快猫短视频

Richard Trethewey thought it would be easy to make a starchier potato. In
this age of genetic engineering, he and his colleagues at the Max Planck
Institute of Molecular Plant Physiology in Golm, Germany, figured they just
needed to insert the right gene into potato plants. And because they knew all
the steps in the biochemical pathway that plant cells use to produce starch,
they thought they knew exactly which gene ought to do the trick. Breaking down
sucrose is the first step in starch production, so the researchers inserted a
yeast gene that codes for invertase, an enzyme that cleaves sucrose.

To their surprise, the engineered potatoes turned out to have a sixth less
starch than before. The invertase, they eventually discovered, had shifted the
plants鈥� biochemistry towards producing glucose, not starch. To correct the
problem, they inserted a second gene, a bacterial glucokinase that would funnel
the excess glucose into the starch pathway. But once again, the plants鈥�
biochemistry shifted in an unexpected direction, and starch yield dropped by
another sixth.

Trethewey鈥檚 experience is not unusual. Many biologists have been frustrated
by their inability to control the behaviour of cells and organisms by tinkering
with one or two genes. 鈥淥ne measure that we don鈥檛 understand biology very well
is that our understanding is not predictive,鈥� says Roger Brent, associate
director of the Molecular Sciences Institute in Berkeley, California. 鈥淵ou can鈥檛
look at the genome of an organism and even tell how many legs it has.鈥�

The problem is that cells are not just collections of independent biochemical
pathways, as the single-gene tweakers implicitly assume. Instead, biochemical
pathways frequently criss-cross, branching and merging to form complex networks.
As Trethewey learnt, ignoring that wider context can lead researchers seriously
astray.

This is why many biologists are now trying to find ways to study how cells鈥�
many components work together, rather than just studying the individual
components. The new approach is being applied to everything from small
biochemical circuits to complicated diseases, and already it鈥檚 yielding
dividends. Some researchers are trying to use the principles of complexity
theory to understand how genes interact鈥攁n approach that promises to solve
long-standing mysteries such as why some cancer cells remain malignant even
after their cancer-causing mutations are repaired.

In the US, leading universities鈥攊ncluding Harvard, Princeton, Caltech
and Johns Hopkins鈥攁re spending tens of millions of dollars setting up
institutes and departments to probe the complex workings of cells. And biotech
companies betting on the success of the new approach are springing up far and
wide. 鈥淭his is the wave of the future,鈥� says Robert Dinerstein, a cell biologist
at the drugs company Hoechst Marion Rousell in Bridgewater, New Jersey.

Simple sketches

Of course, the old approach has had its share of successes鈥攖he
identification of the AIDS virus and the ability to produce large amounts of
insulin and growth hormone, to name just two. But researchers who are willing to
grapple with more complexity are looking for even bigger gains.

Even advocates of the new approach rarely try to tackle the full complexity
of cells, however. Most opt instead for models that are more complex than usual,
but still simple enough to be workable. Building such a model is like painting a
picture, says John Reinitz of the Mount Sinai School of Medicine in New
York鈥攊f it鈥檚 well rendered, even a few simple lines can capture the
essence of a complicated scene.

And sketching those lines might be easier than we think, says Roland Somogyi,
director of neuroscience at Incyte Pharmaceuticals in Palo Alto, California. If
biological networks were vastly interconnected and followed a complicated set of
random rules, you would have to understand each and every detail before
attempting to make any predictions. 鈥淵ou鈥檇 have yourself a big mess,鈥� says
Somogyi. But each gene is usually connected to just a few other genes, which
interact via relatively simple rules. That means you can probably predict a lot
about how a cell will work based on a manageable number of measurements, says
Somogyi.

One of the tricks, of course, is to identify the critical brush strokes. But
you don鈥檛 have to identify them all before you can create a computer model: the
models themselves can sometimes lead the way. For example, Leslie Loew and his
co-workers at the University of Connecticut Health Center in Farmington wanted
to simulate the way the activation of a receptor on the surface of a nerve cell
unleashes a cascade of signals within the cell. At first, their
simulation鈥攚hich used a software package they designed known as the
Virtual Cell鈥攂ehaved nothing like a real neuron. But eventually, the
researchers discovered that if they made their model cell鈥檚 calcium storage
compartments clumped instead of evenly spread, their results suddenly looked
right.

Returning to their microscopes, the scientists looked at a real neuron鈥檚
storage compartments. Sure enough, the body of the cell harboured a large
calcium-storage site, while its branches housed much smaller compartments. The
Virtual Cell had revealed a critical parameter鈥攖he positioning of calcium
storage sites鈥攖hat might have otherwise gone unnoticed. 鈥淚t鈥檚 when the
models fail that we learn the most,鈥� says Loew.

Sometimes, models can even reveal fundamental principles that no one had
thought to look for before. For example, most biologists have assumed that to
keep biochemical systems running smoothly, the activity of every component has
to be precisely tuned. Until recently, hardly anyone had considered questioning
this assumption. But then Stanislas Leibler and his colleagues at Princeton
University started modelling chemotaxis, the process by which certain bacteria
detect food and swim towards it. Chemotaxis begins when an attractant binds to a
receptor protein on a bacterium鈥檚 surface
(快猫短视频, 15 August 1998, p 40).
The receptor is linked to a network of proteins inside the
bacterium, the Che proteins, which control the rotation of the flagella that
propel the bacterium. The result is that the bacterium spends less time tumbling
randomly and more time swimming towards the food source. Eventually, it gets
used to the presence of the attractant and resumes its normal tumbling
behaviour.

Leibler鈥檚 group soon realised that theoretically, there were many ways they
could hook up the Che proteins to simulate chemotaxis. Most allowed bacteria to
resume their normal tumbling rate only if each Che protein was present in
exactly the right concentration. But a few models worked even if the
concentrations varied widely. Since real life is messy and ever-changing, the
researchers began to wonder if nature might have chosen one of these robust
designs.

To answer that question, the researchers genetically engineered bacteria that
made too much or too little of some Che proteins. They found that altered
bacteria returned to normal tumbling rates even when producing 50 times as much
of one Che protein as normal (Nature, vol 397, p 168). 鈥淚t was more
like something you鈥檇 expect a human engineer to design, rather than a generic,
random network,鈥� says lead author Uri Alon, now at the Weizmann Institute of
Science in Israel. Alon鈥檚 work helped to reveal a concept new to most cell
biologists鈥� thinking鈥攖hat the robustness of signalling networks may be an
important trait selected for during evolution.

Success stories such as these are starting to convince researchers from
various walks of biology to add complex models to their toolboxes. The biotech
company Entelos in Menlo Park, California, for example, is developing software
packages to model diseases as diverse as asthma, obesity and AIDS. During its
first nine months of testing, the asthma model has already provided several
insights into drug development. In one case, it predicted that an
anti-inflammatory drug would go through a lag period, perhaps allowing one or
two asthma attacks to occur before kicking in to repress them. Without this
knowledge, researchers might have dismissed the drug as ineffective early on in
clinical trials.

As satisfying as these efforts are, however, they still fall far short of the
complexity of real cells. Researchers have almost finished sequencing the entire
human genome, with its estimated 100 000 genes. The race is on to learn the
functions of the tens of thousands of proteins produced by those genes. That
demands a radically different approach. 鈥淚t鈥檚 something that has never been
taken on in any arena of science before,鈥� says Stuart Kauffman, a pioneer in the
field of complexity, who is now chief scientific officer at Bios Group in Santa
Fe, New Mexico. 鈥淭here鈥檚 no place in science where we鈥檝e ever understood a
system with 100 000 different components regulating themselves in such a
heterogeneous way.鈥�

Complexity tames itself

Thirty years ago, Kauffman showed that a network connected by simple rules
could generate orderly behaviours that are impossible to predict from studying
its individual parts. His work was a striking example of how biologists could be
missing the wood for the trees.

Over the years, he and his colleagues have tinkered with their computer
models to mimic biological reality more accurately. Genes can be thought of as
being either off or on: inactive genes simply sit in the nucleus and have no
effect on the cell, while active genes are 鈥渞ead鈥� and used to make proteins.
Active genes can turn many other genes on or off, including themselves. So with
tens of thousands of genes, patterns of activation could be incredibly complex
and varied.

What Kauffman and his colleagues discovered from their models, however, was
that such systems often tend to settle into just a few stable states in which
most genes remain either on or off, and only a few small islands of activity
remain. They suspected that the same might be true of the genes within real
cells鈥攁nd two years ago they turned up evidence to support that guess.

In a sample of real genes that were regulated by three other genes, for
example, 60 per cent were linked in ways that promoted such stable states. If
the linking rules had been chosen randomly, only 8 per cent of the genes would
be expected to follow such rules. The researchers still don鈥檛 know whether all
or most of the genes in a cell follow similar principles, but so far it looks as
though Kauffman is on the right track. Since real genes don鈥檛 merely turn on and
off, Kauffman鈥檚 team has also refined the models to allow graded responses, with
similar results.

When genetic systems are connected in this way, they tend to be
robust鈥攖hat is, any damage or alteration to a single gene usually does not
have much effect beyond that gene鈥檚 island of activity. At least in some cases,
this robustness seems consistent with biologists鈥� experience. For example,
genetic engineering experiments where single genes are 鈥渒nocked out鈥� often fail
to have dramatic, or even noticeable, effects. The architecture of organisms鈥�
genetic networks may contribute to this resilience.

But even in robust systems, some mutations or groups of mutations can shatter
the normal functioning of the network. This happens in cancer. Sui Huang of
Children鈥檚 Hospital in Boston, for example, thinks that the uncontrolled growth
of cancer cells may be explained as a disruption in a Kauffman-like network of
genes.

Huang views cellular states such as growth or differentiation鈥攖he
process by which a cell turns into a specialised type like a lung cell or a skin
cell鈥攁s stable states in a cellular network. To see what he means, imagine
that cells are marbles and that their states are represented by their position
on a two-dimensional surface. The most stable states will form valleys in this
landscape. Environmental conditions such as the presence of growth factors,
together with a cell鈥檚 history, will determine into which valley a cell 鈥渇alls鈥�
and thus whether it differentiates, commits suicide or divides, for
instance.

But the sizes and shapes of the valleys can change if the network鈥檚 units are
disrupted. Huang suggests that in cancer, the growth valley is broadened, so
that cells are more likely to fall into it, regardless of where they start on
the landscape (see Diagram)
.
This reflects what is seen experimentally. Cancer
cells can grow and divide under a wide variety of conditions where normal cells
cannot. And if the growth valley broadens, others must shrink because the total
landscape is fixed in size. This, too, fits with reality, since cancer cells are
less likely than normal cells to enter or remain in the differentiation
state.

Cancer paradox

Huang鈥檚 model also offers an explanation for a paradox in cancer research.
Often, the removal or inhibition of an oncogene鈥攁 mutated gene that causes
cancer鈥攄oes not cause the cell to revert to its normal, non-cancerous
state. Huang鈥檚 model explains this through the inherent stability of cell
states. The same light push, or mutation, that sends a marble rolling down into
a valley may not be enough to fling it back out. Similarly, his model helps
explain why many different mutations seem to lead to just a few types of cancer.
The same broadened growth valley can be reached by light pushes from many
different locations on the landscape.

To begin bringing molecular details into his model, Huang envisions
classifying cell states by the activity profiles of their genes. He plans to use
screening technologies such as DNA chips
(快猫短视频, 14 November 1998, p 46)
that can monitor the states of thousands of genes at once to
establish whether cellular states are, in fact, like valleys in a landscape. By
building a computer model in which each unit represents the state of a real
gene, he hopes to find out whether his ideas really apply to living cells. If
they do, scientists may someday be able to discover exactly how mutations in
particular sets of genes turn a cell malignant, and maybe even which sets of
genes to tweak to render malignant cells harmless. But Huang stresses that both
his experimental and theoretical proposals are in the earliest stages of
development. Years of hard work still lie ahead.

This type of broad-brush approach has its critics. Brent, for example, says
that attempts to build computer models of complex genetic networks must be
firmly grounded in experimental data and include molecular details from the
outset. And Harley McAdams of Stanford University, whose models of bacterial
viruses are classics in cell biology, adds: 鈥淢y experience showed me how much
modelling depends on having all the information. For most systems, there鈥檚 just
not enough data yet.鈥�

Yet while many researchers are sceptical about applying complexity theory to
genetics, only a few would still argue that molecules should be studied one or
two at a time. But a quick survey of the literature or the latest grant
proposals reveals that this is exactly what most biologists do. 鈥淲e鈥檝e got to
put our money where our mouth is,鈥� says Somogyi. 鈥淚 hear a lot of people saying
`oh, yeah, we鈥檝e got to do it,鈥� but they鈥檙e not.鈥�