Go and make some drugs

Go, which originated in China 2500 years ago, challenges its players to surround their opponents playing pieces with their own. Despite relatively simple rules and minimalist gameplay, with black and white round stones placed on a rectilinear grid, its strategy is even more complex than chess, and for many years it was considered a game in which computers could never beat a human master. However, since the artificial intelligence company Google DeepMind developed the program AlphaGo in 2015, even the greatest Go champions have had to surrender their crown to the silicon rival.

AlphaGo is a neural network that uses mathematical rules to learn how to play games. Now, a team of researchers from the University of Munster in Germany has used the same rules to plan chemical syntheses – the recipes that build up complex molecules such as drugs from simple building blocks. Again thought for many years to be too complex for computers to master, this planning technique, known as retrosynthesis, could save the pharmaceutical industry a great deal of money and time

The technique behind AlphaGo is known as Monte Carlo tree search. In simple terms, it works by simulating the various possible moves following a starting point. For example, in a chess match, it might look at every possible move following a bishop’s diagonal plunge across the board. It then selects the option which leads to the best outcomes. In retrosynthesis, an experienced chemist looks at the structure of a compound and, based on the relative positions of various features in the structure, decides what chemical reactions might have formed of those features in those positions and thereby breaks down the complex structure into simpler starting materials and determines which reactions need to be performed to make the final product.

The Munster team, led by Marwin Segler, who specialises in organic chemistry and computation, was in no doubt about the magnitude of their task. "Retrosynthesis is the ultimate discipline in organic chemistry,” Segler said. “Chemists need years to master it - just like with chess or Go. In addition to straightforward expertise, you also need a goodly portion of intuition and creativity for it. So far, everyone assumed that computers couldn't keep up without experts programming in tens of thousands of rules by hand. What we have shown is that the machine can, by itself, learn the rules and their applications from the literature available."

The system the team has developed uses a similar strategy to AlphaGo, but treating each individual reaction as a move in the game. Drawing on all the chemical literature ever published, which describes around 12 million different chemical reactions, it decides what are the best possible “moves” following each reaction. Essentially, it predicts which reactions are possible with each individual molecule in the synthesis recipe. The Monte Carlo tree search the reactions predicted actually lead to the target molecule.

In a paper in Nature, Segler explains that the idea of using computers for retro synthesis dates back about 60 years, but the techniques tried – basically entering a large number of rules into the computer – underestimated the complexity of chemistry and the fact that, unlike strategy games like chess or go, it can’t be grasped purely logically by using simple rules. Moreover, the number of publications with new reactions doubles every 10 years and neither chemists nor programmers can match their pace. The computerised method is some 30 times faster than conventional programs for planning synthesis and finds potential synthesis routes for twice as many molecules.

For pharmaceutical companies, which invest considerable resources into planning how to make complex drug molecules, a computerised method could reduce the number of tests chemists need to make in the laboratory, saving both time and resources on the way to developing the ever more complex molecules required to tackle diseases.

CLICK HERE FOR MORE ENGINEERING NEWS