The Robot Scientists Are Coming. But That's Not a Bad Thing

This story appeared in the September/October 2020 of Discover magazine as "Robot Scientists Are Coming." We hope you’ll subscribe to Discover and help support science journalism at a time when it’s needed the most. 

In the beginning there was Adam. We’re not talking about the first human, but rather the first machine to fully automate the scientific process and make a discovery on its own.

Adam looks nothing like a human. It resembles a big box, about the size of an office cubicle. It’s equipped with robotic arms, incubators, a freezer, cameras and other parts to help it do work. Everything it needs to conduct its research is there, including the brain to do it.

The man behind the machine is Ross King, a professor of machine intelligence at Chalmers University of Technology in Sweden. He started building Adam in 2004 to study enzymes in yeast, and later created a second robot — aptly named Eve — to search for potential malaria drugs.

“Adam and Eve are what I call robot scientists,” King says. And these types of machines, which marry artificial intelligence with robotic laboratory equipment, are getting savvier with each iteration.

But what makes these robot scientists so special? Automation is becoming more common in modern-day labs, and AI can assist researchers with myriad projects. It’s the combination of both automation and AI to equip machines with the ability to carry out each step of the scientific process — forming hypotheses, conducting experiments, analyzing data and drawing conclusions — that puts these bots in a class of their own.

Though formal attempts to make robots “think” like scientists began in the 1960s, it wasn’t until the past two decades that Adam and other intelligent machines started to effectively carry out experiments from hypothesis to final report. These complex machines are still rare, but a handful of robot scientists in fields from medicine to mathematics have helped their human counterparts with new discoveries that are setting a precedent for the future of scientific research. And you might hear more about these automated researchers in the coming decade, thanks to a worldwide challenge aiming to create a robot capable of winning a Nobel Prize by 2050.

Adam was designed to study the key ingredient in bread, beer and your favorite fluffy desserts: baker’s yeast. The unassuming kitchen essential is a species of single-celled fungi, Saccharomyces cerevisiae, with a structure complicated enough that it can be used as a model for human cells.

“Even though the last common ancestor between humans and yeast was about a billion years ago, biology is incredibly conservative,” King says. “So most of what’s true for yeast cells is true for human cells.”

For decades, researchers have been studying yeast’s DNA with the goal of linking each gene with its function. Some of these genes code for enzymes, proteins that speed up chemical reactions — like the breakdown of glucose. When the organism’s genome was sequenced in 1996, geneticists were given a mountain of new information. 

But confirming a relationship between an enzyme and a gene still requires running physical tests on yeast in the lab. It’s a laborious task that King, who has a background in microbiology, envisioned could be done more efficiently by a machine.

So King equipped Adam with all it would need to execute this process from start to finish. The robot was programmed with a database containing genomes for multiple organisms, information on the enzymes and instructions for how to scan for potential matches. Adam had access to all the lab equipment and thousands of strains of yeast it would need to actually run the tests to confirm potential matches — and knew how to read the results of the experiments and go back to the drawing board if a match was unsuccessful. In the end, Adam formulated and tested 20 hypotheses, eventually proposing 12 new gene-enzyme matches. 

“There are just not enough biologists around to do all the experiments we want to do to understand how even yeast works,” King says. Robots like Adam aren’t designed to take over the world, steal jobs or make human scientists obsolete — rather, it’s the opposite. A robot assistant with the savvy to think like a scientist can fill the gaps where science lacks the hands to do the work.

Adam was the first machine to both form hypotheses and experimentally confirm them, but has since been retired. King says he’s planning to donate the bot to a museum. Eve is still in use, though King says the machine is dormant while he relocates it from the U.K. to Sweden.

Eve’s claim to fame was a study published in Scientific Reports in 2018, in which the bot discovered that triclosan, a common ingredient in toothpaste and soap, could be a potential treatment for malaria. The compound had been identified before as having potential to stop the growth of the malaria parasite, but researchers had difficulty identifying which enzymes in the body would be most responsive to the substance. Eve helped match the compound from a library of FDA-approved substances to an enzyme target that would respond to treatment. King says he’d like to use the machine to continue research on treatments for tropical diseases.

And in the meantime, he’s planning another project: one to study the biochemical makeup of cells. King calls it Genesis; the ambitious project would test and perfect mathematical models that could fill the gaps in understanding of how cells work.

“We understand some of the basic biochemistry [of cells],” he says. “But we can’t really quantitatively predict what will happen if we do an experiment on [something] as even simple as yeast.” 

King’s robotic duo may have been the first to successfully make automated discoveries, but the origins of modern-day robot scientists date back nearly 60 years. Technology still had miles to go, but in 1965, researchers at Stanford University were attempting to automate the scientific process with early computers.

They began to work on a project called Dendral, an AI composed of two main algorithms. The algorithms were used to identify unknown compounds through mass spectrometry data — information on the weight of atoms that can help chemists determine the structure and qualities of a compound.

Dendral paved the way for the earliest expert systems, a type of AI that trains computers to “think” like an expert. New projects popped up in the next several decades: In 1976, there was Automated Mathematician (AM), a program that generated new mathematical theorems, and in 1996, researchers at Wichita State University published a paper on FAHRENHEIT, which automated chemistry research. Employing new advances in AI to aid math-heavy fields spurred computer scientists to focus on building the “brains” of these robot scientists, while lab automation continued to advance as well.

But both the brains and the bodies of these future robot scientists needed time, and lots of human minds tinkering with them, to expand into the projects we see today. AM, while impressive in its ability to seek out patterns, generated many theorems that were deemed useless by mathematicians. And even Dendral had its shortcomings — its search features, for example, weren’t the most effective, and it had limitations on the size of problems that it could compute. The project, in its original form, no longer operates — there wasn’t a group of chemists who were invested enough in the program to carry on its legacy. But a case study written by the original creators of Dendral in 1991 reported that the project had a significant impact on the burgeoning AI community, providing a window into a future where automation was common in science.

Decades of increased computing power, refined algorithms and new robotic equipment has finally led to the dawn of a new class of robot scientists. These bots are mastering new fields and learning to churn through data day and night; one of them is an MIT-based robot, called the Intelligent Towing Tank.

Towing tanks are a common tool in fluid dynamics and engineering research, often large enough to sail a boat through their confines. The long, skinny pools allow researchers to adjust water levels, waves and other parameters to model how the flow of liquid changes. They can use those results to better understand friction, flow and other elements that might act on a vessel or structure.

Since towing tanks are often used to conduct experiments that try to understand complex physics, conducting experiment after incremental experiment is a laborious task for researchers. But the Intelligent Towing Tank’s robotic program can conduct that research on its own and devise its own follow-up experiments without the help of a human.

So far, one of the machine’s biggest challenges is getting experiments off the ground. Currently, a human researcher has to help the tank form its first hypothesis by setting initial parameters. Adam and Eve had a similar shortcoming — each relied on their creator’s expansive background in microbiology to become an expert.

Specifically, the towing tank was designed to study vortex-induced vibrations (VIVs). This area of research focuses on the forces that objects create on their underwater surroundings, with applications for the way engineers design different structures — specifically on ones subjected to high wind and waves. Like cells and genes, scientists understand the basic workings of VIVs, but the physics of how they work in different settings still leaves gaps in knowledge.

George Em Karniadakis, a professor of applied mathematics at Brown University who co-authored a paper on the tank in 2019, says identifying those unknown areas, and allowing the autonomous tank to explore them, is how the machine helps fill in those gaps.

“We [often] view uncertainty as the enemy,” he says. “But here the idea is that uncertainty is our friend.”

The project was led by then-graduate student Dixia Fan, who was automating experiments in fluid mechanics to get work done more efficiently. So efficiently, in fact, that Fan’s collaborators had trouble finding him anywhere near the lab during the day.

“I would go there to try to find him, but he was never in the room,” Karniadakis says. “But the experiments were going on.”

The tank pulls a carriage that can move at a sustained velocity and apply forces, such as vibration, without a human present. It also knows to pause between experiments to let the liquid settle before moving forward with the next one, to avoid cross-contamination of results.

The machine worked 24 hours a day, whipping through 100,000 experiments with little supervision. Like King’s Adam and Eve bots, the tank creates follow-up studies from an initial hypothesis and carries out research until the computer can draw overarching conclusions from the results.

Challenging the computer to explore the unknown makes it grow more intelligent — it’s as if you were to challenge yourself to get better at tennis by playing against athletes who rank higher than you. As Michael Triantafyllou, a professor of ocean science and engineering at MIT, explains, “They’re going to push you into an area that you don’t know yet.”

“If you always play with people who are of the same level or worse than you, it’s like never exploring the space of real difficulty,” he says. The machine has to do the same: Its experiments need to provide a challenge where it will collect new data and find new ways to present it.

The combination of robotics and artificial intelligence to carry out experiments, however, is something that Karniadakis says will likely be compatible with fields beyond his own. In other words, a robot scientist could hold a Ph.D. in just about any subject — it just takes the right humans to build the bot.

“I think this paradigm will apply to any discipline,” Karniadakis says. “From [studying] a molecule to an airplane.”

Robot scientists aren’t exactly commonplace now, but that may change in the next few decades. One project that could get more robot scientists up and running is setting an ambitious goal: Build a machine capable of winning a Nobel Prize by 2050.

The idea was originally proposed by Japanese researcher Hiroaki Kitano in a 2016 report published by the Association for the Advancement of Artificial Intelligence (AAAI). The call to action specified a need to employ AI to push the boundaries of scientific research — specifically in biomedical sciences — and eventually to the greater realm of discovery.

But it wasn’t until 2019 that a formal plan to turn the challenge into a global initiative started to materialize. Ayodeji Coker, a science director for the Office of Naval Research Global, is at the helm. King and Kitano, along with AAAI President Yolanda Gil, are helping to lead the process. The project is still in the planning stages, but Coker says the group had a recent meeting that drew about 30 people from universities, research groups and government agencies.

Coker is hoping the effort can grow to the same scale as one that Kitano spearheaded in 1997: RoboCup. Nearly every year since, researchers around the globe have competed in a challenge with the ultimate goal to automate a team of humanoid robots to beat players in the FIFA World Cup by 2050. But the competition also offers a number of sub-challenges as well, such as building rescue robots and automated assistants for people in their homes.

“I think that the beauty of that whole initiative was the fact that [they] brought a community together,” Coker says. “[They] made this fun for them to learn and to explore these new challenges.”

Last year, RoboCup had over 3,500 participants and saw representation from 40 countries. The event has traversed two decades, igniting new advances in robotics. In a similar way, Coker wants to offer a variety of smaller challenges that will build up to the ultimate goal of automating Nobel-worthy science. He hopes the initiative will bring together experts of varying disciplines to build up and refine each aspect of an automated scientist — from its ability to navigate around a lab to the algorithms it uses to design experiments. And even if a team doesn’t meet the ultimate goal, they’ll still have contributed valuable data to the field, paving the way for the next researchers to make the robot scientists even smarter.

“We’re looking [from] the ground up and saying, ‘OK, what do we need to accomplish right now in terms of natural language processing, in terms of vision, in terms of perception?’ ” Coker says. Building and refining those individual skills would ultimately create a stronger, more stable template for a robot scientist to effectively communicate with a human scientist.

Creating better bots starts with refining each aspect of the automation process in order to make, quite literally, a well-oiled machine. And a global challenge could attract a younger generation of researchers with a smattering of specialties — minds eager to innovate in new ways.

“We need an engine to drive that creativity,” Coker says. “It’s not about going to the moon; it’s about what it takes to go to the moon.”  

Jennifer Walter is an assistant editor at Discover.