For training secondary structure predictors, DSSP is an adequate method.
Stride is about as good, though DSSP and Stride disagree on the
boundaries of helices.
Our group has taken to using more finely divided alphabets, since
there is more predictable from local context than DSSP defines, and
the extra information is useful in fold recognition and alignment.
Our most successful alphabet mainy subdivides the beta strands,
predicting parallel/antiparallel/mixed (see
http://www.soe.ucsc.edu/research/compbio/SAM_T06/STR.html for
information and definitions---str2 is our most successful local
structure alphabet, of the dozens we have tried). The str2 alphabet
can be defined based soley on the full DSSP output.
For burial/exposure, we have found neighborhood-count measures more
conserved and predictable than surface area measures. The simplest
one that works well is CB14 (the number of beta carbons within 14
Angstroms of the beta carbon), but we have another (near-backbone-11)
that is a little better.
------------------------------------------------------------
Kevin Karplus karplus from soe.ucsc.eduhttp://www.soe.ucsc.edu/~karplus
Professor of Biomolecular Engineering, University of California, Santa Cruz
Undergraduate Director, Bioinformatics
Affiliations for identification only.
