TYPICAL. You program a computer, teach it everything you know about drug
design, and then it beats you to the Nobel Prize. But Ashwin Srinivasan is
hoping that, if this does happen, he might at least get an invitation to the
presentation ceremony. After all, he and his computer do their science
together.
Srinivasan鈥檚 machine at Oxford University is just one example of a silent
revolution that is taking place in labs across the world. Computers, commonly
perceived as little more than ultra-fast calculators, are suggesting new ideas
in medicine and chemistry, determining the roles of genes and proposing and
testing new mathematical theorems. They are even helping with the choice of
embryos for IVF implantation. Computers have been promoted from dumb tools to
full research partners, and people working without digital colleagues may soon
begin to fall behind. 鈥淭he future lies in human-computer collaboration,鈥�
Srinivasan says.
The first glimpse of this strange future came in 1980. Ryszard Michalski, now
director of the Machine Learning and Inference Laboratory at George Mason
University in Fairfax, Virginia, broke new ground when he published a paper
describing a computer program that learned to diagnose diseases of soybeans,
based on examples of diagnosis provided by an expert. It was a huge success:
farmers across America bought the program, and it is still in use today.
Advertisement
Michalski made an extraordinary discovery during his research. He initially
fed his program with examples of expert knowledge about soybean diseases. From
those examples, it learned how to recognise a particular disease from a plant鈥檚
symptoms. But Michalski found that the learning program produced better
diagnosis rules than those originally provided by the expert. The machine
eventually became more expert than humans at diagnosis, and it also presented
information about its decisions in a clearer way than humans ever had. 鈥淲e
realised that experts aren鈥檛 very good at explaining why they make their
decisions,鈥� he says.
Srinivasan鈥檚 silicon collaborators use a machine-learning technique known as
Inductive Logic Programming, and produce results that are just as remarkable as
Michalski鈥檚. ILP, initially developed by Stephen Muggleton, professor of
computer science at York University, enables scientists and computers to share
knowledge with ease. If it鈥檚 told something, the computer can assimilate that
knowledge into a theory. It then looks for further implications that arise from
that theory, coming up with ideas that are different from its initial input. And
because the language is transparent to humans and computers, the flesh-and-blood
scientists can see how those ideas came about. This is crucial, Muggleton points
out: after all, this is about collaboration. 鈥淵ou don鈥檛 just want an answer, you
want to know why and how the computer got there,鈥� he says.
Srinivasan and his colleagues used ILP to develop a chemical 鈥渋nhibitor鈥� that
interferes with the operation of ACE, an enzyme associated with high blood
pressure. A renowned chemist at Pfizer set the Oxford group this task to see
whether an ILP-equipped computer could match current chemical skill. The program
analysed an array of chemicals that are known to lock onto ACE molecules and
inhibit them, and tried to extract the feature that was common to them all. It
came up with three-dimensional descriptions of the type of molecule that would
inhibit the enzyme and comply with other necessary restrictions, such as low
toxicity. When the program showed its findings to the Pfizer chemist, one of the
results was the same chemical that he had chosen. But another of the
possibilities was something that neither he, nor other chemists, had ever
thought of. It is still under investigation, but it seems that the Oxford
program may have identified a new way of inhibiting ACE.
Having proved ILP鈥檚 worth, Srinivasan and Muggleton, working with Ross King
from the University of Wales and David Page of the University of Wisconsin, have
used the same process to propose entirely new pharmaceuticals, unexplored by the
human mind. Their program has come up with alternative medicines for treating
many conditions, including Alzheimer鈥檚 disease. The drugs have still to be
tested, but they point to the future of chemical design. 鈥淐hemists will find
their digital assistants to be an inseparable part of their creative work,鈥�
Srinivasan says.
Srinivasan has also used a learning program to improve embryo selection for
IVF implantation. By analysing the characteristics of embryos that successfully
implanted in the womb in the past, the computer has given fertility experts
fresh insights. Because there are so many variables involved, the research would
have proved extremely difficult for anything but a silicon brain.
Another program, a silicon scientist called MECHEM, is making similar
strides. Its creator, Raul Valdes-Perez, a computer scientist at Carnegie-Mellon
University, has applied MECHEM鈥檚 skills to problems in chemistry, particle
physics, cell biology and linguistics, and published its findings in
distinguished journals in each of these fields.
As with the other programs, MECHEM makes its discoveries with the help of a
human researcher. In chemistry, for example, it collaborates with Andrew
Zeigarnik of the Lomonosov Academy of Fine Chemical Technology, Moscow.
Zeigarnik might give MECHEM the chemical properties of carbon monoxide, hydrogen
and some catalyst, in the same way that Srinivasan describes drugs he wants to
design to his computer. Using some pre-programmed knowledge of how chemical
reactions work, MECHEM then works out the products of their interaction at
various temperatures, taking into account all the energy demands involved. If it
suggests reactions that Zeigarnik knows to be impossible at certain
temperatures, he will add that 鈥渞eal world鈥� knowledge into the program. The
process continues until a chemist could have no objections to the program鈥檚
conclusions.
MECHEM has revealed that the number of possible reactions consistent with our
current chemical knowledge is far larger than people had ever thought. 鈥淚t
generates many potential reactions that few, if any, chemists would normally
think of,鈥� Zeigarnik says. 鈥淭his means we get a much better understanding of the
reaction mechanism as a whole.鈥�
But MECHEM has never received any official credit. 鈥淚 would say that MECHEM
should have been the first author on all of the papers it contributed to,鈥� says
Zeigarnik. But, as its friends point out, the scientific community isn鈥檛 ready
to accept a machine as a co-author. 鈥淲hat we鈥檙e doing is pretty outlandish to
most people鈥攚e use a computer to perform the tasks most chemists do,鈥�
Valdes-Perez says. 鈥淲e don鈥檛 want to rub their noses in it by citing MECHEM as
an author as well.鈥�
Simon Colton, a computer scientist at the University of Edinburgh, has no
such qualms about acknowledging a machine as his co-author. 鈥淲henever my
programs invent interesting things I always credit them,鈥� he says. His digital
mathematician (known as HR, after the early-20th-century, flesh-and-blood
mathematicians Godfrey Harold Hardy and Srinivasa Ramanujan) has produced 20 new
number sequences, which have all been accepted into The Encyclopedia of
Integer Sequences, the world鈥檚 biggest repository for interesting
sequences. 鈥淚 credited HR with every one to make it clear I didn鈥檛 invent it,鈥�
he says.
HR begins a new mathematical research project by examining a set of numbers,
for example, in the context of some rules, such as multiplication or division.
Then it starts inventing new 鈥渃oncepts鈥� by combining these rules in novel ways.
An example of a concept might be 鈥渘umbers that only divide by one and
themselves鈥�, in other words, prime numbers.
While all this is going on, HR also looks for connections between the
different concepts it has invented. If it finds a connection between two
seemingly unrelated concepts鈥攕ay between prime numbers and square
numbers鈥攖hen it is 鈥渟urprised鈥� and suggests this connection as an
interesting new theorem.
HR plugs this theorem into another program, called Otter, which tries to
prove that it is true. Only when HR has come up with something interesting and
provable does Colton step in to do the machine鈥檚 dirty work, sending the idea
for publication.
鈥淭he first sequence that Simon Colton sent in鈥攏umbers such that the
number of divisors divides the number鈥攊s quite cute,鈥� says Neil Sloane,
editor of the number sequence encyclopedia at AT&T鈥檚 Shannon Labs in New
Jersey. 鈥淚 had certainly not seen it before in 30 years of collecting sequences,
and the fact that it was found by a computer program makes it especially
颈苍迟别谤别蝉迟颈苍驳.鈥�
With HR鈥檚 success firmly established, Colton is now allowing multiple copies
of the program to collaborate. He runs four copies of HR (he named them
Ramanujan, Littlewood, Wright and Hardy). Each has a distinct 鈥減ersonality鈥� and
goes about the work in a different way. Every time one arrives at an interesting
conclusion it tells its peers, who react according to their personality.
Ramanujan has a short attention span and will fly off on a new tangent every
time he sees a new idea. Littlewood, on the other hand, is pickier and will only
explore ideas that the others suggest if these are novel and interesting. Wright
is a pedantic old soul who tediously examines all his own ideas to the full
before considering any ideas proposed by the others. Hardy prefers to explore
different shades of the same idea before combining it with other ideas. Colton
found that communication is the key to their creativity. 鈥淚n fact, only if they
communicate all the time do they get better results than working on their own,鈥�
he says.
Colton says he is not trying to replace mathematicians, he is simply creating
new ones. It鈥檚 worth doing, he believes, because there鈥檚 a vast amount of work
waiting to be explored. 鈥淭he amount of data in science is becoming so huge that
all scientists, including mathematicians, could do with the help computers can
give,鈥� he says.
It鈥檚 a sentiment shared by Don Swanson, a researcher at the University of
Chicago. In 1988, he stumbled across an indirect link in the scientific
literature between Raynaud鈥檚 disease and dietary fish oil. In Raynaud鈥檚 disease,
the blood supply to the fingers is faulty, leading to attacks of numbness and
discomfort. The oils, it seemed, might be beneficial to Raynaud鈥檚 sufferers. 鈥淚
then searched for, and could not find, any direct connection,鈥� Swanson says. But
two years later, an independent medical research group confirmed the link.
Swanson realised that there might be other vital medical connections hidden
in the undergrowth of literature, and created a program called Arrowsmith that
can trawl through the medical data in the Medline archive held at the US
National Library of Medicine in Bethesda, Maryland. It loves jobs that no human
would ever get round to鈥攐r find time to finish. But they鈥檙e worth doing:
the program has suggested whole new directions for medical research.
Arrowsmith looks at keywords in the papers鈥� titles and abstracts, and
attempts to make links between them. Its most successful piece of work to date
is the link it discovered between magnesium deficiency and migraines. Very few
database records contain both 鈥渕agnesium鈥� and 鈥渕igraine鈥�. But when Arrowsmith
found other keywords in the abstracts of papers on both subjects, it soon made
the connection.
Magnesium has well-documented effects on at least 11 physiological factors
that are separately known to be involved with migraines. After the discovery was
published, at least 12 laboratories independently found direct clinical or
laboratory evidence to support the result.
This kind of research, which would keep a scientist up all night working on a
hunch, will become central to scientific progress. Arrowsmith isn鈥檛 limited to
medicine: it can burn the midnight oil and come up with new insights wherever
there is a mass of data to explore. Srinivasan is excited about the future of
such human-computer collaboration in science. 鈥淚t allows a greater degree of
creativity all round because together we examine things that we would never
bother to look at without the help of a computer,鈥� he says.
To work alongside a silicon scientist who can unearth hidden facts, propose
interesting concepts for further investigation, and even carry out experiments
(see 鈥淭he robot scientist鈥�) is a researcher鈥檚 dream. Srinivasan has a hopeful
glint in his eye. 鈥淥ne day,鈥� he predicts, 鈥渁 man-machine team could share the
Nobel Prize.鈥�
The world鈥檚 first robotic microbiologist will make its laboratory debut in
Britain this month. Based at the University of Wales in Aberystwyth, the robot
will perform experiments designed by a learning program called ASE-Progol, which
runs on a computer in Stephen Muggleton鈥檚 University of York laboratory.
These collaborating machines will initially help to determine the
roles of genes in yeast cells. 鈥淲e are developing a robot that can use a
pipette, move fluids around and grow yeast under different growth conditions,鈥�
says Ross King, who is part of the development team.
The experiments will involve 6000 cultures of mutant brewer鈥檚 yeast, each of
which has had one of its genes deleted. Drawing on its knowledge of the yeast
culture, including descriptions of known metabolic pathways and the probable
function of known genes, ASE-Progol makes conjectures about what kind of protein
the deleted gene is responsible for producing. It then devises a set of
experiments that will test each of the conjectures.
The program sends an e-mail to the robot, telling it what experiments to do.
Each mutant will be deprived of particular nutrients to find out where the
missing protein fits into known metabolic chains: by comparing the mutant
specimen鈥檚 growth rate with a wild, non-mutant strain of yeast, the researchers
hope their machines will decipher the function of the missing gene.
The robot e-mails the results back to ASE-Progol, which analyses them and
discards the conjectures the experiments have shown to be false. Armed with this
new knowledge it then suggests a new set of conjectures and invents new
experiments for the robot to perform. 鈥淭he process continues in this loop until
either there are no consistent conjectures or it finds one conjecture that it
cannot disprove鈥攚hich is therefore the function of that sequence,鈥�
Muggleton explains.