We had enormous trouble getting the beta sheets to form for this target.
Eventually we added constraints taken from some of the Robetta
models, and seeded our optimization with the robetta models.

Without those models, we would probably not have been able to form any
but the simplest anti-parallel sheets.

Most of our effort went into playing with the strand pairing to try to
get a coherent beta sandwich for this protein.  I suspect that if we
knew protein structures the way Dr. Murzin does, we'd have been able
to recognize the structure as it started to from and gotten the strand
ordering right away, but we were forced to try something, look at how it
failed, then try something else, over and over.  There was not enough
time for this slow approach.

Model 2, from try16-opt2, is our second best-scoring model without
sheet constraints.  It has the long hairpin and helix predicted by our
secondary structure neural networks, but does not seem to make a
compact structure.