午夜福利1000集合

JOHANN GASTEIGER thinks that chemists work too hard. In the quest for new
compounds, they have to carry out endless reactions at the laboratory bench. So,
along with other researchers around the world, Gasteiger is training up a lab
assistant that will never tire. He鈥檚 teaching a computer how to do
chemistry.

Discovering how to make chemicals can be an arduous process. It鈥檚 all very
well isolating a new wonder drug from a rare Amazonian plant鈥攖he problems
start when you try to produce it artificially. For decades, chemists have gone
about this the hard way, slaving over hot reaction flasks trying to turn common
chemicals into the final product. Others spend their days scouring scientific
journals for short cuts, or trying to speed up existing routes.

Gasteiger and his colleagues are hoping to cut out all this hard graft by
designing software that predicts how molecules of different chemicals will
react. The software can even suggest the best solvent to use and the optimum
temperature for the reaction. Programs developed by other workers take a newly
discovered molecule, break it into simple pieces, and then tell you how to
create the larger molecule out of its component parts.

By using the programs together, researchers hope to take a drug or pesticide
from nature and work out a synthetic pathway to create it in the test tube. So
will the latest software force chemists to hang up their lab coats? 鈥淚 must
admit I do not like the idea that computer-assisted reaction and synthesis
planning might lead to making some chemists obsolete,鈥� says Gasteiger. 鈥淭here is
certainly such a fear among some chemists and I always do my best to dissipate
this angst.鈥�

In fact, the reality is that computers offer chemists the chance of using
their time much more productively and creatively, as well as bringing down the
cost of developing new drugs and materials.

Gasteiger and his colleagues at the Institute for Organic Chemistry at the
University of Erlangen-N眉rnberg in Germany have developed a neural network
that can predict how one molecule will react and combine with another. It all
comes down to the atoms that take part in joining two molecules together. These
atoms are known as the reaction centre.

Gasteiger tested his method by looking at one reaction centre鈥攖he
coming together of a carbon-carbon double bond with a hydrogen-carbon bond to
produce a chain of three carbons and a hydrogen linked by single bonds. There
are hundreds of reactions in which this happens, often with very different
reagents and conditions. This reaction centre crops up in so-called Michael
addition reactions, in Friedel-Crafts alkylation by alkene reactions, and even
in a photochemical reaction in the presence of oxygen.

So suppose chemists wanted to combine two molecules to create a new drug that
had the carbon-hydrogen chain. Should they try using a Michael addition or a
Friedel-Crafts reaction? With Gasteiger鈥檚 software, they could feed in the
theoretical details of their reaction centre and see which is most likely to
succeed. Naturally, this would cut the workload in the lab dramatically.

In the past, chemists manually searched huge databases of reactions and made
educated guesses about the behaviour of new molecules. But Gasteiger says this
method is inefficient. 鈥淎 single search can lead to a list of several hundred
reactions from a database that can contain millions, so manual analysis is both
laborious and time consuming.鈥�

Neural networks offer a quicker solution, but they must first be trained.
Gasteiger鈥檚 software learns from information about each reaction centre stored
in the huge databases, which also list data on optimum temperature, pressure,
pH, catalysts and reagents.

Learning curve

There are two approaches to teaching a neural network鈥攕upervised and
unsupervised learning. In the first, the network is presented with thousands of
reactions and told which ones work under which conditions. In the second method,
the system learns for itself. 鈥淲e prefer the unsupervised approach,鈥� Gasteiger
says.

What should it learn though? Reactions are usually classified by naming them
after their inventors鈥攖he Michael addition, the Wittig, the Beckman, the
Diels-Alder and so on. Yet this says very little about what controls the
interactions of the molecules involved.

Instead, Gasteiger and his colleagues picked just seven characteristic
properties of the atoms in the reaction centre, including their electrostatic
charges, how easily they are polarised and their ability to attract electrons.
They chose the properties that they thought were most significant in determining
which reaction would happen at the reaction centre.

Gasteiger selected 120 reactions that use the carbon-carbon plus
carbon-hydrogen reaction centre from databases. He then set about educating the
neural net with these reactions, and each of their seven variables, using what
is called a Kohonen network. This gets reactions that are similar in seven
dimensions鈥攖hat have similar sets of seven variables鈥攖o group
together on a 2D grid.

Gasteiger used a 12 脳 12 grid. Since the network must learn from scratch,
each grid square is initially assigned a random set of seven values,
corresponding to the seven properties of the reaction centre. One by one, each
reaction is placed into the square with the most similar set of properties to
its own. Then the properties of all the other squares are adjusted so that the
squares nearest the 鈥渨inning鈥� square have properties slightly closer to it.

The researchers did this in turn for each of the 120 reactions, and then
repeated the process several times鈥攖his took a mere 20 seconds on their
workstation. The repetitions of the teaching process fine-tune the grid,
creating a 2D landscape that reflects the relationships between all the
reactions
(see Diagram). Gasteiger鈥檚 team was pleased to discover that the
neural network sorted the reactions into groups which had already been
identified by chemists in the past鈥攁ll the Michael additions grouped
together, as did the Friedel-Crafts alkylation by alkene reactions. Reactions
far apart in the landscape were very different and isolated parts of the
landscape were uncommon reactions.

With the neural network trained, it can now be used to predict how an
untested compound with a carbon-carbon double bond will embrace another molecule
with a carbon-hydrogen group. When the seven properties of the new reaction
centre are fed in, the network decides where to place it on the map. If a
reaction is placed at the centre of the area of the map covered by Michael
additions, then it is very likely to react by a standard Michael addition. If it
is more isolated on the map, it will probably react in a more exotic way.

To test the network鈥檚 powers of prediction, Gasteiger and his colleagues fed
it details of reactions using the same reaction centre, but with molecules from
another database that the network had not seen before. Chemists had already
classified how these molecules would react鈥攙ia Michael additions or
whatever. The network gave a hit rate as high as 86 per cent.

This means chemists could use the system with some confidence to predict how
a new molecule containing a carbon-carbon double bond will react with a
carbon-hydrogen group in another molecule, and so avoid trying to carry out
reactions that are doomed to failure. Predicting reactions for different
reaction centres would require separate neural networks, trained for each
reaction centre.

Cameo role

Another program that predicts the outcome of reactions is CAMEO鈥攐r
computer-aided mechanistic evaluation of organic reactions. But this works in a
completely different way from Gasteiger鈥檚 network. CAMEO predicts the course of
a reaction by piecing together lots of fundamental processes, such as transfers
of electrons and hydrogen atoms between reacting molecules. 鈥淐AMEO avoids the
use of large databases of specific reaction classes,鈥� says its developer,
William Jorgensen of Yale University in New Haven, Connecticut. 鈥淩ather, it
assembles reactions from mechanistic steps such as addition and substitution of
atoms and groups.鈥�

CAMEO is like a collective brain of chemists past. It uses hundreds of rules
born out of decades of laboratory work. Baldwin鈥檚 rules, for example, predict
whether or not a carbon chain containing an oxygen or nitrogen atom will curl up
to form a ring or not. Cram鈥檚 rules tell the chemist how a carbon-oxygen double
bond in an aldehyde or ketone will react鈥攖his depends on the size of
neighbouring groups. Jorgensen says that most organic reactions are just
different combinations of various fundamental steps.

The chemist feeds the starting materials and reaction conditions鈥攕uch
as temperature and pH鈥攊nto CAMEO, and the program looks at its
rules to try to predict what might happen. Sometimes CAMEO predicts that a
particular reaction will generate no product at all because, say, Baldwin鈥檚
rules do not permit the reaction to occur. If this happens, chemists can use
CAMEO to simulate the reaction using a different solvent or at a higher
temperature to see if that would change its path or improve its yield of
product. Being able to do this on computer before setting foot in the lab saves
researchers time and chemicals.

Andrew Holmes of the University of Cambridge is all in favour of such
computerised help. But he thinks that human creativity will always have a part
to play in predicting how reactions will work. 鈥淲hat works for one molecule may
not necessarily translate into the optimum conditions for a related compound,鈥�
he says. In other words, computers cannot suggest anything new and simply rely
on what has worked in the past.

The ultimate machine for the synthetic chemist would be a computer that could
provide a complete recipe for cooking up a target molecule. Progress on such a
machine is already being made, thanks to pioneering work in the 1960s by Elias
J. Corey, now at Harvard University. Corey鈥檚 idea was to dismantle the target
molecule into simple pieces that could be made from off-the-shelf ingredients
with easy reactions. These pieces could then be fixed together in the reaction
flask to make the finished product.

This method has two main difficulties. The first is finding the reactions to
fix the pieces back together鈥攁 job for which CAMEO and Gasteiger鈥檚 neural
network might be handy. But the second problem is knowing where to snap apart
the target molecule in the first place.

Breaking up

To work this out, Corey developed a computer program called LHASA鈥攍ogic
and heuristics applied to synthetic analysis. 鈥淟HASA is a knowledge-based expert
system not a reaction database,鈥� according to Nigel Greene of LHASA UK, based at
the University of Leeds. It uses what he calls 鈥渢ransforms鈥� to describe classes
of chemical reactions. Each transform represents one way in which two molecules
can react鈥攆or example, joining a carbon-carbon double bond and a
carbon hydrogen bond using the Michael addition is one transform.

LHASA tries to break the target molecule in two pieces that can be joined
together again by a single transform. The software then tries to break apart the
pieces in the same way, until it gets down to simple, readily available
chemicals. These can then be reacted together in reverse order, via the series
of transforms, to produce the target molecule.

The potential number of paths could be enormous, says James Hendrickson of
Brandeis University, Waltham, Massachusetts. 鈥淭here are literally millions of
different routes possible, from different starting materials, to any substance
of interest,鈥� he says. It is rather astonishing, he thinks, how little chemists
really know about designing the best synthesis of a new molecule.

Hendrickson and his team have devised a program called SYNGEN, which tries to
find the shortest and simplest synthetic paths to a product using readily
available starting materials. Like LHASA, SYNGEN looks for ways to dismantle the
target molecule into small pieces. But while LHASA suggests the general classes
of reactions that could recombine the pieces and then lets the chemists work out
the details, SYNGEN gives specific reactions for combining one molecule with
another.

The program draws on a large reaction database to make a complete synthetic
path, replete with necessary reaction conditions such as the temperature,
pH and any catalyst needed. Early results suggest the computer is on the
right track. 鈥淚n a number of cases to date, the computer has generated the
current industrial routes to several pharmaceuticals,鈥� says Hendrickson. SYNGEN
has also proposed more efficient routes to numerous compounds such as lysergic
acid, the precursor to LSD and some medicinal drugs. Hendrickson and his team
are currently working on a more advanced version of the software.

So the future could perhaps see chemists huddled over computers feeding in
the structures of exotic drugs and chemicals to discover how to make them. But
might that mean fewer researchers are needed in the labs to explore possible
synthetic routes to compounds?

鈥淚 have little fear that synthetic chemistry will ever reach the stage where
one could sit at a computer terminal and plan the exact route and reaction
parameters . . . to obtain a chemical product,鈥� says Al Meyers of Colorado State
University at Fort Collins. He does believe that software has its role to play,
however. 鈥淭he synthesis programs will, on many occasions, save time in the
library searching for the various ways to transform one molecule or one
functional group into another.鈥� Software could also calculate the efficiencies,
and expected cost of alternate routes.

At Imperial College, London, organic chemist Tony Barrett doesn鈥檛 feel his
job is under threat either. 鈥淎t the time LHASA came out, I can clearly remember
all the angst amongst the weak in the synthetic community,鈥� he says. But the
program didn鈥檛 cause great upheaval. 鈥淚 doubt very much that any software will
lead to droves of synthetic chemists on welfare.鈥�