The SAM-T04 human predictions for CASP6 use a very similar method to
the SAM-T02 method in CASP5.

We start with a fully automated method:
   
    Use the SAM-T2K and SAM-T04 methods for finding homologs of the
      target and aligning them.

    Make local structure predictions using neural nets and the
      multiple alignments.  Different neural nets are used for the
      SAM-T2K alignments and the SAM-T04 alignments.
      We currently use 7 local-structure alphabets: 
        DSSP
	STRIDE
	STR2	an extended version of DSSP that splits the beta strands
		into multiple classes (parallel/antiparallel/mixed,
					edge/center)
	ALPHA	an discretization of the alpha torsion angle:
		CA(i-i), CA(i), CA(i+1), CA(i+2)
        BYS	a discretization of Ramachandran plots, due to Bystroff
	CB_burial_14_7	a 7-state discretization of the number of C_beta
		atoms in a 14 Angstrom radius sphere around the C_beta.
	DSSP_EHL2	CASP's collapse of the DSSP alphabet
			DSSP_EHL2 is not predicted directly by a
			neural net, but is computed as a weighted
			average of the other backbone alphabet predictions. 
    
    We make 2-track HMMs with each alphabet (1.0 amino acid + 0.3 local structure)
      and use them to score a template library of 6400 (t04) or 9900 (t2k) templates
      We also used a single-track HMM to score not just the template
      library, but a non-redundant copy of the entire PDB.

    We also made a few 3-track HMMs (AA, str2, CB_burial_14_7) for
      finding and aligning more remote homologs.
    
    One-track HMMs built from the template library multiple alignments
       were used to score the target sequence (for early targets, only
       t2k template library was searched this way).

    All the logs of e-values were combined in a weighted average (with
       rather arbitrary weights, since we did not have time to
       optimize them), and the best templates ranked.  Ranking was
       separate for predictions from the t2k and t04 multiple alignments.
    
    Alignments of the target to the top templates were made using
      several different alignment methods (mainly using the SAM hmmscore
      program, but a few alignments were made with Bob Edgar's MUSCLE
      profile-profile aligner).

    Generate fragments (short 9-residue alignments for each position)
      using SAM's "fragfinder" program and the 3-track HMM.

    Then the "undertaker" program (named because it optimizes burial)
      is used to try to combine the alignments and the fragments into
      a consistent 3D model.  No single alignment or parent template
      was used, though in many cases one had much more influence than
      the others.  The alignment scores were not passed to undertaker,
      but were used only to pick the set of alignments and fragments
      that undertaker would see.

After the initial automatic run was finished, the results were
examined by hand, and various tweaks were made to the undertaker
cost function to improve the models.  Many of the tweaks consisted of
adding specific strand pairings and distance constraints, to make the
model look better to us.

Undertaker uses a genetic algorithm with about 28 different operators
to minimize its cost function.  The cost function has many components,
including various definitions of burial and compactness, sidechain
rotamer preferences, steric clashes, chain breaks, predicted local
backbone conformation, hydrogen bonding, disulfide bonds, and user
specified constraints. The relative weights of these components were
tweaked for each target, as we have not found a generally applicable
set of weights.

For T0197, the sam-t2k and sam-t04 multiple alignments resulted in
different secondary structure predictions.  We preferred the sam-t04
predictions, which resulted in 8 strands.  These eight strands cried
out to be arranged into a barrel, and two of our top three
fold-recognition hits were for barrels (SCOP domain d.63.1,
exemplified by 1d8hiA and 1d8iA, SCOP domian b.60.1 exemplified by
1gm6A and 1ew3A) and the other formed a half-barrel (d.32.1
exemplified by 1cjxA and 1ewjA).

None of the automatically generated alignments were full-length,
generally getting only 3 or 4 strands in a sheet.  We did not like the
all-antiparallel barrel of b.60.1, thinking that the helices implied
some parallel strands, and tried creating an ab-initio model that
had parallel connections for strand-helix-strand connections, and
antiparallel for strand-strand connections, but keeping an
s1-s2-s3-s4-s5-s6-s7-s8 order around the barrel.  This failed, though
not spectacularly.  One of the attempts is reported as model 3.

Noticing that the attempt to form the barrel resulted in a cluster of
charged residues in the center of the barrel, we tried finding an
alignment to d.63.1, which also has such a cluster.  This involved
building an ab initio model with a guess at the strand pairing, doing
a structural alignment to 1d8iA, then modifying the alignment to flip
over strands that appeared to be facing the wrong way.  We tried also
putting in some bulges, but this may have been a mistake.

Had we not been overwhelmed by too many targets for CASP6, we could
probably have produced a more convincing model for this target, but
time pressure forced us to stop well before we were done.