Melbourne
THE E-MAIL messages arrive from all over the world鈥攖he US, Korea,
Germany, Mexico鈥攁ll asking for help from Vladimir Brusic. A typical
request came from Milan: 鈥淲e are cloning cell lines from ovarian cancers,鈥� the
message said. 鈥淭he cells produce a characteristic protein, and we want to find a
fragment of that protein [a peptide] that will trigger the immune system. Can
you provide us with five candidate peptides we can try? We can afford neither
the time nor the money to make and test any more.鈥� Brusic fed the details into
his computer, and came up with several potentially useful peptides to test. 鈥淚
have just had word,鈥� he says. 鈥淲e were not unsuccessful.鈥�
Brusic has saved the Italian lab possibly hundreds of hours of experimental
time, as he has many others. His success relies on the careful application of
the power of computing鈥攂y allowing the complexity of the biology to remain
intact, rather than attempting to simplify it. His work may have lessons for
many other areas of biological and medical science.
Advertisement
Brusic is a leading light in the field of bioinformatics鈥攗sing
computers to help molecular biologists deal with their complex data. 鈥淧roblems
in biology are several orders of magnitude more complex than anything in
engineering,鈥� says Brusic from his base at the Walter and Eliza Hall Institute
(WEHI) of Medical Research in Melbourne, Australia. In cellular immunology, for
instance, there are about 100 billion peptide sequences to which the immune
system can respond, each targeted by a small set of white blood cells, or T
lymphocytes鈥攁s many types of T cells in one human being as there are stars
in our Galaxy.
Brusic helps to locate the crucial parts of proteins that trigger the immune
system. 鈥淲e want to be able to take a protein and find the peptide sequences
which are of immunological interest,鈥� he says.
Such information is crucial in the fight against disease, especially in
vaccine development. The work may also help to find which fragments of proteins
trigger the immune system to attack the body, as in autoimmune diseases such as
diabetes and rheumatoid arthritis. There is also the question of why some
cancers are able to avoid destruction by the immune system.
But picking out the important peptide sequence in a protein is like trying to
find a needle in a haystack. Protein molecules are long chains of hundreds of
amino acid units鈥攁nd any of the 20 different types of amino acid might be
found in any position on the chain. The sequences that trigger the immune system
are small peptide fragments鈥攂etween eight and 40 amino acids long. In the
past, researchers chopped the chain up into hundreds of different peptide
fragments and tested each one to see if it would trigger an immune response.
Brusic is trying to cut this workload鈥攁nd dramatically. He uses a
combination of his own programmes and commercially available neural network
software to search along a protein sequence and pick out the peptide fragments
most likely to bind with key immune system molecules. He claims his system
allows researchers to eliminate more than 80 per cent of the possible peptides,
letting them concentrate on the promising 20 per cent. Better still, says
Brusic, the system should improve with time. Neural networks 鈥渓earn鈥� as they go
along (see 鈥淭uning up鈥�). The more data they are fed, the more accurate and
selective they become.
Brusic is passionate about creating a successful marriage between computer
science and biology. Each discipline has its weaknesses when it comes to the
other, but Brusic thinks he can bridge the gaps. 鈥淏iologists do not understand
how to handle complex data,鈥� he says. To make best use of it, they need computer
scientists to help them. But computer scientists 鈥渄on鈥檛 have any deep
understanding of what biological data means鈥�, Brusic says. 鈥淭hey are always
trying to make assumptions about the data to simplify their task. But to
understand the biology it is important to retain its complexity.鈥�
Born in Yugoslavia, Brusic studied biomedical engineering at the Centre for
Multidisciplinary Studies at the University of Belgrade. 鈥淚t gave me a
background in cell biology and physiology, as well as engineering,鈥� he says. But
he thinks Australia is one of the few places where he could have developed his
techniques from scratch. 鈥淚t would have been difficult anywhere else,鈥� he
claims. In Europe, the hierarchical system makes it hard for young researchers
to try anything so radical. And in North America, researchers are too concerned
with achieving quick results to be bothered with the slow accumulation of data,
he says.
Surface display
Brusic was inspired to develop his peptide prediction system when working
with Len Harrison, the head of WEHI鈥檚 diabetes laboratory. Harrison studies
insulin-dependent diabetes, sometimes known as juvenile-onset diabetes. This is
an autoimmune disease in which the immune system attacks the
insulin-secreting cells of the pancreas.
A healthy immune system attacks foreign cells, viruses and compounds in the
body. The two key players in the recognition system are a group of large
glycoproteins belonging to what is called the major histocompatibility complex
(MHC), and the T lymphocytes. MHC molecules are located on the outer membranes
of cells, and recognise characteristic peptide fragments produced when foreign
proteins break down. A fragment, normally about nine amino acids long, is drawn
to the cell surface where it binds to a groove or cleft in the MHC molecule and
is held on display or 鈥減resented鈥� to the T lymphocytes.
Among the billions of different T lymphocytes, a few will have a receptor
that is complementary to the peptide fragment on display and can bind to it
(see Diagram).
Once a T lymphocyte has found its target, it multiplies rapidly in the
bloodstream鈥攁nd soon millions of T lymphocytes with the correct
complementary receptor are seeking out cells displaying the fragment. They then
kill any cell鈥攏ormally those infected with foreign invaders such as
bacteria or viruses or the bacteria and viruses themselves鈥攖o which the
protein is attached.
There are so many peptides that bind to MHC molecules, and their interactions
are so vital to immunology, that in 1993 Harrison decided to create a database
to pull together all the published information. The computer scientist assigned
to the job was Brusic. The database, called MHCPEP, was put onto the Web in 1994
and now contains information on more than 10 000 peptides known to bind to MHC
molecules. Brusic browses the literature every day for new data, and updates the
database once a year.
Around the time the database was being assembled, Harrison was trying to
develop a blood test that could identify people susceptible to diabetes long
before the standard medical symptoms emerged. The test determines whether an
enzyme called glutamic acid decarboxylase (GAD), responsible for the burning of
sugar in our bodies, is recognised and attacked by the immune system. In
insulin-dependent diabetes, peptide fragments of GAD bind to the MHC molecule
DR4鈥攂ut for his test, Harrison needed to know which fragments were
involved.
Harrison鈥檚 group chopped up the entire GAD protein of about 300 amino acids
into fragments only 15 amino acids long. Each fragment overlapped its neighbours
by four amino acids. The researchers then tested each of the fragments with the
MHC molecule, and those that bound most strongly were selected for further
trials with the diabetes test.
Not surprisingly during this laborious task, Harrison wondered if a computer
might help recognise any common pattern among those peptides that bound well to
the MHC molecule. 鈥淚t occurred to me that perhaps we could even develop a method
of prediction which could scan a protein for DR4-binding peptide sequences,鈥� he
says. So Brusic set to work with George Rudy, one of Harrison鈥檚 colleagues,
training a neural network with the GAD binding data, and other data from
MHCPEP.
Other computer-based techniques were being developed elsewhere in the world,
but none seems to match the power of Brusic鈥檚 approach. The earliest attempts
stemmed from the work of researchers who noticed that peptides that bound to a
particular MHC molecule shared common patterns of amino acids at specific
positions, or 鈥渕otifs鈥�. In certain positions in strongly binding peptide
fragments, particular amino acids are very likely, whereas others never crop up
at all. It was suspected these were key anchor points鈥攖he positions where
the fragment would bind particularly strongly to the MHC molecule.
Prominent among these researchers was Hans-Georg Rammensee, now at the
University of T眉bingen in Germany, who developed experimental methods to
determine motifs on a massive scale. After he and others started publishing
these motifs, computer software was developed by many groups to search protein
sequences for motifs. There was some success, but less than half the peptides
selected by the software actually bound to the particular MHC molecule. And many
other peptides that had no motif were found to bind quite well.
Brusic says there is a problem in relying too heavily on motifs. 鈥淚t can bias
research,鈥� he says. Not only could researchers miss a key peptide, but
restricting the search only to those peptides with motifs means the data isn鈥檛
representative of the whole gamut of peptides鈥攖he entire data set will be
biased towards peptides containing motifs.
To get round this problem, Ken Parker of the US National Institutes of Health
in Bethesda, Maryland, tried a second method. This attempts to determine the
effect that each individual amino acid has on the binding strength of a peptide
fragment to a particular MHC molecule. The effect is calculated in each of the
nine positions in a typical peptide sequence, and added up to give an overall
measure of binding strength for the whole fragment. A computer programme then
picks out the fragments that should have the highest binding strengths.
Although the results are generally better using this method than using
computer searches for motifs, the predictions are still only correct in up to
two-thirds of cases when tested. And Brusic thinks this may be the limit using
this approach. The theory assumes that each amino acid in the fragment binds
independently of the others. But this is not the case, says Brusic, because they
interact and can affect one another. The initial assumption is too
simplistic.
Brusic thinks his system gives more accurate predictions because it does not
assume that it is looking for anything in particular, such as a motif or a
strong individual amino acid binding site. 鈥淢ethods which do not make
assumptions about the data will eventually come out on top,鈥� he says. 鈥淚f you
make assumptions, you need to be very sure you are correct, or you will end up
biasing your information.鈥�
Open-minded
Brusic鈥檚 neural network doesn鈥檛 make any assumptions. When the sequence of
amino acids in an unknown peptide fragment is fed in, the network gives a
measure of the fragment鈥檚 binding strength to a MHC molecule based on what is
has already learnt from previous peptides and their binding strengths.
Gary Stormo of the University of Colorado, who edits the journal Computer
Applications in the Biological Sciences, is impressed by the technique.
鈥淏rusic has taken a fairly novel approach which looks reasonably promising,鈥� he
says. 鈥淚t should save a lot of time-consuming and expensive experiments.鈥�
One of the major benefits of the system is that it can learn from any
researcher鈥檚 data鈥攊t is not picky about how the measurements of binding
strengths are made, or even if a few are simply wrong. Thousands of results are
effectively 鈥渁veraged together鈥� by the neural network, so stray results have
very little impact on the overall accuracy of the system. Biologists develop
their own methods for measuring binding strengths, says Brusic. 鈥淓very method is
biased鈥攖he measurements are skewed [with] too much of one thing, too
little of something else.鈥� But ask anyone whose method is best, says Brusic, and
they鈥檒l all tell you it鈥檚 theirs.
To combine the data into one database, Brusic and Rudy use their own 鈥渇uzzy鈥�
measures of binding strength based on all the published data about a particular
peptide fragment and MHC molecule. The researchers sort results into high
strength, moderate strength, low strength and non-binding groups, each covering
a broad range of values. These broad categories are then assigned standard
average values for the computer to work with. Although Brusic鈥檚 software can鈥檛
give pinpoint predictions of binding strength, it does select peptides that bind
with more than a certain threshold strength and so might be immunologically
interesting.
Rammensee鈥檚 main concern is about Brusic鈥檚 method of combining data from
different sources. In his own work, he says, he only trusts data he has
鈥渢ouched鈥�, and is not convinced that what he would consider bad laboratory
technique can be overcome by slick data handling.
In early trials, Harrison, Rudy and Brusic tested their system with
combinations of peptides and MHC molecules for which they already knew the
binding strengths鈥攖he data from the GAD work plus other data from MHCPEP.
They found they could achieve prediction rates of about 80 per cent or better.
But to be sure the system was really useful, they needed to predict some
experimental results in advance.
Power of prediction
As they already had data from the GAD experiments on some of the peptide
sequences that would bind to the MHC molecule DR4, the group decided to predict
which peptide sequences from the enzyme tyrosine phosphatase鈥攁nother of
the immune system鈥檚 targets in insulin-dependent diabetes鈥攚ould bind to
DR4. Laboratory binding studies confirmed that the neural network correctly
predicted all the strongly binding fragments, with only a few false positives.
For moderate binding fragments, the prediction was 84 per cent correct. Brusic,
Rudy and Harrison were delighted.
Another researcher happy with Brusic鈥檚 results is Venkatesh Ramakrishna at
Italy鈥檚 National Cancer Institute in Milan. He has been collaborating with
Brusic for two years. 鈥淚t is very productive and cost effective,鈥� he says. 鈥淚t
can reduce the number of experiments we have to do from a thousand to ten
critical ones鈥攖hat鈥檚 the power of prediction.鈥� Soren Brunak of the
Technical University of Denmark in Lyngby agrees that the performance of
Brusic鈥檚 system is good, but adds a note of caution. 鈥淭his type of prediction is
a data-driven method,鈥� he says. 鈥淚t is limited by the amount of data available.
At present, data are pouring out of machines and Brusic is very good at
gathering them.鈥�
As long as these data for training and refining his network keep coming,
Brusic should find himself increasingly in demand. For the foreseeable future,
his database seems set to continue expanding. He is discussing collaborations
with both Stormo and Brunak and there is a growing tide of people knocking at
his door, hoping he can locate that elusive peptide fragment. Brusic鈥檚 marriage
between computer science biology looks set for a long and productive
honeymoon.
* * *
Tuning up
NEURAL networks are learning systems. They pick up on patterns in data they
are fed, and use them to predict results of similar experiments. The networks
consist of a set of points or 鈥渘odes鈥� to which are connected lots of inputs and
outputs. In Brusic鈥檚 system, the inputs are the positions of amino acids in
peptide fragments and the MHC molecule to which they bind. The output is the
measure of the binding strength.
The job of the nodes is to make the connections between the inputs and the
output, and to remember and learn from the patterns of connections. The system
is set up using 鈥渢raining data鈥� for which both inputs and output are
known鈥攖he bigger the set of training data used, the more accurate the
system becomes. Mathematical representations of peptide sequences are fed in as
inputs, and the system is also given the outputs鈥攖he binding
strengths.
The network is adjusted, or 鈥渢uned鈥�, by trial and error until the system
matches up the inputs to the correct output. This is carried out with the
training data over and over again, the network being fine-tuned at each stage
until the system has 鈥渓earnt鈥� which inputs, or patterns of inputs, give certain
outputs. The network can then be used to predict an unknown output from a set of
inputs. For Brusic, it predicts binding strength from a particular peptide
sequence and MHC molecule.