SAM-T08-human
Kevin Karplus

The SAM-T08 hand predictions used methods similar to SAM_T06 in CASP7
and SAM_T04 in CASP6 [2].

We start with a fully automated method that was essentially the same
as the SAM-T08-server, though we froze the code for the server but had
several bug fixes and minor improvements for the version used in hand
prediction during the summer.  The automated method includes improved
neural networks for local structure prediction [3] and improved
residue-residue contact prediction (see SAM-T08-2stage) [5].
    
One major change for the method this time was the use of C-beta
distance constraints derived from the alignments to templates.  These
were used to select among the initial alignments and during at least
the first run of optimization.  The addition of these constraints kept
all-alpha structures correctly copied from alignments from being
pulled apart by the optimization---a problem that we had in CASP7 and
earlier experiments.

After the automatic prediction was done, we examined it visually and
tried to fix any flaws that we saw.  This generally involved rerunning
undertaker with new cost functions, increasing the weights for
features we wanted to see and decreasing the weights where we thought
the optimization has gone overboard.  Sometimes we added new templates
or removed ones that we thought were misleading the optimization process.
We often did "polishing" runs, where all the current models were read in
and optimization with undertaker's genetic algorithm was done with high
crossover.  These did not usually make much difference to the
appearance of the model, but often resolved small clashes or breaks.

Some improvements in undertaker since CASP7 include better
communication with SCWRL for initial model building form alignments
(now using the standard protocol that identical residues have fixed
rotamers, rather than being re-optimized by SCWRL), more cost functions
based on the neural net predictions, multiple constraint sets (for
easier weighting of the importance of different constraints), and some
new conformation-change operators (Backrub and BigBackrub).

We also created model-quality-assessment methods for CASP8, which we
applied to the server predictions to get metaserver results.  For each
target, we did two undertaker optimizations from the top 10 models
with two of the MQA methods (SAM-T08-MQAU and SAM-T08-MQAC [1,4]), and
considered these models as possible alternatives to our
natively-generated models.  For some of the targets, we did even more
meta-server runs, optimizing from some or all of the server models
with various cost functions.

For some targets, we tried breaking the protein up into domains, in an
attempt to get more structure searching for domains with few homologs,
avoiding contamination of the multiple alignments by neighboring domains.

All results, intermediate files, and working notes are available on
the web at http://www.soe.ucsc.edu/~karplus/casp8/

Note: for almost all the targets this summer "we" means Kevin
Karplus---students provided some assistance on only 9 targets:T0387,
T0388, T0419, T0437, T0443, T0465, T0476, T0484, and T0500.

References

1 John Archie and Kevin Karplus. Applying Undertaker Cost Functions to
  Model Quality Assessment. Proteins: Structure, Function, and Bioinformatics
  accepted. 

2 Kevin Karplus, Sol Katzman, George Shackelford, Martina Koeva,
  Jenny Draper, Bret Barnes, Marcia Soriano, and Richard
  Hughey. SAM-T04: what's new in protein-structure prediction for
  CASP6.  Proteins: Structure, Function, and Bioinformatics,
  2005. 61(S7):135-142. doi:doi:10.1002/prot.20730 

3 Sol Katzman, Christian Barrett, Grant Thiltgen, Rachel Karchin,
  and Kevin Karplus. Predict-2nd: a tool for generalized protein local
  structure prediction. Bioinformatics
  (advanced access 30 Aug 2008) Supplementary material
  doi:10.1093/bioinformatics/btn438

4 Martin Paluszewski and Kevin Karplus. Model Quality Assessment
  using Distance Constraints from Alignments. Proteins: Structure,
  Function, and Bioinformatics
  in press. 

5 George Shackelford and Kevin Karplus. Contact Prediction using
  Mutual Information and Neural Nets. Proteins: Structure, Function,
  and Bioinformatics, 69(S8):159-164, 2007. (CASP7 special issue).
  doi:10.1002/prot.21791