ÐÓ°ÉÔ­´´

You are walking round the brain. Nerve cells tower over you on all sides,
pulsating in hues of blue, green, and yellow with occasional flashes of
red. Behind you, you can hear the pop-pop-popping of the cluster of cells
you just passed. They are sending their rhythmic electrical message to some
far-off region of the brain to be turned into memories, perhaps, or to generate
a thought. You point to a nerve branchlet that begins off to your right,
running across your field of view to twist around the nerve cell directly
ahead. ‘What happens if I move this over here?’ you ask, pointing one cell
further to your left. Immediately, the tendril jumps over to its new position,
and the pop-pop-popping behind you slows down.

Welcome to one of the newest – and certainly the flashiest – facets
of neuroscience. Thanks to the technology of virtual reality, which creates
3D worlds within a computer and displays them for users to explore, researchers
can now wander through simulations that mimic microscopic fragments of the
brain. These simulations let researchers develop a feel for the intricate
workings of realistic nerve networks. Indeed, with computer models, researchers
can now explore details too fine and networks too expansive to observe
directly in live brains. At its best, this simulation approach – known as
computational neuroscience – dances an intimate pas de deux with more conventional
experiments that record electrical activity in the brains of real organisms.
In this new ballet, one partner leads, then another, and occasionally each
lifts the other to new heights.

Mysterious machine

Such a partnership, many neuroscientists feel, may be the only way to
make sense of what happens in the brain. ‘The basic point, which I believe
more and more, is that we really don’t know what kind of machine the brain
is,’ says James Bower, a neurobiologist at Caltech in Pasadena, California.
‘If you believe, as I do, that the nervous system is immaculately engineered
– that there’s a tight association between its structure and its function
– then the closer you stick to the anatomy and physiology when building
models, the more likely the system will tell you something you didn’t know.’
These serendipities, which turn up frequently in Bower’s work, may provide
the key to understanding how the brain works, he believes.

But the nervous system’s tremendous complexity means that building a
realistic model of even a small part of the brain is a daunting task. Bower,
for example, studies a brain region called the piriform cortex, which in
humans is located on the underside of the brain, just above and behind
the nose. Neurobiologists think this region’s job is to recognise odours,
synthesising the raw data from the olfactory cells into outputs that say,
in effect, ‘banana’ or ‘brown ale’. Yet even this apparently simple task
– which seems a pushover compared to proving a geometric theorem or understanding
a Beethoven string quartet – involves about 6 million neurons, each one
receiving perhaps 10 000 inputs from its mates.

Even within a single cell, the details mount up rapidly. Each neuron
receives input from thousands of other neurons through connections called
synapses. When a signal arrives at a synapse, it causes a tiny electrical
impulse in the receiving cell. All the impulses the cell receives travel
down the intricate branchings of the neuron like ripples in a canal system,
summing together whenever they meet. If the resulting ripple is big enough
at a certain critical point in the cell near the confluence of all the canals,
the neuron fires, sending a powerful electrical pulse. This pulse, or spike,
generates tiny impulses in each of the thousands of cells the neuron contacts
with its output synapses.

This complex tangle of thousands of tiny inputs on tiny nerve branchlets
means that researchers can’t come close to building a fully detailed model
of electrical activity in even a simple system like the piriform cortex,
Bower says. Computational neuroscientists find themselves starved for computing
power, always waiting for the next development in computer design. But these
improvements are never enough. ‘Every new machine, every new increase in
speed, the first thing we do is move to increase the complexity of the models,’
says Bower.

Bower and his colleagues are already buried in an avalanche of data.
Their current simulation of the piriform cortex, for example, keeps track
of electrical signals at five points on each of hundreds of neurons and
generates gigabytes of data in simulating barely two minutes of electrical
activity in response to simulated synaptic inputs. The next-generation model,
which he and his students are building on Caltech’s most powerful parallel-processing
computer, will include 1000 points on each of 5000 neurons and churn out
terabytes of data. Each of these simulated neurons fires according to its
own inputs, and the ensemble can interact to produce larger-scale patterns
such as bursts, in which many neurons fire together for a period of time,
or oscillations, in which neurons fire alternately in rhythmic sequences.
How can researchers sift through the billions or trillions of bytes of data
on electrical activity to pick out the meaningful patterns?

The answer, Bower says, is to turn to the most sophisticated pattern-recognition
system on the market: the human brain. But to work at its best, the brain
demands data in 3D form. ‘The brain is a tremendously sophisticated device
for moving through 3D space. It’s much less powerful in two dimensions,’
says Bower. From infancy, all of us practise constantly at making sense
of spatial data, he notes. Even as adults, we might puzzle over a map or
a photograph, but we understand even complex 3D scenes immediately. And
we localise sounds in three dimensions, as well, combining them with visual
information to form a strong sense of our surroundings. To capitalise on
this strength, therefore, Bower and his colleagues are increasingly using
virtual reality to display their data in 3D form.

Instead of data tables and graphs, Bower’s team organise their data
in a 3D graphical image of the cells in their model of the piriform cortex,
using software developed with researchers in the Scientific Visualization
Laboratory at the University of Illinois at Chicago. To view the data, a
researcher sits in front of a video screen wearing a high-tech version of
the familiar 3D movie glasses of the 1950s. The glasses electronically black
out one eye and then the other, many times per second, while the screen
alternates between two slightly offset views of the object. In effect, each
eye sees from its own angle, and the brain assembles a 3D image.

In this display, the neurons in the image indicate their electrical
state by their colour – blue for inactive regions, and moving down the
spectrum to green, yellow, orange, and red as the cell gets closer and closer
to firing. When a cell fires, it produces a popping noise.

The researcher can also navigate around the model by means of a hand-held
pointer that resembles the familiar ‘mouse’ attached to most personal computers.
This mouse, however, moves in 3D – and is called, logically enough, a ‘bat.’
‘Not only is it a mouse that flies, but it works by emitting sonic signals.
So it really is a bat,’ Bower says. Using the bat to enter the model, the
researcher can look around, zoom in on regions of interest, and point to
a part of a cell to monitor its electrical activity. He or she can make
individual synaptic connections stronger or weaker, or move connections,
and observe the effect on the neurons’ colour and firing frequency.

‘In 3D, the force of what you’re looking at is much stronger,’ says
Bower. ‘It’s almost an emotional impact. Even though you know perfectly
well intellectually that the image is sitting there on the screen, if you
swing (the image) around, people back up so as not to get hit by it. That
suggests that the information is being interpreted in a much more powerful
way. You are engaged with it, not just interpreting it.’

Sliced neurons

Flashy and informative as such models are, however, they are only as
accurate as the anatomical data used to build them. And it’s no easy matter
to build an accurate, detailed picture of the 3D structure of even a single
neuron – much less hundreds or thousands, complete with their countless
synaptic connections. Indeed, the key features that mark a synapse are so
small they appear only under the electron microscope.

Electron microscopy requires very thin slices of tissue – ideally about
one ten-thousandth of a millimetre thick. To reconstruct a 3D nerve cell,
therefore, micro-scopists must examine and keep track of several hundred
slices, stacking images from successive slices much as one could recon-struct
a sliced head of broccoli by stacking up the slices in sequence. But such
extraordinarily thin slices distort easily during cutting and under the
heat of the microscope’s electron beam, so the cell outlines in adjacent
slices often don’t match perfectly, leaving researchers to guess about some
of the fine details.

An alternative approach pioneered by Mark Ellisman of the University
of California at San Diego uses much thicker slices to reduce both the distortion
and the bookkeeping involved in reconstructing nerve cells. Ellisman’s technique
is essentially a microscopic version of medicine’s computerised tomography
– the now familiar ‘CAT scan’, which uses a moving camera to generate a
3D X-ray image of a patient’s body. Using a similar approach, Ellisman looks
at ‘thick’ slices (actually only two micrometres thick) from several angles.
A computer calculates the 3D structure of this thick section. ‘Then we link
those together like a stack of books, instead of a stack of pages from a
book,’ says Ellisman.

Because researchers need to process fewer slices, Ellisman’s technique
is much faster than the standard method: less than a week to reconstruct
a large neuron, he estimates, as opposed to perhaps two months using thin
sections. In addition, as any veteran salami-slicer knows, thick slices
are easier to cut without distortion than thin ones.

The main drawback of the technique, as Ellisman himself admits, is
that only his laboratory and three others in the US have electron microscopes
powerful enough to punch through such thick slices of tissue. Moreover,
computing the 3D position of every structure visible under the microscope
requires a huge amount of number-crunching – for each slice, typically about
six hours on a powerful computer, Ellisman says.

Ellisman and his coworkers are now working to provide remote access
to the UCSD microscope facility so that more researchers can use their
technique. Eventually, they hope, researchers anywhere in the world will
be able to ship specimens to San Diego, where technicians will load them
into the microscope. Then, using the Internet, the global computer network,
the far-off researcher could view and manipulate the image on their own
computer terminal. Ellisman believes a prototype system should be in operation
within 12 to 18 months.

Using anatomical reconstructions created by colleagues in Israel, Bower’s
lab at Caltech has modeled the workings of a single Purkinje cell, a giant
type of neuron found in the cerebellum of the brain whose function is still
not understood, although most researchers believe it plays a role in motor
control. The most distinctive feature of Purkinje cells is a large, treelike
projection whose int-ricate branches collect synaptic inputs from as many
as 150 000 other neurons and conduct the resulting small electrical signals
down to the cell body, located at the base of the ‘tree’s’ trunk. Bower’s
model represents this projection in a somewhat simplified form, but using
anatomically realistic branching patterns and branchlet diameters derived
from microscopy.

Using this reconstruction as a starting point, Bower and his collaborator
Erik De Schutter then built a computer simulation of a Purkinje cell to
study how it integrates its vast array of inputs to generate its output.
The simulated neuron, which resembles a sewer system designed by a hyperactive
plumber, consists of 8000 cylindrical segments, with the smallest pipes
feeding into larger ones, then into larger ones still, until a single, large-bore
main line enters the cell body. Equations based on experiments with real
neurons describe how each seg-ment conducts electrical charges, given the
segment’s diameter, its position in the system, and other details. The simulation
is the largest, most complex single-neuron model ever built, says Bower,
‘and I know that for a fact, because we run it on the biggest computer on
the planet’.

Despite all the effort that goes into making such a simulation, the
model would be fundamentally boring if it merely mimicked the known behaviour
of real cells and suggested no unexpected new ideas about how they work,
says Bower. But that is not a problem he has encountered so far. ‘In this
case, as in every other case, something has fallen out of the model that
we had no way of anticipating,’ he says. For the Purkinje cell, the model
suggests synap-tic inputs from cells known as the parallel fibre system
– which most researchers thought were the main stimulus to the Purkinje
cell – may merely regulate the stimulus sent by other synapses.

Bower’s simulation of the piriform cortex has proved equally fruitful.
Bower’s research team, for example, used the model to guide their search
for the neurons responsible for developing memories of odours. Memories
develop when frequently used synapses increase in strength, and this reinforcement
makes the same sequence of neurons more likely to fire in the future. But
which synapses hold the memories of odours? Bower and his student Michael
Hasselmo experimented with the model, allowing reinforcement of synaptic
strength to occur along several different pathways in the piriform cortex
and looking at the result. They found that memory depended on reinforcement
occurring in one particular set of synapses. Almost at the same time, neurobiologist
Lewis Haberly of the University of Wisconsin published results of electrical
recordings from brain cell preparations showing that these synapses were
indeed capable of reinforcement.

Bower and Hasselmo’s model also helped resolve an odd paradox relating
to the learning of odours. In experiments with rat brain tissue, Bower
and Hasselmo observed that acetylcholine – a common brain chemical involved
in memory function – suppressed these same synapses rather than reinforcing
them. Puzzled, they incorporated these new results into their model and
began reexamining its behaviour. They found that acetylcholine’s inhibitory
effect did, indeed, contribute to learning, but in an unexpected way – by
muffling existing memories to keep them from interfering with those just
being learned.

Memory maintenance

This result could have far-reaching consequences, Bower suggests. If
this memory-muffler design is widespread in the brain, it could shed light
on the whole process of acquiring and maintaining memories. Indeed, says
Bower, afflictions such as Alzheimer’s disease are known to involve a disruption
in acetyl-choline production, so the model may provide insight into changes
that take place as this disease progresses.

Recently, Bower’s models have been leading him towards an unorthodox
view of the brain. Within the olfactory system, new results show few stable
patterns of activity. At one moment, a given odour – Chanel No. 5, say –
might produce a response in one set of cells; ten minutes later, the same
odour might stimulate an entirely different set. To Bower, this suggests
that neurons don’t simply gather information from below and pass it on up
the ladder to the brain. Instead, higher levels of the brain must actively
solicit certain bits of information and recruit neurons to provide it.

‘Most theories of brain function are what are called feed-forward theories,’
he says. ‘You receive the information and pass it on and on. At each step,
the information gets represented in a higher form. (But) our data suggests
that’s the wrong way to think about the brain. Expectation and context determine
what information the neurons are conducting. We’re basically model-driven
systems. We have this model running inside of what the world’s like, and
we collect data accordingly.’

The next step, Bower says, is to test this new conception with measurements
of electrical activity in real brains. Keeping track of such an ethereal,
always-changing flux of signals, however, will require recording the electrical
activity of several hundred neurons simultaneously. With colleagues at Stanford
University, he and his students are now developing a tiny computer chip
that will allow them to do just that.

Until then, however, Bower finds it difficult to convince colleagues
of his unorthodox ideas. Whether Bower’s ideas ultimately prove right or
wrong, however, the close interplay between virtual reality and the ‘wetware’
– the real brain – has already shown itself to be valuable and thought-provoking.

Bob Holmes is a freelance science writer based in Santa Cruz, California.