Please cite this article in press as: Gajewska et al., Algorithmic Discovery of Tactical Combinations for Advanced Organic Syntheses, Chem
Article
Algorithmic Discovery of Tactical
Combinations for Advanced
Organic Syntheses
Ewa P. Gajewska,1,5 Sara Szymkuc, Piotr Dittwald,1 Michał Startek,2 Oskar Popik,1 Jacek Mlynarski,1,
and Bartosz A. Grzybowski1,3,4,6,
*
SUMMARY
The Bigger Picture
Although computers have
Whereas most organic molecules can be synthesized from progressively simpler
substrates, syntheses of complex organic targets often involve counterintuitive
sequence of steps that first complexify the structure but, by doing so, open up
possibilities for pronounced structural simplification in subsequent, down-
stream steps. Such complexifying/simplifying reaction sequences, called
tactical combinations (TCs), can be quite powerful and elegant but also inher-
ently hard to spot—indeed, only some 500 TCs have so far been cataloged,
and even fewer are routinely used in synthetic practice. This paper describes
computer-driven discovery of large numbers of viable TCs (over 46,000 combi-
nations of reaction classes and ꢀ4.85 million combinations of reaction variants),
the vast majority of which have no prior literature precedent. Examples—
including a concise wet lab synthesis of a small natural product—are provided
to illustrate how the use of these newly discovered TCs can streamline the
design of syntheses leading to important drugs and/or natural products.
recently made remarkable
progress in autonomous synthetic
planning, their ability to strategize
over multiple steps, essential for
the synthesis of complex targets,
is still limited. One form of such
‘‘strategizing’’ is the so-called
tactical combinations (TCs)—
which are sequences of steps that
first complexify the target but, by
doing so, enable elegant and
simplifying downstream
disconnections. TCs are often
counterintuitive even to human
experts and, over several
INTRODUCTION
decades, only about 500 of them
have been cataloged. Here, we
show that computers can
When planning syntheses of complex organic molecules, it is often not sufficient to
gradually simplify the structure—instead, it may be beneficial to go through an inter-
mediate that does not, per se, produce any immediate gain (or even intermittently
increases structural complexity) but sets the scene for a ‘‘downstream’’ disconnec-
tion offering a significant structural simplification (Figures 1 and 2A). A few decades
ago, Corey and Cheng1 christened such sequences as ‘‘tactical combinations’’
(TCs)—since then, some of them have become a part of mainstream retrosynthetic
thinking2–6 (Figures 1A and 1B), some are less obvious and require a trained eye
to spot7,8 (Figures 1C–1E), and yet some others, used as key steps in syntheses of
complex natural products, are truly remarkable, making one wonder how the authors
were inspired to identify such an elegant combination9,10 (Figures 1F and 1G).
Indeed, the notion that TCs are ‘‘inspired’’ rather than ‘‘discovered’’ is reinforced
by the fact that the largest collection cataloged in Ott11 provides only ca. 500 com-
binations of suitable reaction types (see Section S3). On the other hand, TCs are
composed of known reaction types, and one might reasonably hypothesize that
more than just 500 are to be found among the myriad of possible two-step reaction
combinations. Here, we validate this hypothesis via big-data analyses that allow us
to enumerate and discover TCs in a systematic manner. Inspecting close to a billion
two-step putative reaction sequences, we identify and rank over 46,000 combina-
tions of reaction classes (and ca. 4.85 million combinations of reaction variants)
that meet the definition of a TC. Remarkably, the vast majority of these TCs
systematically discover much
larger numbers (millions) of
previously unreported yet valid
TCs that unlock new and elegant
synthetic approaches. These
results and accompanying
experimental demonstrations
indicate that computers can now
assist chemists not only by
processing and adapting existing
synthetic approaches (e.g., via
artificial intelligence tools
learning from prior art) but also by
uncovering new ones.
Chem 6, 1–14, January 9, 2020 ª 2019 The Authors. Published by Elsevier Inc.
1