry with emerging technologies.
Ultimately, it may impact on molecular medicine pipelines by
expediting the access to high value chemical matter for
screening purposes. We foresee this and similar integrations
working as research assistants, promoting probabilistically
informed decisions and democratizing organic syntheses in the
digital chemistry era.
Synthesis of 2-Aminobenzaldehyde. A solution of 2-nitro-
benzaldehyde (0.50 g, 3.3 mmol) in ethanol (9.4 mL) was stirred
for approximately 1 min. Iron powder (0.554 g, 9.9 mmol) and dilute
hydrochloric acid (3.3 mL of 1.0 M HCl, 0.5 mmol) were added to
the stirred solution, and the reaction was heated to reflux for 2 h in an
Asynt DrySyn Single Position Blocks system. The reaction mixture
was cooled to room temperature, diluted with ethyl acetate (27.0
mL), and stirred for 5 min before being filtered through a short Celite
plug. The filtrate was evaporated under reduced pressure to yield a
yellow oil. The product was stored at −20 °C. Yellow oil, 100% (0.40
1
g), Rf 0.44 (1/5 v/v ethyl acetate/hexane). H NMR (300 MHz,
(CD3)2CO) δ 9.88 (d, J = 0.6 Hz, 1H), 7.55 (dd, J = 7.8, 1.6 Hz, 1H),
7.31 (ddd, J = 8.5, 7.0, 1.6 Hz, 1H), 6.81 (dq, J = 8.3, 0.7 Hz, 2H),
6.74−6.66 (m, 1H). 13C{1H} NMR (75 MHz, (CD3)2CO) δ 193.7,
1508, 135.6, 134.9, 118.6, 115.9, 115.4. As described in literature.48
Synthesis of Isomacroin (1). The ketone (0.10 g, 0.80 mmol, 1
molar equiv) and potassium hydroxide (0.057 g, 1.0 mmol, 1.25 molar
equiv) were dissolved in ethanol (6 mL/mmol). Then the 2-
aminobenzaldehyde (0.098g, 0.8 mmol, 1 molar equiv) was added to
the solution and the reaction mixture refluxed 3 h in a Asynt DrySyn
Single Position Blocks system. The solvent was evaporated, and the
crude was purified by flash column chromatography with heptane/
ethyl acetate (3:1) eluent. Yellow solid, 80% (0.135 g), Rf 0.40 (1/1
EXPERIMENTAL SECTION
■
General Methods. Starting materials and reagents were
purchased from Sigma-Aldrich, Alfa Aesar, Fluka, or Acros and used
without further purification. Reactions were carried out on a Radleys
1
Carousel 6 Plus Reaction Station. H NMR spectra were obtained at
on a Bruker Avance 300 MHz in CD3OD and (CD3)2CO with
chemical shift values (δ) in parts per million using residual solvent
peaks as the internal standard, and 13C NMR spectra were obtained at
75 MHz in the same deuterated solvents. Coupling constants (J) are
reported in hertz with the following splitting abbreviations: s = singlet,
d = doublet, t = triplet, dd = doublet of doublets, ddd = doublet of
D
https://doi.org/10.1021/acs.joc.1c01038
J. Org. Chem. XXXX, XXX, XXX−XXX

The Journal of Organic Chemistry
pubs.acs.org/joc
Note
v/v ethyl acetate/hexane). 1H NMR (300 MHz, CD3OD): δ 8.34 (d,
J = 8.7 Hz, 1H), 8.18 (dd, J = 8.6, 0.8 Hz, 1H), 8.05 (ddt, J = 8.5, 1.4,
0.7 Hz, 1H), 7.82−7.77 (m, 1H), 7.69 (ddd, J = 8.5, 6.9, 1.5 Hz, 1H),
7.51 (ddd, J = 8.1, 6.9, 1.2 Hz, 1H), 7.18 (d, J = 1.1 Hz, 1H), 7.02 (d,
J = 1.1 Hz, 1H), 4.30 (s, 3H). 13C{1H} NMR (75 MHz, CD3OD) δ
150.6, 147.3, 145.0, 136.3, 129.5, 129.4, 128.6, 127.6, 127.2, 126.5,
125.2, 120.6, 36.6.
AUTHOR INFORMATION
Corresponding Authors
■
Filipe A. P. Ataíde − R&D, Process Chemistry Development,
̂
Hovione FarmaCiencia S.A, 1649-038 Lisboa, Portugal;
orcid.org/0000-0002-8288-2152; Email: fataide@
hovione.com
Tiago Rodrigues − Research Institute for Medicines
(iMed.Ulisboa), Faculty of Pharmacy, Universidade de
Lisboa, 1649-003 Lisboa, Portugal; orcid.org/0000-
0002-1581-5654; Email: tiago.rodrigues@ff.ulisboa.pt
Machine Learning. We used random forest regressors and probed
several hyperparameters [number of trees (100−1000), tree depth
(none, 2, 4) and number of features (auto or sqrt)] to create a
prediction model. A 10-fold cross-validation was applied, splitting the
data set into a training group with 90% of data and test group with the
remaining 10%. The selection of best hyperparameters was guided by
the calculation of the mean absolute error (MAE). With these
parameters, the model predicted the product yield for the remaining
possible reaction conditions, and then the next reaction was selected.
An exploration approach was applied for the first three iterations to
allow model improvement. For the next iterations, an exploitative
approach was applied by selecting the reaction with the lowest
variance among the predicted top-5 high-yielding reactions. The
model develops with each added data point by improving its
predictive model. The model and data analyses were fully
implemented in Python 3.7.3 using the NumPy 1.16.4, Pandas
0.24.2 and Scikit-learn 0.21.2 libraries and was run (5−10 min) on an
HP ProBook G3 (2.40 GHz 2 core processor, 4 Gb RAM). For
initialization, 20 random reactions were performed in parallel (Table
S1) and analyzed through HPLC to create the first training data set.
Using a calibration curve, it was possible to quantify the product yield
obtained in each reaction. Based on this data, the machine learning
routine selected one additional set of conditions, at a time, for
experimental validation. The output for the selected experiment was
added to training set and an iterative process involving design−
make−test was pursued over five iterations. Code for the optimization
routine is accessible at https://github.com/tcorodrigues/LabMate.
ML.
Machine Learning Controls. To evaluate the robustness of the
method, other models, such as linear regression, ridge, and lasso, were
computed. In parallel, we performed y-randomization studies to rule
out artifacts in the decision process of our random forests. Error
metrics, such as the root mean square error (RMSE), mean absolute
errors (MAE), and coefficient of determination (r2) were calculated
and used for comparison between the different methods.
Kinetic Method. The DynoChem software (Scale-Up Systems;
Build: 1.3.20b340, Data Version: 1.0) was used for reaction modeling.
The software includes model templates, provides simulator, fitting,
and optimization add-in options to run the models. In this work, we
used a single model template to fit all experiments and calculate
kinetic parameters. We employed the Levenberg−Marquardt
algorithm for fitting, during minimization of SSQ, and the Backward
Euler (BE) integration method to solve the mass balance differential
and algebraic equations associated with the reaction. The kinetic
parameters were calculated at a defined reference temperature (78
°C). All reactions were run for 360 min and aliquots collected at
specified time points. Analyses were performed in HPLC and based
on calibration curves.
Authors
A. Filipa Almeida − R&D, Process Chemistry Development,
̂
Hovione FarmaCiencia S.A, 1649-038 Lisboa, Portugal;
Research Institute for Medicines (iMed.Ulisboa), Faculty of
Pharmacy, Universidade de Lisboa, 1649-003 Lisboa,
Portugal
Rui M. S. Loureiro − R&D, Process Chemistry Development,
̂
Hovione FarmaCiencia S.A, 1649-038 Lisboa, Portugal
Rui Moreira − Research Institute for Medicines
(iMed.Ulisboa), Faculty of Pharmacy, Universidade de
Lisboa, 1649-003 Lisboa, Portugal
Complete contact information is available at:
https://pubs.acs.org/10.1021/acs.joc.1c01038
Author Contributions
A.F.A. performed chemistry and modeling. All authors
designed experiments and analyzed data. A.F.A and T.R.
wrote the manuscript with contributions from the remaining
authors. F.A.P.A., R.M.S.L., R.M., and T.R. supervised the
research. F.A.P.A. and T.R. coordinated the study. All authors
have given approval to the final version of the manuscript.
Notes
The authors declare the following competing financial
interest(s): T.R. is co-founder and shareholder of TargTex S.A.
ACKNOWLEDGMENTS
A.F.A. and T.R. acknowledge Fundaca
■
̂
̧
o
̃
para a Ciencia e
Tecnologia (FCT) Portugal for financial support through PD/
BD/143125/2019 and CEECIND/00684/2018, respectively.
This research received funding from European Structural &
Investment Funds through the COMPETE Program
Programa Operacional Regional de LisboaProgram Grant
LISBOA-01-0145- FEDER-016405, and from National Funds
̂
through the FCTFundaca para a Ciencia e a Tecnologia
̧
o
̃
Program Grant SAICTPAC/0019/2015. FCT is also acknowl-
edged for support of the MedChemTrain PhD programme
(PD147-2013-AA). The authors acknowledge Ricardo Gon-
ca̧ lves (Hovione) for great support in designing the analytical
method to quantify all species.
REFERENCES
■
(1) Hayashi, Y. Time Economy in Total Synthesis. J. Org. Chem.
2021, 86 (1), 1−23.
ASSOCIATED CONTENT
■
sı
(2) Peiretti, F.; Brunel, J. M. Artificial Intelligence: The Future for
Organic Chemistry? ACS Omega 2018, 3 (10), 13263−13266.
* Supporting Information
The Supporting Information is available free of charge at
https://pubs.acs.org/doi/10.1021/acs.joc.1c01038.
(3) Pfluger, P. M.; Glorius, F. Molecular Machine Learning: The
̈
Future of Synthetic Chemistry? Angew. Chem., Int. Ed. 2020, 59 (43),
18860−18865.
1H and 13C NMR spectra for all compounds, UPLC-
QDa chromatograms for all compounds, data and
conditions used in machine learning model and kinetic
model (PDF)
(4) Harvey, A. L. Natural Products in Drug Discovery. Drug
Discovery Today 2008, 13 (19−20), 894−901.
(5) Atanasov, A. G.; Zotchev, S. B.; Dirsch, V. M.; Orhan, I. E.;
Banach, M.; Rollinger, J. M.; Barreca, D.; Weckwerth, W.; Bauer, R.;
E
https://doi.org/10.1021/acs.joc.1c01038
J. Org. Chem. XXXX, XXX, XXX−XXX

The Journal of Organic Chemistry
pubs.acs.org/joc
Note
Bayer, E. A.; Majeed, M.; Bishayee, A.; Bochkov, V.; Bonn, G. K.;
Braidy, N.; Bucar, F.; Cifuentes, A.; D’Onofrio, G.; Bodkin, M.;
Diederich, M.; Dinkova-Kostova, A. T.; Efferth, T.; El Bairi, K.;
Arkells, N.; Fan, T.-P.; Fiebich, B. L.; Freissmuth, M.; Georgiev, M. I.;
Gibbons, S.; Godfrey, K. M.; Gruber, C. W.; Heer, J.; Huber, L. A.;
Ibanez, E.; Kijjoa, A.; Kiss, A. K.; Lu, A.; Macias, F. A.; Miller, M. J. S.;
(17) Rodrigues, T.; Reker, D.; Welin, M.; Caldera, M.; Brunner, C.;
Gabernet, G.; Schneider, P.; Walse, B.; Schneider, G. De Novo
Fragment Design for Drug Discovery and Chemical Biology. Angew.
Chem., Int. Ed. 2015, 54 (50), 15079−15083.
(18) Reker, D.; Hoyt, E. A.; Bernardes, G. J. L.; Rodrigues, T.
Adaptive Optimization of Chemical Reactions with Minimal
Experimental Information. Cell Reports Phys. Sci. 2020, 1 (11),
100247.
(19) Häse, F.; Roch, L. M.; Kreisbeck, C.; Aspuru-Guzik, A.
Phoenics: A Bayesian Optimizer for Chemistry. ACS Cent. Sci. 2018, 4
(9), 1134−1145.
(20) Strieth-Kalthoff, F.; Sandfort, F.; Segler, M. H. S.; Glorius, F.
Machine Learning the Ropes: Principles, Applications and Directions
in Synthetic Chemistry. Chem. Soc. Rev. 2020, 49 (17), 6154−6168.
(21) Shields, B. J.; Stevens, J.; Li, J.; Parasram, M.; Damani, F.;
Alvarado, J. I. M.; Janey, J. M.; Adams, R. P.; Doyle, A. G. Bayesian
Reaction Optimization as a Tool for Chemical Synthesis. Nature
2021, 590 (7844), 89−96.
(22) Blackmond, D. G. Reaction Progress Kinetic Analysis: A
Powerful Methodology for Mechanistic Studies of Complex Catalytic
Reactions. Angew. Chem., Int. Ed. 2005, 44 (28), 4302−4320.
(23) Wang, K.; Han, L.; Mustakis, J.; Li, B.; Magano, J.; Damon, D.
B.; Dion, A.; Maloney, M. T.; Post, R.; Li, R. Kinetic and Data-Driven
Reaction Analysis for Pharmaceutical Process Development. Ind. Eng.
Chem. Res. 2020, 59 (6), 2409−2421.
(24) Nielsen, C. D.-T.; Burés, J. Visual Kinetic Analysis. Chem. Sci.
2019, 10 (2), 348−353.
(25) Marco-Contelles, J.; Pérez-Mayoral, E.; Samadi, A.; Carreiras,
̀
Mocan, A.; Muller, R.; Nicoletti, F.; Perry, G.; Pittala, V.; Rastrelli, L.;
̈
Ristow, M.; Russo, G. L.; Silva, A. S.; Schuster, D.; Sheridan, H.;
Skalicka-Woźniak, K.; Skaltsounis, L.; Sobarzo-Sánchez, E.; Bredt, D.
S.; Stuppner, H.; Sureda, A.; Tzvetkov, N. T.; Vacca, R. A.; Aggarwal,
B. B.; Battino, M.; Giampieri, F.; Wink, M.; Wolfender, J.-L.; Xiao, J.;
Yeung, A. W. K.; Lizard, G.; Popp, M. A.; Heinrich, M.; Berindan-
Neagoe, I.; Stadler, M.; Daglia, M.; Verpoorte, R.; Supuran, C. T.
Taskforce, the I. N. P. S. Natural Products in Drug Discovery:
Advances and Opportunities. Nat. Rev. Drug Discovery 2021, 20, 200−
216.
(6) Karageorgis, G.; Foley, D. J.; Laraia, L.; Waldmann, H. Principle
and Design of Pseudo-Natural Products. Nat. Chem. 2020, 12 (3),
227−235.
(7) Newman, D. J.; Cragg, G. M. Natural Products as Sources of
New Drugs over the Nearly Four Decades from 01/1981 to 09/2019.
J. Nat. Prod. 2020, 83 (3), 770−803.
(8) Over, B.; Wetzel, S.; Grutter, C.; Nakai, Y.; Renner, S.; Rauh, D.;
̈
Waldmann, H. Natural-Product-Derived Fragments for Fragment-
Based Ligand Discovery. Nat. Chem. 2013, 5 (1), 21−28.
(9) Rodrigues, T.; Reker, D.; Schneider, P.; Schneider, G. Counting
on Natural Products for Drug Design. Nat. Chem. 2016, 8 (6), 531−
541.
(10) Arndt, R. R.; Jordaan, A.; Joynt, V. P. Alkaloids of Macrorungia
Longistrobus C.B. Cl. J. Chem. Soc. 1964, 5969−5975.
M.; do, C.; Soriano, E. Recent Advances in the Friedlan
Chem. Rev. 2009, 109 (6), 2652−2671.
̈
der Reaction.
(11) Rodrigues, T.; Reker, D.; Kunze, J.; Schneider, P.; Schneider, G.
Revealing the Macromolecular Targets of Fragment-Like Natural
Products. Angew. Chem., Int. Ed. 2015, 54, 10516−10520.
(12) Reker, D.; Rodrigues, T.; Schneider, P.; Schneider, G.
Identifying the Macromolecular Targets of de Novo-Designed
Chemical Entities through Self-Organizing Map Consensus. Proc.
Natl. Acad. Sci. U. S. A. 2014, 111 (11), 4067−4072.
(13) Hicken, E. J.; Marmsater, F. P.; Munson, M. C.; Schlachter, S.
T.; Robinson, J. E.; Allen, S.; Burgess, L. E.; DeLisle, R. K.; Rizzi, J. P.;
Topalov, G. T.; Zhao, Q.; Hicks, J. M.; Kallan, N. C.; Tarlton, E.;
Allen, A.; Callejo, M.; Cox, A.; Rana, S.; Klopfenstein, N.; Woessner,
R.; Lyssikatos, J. P. Discovery of a Novel Class of Imidazo[1,2- a
]Pyridines with Potent PDGFR Activity and Oral Bioavailability. ACS
Med. Chem. Lett. 2014, 5 (1), 78−83.
(26) Lan, X.-B.; Ye, Z.; Huang, M.; Liu, J.; Liu, Y.; Ke, Z.
Nonbifunctional Outer-Sphere Strategy Achieved Highly Active α-
Alkylation of Ketones with Alcohols by N-Heterocyclic Carbene
Manganese (NHC-Mn). Org. Lett. 2019, 21 (19), 8065−8070.
(27) Kumar, P.; Garg, V.; Kumar, M.; Verma, A. K. Rh(Iii)-
Catalyzed Alkynylation: Synthesis of Functionalized Quinolines from
Aminohydrazones. Chem. Commun. 2019, 55 (81), 12168−12171.
(28) Martínez, R.; Ramón, D. J.; Yus, M. Transition-Metal-Free
̈
Indirect Friedlander Synthesis of Quinolines from Alcohols. J. Org.
Chem. 2008, 73 (24), 9778−9780.
(29) Harry, N. A.; Ujwaldev, S. M.; Anilkumar, G. Recent Advances
and Prospects in the Metal-Free Synthesis of Quinolines. Org. Biomol.
Chem. 2020, 18 (48), 9775−9790.
̈
(30) Ghobadi, N.; Nazari, N.; Gholamzadeh, P. The Friedlander
(14) Effendi, N.; Mishiro, K.; Takarada, T.; Yamada, D.; Nishii, R.;
Shiba, K.; Kinuya, S.; Odani, A.; Ogawa, K. Design, Synthesis, and
Biological Evaluation of Radioiodinated Benzo[d]Imidazole-Quino-
line Derivatives for Platelet-Derived Growth Factor Receptor β
(PDGFRβ) Imaging. Bioorg. Med. Chem. 2019, 27 (2), 383−393.
(15) Cushing, T. D.; Baichwal, V.; Berry, K.; Billedeau, R.;
Bordunov, V.; Broka, C.; Cardozo, M.; Cheng, P.; Clark, D.;
Dalrymple, S.; DeGraffenreid, M.; Gill, A.; Hao, X.; Hawley, R. C.;
He, X.; Jaen, J. C.; Labadie, S. S.; Labelle, M.; Lehel, C.; Lu, P.-P.;
McIntosh, J.; Miao, S.; Parast, C.; Shin, Y.; Sjogren, E. B.; Smith, M.-
L.; Talamas, F. X.; Tonn, G.; Walker, K. M.; Walker, N. P. C.;
Wesche, H.; Whitehead, C.; Wright, M.; Browner, M. F. A Novel
Series of IKKβ Inhibitors Part I: Initial SAR Studies of a HTS Hit.
Bioorg. Med. Chem. Lett. 2011, 21 (1), 417−422.
(16) Cushing, T. D.; Baichwal, V.; Berry, K.; Billedeau, R.;
Bordunov, V.; Broka, C.; Browner, M. F.; Cardozo, M.; Cheng, P.;
Clark, D.; Dalrymple, S.; DeGraffenreid, M.; Gill, A.; Hao, X.;
Hawley, R. C.; He, X.; Labadie, S. S.; Labelle, M.; Lehel, C.; Lu, P.-P.;
McIntosh, J.; Miao, S.; Parast, C.; Shin, Y.; Sjogren, E. B.; Smith, M.-
L.; Talamas, F. X.; Tonn, G.; Walker, K. M.; Walker, N. P. C.;
Wesche, H.; Whitehead, C.; Wright, M.; Jaen, J. C. A Novel Series of
IKKβ Inhibitors Part II: Description of a Potent and Pharmacolog-
ically Active Series of Analogs. Bioorg. Med. Chem. Lett. 2011, 21 (1),
423−426.
Reaction: A Powerful Strategy for the Synthesis of Heterocycles; Scriven,
E. F. V, Ramsden, C. A. B. T.-A. H. C., Eds.; Academic Press, 2020;
Vol. 132, Chapter 2, pp 85−134.
(31) Chen, Y.; Stork, C.; Hirte, S.; Kirchmair, J. NP-Scout: Machine
Learning Approach for the Quantification and Visualization of the
Natural Product-Likeness of Small Molecules. Biomolecules 2019, 9
(2), 43.
(32) Reker, D.; Schneider, P.; Schneider, G. Multi-Objective Active
Machine Learning Rapidly Improves Structure-Activity Models and
Reveals New Protein-Protein Interaction Inhibitors. Chem. Sci. 2016,
7 (6), 3919−3927.
(33) Herrera, V. M.; Khoshgoftaar, T. M.; Villanustre, F.; Furht, B.
Random Forest Implementation and Optimization for Big Data
Analytics on LexisNexis’s High Performance Computing Cluster
Platform. J. Big Data 2019, 6 (1), 68.
(34) Ao, Y.; Li, H.; Zhu, L.; Ali, S.; Yang, Z. The Linear Random
Forest Algorithm and Its Advantages in Machine Learning Assisted
Logging Regression Modeling. J. Pet. Sci. Eng. 2019, 174, 776−789.
(35) Rodrigues, T. The Good, the Bad, and the Ugly in Chemical
and Biological Data for Machine Learning. Drug Discovery Today:
Technol. 2019, 32−33, 3−8.
(36) Kaneko, H. Estimation of Predictive Performance for Test Data
in Applicability Domains Using Y-Randomization. J. Chemom. 2019,
33 (9), 1−12.
F
https://doi.org/10.1021/acs.joc.1c01038
J. Org. Chem. XXXX, XXX, XXX−XXX

The Journal of Organic Chemistry
pubs.acs.org/joc
Note
(37) Rucker, C.; Rucker, G.; Meringer, M. Y-Randomization and Its
̈ ̈
Variants in QSPR/QSAR. J. Chem. Inf. Model. 2007, 47 (6), 2345−
2357.
(38) Chuang, K. V.; Keiser, M. J. Adversarial Controls for Scientific
Machine Learning. ACS Chem. Biol. 2018, 13 (10), 2819−2821.
(39) Lundberg, S.; Lee, S.-I. A Unified Approach to Interpreting
Model Predictions. Adv. Neural Inf. Process. Syst. 2017, No. Section 2,
4766−4775.
(40) Pollice, R.; dos Passos Gomes, G.; Aldeghi, M.; Hickman, R. J.;
Krenn, M.; Lavigne, C.; Lindner-D’Addario, M.; Nigam, A.; Ser, C.
T.; Yao, Z.; Aspuru-Guzik, A. Data-Driven Strategies for Accelerated
Materials Design. Acc. Chem. Res. 2021, 54 (4), 849−860.
(41) Saal, J. E.; Oliynyk, A. O.; Meredig, B. Machine Learning in
Materials Discovery: Confirmed Predictions and Their Underlying
Approaches. Annu. Rev. Mater. Res. 2020, 50 (1), 49−69.
(42) Haghighatlari, M.; Li, J.; Heidar-Zadeh, F.; Liu, Y.; Guan, X.;
Head-Gordon, T. Learning to Make Chemical Predictions: The
Interplay of Feature Representation, Data, and Machine Learning
Methods. Chem. 2020, 6 (7), 1527−1542.
(43) Jiménez-Luna, J.; Grisoni, F.; Schneider, G. Drug Discovery
with Explainable Artificial Intelligence. Nat. Mach. Intell. 2020, 2 (10),
573−584.
(44) Lundberg, S. M.; Nair, B.; Vavilala, M. S.; Horibe, M.; Eisses,
M. J.; Adams, T.; Liston, D. E.; Low, D. K.-W.; Newman, S.-F.; Kim,
J.; Lee, S.-I. Explainable Machine-Learning Predictions for the
Prevention of Hypoxaemia during Surgery. Nat. Biomed. Eng. 2018,
2 (10), 749−760.
(45) McInnes, L.; Healy, J.; Saul, N.; Großberger, L. UMAP:
Uniform Manifold Approximation and Projection. J. Open Source
Softw. 2018, 3 (29), 861.
(46) Scale-up Systems, https://www.scale-up.com/dynochem.
(47) Tillekeratne, V.; Al-Hamashi, A.; Dlamini, S.; Alqahtani, A. S.;
Karaj, E. Imidazole-Based Anticancer Agents and Derivatives Thereof,
And Methods of Making and Using Same. WO2019036607, US
16638587, 2018.
(48) Zhu, H.; Zhang, M.; Xia, L.; Wang, B. Process for Synthesis of
Quinoline Containing Fluorescent Core Compound. CN 107417609,
2017.
G
https://doi.org/10.1021/acs.joc.1c01038
J. Org. Chem. XXXX, XXX, XXX−XXX