SCI: Self-driving labs are coming: What does that mean for chemistry?

BY MARIA BURKE

Self-driving labs based on robotics and AI are starting to perform experiments traditionally done by researchers. Maria Burke looks at what this means for the future of chemistry.

Self-driving laboratories (SDLs) use robotics, automation, AI and machine language algorithms to autonomously conduct experiments, gather and analyse data, and refine decision-making processes[1]. Researchers specify the objective, such as optimising a specific catalyst. AI designs the experiment, robots perform the work, instruments take measurements and data are analysed automatically. AI then updates its model and selects the next experiment.

‘SDLs accelerate scientific discovery through closed-loop processes [which] speed up research by focusing on the most informative experiments,’ says Sean Caffrey, executive director of the Acceleration Consortium at the University of Toronto, Canada, which coordinates materials discovery using SDLs. ‘Originally, SDLs were designed for specific problems. Today’s more versatile labs use ‘plug-and play’ components, along with advanced orchestration software, allowing easy reconfiguration for various chemistry projects.’

SDLs can now collect comprehensive data – environmental factors, reagent batches, and equipment settings – which enhances reproducibility and data integrity, Caffrey says.

‘The real advance is not simply that a robot can perform experiments, but that every experiment can inform what should be done next,’ says Timothy Noël of the Van’t Hoff Institute for Molecular Sciences at the University of Amsterdam, the Netherlands. ‘This allows us to explore complex chemical spaces faster, generate highly reproducible and machine-readable datasets, and increasingly reuse knowledge from previous experiments rather than starting every new project from scratch.’

The latter is called transfer learning – a machine learning technique that uses what has been learned in one setting to improve generalisation in another setting – and the team used it for the first time in a study reported recently[2]. This paper presents six case studies that demonstrate the team’s SDL automated flow-chemistry system called RoboChem as it tackled different chemical optimisation challenges, including photocatalytic transformations, biocatalysis, thermal cross-couplings and enantioselective catalysis. The team tailored algorithmic workflows and hardware modules to each problem and showed the optimised solutions could be applied to real world examples.

John Gregoire, chief autonomous science officer at Lila, a US company working on integrating generative AI with lab robotics, believes SDLs offer a new paradigm for chemistry and related sciences. ‘Empowering AI to run its own experiment designs is the biggest change to science in a century,’ he says.

Ahmed Ismail, co-founder and COO of Dunia, a Berlin-based company working on SDLs for materials discovery, agrees change is afoot. ‘[In materials science] traditional breakthroughs take around 15-20 years and anywhere between $100-$500m to discover and scale. Each step in a material discovery could take one to two PhD students around eight years. But when you automate and run hundreds of experiments in a short time period, you can quickly and cheaply know if you need to change strategy.’

Workforce worries

As practical work becomes increasingly automated, some scientists are concerned there will be fewer research jobs[3]. However, Noël regards an autonomous platform as another member of the team. SDLs allow researchers to concentrate on creativity and decision-making, he says. ‘Every chemist should learn how to perform experiments and develop experimental intuition. But that doesn’t mean they should manually perform hundreds of repetitive experiments when there is a better tool available.’

Gregoire agrees: ‘AI-powered SDLs shift the role of researchers upstream.’ The tasks now are more intellectually demanding, like framing the right question, designing the experiment and interpreting results generated far faster than by hand. ‘Training programmes need to change,’ he says, ‘and that transition is genuinely uncomfortable for people who built their identity around traditional methods.’

‘What we’ve found in practice is that our chemists end up doing more chemistry, not less,’ Ismail says. ‘Only the repetitions, pipetting, and similar steps are not being done by them.’

Caffrey says we will still need the brightest minds to ask big scientific questions and decide which problems are worth pursuing, but notes: “In the long run, the structure of research, labs, and the scientific process may change significantly.’

Safety and security

Another worry is that SDLs will compromise safety if they struggle to cope with issues such as fires or spills[4]. But because safety protocols can be encoded in automated labs, researchers using SDLs contend this creates a safety system more robust than traditional labs. The fact that hazardous reactions can be carried out without human personnel performing them is another safety gain.

Caffrey’s team is specifically developing systems that can monitor and oversee the operation of safe SDLs[5]. Such agents can modify experimental procedures, enforce restrictions on prohibited or dual-use chemicals, and generate safety documentation for autonomous workflows.

But what’s to stop SDLs undertaking rogue reactions, or being taken over by a hostile agent?

As SDLs become cheaper and easier to use, it will become very difficult to prevent rogue actors from building their own SDLs, Caffrey says. While acknowledging that IT security systems can prevent access: ‘Controlling rogue actors would require government and international cooperation to control SDL components,’ he says. ‘Low-cost, custom-built SDLs will be impossible to control.’

Rogue actors are also a problem in synthetic biology and other fields, he points out.

Ismail believes industry standards are needed to control newer AI models that can bypass, at least partially, some of the controls, limiting their ability to produce uncontrolled results. ‘Unfortunately, the current geopolitical landscape doesn’t encourage companies and countries to collaborate more. It has pushed everyone to compete, which hinders industry’s ability to agree on a protocol of protection.’ But he’s hopeful that such agreement will happen soon.

Other factors that may limit the widespread use of SDLs include the requirement for substantial financial investment and increased technical expertise.

Initial SDLs have been niche, developed to do a specific type of research and to solve a specific problem, says Gregoire. ‘Since SDLs require substantial capital, nicheness makes it hard to realise a return on investment because the returns are limited to those of the specific problem. Generalisation is the key to success, and generalisation requires scaling beyond the scope of initial SDLs.’

Noël’s team says the second version of their automated flow-chemistry system RoboChem – the first came out in 2024 – addresses cost and complexity. It combines readily available parts with those that can be made through cost-effective in-house 3D printing, offering flexibility and affordability[2].

RoboChem-Flex can be built for around €5000, he says. Although this doesn’t include expensive analytical equipment, he points out that labs already share infrastructure such as NMR, so several inexpensive robotic platforms could share analytical equipment.

As a result, Noël expects SDLs to become much more widespread. ‘The remaining challenge is not only hardware; it’s demonstrating compelling applications that convince mainstream chemists that these systems genuinely improve their everyday workflow. Our next goal is therefore to make the systems increasingly portable and demonstrate them in laboratories beyond our own.’

Caffrey says the growth of SDLs in academia, and the significant government and private-sector investments, suggest that demand will increase, either for ‘low-cost’ automation or off-the-shelf flexible SDLs. His team is building ‘a catalogue’ of low-cost lab-built automation unit operations that can be used in a plug-and-play fashion to create a self-driving workflow. They’re also developing open-source software that can link these unit operations to an AI system which can use them to run experiments.

Most instrument companies have bought into automation and are increasing instrument interoperability, allowing them to connect into workflows, says Caffrey. In future, he sees this trend for off-the-shelf packages continuing, providing more options for researchers at both the low and high ends of the market.

Defining IPR

Enshrining intellectual property rights (IPR) in national patent laws has been a key facilitator of technology commercialisation. Patent office guidance on SDL-generated inventions is that they are excluded from patent protection. But could these restrictions limit the field’s development if there is no path to patent protection for non-traditional inventions?

‘[To receive a patent], a physical person needs to be the inventor,’ explains Rafa Gómez-Bombarelli, chief science officer of physical science at Lila Sciences. ‘A corporation cannot be an inventor in the same way that AI cannot. [With SDLs], scientists will sign the inventions as the people that operated the AI and instructed it.’

However, Ismail acknowledges that there are ‘grey areas’, which will get harder to define as systems become more autonomous. ‘My expectation is jurisdictions will clarify what counts as a sufficient human contribution rather than grant inventorship to machines. That’s the path of least disruption to existing law and will have to be implemented at some point.’

If companies are unsure in future whether their filings will hold, they may move towards making their inventions trade secrets, he adds. ‘On the other hand, if SDLs produce so much data, the scientific community may favour more data sharing and open innovation.’

But could a time come when SDLs will be able to make truly conceptual advances? Gregoire is cautiously optimistic. ‘Today’s systems are excellent at optimising within a space a human has defined,’ he says. ‘Whether they can originate an entirely new theoretical framework the way a human scientist does is a much harder, more open question. We are seeing glimmers of this type of emergent behaviour, but we have a way to go before this becomes commonplace.’

Ismail points out that current SDLs do not reframe the problem by themselves or think critically outside the boundaries set for them. The problem is the lack of enough high quality (unbiased) data to enable AI agents to learn and explore new spaces on their own. ‘In all the last five years of experimentation and after running thousands of experiments and building two SDLs, we have seen only two examples at Dunia where this happened,’ Ismael says. To get the best data possible, the Dunia team has built their own database from scratch rather than relying on data from literature searches, for example.

However, Noël says systems are moving from automation towards learning. ‘Our recent work on transfer learning, for example, asks whether knowledge gained from one optimisation can make the next one more informed. The next frontier is to move beyond asking, “Which experiment will give the highest yield?” towards, “Which experiment will teach us the most?” Whether machines will eventually make truly conceptual scientific advances remains open, but I think human-machine collaboration will certainly help us reach insights that neither could reach as efficiently alone.’

AI can explore regions that scientists might otherwise overlook – a form of engineered serendipity, adds Caffrey.

SDLs might discover a new class of materials with unusual properties, then study why those properties appear, leading to new explanations and concepts, not just better materials. But, he says, genuine conceptual advances will also require models that can make reliable predictions beyond their training data and reason about causes and mechanisms. ‘[However], interpreting and validating those advances will probably remain a joint effort between humans and machines for some time to come.’