30 November 2001 Kevin Karplus

The "dssp" directory contains networks, scripts, and quality reports
for neural nets attempting to predict the secondary structure defined
by DSSP from a multiple alignment.

There are also some networks for predicting DSSP 2ry sequence from the
STRIDE 2ry sequence, to see how much information is in the DSSP string
that is not in the STRIDE string.

Most of the scripts are still set up to refer to an older subdirectory
organization (in which the subdirectories of testing/dssp/ were just
subdirectories of testing/).

6 December 2001 Rachel Karchin

Using t2k-thin90 alignments, the best EBGHSTL network (predicting 7-class DSSP
from amino acid multiple alignment) is:
        dunbrack-in-scop-IDaa13-7-10-9-10-9-ebghstl-seeded-dssp-trained3.net

   null bits=2.39743

   alphabet   bits_saved   Q-n    SOV(E)  SOV(H)
     EBGHSTL     0.9357   0.6133  0.7973  0.8345

8 March 2002 Kevin Karplus

Did one training from dunbrack-1795-1+2 set, tested on dunbrack-1795-3:

# dunbrack-1795-1+2-IDaaHr-5-15-7-15-9-15-13-ebghstl-from-empty.net (6070 degrees of freedom)
          Bits_saved      Q7     SOV  object   SOV_E   SOV_B   SOV_G   SOV_H   SOV_S   SOV_T   SOV_C   
train	      1.0183  0.6563  0.6345  1.9918  0.8040  0.3495  0.3048  0.8534  0.1280  0.4721  0.5209
test	      0.8994  0.6290  0.6036  1.8302  0.7847  0.3308  0.2910  0.8310  0.1240  0.4206  0.4991

There are more bits saved here than for stride on the same test and
using the same training protocol, so DSSP seems to be more predictable
than STRIDE.