Robot automates science
Technology Research News
What better entity to assign repetitive
scientific tasks, like working out the function of specific genes, than
A group of researchers from from the University of Wales, Robert
Gordon University in Scotland, and the University of Manchester in England
have put together a robot scientist that can devise a theory, come up
with experiments to test the theory, carry out the experiments, and interpret
The researchers put the system through its paces testing yeast
genes, and also had a control group of computer scientists and biologists
perform the same task. "The robot performed as well as the best humans,"
said Ross King, a professor of computer science at the University of Wales.
The researchers also showed that the robot scientist's method
of selecting experiments was both faster and cheaper than choosing the
cheapest experiment or simply choosing experiments randomly, said King.
The robot scientist was three times cheaper than choosing the cheapest
experiment and 100 times cheaper than random selection, according to King.
The approach could make scientific research less expensive, and
could be applied within a few years in areas where the level of laboratory
automation is already high, like drug design, said King. Today's state-of-the-art
drug design uses brute force automation.
The robot scientist consists of a computer running artificial
intelligence software, a fluid-handling robot, and a plate reader that
checks the experimental results for variables like color.
The software allows the system to "infer new scientific hypotheses
and plan efficient experiments to test those hypotheses," said King. The
robot conducts experiments by dispensing and mixing liquids, then measuring
the growth of yeast using the plate reader, which feeds the results back
into the system, he said. There is no human input in the design of experiments
or interpretation of data, King added.
The researchers gave the robot the task of testing gene functions
in Saccharomyces cerevisiae, also known as baker's yeast. "The robot scientist
generates a set of hypotheses from what it knows about biochemistry and
then plans an experiment that will eliminate as many hypotheses as possible
as fast and as cheaply as possible," said King.
The robot scientist looks for the function of a given gene using
knockout strains of yeast that have had one gene removed. Observing how
yeast grows, or does not grow, on surfaces that contain specific chemicals
gives the investigator clues about different possible functions for the
gene, he said. "This is like trying to understand what the different components
in a car do by removing them one by one."
The robot evaluates the results against the set of hypotheses,
interprets the results to eliminate hypotheses that are inconsistent with
the data, generates new hypotheses, and repeats the process, said King.
This is the same type of cycle human scientists use to understand the
world, he said.
This standard process is relatively tedious for humans to carry
out, however. The functions of about 30 percent of the 6,000 genes contained
in baker's yeast are still unknown, said King. "With many of these genes
thought to be common to the human genome, they could prove to be medically
important," he said.
The software challenges involved in constructing the robot scientist
included encoding all the relevant background information in a form that
the system could use, developing a way of inferring possible hypotheses,
and developing a way of inferring efficient experiments, said King. The
engineering challenge was to put everything together into a working system,
The researchers have demonstrated the system as a proof of principal.
The next step is to show that the system can discover the function of
genes that are currently unknown, said King.
The researchers drew on a 30-year history of research in artificial
intelligence to make the system, said King. "The application of artificial
intelligence to science is known as the field of scientific discovery,"
King said. " I think the main step forward of our work is... connecting
such programs to a physical robotic system."
The work is solid, and important, according to Pat Langley, director
of the Institute for the Study of Learning and Expertise located at Stanford
University. It differs from previous systems that use artificial intelligence
to control robotic equipment because it takes advantage of background
knowledge, is aimed at designing efficient experiments, and uses a closed
experimental loop so that the results inform the next hypothesis.
The work is part of a branch of artificial intelligence, dubbed
active learning, that develops algorithms that weigh the odds of hypotheses
being correct and the costs of potential experiments to determine the
optimal series of experiments to eliminate all but the correct hypothesis.
In general, techniques for cognitive science and artificial intelligence
should be as applicable to modeling and replacing scientific discovery
and experiment design as for more mundane tasks, said Langley. The researchers
work is a step along these lines, he said.
The ideas have precedents, said Raul Valdes-Perez, president
of Vivisimo, Inc. and an adjunct associate professor of computer science
at Carnegie Mellon University. But "I would say that this is the
first convincing demonstration of a link between completely automated
physical experimentation and hypothesis generation and testing,"
The robot scientist could be ready for practical use in three
to six years, according to King. The first practical use is likely to
be making drug design more efficient, he said.
Ross's research colleagues were Kenneth E. Whelan, Ffion M. Jones
and Philip G. K. Reiser of the University of Wales, Christopher H. Bryant
of the Robert Gordon University in Scotland, Stephen H. Muggleton of Imperial
College, London in England, Douglas B. Kell of The University of Manchester
Institute of Science and Technology (UMIST) in England, and Steve Oliver
of the University of Manchester in England. The work appeared in the January
15, 2004 issue of Nature. The research was funded by the UK Biotechnology
and Biological Sciences Research Council, and the UK Engineering and Physical
Sciences Research Council.
Timeline: 3-6 years
TRN Categories: Artificial Intelligence; Robotics; Applied
Story Type: News
Related Elements: Technical paper, "Functional Genomics
Hypothesis Generation and Experimentation by a Robot Scientist," Nature,
January 15, 2004
January 28/February 4, 2004
Robot automates science
Chemicals map nanowire
Badge controls displays
Inkjet goes 3D
tied to silicon circuit
have ice grip
Research News Roundup
Research Watch blog
View from the High Ground Q&A
How It Works
News | Blog
Buy an ad link