conditions that work for these complex substrates.
Additionally, HTE has recently been leveraged
to benchmark emerging methods against dif-
ferent catalytic procedures through the cre-
ation of arrays of complex, drug-like substrates
known as informer libraries ( 59 )orthrough
addition of diverse molecular fragments that
can disrupt catalysis ( 60 , 61 ). The use of these
diagnostic methods allows exploration of the
relationship between reaction types and diverse
complex substrate structures, thus enabling
synthetic practitioners to make better decisions
about which synthetic methods to prioritize in
their problem-solving. Additionally, miniatur-
ization of HTE to nanomole scale—for example,
by automated nanomole-scale batch ( 62 ) and
flow ( 63 ) approaches—now enables the execution
of more than 1500 simultaneous experiments
at microgram scale in 1 day for rapid identifi-
cation of suitable reaction conditions to explore
chemical space and accelerate drug discovery.
This capability is augmented by advances in
rapid high-throughput analytics, such as MISER
(multiple injections in a single experimental run)
and MALDI (matrix-assisted laser desorption/
ionization) mass spectrometry techniques ( 64 ),
which have enabled the analysis of as many as
1536 reactions in very short time frames. Finally,
nanomole HTE can also expedite the preparation
of diverse, complex arraysof molecules and, when
coupled directly with biological testing, can rad-
ically alter how drug discovery is performed ( 65 ).
Computational methods
The use of computer-assisted methods to guide
synthetic chemistry is emerging as an important
component in the practice of drug discovery.
Advances in computational chemistry and ma-
chine learning in the past decade are delivering
real impact in areas such as new catalyst design
( 66 ) or showing considerable promise in others
such as reaction prediction ( 67 ). The application
of deep learning methods has the potential to
uncover new chemical reactions, expanding the
access to new pharmaceutical chemical matter.
Grandaet al.( 68 ) have reported promising results
toward this end. By combining automated syn-
thesis with machine learning, they reported the
discovery of four chemical transformations with
differentiated novelty.
Recently, computer-guided design has been
successfully applied to the preparation of cat-
alysts that provide asymmetric control of a cy-
cloisomerization reaction ( 69 ). Computational
methods were used to evaluate the catalytic path-
way of a previously unknown reaction, leading
to the hypothesis that the electronics of the cat-
alyst ligand influence both the rate and stereo-
selectivity of the transformation. Application of
quantum methods such as density functional
theory (DFT) provided optimal ligand designs
with markedly enhanced rate and selectivity over
the original ligand. A second example where the
use of computational methods aided in the de-
sign of a superior catalyst is reported in the syn-
thesis of a pronucleotide (ProTide, Fig. 6) ( 70 ).
Achieving selective phosphoramidation of a nu-
cleoside at the 5′hydroxyl over the 3′hydroxyl with
stereocontrol at the phosphorus center is highly
challenging. A combination of mechanistic studies
using a variety of chiral catalysts and DFT calcula-
tions of a proposed transition state further informed
by experimental observations led to the rational
design of a dimeric phosphoramidation catalyst
with an improved rate and excellent stereoselectivity.
Despite these successes, the process for ratio-
nal computational design of a catalyst is arduous,
requiring the modeling of multiple mechanistic
pathways and refinement of numerous molecules
and transition states. A program for automating
much of this process has been reported ( 71 ), and
theadvancementofsuchmethodsaswellasthe
continual increase in processing power will drive
further use of these tools in the future.
The application of machine learning to syn-
thetic problems has also generated considera-
ble interest and excitement. One area of active
research is the use of algorithms for synthetic
route planning to a target molecule ( 72 , 73 ).
Segleret al. combined Monte Carlo tree search
and three neural networks to identify potential
synthetic routes ( 74 ).Thesuccessoftheapproach
was qualitatively evaluated through a double-
blind A/B test, where 45 chemistry students
showed no preference between machine-suggested
synthetic routes versus literature routes for repre-
sentative target molecules. Machine learning has
additionally been applied to forward reaction
prediction ( 75 ). Neural networks were used to
predict the major product of a reaction using an
algorithm that assigns a probability and rank to
potential products. Additionally, machine learn-
ing was used to successfully predict the perform-ance of a single reaction, a Buchwald-Hartwig
amination, against multiple variables: reactants,
catalysts, bases, and additives ( 76 ). Application of
machine learning holds considerable promise for
synthetic optimization of targets far exceeding
those described herein, toward predicting routes,
main products, side products, and optimal con-
ditions, among others. The continued advance-
ment of these methods leverages the wealth of
public information in the scientific and patent
literature as well as within pharmaceutical insti-
tutions. The quality, breadth, depth, and density
of the data within the domain of the predictions is
critical for driving toward high-accuracy models.
Inclusion of examples of both successful and un-
successful transformationsisalsohighlyimportant.
HTE is a highly attractive, complementary tech-
nology for augmenting existing datasets by
generating model-suitable data, maximizing in-
formation content through careful design of exper-
iments and capacity to deliver large volumes of
data in a rapid and cost-effective manner.Future directions
As we have discussed, breakthroughs in syn-
thetic chemistry have proven to be the inspi-
ration for the discovery and development of new
medicines of important therapeutic value. Despite
themanyadvancesdescribedabove,thepaceand
breadth of molecule design is still constrained
because of unsolved problems in synthetic chem-
istry. Many opportunities still remain to advance
the field, such that synthetic chemistry will never
constrain compound design or program pace,
and should actually inspire access to uncharted
chemical space in the pharmaceutical industry.Camposet al.,Science 363 , eaat0805 (2019) 18 January 2019 6of8
Fig. 7. Molecular editing to enable drug discovery.RESEARCH | REVIEW
on January 19, 2019^http://science.sciencemag.org/Downloaded from