For the refinement models, we finally thought of a way to keep the
models from collapsing---we copied in a peptide ligand from one of the
templates (GPTIEEVD from 1elw).  Although undertaker has no general
understanding of ligands, peptide ligands can be handled (we pretend
that we have a longer chain with a known break in the chain, just as
we do for multimers).

Optimization with the ligand in place was not quite enough to get a
good model---we had to add some constraints to keep the helices from
rearranging around the peptide ligand. 

Once we had done some optimization with the ligand, we tried removing
the constraints and found that the model was now stable without
constraints and without the ligand, so we did further optimization
unconstrained and without the ligand.


Refinement model 1 is T0368.try34-opt2 which scored the best with the
	try35 costfcn in undertaker.  It was optimized by undertaker
	without ligands primarily from
	try30-opt2.gromacs0.repack-nonPC.  try30-opt2 was optimized by
	undertaker from tr368 with a ligand added and constraints to
	hold the ligand in place.
	
Refinement model 2 is T0368.try35-opt2.gromacs0.repack-nonPC, which
	scores the best of any of the refinement models with Rosetta's
	energy function.  It was optimized by undertaker primarily
	from try34-opt2.gromacs0, then re-optimized with gromacs and
	with the sidechains (except PRO and CYS) repacked by rosetta.

Refinement model 3 is T0368.try31-opt2.gromacs0.  Of the models
	re-optimized by gromacs (other than try34-opt2 and try35-opt2)
	try31-opt2.gromacs0 scores the best with the try35
	cost function.  Try31-opt2 was optimized by undertaker with a
	ligand and constraints from try29-opt2, which in turn was
	optimized by undertaker from tr368 with a ligand added and
	constraints to hold the ligand in place.
	
Refinement model 4 is  T0368.try33-opt2.repack-nonPC.  Of the models
	with sidechains repacked by rosetta directly from an
	undertaker-optimized backbone (without re-optimization by gromacs),
	this model scored the best (other than try35-opt2.repack-nonPC).
	It was optimized with a ligand but without constraints.  The
	ligand moved substantially, and was no longer in the binding
	pocket, which is what clued us in that the ligand was no
	longer needed for the stability of the structure during optimization.

Refinement model 5 is  T0368.try32-opt2 is another decent choice.
	Try32-opt2 was optimized by undertaker with a
	ligand and constraints from try30-opt2, which in turn was
	optimized by undertaker from tr368 with a ligand added and
	constraints to hold the ligand in place.