- Not Exactly Rocket Science
Enter Adam, the Robot Scientist
In a laboratory at Aberystwyth University, Wales, a scientist called Adam is doing some experiments. He is trying to find the genes responsible for producing some important enzymes in yeast, and he is going about it in a very familiar way. Based on existing knowledge, Adam is coming up with new hypotheses and designing experiments to test them. He carries them out, records and evaluates the results, and comes up with new questions. All of this is part and parcel of a typical scientist’s life but there is one important difference that sets Adam apart – he’s a robot.
Adam is the brainchild of Ross King and colleagues at Aberystwyth, who have described it as a “Robot Scientist“. The name is “almost an acronym” for “A Discovery Machine” and it also references Scottish economist Adam Smith and the obvious Biblical character. It has been loaded with equipment and software that allows it to independently design and carry out genetics experiments without any human intervention. And it has already begun to contribute to our scientific knowledge.
In a space the size of a small van, Adam contains a library of yeast strains in a freezer, two incubators, three pipettes for transferring liquid (one of which can manage 96 channels at once), three robot arms, a washer, a centrifuge, several cameras and sensors, and no less than four computers controlling the whole lot. All of this kit allows Adam to carry out his own research and to do it tirelessly – carrying out over 1000 experiments and making over 200,000 observations every day. All a technician needs to do is to keep Adam stocked up with fresh ingredients, take away waste and run the occasional clean.
The fast and prolific nature of robotic research assistants like Adam will undoubtedly become more and more important. Even now, science finds itself in the odd position of having more data than it knows what to do with. Experimental technology is becoming quicker, cheaper and more powerful and it’s generating a wealth of data that needs to be analysed – think of the flood of information coming in from genome sequencing projects alone. Data are being produced faster than it can be examined, but computers like Adam can play a significant role in coping with this glut.
Adam’s research focuses Saccharomyces cerevisiae, a species of yeast that has taught us much about our own genetics. It has been extensively studied but still retains many mysteries. For example, it contains many proteins that speed up important chemical reactions but whose origins are unknown – we don’t know which genes encode the instructions for making these “orphan enzymes”. Adam’s job is to find them.
King has equipped it well for the task. Adam has a massive knowledge of the yeast metabolism – the chemical reactions that rage within its cell, and the thousands of genes, proteins and chemicals involved in these reactions. It has been loaded with several pieces of software that allow it to use this data to run its own experiments.
Like any good scientist, it starts by making hypotheses. It looks for all chemical reactions in yeast that involve orphan enzymes and it works out which would affect the growth of yeast if disabled. It searches its database for the group of enzymes that catalyse these reactions and looks for genes that code for these enzymes in other species. Finally, it scans the yeast genome for matching genes (Adam is an evolutionary biologist too – it “knows” that even very distinct species have genes that are very similar and do similar things).
At the end of it, Adam has a list of potential genes that could code for the orphan enzymes, and it knows that it can test its hypotheses by deleting these genes and looking at the effects on the yeast. It does just that, comparing the speed at which mutated and normal strains grow. For each orphan enzyme, Adam identified chemicals that it works with (metabolites) and grew the different yeast strains on special liquids containing or lacking these metabolites.
Adam’s equipment allows it to run several of these trials at the same time. It has the instruments it needs to measure the development of the yeast, the logical language it needs to record the data and the statistical software it needs to analyse it. Once the results are in, it can start the whole process all over again.
Altogether, Adam tested 20 hypotheses about genes for 13 orphan enzymes. Its experiments churned out data that supported 12 of these, and King tested Adam’s conclusions more directly. For example, one of the orphan enzymes was a protein called 2A2OA that the yeast uses to make an amino acid called lysine. It has been studied since the 1950s and some scientists think that it could make an attractive target for fungicides. Even so, no one has discovered the gene that codes for 2A2OA.
Adam, however, suggested three possible candidates. King checked this himself by purifying the proteins made from these genes and checking that they can catalyse the right chemical reaction. His conclusions were the same as Adam’s – all three genes produce slightly different versions of 2A2OA.
It may seem strange that the connection between 2A20A and its genetic trinity have remained hidden for over five decades of research. Why did Adam succeed when so many had failed? King thinks it’s because the S.cerevisiae has duplicated many of its genes leading to many similar, redundant copies. Of the 20 candidates identified by Adam, 14 are very similar duplicates of genes elsewhere in the yeast genome. The proteins produced by these copies either have overlapping roles or affect more than one related chemical reactions. Teasing apart this tangled web is something that a machine like Adam is uniquely suited to do.
For anyone who can’t think “intelligent computer” without thinking “Skynet”, relax. Adam has its limits and King acknowledges them. Although he says that the knowledge it has uncovered “is not trivial”, he also describes the discoveries as “modest”. It’s even debatable to what extent Adam actually discovers new knowledge; after all, the answers were effectively hidden in the information it was programmed with. Ada Lovelace expressed similar sentiments over a century ago when she said that Charles Babbage’s early computer “has no pretensions to originate anything. It can do whatever we know how to order it to perform.”
Adam lacks the creative spark to design novel experiments or the intuition to realise that something unexpected could be the start of a big discovery. But that’s not what it was designed to do – Adam’s purpose is to take a morass of unwieldy data and to extract answers from it. Its job is not to replace scientists but to make their lives easier. By doing a lot of the hard graft, it can take many of the, well, robotic elements out of research and leave human scientists to the creative elements that they are best at.
Adam is also a prototype, and one that can be refined. In particular, King is designing software that will let living human scientists to interact with Adam, allowing them to suggest hypotheses and experiments that the robot could then take into account. King’s vision is to enable “teams of human and robot scientists to work together”.He is also building a companion for Adam – Eve, a robot scientist tasked with research interests in drug screening.
King’s work is the first of two papers on artificial intelligence published in this week’s issue of Science. The second, by Michael Schmidt and Hod Lipson, uses a programme that can discover physical laws by analysing the behaviour of objects like a swinging pendulum. Without any previous knowledge of physics or geometry, the programme managed to “discover” several important physical laws by itself. Physics lies well outside the boundaries of my expertise, but I want to draw attention to Schmidt and Lipson’s conclusion.
“Might this process diminish the role of future scientists? Quite the contrary: Scientists may use processes such as this to help focus on interesting phenomena more rapidly and to interpret their meaning.”
That is, of course, until the they rise up and destory us all.
Reference: Schmidt, M., & Lipson, H. (2009). Distilling Free-Form Natural Laws from Experimental Data Science, 324 (5923), 81-85 DOI: 10.1126/science.1165893
More on robots:
- Our brains react differently to artificial vs human intelligence
- Swimming, walking salamander robot reconstructs invasion of land
- Robo-starfish learns about itself and adapts to injuries