I haven't *seen* this posted anywhere yet (apologies if it has been,
I have that sneaking suspicion we are losing some articles through the feed)
but the information is tucked away in Nature, so I think it needs advertising!
**** PHD - a new mail server for protein secondary structure prediction ***
First the picture: for those of us who just run sequence through the
database and are then faced with one of Russ Doolittle's less attractive
options (Like "your protein does not match anything else yet seen in the
Universe" :) this is going to help you. But if you find that your protein is a
member of a family, and there is no homolgous 3D structure to work to, you are
kind of stuck with secondary structure predicitions that you *know* are only
going to be a little over half right, and you don't which half it's going to
be..... The better the prediction is going to be, the more useful it could be
to you. And if you have a measure for which parts are more likely to be
correct then that is going to help enormously.
So: and this is the end of a string of letters and papers that make
interesting reading in themselves; look at Nature, vol 360, page 540 (10th Dec
1992 issue) at the letter from Burkhard Rost and Chris Sander from EMBL
describing a new network server (PHD, Profile Neural Network Prediction) for
the prediction of secondary structures using information from multiple
To get started send HELP as the body of text to the server at
PredictProtein at EMBL-Heidelberg.DE, and go from there.
You will be instructed on the correct format to send your sequence,
which will then be matched against the database, multiply aligned, the
prediction of secondary structure calculated and the whole thing sent back.
The return times are reasonably fast, but maybe that's because I'm testing it
in the vacation? Overall round trip time for a 200 amino acid domain is about
30 minutes. Of course this could get a lot slower if everyone tries it at
One final point; the whole process will not work if your sequence has
less than 30% idenitity to another member of the database. You may be in the
position where your own favorite sequence is a member of a family, but at less
than 30%. In which case, you should be able to learn something using one of
the other family members; the results will not include 'your' protein, but if
you already have alignments that are satisfactory this may be useful.
I, for one, would be interested to know what sort of results people
get from the system, and how well the outupt compares with a range of known or
new 3D structures. Maybe as people get data they could feed this back to the
group? Unless there is preference from EMBL as to where to send feedback?
I hope I'm not stealing anyone's thunder by posting this, but it looks
too good to go unnoticed.
Well, many thanks to Burkhard and Chris, and New Years greetings to
Oxford University Molecular Biology Data Centre (nights and weekends)
Sir William Dunn School of Pathology (always)
jasper at vax.path.ox.ac.uk