Planning backwards
A medicinal chemist sketches a molecule that might block an enzyme. Before anyone can test it, someone has to make it, from compounds that can be bought, using reactions that actually work, in as few steps as possible, without destroying the parts of the molecule already built. Organic synthesis is that craft, and since the 1960s it has had an explicit logic.
The logic is to plan backwards. Starting from the target, a chemist asks: which bond could I have made last, and from what? Then the same question of each precursor, and so on, until every branch ends in something purchasable. E. J. Corey at Harvard set out this retrosynthetic analysis as a formal method in the 1960s, used it to make dozens of natural products, including prostaglandins and ginkgolide B, and received the 1990 Nobel Prize in Chemistry.
This lesson teaches the method with the reactions of Reaction mechanisms and Carbonyl and aromatic chemistry, then looks at what makes a route good: selectivity, protecting groups, step count, yield and waste. It is about strategy, as chemists discuss it on a whiteboard. How each step is carried out safely at the bench or in a plant is a separate, professional skill, learnt in a supervised laboratory.
Analogy, not a formal construction
Planning a route through a maze is easier from the exit, because the exit is one place and the entrances are many. Retrosynthesis works the same way: the target is fixed, while the possible starting materials number in the millions. The analogy has one limit: in a maze every corridor works in both directions, but a disconnection is only useful if the corresponding forward reaction really works.
Disconnections, synthons and equivalents
Definition— retrosynthetic vocabulary
- A disconnection is the imaginary breaking of a bond in the target, the reverse of a bond-forming reaction. It is written with a double-shafted retrosynthetic arrow, target ⇒ precursors, and the broken bond is marked with a wavy line.
- A synthon is an idealised fragment produced by a disconnection, usually an ion such as CH₃CO⁺ or a carbanion. Synthons are bookkeeping devices and need not exist.
- A synthetic equivalent is a real reagent that behaves like a synthon: ethanoic anhydride or ethanoyl chloride for the acylium synthon CH₃CO⁺, a Grignard reagent for a carbanion.
- A functional-group interconversion (FGI) changes one functional group into another without changing the skeleton (an alcohol into a ketone, a nitro group into an amine), and is written with the same arrow.
Good disconnections are not arbitrary. The safest place to cut is next to a heteroatom, especially the C–N bond of an amide and the C–O bond of an ester or ether, because the forward reactions (acyl substitution, alkylation) are reliable. Carbon–carbon bonds are best cut next to a functional group that can supply the necessary polarity: a bond between the α- and β-carbons of a β-hydroxy ketone points to an aldol reaction, a bond next to a tertiary alcohol points to a Grignard addition to a ketone, and a bond between two aromatic rings points to a cross-coupling.
Example— A tertiary alcohol: two disconnections, one choice
Take 2-phenylbutan-2-ol, C₆H₅C(OH)(CH₃)CH₂CH₃. The carbon carrying the OH has three carbon groups, and any of the three bonds to it can be disconnected to a ketone plus a carbanion synthon, whose equivalent is a Grignard reagent. Cutting the ethyl bond gives 1-phenylethan-1-one plus an ethyl carbanion; cutting the methyl bond gives 1-phenylpropan-1-one plus a methyl carbanion; cutting the phenyl bond gives butan-2-one plus a phenyl carbanion. All three are sound. A chemist picks by asking which starting materials are cheap, which Grignard reagent is easiest to handle, and whether the rest of the target has groups that the Grignard reagent would attack.
Example— A β-hydroxy ketone: recognising a pattern
4-Hydroxy-4-methylpentan-2-one, (CH₃)₂C(OH)CH₂COCH₃, has an OH on the carbon β to a ketone: the signature of an aldol product. Disconnecting the bond between the α-carbon and the β-carbon gives an enolate synthon of propanone and a second propanone as the electrophile. So the target is a "dimer" of propanone, and the forward reaction is an aldol addition of propanone to itself. Recognising such patterns, called retrons in Corey's language, is most of the skill.
Selectivity and protecting groups
A real target has several functional groups, and every reagent must touch only the one intended. Chemists distinguish three kinds of selectivity:
- Chemoselectivity: which functional group reacts. Sodium borohydride reduces a ketone but leaves an ester alone; lithium aluminium hydride reduces both.
- Regioselectivity: which position of a group reacts. The directing effects of Carbonyl and aromatic chemistry, the choice between kinetic and thermodynamic enolates, and Zaitsev or Hofmann elimination are all regioselectivity.
- Stereoselectivity: which stereoisomer forms. Most medicines are single enantiomers (Stereochemistry), so modern synthesis relies on asymmetric catalysis, recognised by the 2001 Nobel Prize to William Knowles, Ryoji Noyori and Barry Sharpless and the 2021 prize below, and on enzymes.
When no reagent is selective enough, the chemist protects the group that must not react: converts it temporarily into an unreactive form, does the reaction elsewhere, then removes the protection.
Example— Protecting a ketone
Suppose a molecule contains both a ketone and an ester, and the plan is to add a Grignard reagent to the ester only. The Grignard reagent would attack the ketone first, since ketones are more electrophilic than esters. The answer is to convert the ketone into a cyclic acetal with ethane-1,2-diol, under acid catalysis: an acetal is an ether on both sides and ignores strongly basic nucleophiles. After the Grignard step, aqueous acid hydrolyses the acetal back to the ketone. Peptide synthesis relies on the same idea on a grand scale: each amino acid's amine is protected while its acid end is coupled, then unprotected for the next coupling, the basis of Bruce Merrifield's solid-phase method (Nobel Prize 1984), which machines now run automatically.
Counting steps, yield and waste
Every step loses material. If step has yield , a linear synthesis, in which each step feeds the next, delivers an overall yield . A convergent synthesis makes two or more fragments separately and joins them late, so the number of steps the material passes through, the longest linear sequence, is much shorter than the total number of steps.
import numpy as np
print("overall yield (%) of a linear synthesis")
print("steps 99%/step 95%/step 90%/step 80%/step")
for n in (5, 10, 20, 30, 40):
print(f"{n:5d} " + "".join(f"{100 * y**n:10.1f} " for y in (0.99, 0.95, 0.90, 0.80)))
# Twelve steps at 90% each, arranged two ways.
y = 0.90
linear = y**12 # A -> ... -> target in one chain of 12 steps
convergent = y**(5 + 2) # two 5-step branches, then coupling + one last step: 7 in the longest chain
print(f"\n12 steps, linear: longest chain 12 steps, overall yield {100 * linear:.1f}%,"
f" starting material per mol of product {1 / linear:.2f} mol")
print(f"12 steps, convergent: longest chain 7 steps, overall yield {100 * convergent:.1f}%,"
f" each starting material per mol of product {1 / convergent:.2f} mol")
# How good must each step be for a 25-step synthesis to deliver 10%?
need = 0.10 ** (1 / 25)
print(f"\na 25-step linear route needs an average of {100 * need:.1f}% per step to give 10% overall")The table is the most important arithmetic in synthesis. Twenty steps at a very respectable 90% leave 12% of the material; at 80% per step, a little over 1%. Arranging the same twelve steps convergently raises the overall yield from 28% to 48%, and each starting material is needed in 2.1 rather than 3.5 times the final amount. A 25-step synthesis that is to deliver 10% needs an average of 91% per step.
Yield measures how much of the intended product survives; it says nothing about everything else that went into the flask. Barry Trost's atom economy (1991) asks what fraction of the mass of all the reactants ends up in the product, assuming every step goes perfectly. The classic comparison is ibuprofen. The original route of the Boots Company, from the 1960s, took six steps; the BHC route, introduced in 1992, takes three catalytic ones. Both are read here from balanced equations.
import re
MASSES = {"H": 1.008, "C": 12.011, "N": 14.007, "O": 15.999, "Na": 22.990, "Cl": 35.45}
def molar_mass(formula):
"""Molar mass in g/mol of a formula such as 'C13H18O2' (no brackets)."""
return sum(MASSES[el] * (int(n) if n else 1) for el, n in re.findall(r"([A-Z][a-z]?)(\d*)", formula))
def atom_economy(reactants, product):
return molar_mass(product) / sum(molar_mass(f) for f in reactants)
ibuprofen = "C13H18O2"
# Every reagent whose atoms appear in the balanced equations of each route (the usual bookkeeping).
routes = {
"Boots route, 6 steps": ["C10H14", "C4H6O3", "C4H7ClO2", "C2H5ONa", "H3O", "NH3O", "H4O2"],
"BHC route, 3 steps": ["C10H14", "C4H6O3", "H2", "CO"],
}
print(f"ibuprofen {ibuprofen}: {molar_mass(ibuprofen):.2f} g/mol\n")
for name, reactants in routes.items():
total = sum(molar_mass(f) for f in reactants)
ae = atom_economy(reactants, ibuprofen)
print(f"{name}: reactants {total:6.1f} g/mol, atom economy {100 * ae:5.1f}%, "
f"by-products {total - molar_mass(ibuprofen):5.1f} g per mol of ibuprofen")
# In the BHC route the only by-product is ethanoic acid, C2H4O2, which can be recovered and used.
used = (molar_mass(ibuprofen) + molar_mass("C2H4O2")) / sum(molar_mass(f) for f in routes["BHC route, 3 steps"])
print(f"BHC route counting the recovered ethanoic acid as product: {100 * used:.1f}%")Even with perfect yields, the six-step route turns only 40% of the mass of its reagents into ibuprofen, and makes about 309 g of by-products per mole (206 g) of drug. The three-step route reaches 77%, and its one by-product, ethanoic acid, is recovered, so nearly every atom bought ends up in something useful. Fewer steps, catalysts instead of stoichiometric reagents, and reactions that add whole molecules (hydrogen, carbon monoxide) rather than swap groups are what make the difference. The large-scale steps use hazardous gases such as carbon monoxide, and run only in closed, monitored industrial plant; the strategy is what transfers to other problems.
Retrosynthesis as search
Retrosynthesis is a search problem. Each molecule is a node with several possible disconnections (an "or": any one will do), and each disconnection leads to precursors that must all be made (an "and"). The whole space is an AND/OR tree, and a synthetic plan is a subtree whose leaves are all purchasable. The two ibuprofen routes can be found by searching such a tree. Below, a hand-written "database" holds eight known reactions, each read backwards as a rule. A recursive function expands the target, tries every rule for every intermediate, and combines the routes to all the precursors a rule needs.
# A tiny reaction "database" for ibuprofen, 2-[4-(2-methylpropyl)phenyl]propanoic acid.
# Each entry reads a known forward reaction backwards: product <= (reaction, precursors).
# Short labels: ketone = 1-[4-(2-methylpropyl)phenyl]ethan-1-one, alcohol = the corresponding
# ethan-1-ol, and aldehyde / oxime / nitrile / epoxy ester are the intermediates of the Boots route.
RULES = {
"ibuprofen": [("Pd-catalysed carbonylation", ["alcohol", "carbon monoxide"]),
("nitrile hydrolysis", ["nitrile", "water"])],
"alcohol": [("hydrogenation", ["ketone", "hydrogen"])],
"nitrile": [("oxime dehydration", ["oxime"])],
"oxime": [("oxime formation", ["aldehyde", "hydroxylamine"])],
"aldehyde": [("hydrolysis and decarboxylation", ["epoxy ester", "water"])],
"epoxy ester": [("Darzens condensation", ["ketone", "ethyl chloroacetate"])],
"ketone": [("Friedel-Crafts acylation", ["(2-methylpropyl)benzene", "ethanoic anhydride"])],
}
BUYABLE = {"(2-methylpropyl)benzene", "ethanoic anhydride", "ethyl chloroacetate",
"hydroxylamine", "hydrogen", "carbon monoxide", "water"}
def routes(molecule):
"""Every way to make `molecule` from buyable compounds: a list of routes, each a list of steps
in forward order. A molecule with several rules is an OR node; a rule with several precursors
is an AND node, because every precursor must be made."""
if molecule in BUYABLE:
return [[]]
found = []
for reaction, precursors in RULES.get(molecule, []):
partial = [[]]
for p in precursors:
partial = [a + b for a in partial for b in routes(p)]
step = f"{' + '.join(precursors)} -> {molecule} [{reaction}]"
found += [steps + [step] for steps in partial]
return found
all_routes = sorted(routes("ibuprofen"), key=len)
for k, route in enumerate(all_routes, 1):
print(f"route {k}: {len(route)} steps")
for i, step in enumerate(route, 1):
print(f" {i}. {step}")
n_rules = sum(len(r) for r in RULES.values())
print(f"\n{len(all_routes)} routes from {n_rules} rules; shortest: {len(all_routes[0])} steps")The search rediscovers both routes and ranks the BHC route, three steps, first. Both begin with the same Friedel–Crafts acylation; they differ in how they add the last carbon. The older route builds it through a chain of carbonyl chemistry (an epoxy ester, an aldehyde, an oxime and a nitrile), while the newer one inserts carbon monoxide directly with a palladium catalyst.
Real planning programs work the same way at a vastly larger scale. Corey and Todd Wipke's LHASA program (1969) encoded hand-written rules; today's planners extract hundreds of thousands of rules automatically from published and patented reactions, use neural networks to rank which rules to try at each node, and search with methods such as Monte Carlo tree search (Segler, Preuss and Waller, 2018). Open-source examples include AiZynthFinder and ASKCOS. How well they work, and where they fail, is a subject of AI for chemistry; joining a planner to robotic hardware is a subject of Self-driving laboratories.
Landmark syntheses, as strategy
Synthesis became an intellectual discipline through a series of targets that seemed impossible until someone found the strategy.
- Tropinone, 1917. Robert Robinson in Liverpool made this alkaloid skeleton in a single pot from three simple components, by imagining how a plant might assemble it: an early triumph of biomimetic thinking, in which the strategy follows a guess about biosynthesis.
- Robert Burns Woodward, 1940s–1970s. Woodward, at Harvard, showed that molecules of great complexity could be planned with the logic of mechanism and conformation: the antihypertensive reserpine (1956), with five adjacent stereocentres on one ring set by exploiting the shape of rigid intermediates; chlorophyll (1960); and, with Albert Eschenmoser at ETH Zurich, vitamin B12 (1972), a project of about a hundred chemists over more than a decade. Puzzles met during the B12 work led to the Woodward–Hoffmann rules of Molecular orbitals. He received the 1965 Nobel Prize "for his outstanding achievements in the art of organic synthesis".
- Corey, 1960s–1990s. Corey made retrosynthesis explicit and teachable, introduced many reagents built for selectivity, and used them in syntheses such as the prostaglandins, hormones that became medicines.
- Paclitaxel (Taxol), 1994. Two groups, led by Robert Holton and K. C. Nicolaou, completed this anticancer natural product almost simultaneously, by different strategies. The medicine itself is now made from plant cell cultures and by semi-synthesis from a related natural product, a reminder that the best supply route need not be a total synthesis.
- Eribulin, 2010. This anticancer medicine, developed by the Japanese company Eisai from Yoshito Kishi's synthesis of the marine natural product halichondrin B, is manufactured by total synthesis in more than sixty steps, made practical by a highly convergent plan.
- Artemisinin, 2013. A semi-synthetic route to the antimalarial joins synthetic biology and chemistry: engineered yeast makes a precursor, artemisinic acid, and a short chemical sequence finishes the job (Paddon and colleagues, 2013).
Methods that changed the plan
A new reaction changes synthesis when it creates a new kind of disconnection. Three recent Nobel Prizes recognised exactly that.
Palladium-catalysed cross-coupling (2010: Richard Heck, Ei-ichi Negishi, Akira Suzuki). Joining two aromatic rings, or an aromatic ring to an alkene, used to be awkward. Cross-coupling makes the C–C bond between two pieces that carry a halogen and a metal or boron, with a palladium catalyst that runs round a cycle many times.
Akira Suzuki and Norio Miyaura's version, with boronic acids that are stable, mostly non-toxic and tolerant of water, became so popular that it is now the second most used reaction in medicinal chemistry, after amide formation (Brown and Boström, 2016). It created the biaryl disconnection: any bond between two rings can be cut into an aryl halide and an arylboronic acid. The biphenyl core of the blood-pressure medicine losartan, for example, is made this way.
Alkene metathesis (2005: Yves Chauvin, Robert Grubbs, Richard Schrock). Metathesis swaps the partners of two C=C bonds: RCH=CHR + R′CH=CHR′ ⇌ 2 RCH=CHR′. Chauvin proposed the mechanism, through a four-membered ring containing a metal, in 1971; Schrock's molybdenum and tungsten catalysts and Grubbs's ruthenium catalysts made it practical. Its most important strategic gift is ring-closing metathesis, which joins two alkene ends of one chain into a ring and releases ethene, the easiest way to make medium and large rings. It is used to manufacture macrocyclic antiviral medicines and, in its ring-opening form, tough polymers.
Asymmetric organocatalysis (2021: Benjamin List, David MacMillan). In 2000 List showed that the amino acid proline catalyses an asymmetric aldol reaction, and MacMillan that a small chiral amine catalyses an asymmetric Diels–Alder reaction. Both work through the carbonyl chemistry of the previous lesson: the catalyst's amine condenses with a carbonyl compound to form an enamine (a nitrogen version of an enol) or an iminium ion (a more electrophilic version of the carbonyl), inside a chiral environment that favours one enantiomer. Organocatalysts are cheap, stable in air and free of metals, which matters for medicines.
The professional view
Two kinds of chemist plan syntheses, with different goals. In drug discovery, the aim is to make hundreds of analogues quickly, so routes favour reliable reactions (amide formation, cross-coupling) and late-stage diversification. In process chemistry, the aim is to make tonnes of one compound safely, cheaply, reproducibly and with little waste, so a route is judged by its step count, overall yield, the process mass intensity (kilograms of all materials used per kilogram of product, often well over a hundred for a medicine), the hazards of its reagents and intermediates, and how robust each step is to small changes. The two often lead to completely different routes for the same molecule.
The process chemist's toolbox now includes enzymes. For the diabetes medicine sitagliptin, Merck and Codexis replaced a rhodium-catalysed asymmetric hydrogenation with an engineered transaminase enzyme, evolved in the laboratory to accept the drug's ketone precursor, which increased the yield and removed the metal (Savile and colleagues, 2010); directed evolution of enzymes was recognised by the 2018 Nobel Prize to Frances Arnold. Continuous flow chemistry lets hazardous steps run in small volumes, and retrosynthesis software, as of 2026, routinely proposes routes for chemists to judge. The judgement remains human: whether a step will be selective on this particular molecule, scale safely, or meet regulatory standards for impurities. These themes continue in Catalysis and green chemistry and Medicinal chemistry.
Common misconceptions
Misconception
"Retrosynthesis is just the synthesis written backwards." A disconnection is only as good as the forward reaction it stands for. Synthons such as CH₃CO⁺ are fictions that must be replaced by real reagents, and a disconnection that has no reliable forward reaction, or that would be ruined by another group in the molecule, is useless however neat it looks.
Misconception
"If every step gives 99%, length does not matter." Fifty steps at 99% give 61%; at 95%, under 8%. And each step still costs time, solvent, energy and purification, whatever its yield.
Misconception
"The shortest route is always best." A short route through a hazardous intermediate, an expensive catalyst or an unselective step can lose to a longer, safer and more robust one. Routes are compared on yield, cost, safety, waste and robustness together.
Misconception
"Total synthesis is finished: anything can be made." Almost any structure can be made by a large enough effort, but making it efficiently is still hard, and targets continue to drive the invention of new reactions. Many medicines still come from fermentation, plants or semi-synthesis because no synthetic route is competitive.
Exercises
1. Synthesis and retrosynthesis
Warm-upHumanDisconnecting the amide C–N bond of paracetamol gives two fragments. Which of the following is a synthon rather than a real reagent?
2. Synthesis and retrosynthesis
CoreHumanBenzocaine, ethyl 4-aminobenzoate (H₂N–C₆H₄–COOCH₂CH₃), is a local anaesthetic used in throat lozenges. Which first disconnection (the last step of the synthesis) is the most sensible?
3. Synthesis and retrosynthesis
CoreHumanA molecule contains a ketone and an ester. The plan needs a Grignard reagent to react with the ester only. What is the standard solution?
4. Synthesis and retrosynthesis
Warm-upHumanA linear synthesis has 8 steps, each with a yield of 85%. What is the overall yield, as a percentage to the nearest whole number?
5. Synthesis and retrosynthesis
CoreHumanIn a convergent synthesis, fragment A is made from a purchased compound in 6 steps and fragment B in 4 steps. A and B are then coupled (1 step), and the coupled product needs 2 more steps to become the target. What is the longest linear sequence, in steps?
6. Synthesis and retrosynthesis
CoreHumanA target contains two benzene rings joined directly by a C–C bond. Which Nobel-recognised method gives the most direct disconnection of that bond?
7. Synthesis and retrosynthesis
CoreComputationalRepresent a synthetic route as a tree: a purchased material is the string "buy", and a step is a dictionary {"yield": y, "from": [...]} listing the routes to its precursors. Write route_stats(route) returning (total_steps, longest_linear_sequence, worst_path_yield), where the last is the smallest product of step yields along any path from a purchased material to the target (the overall yield based on the material that suffers most losses).
8. Synthesis and retrosynthesis
CoreComputationalWrite molar_mass(formula) for formulas without brackets (such as 'C8H9NO2' or 'C2H5ONa'), using the atomic masses given, and atom_economy(reactants, product), the fraction of the total mass of the reactants that ends up in the product. The tests include paracetamol from 4-aminophenol (C₆H₇NO) and ethanoic anhydride (C₄H₆O₃).
Further reading and resources
- The Logic of Chemical Synthesis: Multistep Synthesis of Complex Carbogenic Molecules (Nobel Lecture) · E. J. Corey · 1991Retrosynthetic analysis explained by the chemist who formalised it.
- Computer-Assisted Design of Complex Organic Syntheses · E. J. Corey, W. Todd Wipke · 1969The first program for planning syntheses.
- The total synthesis of vitamin B12 · R. B. Woodward · 1973
- The Atom Economy — A Search for Synthetic Efficiency · Barry M. Trost · 1991
- Aiming for the Ideal Synthesis · Tanja Gaich, Phil S. Baran · 2010
- Total synthesis of marine natural products without using protecting groups · Phil S. Baran, Thomas J. Maimone, Jeremy M. Richter · 2007
- Palladium-Catalyzed Cross-Coupling Reactions of Organoboron Compounds · Norio Miyaura, Akira Suzuki · 1995
- Proline-Catalyzed Direct Asymmetric Aldol Reactions · Benjamin List, Richard A. Lerner, Carlos F. Barbas III · 2000
- New Strategies for Organic Catalysis: The First Highly Enantioselective Organocatalytic Diels–Alder Reaction · Kateri A. Ahrendt, Christopher J. Borths, David W. C. MacMillan · 2000
- The Nobel Prize in Chemistry 2010 (palladium-catalysed cross couplings)
- The Nobel Prize in Chemistry 2005 (metathesis)
- The Nobel Prize in Chemistry 2021 (asymmetric organocatalysis)
- Biocatalytic Asymmetric Synthesis of Chiral Amines from Ketones Applied to Sitagliptin Manufacture · Christopher K. Savile and colleagues · 2010
- High-level semi-synthetic production of the potent antimalarial artemisinin · C. J. Paddon and colleagues · 2013
- Planning chemical syntheses with deep neural networks and symbolic AI · Marwin H. S. Segler, Mike Preuss, Mark P. Waller · 2018
- AiZynthFinderAn open-source retrosynthesis planner.