CHEMISTS have a unique power to manipulate matter. Imagine any arrangement of atoms you like and a chemist will have a good shot at stitching them together. Over the decades, their round-bottomed flasks have helped bring all sorts of new compounds into being, from dazzling pigments to miracle pills and wonder materials. But they don’t come easy, not least because chemists must do it all backwards.
The tried-and-tested method for planning how to create a sophisticated molecule starts where you would like to end up. You draw out the web of connected atoms you want to make, then pick it apart, working backwards to plot out a series of reactions that, if performed in the reverse order, will get you to your goal.
Advertisement
It is a simple, old and indispensable idea that won its inventor a Nobel prize. Plenty of the last century’s finest drugs have chemical structures so fiendishly complicated that they could never have been made without a logical reverse engineering.
“Chess and chemistry are very similar. They’re both about plotting moves”
Yet with thousands of possible ways to make compounds of even middling complexity, it is tough for humans to spot the best routes. That is why a few chemists think the quickest path to molecules more wondrous than ever lies in taking themselves out of the equation.
Most of the biological world is built of organic, or carbon-containing, molecules. From hormones to vitamins to poisons, organic chemists have long tried to both divine their structures and find ways to make them in the laboratory.
In the middle of the 20th century, chemists generally tried to synthesise new compounds by starting from structures that looked similar to the target. That yielded handy compounds right enough, from synthetic versions of penicillin to the progesterone hormone used in the first birth control pill. But there was a large and exotic landscape of potentially useful compounds that no one had the faintest idea how to produce in the lab.
In the 1960s, the Harvard University chemist decided to make the planning of syntheses more logical. He realised a good way to do that would be to work backwards, which led to the name retrosynthesis.
Take a blank piece of paper and draw the target molecule at the top. Examining the bonds holding it together, the chemist will pick one and break it. Choosing the right bond is where the years of training come in. The bond you break on paper has to be one you think you could make in the laboratory, in what would be the final step of the synthesis.
First bond disconnected, you go again. Step by step, you walk your molecule backwards, stripping away complexity until you reach a structure so simple you can buy it (see “Diagram”). A particularly fiendish structure might take 30 steps to deconstruct to this point.
“You don’t work back to a specific precursor, just to simpler and simpler structures,” says , a synthetic organic chemist at the Australian National University in Canberra. “If you do it without bias, then you end up at starting materials you perhaps wouldn’t have considered.” It is like planning the ascent of a never-before-scaled mountain. Starting from the peak, reverse-plotting the route eventually reveals the base camp from which an ascent has the best chance of success.
“Coming up with a molecule recipe is one thing. Unless you cook it, it doesn’t exist”
The moment of truth comes when you crack open a bottle of your starting chemical. Not every route planned on paper works first time, because this is virgin chemical territory. Still, retrosynthesis has become routine for planning the most difficult chemical targets (see “Totally synthetic”).
Few chemists go back and read Corey’s original papers on the subject, but if you do there’s a surprise, says Matthew Todd, chair of drug discovery at University College London. “I thought Corey was codifying a very human activity, but right from the get-go his explicit aim in developing retrosynthesis was to tell a computer how to do it,” says Todd.
Corey and his team even developed such a program, called LHASA for Logic and Heuristics Applied to Synthetic Analysis. But it wasn’t a genuinely useful tool, being hamstrung by the limited speed and memory of 1970s computers.
Computers have come a long way since then. For Todd, a key moment came in 1997 when chess world champion Garry Kasparov lost to IBM supercomputer Deep Blue. “We often talk about chess and organic synthesis as being similar,” says Todd. Both involve strategically plotting moves. “The fact one area had fallen to a computer made me think about the other.”
When Todd eventually decided to write a paper on the subject in 2005, he was startled at how little had been done since Corey’s work on LHASA. He ended up speaking with Deep Blue’s developers, who told him that formidable computing power wasn’t enough. Chess has so many possible permutations that even a potent computer can’t crunch through them all. The essential ingredient of Deep Blue’s success was the software that encoded the heuristics of chess, the rules of thumb that allowed it to quickly discard bad moves.
At the time, no one was interested in encoding equivalent rules for chemical synthesis. That’s partly because those rules are extremely complex. In chess, there are tens of potential moves from any position. In chemistry, the number of possible transformations for a single step of a synthesis can range from about 80 to several thousand. Even using the conservative estimate of 100 choices per step, a 15-stage synthesis becomes a tree of possibilities with 100 million billion branches. As one industry insider put it to Todd, it was cheaper to pay consultancy fees and get top academics to do the retrosynthesis manually.
But what if we could get machines to teach themselves the rules? That is the promise of the burgeoning field of machine learning. Let’s say we want an algorithm to find films that certain types of people will enjoy. Simply give the algorithm a long list of films and information about them, together with a list of the people who liked them. The machine can then learn to recommend films with certain characteristics to people who like those qualities.
These algorithms are not smart as such, they only learn to reproduce relationships discovered in the training data. The effects can be tremendous, however. An algorithm created by artificial intelligence firm DeepMind taught itself to play the strategy board game Go better than any human.
The most promising attempt at machine learning retrosynthesis comes from Mark Waller at Shanghai University in China and Marwin Segler at the University of Münster in Germany. They developed a neural network, a computing system inspired by the brain, that taught itself the rules of organic chemistry by sifting through a major database of reactions. The program then that typically required six synthetic steps to make. Although the syntheses were not tested in the lab, Waller and Segler asked chemists to distinguish the computer syntheses from human efforts in a double-blind test. They couldn’t.
The trend for artificial retrosynthesis is catching on with big players, including the publishing giant Wiley. It has a commercial database called SciFinder that chemists use to find recipes for individual reactions. Recently, the firm added a computer-aided tool called to this package that suggests whole synthetic routes.
Yet there is no published evidence as to how impressive ChemPlanner’s abilities are. And for their part, Waller and Segler say their system is not ready to tackle the most prized and challenging targets, such as “natural products”, the often medicinally useful and intricately structured molecules isolated from natural sources such as plants or microbes.
One man who hopes he can do better is Bartosz Grzybowski at the Ulsan National Institute of Science and Technology (UNIST) in South Korea. He began his career studying the physics of chemical systems, not cooking up molecules. But, intrigued by the challenge of artificial retrosynthesis, he began exploring the mathematics that would allow a computer to efficiently navigate one of those vast trees with millions of branches. People were mildly bemused when, from 2005, his hardcore maths papers started turning up in chemistry journals, Grzybowski recalls. But it was an essential first step.
From there, Grzybowski encountered the same problem that Todd had identified: he needed a way to help his program quickly discount bad moves. He investigated feeding it a database of reactions, but “the quality of chemistry you get out is pretty pathetic”, he says. So he and his team took a path others had written off and spent years teaching the program 50,000 rules describing why bonds change as they do.
Pulling the rules together took a long time, but by 2012, he had shown that his fledgling program, Chematica, worked in principle. But coming up with a recipe for a molecule is one thing. “Unless you cook something, it doesn’t exist,” says Grzybowski.
He began negotiating to sell the software to chemical supplies firm MilliporeSigma in 2017. The company wanted to test the program on molecules its chemists could produce only in low yield or not at all. The team went on to , including one of a natural product. They all worked and the sale went through, with the program being renamed Synthia.
It is a landmark result. But despite the potential, not everyone is as enthusiastic as Grzybowski.
In fact, the artificial retrosynthesis concept has proved polarising. “Some believe synthesis is mainly about artistry, and that human imagination and intellect, creativity and knowledge, can never be beaten by a computer,” says Sherburn. But then, they said that about chess. Chemistry may be more complicated for algorithms to master, but that day could come. In the nearer future, Sherburn says, he can imagine using programs as he would a colleague, to get new ideas or ask for advice.
Perhaps it won’t just be chemists asking for advice. A few years ago, Grzybowski warned that his program could . But in reality it won’t help much. All the program does is plan a synthesis. It still takes a highly trained chemist with specialist equipment to realise it.
Fears of chemists losing their jobs to AIs are probably also overblown. It is more likely that synthesis will become like playing advanced chess where grand masters face off armed with laptops, says Todd. The computer checks for blunders, which are just as important to avoid in synthesis as in chess. “You want to make sure you’re not missing something,” he says. Synthia proved particularly adept at spotting reactions in which three or more simple molecules zip together in one go.
Even if some see Grzybowski’s work as removing the artistry from chemistry, he can at least take comfort that his work in a sense completes retrosynthesis and brings it back to Corey’s grand vision to drive chemistry forward – by doing it backwards with a computer.
Totally synthetic
Chemists whip up all manner of useful molecules, many of which are effective drugs. Through the years, we’ve scaled more and more difficult synthetic peaks
1964
Longifolene
Chemist Elias Corey at Harvard University came up with retrosynthesis, a method of logically planning how to stitch together chemicals. One of the first molecules he tried it on was longifolene, a component of pine resin, and one of the aromatic molecules found in lapsang souchong tea, which is smoked over pine fires. These days it is not considered difficult, but at the time no one had synthesised it before.
1977
L-Dopa
Starting in the 1950s, thalidomide was prescribed as a cure for morning sickness. It was a mixture of two mirror-image forms, like right and left hands. Doctors didn’t know at the time, but one of these forms caused birth defects. That led chemists to find ways of making the mirror-image forms separately. William Knowles was the first to find a way of mass producing just one of these forms. It was the amino acid derivative L-dopa, used to treat Parkinson’s disease. Since then, such methods have become crucial in drug manufacturing.
1989
Palytoxin
Made by corals, palytoxin is the second most toxic non-protein we know of. Just 3 micrograms is thought to be enough to kill a human. Its production was desired because its huge size and complexity made it the Mount Everest of chemical synthesis. Whereas thalidomide came in two stereoisomers, palytoxin has 1021, only one of which is the correct structure. It was built in eight parts by a team led by Yoshito Kishi at Harvard University before they were all connected together.
1994
Taxol
This chemotherapy drug was first synthesised by two research teams, led by Kyriacos Nicolaou and Robert Holton, within a month of each other. Before that, the drug, also called paclitaxel, was harvested from the sap of rare Pacific yew trees, making it scarce and costly. There are stories from the 1990s of the relatives of people with cancer going into forests looking for the trees. Synthesis helped lessen pressure on the species.
2017
Resiniferatoxin
Resiniferatoxin is another poisonous molecule. At low doses, it binds a protein found in nerves and dials down chronic pain. Synthesising it was daunting, because it contains several rings of atoms fused together. These rings are hard to build and chemically sensitive, so they can easily split apart.
It was prepared in 44 steps in 1997, but renewed interest in the drug prompted Masayuki Inoue of the University of Tokyo to devise a shorter way. There is still a way to go before the drug can be affordably manufactured.
This article appeared in print under the headline “The disassembly line”
