The automated lab of tomorrow

By combining automation and AI, labs could see big boosts in speed, efficiency, and even creativity.

In 1875, Thaddeus M. Stevens devised and described one of the first pieces of automated lab equipment—a sealed vessel with a controlled drip rate to wash filter papers (1). He’d likely be surprised to see how many tasks academic scientists still carry out manually 150 years later.

The increasingly common marriage of AI and automation in basic science labs could bring much better efficiencies—and potentially new paths to discovery. Image credit: Dave Cutler (artist).

Unlike industry and clinical research, which have both embraced and benefited from automation, bench science still depends to a large degree on the low-cost labor provided by postgraduate students and postdocs. That leaves academic science with an efficiency problem. Labs often sit empty at night and on weekends, while the short contracts and high turnover of junior staff lead to the loss of valuable knowledge and experience.

AI can help. Researchers across disciplines are finding ways to integrate machine learning into their experimental setups. Combined with automation, these systems—called self-driving labs—are being used in fields across biology, chemistry, and materials science to design and execute repetitive steps, analyze data, and then tweak the next cycle of experiments to build on the results. Self-driving labs have already been used to design smart materials and find enzymes that human scientists would have struggled to produce. Meanwhile, universities, funders, and the private sector are investing in so-called cloud labs, which offer remote access to robot workers and lab equipment.

How far could this marriage of AI and automation take basic science? The founders of the so-called Nobel Turing Challenge, a contest computer scientists started in 2020, believe that in the next 30 years, autonomous laboratories could yield discoveries worthy of winning a Nobel Prize—with little or no help from people. Critics, though, worry about hype and, worse, that in the rush to leverage machine learning to automate research, we could be introducing new ways for critical systems to fail. “AI is an important tool that can be used by humans to support humans,” says Missy Cummings, director of the Autonomy and Robotics Center at George Mason University in Fairfax, VA. “But it needs to be supervised by humans.”

Improved Automation

Academic research hasn’t traditionally lent itself to loads of automation. Protocols that change frequently, for example, are harder to perform with robots. As a result, true self-driving labs with fully automated systems remain rare. “As researchers, what interests us today might not interest us a year from now,” says Curtis Berlinguette, a chemistry researcher who runs a self-driving lab at the University of British Columbia in Vancouver, Canada. “So, if you want to get good value from your automated system, you want to make sure that your robot evolves with your experiments.”

Another sticking point for many academic labs is experiments that rely on multiple steps that are difficult to link together. “You can go out and buy commercial equipment that does very specific tasks, like an autosampler for a gas chromatograph or an autosampler for an NMR experiment,” Berlinguette explains. Few automated tools can link together commercially available automated experiments. “So, if you want to develop a fully automated workflow, connecting each one of those tools or modules is currently really challenging,” he says.

While AI receives most of the attention, Berlinguette argues that a focus on improved automation is essential for researchers to leverage machine learning in science. That’s because robots are the only feasible way for many labs to boost production of the sort of reliable data that AI feeds on. “Machine learning really requires automated workflows,” he says. “When done properly, they allow you to catalog and process data far more effectively than you could through manual experiments.”

Such better-quality and annotated data streams can then allow AI to analyze the results and decide what experiment the robot should do next, which produces more data for the AI to analyze, and so on. “This is what separates a self-driving lab from high-throughput experimentation,” Berlinguette says. “High-throughput experimentation creates these massive, massive datasets, but you’re not necessarily processing that data on the fly.”

Material Differences

Berlinguette’s lab features an automated system called Ada, named after Ada Lovelace, one of the pioneers of computer programming. Ada uses a robot arm to make thin-film materials used in perovskite solar cells. The performance of these films is highly sensitive to thickness, morphology, and defect chemistry, which are affected, in turn, by process conditions, such as additive concentration and annealing time and temperature.

Ada can vary these conditions and, by comparing the properties of the resulting films to a set goal, can repeat and iterate the experiment until it produces the best-performing material. It’s not a trivial task. Each run must draw a precise amount of liquid solution from a vial and drop it onto a slide, spread the solution into a thin film, and bake it with hot air.

Ada then autonomously measures the physical properties of the resulting film, such as electrical conductivity and optical spectra. Next, it passes the results to an open-source optimization algorithm, called Phoenics, which plans the protocol for the next experimental round and issues instructions on how to set up the conditions. Each cycle takes about 20 minutes.

In 2020, Berlinguette’s team asked Ada to find the concentration of cobalt additive (called the dopant) that gave the best performance, measured by how well the oxide material in the film conducted electricity and absorbed radiation. It took 35 cycles—about 30 hours—to determine the best combination: add 0.4 equivalents of cobalt and anneal for less than 75 seconds (2). “That’s more effective than a human experimentalist, who has to take breaks for coffee and lunch and dinner and sleep,” he says. It could also improve the reproducibility of results by ensuring that the experiment is done the same way every time.

By marrying automation and AI, other self-driving setups have had similar successes. Late last year, researchers from the Lawrence Berkeley National Laboratory (LBNL) and Google DeepMind reported how they paired an algorithm that conceives and designs the structures of new inorganic materials with a robot system called the A-Lab that can make and analyze them (3).

Of the hundreds of thousands of possible new materials identified in theory, the project fed the A-Lab’s AI a shortlist of 58 targets. The algorithm, trained on a database of 30,000 existing synthesis protocols, gauges how closely the targets match existing materials. Drawing on the established procedures for making those known materials, the algorithm then selects reagents and reaction conditions that it thinks will be able to make the new, theoretically possible materials and then attempts to do so. If these methods don’t work, the AI compares the selected procedure with successful recipes devised by the algorithm and suggests another procedure.

In just over two weeks, this LBNL self-driving lab (set up over 18 months with $2 million from the US Department of Energy) produced 41 new inorganic materials; 9 of these were made after the “active learning” algorithm invented new synthesis methods that went beyond its training data. Some of the materials that the A-Lab failed to produce on its own were successfully created after human scientists reintroduced themselves into the loop—regrinding a mixture midreaction, for example.

Carnegie Mellon University claims to have the world’s first academic cloud lab, pictured here and opening this year. At Emerald Cloud Lab, researchers and students can order experiments online. A combination of robotic instrumentation and trained technicians will perform the experiments as specified. Image credit: Jonah Bayer (Carnegie Mellon University, Pittsburgh, PA).

Beyond Trial and Error

AI systems that essentially learn as they go can optimize work—though often without revealing to humans the secrets of their success. Case in point: Biochemists at the University of Wisconsin–Madison have built an automated AI system that designs, makes, and tests proteins—including steps to combine strands of DNA and express proteins from them in cell-free culture.

In a paper published in January (4), the researchers used the system, called SAMPLE, to engineer glycoside hydrolase enzymes more tolerant to heat. In four different runs, the autonomous scientist quickly produced sequences that were stable at 12 °C higher than previously known versions. “It explores the space and tries new combinations and learns from the data to refine its understanding over time,” says Philip Romero, a researcher at the University of Wisconsin–Madison, who worked on the project.

A lab in the cloud

Cloud labs are remote facilities packed with equipment to which researchers can outsource their experimental work. Working 24/7 and 365 days a year, such labs offer scientists anywhere the opportunity to gather data without needing access to their expensive in-house apparatus. They can design experiments and upload the protocols, and the cloud lab will carry out the procedures and send back the results and data.

Researchers at the University of Wisconsin–Madison have used a cloud lab to discover new heat-tolerant enzymes—using the researchers’ own algorithm to design proteins, but then sending the instructions to make them to a cloud lab some 2,000 miles away in California, which used robots to carry out the enzyme synthesis and analysis. “The cloud lab would send back raw data, and a computer on our end was putting it into the machine learning model to update the system’s understanding,” says Philip Romero, a researcher at the University of Wisconsin–Madison, who worked on the project.

Only a few cloud labs have been set up, and it’s a tough business model to get right. The company Strateos, which owned the fully automated lab in Menlo Park that Romero’s team used for their project, folded last year. (The company announced it was restructuring in April 2023, but couldn’t be reached for comment regarding its current operating status.) That leaves Emerald Cloud Labs (ECL), which runs a remote facility in Austin, TX.

Working with ECL, Carnegie Mellon University in Pittsburgh, PA, has built its own $40 million cloud lab, which features more than 200 automated scientific instruments, from pressurized reaction vessels and mass spectrometers to nuclear magnetic resonance and Western blotting analysis.

Federal funders have shown interest. The NSF has sponsored a series of workshops to discuss establishing a national network of cloud and self-driving labs, under its new Directorate for Technology, Innovation and Partnerships (TIP) that aims to boost US competitiveness.

But if automation and AI are to work effectively in concert, researchers will have to agree on common methods, says Milad Abolhasani, a chemical engineer who works on self-driving labs at North Carolina State University in Raleigh, and who organized an NSF-sponsored workshop on autonomous labs in January. “Everybody in academia and industry thinks their way of doing hardware development is the right way,” he says. “It’s going to take us probably another 7 or 10 years for people to eventually give up on something that might not be the most efficient way of pursuing that automation.”

The AI works with more than trial and error though. “The AI also builds an internal model of the relationship between protein sequence and function. And that allows it to kind of accelerate the search process,” he adds. Once optimized, the automated system should be able to carry out in two months the same number of experimental runs that would take a human scientist up to a year to perform, the team says.

“Most things that we thought AI could do have eventually turned out to be possible. It’s just whether it’s going be next year or a hundred years is the hard part to say.”

—Ross King

Working out the links between DNA sequences, and the structure and function of proteins for which they code, is a notoriously hard problem. Romero admits that his team does not understand what clues their AI robot identified to help it complete its task. They just know that they couldn’t have done it themselves. “It is so far beyond what humans can comprehend, it’s amazing,” he says. “AI and machine learning have completely transformed the field of protein science over the last five years.” Probably the most high-profile project in this fast-growing application of AI is AlphaFold. Another Google DeepMind effort, this algorithm predicts protein structure based on its amino acid sequence. (See Researchers turn to deep learning to decode protein structures.)

Although these kinds of problem-solving algorithms accelerate scientific projects, critics have questioned how much they contribute to scientific knowledge. Despite headlines to the contrary, structural biologists have pointed out that protein folding itself is not solved (5). (See also Protein folds vs. protein folding: Differing questions, different challenges.) That would take a fully outlined mechanism that links the amino acid sequence to folding properties. At present, that explanation, in whatever state, resides only in the circuitry of AlphaFold.

Still, Romero says he is less interested in the theory of how his SAMPLE system connects the dots to build better enzymes than in how those enzymes could be used in practical applications. “You get something that’s better and it’s good enough, and then you move on to the next,” he says. (The thermally stable enzymes could find use in biomass deconstruction, he says—for example, degrading plant material into sugars that can then be converted into fuels or chemicals.)

Toward New Science

Some experts think AI can offer more to scientific investigations than such practical demonstrations of its exemplary problem-solving abilities. In a provocative 2021 article (6), Hiroaki Kitano, head of the Systems Biology Institute and AI chief at the electronics giant Sony, claimed that AI could develop “an alternative form of science that will break the limitation of current scientific practice largely hampered by human cognitive limitation and sociological constraints.”

Such a system could empower an “AI scientist” to make discoveries worthy of a Nobel Prize, he said, a goal formalized in the Nobel Turing Challenge project, which has set a deadline of 2050 to make it happen (6). “Twenty years ago, most people thought it was a crazy idea,” says Ross King, a computer scientist at the University of Cambridge in the UK and one of the organizers of the initiative, which held its first workshop in 2022. “Now it seems most people I speak to think it’s probably going to happen sooner or later.” Thus far, the Nobel Turing Challenge consists of a series of workshops and meetings to discuss progress and priorities in AI as they related to laboratory science; the most recent took place in Tokyo in February.

An AI scientist working at that level would have to do much more than execute the kind of iterative experiments seen in existing self-driving labs. They would have to make strategic choices on the topic of research to pursue, communicate the reasons why and mechanisms involved, integrate the results into the scientific literature, and make judgments about both the applications and the social implications.

Such an AI scientist would effectively generate progressive hypotheses. But there’s a major hurdle to cross before that could happen: developing machine learning techniques that can represent and extract knowledge themselves—by, for example, identifying how an idea in one discipline might apply to something quite different. The most ambitious way to achieve this—and some say the best way—would be to build so-called general-purpose AI (GPAI), a much-hyped versatile tool that would be across all areas of science and could be applied to any task.

Many experts think that’s not a realistic goal for anytime soon. But King says recent progress using AI to play chess and Go shows that general-purpose capabilities might not be essential for devising AI that’s able to wield something resembling creative thinking. “GPAI turned out to be unnecessary, and it was possible to build machines that are world-class at chess and Go but able to do nothing else intelligent,” he says. Any future AI system that could do Nobel Prize-quality research would certainly need to be more general and more intelligent than these game-playing machines. “But it’s not clear if they need to be as general and intelligent as a GPAI,” King says.

As AI technology improves, and even if general AI can be achieved, King, like others in this field, emphasizes the importance of wider and more reliable access to automation of lab procedures. That, he says, holds the key to converting the promise of machine learning into world-class science. “It’s much easier to make radical advances in software than in hardware,” he says. “One of the reasons that machine learning has become so popular is it’s becoming easier for people to just take tools and use it. But it’s much harder to control robots and things like that.” Cloud labs, remote, automated facilities on which researchers can buy time instead of acquiring their own equipment, could help (see A lab in the cloud).

Tracking AI Capabilities

Given the rate of progress, even seasoned experts like King are unwilling to speculate on how far toward the Nobel Prize-winning goal of AI-powered, autonomous science may be reached within the next decade.

But while it’s difficult to predict development, King argues that it’s important to have a way to measure and track it, to help steer research in the right direction. In June of last year, he and Hector Zenil, a colleague at the University of Cambridge, outlined a means of benchmarking that progress, based on a series of steps toward what they call Level Five: a fully automated AI system that would be equivalent, if not superior, to a human scientist (7).

The phases are marked by a steady decrease in the need for human curation. Current capabilities in science are moving toward Level Three, the duo argued, with AIs able to choose the best models for data, for example, but with humans still outlining the hypothesis and the type of solution required.

How much longer will the science enterprise need that personal touch? “I don’t think there’s anything deeply different about humans that we can’t get a machine to do it,” King says. “Most things that we thought AI could do have eventually turned out to be possible. It’s just whether it’s going be next year or a hundred years is the hard part to say.”