“If I ask an organic chemist to make me a random new molecule, they can’t do it,” says Lee Cronin, a professor of chemistry at the University of Glasgow. “It’s not because they are stupid. It’s because it’s a stupid request. They will ask me what type of molecule and what specification. It could take them one week or ten years. It’s a waste of their time.”
Cronin realised that even though that’s a difficult ask for a human being, it probably wasn’t such a difficult project for a machine learning robot to undertake. “Could we get a robot to start mixing together random chemicals, without any previous knowledge and see what happens? In other words, suck it and see?” Cronin asks. So he decided to build one.
Developing the robot took a couple of years and the first results have been published in the journal Nature. The robot is controlled by a machine learning algorithm and is capable of running six experiments in parallel. The reactions are assessed in real-time by mass spectrometers, nuclear magnetic resonance machines and infrared spectrometers. “I was inspired by how drones use cheap sensors to fly,” Cronin says. “So I equipped my chemical robot with these sensors.” An algorithm then classifies the chemical reactions.
Using nucleophiles and electrophiles, a class of chemical reagents that can easily react with each other, one of the chemists in Cronin’s team first trained the robot to spot reactions using a few examples. This was often a challenge. “Chemists want to give the robot their ideas,” Cronin says “I was very adamant that I didn’t want to. I wanted to see what happened when I remove the bias because I am pretty sure now that we are going to discover completely new reactions and completely new reactivity the chemist could not imagine.”
In the new paper, Cronin and his team report that after the initial training the robot ran simulations — without any chemical data — to predict which chemical combinations were more reactive and then predicted the reactivity of approximately 1,000 chemical reactions with an accuracy of 86 per cent. “What the robot allows us to do is basically make these most reactive discoveries many times faster,” Cronin says. “If this robot gives us nothing else what we’re able to do is cut down the workload of the chemist by 90 per cent.”
These predictions were then confirmed by a chemist and, to Cronin’s surprise, also led to the discovery of four new reactions. “I was asking my co-worker if he was sure about these discoveries because my knowledge of chemical organic chemistry is quite poor," Cronin says. "One of the reasons why I genuinely have little bias because I'm such a idiot. In the future, Cronin would like to run a competition akin to Gary Kasparov's chess match versus Deep Blue, but instead pitting his robot against the best organic chemist in the world.
Cronin’s robot is the indirect result of a very different original project: the investigation of the origins of life. In other words, how the first molecules that could assemble, replicate and evolve were spontaenously created from inorganic matter. “I asked questions like: what was the first organic molecule that could self-replicate, the first protein? It was very hard to answer those questions so I needed to develop a chemical search engine.” After he obtained the first results, he realised that the robot could be used for a much more more practical purpose, like discovering new drugs and chemicals.
Cronin believes the updated version of the robot is, in some regards, better than the AI algorithms developed by DeepMind. “It’s not that DeepMind is not good but it’s purely algorithmic and virtual. That’s a very different problem. Our robot can physically find something new. We've got a robot that actually displays human observation and creativity because it finds new things we weren't expecting,” he says. The new version of the robot is equipped with additional sensors — like enzyme assays and light detectors — that assist in making chemical discoveries with a specific application.
In the future, Cronin expects his invention to become a bona-fide digital chemist: a robot that will be able not only to that can discover new chemistry, optimise molecules and make them purer, and take molecular code and produce chemicals on demand. The applications, of course, are enormous, in theory allowing us to discover and produce everything from genetically tailored medicine to new plastics that don’t pollute the environment.
Contrary to expectations, this particular AI robot will not put organic chemists out of a job. Quite the opposite. “There are a lot of people overhyping AI as some kind of sentient thing,” Cronin says. “The fact is that AI is little more than a regression algorithm. The training comes from the chemist. No chemist, no AI.”
Instead, Cronin believes that the robot will not only save unnecessary labour but ensure that organic chemists will be able to spend more time doing fruitful research. For instance, Cronin and his team have started compiling a database of failed reactions, which in the future will save time by preventing chemists from repeating experiments that have been shown not to work. “Historically, we have done nothing in terms of organic chemistry,” he says. “We have only just started skimming the surface.” He quantifies this by saying that scientists only know about 100 million molecules (with a molecular weight below 500) out of a possible 1,060.
“This means we have made less than 0.000000000000000000000000000000000000000000000000001(10-52) per cent of these molecules in over two hundred years.” Cronin wrote in his cover letter to Nature. “Even if all the 200,000 chemists working at the bench worldwide were each able to explore three reaction conditions per day on average, and each chemist worked around the clock all year round, the maximum number of new conditions that could be explored is around ca. 200 M per year, and only a fraction of these would yield new molecules. Indeed, if we exaggerate this dramatically by assuming that the entire population of the Earth became an organic chemist, and explored reaction conditions full time for the lifetime of the Universe, still less than 1030 compounds could be synthesised.”
In the meantime, Cronin is also making progress with his initial research about the origins of life. In his lab, one of the robots is working with simple molecules in an attempt to spontaneously make more complex ones. These molecules would, in theory, exhibit what Cronin calls “unreasonable complexity” — molecules that “are more complicated then it could be reasonable to occur naturally without a living system like myself intervening”.
“In life is there was no maker, there was no creationist, it was just the world and its chemistry,” Cronin says. “So with our robot we cheated insofar as we put in these sophisticated chemicals that we knew would react and give us interesting molecules. What we are now doing is making those molecules more and more simple and seeing how we could get to complexity.”
Cronin equates what he is trying to do for chemistry to what the Higgs boson has done for particle physics. “I’m building a Large Hadron Collider for the origin of life,” he says. And this might come as a surprise to his team of computer experts. “I didn’t tell them right away that this was the goal of the project,” Cronin says. “Imagine telling a post-doc that they were going to solve the problem of the origins of life problem and that we’re going to build a robot to make organic molecules. They would run away!”