In article <Pine.A41.4.58.0310091049310.1531924 at mead11.u.washington.edu>,
E. Wijsman wrote:
> On Thu, 9 Oct 2003, Mulatz wrote:
>>> I'm looking for software to predict protein folding.
>>>> I have the sequence of a gene and I would like to view the protein that the
>> sequence codes for. I have searched the internet and I cannot find any
>> software (commercial or otherwise) that can predict the structure.
>>>> I a new to this type of software so I my have overlooked or misunderstood
>> the features of programs that I have already seen.
>>>> I would appreciate any suggestions as to which programs I should take a look
>> at.
>> This is one of the big unsolved problems in biology. There are programs
> to predict structure from sequence data but the results aren't very good,
> or rather, at this point they are pretty bad.
>> Ellen M. Wijsman
> Research Professor
With all due respect to Ellen Wijsman, the results are not uniformly bad.
Protein-structure prediction varies enormously in quality depending on
how similar a protein has already been solved experimentally. That
is, prediction is pretty good for very close homologs of known
structures and gets gradually worse as the sequence similarity decreases.
Detection of remote homology has gotten pretty good, so we can find
reasonable folds for about half the globular protein domains.
None of the predicted structures are of high enough quality to do drug
design, but if you are just trying to find out if you have a
TIM-barrel or a Rossmann fold, and to get some idea where the active
site is likely to be, protein structure prediction may be quite useful.
The best predictors do not all work identically, and they each have
their own strengths and weaknesses, so the most practical approach is
to hand your protein to a meta-predictor (I usually recommend the one
at http://BioInfo.PL/Meta/ but there are now several). The
meta-predictor hands your sequence off to many predictors and collects
the results. They try to add value by selecting among the predictions
to provide the one most likely to be good (the meta predictors vary a
lot in how much value they really add, but even the service of
collecting the results from several different servers is valuable).
In recent tests, the automated meta-predictors have been doing as well
as all but a handful of human predictors, and better than any single
primary server. Our limitation does not seem to be automation of
methods, but in doing predictions at all.
Disclaimer: my current main research effort is in protein-structure
prediction, but I don't create metaservers, just primary servers,
since I see the progress in prediction as depending more on
improvement in the primary servers.
--
Kevin Karplus karplus at soe.ucsc.eduhttp://www.soe.ucsc.edu/~karplus
Professor of Computer Engineering, University of California, Santa Cruz
Undergraduate and Graduate Director, Bioinformatics
Affiliations for identification only.
