UvA researchers build ChatGPT for chemistry
For five years, UvA chemists worked on a robot that can take over the work of an organic chemist. On January 25th, those results were published in the leading journal Science. With the data generated by the robot, AI will soon be able to predict which reaction works best. “Getting it right the first time would be fantastic.”
Loud music, crowded lab cabinets full of molecules, and students and PhD students performing chemical reactions in round flasks in hot or ice baths. This is the lab of organic chemists, the chemists who spend months in the lab making a specific molecule, including for the pharmaceutical industry.
They recently added a colleague, the synthesis robot RoboChem, which takes up exactly one workbench. RoboChem works 24 hours a day and optimizes the synthesis of some 10 to 20 molecules in one week—a PhD student would take months.
“That was one of the reasons for building the robot,” says Professor Timothy Noël of the Van ‘t Hoff Institute for Molecular Sciences (HIMS), who worked on the robot with software engineers, AI specialists, and organic chemists over the past five years. Organic chemists spend half their time trying to increase the efficiency of reactions, which often starts at five percent of the starting substance converted to product but often needs to move toward 80 percent for publication.
Ratatouille
That task can now be taken over by the robot. The organic chemist is still in charge, though, and determines the ingredients with which the robot will perform the reaction. Noël likes to use cooking metaphors and compares it to the Disney movie Ratatouille: “The computer armed with AI is the rat Remy, who directs the cook from under the chef’s hat to make the dish.”
The chemist feeds the ingredients into the computer, which then gives instructions to the robot. Using a needle, the robot sucks chemicals from test tubes and combines them through thin tubing into the reaction mixture.
That reaction mixture goes to the photochemical reactor, which initiates the reaction using light. Noël says: “Photochemistry is the specialization of our research group, but here you could just as easily put in a heating plate or burner that starts the reaction.”
A measuring device then checks how much product has been formed, and that data goes back to the computer. The algorithm adjusts the reaction conditions based on that information, after which the process starts all over again. Noël explains: “The computer learns from those iterations, and in almost every round the efficiency goes up. In just 20 iterations—depending on the complexity of the molecule—the computer will have determined the optimal reaction conditions to make a new molecule.”
No rush
Last year, the robot was already working properly, but the researchers did not rush into anything. Over the past year, they manually checked all of the robot’s results. “So now we don’t have too many concerns about reproducibility, because we have done our homework.”
The robot also performed reactions previously described in other papers. “And it turns out that in all cases the robot matches and in many cases even exceeds the chemist. The robot finds reaction conditions that work even better, and sometimes twice as well.” Noël is still amazed by this. “You hope it works, but I couldn’t have predicted it would work so well.”
“All this is not so trivial.” He gives an example. “Some reactions use a lot of light. But for some reactions, the light is reduced to two watts. In retrospect, I can explain that, because these molecules are quite sensitive. But I don’t know if I would have figured that out beforehand. The robot discovers this experimentally and learns what the sensitivities of chemistry are.”
AI infiltrates chemistry
For now, AI has infiltrated the world of organic chemistry only with difficulty. Despite the mountains of literature out there on chemical reactions, the algorithm could not predict the outcome of new reactions.
There are two reasons for this. First, each organic chemistry lab has its own recipes. Noël says: “Organic synthesis is one of the most conservative disciplines in organic chemistry. Once the ‘cookbook’ works, it is hardly ever modified, but those recipes are sometimes incomplete and not always reproducible because they include human error.”
Second, reactions are often published only when they produce yields above 50 percent. “An algorithm then quickly concludes that all reactions work, but that is obviously not true. RoboChem collects the data from all experiments and learns precisely from the failures.”
ChatGPT for chemistry
Once the robot has generated a stack of data, the next step is for AI to continue training on it. AI then learns to understand the logic of chemical reactions in the same way Chat GPT can “understand” language. “Once we have that data, we can start using AI to predict what the best reaction conditions are and how much it will yield. That’s the future. Then you only have to perform one reaction, the correct one right one right away. Getting it right the first time would be fantastic.”
RoboChem does not yet have much competition, according to Noël. “There are more synthesis robots, but rarely one that can capture the imagination of a classical organic chemist. You need software engineers and AI specialists as well as chemists; there are only a few research groups in the world that can handle this kind of research right now.”
Companies in the pharmaceutical industry have already shown interest in the robot. “For companies in contract research, the robot is ideal. When a customer needs a certain molecule, the robot can not only find the best reaction for it but also scale up to larger quantities at lightning speed. One company has already invested in a PhD to work with the robot.”
Does that mean the end of organic chemists? Noël doesn’t think so. “There are still people who need to understand how these devices work. But I think the modern chemist will become a jack-of-all-trades: someone who knows robotics, chemistry, and AI.”
