Meet the scientists behind the robo-scientists

THERE WAS A MOMENT of scientific synchronicity this week when Science ran not one but two reports from teams building robot scientists.

Cornell University's Hod Lipson announced that its software robot can derive fundamental laws of nature from raw data. And Aberystwyth's Ross King's robot scientist, Adam, can not only think up hypotheses about the workings of genes in baker's yeast and devise experiments to test them but drive automated equipment to carry them out, assess the results, and revise the hypotheses and run new tests.

"I would agree it's the beginning of a new way of doing science," says Lipson. "As we move into these increasingly complex systems there's no way a single human can wrap their mind around all the information and distill it into a compact model. It might be that there is no compact model – and that we need to move into different ways of representing knowledge, or that there are but finding them is very, very difficult and we need better tools."

Unlike today's computer models of complex systems such as weather, Lipson's robot scientist isn't limited to the assumptions held by humans. Instead, it engages in what Lipson calls "active learning". Like a good physics student, it analyses data – say, how a pendulum moves in specific circumstances – and then asks questions: if you put a pendulum in this position, what does it do? Its humans then carry out the experiments it's asked for and feed it the results.

"It's nice to see what questions it asks, how it explores," says Lipson. The most difficult challenge, he says, was finding a way to define for the computer what is and is not trivial in its search for constants. You want it to focus on underlying mathematical constants and physical laws, things that will be general principles, not the length of string on that particular pendulum. "There are an infinite number of trivial things to say about the data, so definining what's interesting was holding us back for a while. Once we were able to do that, results starting coming out very quickly."

King's Adam is more self-determined; it's built on existing laboratory automation systems that can shake plates, control temperatures, and so on. King has been working on this project for ten years and says they showed five years ago it could generate hypotheses and test them. Yeast cells are some of the best-understood cells on the planet; the genes that have obvious effects have been discovered. Adam has been constrained to look for small effects.

"When we started," says King, "we thought it would be the hypothesis formation that would be the hard part." Instead, "It actually turned out to be analysing results." They had to develop, he says, "quite complicated statistical methods to tell whether we were getting a real difference or not, similar to techniques people used in the 1920s to decide on new varieties of crops."

Both King and Lipson are talking about these robot scientists as tools for humans. "It's a branch of artificial intelligence, yes," says King. "But I wouldn't like to call it sentient." Besides, it doesn't need to be: "It doesn't have to understand it's doing science to do science."

What these robots will do, he hopes, is automate most laboratory work: robots keep better, more complete records, and can be more precise about controlling environments and carrying out lengthy series of repetitive experiments. "Biologists are probably one of the hardest working groups of scientists," he says, "because success depends on doing a lot of experiments because you can't guarantee what's going to work – and biological systems are so complicated that there aren't enough biologists on the planet to do all the experiments we need to do to understand how it works." The hope is to "free up biologists to spend more time thinking rather than pipetting."

For Lipson, the robot is suggesting new avenues of research. "We're working on a different problem we haven't published yet, where we have an answer equation – and don't know what it means." µ