In article <uvacj1h21cskff at corp.supernews.com>, Anton Karnoup wrote:
> Very immature statements.
>> The protein folding is a very tough problem, and the world's best minds are
> working on it for several years. It's not the fault of the scientists who
> are trying very hard to solve it: the problem fundamentally is very complex.
> Relativity theory was also not created overnight, but was a fruit of several
> years of hard work of a genius.
> Don't tell me "there is no beef there"; once the working universal folding
> theory is created - and I'm sure it will be created, maybe via several more
> years of hard theoretical and experimental work - it will enable lots of
> practical applications (and it will be a true revolution), from medicine to
> creation of new materials.
Actually we know the theory of protein folding. It's just quantum
mechanics, and there is no way to sample the energy function enough with
today's computers or even tomorrow's computers to solve the protein
folding problem for even the smallest stable proteins in acqueous
solution. This is also ignoring effects like solvents, ligands,
unstructured/flexible regions, disordered proteins, or proteins which
have multiple minima.
There is very little current success in folding up proteins using ab initio
(IE, physically based) force fields. The typical problems are the quality
of the force field and the amount of sampling that can be done in a reasonable
time. IBM's stated approach is to just throw much, much larger computers
at the problem, but this is not a cost-effective way to solve the problem.
It's often useful to have an accurate theory of how something works,
but it doesn't always mean you can simulate the theory accurately
enough to make good predictions. Heuristics are commonly used in protein
folding; some of these make little or no reference to the actual folding of
proteins but merely extract statistical information from known examples and
apply it to the protein of interest.
Most useful protein folding programs follow several methods:
1) finding a homologous or analgous protein and using it as a structural
template on which the target protein is modelled ("comparative modeling"
and "fold recognition"). This can be quite successful at predicting the general
features of the protein, although the fine details tend to be unspecified
or inaccurate. Also, if no templates exist the method is pretty useless.
2) fragment methods. These find small regions of homology/analogy between
parts of the target protein sequence and known fragments from existing
structures, and combine those fragments in either combinatorial or
random (esp. monte carlo) ways. Various pre- and post-processing is
done to filter out the obviously bad solutions, and the resulting predictions
are analyzed and sorted. This method is fairly effective at dealing with
small, stable single-domain proteins.
Dave
